- APs configure local timers
- while configuring local APIC timer the CPUs fiddle with the interrupt
handlers. As the interrupt table is shared the BSP must not run
- APs wait until BSP turns paging on, it is not possible to safely
execute any code on APs until we can turn paging on as well as it
must be done synchronously everywhere
- APs turn paging on but do not continue and wait
- to isolate execution inside kernel we use a big kernel lock
implemented as a spinlock
- the lock is acquired asap after entering kernel mode and released as
late as possible. Only one CPU as a time can execute the core kernel
code
- measurement son real hw show that the overhead of this lock is close
to 0% of kernel time for the currnet system
- the overhead of this lock may be as high as 45% of kernel time in
virtual machines depending on the ratio between physical CPUs
available and emulated CPUs. The performance degradation is
significant
- kernel detects CPUs by searching ACPI tables for local apic nodes
- each CPU has its own TSS that points to its own stack. All cpus boot
on the same boot stack (in sequence) but switch to its private stack
as soon as they can.
- final booting code in main() placed in bsp_finish_booting() which is
executed only after the BSP switches to its final stack
- apic functions to send startup interrupts
- assembler functions to handle CPU features not needed for single cpu
mode like memory barries, HT detection etc.
- new files kernel/smp.[ch], kernel/arch/i386/arch_smp.c and
kernel/arch/i386/include/arch_smp.h
- 16-bit trampoline code for the APs. It is executed by each AP after
receiving startup IPIs it brings up the CPUs to 32bit mode and let
them spin in an infinite loop so they don't do any damage.
- implementation of kernel spinlock
- CONFIG_SMP and CONFIG_MAX_CPUS set by the build system
- most global variables carry information which is specific to the
local CPU and each CPU must have its own copy
- cpu local variable must be declared in cpulocal.h between
DECLARE_CPULOCAL_START and DECLARE_CPULOCAL_END markers using
DECLARE_CPULOCAL macro
- to access the cpu local data the provided macros must be used
get_cpu_var(cpu, name)
get_cpu_var_ptr(cpu, name)
get_cpulocal_var(name)
get_cpulocal_var_ptr(name)
- using this macros makes future changes in the implementation
possible
- switching to ELF will make the declaration of cpu local data much
simpler, e.g.
CPULOCAL int blah;
anywhere in the kernel source code
- kernel turns on IO APICs if no_apic is _not_ set or is equal 0
- pci driver must use the acpi driver to setup IRQ routing otherwise
the system cannot work correctly except systems like KVM that use
only legacy (E)ISA IRQs 0-15
- kernel exports DSDP (the root pointer where ACPI parsing starts) and
apic_enabled in the machine structure.
- ACPI driver uses DSDP to locate ACPI in memory. acpi_enabled tell
PCI driver to query ACPI for IRQ routing information.
- the ability for kernel to use ACPI tables to detect IO APICs. It is
the bare minimum the kernel needs to know about ACPI tables.
- it will be used to find out about processors as the MPS tables are
deprecated by ACPI and not all vendorsprovide them.
- kernel compile was broken with gcc as putchar() was added by gcc in
stacktrace.c
- add -fno-builtin everywhere to avoid such problems in the future
- -fno-builtin in kernel now redundant
- for better readability xpp is substitued by sender
- makes sure that the dequeued sender has p_q_link == NULL and that
this condition holds when enqueuing the sender again. This is a
sanity check to make sure that the new sender is not enqueued
already
- Before this change the dequeued sender's p_q_link may not be NULL
and it was only set to NULL when enqueued again
- removes p_delivermsg_lin item from the process structure and code
related to it
- as the send part, the receive does not need to use the
PHYS_COPY_CATCH() and umap_local() couple.
- The address space of the target process is installed before
delivermsg() is called.
- unlike the linear address, the virtual address does not change when
paging is turned on nor after fork().
- FPU context is stored only if conflict between 2 FPU users or while
exporting context of a process to userspace while it is the active
user of FPU
- FPU has its owner (fpu_owner) which points to the process whose
state is currently loaded in FPU
- the FPU exception is only turned on when scheduling a process which
is not the owner of FPU
- FPU state is restored for the process that generated the FPU
exception. This process runs immediately without letting scheduler
to pick a new process to resolve the FPU conflict asap, to minimize
the FPU thrashing and FPU exception hadler execution
- faster all non-FPU-exception kernel entries as FPU state is not
checked nor saved
- removed MF_USED_FPU flag, only MF_FPU_INITIALIZED remains to signal
that a process has used FPU in the past
There seems to have been a broken assumption in the fpu context
restoring code. It restores the context of the running process, without
guarantee that the current process is the one that will be scheduled.
This caused fpu saving for a different process to be triggered without
fpu hardware being enabled, causing an fpu exception in the kernel. This
practically only shows up with DEBUG_RACE on. Fix my thruby+me.
The fix
. is to only set the fpu-in-use-by-this-process flag in the
exception handler, and then take care of fpu restoring when
actually returning to userspace
And the patch
. translates fpu saving and restoring to c in arch_system.c,
getting rid of a juicy chunk of assembly
. makes osfxsr_feature private to arch_system.c
. removes most of the arch dependent code from do_sigsend
- substituted the use of the m_source message field by
caller->p_endpoint in kernel calls. It is the same information, just
passed more intuitively.
- the last dependency on m_type field is removed.
- do_unused() is substituted by a check for NULL.
- this pretty much removes the depency of kernel calls on the general
message format. In the future this may be used to pass the kcall
arguments in a different structure or registers (x86-64, ARM?) The
kcall number may be passed in a register already.
- removes dependency of do_safecopy() on the m_type field of the kcall
messages.
- instead of do_safecopy() figuring out what action is requested, the
correct safecopy method is called right away.
- Currently the cpu time quantum is timer-ticks based. Thus the
remaining quantum is decreased only if the processes is interrupted
by a timer tick. As processes block a lot this typically does not
happen for normal user processes. Also the quantum depends on the
frequency of the timer.
- This change makes the quantum miliseconds based. Internally the
miliseconds are translated into cpu cycles. Everytime userspace
execution is interrupted by kernel the cycles just consumed by the
current process are deducted from the remaining quantum.
- It makes the quantum system timer frequency independent.
- The boot processes quantum is loosely derived from the tick-based
quantas and 60Hz timer and subject to future change
- the 64bit arithmetics is a little ugly, will be changes once we have
compiler support for 64bit integers (soon)
-Makefile updates
-Update mkdep
-Build fixes/warning cleanups for some programs
-Restore leading underscores on global syms in kernel asm files
-Increase ramdisk size
ask to map in oxpcie i/o memory and support serial i/o for it in the
kernel. set oxpcie=<address> in boot monitor (retrieve address using
pci_debug=1 output). (no sanity checking is done on the address
currently.) disabled by default.
The change also contains some other minor cleanup (a new serial.h to set
register info common to UART and the OXPCIe card, in-kernel memory
mapping a little more structured and env_get() to get sysenv variables
without knowing about the params_buffer).
In this second phase, scheduling is moved from PM to its own
scheduler (see r6557 for phase one). In the next phase we hope to a)
include useful information in the "out of quantum" message and b)
create some simple scheduling policy that makes use of that
information.
When the system starts up, PM will iterate over its process table and
ask SCHED to take over scheduling unprivileged processes. This is
done by sending a SCHEDULING_START message to SCHED. This message
includes the processes endpoint, the parent's endpoint and its nice
level. The scheduler adds this process to its schedproc table, issues
a schedctl, and returns its own endpoint to PM - as the endpoint of
the effective scheduler. When a process terminates, a SCHEDULING_STOP
message is sent to the scheduler.
The reason for this effective endpoint is for future compatibility.
Some day, we may have a scheduler that, instead of scheduling the
process itself, forwards the SCHEDULING_START message on to another
scheduler.
PM has information on who schedules whom. As such, scheduling
messages from user-land are sent through PM. An example is when
processes change their priority, using nice(). In that case, a
getsetpriority message is sent to PM, which then sends a
SCHEDULING_SET_NICE to the process's effective scheduler.
When a process is forked through PM, it inherits its parent's
scheduler, but is spawned with an empty quantum. As before, a request
to fork a process flows through VM before returning to PM, which then
wakes up the child process. This flow has been modified slightly so
that PM notifies the scheduler of the new process, before waking up
the child process. If the scheduler fails to take over scheduling,
the child process is torn down and the fork fails with an erroneous
value.
Process priority is entirely decided upon using nice levels. PM
stores a copy of each process's nice level and when a child is
forked, its parent's nice level is sent in the SCHEDULING_START
message. How this level is mapped to a priority queue is up to the
scheduler. It should be noted that the nice level is used to
determine the max_priority and the parent could have been in a lower
priority when it was spawned. To prevent a CPU intensive process from
hawking the CPU by continuously forking children that get scheduled
in the max_priority, the scheduler should determine in which queue
the parent is currently scheduled, and schedule the child in that
same queue.
Other fixes: The USER_Q in kernel/proc.h was incorrectly defined as
NR_SCHED_QUEUES/2. That results in a "off by one" error when
converting priority->nice->priority for nice=0. This also had the
side effect that if someone were to set the MAX_USER_Q to something
else than 0, then USER_Q would be off.
- this patch only renames schedcheck() to switch_to_user(),
cycles_accounting_stop() to context_stop() and restart() to
+restore_user_context()
- the motivation is that since the introduction of schedcheck() it has
been abused for many things. It deserves a better name. It should
express the fact that from the moment we call the function we are in
the process of switching to user.
- cycles_accounting_stop() was originally a single purpose function.
As this function is called at were convenient places it is used in
for other things too, e.g. (un)locking the kernel. Thus it deserves
a better name too.
- using the old name, restart() does not call schedcheck(), however
calls to restart are replaced by calls to schedcheck()
[switch_to_user] and it calls restart() [restore_user_context]
it does this by
- making all processes interruptible by running out of quantum
- giving all processes a single tick of quantum
- picking a random runnable process instead of in order, and
from a single pool of runnable processes (no priorities)
This together with very high HZ values currently provokes some race conditions
seen earlier only when running with SMP.
- this patch substitutes *xpp for sender to increase readability of
mini_receive().
- makes sure that the dequeued sender has p_q_link == NULL and that
this condition holds when enqueuing the sender again.
- it is a sanity check to make sure that the new sender is not
enqueued already. Before this change the dequeued sender's p_q_link
may not be NULL and it was only set to NULL when enqueued again.
- deadlock() is more verbose in case of a detected deadlock. First, it
lists all processses in the deadlock group. Then it prints the proc
extra info, not only stack trace and register dump
- this patch moves the former printslot() from arch_system.c to
debug.c and reimplements it slightly. The output is not changed,
however, the process information is printed in a separate function
print_proc() in debug.c as such a function is also handy in other
situations and should be publicly available when debugging.
this patch changes the way pagefaults are delivered to VM. It adopts
the same model as the out-of-quantum messages sent by kernel to a
scheduler.
- everytime a userspace pagefault occurs, kernel creates a message
which is sent to VM on behalf of the faulting process
- the process is blocked on delivery to VM in the standard IPC code
instead of waiting in a spacial in-kernel queue (stack) and is not
runnable until VM tell kernel that the pagefault is resolved and is
free to clear the RTS_PAGEFAULT flag.
- VM does not need call kernel and poll the pagefault information
which saves many (1/2?) calls and kernel calls that return "no more
data"
- VM notification by kernel does not need to use signals
- each entry in proc table is by 12 bytes smaller (~3k save)
- This patch removes the time slice split between parent and child in
fork.
- The time slice of the parent remains unchanged and the child does
not have any.
- If the process has a scheduler, the scheduler must assign the
quantum and priority of the new process and let it run.
- If the child does not inherit a scheduler, it is scheduled by the
dummy default kernel policy. (servers, drivers, etc.)
- In theory, the scheduler can change the quantum even of the parent
process and implement any policy for splitting the quantum as
neither the parent nor the child are runnable. Sending the
out-of_quantum message on behalf of the processes may look like the
right solution, however, the scheduler would probably handle the
message before the whole fork protocol is finished. This way the
scheduler has absolute control when the process should become
runnable.
- this is a small addition to the userspace scheduling.
proc_kernel_scheduler() tests whether to use the default scheduling
policy in kernel. It is true if the process' scheduler is NULL _or_
self. Currently none of the tests was complete.
- it is not neccessary to test whether the scheduler is a system
process as the process already head permissions to make this call.
- it is better to test whether the scheduler has permission to make
changes to this process before testing whether the values are valid.
SYSLIB CHANGES:
- DS calls to publish / retrieve labels consider endpoints instead of u32_t.
VFS CHANGES:
- mapdriver() only adds an entry in the dmap table in VFS.
- dev_up() is only executed upon reception of a driver up event.
INET CHANGES:
- INET no longer searches for existing drivers instances at startup.
- A newtwork driver is (re)initialized upon reception of a driver up event.
- Networking startup is now race-free by design. No need to waste 5 seconds
at startup any more.
DRIVER CHANGES:
- Every driver publishes driver up events when starting for the first time or
in case of restart when recovery actions must be taken in the upper layers.
- Driver up events are published by drivers through DS.
- For regular drivers, VFS is normally the only subscriber, but not necessarily.
For instance, when the filter driver is in use, it must subscribe to driver
up events to initiate recovery.
- For network drivers, inet is the only subscriber for now.
- Every VFS driver is statically linked with libdriver, every network driver
is statically linked with libnetdriver.
DRIVER LIBRARIES CHANGES:
- Libdriver is extended to provide generic receive() and ds_publish() interfaces
for VFS drivers.
- driver_receive() is a wrapper for sef_receive() also used in driver_task()
to discard spurious messages that were meant to be delivered to a previous
version of the driver.
- driver_receive_mq() is the same as driver_receive() but integrates support
for queued messages.
- driver_announce() publishes a driver up event for VFS drivers and marks
the driver as initialized and expecting a DEV_OPEN message.
- Libnetdriver is introduced to provide similar receive() and ds_publish()
interfaces for network drivers (netdriver_announce() and netdriver_receive()).
- Network drivers all support live update with no state transfer now.
KERNEL CHANGES:
- Added kernel call statectl for state management. Used by driver_announce() to
unblock eventual callers sendrecing to the driver.
this patch does not add or change any functionality of do_ipc(), it
only makes things a little cleaner (hopefully).
Until now do_ipc() was responsible for handling all ipc calls. The
catch is that SENDA is fairly different which results in some ugly
code like this typecasting and variables naming which does not make
much sense for SENDA and makes the code hard to read.
result = mini_senda(caller_ptr, (asynmsg_t *)m_ptr, (size_t)src_dst_e);
As it is called directly from assembly, the new do_ipc() takes as
input values of 3 registers in reg_t variables (it used to be 4,
however, bit_map wasn't used so I removed it), does the checks common
to all ipc calls and call the appropriate handler either for
do_sync_ipc() (all except SENDA) or mini_senda() (for SENDA) while
typecasting the reg_t values correctly. As a result, handling SENDA
differences in do_sync_ipc() is no more needed. Also the code that
uses msg_size variable is improved a little bit.
arch_do_syscall() is simplified too.
- IPC_FLG_MSG_FROM_KERNEL status flag is returned to userspace if the
receive was satisfied by s message which was sent by the kernel on
behalf of a process. This perfectly reliale information.
- MF_SENDING_FROM_KERNEL flag added to processes to be able to set
IPC_FLG_MSG_FROM_KERNEL when finishing receive if the receiver
wasn't ready to receive immediately.
- PM is changed to use this information to confirm that the scheduling
messages are indeed from the kernel and not faked by a process.
PM uses sef_receive_status()
- get_work() is removed from PM to make the changes simpler
- there are cycles wasted in the IPC call due to a fairly compliacted
way of copying messages from userland to kernel. Sometimes this
complicated way (generic though) is used even for copying within the
kernel address space, sometimes it is used for copying in case _no_
copying is necessary. The goal of this patch is to improve this a
little bit.
- the places where a copy is from user to kernel use the
copy_msg_from_user() kernel-kernel copies are turned into
assignments and BuildNotifyMessage uses the delivery buffers to
avoid copying.
- copy_msg_from_user() was introduced when removing the system task
and is about 2/3 faster then using the current mechanism
(phys_copy). It also avoids the PHYS_COPY_CATCH macro. Assignment is
also faster and no copy is the fastest ;-) so perhaps there will be
some hardly noticable performance gain besides the clean up.
- cotributed by Bjorn Swift
- In this first phase, scheduling is moved from the kernel to the PM
server. The next steps are to a) moving scheduling to its own server
and b) include useful information in the "out of quantum" message,
so that the scheduler can make use of this information.
- The kernel process table now keeps record of who is responsible for
scheduling each process (p_scheduler). When this pointer is NULL,
the process will be scheduled by the kernel. If such a process runs
out of quantum, the kernel will simply renew its quantum an requeue
it.
- When PM loads, it will take over scheduling of all running
processes, except system processes, using sys_schedctl().
Essentially, this only results in taking over init. As children
inherit a scheduler from their parent, user space programs forked by
init will inherit PM (for now) as their scheduler.
- Once a process has been assigned a scheduler, and runs out of
quantum, its RTS_NO_QUANTUM flag will be set and the process
dequeued. The kernel will send a message to the scheduler, on the
process' behalf, informing the scheduler that it has run out of
quantum. The scheduler can take what ever action it pleases, based
on its policy, and then reschedule the process using the
sys_schedule() system call.
- Balance queues does not work as before. While the old in-kernel
function used to renew the quantum of processes in the highest
priority run queue, the user-space implementation only acts on
processes that have been bumped down to a lower priority queue.
This approach reacts slower to changes than the old one, but saves
us sending a sys_schedule message for each process every time we
balance the queues. Currently, when processes are moved up a
priority queue, their quantum is also renewed, but this can be
fiddled with.
- do_nice has been removed from kernel. PM answers to get- and
setpriority calls, updates it's own nice variable as well as the
max_run_queue. This will be refactored once scheduling is moved to a
separate server. We will probably have PM update it's local nice
value and then send a message to whoever is scheduling the process.
- changes to fix an issue in do_fork() where processes could run out
of quantum but bypassing the code path that handles it correctly.
The future plan is to remove the policy from do_fork() and implement
it in userspace too.
Currently a sequence of messages between a sender A and a receiver B of the
form: A.asynsend(M1, B); A.send(M2, B) may result in the receiver receiving
M1 first and then M2 or viceversa. This patch makes sure that the original
order M1, M2 is always preserved.
Note that the order of a hypotetical sequence A.asynsend(M1, B);
A.asynsend(M2, B) is already guaranteed by the implementation of
asynsend by design. Other senda-based wrappers can define their own
semantics.
- ack assumes that the direction flag in eflags is clear when
assigning two structures. It is implemented by a call to a built-in
function which is like memcpy but needs the flag to be clear
otherwise rubish is copied. This patch fixes the kernel entries.
- When the cpu halts, the interrupts are enable so the cpu may be
woken up. When the interrupt handler returns but another interrupt
is available it is also serviced immediately. This is not a problem
per-se. It only slightly breaks time accounting as idle accounted is
for the kernel time in the interrupt handler.
- As the big kernel lock is lock/unlocked in the smp branch in the
time acounting functions as they are called exactly at the places
we need to take the lock) this leads to a deadlock.
- we make sure that once the interrupt handler returns from the nested
trap, the interrupts are disabled. This means that only one
interrupt is serviced after idle is interrupted.
- this requires the loop in apic timer calibration to keep reenabling
the interrupts. I admit it is a little bit hackish (one line),
however, this code is a stupid corner case at the boot time.
Hopefully it does not matter too much.
IPC changes:
- receive() is changed to take an additional parameter, which is a pointer to
a status code.
- The status code is filled in by the kernel to provide additional information
to the caller. For now, the kernel only fills in the IPC call used by the
sender.
Syslib changes:
- sef_receive() has been split into sef_receive() (with the original semantics)
and sef_receive_status() which exposes the status code to userland.
- Ideally, every sys process should gradually switch to sef_receive_status()
and use is_ipc_notify() as a dependable way to check for notify.
- SEF has been modified to use is_ipc_notify() and demonstrate how to use the
new status code.
