|Age||Commit message (Collapse)||Author||Lines|
the omission of the flag here seems to have been an oversight when the
function was added in 8fb28b0b3e7a5e958fb844722a4b2ef9bc244af1
this change should have been made when priority inheritance mutex
support was added. if priority protection is also added at some point
the implementation will need to change and will probably no longer be
a simple bit shuffling.
both __clone and __syscall_cp_asm failed to restore the original value
of r6 after using it as a syscall argument register. the extent of
breakage is not known, and in some cases may be mitigated by the only
callers being internal to libc; if they used r6 but no longer needed
its value after the call, they may not have noticed the problem.
however at least posix_spawn (which uses __clone) was observed
returning to the application with the wrong value in r6, leading to
since the call frame ABI already provides a place to spill registers,
fixing this is just a matter of using it. in __clone, we also
spuriously restore r6 in the child, since the parent branch directly
returns to the caller. this takes the value from an uninitialized slot
of the child's stack, but is harmless since there is no caller to
return to in the child.
commit d26e0774a59bb7245b205bc8e7d8b35cc2037095 moved the detach state
transition at exit before the thread list lock was taken. this
inadvertently allowed pthread_join to race to take the thread list
lock first, and proceed with unmapping of the exiting thread's memory.
we could fix this by just revering the offending commit and instead
performing __vm_wait unconditionally before taking the thread list
lock, but that may be costly. instead, bring back the old DT_EXITING
vs DT_EXITED state distinction that was removed in commit
8f11e6127fe93093f81a52b15bb1537edc3fc8af, and don't transition to
DT_EXITED (a value of 0, which is what pthread_join waits for) until
after the lock has been taken.
after a non-normal-type process-shared mutex is unlocked, it's
immediately available to another thread to lock, unlock, and destroy,
but the first unlocking thread may still have a pointer to it in its
robust_list pending slot. this means, on async process termination,
the kernel may attempt to access and modify the memory that used to
contain the mutex -- memory that may have been reused for some other
purpose after the mutex was destroyed.
setting up for this kind of race to occur is difficult to begin with,
requiring dynamic use of shared memory maps, and actually hitting the
race is very difficult even with a suitable setup. so this is mostly a
theoretical fix, but in any case the cost is very low.
the __vm_wait operation can delay forward progress arbitrarily long if
a thread holding the lock is interrupted by a signal. in a worst case
this can deadlock. any critical section holding the thread list lock
must respect lock ordering contracts and must not take any lock which
is not AS-safe.
to fix, move the determination of thread joinable/detached state to
take place before the killlock and thread list lock are taken. this
requires reverting the atomic state transition if we determine that
the exiting thread is the last thread and must call exit, but that's
easy to do since it's a single-threaded context with application
as the outcome of Austin Group tracker issue #62, future editions of
POSIX have dropped the requirement that fork be AS-safe. this allows
but does not require implementations to synchronize fork with internal
locks and give forked children of multithreaded parents a partly or
fully unrestricted execution environment where they can continue to
use the standard library (per POSIX, they can only portably use
up until recently, taking this allowance did not seem desirable.
however, commit 8ed2bd8bfcb4ea6448afb55a941f4b5b2b0398c0 exposed the
extent to which applications and libraries are depending on the
ability to use malloc and other non-AS-safe interfaces in MT-forked
children, by converting latent very-low-probability catastrophic state
corruption into predictable deadlock. dealing with the fallout has
been a huge burden for users/distros.
while it looks like most of the non-portable usage in applications
could be fixed given sufficient effort, at least some of it seems to
occur in language runtimes which are exposing the ability to run
unrestricted code in the child as part of the contract with the
programmer. any attempt at fixing such contracts is not just a
technical problem but a social one, and is probably not tractable.
this patch extends the fork function to take locks for all libc
singletons in the parent, and release or reset those locks in the
child, so that when the underlying fork operation takes place, the
state protected by these locks is consistent and ready for the child
to use. locking is skipped in the case where the parent is
single-threaded so as not to interfere with legacy AS-safety property
of fork in single-threaded programs. lock order is mostly arbitrary,
but the malloc locks (including bump allocator in case it's used) must
be taken after the locks on any subsystems that might use malloc, and
non-AS-safe locks cannot be taken while the thread list lock is held,
imposing a requirement that it be taken last.
this change lifts undocumented restrictions on calls by replacement
mallocs to libc functions that might take these locks, and sets the
stage for lifting restrictions on the child execution environment
after multithreaded fork.
care is taken to #define macros to replace all four functions (malloc,
calloc, realloc, free) even if not all of them will be used, using an
undefined symbol name for the ones intended not to be used so that any
inadvertent future use will be caught at compile time rather than
directed to the wrong implementation.
introduced in commit 27b2fc9d6db956359727a66c262f1e69995660aa.
the reasoning in commit 2d0bbe6c788938d1332609c014eeebc1dff966ac was
not entirely correct. while it's true that setting the waiters flag
ensures that the next unlock will perform a wake, it's possible that
the wake is consumed by a mutex waiter that has no relationship with
the condvar wait queue being processed, which then takes the mutex.
when that thread subsequently unlocks, it sees no waiters, and leaves
the rest of the condvar queue stuck.
bring back the waiter count adjustment, but skip it for PI mutexes,
for which a successful lock-after-waiting always sets the waiters bit.
if future changes are made to bring this same waiters-bit contract to
all lock types, this can be reverted.
sem_open is required to return the same sem_t pointer for all
references to the same named semaphore when it's opened more than once
in the same process. thus we keep a table of all the mapped semaphores
and their reference counts. the code path for sem_close checked the
reference count, but then proceeded to unmap the semaphore regardless
of whether the count had reached zero.
add an immediate unlock-and-return for the nonzero refcnt case so the
property of performing the munmap syscall after releasing the lock can
pthread_cond_wait arranged for requeued waiters to wake when the mutex
is unlocked by temporarily adjusting the mutex's waiter count. commit
54ca677983d47529bab8752315ac1a2b49888870 broke this when introducing
PI mutexes by repurposing the waiter count field of the mutex
structure. since then, for PI mutexes, the waiter count adjustment was
misinterpreted by the mutex locking code as indicating that the mutex
is non a non-recoverable state.
it would be possible to special-case PI mutexes here, but instead just
drop all adjustment of the waiters count, and instead use the lock
word waiters bit for all mutex types. since the mutex is either held
by the caller or in unrecoverable state at the time the bit is set, it
will necessarily still be set at the time of any subsequent valid
unlock operation, and this will produce the desired effect of waking
the next waiter.
if waiter counts are entirely dropped at some point in the future this
code should still work without modification.
this makes the code slightly smaller and eliminates these functions
from relevance to possible future changes to multithreaded fork.
the barrier of a_store isn't technically needed here, but a_store is
used anyway for internal consistency of the memory model.
taking the deprecated/dropped vfork spec strictly, doing pretty much
anything but execve in the child is wrong and undefined. however,
these are commonly needed operations to setup the child state before
exec, and historical implementations tolerated them.
for single-threaded parents, these operations already worked as
expected in the vforked child. however, due to the need for __synccall
to synchronize id/resource limit changes among all threads, calling
these functions in the vforked child of a multithreaded parent caused
a misdirected broadcast signaling of all threads in the parent. these
signals could kill the parent entirely if the synccall signal handler
had never been installed in the parent, or could be ignored if it had,
or could signal/kill one or more utterly wrong processes if the parent
already terminated (due to vfork semantics, only possible via fatal
signal) and the parent tids were recycled. in any case, the expected
number of semaphore posts would never happen, so the child would
permanently hang (with all signals blocked) waiting for them.
to mitigate this, and also make the normal usage case work as
intended, treat the condition where the caller's actual tid does not
match the tid in its thread structure as single-threaded, and bypass
the entire synccall broadcast operation.
this code is only needed for pre-2.6 kernels, which are not actually
supported anyway, and was never tested. the fallback path using
SYS_modify_ldt failed to clear the upper bits of %eax (all ones due to
SYS_set_thread_area's return value being an error) before modifying
%al to attempt a new syscall.
dtv_copy, canary2, and canary_at_end existed solely to match multiple
ABI and asm-accessed layouts simultaneously. now that pthread_arch.h
can be included before struct __pthread is defined, the struct layout
can depend on macros defined by pthread_arch.h.
the previous commit addressing async-signal-safety issues around
pthread_kill did not fully fix pthread_cancel, which is also required
(albeit rather irrationally) to be async-cancel-safe.
without blocking implementation-internal signals, it's possible that,
when async cancellation is enabled, a cancel signal sent by another
thread interrupts pthread_kill while the killlock for a targeted
thread is held. as a result, the calling thread will terminate due to
cancellation without ever unlocking the targeted thread's killlock,
and thus the targeted thread will be unable to exit.
pthread_kill is required to be AS-safe. that requirement can't be met
if the target thread's killlock can be taken in contexts where
application-installed signal handlers can run.
block signals around use of this lock in all pthread_* functions which
target a tid, and reorder blocking/unblocking of signals in
pthread_exit so that they're blocked whenever the killlock is held.
the design used here relies on the barrier provided by the first lock
operation after the process returns to single-threaded state to
synchronize with actions by the last thread that exited. by storing
the intent to change modes in the same object used to detect whether
locking is needed, it's possible to avoid an extra (possibly costly)
memory load after the lock is taken.
after all but the last thread exits, the next thread to observe
libc.threads_minus_1==0 and conclude that it can skip locking fails to
synchronize with any changes to memory that were made by the
last-exiting thread. this can produce data races.
on some archs, at least x86, memory synchronization is unlikely to be
a problem; however, with the inline locks in malloc, skipping the lock
also eliminated the compiler barrier, and caused code that needed to
re-check chunk in-use bits after obtaining the lock to reuse a stale
value, possibly from before the process became single-threaded. this
in turn produced corruption of the heap state.
some uses of libc.threads_minus_1 remain, especially for allocation of
new TLS in the dynamic linker; otherwise, it could be removed
entirely. it's made non-volatile to reflect that the remaining
accesses are only made under lock on the thread list.
instead of libc.threads_minus_1, libc.threaded is now used for
skipping locks. the difference is that libc.threaded is permanently
true once an additional thread has been created. this will produce
some performance regression in processes that are mostly
single-threaded but occasionally creating threads. in the future it
may be possible to bring back the full lock-skipping, but more care
needs to be taken to produce a safe design.
since the backend for LOCK() skips locking if single-threaded, it's
unsafe to make the process appear single-threaded before the last use
this fixes potential unsynchronized access to a linked list via
commit 8a544ee3a2a75af278145b09531177cab4939b41 introduced a
dependency of the failure path for explicit scheduling at thread
creation on __clone's handling of the start function returning, which
should result in SYS_exit.
as noted in commit 05870abeaac0588fb9115cfd11f96880a0af2108, the arm
version of __clone was broken in this case. in the past, the mips
version was also broken; it was fixed in commit
since this code path is pretty much entirely untested (previously only
reachable in applications that call the public clone() and return from
the start function) and consists of fragile per-arch asm, don't assume
it works, at least not until it's been thoroughly tested. instead make
the SYS_exit syscall from the start function's failure path.
as noted in commit 05870abeaac0588fb9115cfd11f96880a0af2108, mov lr,pc
is not a valid method for saving the return address in code that might
be built as thumb.
this one is unlikely to matter, since any ISA level that has thumb2
should also have native implementations of atomics that don't involve
kuser_helper, and the affected code is only used on very old kernels
to begin with.
mov lr,pc is not a valid way to save the return address in thumb mode
since it omits the thumb bit. use a chain of bl and bx to emulate blx.
this could be avoided by converting to a .S file with preprocessor
conditions to use blx if available, but the time cost here is
dominated by the syscall anyway.
while making this change, also remove the remnants of support for
pre-bx ISA levels. commit 9f290a49bf9ee247d540d3c83875288a7991699c
removed the hack from the parent code paths, but left the unnecessary
code in the child. keeping it would require rewriting two code paths
rather than one, and is useless for reasons described in that commit.
previously, when pthread_create failed due to inability to set
explicit scheduling according to the requested attributes, the nascent
thread was detached and made responsible for its own cleanup via the
standard pthread_exit code path. this left it consuming resources
potentially well after pthread_create returned, in a way that the
application could not see or mitigate, and unnecessarily exposed its
existence to the rest of the implementation via the global thread
instead, attempt explicit scheduling early and reuse the failure path
for __clone failure if it fails. the nascent thread's exit futex is
not needed for unlocking the thread list, since the thread calling
pthread_create holds the thread list lock the whole time, so it can be
repurposed to ensure the thread has finished exiting. no pthread_exit
is needed, and freeing the stack, if needed, can happen just as it
would if __clone failed.
if setting scheduling properties succeeds, the new thread may end up
with lower priority than the caller, and may be unable to continue
running due to another intermediate-priority thread. this produces a
priority inversion situation for the thread calling pthread_create,
since it cannot return until the new thread reports success.
originally, the parent was responsible for setting the new thread's
priority; commits b8742f32602add243ee2ce74d804015463726899 and
40bae2d32fd6f3ffea437fa745ad38a1fe77b27e changed it as part of
trimming down the pthread structure. since then, commit
04335d9260c076cf4d9264bd93dd3b06c237a639 partly reversed the changes,
but did not switch responsibilities back. do that now.
commit 8f11e6127fe93093f81a52b15bb1537edc3fc8af wrongly documented
that all changes to libc.threads_minus_1 were guarded by the thread
list lock, but the decrement for failed SYS_clone took place after the
thread list lock was released.
The only reason we needed to preserve the link register was because we
were using a branch-link instruction to branch to __cp_cancel.
Replacing this with a branch means we can avoid the save/restore as
the link register is no longer modified.
per policy, define the feature test macro to get declarations for the
pthread_tryjoin_np and pthread_timedjoin_np functions. in the past
this has been only for checking; with 32-bit archs getting 64-bit
time_t it will also be necessary for symbols to get redirected
the x32 syscall interfaces treat timespec's tv_nsec member as 64-bit
despite the API type being long and long being 32-bit in the ABI. this
is no problem for syscalls that store timespecs to userspace as
results, but caused uninitialized padding to be misinterpreted as the
high bits in syscalls that take timespecs as input.
since the beginning of the port, we've dealt with this situation with
hacks in syscall_arch.h, and injected between __syscall_cp_c and
__syscall_cp_asm, to special-case the syscall numbers that involve
timespecs as inputs and copy them to a form suitable to pass to the
commit 40aa18d55ab763e69ad16d0cf1cebea708ffde47 set the stage for
removal of these hacks by letting us treat the "normal" x32 syscalls
dealing with timespec as if they're x32's "time64" syscalls,
effectively making x32 ax "time64-only 32-bit arch" like riscv32 will
be when it's added. since then, all users of syscalls that x32's
syscall_arch.h had hacks for have been updated to use time64 syscalls,
so the hacks can be removed.
there are still at least a few other timespec-related syscalls broken
on x32, which were overlooked when the x32 hacks were done or added
later. these include at least recvmmsg, adjtimex/clock_adjtime, and
timerfd_settime, and they will be fixed independently later on.
thanks to the original factorization using the __timedwait function,
there are no FUTEX_WAIT calls anywhere else, giving us a single point
of change to make nearly all the timed thread primitives time64-ready.
the one exception is the FUTEX_LOCK_PI command for PI mutex timedlock.
I haven't tried to make these two points share code, since they have
different fallbacks (no non-private fallback needed for PI since PI
was added later) and FUTEX_LOCK_PI isn't a cancellation point (thus
allowing the whole code path to inline into pthread_mutex_timedlock).
as for other changes in this series, the time64 syscall is used only
if it's the only one defined for the arch, or if the requested timeout
does not fit in 32 bits. on current 32-bit archs where time_t is a
32-bit type, this makes it statically unreachable.
on 64-bit archs, there are only superficial changes to the code after
preprocessing. on current 32-bit archs, the time is passed via an
intermediate copy to remove the assumption that time_t is a 32-bit
for namespace-safety with thrd_sleep, this requires an alias, which is
also added. this eliminates all but one direct call point for
nanosleep syscalls, and arranges that 64-bit time_t conversion logic
will only need to exist in one file rather than three.
as a bonus, clock_nanosleep with CLOCK_REALTIME and empty flags is now
implemented as SYS_nanosleep, thereby working on older kernels that
may lack POSIX clocks functionality.
Author: Alex Suykov <email@example.com>
Author: Aric Belsito <firstname.lastname@example.org>
Author: Drew DeVault <email@example.com>
Author: Michael Clark <firstname.lastname@example.org>
Author: Michael Forney <email@example.com>
Author: Stefan O'Rear <firstname.lastname@example.org>
This port has involved the work of many people over several years. I
have tried to ensure that everyone with substantial contributions has
been credited above; if any omissions are found they will be noted
later in an update to the authors/contributors list in the COPYRIGHT
The version committed here comes from the riscv/riscv-musl repo's
commit 3fe7e2c75df78eef42dcdc352a55757729f451e2, with minor changes by
me for issues found during final review:
- a_ll/a_sc atomics are removed (according to the ISA spec, lr/sc
are not safe to use in separate inline asm fragments)
- a_cas[_p] is fixed to be a memory barrier
- the call from the _start assembly into the C part of crt1/ldso is
changed to allow for the possibility that the linker does not place
them nearby each other.
- DTP_OFFSET is defined correctly so that local-dynamic TLS works
- reloc.h LDSO_ARCH logic is simplified and made explicit.
- unused, non-functional crti/n asm files are removed.
- an empty .sdata section is added to crt1 so that the
__global_pointer reference is resolvable.
- indentation style errors in some asm files are fixed.
the weak version of __syscall_cp_c was using a tail call to __syscall
to avoid duplicating the 6-argument syscall code inline in small
static-linked programs, but now that __syscall no longer exists, the
inline expansion is no longer duplication.
the syscall.h machinery suppported up to 7 syscall arguments, only via
an external __syscall function, but we presently have no syscall call
points that actually make use of that many, and the kernel only
defines 7-argument calling conventions for arm, powerpc (32-bit), and
sh. if it turns out we need them in the future, they can easily be
this is the first part of a series of patches intended to make
__syscall fully self-contained in the object file produced using
syscall.h, which will make it possible for crt1 code to perform
the (confusingly named) i386 __vsyscall mechanism, which this commit
removes, was introduced before the presence of a valid thread pointer
was mandatory; back then the thread pointer was setup lazily only if
threads were used. the intent was to be able to perform syscalls using
the kernel's fast entry point in the VDSO, which can use the sysenter
(Intel) or syscall (AMD) instruction instead of int $128, but without
inlining an access to the __syscall global at the point of each
syscall, which would incur a significant size cost from PIC setup
everywhere. the mechanism also shuffled registers/calling convention
around to avoid spills of call-saved registers, and to avoid
allocating ebx or ebp via asm constraints, since there are plenty of
broken-but-supported compiler versions which are incapable of
allocating ebx with -fPIC or ebp with -fno-omit-frame-pointer.
the new mechanism preserves the properties of avoiding spills and
avoiding allocation of ebx/ebp in constraints, but does it inline,
using some fairly simple register shuffling, and uses a field of the
thread structure rather than global data for the vdso-provided syscall
for now, the external __syscall function is refactored not to use the
old __vsyscall so it can be kept, but the intent is to remove it too.
commit 54ca677983d47529bab8752315ac1a2b49888870 inadvertently
introduced bitwise and where logical and was intended. since the
right-hand operand is always 0 or -1 whenever the left-hand operand is
nonzero, the behavior happened to be equivalent.
priority inheritance is a feature to mitigate priority inversion
situations, where a execution of a medium-priority thread can
unboundedly block forward progress of a high-priority thread when a
lock it needs is held by a low-priority thread.
the natural way to do priority inheritance would be with a simple
futex flag to donate the calling thread's priority to a target thread
while it waits on the futex. unfortunately, linux does not offer such
an interface, but instead insists on implementing the whole locking
protocol in kernelspace with special futex commands that exist solely
for the purpose of doing PI mutexes. this would require the entire
"trylock" logic to be duplicated in the timedlock code path for PI
mutexes, since, once the previous lock holder releases the lock and
the futex call returns, the lock is already held by the caller.
obviously such code duplication is undesirable.
instead, I've made the PI timedlock success path set the mutex lock
count to -1, which can be thought of as "not yet complete", since a
lock count of 0 is "locked, with no recursive references". a simple
branch in a non-hot path of pthread_mutex_trylock can then see and act
on this state, skipping past the code that would check and take the
lock to the same code path that runs after the lock is obtained for a
because we're forced to let the kernel perform the actual lock and
unlock operations whenever the mutex is contended, we have to patch
things up when it does the wrong thing:
1. the lock operation is not aware of whether the mutex is
error-checking, so it will always fail with EDEADLK rather than
2. the lock operation is not aware of whether the mutex is robust, so
it will successfully obtain mutexes in the owner-died state even if
they're non-robust, whereas this operation should deadlock.
3. the unlock operation always sets the lock value to zero, whereas
for robust mutexes, we want to set it to a special value indicating
that the mutex obtained after its owner died was unlocked without
marking it consistent, so that future operations all fail with
the first of these is easy to solve, just by performing a futex wait
on a dummy futex address to simulate deadlock or ETIMEDOUT as
appropriate. but problems 2 and 3 interact in a nasty way. to solve
problem 2, we need to back out the spurious success. but if waiters
are present -- which we can't just ignore, because even if we don't
want to wake them, the calling thread is incorrectly inheriting their
priorities -- this requires using the kernel's unlock operation, which
will zero the lock value, thereby losing the "owner died with lock
to solve these problems, we overload the mutex's waiters field, which
is unused for PI mutexes since they don't call the normal futex wait
functions, as an indicator that the PI mutex is permanently
non-lockable. originally I wanted to use the count field, but there is
one code path that needs to access this flag without synchronization:
trylock's CAS failure path needs to be able to decide whether to fail
with EBUSY or ENOTRECOVERABLE, the waiters field is already treated as
a relaxed-order atomic in our memory model, so this works out nicely.
there was no point in masking off the pshared bit when first loading
the type, since every subsequent access involves a mask anyway. not
masking it may avoid a subsequent load to check the pshared flag, and
it's just simpler.
commit 84d061d5a31c9c773e29e1e2b1ffe8cb9557bc58 wrongly moved the
access to the global next_key outside of the scope of the lock. the
error manifested as spurious failure to find an available key slot
under concurrent calls to pthread_key_create, since the stopping
condition could be met after only a small number of slots were
commit 2de29bc994029b903a366b8a4a9f8c3c3ee2be90 left behind one
reference to pthread_mutex_trylock. fixing this also improves code
generation due to the namespace-safe version being hidde.
the motivation for this change is twofold. first, it gets the fallback
logic out of the dynamic linker, improving code readability and
organization. second, it provides application code that wants to use
the membarrier syscall, which depends on preregistration of intent
before the process becomes multithreaded unless unbounded latency is
acceptable, with a symbol that, when linked, ensures that this
this is a prerequisite for factoring the membarrier fallback code into
a function that can be called from a context with the thread list
already locked or independently.
previously, dynamic loading of new libraries with thread-local storage
allocated the storage needed for all existing threads at load-time,
precluding late failure that can't be handled, but left installation
in existing threads to take place lazily on first access. this imposed
an additional memory access and branch on every dynamic tls access,
and imposed a requirement, which was not actually met, that the
dynamic tlsdesc asm functions preserve all call-clobbered registers
before calling C code to to install new dynamic tls on first access.
the x86[_64] versions of this code wrongly omitted saving and
restoring of fpu/vector registers, assuming the compiler would not
generate anything using them in the called C code. the arm and aarch64
versions saved known existing registers, but failed to be future-proof
against expansion of the register file.
now that we track live threads in a list, it's possible to install the
new dynamic tls for each thread at dlopen time. for the most part,
synchronization is not needed, because if a thread has not
synchronized with completion of the dlopen, there is no way it can
meaningfully request access to a slot past the end of the old dtv,
which remains valid for accessing slots which already existed.
however, it is necessary to ensure that, if a thread sees its new dtv
pointer, it sees correct pointers in each of the slots that existed
prior to the dlopen. my understanding is that, on most real-world
coherency architectures including all the ones we presently support, a
built-in consume order guarantees this; however, don't rely on that.
instead, the SYS_membarrier syscall is used to ensure that all threads
see the stores to the slots of their new dtv prior to the installation
of the new dtv. if it is not supported, the same is implemented in
userspace via signals, using the same mechanism as __synccall.
the __tls_get_addr function, variants, and dynamic tlsdesc asm
functions are all updated to remove the fallback paths for claiming
new dynamic tls, and are now all branch-free.
access to clear the entry in each thread's tsd array for the key being
deleted was not synchronized with __pthread_tsd_run_dtors. I probably
made this mistake from a mistaken belief that the thread list lock was
held during the latter, which of course is not possible since it
executes application code in a still-live-thread context.
while we're at it, expand the interval during which signals are
blocked to cover taking the write lock on key_lock, so that a signal
at an inopportune time doesn't block forward progress of readers.
commit 84d061d5a31c9c773e29e1e2b1ffe8cb9557bc58 inadvertently
introduced namespace violations by using the pthread-namespace rwlock
functions in pthread_key_create, which is in turn used for C11 tss.
fix that and possible future uses of rwlocks elsewhere.
with the availability of the thread list, there is no need to mark tsd
key slots dirty and clean them up only when a free slot can't be
found. instead, directly iterate threads and clear any value
associated with the key being deleted.
no synchronization is necessary for the clearing, since there is no
way the slot can be accessed without having synchronized with the
creation of a new key occupying the same slot, which is already
sequenced after and synchronized with the deletion of the old key.
the __synccall mechanism provides stop-the-world synchronous execution
of a callback in all threads of the process. it is used to implement
multi-threaded setuid/setgid operations, since Linux lacks them at the
kernel level, and for some other less-critical purposes.
this change eliminates dependency on /proc/self/task to determine the
set of live threads, which in addition to being an unwanted dependency
and a potential point of resource-exhaustion failure, turned out to be
inaccurate. test cases provided by Alexey Izbyshev showed that it
could fail to reflect newly created threads. due to how the
presignaling phase worked, this usually yielded a deadlock if hit, but
in the worst case it could also result in threads being silently
missed (allowed to continue running without executing the callback).
the hard problem here is unlinking threads from a list when they exit
without creating a window of inconsistency where the kernel task for a
thread still exists and is still executing instructions in userspace,
but is not reflected in the list. the magic solution here is getting
rid of per-thread exit futex addresses (set_tid_address), and instead
using the exit futex to unlock the global thread list.
since pthread_join can no longer see the thread enter a detach_state
of EXITED (which depended on the exit futex address pointing to the
detach_state), it must now observe the unlocking of the thread list
lock before it can unmap the joined thread and return. it doesn't
actually have to take the lock. for this, a __tl_sync primitive is
offered, with a signature that will allow it to be enhanced for quick
return even under contention on the lock, if needed. for now, the
exiting thread always performs a futex wake on its detach_state. a
future change could optimize this out except when there is already a
initial/dynamic variants of detached state no longer need to be
tracked separately, since the futex address is always set to the
global list lock, not a thread-local address that could become invalid
on detached thread exit. all detached threads, however, must perform a
second sigprocmask syscall to block implementation-internal signals,
since locking the thread list with them already blocked is not
the arch-independent C version of __unmapself no longer needs to take
a lock or setup its own futex address to release the lock, since it
must necessarily be called with the thread list lock already held,
guaranteeing exclusive access to the temporary stack.
changes to libc.threads_minus_1 no longer need to be atomic, since
they are guarded by the thread list lock. it is largely vestigial at
this point, and can be replaced with a cheaper boolean indicating
whether the process is multithreaded at some point in the future.
whether signals need to be blocked at thread start, and whether
unblocking is necessary in the entry point function, has historically
depended on intricacies of the cancellation design and on whether
there are scheduling operations to perform on the new thread before
its successful creation can be committed. future changes to track an
AS-safe list of live threads will require signals to be blocked
whenever changes are made to the list, so ...
prior to commits b8742f32602add243ee2ce74d804015463726899 and
40bae2d32fd6f3ffea437fa745ad38a1fe77b27e, a signal mask for the entry
function to restore was part of the pthread structure. it was removed
to trim down the size of the structure, which both saved a small
amount of stack space and improved code generation on archs where
small immediate displacements are less costly than arbitrary ones, by
limiting the range of offsets between the base of the thread
structure, its members, and the thread pointer. these commits moved
the saved mask to a special structure used only when special
scheduling was needed, in which case the pthread_create caller and new
thread had to synchronize with each other and could use this memory to
pass a mask.
this commit partially reverts the above two commits, but instead of
putting the mask back in the pthread structure, it moves all "start
argument" members out of the pthread structure, trimming it down
further, and puts them in a separate structure passed on the new
thread's stack. the code path for explicit scheduling of the new
thread is also changed to synchronize with the calling thread in such
a way to avoid spurious futex wakes.
in order to implement ENOTRECOVERABLE, the implementation has
traditionally used a bit of the mutex type field to indicate that it's
recovered after EOWNERDEAD and will go into ENOTRECOVERABLE state if
pthread_mutex_consistent is not called before unlocking. while it's
only the thread that holds the lock that needs access to this
information (except possibly for the sake of pthread_mutex_consistent
choosing between EINVAL and EPERM for erroneous calls), the change to
the type field is formally a data race with all other threads that
perform any operation on the mutex. no individual bits race, and no
write races are possible, so things are "okay" in some sense, but it's
still not good.
this patch moves the recovery/consistency state to the mutex
owner/lock field which is rightfully mutable. bit 30, the same bit the
kernel uses with a zero owner to indicate that the previous owner died
holding the lock, is now used with a nonzero owner to indicate that
the mutex is held but has not yet been marked consistent. note that
the kernel ABI also reserves bit 29 not to appear in any tid, so the
sentinel value we use for ENOTRECOVERABLE, 0x7fffffff, does not clash
with any tid plus bit 30.