There is a deadlock vulnerability when there are no worker threads
available and all of them blocked on a worker thread that's waiting for a
reply from a driver or a reply from an FS that needs to make a back call. In
these cases the deadlock resolver thread should kick in, but didn't in all
cases. Moreover, POSIX calls from File Servers weren't handled properly
anymore, which also could lead to deadlocks.
The check_bsf() macro uses assert(mutex_trylock(&bsf_lock)) and
assumes bsf_lock is locked afterwards. This breaks when compiling
with NOASSERTS="yes". Also: macro to function transition.
. whenever this function is called, pm will expect
the process to be cleaned up
. so don't abort the process entirely on error
. fixes a later 'forking on top of in-use child' vfs panic
By decoupling synchronous drivers from VFS, we are a big step closer to
supporting driver crashes under all circumstances. That is, VFS can't
become stuck on IPC with a synchronous driver (e.g., INET) and can
recover from crashing block drivers during open/close/ioctl or during
communication with an FS.
In order to maintain serialized communication with a synchronous driver,
the communication is wrapped by a mutex on a per driver basis (not major
numbers as there can be multiple majors with identical endpoints). Majors
that share a driver endpoint point to a single mutex object.
In order to support crashes from block drivers, the file reopen tactic
had to be changed; first reopen files associated with the crashed
driver, then send the new driver endpoint to FSes. This solves a
deadlock between the FS and the block driver;
- VFS would send REQ_NEW_DRIVER to an FS, but he FS only receives it
after retrying the current request to the newly started driver.
- The block driver would refuse the retried request until all files
had been reopened.
- VFS would reopen files only after getting a reply from the initial
REQ_NEW_DRIVER.
When a character special driver crashes, all associated files have to
be marked invalid and closed (or reopened if flagged as such). However,
they can only be closed if a thread holds exclusive access to it. To
obtain exclusive access, the worker thread (which handles the new driver
endpoint event from DS) schedules a new job to garbage collect invalid
files. This way, we can signal the worker thread that was talking to the
crashed driver and will release exclusive access to a file associated
with the crashed driver and prevent the garbage collecting worker thread
from dead locking on that file.
Also, when a character special driver crashes, RS will unmap the driver
and remap it upon restart. During unmapping, associated files are marked
invalid instead of waiting for an endpoint up event from DS, as that
event might come later than new read/write/select requests and thus
cause confusion in the freshly started driver.
When locking a filp, the usage counters are no longer checked. The usage
counter can legally go down to zero during filp invalidation while there
are locks pending.
DS events are handled by a separate worker thread instead of the main
thread as reopening files could lead to another crash and a stuck thread.
An additional worker thread is then necessary to unlock it.
Finally, with everything asynchronous a race condition in do_select
surfaced. A select entry was only marked in use after succesfully sending
initial select requests to drivers and having to wait. When multiple
select() calls were handled there was opportunity that these entries
were overwritten. This had as effect that some select results were
ignored (and select() remained blocking instead if returning) or do_select
tried to access filps that were not present (because thrown away by
secondary select()). This bug manifested itself with sendrecs, but was
very hard to reproduce. However, it became awfully easy to trigger with
asynsends only.
. ld.so is linked at 0 but it can relocate itself; we
wish to load ld.so higher though to trap NULL dereferences.
if we know we have to execute ld.so, vfs tells libexec to put it
higher.
When VFS runs out of vnodes after closing a vnode in opcl, common_open
will try to unlock a vnode through unlock_filp that has already been
unlocked in clone_opcl. By first obtaining and locking a new vnode this
situation is prevented; if there are no free vnodes, common_open will
unlock a still locked vnode.
. some strncpy/strcpy to strlcpy conversions
. new <minix/param.h> to avoid including other minix headers
that have colliding definitions with library and commands code,
causing parse warnings
. removed some dead code / assignments
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.
. sys_vircopy always uses D for both src and dst
. sys_physcopy uses PHYS_SEG if and only if corresponding
endpoint is NONE, so we can derive the mode (PHYS_SEG or D)
from the endpoint arg in the kernel, dropping the seg args
. fields in msg still filled in for backwards compatability,
using same NONE-logic in the library
. all invocations were S or D, so can safely be dropped
to prepare for the segmentless world
. still assign D to the SCP_SEG field in the message
to make previous kernels usable
. new mode for sys_memset: include process so memset can be
done in physical or virtual address space.
. add a mode to mmap() that lets a process allocate uninitialized
memory.
. this allows an exec()er (RS, VFS, etc.) to request uninitialized
memory from VM and selectively clear the ranges that don't come
from a file, leaving no uninitialized memory left for the process
to see.
. use callbacks for clearing the process, clearing memory in the
process, and copying into the process; so that the libexec code
can be used from rs, vfs, and in the future, kernel (to load vm)
and vm (to load boot-time processes)
. make exec() callers (i.e. vfs and rs) determine the
memory layout by explicitly reserving regions using
mmap() calls on behalf of the exec()ing process,
i.e. handling all of the exec logic, thereby eliminating
all special exec() knowledge from VM.
. the new procedure is: clear the exec()ing process
first, then call third-party mmap()s to reserve memory, then
copy the executable file section contents in, all using callbacks
tailored to the caller's way of starting an executable
. i.e. no more explicit EXEC_NEWMEM-style calls in PM or VM
as with rigid 2-section arguments
. this naturally allows generalizing exec() by simply loading
all ELF sections
. drop/merge of lots of duplicate exec() code into libexec
. not copying the code sections to vfs and into the executable
again is a measurable performance improvement (about 3.3% faster
for 'make' in src/servers/)
justification: soon we won't be able to execute sep I&D aouts at
all (because of the vanishing segments), which was the default mode
to generate them so most binaries will be sep I&D.
this makes the vfs/rs exec() unification work simpler.
after unification, common I&D aout could be added back quite simply.
Only attempt to release blocked processes that are blocked. There is
no use in trying to find more blocked processes than we know that are
blocked (on a pipe).
According to POSIX the st_size field of struct stat is undefined for
fifos and anonymous pipes. Thus we can do anything we want. We save a
copy by not being accurate on pipe sizes.
. vfs: pass execname in aux vectors
. ld.elf_so: use this to expand $ORIGIN
. this requires the executable to reserve more
space at exec() calling time
POSIX mandates that a file's modification and change time be left
untouched upon truncate/ftruncate iff the file size does not change.
However, an open(O_TRUNC) call must always update the modification and
change time of the file, even if it was already zero-sized. VFS uses
the file systems' truncate call to implement O_TRUNC. This patch
replaces git-255ae85, which did not take into account the open case.
The size check is now moved into VFS, so that individual file systems
need not check for this case anymore.
. generalize libexec slightly to get some more necessary information
from ELF files, e.g. the interpreter
. execute dynamically linked executables when exec()ed by VFS
. switch to netbsd variant of elf32.h exclusively, solves some
conflicting headers
Pipes consist of two filps (read filp and write filp) and a shared
vnode. When the writer leaves the filp reference count drops to
zero and subsequent find_filp()s should not find the filp when a
reader looks for it and the reader gets EOF. However, the pipe()
system call tries to find two filps, marks them in use, and only
after a successful node creation on PFS, overwrites the shared
vnode with the new vnode. Consequently, this leaves a small window
where a just closed 'pipe write filp' gets reused and marked as
present, before becoming the actual new 'pipe write filp' for a new
pipe. A reader for the old pipe will think a writer is present and
wait for that writer to write something or to leave; both actions
should revive the suspended reader. This will never happen and the
reader will be stuck forever.
When running out of worker threads to handle device replies a dead
lock resolver thread is used. However, it was only used for FS
endpoints; it is now used for "system processes" (drivers and FS
endpoints). Also, drivers were marked as system process when they
were not "forced" to map (i.e., mapping was done before endpoint was
alive).
By making m_in job local (i.e., each job has its own copy of m_in instead
of refering to the global m_in) we don't have to store and restore m_in
on every thread yield. This reduces overhead. Moreover, remove the
assumption that m_in is preserved. Do_XXX functions have to copy the
system call parameters as soon as possible and only pass those copies to
other functions.
Furthermore, this patch cleans up some code and uses better types in a lot
of places.
