minix/servers/vm/region.c

2534 lines
63 KiB
C
Raw Normal View History

#define _SYSTEM 1
#include <minix/com.h>
#include <minix/callnr.h>
#include <minix/type.h>
#include <minix/config.h>
#include <minix/const.h>
#include <minix/sysutil.h>
#include <minix/syslib.h>
#include <minix/debug.h>
#include <minix/bitmap.h>
2010-10-15 11:10:14 +02:00
#include <minix/hash.h>
#include <machine/multiboot.h>
#include <sys/mman.h>
#include <limits.h>
#include <string.h>
#include <errno.h>
#include <assert.h>
#include <stdint.h>
#include <memory.h>
#include <sys/param.h>
#include "vm.h"
#include "proto.h"
#include "util.h"
#include "glo.h"
#include "region.h"
#include "sanitycheck.h"
#include "physravl.h"
#include "memlist.h"
/* LRU list. */
2012-03-25 20:25:53 +02:00
static yielded_t *lru_youngest = NULL, *lru_oldest = NULL;
/* Should a physblock be mapped writable? */
#define WRITABLE(r, pb) \
(((r)->flags & VR_WRITABLE) && \
(((r)->flags & (VR_DIRECT | VR_SHARED)) || \
(pb)->refcount == 1))
2012-03-25 20:25:53 +02:00
static int map_new_physblock(struct vmproc *vmp, struct vir_region
*region, vir_bytes offset, vir_bytes length, phys_bytes what, u32_t
allocflags, int written);
2012-03-25 20:25:53 +02:00
static int map_ph_writept(struct vmproc *vmp, struct vir_region *vr,
struct phys_region *pr);
2012-03-25 20:25:53 +02:00
static phys_bytes freeyieldednode(yielded_t *node, int freemem);
2012-03-25 20:25:53 +02:00
static struct vir_region *map_copy_region(struct vmproc *vmp, struct
vir_region *vr);
2012-03-25 20:25:53 +02:00
static struct phys_region *map_clone_ph_block(struct vmproc *vmp,
struct vir_region *region, struct phys_region *ph, physr_iter *iter);
#if SANITYCHECKS
2012-03-25 20:25:53 +02:00
static void lrucheck(void);
#endif
2010-10-15 11:10:14 +02:00
/* hash table of yielded blocks */
#define YIELD_HASHSIZE 65536
2012-03-25 20:25:53 +02:00
static yielded_avl vm_yielded_blocks[YIELD_HASHSIZE];
2010-10-15 11:10:14 +02:00
2012-03-25 20:25:53 +02:00
static int avl_inited = 0;
2010-10-15 11:10:14 +02:00
2012-03-25 20:25:53 +02:00
void map_region_init(void)
2010-10-15 11:10:14 +02:00
{
int h;
assert(!avl_inited);
for(h = 0; h < YIELD_HASHSIZE; h++)
yielded_init(&vm_yielded_blocks[h]);
avl_inited = 1;
}
2012-03-25 20:25:53 +02:00
static yielded_avl *get_yielded_avl(block_id_t id)
2010-10-15 11:10:14 +02:00
{
u32_t h;
assert(avl_inited);
hash_i_64(id.owner, id.id, h);
h = h % YIELD_HASHSIZE;
assert(h >= 0);
assert(h < YIELD_HASHSIZE);
return &vm_yielded_blocks[h];
}
2012-03-25 20:25:53 +02:00
static char *map_name(struct vir_region *vr)
{
static char name[100];
char *typename, *tag;
int type = vr->flags & (VR_ANON|VR_DIRECT);
switch(type) {
case VR_ANON:
typename = "anonymous";
break;
case VR_DIRECT:
typename = "direct";
break;
default:
panic("unknown mapping type: %d", type);
}
switch(vr->tag) {
case VRT_TEXT:
tag = "text";
break;
case VRT_STACK:
tag = "stack";
break;
case VRT_HEAP:
tag = "heap";
break;
case VRT_NONE:
tag = "untagged";
break;
default:
tag = "unknown tag value";
break;
}
sprintf(name, "%s, %s", typename, tag);
return name;
}
2012-03-25 20:25:53 +02:00
void map_printregion(struct vmproc *vmp, struct vir_region *vr)
{
physr_iter iter;
struct phys_region *ph;
printf("map_printmap: map_name: %s\n", map_name(vr));
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
printf("\t%lx (len 0x%lx, %lukB), %p\n",
vr->vaddr, vr->length, vr->length/1024, map_name(vr));
printf("\t\tphysblocks:\n");
physr_start_iter_least(vr->phys, &iter);
while((ph = physr_get_iter(&iter))) {
printf("\t\t@ %lx (refs %d): phys 0x%lx\n",
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
(vr->vaddr + ph->offset),
ph->ph->refcount, ph->ph->phys);
physr_incr_iter(&iter);
}
}
/*===========================================================================*
* map_printmap *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
void map_printmap(vmp)
struct vmproc *vmp;
{
struct vir_region *vr;
region_iter iter;
printf("memory regions in process %d:\n", vmp->vm_endpoint);
region_start_iter_least(&vmp->vm_regions_avl, &iter);
while((vr = region_get_iter(&iter))) {
map_printregion(vmp, vr);
region_incr_iter(&iter);
}
}
2012-03-25 20:25:53 +02:00
static struct vir_region *getnextvr(struct vir_region *vr)
{
struct vir_region *nextvr;
region_iter v_iter;
SLABSANE(vr);
region_start_iter(&vr->parent->vm_regions_avl, &v_iter, vr->vaddr, AVL_EQUAL);
assert(region_get_iter(&v_iter));
assert(region_get_iter(&v_iter) == vr);
region_incr_iter(&v_iter);
nextvr = region_get_iter(&v_iter);
if(!nextvr) return NULL;
SLABSANE(nextvr);
assert(vr->parent == nextvr->parent);
assert(vr->vaddr < nextvr->vaddr);
assert(vr->vaddr + vr->length <= nextvr->vaddr);
return nextvr;
}
#if SANITYCHECKS
/*===========================================================================*
* map_sanitycheck_pt *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static int map_sanitycheck_pt(struct vmproc *vmp,
struct vir_region *vr, struct phys_region *pr)
{
struct phys_block *pb = pr->ph;
int rw;
int r;
if(WRITABLE(vr, pb))
rw = PTF_WRITE;
else
2012-08-16 23:34:15 +02:00
rw = PTF_READ;
r = pt_writemap(vmp, &vmp->vm_pt, vr->vaddr + pr->offset,
pb->phys, VM_PAGE_SIZE, PTF_PRESENT | PTF_USER | rw, WMF_VERIFY);
if(r != OK) {
printf("proc %d phys_region 0x%lx sanity check failed\n",
vmp->vm_endpoint, pr->offset);
map_printregion(vmp, vr);
}
return r;
}
/*===========================================================================*
* map_sanitycheck *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
void map_sanitycheck(char *file, int line)
{
struct vmproc *vmp;
lrucheck();
/* Macro for looping over all physical blocks of all regions of
* all processes.
*/
#define ALLREGIONS(regioncode, physcode) \
for(vmp = vmproc; vmp < &vmproc[VMP_NR]; vmp++) { \
region_iter v_iter; \
struct vir_region *vr; \
if(!(vmp->vm_flags & VMF_INUSE)) \
continue; \
region_start_iter_least(&vmp->vm_regions_avl, &v_iter); \
while((vr = region_get_iter(&v_iter))) { \
physr_iter iter; \
struct phys_region *pr; \
regioncode; \
physr_start_iter_least(vr->phys, &iter); \
while((pr = physr_get_iter(&iter))) { \
physcode; \
physr_incr_iter(&iter); \
} \
region_incr_iter(&v_iter); \
} \
}
#define MYSLABSANE(s) MYASSERT(slabsane_f(__FILE__, __LINE__, s, sizeof(*(s))))
/* Basic pointers check. */
ALLREGIONS(MYSLABSANE(vr),MYSLABSANE(pr); MYSLABSANE(pr->ph);MYSLABSANE(pr->parent));
ALLREGIONS(/* MYASSERT(vr->parent == vmp) */,MYASSERT(pr->parent == vr););
/* Do counting for consistency check. */
ALLREGIONS(;,USE(pr->ph, pr->ph->seencount = 0;););
ALLREGIONS(;,USE(pr->ph, pr->ph->seencount++;);
if(pr->ph->seencount == 1) {
if(!(pr->parent->flags & VR_DIRECT)) {
MYASSERT(usedpages_add(pr->ph->phys,
VM_PAGE_SIZE) == OK);
}
}
);
/* Do consistency check. */
ALLREGIONS({ struct vir_region *nextvr = getnextvr(vr);
if(nextvr) {
MYASSERT(vr->vaddr < nextvr->vaddr);
MYASSERT(vr->vaddr + vr->length <= nextvr->vaddr);
}
}
MYASSERT(!(vr->vaddr % VM_PAGE_SIZE));,
if(pr->ph->refcount != pr->ph->seencount) {
map_printmap(vmp);
printf("ph in vr %p: 0x%lx refcount %u "
"but seencount %u\n",
vr, pr->offset,
pr->ph->refcount, pr->ph->seencount);
}
{
int n_others = 0;
struct phys_region *others;
if(pr->ph->refcount > 0) {
MYASSERT(pr->ph->firstregion);
if(pr->ph->refcount == 1) {
MYASSERT(pr->ph->firstregion == pr);
}
} else {
MYASSERT(!pr->ph->firstregion);
}
for(others = pr->ph->firstregion; others;
others = others->next_ph_list) {
MYSLABSANE(others);
MYASSERT(others->ph == pr->ph);
n_others++;
}
MYASSERT(pr->ph->refcount == n_others);
}
MYASSERT(pr->ph->refcount == pr->ph->seencount);
MYASSERT(!(pr->offset % VM_PAGE_SIZE)););
ALLREGIONS(,MYASSERT(map_sanitycheck_pt(vmp, vr, pr) == OK));
}
#define LRUCHECK lrucheck()
2012-03-25 20:25:53 +02:00
static void lrucheck(void)
{
yielded_t *list;
/* list is empty and ok if both ends point to null. */
if(!lru_youngest && !lru_oldest)
return;
/* if not, both should point to something. */
SLABSANE(lru_youngest);
SLABSANE(lru_oldest);
assert(!lru_youngest->younger);
assert(!lru_oldest->older);
for(list = lru_youngest; list; list = list->older) {
SLABSANE(list);
if(list->younger) {
SLABSANE(list->younger);
assert(list->younger->older == list);
} else assert(list == lru_youngest);
if(list->older) {
SLABSANE(list->older);
assert(list->older->younger == list);
} else assert(list == lru_oldest);
}
}
void blockstats(void)
{
yielded_t *list;
int blocks = 0;
phys_bytes mem = 0;
clock_t ticks;
int s;
s = getuptime(&ticks);
assert(s == OK);
LRUCHECK;
for(list = lru_youngest; list; list = list->older) {
mem += VM_PAGE_SIZE;
blocks++;
}
if(blocks > 0)
2010-07-05 15:58:57 +02:00
printf("%d blocks, %lukB; ", blocks, mem/1024);
printmemstats();
}
#else
#define LRUCHECK
#endif
/*=========================================================================*
* map_ph_writept *
*=========================================================================*/
2012-03-25 20:25:53 +02:00
static int map_ph_writept(struct vmproc *vmp, struct vir_region *vr,
struct phys_region *pr)
{
int rw;
struct phys_block *pb = pr->ph;
assert(!(vr->vaddr % VM_PAGE_SIZE));
assert(!(pr->offset % VM_PAGE_SIZE));
assert(pb->refcount > 0);
if(WRITABLE(vr, pb))
rw = PTF_WRITE;
else
2012-08-16 23:34:15 +02:00
rw = PTF_READ;
if(pt_writemap(vmp, &vmp->vm_pt, vr->vaddr + pr->offset,
pb->phys, VM_PAGE_SIZE, PTF_PRESENT | PTF_USER | rw,
#if SANITYCHECKS
!pr->written ? 0 :
#endif
WMF_OVERWRITE) != OK) {
printf("VM: map_writept: pt_writemap failed\n");
return ENOMEM;
}
#if SANITYCHECKS
USE(pr, pr->written = 1;);
#endif
return OK;
}
#define SLOT_FAIL ((vir_bytes) -1)
/*===========================================================================*
* region_find_slot_range *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static vir_bytes region_find_slot_range(struct vmproc *vmp,
vir_bytes minv, vir_bytes maxv, vir_bytes length)
{
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
struct vir_region *lastregion;
vir_bytes startv = 0;
int foundflag = 0;
region_iter iter;
SANITYCHECK(SCL_FUNCTIONS);
/* Length must be reasonable. */
assert(length > 0);
/* Special case: allow caller to set maxv to 0 meaning 'I want
* it to be mapped in right here.'
*/
if(maxv == 0) {
maxv = minv + length;
/* Sanity check. */
if(maxv <= minv) {
printf("region_find_slot: minv 0x%lx and bytes 0x%lx\n",
minv, length);
return SLOT_FAIL;
}
}
/* Basic input sanity checks. */
assert(!(length % VM_PAGE_SIZE));
if(minv >= maxv) {
printf("VM: 1 minv: 0x%lx maxv: 0x%lx length: 0x%lx\n",
minv, maxv, length);
}
assert(minv < maxv);
if(minv + length > maxv)
return SLOT_FAIL;
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
#define FREEVRANGE_TRY(rangestart, rangeend) { \
vir_bytes frstart = (rangestart), frend = (rangeend); \
frstart = MAX(frstart, minv); \
frend = MIN(frend, maxv); \
if(frend > frstart && (frend - frstart) >= length) { \
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
startv = frend-length; \
foundflag = 1; \
} }
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
#define FREEVRANGE(start, end) { \
assert(!foundflag); \
FREEVRANGE_TRY(((start)+VM_PAGE_SIZE), ((end)-VM_PAGE_SIZE)); \
if(!foundflag) { \
FREEVRANGE_TRY((start), (end)); \
} \
}
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
/* find region after maxv. */
region_start_iter(&vmp->vm_regions_avl, &iter, maxv, AVL_GREATER_EQUAL);
lastregion = region_get_iter(&iter);
if(!lastregion) {
/* This is the free virtual address space after the last region. */
region_start_iter(&vmp->vm_regions_avl, &iter, maxv, AVL_LESS);
lastregion = region_get_iter(&iter);
FREEVRANGE(lastregion ?
lastregion->vaddr+lastregion->length : 0, VM_DATATOP);
}
if(!foundflag) {
struct vir_region *vr;
while((vr = region_get_iter(&iter)) && !foundflag) {
struct vir_region *nextvr;
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
region_decr_iter(&iter);
nextvr = region_get_iter(&iter);
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
FREEVRANGE(nextvr ? nextvr->vaddr+nextvr->length : 0,
vr->vaddr);
}
}
if(!foundflag) {
printf("VM: region_find_slot: no 0x%lx bytes found for %d between 0x%lx and 0x%lx\n",
length, vmp->vm_endpoint, minv, maxv);
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
util_stacktrace();
return SLOT_FAIL;
}
/* However we got it, startv must be in the requested range. */
assert(startv >= minv);
assert(startv < maxv);
assert(startv + length <= maxv);
/* remember this position as a hint for next time. */
vmp->vm_region_top = startv + length;
return startv;
}
/*===========================================================================*
* region_find_slot *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static vir_bytes region_find_slot(struct vmproc *vmp,
vir_bytes minv, vir_bytes maxv, vir_bytes length)
{
vir_bytes v, hint = vmp->vm_region_top;
/* use the top of the last inserted region as a minv hint if
* possible. remember that a zero maxv is a special case.
*/
if(maxv && hint < maxv && hint >= minv) {
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
v = region_find_slot_range(vmp, minv, hint, length);
if(v != SLOT_FAIL)
return v;
}
return region_find_slot_range(vmp, minv, maxv, length);
}
struct vir_region *region_new(struct vmproc *vmp, vir_bytes startv, vir_bytes length, int flags)
{
physr_avl *phavl;
struct vir_region *newregion;
if(!(SLABALLOC(newregion))) {
printf("vm: region_new: could not allocate\n");
return NULL;
}
/* Fill in node details. */
USE(newregion,
newregion->vaddr = startv;
newregion->length = length;
newregion->flags = flags;
newregion->tag = VRT_NONE;
newregion->lower = newregion->higher = NULL;
newregion->parent = vmp;);
SLABALLOC(phavl);
if(!phavl) {
printf("VM: region_new: allocating phys avl failed\n");
SLABFREE(newregion);
return NULL;
}
USE(newregion, newregion->phys = phavl;);
physr_init(newregion->phys);
return newregion;
}
/*===========================================================================*
* map_page_region *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
struct vir_region *map_page_region(vmp, minv, maxv, length,
what, flags, mapflags)
struct vmproc *vmp;
vir_bytes minv;
vir_bytes maxv;
vir_bytes length;
vir_bytes what;
u32_t flags;
int mapflags;
{
struct vir_region *newregion;
vir_bytes startv;
assert(!(length % VM_PAGE_SIZE));
SANITYCHECK(SCL_FUNCTIONS);
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
if((flags & VR_CONTIG) && !(mapflags & MF_PREALLOC)) {
printf("map_page_region: can't make contiguous allocation without preallocating\n");
return NULL;
}
startv = region_find_slot(vmp, minv, maxv, length);
if (startv == SLOT_FAIL)
return NULL;
/* Now we want a new region. */
if(!(newregion = region_new(vmp, startv, length, flags))) {
printf("VM: map_page_region: allocating region failed\n");
return NULL;
}
/* If we know what we're going to map to, map it right away. */
if(what != MAP_NONE) {
assert(!(what % VM_PAGE_SIZE));
assert(!(startv % VM_PAGE_SIZE));
assert(!(mapflags & MF_PREALLOC));
if(map_new_physblock(vmp, newregion, 0, length,
what, PAF_CLEAR, 0) != OK) {
printf("VM: map_new_physblock failed\n");
USE(newregion,
SLABFREE(newregion->phys););
SLABFREE(newregion);
return NULL;
}
}
if((flags & VR_ANON) && (mapflags & MF_PREALLOC)) {
if(map_handle_memory(vmp, newregion, 0, length, 1) != OK) {
printf("VM: map_page_region: prealloc failed\n");
USE(newregion,
SLABFREE(newregion->phys););
SLABFREE(newregion);
return NULL;
}
}
/* Pre-allocations should be uninitialized, but after that it's a
* different story.
*/
USE(newregion, newregion->flags &= ~VR_UNINITIALIZED;);
/* Link it. */
region_insert(&vmp->vm_regions_avl, newregion);
#if SANITYCHECKS
assert(startv == newregion->vaddr);
{
struct vir_region *nextvr;
if((nextvr = getnextvr(newregion))) {
assert(newregion->vaddr < nextvr->vaddr);
}
}
#endif
SANITYCHECK(SCL_FUNCTIONS);
return newregion;
}
2012-03-25 20:25:53 +02:00
static struct phys_region *reset_physr_iter(struct vir_region *region,
physr_iter *iter, vir_bytes offset)
{
struct phys_region *ph;
physr_start_iter(region->phys, iter, offset, AVL_EQUAL);
ph = physr_get_iter(iter);
assert(ph);
assert(ph->offset == offset);
return ph;
}
/*===========================================================================*
* map_subfree *
*===========================================================================*/
static int map_subfree(struct vir_region *region,
vir_bytes start, vir_bytes len)
{
struct phys_region *pr;
physr_iter iter;
vir_bytes end = start+len;
int full = 0;
#if SANITYCHECKS
{
SLABSANE(region);
SLABSANE(region->phys);
physr_start_iter_least(region->phys, &iter);
while((pr = physr_get_iter(&iter))) {
struct phys_region *others;
struct phys_block *pb;
pb = pr->ph;
for(others = pb->firstregion; others;
others = others->next_ph_list) {
assert(others->ph == pb);
}
physr_incr_iter(&iter);
}
}
#endif
if(start == 0 && len == region->length)
full = 1;
2012-09-18 15:10:59 +02:00
physr_init_iter(&iter);
physr_start_iter(region->phys, &iter, start, AVL_GREATER_EQUAL);
while((pr = physr_get_iter(&iter))) {
physr_incr_iter(&iter);
if(pr->offset >= end)
break;
pb_unreferenced(region, pr, !full);
if(!full) {
physr_start_iter(region->phys, &iter,
pr->offset, AVL_GREATER_EQUAL);
}
SLABFREE(pr);
}
if(full)
physr_init(region->phys);
return OK;
}
/*===========================================================================*
* map_free *
*===========================================================================*/
static int map_free(struct vir_region *region)
{
int r;
if((r=map_subfree(region, 0, region->length)) != OK) {
printf("%d\n", __LINE__);
return r;
}
USE(region,
SLABFREE(region->phys););
SLABFREE(region);
return OK;
}
2010-10-15 11:10:14 +02:00
/*===========================================================================*
* yielded_block_cmp *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int yielded_block_cmp(struct block_id *id1, struct block_id *id2)
2010-10-15 11:10:14 +02:00
{
if(id1->owner < id2->owner)
return -1;
if(id1->owner > id2->owner)
return 1;
return cmp64(id1->id, id2->id);
}
/*===========================================================================*
* free_yielded_proc *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static vir_bytes free_yielded_proc(struct vmproc *vmp)
{
vir_bytes total = 0;
2010-10-15 11:10:14 +02:00
int h;
SANITYCHECK(SCL_FUNCTIONS);
/* Free associated regions. */
2010-10-15 11:10:14 +02:00
for(h = 0; h < YIELD_HASHSIZE && vmp->vm_yielded > 0; h++) {
yielded_t *yb;
yielded_iter iter;
yielded_avl *avl = &vm_yielded_blocks[h];
yielded_start_iter_least(avl, &iter);
while((yb = yielded_get_iter(&iter))) {
yielded_t *next_yb;
SLABSANE(yb);
yielded_incr_iter(&iter);
if(yb->id.owner != vmp->vm_endpoint)
continue;
next_yb = yielded_get_iter(&iter);
total += freeyieldednode(yb, 1);
/* the above removal invalidated our iter; restart it
* for the node we want to start at.
*/
if(!next_yb) break;
yielded_start_iter(avl, &iter, next_yb->id, AVL_EQUAL);
assert(yielded_get_iter(&iter) == next_yb);
}
}
return total;
}
2012-03-25 20:25:53 +02:00
static phys_bytes freeyieldednode(yielded_t *node, int freemem)
{
yielded_t *older, *younger, *removed;
2010-10-15 11:10:14 +02:00
yielded_avl *avl;
int p;
SLABSANE(node);
LRUCHECK;
/* Update LRU. */
younger = node->younger;
older = node->older;
if(younger) {
SLABSANE(younger);
assert(younger->older == node);
USE(younger, younger->older = node->older;);
} else {
assert(node == lru_youngest);
lru_youngest = node->older;
}
if(older) {
SLABSANE(older);
assert(older->younger == node);
USE(older, older->younger = node->younger;);
} else {
assert(node == lru_oldest);
lru_oldest = node->younger;
}
LRUCHECK;
/* Update AVL. */
2010-10-15 11:10:14 +02:00
if(vm_isokendpt(node->id.owner, &p) != OK)
panic("out of date owner of yielded block %d", node->id.owner);
avl = get_yielded_avl(node->id);
removed = yielded_remove(avl, node->id);
assert(removed == node);
2010-10-15 11:10:14 +02:00
assert(vmproc[p].vm_yielded > 0);
vmproc[p].vm_yielded--;
/* Free associated memory if requested. */
if(freemem) {
free_mem(ABS2CLICK(node->physaddr), node->pages);
}
/* Free node. */
SLABFREE(node);
return VM_PAGE_SIZE;
}
/*========================================================================*
* free_yielded *
*========================================================================*/
2012-03-25 20:25:53 +02:00
vir_bytes free_yielded(vir_bytes max_bytes)
{
/* PRIVATE yielded_t *lru_youngest = NULL, *lru_oldest = NULL; */
vir_bytes freed = 0;
int blocks = 0;
while(freed < max_bytes && lru_oldest) {
SLABSANE(lru_oldest);
freed += freeyieldednode(lru_oldest, 1);
blocks++;
}
return freed;
}
/*========================================================================*
* map_free_proc *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_free_proc(vmp)
struct vmproc *vmp;
{
struct vir_region *r;
while((r = region_search_root(&vmp->vm_regions_avl))) {
SANITYCHECK(SCL_DETAIL);
#if SANITYCHECKS
nocheck++;
#endif
region_remove(&vmp->vm_regions_avl, r->vaddr); /* For sanity checks. */
map_free(r);
#if SANITYCHECKS
nocheck--;
#endif
SANITYCHECK(SCL_DETAIL);
}
region_init(&vmp->vm_regions_avl);
/* Free associated yielded blocks. */
free_yielded_proc(vmp);
SANITYCHECK(SCL_FUNCTIONS);
return OK;
}
/*===========================================================================*
* map_lookup *
*===========================================================================*/
struct vir_region *map_lookup(vmp, offset, physr)
struct vmproc *vmp;
vir_bytes offset;
struct phys_region **physr;
{
struct vir_region *r;
SANITYCHECK(SCL_FUNCTIONS);
#if SANITYCHECKS
if(!region_search_root(&vmp->vm_regions_avl))
panic("process has no regions: %d", vmp->vm_endpoint);
#endif
if((r = region_search(&vmp->vm_regions_avl, offset, AVL_LESS_EQUAL))) {
vir_bytes ph;
if(offset >= r->vaddr && offset < r->vaddr + r->length) {
ph = offset - r->vaddr;
if(physr) {
*physr = physr_search(r->phys, ph, AVL_EQUAL);
assert((*physr)->offset == ph);
}
return r;
}
}
SANITYCHECK(SCL_FUNCTIONS);
return NULL;
}
2012-03-25 20:25:53 +02:00
static u32_t vrallocflags(u32_t flags)
{
u32_t allocflags = 0;
if(flags & VR_PHYS64K)
allocflags |= PAF_ALIGN64K;
if(flags & VR_LOWER16MB)
allocflags |= PAF_LOWER16MB;
if(flags & VR_LOWER1MB)
allocflags |= PAF_LOWER1MB;
if(flags & VR_CONTIG)
allocflags |= PAF_CONTIG;
return allocflags;
}
/*===========================================================================*
* map_new_physblock *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static int map_new_physblock(vmp, region, start_offset, length,
what_mem, allocflags, written)
struct vmproc *vmp;
struct vir_region *region;
vir_bytes start_offset;
vir_bytes length;
phys_bytes what_mem;
u32_t allocflags;
int written;
{
struct memlist *memlist, *ml;
int r;
vir_bytes mapped = 0;
vir_bytes offset = start_offset;
SANITYCHECK(SCL_FUNCTIONS);
assert(!(length % VM_PAGE_SIZE));
if((region->flags & VR_CONTIG) &&
(start_offset > 0 || length < region->length)) {
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
printf("VM: region length 0x%lx, offset 0x%lx length 0x%lx\n",
region->length, start_offset, length);
map_printmap(vmp);
printf("VM: map_new_physblock: non-full contig allocation requested\n");
return EFAULT;
}
/* Memory for new physical block. */
allocflags |= vrallocflags(region->flags);
if(allocflags & PAF_CONTIG) {
assert(what_mem == MAP_NONE);
if((what_mem = alloc_mem(length/VM_PAGE_SIZE, allocflags)) == NO_MEM) {
return ENOMEM;
}
what_mem = CLICK2ABS(what_mem);
allocflags &= ~PAF_CONTIG;
assert(what_mem != MAP_NONE);
}
if(!(memlist = alloc_mem_in_list(length, allocflags, what_mem))) {
printf("map_new_physblock: couldn't allocate\n");
return ENOMEM;
}
r = OK;
for(ml = memlist; ml; ml = ml->next) {
struct phys_region *newphysr = NULL;
struct phys_block *newpb = NULL;
/* Allocate things necessary for this chunk of memory. */
if(!(newpb = pb_new(ml->phys)) ||
!(newphysr = pb_reference(newpb, offset, region))) {
printf("map_new_physblock: no memory for the ph slabs\n");
2012-09-18 15:10:59 +02:00
assert(!newphysr);
if(newpb) SLABFREE(newpb);
r = ENOMEM;
break;
}
/* Update pagetable. */
if(map_ph_writept(vmp, region, newphysr) != OK) {
printf("map_new_physblock: map_ph_writept failed\n");
r = ENOMEM;
break;
}
offset += VM_PAGE_SIZE;
mapped += VM_PAGE_SIZE;
}
if(r != OK) {
offset = start_offset;
/* Things did not go well. Undo everything. */
for(ml = memlist; ml; ml = ml->next) {
struct phys_region *physr;
if((physr = physr_search(region->phys, offset,
AVL_EQUAL))) {
assert(physr->ph->refcount == 1);
pb_unreferenced(region, physr, 1);
SLABFREE(physr);
}
offset += VM_PAGE_SIZE;
}
} else assert(mapped == length);
/* Always clean up the memlist itself, even if everything
* worked we're not using the memlist nodes any more. And
* the memory they reference is either freed above or in use.
*/
free_mem_list(memlist, 0);
SANITYCHECK(SCL_FUNCTIONS);
return r;
}
/*===========================================================================*
* map_clone_ph_block *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static struct phys_region *map_clone_ph_block(vmp, region, ph, iter)
struct vmproc *vmp;
struct vir_region *region;
struct phys_region *ph;
physr_iter *iter;
{
vir_bytes offset;
u32_t allocflags;
phys_bytes physaddr;
struct phys_region *newpr;
int region_has_single_block;
int written = 0;
#if SANITYCHECKS
written = ph->written;
#endif
SANITYCHECK(SCL_FUNCTIONS);
/* Warning: this function will free the passed
* phys_region *ph and replace it (in the same offset)
* with another! So both the pointer to it
* and any iterators over the phys_regions in the vir_region
* will be invalid on successful return. (Iterators over
* the vir_region could be invalid on unsuccessful return too.)
*/
/* This is only to be done if there is more than one copy. */
assert(ph->ph->refcount > 1);
/* This function takes a physical block, copies its contents
* into newly allocated memory, and replaces the single physical
* block by one or more physical blocks with refcount 1 with the
* same contents as the original. In other words, a fragmentable
* version of map_copy_ph_block().
*/
/* Remember where and how much. */
offset = ph->offset;
physaddr = ph->ph->phys;
/* Now unlink the original physical block so we can replace
* it with new ones.
*/
SLABSANE(ph);
SLABSANE(ph->ph);
assert(ph->ph->refcount > 1);
pb_unreferenced(region, ph, 1);
assert(ph->ph->refcount >= 1);
SLABFREE(ph);
SANITYCHECK(SCL_DETAIL);
/* Put new free memory in. */
allocflags = vrallocflags(region->flags);
region_has_single_block = (offset == 0 && region->length == VM_PAGE_SIZE);
assert(region_has_single_block || !(allocflags & PAF_CONTIG));
assert(!(allocflags & PAF_CLEAR));
if(map_new_physblock(vmp, region, offset, VM_PAGE_SIZE,
MAP_NONE, allocflags, written) != OK) {
/* XXX original range now gone. */
printf("VM: map_clone_ph_block: map_new_physblock failed.\n");
return NULL;
}
/* Copy the block to the new memory.
* Can only fail if map_new_physblock didn't do what we asked.
*/
if(copy_abs2region(physaddr, region, offset, VM_PAGE_SIZE) != OK)
panic("copy_abs2region failed, no good reason for that");
newpr = physr_search(region->phys, offset, AVL_EQUAL);
assert(newpr);
assert(newpr->offset == offset);
if(iter) {
physr_start_iter(region->phys, iter, offset, AVL_EQUAL);
assert(physr_get_iter(iter) == newpr);
}
SANITYCHECK(SCL_FUNCTIONS);
return newpr;
}
/*===========================================================================*
* map_pf *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int map_pf(vmp, region, offset, write)
struct vmproc *vmp;
struct vir_region *region;
vir_bytes offset;
int write;
{
vir_bytes virpage;
struct phys_region *ph;
int r = OK;
assert(offset >= 0);
assert(offset < region->length);
assert(region->flags & VR_ANON);
assert(!(region->vaddr % VM_PAGE_SIZE));
virpage = offset - offset % VM_PAGE_SIZE;
SANITYCHECK(SCL_FUNCTIONS);
if((ph = physr_search(region->phys, offset, AVL_LESS_EQUAL)) &&
(ph->offset <= offset && offset < ph->offset + VM_PAGE_SIZE)) {
/* Pagefault in existing block. Do copy-on-write. */
assert(write);
assert(region->flags & VR_WRITABLE);
assert(ph->ph->refcount > 0);
if(WRITABLE(region, ph->ph)) {
r = map_ph_writept(vmp, region, ph);
if(r != OK)
printf("map_ph_writept failed\n");
} else {
if(ph->ph->refcount > 0
&& ph->ph->share_flag != PBSH_COW) {
printf("VM: write RO mapped pages.\n");
return EFAULT;
} else {
if(!map_clone_ph_block(vmp, region, ph, NULL))
r = ENOMEM;
}
}
} else {
/* Pagefault in non-existing block. Map in new block. */
if(map_new_physblock(vmp, region, virpage,
VM_PAGE_SIZE, MAP_NONE, PAF_CLEAR, 0) != OK) {
printf("map_new_physblock failed\n");
r = ENOMEM;
}
}
SANITYCHECK(SCL_FUNCTIONS);
if(r != OK) {
printf("VM: map_pf: failed (%d)\n", r);
return r;
}
#if SANITYCHECKS
if(OK != pt_checkrange(&vmp->vm_pt, region->vaddr+virpage,
VM_PAGE_SIZE, write)) {
panic("map_pf: pt_checkrange failed: %d", r);
}
#endif
return r;
}
2010-06-28 23:53:37 +02:00
/*===========================================================================*
* map_pin_memory *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int map_pin_memory(struct vmproc *vmp)
2010-06-28 23:53:37 +02:00
{
struct vir_region *vr;
2010-07-01 10:54:25 +02:00
int r;
region_iter iter;
region_start_iter_least(&vmp->vm_regions_avl, &iter);
2010-06-28 23:53:37 +02:00
/* Scan all memory regions. */
while((vr = region_get_iter(&iter))) {
2010-07-01 10:54:25 +02:00
/* Make sure region is mapped to physical memory and writable.*/
r = map_handle_memory(vmp, vr, 0, vr->length, 1);
if(r != OK) {
panic("map_pin_memory: map_handle_memory failed: %d", r);
2010-06-28 23:53:37 +02:00
}
region_incr_iter(&iter);
2010-06-28 23:53:37 +02:00
}
return OK;
}
/*===========================================================================*
* map_handle_memory *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int map_handle_memory(vmp, region, offset, length, write)
struct vmproc *vmp;
struct vir_region *region;
vir_bytes offset, length;
int write;
{
struct phys_region *physr, *nextphysr;
int changes = 0;
physr_iter iter;
u32_t allocflags = 0;
if(!(region->flags & VR_UNINITIALIZED)) {
allocflags = PAF_CLEAR;
}
#define FREE_RANGE_HERE(er1, er2) { \
struct phys_region *r1 = (er1), *r2 = (er2); \
vir_bytes start = offset, end = offset + length; \
if(r1) { \
start = MAX(start, r1->offset + VM_PAGE_SIZE); } \
if(r2) { \
end = MIN(end, r2->offset); } \
if(start < end) { \
SANITYCHECK(SCL_DETAIL); \
if(map_new_physblock(vmp, region, start, \
end-start, MAP_NONE, allocflags, 0) != OK) { \
SANITYCHECK(SCL_DETAIL); \
return ENOMEM; \
} \
changes++; \
} }
SANITYCHECK(SCL_FUNCTIONS);
assert(region->flags & VR_ANON);
assert(!(region->vaddr % VM_PAGE_SIZE));
assert(!(offset % VM_PAGE_SIZE));
assert(!(length % VM_PAGE_SIZE));
assert(!write || (region->flags & VR_WRITABLE));
physr_start_iter(region->phys, &iter, offset, AVL_LESS_EQUAL);
physr = physr_get_iter(&iter);
if(!physr) {
physr_start_iter(region->phys, &iter, offset, AVL_GREATER_EQUAL);
physr = physr_get_iter(&iter);
}
FREE_RANGE_HERE(NULL, physr);
if(physr) {
physr = reset_physr_iter(region, &iter, physr->offset);
if(physr->offset + VM_PAGE_SIZE <= offset) {
physr_incr_iter(&iter);
physr = physr_get_iter(&iter);
FREE_RANGE_HERE(NULL, physr);
if(physr) {
physr = reset_physr_iter(region, &iter,
physr->offset);
}
}
}
while(physr) {
int r;
SANITYCHECK(SCL_DETAIL);
if(write) {
assert(physr->ph->refcount > 0);
if(!WRITABLE(region, physr->ph)) {
if(!(physr = map_clone_ph_block(vmp, region,
physr, &iter))) {
printf("VM: map_handle_memory: no copy\n");
return ENOMEM;
}
changes++;
} else {
SANITYCHECK(SCL_DETAIL);
if((r=map_ph_writept(vmp, region, physr)) != OK) {
printf("VM: map_ph_writept failed\n");
return r;
}
changes++;
SANITYCHECK(SCL_DETAIL);
}
}
SANITYCHECK(SCL_DETAIL);
physr_incr_iter(&iter);
nextphysr = physr_get_iter(&iter);
FREE_RANGE_HERE(physr, nextphysr);
SANITYCHECK(SCL_DETAIL);
if(nextphysr) {
if(nextphysr->offset >= offset + length)
break;
nextphysr = reset_physr_iter(region, &iter,
nextphysr->offset);
}
physr = nextphysr;
}
SANITYCHECK(SCL_FUNCTIONS);
if(changes < 1) {
#if VERBOSE
printf("region start at 0x%lx offset 0x%lx len 0x%lx write %d\n",
region->vaddr, offset, length, write);
printf("no changes in map_handle_memory\n");
#endif
return EFAULT;
}
#if SANITYCHECKS
if(OK != pt_checkrange(&vmp->vm_pt, region->vaddr+offset, length, write)) {
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
printf("handle mem 0x%lx-0x%lx failed\n",
region->vaddr+offset,region->vaddr+offset+length);
map_printregion(vmp, region);
panic("checkrange failed");
}
#endif
return OK;
}
#if SANITYCHECKS
static int count_phys_regions(struct vir_region *vr)
{
int n = 0;
struct phys_region *ph;
physr_iter iter;
physr_start_iter_least(vr->phys, &iter);
while((ph = physr_get_iter(&iter))) {
n++;
physr_incr_iter(&iter);
}
return n;
}
#endif
/*===========================================================================*
* map_copy_region *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static struct vir_region *map_copy_region(struct vmproc *vmp, struct vir_region *vr)
{
/* map_copy_region creates a complete copy of the vir_region
* data structure, linking in the same phys_blocks directly,
* but all in limbo, i.e., the caller has to link the vir_region
* to a process. Therefore it doesn't increase the refcount in
* the phys_block; the caller has to do this once it's linked.
* The reason for this is to keep the sanity checks working
* within this function.
*/
struct vir_region *newvr;
struct phys_region *ph;
physr_iter iter;
#if SANITYCHECKS
int cr;
cr = count_phys_regions(vr);
#endif
if(!(newvr = region_new(vr->parent, vr->vaddr, vr->length, vr->flags)))
return NULL;
physr_start_iter_least(vr->phys, &iter);
while((ph = physr_get_iter(&iter))) {
struct phys_region *newph = pb_reference(ph->ph, ph->offset, newvr);
if(!newph) { map_free(newvr); return NULL; }
#if SANITYCHECKS
USE(newph, newph->written = 0;);
assert(count_phys_regions(vr) == cr);
#endif
physr_incr_iter(&iter);
}
#if SANITYCHECKS
assert(count_phys_regions(vr) == count_phys_regions(newvr));
#endif
return newvr;
}
/*===========================================================================*
* copy_abs2region *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int copy_abs2region(phys_bytes abs, struct vir_region *destregion,
phys_bytes offset, phys_bytes len)
{
assert(destregion);
assert(destregion->phys);
while(len > 0) {
phys_bytes sublen, suboffset;
struct phys_region *ph;
assert(destregion);
assert(destregion->phys);
if(!(ph = physr_search(destregion->phys, offset, AVL_LESS_EQUAL))) {
printf("VM: copy_abs2region: no phys region found (1).\n");
return EFAULT;
}
assert(ph->offset <= offset);
if(ph->offset+VM_PAGE_SIZE <= offset) {
printf("VM: copy_abs2region: no phys region found (2).\n");
return EFAULT;
}
suboffset = offset - ph->offset;
assert(suboffset < VM_PAGE_SIZE);
sublen = len;
if(sublen > VM_PAGE_SIZE - suboffset)
sublen = VM_PAGE_SIZE - suboffset;
assert(suboffset + sublen <= VM_PAGE_SIZE);
if(ph->ph->refcount != 1) {
printf("VM: copy_abs2region: refcount not 1.\n");
return EFAULT;
}
if(sys_abscopy(abs, ph->ph->phys + suboffset, sublen) != OK) {
printf("VM: copy_abs2region: abscopy failed.\n");
return EFAULT;
}
abs += sublen;
offset += sublen;
len -= sublen;
}
return OK;
}
/*=========================================================================*
* map_writept *
*=========================================================================*/
2012-03-25 20:25:53 +02:00
int map_writept(struct vmproc *vmp)
{
struct vir_region *vr;
struct phys_region *ph;
int r;
region_iter v_iter;
region_start_iter_least(&vmp->vm_regions_avl, &v_iter);
while((vr = region_get_iter(&v_iter))) {
physr_iter ph_iter;
physr_start_iter_least(vr->phys, &ph_iter);
while((ph = physr_get_iter(&ph_iter))) {
physr_incr_iter(&ph_iter);
/* If this phys block is shared as SMAP, then do
* not update the page table. */
if(ph->ph->refcount > 1
&& ph->ph->share_flag == PBSH_SMAP) {
continue;
}
if((r=map_ph_writept(vmp, vr, ph)) != OK) {
printf("VM: map_writept: failed\n");
return r;
}
}
region_incr_iter(&v_iter);
}
return OK;
}
/*========================================================================*
* map_proc_copy *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_proc_copy(dst, src)
struct vmproc *dst;
struct vmproc *src;
{
/* Copy all the memory regions from the src process to the dst process. */
region_init(&dst->vm_regions_avl);
return map_proc_copy_from(dst, src, NULL);
}
/*========================================================================*
* map_proc_copy_from *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_proc_copy_from(dst, src, start_src_vr)
struct vmproc *dst;
struct vmproc *src;
struct vir_region *start_src_vr;
{
struct vir_region *vr;
region_iter v_iter;
if(!start_src_vr)
start_src_vr = region_search_least(&src->vm_regions_avl);
assert(start_src_vr);
assert(start_src_vr->parent == src);
region_start_iter(&src->vm_regions_avl, &v_iter,
start_src_vr->vaddr, AVL_EQUAL);
assert(region_get_iter(&v_iter) == start_src_vr);
/* Copy source regions after the destination's last region (if any). */
SANITYCHECK(SCL_FUNCTIONS);
while((vr = region_get_iter(&v_iter))) {
physr_iter iter_orig, iter_new;
struct vir_region *newvr;
struct phys_region *orig_ph, *new_ph;
if(!(newvr = map_copy_region(dst, vr))) {
map_free_proc(dst);
return ENOMEM;
}
USE(newvr, newvr->parent = dst;);
region_insert(&dst->vm_regions_avl, newvr);
physr_start_iter_least(vr->phys, &iter_orig);
physr_start_iter_least(newvr->phys, &iter_new);
while((orig_ph = physr_get_iter(&iter_orig))) {
struct phys_block *pb;
new_ph = physr_get_iter(&iter_new);
/* Check two physregions both are nonnull,
* are different, and match physblocks.
*/
assert(new_ph);
assert(orig_ph);
assert(orig_ph != new_ph);
pb = orig_ph->ph;
assert(orig_ph->ph == new_ph->ph);
/* If the phys block has been shared as SMAP,
* do the regular copy. */
if(pb->refcount > 2 && pb->share_flag == PBSH_SMAP) {
map_clone_ph_block(dst, newvr,new_ph,
&iter_new);
} else {
USE(pb, pb->share_flag = PBSH_COW;);
}
/* Get next new physregion */
physr_incr_iter(&iter_orig);
physr_incr_iter(&iter_new);
}
assert(!physr_get_iter(&iter_new));
region_incr_iter(&v_iter);
}
map_writept(src);
map_writept(dst);
SANITYCHECK(SCL_FUNCTIONS);
return OK;
}
int map_region_extend_upto_v(struct vmproc *vmp, vir_bytes v)
{
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
vir_bytes offset = v, end;
struct vir_region *vr, *nextvr;
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
int r = OK;
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
if(!(vr = region_search(&vmp->vm_regions_avl, offset, AVL_LESS))) {
printf("VM: nothing to extend\n");
return ENOMEM;
}
if(!(vr->flags & VR_ANON)) {
printf("VM: memory range to extend not anonymous\n");
return ENOMEM;
}
assert(vr->vaddr <= offset);
if((nextvr = getnextvr(vr))) {
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
assert(offset <= nextvr->vaddr);
}
end = vr->vaddr + vr->length;
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
offset = roundup(offset, VM_PAGE_SIZE);
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
if(end < offset)
r = map_region_extend(vmp, vr, offset - end);
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
return r;
}
/*========================================================================*
* map_region_extend *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_region_extend(struct vmproc *vmp, struct vir_region *vr,
vir_bytes delta)
{
vir_bytes end;
struct vir_region *nextvr;
assert(vr);
assert(vr->flags & VR_ANON);
assert(!(delta % VM_PAGE_SIZE));
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
if(vr->flags & VR_CONTIG) {
printf("VM: can't grow contig region\n");
return EFAULT;
}
if(!delta) return OK;
end = vr->vaddr + vr->length;
assert(end >= vr->vaddr);
if(end + delta <= end) {
printf("VM: strange delta 0x%lx\n", delta);
return ENOMEM;
}
nextvr = getnextvr(vr);
if(!nextvr || end + delta <= nextvr->vaddr) {
USE(vr, vr->length += delta;);
return OK;
}
return ENOMEM;
}
/*========================================================================*
* map_region_shrink *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_region_shrink(struct vir_region *vr, vir_bytes delta)
{
assert(vr);
assert(vr->flags & VR_ANON);
assert(!(delta % VM_PAGE_SIZE));
#if 0
printf("VM: ignoring region shrink\n");
#endif
return OK;
}
2012-03-25 20:25:53 +02:00
struct vir_region *map_region_lookup_tag(vmp, tag)
struct vmproc *vmp;
u32_t tag;
{
struct vir_region *vr;
region_iter v_iter;
region_start_iter_least(&vmp->vm_regions_avl, &v_iter);
while((vr = region_get_iter(&v_iter))) {
if(vr->tag == tag)
return vr;
region_incr_iter(&v_iter);
}
return NULL;
}
2012-03-25 20:25:53 +02:00
void map_region_set_tag(struct vir_region *vr, u32_t tag)
{
USE(vr, vr->tag = tag;);
}
2012-03-25 20:25:53 +02:00
u32_t map_region_get_tag(struct vir_region *vr)
{
return vr->tag;
}
/*========================================================================*
* map_unmap_region *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_unmap_region(struct vmproc *vmp, struct vir_region *r,
vir_bytes offset, vir_bytes len)
{
/* Shrink the region by 'len' bytes, from the start. Unreference
* memory it used to reference if any.
*/
vir_bytes regionstart;
SANITYCHECK(SCL_FUNCTIONS);
if(offset+len > r->length || (len % VM_PAGE_SIZE)) {
printf("VM: bogus length 0x%lx\n", len);
return EINVAL;
}
if(!(r->flags & (VR_ANON|VR_DIRECT))) {
printf("VM: only unmap anonymous or direct memory\n");
return EINVAL;
}
regionstart = r->vaddr + offset;
/* unreference its memory */
map_subfree(r, offset, len);
/* if unmap was at start/end of this region, it actually shrinks */
if(offset == 0) {
struct phys_region *pr;
physr_iter iter;
region_remove(&vmp->vm_regions_avl, r->vaddr);
USE(r,
r->vaddr += len;
r->length -= len;);
region_insert(&vmp->vm_regions_avl, r);
/* vaddr has increased; to make all the phys_regions
* point to the same addresses, make them shrink by the
* same amount.
*/
2012-09-18 15:10:59 +02:00
physr_init_iter(&iter);
physr_start_iter(r->phys, &iter, offset, AVL_GREATER_EQUAL);
while((pr = physr_get_iter(&iter))) {
assert(pr->offset >= offset);
USE(pr, pr->offset -= len;);
physr_incr_iter(&iter);
}
} else if(offset + len == r->length) {
assert(len <= r->length);
r->length -= len;
}
if(r->length == 0) {
/* Whole region disappears. Unlink and free it. */
region_remove(&vmp->vm_regions_avl, r->vaddr);
map_free(r);
}
SANITYCHECK(SCL_DETAIL);
if(pt_writemap(vmp, &vmp->vm_pt, regionstart,
MAP_NONE, len, 0, WMF_OVERWRITE) != OK) {
printf("VM: map_unmap_region: pt_writemap failed\n");
return ENOMEM;
}
SANITYCHECK(SCL_FUNCTIONS);
return OK;
}
/*========================================================================*
* map_remap *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_remap(struct vmproc *dvmp, vir_bytes da, size_t size,
struct vir_region *region, vir_bytes *r, int readonly)
{
struct vir_region *vr;
struct phys_region *ph;
vir_bytes startv, dst_addr;
physr_iter iter;
SANITYCHECK(SCL_FUNCTIONS);
assert(region->flags & VR_SHARED);
/* da is handled differently */
if (!da)
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
dst_addr = 0;
else
dst_addr = da;
/* round up to page size */
assert(!(size % VM_PAGE_SIZE));
startv = region_find_slot(dvmp, dst_addr, VM_DATATOP, size);
if (startv == SLOT_FAIL) {
return ENOMEM;
}
/* when the user specifies the address, we cannot change it */
if (da && (startv != dst_addr))
return EINVAL;
vr = map_copy_region(dvmp, region);
if(!vr)
return ENOMEM;
USE(vr,
vr->vaddr = startv;
vr->length = size;
vr->flags = region->flags;
vr->tag = VRT_NONE;
vr->parent = dvmp;
if(readonly) {
vr->flags &= ~VR_WRITABLE;
}
);
assert(vr->flags & VR_SHARED);
region_insert(&dvmp->vm_regions_avl, vr);
physr_start_iter_least(vr->phys, &iter);
while((ph = physr_get_iter(&iter))) {
if(map_ph_writept(dvmp, vr, ph) != OK) {
panic("map_remap: map_ph_writept failed");
}
physr_incr_iter(&iter);
}
*r = startv;
SANITYCHECK(SCL_FUNCTIONS);
return OK;
}
/*========================================================================*
* map_get_phys *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_get_phys(struct vmproc *vmp, vir_bytes addr, phys_bytes *r)
{
struct vir_region *vr;
struct phys_region *ph;
physr_iter iter;
if (!(vr = map_lookup(vmp, addr, NULL)) ||
(vr->vaddr != addr))
return EINVAL;
if (!(vr->flags & VR_SHARED))
return EINVAL;
physr_start_iter_least(vr->phys, &iter);
ph = physr_get_iter(&iter);
assert(ph);
assert(ph->ph);
if (r)
*r = ph->ph->phys;
return OK;
}
/*========================================================================*
* map_get_ref *
*========================================================================*/
2012-03-25 20:25:53 +02:00
int map_get_ref(struct vmproc *vmp, vir_bytes addr, u8_t *cnt)
{
struct vir_region *vr;
struct phys_region *ph;
physr_iter iter;
if (!(vr = map_lookup(vmp, addr, NULL)) ||
(vr->vaddr != addr))
return EINVAL;
if (!(vr->flags & VR_SHARED))
return EINVAL;
physr_start_iter_least(vr->phys, &iter);
ph = physr_get_iter(&iter);
assert(ph);
assert(ph->ph);
if (cnt)
*cnt = ph->ph->refcount;
return OK;
}
2010-09-14 23:22:56 +02:00
/*========================================================================*
* get_stats_info *
*========================================================================*/
2012-03-25 20:25:53 +02:00
void get_stats_info(struct vm_stats_info *vsi)
2010-09-14 23:22:56 +02:00
{
yielded_t *yb;
vsi->vsi_cached = 0L;
for(yb = lru_youngest; yb; yb = yb->older)
vsi->vsi_cached++;
2010-09-14 23:22:56 +02:00
}
void get_usage_info_kernel(struct vm_usage_info *vui)
{
memset(vui, 0, sizeof(*vui));
vui->vui_total = kernel_boot_info.kernel_allocated_bytes;
}
static void get_usage_info_vm(struct vm_usage_info *vui)
{
memset(vui, 0, sizeof(*vui));
vui->vui_total = kernel_boot_info.vm_allocated_bytes +
get_vm_self_pages() * VM_PAGE_SIZE;
}
2010-01-19 22:00:20 +01:00
/*========================================================================*
* get_usage_info *
*========================================================================*/
2012-03-25 20:25:53 +02:00
void get_usage_info(struct vmproc *vmp, struct vm_usage_info *vui)
2010-01-19 22:00:20 +01:00
{
struct vir_region *vr;
physr_iter iter;
struct phys_region *ph;
region_iter v_iter;
region_start_iter_least(&vmp->vm_regions_avl, &v_iter);
2010-01-19 22:00:20 +01:00
memset(vui, 0, sizeof(*vui));
if(vmp->vm_endpoint == VM_PROC_NR) {
get_usage_info_vm(vui);
return;
}
if(vmp->vm_endpoint < 0) {
get_usage_info_kernel(vui);
return;
}
while((vr = region_get_iter(&v_iter))) {
2010-01-19 22:00:20 +01:00
physr_start_iter_least(vr->phys, &iter);
while((ph = physr_get_iter(&iter))) {
/* All present pages are counted towards the total. */
vui->vui_total += VM_PAGE_SIZE;
2010-01-19 22:00:20 +01:00
if (ph->ph->refcount > 1) {
/* Any page with a refcount > 1 is common. */
vui->vui_common += VM_PAGE_SIZE;
2010-01-19 22:00:20 +01:00
/* Any common, non-COW page is shared. */
if (vr->flags & VR_SHARED ||
ph->ph->share_flag == PBSH_SMAP)
vui->vui_shared += VM_PAGE_SIZE;
2010-01-19 22:00:20 +01:00
}
physr_incr_iter(&iter);
}
region_incr_iter(&v_iter);
2010-01-19 22:00:20 +01:00
}
}
/*===========================================================================*
* get_region_info *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int get_region_info(struct vmproc *vmp, struct vm_region_info *vri,
2010-01-19 22:00:20 +01:00
int max, vir_bytes *nextp)
{
struct vir_region *vr;
vir_bytes next;
int count;
region_iter v_iter;
2010-01-19 22:00:20 +01:00
next = *nextp;
if (!max) return 0;
region_start_iter(&vmp->vm_regions_avl, &v_iter, next, AVL_GREATER_EQUAL);
if(!(vr = region_get_iter(&v_iter))) return 0;
2010-01-19 22:00:20 +01:00
for(count = 0; (vr = region_get_iter(&v_iter)) && count < max; count++, vri++) {
struct phys_region *ph1, *ph2;
/* Report part of the region that's actually in use. */
/* Get first and last phys_regions, if any */
ph1 = physr_search_least(vr->phys);
ph2 = physr_search_greatest(vr->phys);
if(!ph1 || !ph2) { assert(!ph1 && !ph2); continue; }
/* Report start+length of region starting from lowest use. */
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
vri->vri_addr = vr->vaddr + ph1->offset;
vri->vri_prot = 0;
vri->vri_length = ph2->offset + VM_PAGE_SIZE - ph1->offset;
2010-01-19 22:00:20 +01:00
/* "AND" the provided protection with per-page protection. */
if (!(vr->flags & VR_WRITABLE))
vri->vri_prot &= ~PROT_WRITE;
vri->vri_flags = (vr->flags & VR_SHARED) ? MAP_IPC_SHARED : 0;
2010-01-19 22:00:20 +01:00
next = vr->vaddr + vr->length;
region_incr_iter(&v_iter);
2010-01-19 22:00:20 +01:00
}
*nextp = next;
return count;
}
/*========================================================================*
* regionprintstats *
*========================================================================*/
2012-03-25 20:25:53 +02:00
void printregionstats(struct vmproc *vmp)
{
struct vir_region *vr;
struct phys_region *pr;
physr_iter iter;
vir_bytes used = 0, weighted = 0;
region_iter v_iter;
region_start_iter_least(&vmp->vm_regions_avl, &v_iter);
while((vr = region_get_iter(&v_iter))) {
region_incr_iter(&v_iter);
if(vr->flags & VR_DIRECT)
continue;
physr_start_iter_least(vr->phys, &iter);
while((pr = physr_get_iter(&iter))) {
physr_incr_iter(&iter);
used += VM_PAGE_SIZE;
weighted += VM_PAGE_SIZE / pr->ph->refcount;
}
}
2010-07-05 15:58:57 +02:00
printf("%6lukB %6lukB\n", used/1024, weighted/1024);
return;
}
/*===========================================================================*
* do_map_memory *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static int do_map_memory(struct vmproc *vms, struct vmproc *vmd,
struct vir_region *vrs, struct vir_region *vrd,
vir_bytes offset_s, vir_bytes offset_d,
vir_bytes length, int flag)
{
struct phys_region *prs;
struct phys_region *newphysr;
struct phys_block *pb;
physr_iter iter;
u32_t pt_flag = PTF_PRESENT | PTF_USER;
vir_bytes end;
/* Search for the first phys region in the source process. */
physr_start_iter(vrs->phys, &iter, offset_s, AVL_EQUAL);
prs = physr_get_iter(&iter);
if(!prs)
panic("do_map_memory: no aligned phys region: %d", 0);
/* flag: 0 -> read-only
* 1 -> writable
* -1 -> share as COW, so read-only
*/
if(flag > 0)
pt_flag |= PTF_WRITE;
2012-08-16 23:34:15 +02:00
else
pt_flag |= PTF_READ;
/* Map phys blocks in the source process to the destination process. */
end = offset_d + length;
while((prs = physr_get_iter(&iter)) && offset_d < end) {
/* If a SMAP share was requested but the phys block has already
* been shared as COW, copy the block for the source phys region
* first.
*/
pb = prs->ph;
if(flag >= 0 && pb->refcount > 1
&& pb->share_flag == PBSH_COW) {
if(!(prs = map_clone_ph_block(vms, vrs, prs, &iter)))
return ENOMEM;
pb = prs->ph;
}
/* Allocate a new phys region. */
if(!(newphysr = pb_reference(pb, offset_d, vrd)))
return ENOMEM;
/* If a COW share was requested but the phys block has already
* been shared as SMAP, give up on COW and copy the block for
* the destination phys region now.
*/
if(flag < 0 && pb->refcount > 1
&& pb->share_flag == PBSH_SMAP) {
if(!(newphysr = map_clone_ph_block(vmd, vrd,
newphysr, NULL))) {
return ENOMEM;
}
}
else {
/* See if this is a COW share or SMAP share. */
if(flag < 0) { /* COW share */
pb->share_flag = PBSH_COW;
/* Update the page table for the src process. */
pt_writemap(vms, &vms->vm_pt, offset_s + vrs->vaddr,
pb->phys, VM_PAGE_SIZE,
pt_flag, WMF_OVERWRITE);
}
else { /* SMAP share */
pb->share_flag = PBSH_SMAP;
}
/* Update the page table for the destination process. */
pt_writemap(vmd, &vmd->vm_pt, offset_d + vrd->vaddr,
pb->phys, VM_PAGE_SIZE, pt_flag, WMF_OVERWRITE);
}
physr_incr_iter(&iter);
offset_d += VM_PAGE_SIZE;
offset_s += VM_PAGE_SIZE;
}
return OK;
}
/*===========================================================================*
* unmap_memory *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int unmap_memory(endpoint_t sour, endpoint_t dest,
vir_bytes virt_s, vir_bytes virt_d, vir_bytes length, int flag)
{
struct vmproc *vmd;
struct vir_region *vrd;
struct phys_region *pr;
struct phys_block *pb;
physr_iter iter;
vir_bytes off, end;
int p;
/* Use information on the destination process to unmap. */
if(vm_isokendpt(dest, &p) != OK)
panic("unmap_memory: bad endpoint: %d", dest);
vmd = &vmproc[p];
vrd = map_lookup(vmd, virt_d, NULL);
assert(vrd);
/* Search for the first phys region in the destination process. */
off = virt_d - vrd->vaddr;
physr_start_iter(vrd->phys, &iter, off, AVL_EQUAL);
pr = physr_get_iter(&iter);
if(!pr)
panic("unmap_memory: no aligned phys region: %d", 0);
/* Copy the phys block now rather than doing COW. */
end = off + length;
while((pr = physr_get_iter(&iter)) && off < end) {
pb = pr->ph;
assert(pb->refcount > 1);
assert(pb->share_flag == PBSH_SMAP);
if(!(pr = map_clone_ph_block(vmd, vrd, pr, &iter)))
return ENOMEM;
physr_incr_iter(&iter);
off += VM_PAGE_SIZE;
}
return OK;
}
/*===========================================================================*
* rm_phys_regions *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static void rm_phys_regions(struct vir_region *region,
vir_bytes begin, vir_bytes length)
{
/* Remove all phys regions between @begin and @begin+length.
*
* Don't update the page table, because we will update it at map_memory()
* later.
*/
struct phys_region *pr;
physr_iter iter;
physr_start_iter(region->phys, &iter, begin, AVL_GREATER_EQUAL);
while((pr = physr_get_iter(&iter)) && pr->offset < begin + length) {
pb_unreferenced(region, pr, 1);
physr_start_iter(region->phys, &iter, begin,
AVL_GREATER_EQUAL);
SLABFREE(pr);
}
}
/*===========================================================================*
* map_memory *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int map_memory(endpoint_t sour, endpoint_t dest,
vir_bytes virt_s, vir_bytes virt_d, vir_bytes length, int flag)
{
/* This is the entry point. This function will be called by handle_memory() when
* VM recieves a map-memory request.
*/
struct vmproc *vms, *vmd;
struct vir_region *vrs, *vrd;
vir_bytes offset_s, offset_d;
int p;
int r;
if(vm_isokendpt(sour, &p) != OK)
panic("map_memory: bad endpoint: %d", sour);
vms = &vmproc[p];
if(vm_isokendpt(dest, &p) != OK)
panic("map_memory: bad endpoint: %d", dest);
vmd = &vmproc[p];
vrs = map_lookup(vms, virt_s, NULL);
assert(vrs);
vrd = map_lookup(vmd, virt_d, NULL);
assert(vrd);
/* Linear address -> offset from start of vir region. */
offset_s = virt_s - vrs->vaddr;
offset_d = virt_d - vrd->vaddr;
/* Make sure that the range in the source process has been mapped
* to physical memory.
*/
map_handle_memory(vms, vrs, offset_s, length, 0);
/* Prepare work. */
rm_phys_regions(vrd, offset_d, length);
/* Map memory. */
r = do_map_memory(vms, vmd, vrs, vrd, offset_s, offset_d, length, flag);
return r;
}
/*===========================================================================*
* get_clean_phys_region *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
static struct phys_region *
get_clean_phys_region(struct vmproc *vmp, vir_bytes vaddr, struct vir_region **ret_region)
{
struct vir_region *region;
vir_bytes mapaddr;
struct phys_region *ph;
No more intel/minix segments. This commit removes all traces of Minix segments (the text/data/stack memory map abstraction in the kernel) and significance of Intel segments (hardware segments like CS, DS that add offsets to all addressing before page table translation). This ultimately simplifies the memory layout and addressing and makes the same layout possible on non-Intel architectures. There are only two types of addresses in the world now: virtual and physical; even the kernel and processes have the same virtual address space. Kernel and user processes can be distinguished at a glance as processes won't use 0xF0000000 and above. No static pre-allocated memory sizes exist any more. Changes to booting: . The pre_init.c leaves the kernel and modules exactly as they were left by the bootloader in physical memory . The kernel starts running using physical addressing, loaded at a fixed location given in its linker script by the bootloader. All code and data in this phase are linked to this fixed low location. . It makes a bootstrap pagetable to map itself to a fixed high location (also in linker script) and jumps to the high address. All code and data then use this high addressing. . All code/data symbols linked at the low addresses is prefixed by an objcopy step with __k_unpaged_*, so that that code cannot reference highly-linked symbols (which aren't valid yet) or vice versa (symbols that aren't valid any more). . The two addressing modes are separated in the linker script by collecting the unpaged_*.o objects and linking them with low addresses, and linking the rest high. Some objects are linked twice, once low and once high. . The bootstrap phase passes a lot of information (e.g. free memory list, physical location of the modules, etc.) using the kinfo struct. . After this bootstrap the low-linked part is freed. . The kernel maps in VM into the bootstrap page table so that VM can begin executing. Its first job is to make page tables for all other boot processes. So VM runs before RS, and RS gets a fully dynamic, VM-managed address space. VM gets its privilege info from RS as usual but that happens after RS starts running. . Both the kernel loading VM and VM organizing boot processes happen using the libexec logic. This removes the last reason for VM to still know much about exec() and vm/exec.c is gone. Further Implementation: . All segments are based at 0 and have a 4 GB limit. . The kernel is mapped in at the top of the virtual address space so as not to constrain the user processes. . Processes do not use segments from the LDT at all; there are no segments in the LDT any more, so no LLDT is needed. . The Minix segments T/D/S are gone and so none of the user-space or in-kernel copy functions use them. The copy functions use a process endpoint of NONE to realize it's a physical address, virtual otherwise. . The umap call only makes sense to translate a virtual address to a physical address now. . Segments-related calls like newmap and alloc_segments are gone. . All segments-related translation in VM is gone (vir2map etc). . Initialization in VM is simpler as no moving around is necessary. . VM and all other boot processes can be linked wherever they wish and will be mapped in at the right location by the kernel and VM respectively. Other changes: . The multiboot code is less special: it does not use mb_print for its diagnostics any more but uses printf() as normal, saving the output into the diagnostics buffer, only printing to the screen using the direct print functions if a panic() occurs. . The multiboot code uses the flexible 'free memory map list' style to receive the list of free memory if available. . The kernel determines the memory layout of the processes to a degree: it tells VM where the kernel starts and ends and where the kernel wants the top of the process to be. VM then uses this entire range, i.e. the stack is right at the top, and mmap()ped bits of memory are placed below that downwards, and the break grows upwards. Other Consequences: . Every process gets its own page table as address spaces can't be separated any more by segments. . As all segments are 0-based, there is no distinction between virtual and linear addresses, nor between userspace and kernel addresses. . Less work is done when context switching, leading to a net performance increase. (8% faster on my machine for 'make servers'.) . The layout and configuration of the GDT makes sysenter and syscall possible.
2012-05-07 16:03:35 +02:00
mapaddr = vaddr;
if(!(region = map_lookup(vmp, mapaddr, &ph)) || !ph) {
printf("VM: get_clean_phys_region: 0x%lx not found\n", vaddr);
return NULL;
}
assert(mapaddr >= region->vaddr);
assert(mapaddr < region->vaddr + region->length);
/* If it's mapped more than once, make a copy. */
assert(ph->ph->refcount > 0);
if(ph->ph->refcount > 1) {
if(!(ph = map_clone_ph_block(vmp, region,
ph, NULL))) {
printf("VM: get_clean_phys_region: ph copy failed\n");
return NULL;
}
}
assert(ph->ph->refcount == 1);
*ret_region = region;
return ph;
}
static int getblock(struct vmproc *vmp, u64_t id, vir_bytes vaddr, int pages)
{
yielded_t *yb;
struct phys_region *ph;
struct vir_region *region;
2010-10-15 11:10:14 +02:00
yielded_avl *avl;
block_id_t blockid;
phys_bytes phaddr;
int p;
/* Try to get the yielded block */
2010-10-15 11:10:14 +02:00
blockid.owner = vmp->vm_endpoint;
blockid.id = id;
avl = get_yielded_avl(blockid);
if(!(yb = yielded_search(avl, blockid, AVL_EQUAL))) {
return ESRCH;
}
if(yb->pages != pages) {
printf("VM: getblock: length mismatch (%d != %d)\n",
pages, yb->pages);
return EFAULT;
}
phaddr = yb->physaddr;
for(p = 0; p < pages; p++) {
/* Get the intended phys region, make sure refcount is 1. */
if(!(ph = get_clean_phys_region(vmp, vaddr, &region))) {
printf("VM: getblock: not found for %d\n", vmp->vm_endpoint);
return EINVAL;
}
assert(ph->ph->refcount == 1);
/* Free the block that is currently there. */
free_mem(ABS2CLICK(ph->ph->phys), 1);
/* Set the phys block to new addr and update pagetable. */
USE(ph->ph, ph->ph->phys = phaddr;);
if(map_ph_writept(vmp, region, ph) != OK) {
/* Presumably it was mapped, so there is no reason
* updating should fail.
*/
panic("do_get_block: couldn't write pt");
}
vaddr += VM_PAGE_SIZE;
phaddr += VM_PAGE_SIZE;
}
/* Forget about the yielded block and free the struct. */
freeyieldednode(yb, 0);
return OK;
}
2012-03-25 20:25:53 +02:00
static int yieldblock(struct vmproc *vmp, u64_t id,
vir_bytes vaddr, yielded_t **retyb, int pages)
{
yielded_t *newyb;
vir_bytes mem_clicks, v, p, new_phaddr;
struct vir_region *region;
struct phys_region *ph = NULL, *prev_ph = NULL, *first_ph = NULL;
2010-10-15 11:10:14 +02:00
yielded_avl *avl;
block_id_t blockid;
/* Makes no sense if yielded block ID already exists, and
* is likely a serious bug in the caller.
*/
2010-10-15 11:10:14 +02:00
blockid.id = id;
blockid.owner = vmp->vm_endpoint;
avl = get_yielded_avl(blockid);
if(yielded_search(avl, blockid, AVL_EQUAL)) {
printf("!");
return EINVAL;
}
if((vaddr % VM_PAGE_SIZE) || pages < 1) return EFAULT;
v = vaddr;
for(p = 0; p < pages; p++) {
if(!(region = map_lookup(vmp, v, &ph)) || !ph) {
printf("VM: do_yield_block: not found for %d\n",
vmp->vm_endpoint);
return EINVAL;
}
if(!(region->flags & VR_ANON)) {
printf("VM: yieldblock: non-anon 0x%lx\n", v);
return EFAULT;
}
if(ph->ph->refcount != 1) {
printf("VM: do_yield_block: mapped not once for %d\n",
vmp->vm_endpoint);
return EFAULT;
}
if(prev_ph) {
if(ph->ph->phys != prev_ph->ph->phys + VM_PAGE_SIZE) {
printf("VM: physically discontiguous yield\n");
return EINVAL;
}
}
prev_ph = ph;
if(!first_ph) first_ph = ph;
v += VM_PAGE_SIZE;
}
/* Make a new block to record the yielding in. */
if(!SLABALLOC(newyb)) {
return ENOMEM;
}
assert(!(ph->ph->phys % VM_PAGE_SIZE));
if((mem_clicks = alloc_mem(pages, PAF_CLEAR)) == NO_MEM) {
SLABFREE(newyb);
return ENOMEM;
}
/* Update yielded block info. */
USE(newyb,
2010-10-15 11:10:14 +02:00
newyb->id = blockid;
newyb->physaddr = first_ph->ph->phys;
newyb->pages = pages;
newyb->younger = NULL;);
new_phaddr = CLICK2ABS(mem_clicks);
/* Set new phys block to new addr and update pagetable. */
v = vaddr;
for(p = 0; p < pages; p++) {
region = map_lookup(vmp, v, &ph);
assert(region && ph);
assert(ph->ph->refcount == 1);
USE(ph->ph,
ph->ph->phys = new_phaddr;);
if(map_ph_writept(vmp, region, ph) != OK) {
/* Presumably it was mapped, so there is no reason
* updating should fail.
*/
panic("yield_block: couldn't write pt");
}
v += VM_PAGE_SIZE;
new_phaddr += VM_PAGE_SIZE;
}
/* Remember yielded block. */
2010-10-15 11:10:14 +02:00
yielded_insert(avl, newyb);
vmp->vm_yielded++;
/* Add to LRU list too. It's the youngest block. */
LRUCHECK;
if(lru_youngest) {
USE(lru_youngest,
lru_youngest->younger = newyb;);
} else {
lru_oldest = newyb;
}
USE(newyb,
newyb->older = lru_youngest;);
lru_youngest = newyb;
LRUCHECK;
if(retyb)
*retyb = newyb;
return OK;
}
/*===========================================================================*
* do_forgetblocks *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int do_forgetblocks(message *m)
{
int n;
struct vmproc *vmp;
endpoint_t caller = m->m_source;
if(vm_isokendpt(caller, &n) != OK)
panic("do_yield_block: message from strange source: %d",
m->m_source);
vmp = &vmproc[n];
free_yielded_proc(vmp);
return OK;
}
/*===========================================================================*
* do_forgetblock *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int do_forgetblock(message *m)
{
int n;
struct vmproc *vmp;
endpoint_t caller = m->m_source;
yielded_t *yb;
u64_t id;
2010-10-15 11:10:14 +02:00
block_id_t blockid;
yielded_avl *avl;
if(vm_isokendpt(caller, &n) != OK)
panic("do_yield_block: message from strange source: %d",
m->m_source);
vmp = &vmproc[n];
id = make64(m->VMFB_IDLO, m->VMFB_IDHI);
2010-10-15 11:10:14 +02:00
blockid.id = id;
blockid.owner = vmp->vm_endpoint;
avl = get_yielded_avl(blockid);
if((yb = yielded_search(avl, blockid, AVL_EQUAL))) {
freeyieldednode(yb, 1);
}
return OK;
}
/*===========================================================================*
* do_yieldblockgetblock *
*===========================================================================*/
2012-03-25 20:25:53 +02:00
int do_yieldblockgetblock(message *m)
{
u64_t yieldid, getid;
int n;
endpoint_t caller = m->m_source;
struct vmproc *vmp;
yielded_t *yb = NULL;
int r = ESRCH;
int pages;
if(vm_isokendpt(caller, &n) != OK)
panic("do_yieldblockgetblock: message from strange source: %d",
m->m_source);
vmp = &vmproc[n];
pages = m->VMYBGB_LEN / VM_PAGE_SIZE;
if((m->VMYBGB_LEN % VM_PAGE_SIZE) || pages < 1) {
static int printed;
if(!printed) {
printed = 1;
printf("vm: non-page-aligned or short block length\n");
}
return EFAULT;
}
yieldid = make64(m->VMYBGB_YIELDIDLO, m->VMYBGB_YIELDIDHI);
getid = make64(m->VMYBGB_GETIDLO, m->VMYBGB_GETIDHI);
if(cmp64(yieldid, VM_BLOCKID_NONE) != 0) {
/* A block was given to yield. */
yieldblock(vmp, yieldid, (vir_bytes) m->VMYBGB_VADDR, &yb,
pages);
}
if(cmp64(getid, VM_BLOCKID_NONE) != 0) {
/* A block was given to get. */
r = getblock(vmp, getid, (vir_bytes) m->VMYBGB_VADDR, pages);
}
return r;
}
void map_setparent(struct vmproc *vmp)
{
region_iter iter;
struct vir_region *vr;
region_start_iter_least(&vmp->vm_regions_avl, &iter);
while((vr = region_get_iter(&iter))) {
USE(vr, vr->parent = vmp;);
region_incr_iter(&iter);
}
}