in the case where malloc is being replaced, it's not valid to call
malloc between final relocations and main app's crt1 entry point; on
fdpic archs the main app's entry point will not yet have performed the
self-fixups necessary to call its code.

to fix, reorder queue_ctors before final relocations. an alternative
solution would be doing the allocation from __libc_start_init, after
the entry point but before any ctors run. this is less desirable,
since it would leave a call to malloc that might be provided by the
application happening at startup when doing so can be easily avoided.

previously, going way back, there was simply no synchronization here.
a call to exit concurrent with ctor execution from dlopen could cause
a dtor to execute concurrently with its corresponding ctor, or could
cause dtors for newly-constructed libraries to be skipped.

introduce a shutting_down state that blocks further ctor execution,
producing the quiescence the dtor execution loop needs to ensure any
kind of consistency, and that blocks further calls to dlopen so that a
call into dlopen from a dtor cannot deadlock.

better approaches to some of this may be possible, but the changes
here at least make things safe.

previously, shared library constructors at program start and dlopen
time were executed in reverse load order. some libraries, however,
rely on a depth-first dependency order, which most other dynamic
linker implementations provide. this is a much more reasonable, less
arbitrary order, and it turns out to have much better properties with
regard to how slow-running ctors affect multi-threaded programs, and
how recursive dlopen behaves.

this commit builds on previous work tracking direct dependencies of
each dso (commit 40355569), and
performs a topological sort on the dependency graph at load time while
the main ldso lock is held and before success is committed, producing
a queue of constructors needed by the newly-loaded dso (or main
application). in the case of circular dependencies, the dependency
chain is simply broken at points where it becomes circular.

when the ctor queue is run, the init_fini_lock is held only for
iteration purposes; it's released during execution of each ctor, so
that arbitrarily-long-running application code no longer runs with a
lock held in the caller. this prevents a dlopen with slow ctors in one
thread from arbitrarily delaying other threads that call dlopen.
fully-independent ctors can run concurrently; when multiple threads
call dlopen with a shared dependency, one will end up executing the
ctor while the other waits on a condvar for it to finish.

another corner case improved by these changes is recursive dlopen
(call from a ctor). previously, recursive calls to dlopen could cause
a ctor for a library to be executed before the ctor for its
dependency, even when there was no relation between the calling
library and the library it was loading, simply due to the naive
reverse-load-order traversal. now, we can guarantee that recursive
dlopen in non-circular-dependency usage preserves the desired ctor
execution order properties, and that even in circular usage, at worst
the libraries whose ctors call dlopen will fail to have completed
construction when ctors that depend on them run.

init_fini_lock is changed to a normal, non-recursive mutex, since it
is no longer held while calling back into application code.

this makes calling dlsym on the main app more consistent with the
global symbol table (load order), and is a prerequisite for
dependency-order ctor execution to work correctly with LD_PRELOAD.

commit 40355569 introduced runtime
realloc of an array that may have been allocated before symbols were
resolved outside of libc, which is invalid if the allocator has been
replaced. track this condition and manually copy if needed.

dlsym with an explicit handle is specified to use "dependency order",
a breadth-first search rooted at the argument. this has always been
implemented by iterating a flattened dependency list built at dlopen
time. however, the logic for building this list was completely wrong
except in trivial cases; it simply used the list of libraries loaded
since a given library, and their direct dependencies, as that
library's dependencies, which could result in misordering, wrongful
omission of deep dependencies from the search, and wrongful inclusion
of unrelated libraries in the search.

further, libraries did not have any recorded list of resolved
dependencies until they were explicitly dlopened, meaning that
DT_NEEDED entries had to be resolved again whenever a library
participated as a dependency of more than one dlopened library.

with this overhaul, the resolved direct dependency list of each
library is always recorded when it is first loaded, and can be
extended to a full flattened breadth-first search list if dlopen is
called on the library. the extension is performed using the direct
dependency list as a queue and appending copies of the direct
dependency list of each dependency in the queue, excluding duplicates,
until the end of the queue is reached. the direct deps remain
available for future use as the initial subarray of the full deps
array.

first-load logic in dlopen is updated to match these changes, and
clarified.

code introduced in commit 9d44b646
wrongly attempted to read past the end of the currently-installed dtv
to determine if a dso provides new, not-already-installed tls. this
logic was probably leftover from an earlier draft of the code that
wrongly installed the new dtv before populating it.

it would work if we instead queried the new, not-yet-installed dtv,
but instead, replace the incorrect check with a simple range check
against old_cnt. this also catches modules that have no tls at all
with a single condition.

code introduced in commit 9d44b646
wrongly assumed the dso list tail was the right place to find new dtv
storage. however, this is only true if the last-loaded dependency has
tls. the correct place to get it is the dso corresponding to the tls
module list tail. introduce a container_of macro to get it, and use
it.

ultimately, dynamic tls allocation should be refactored so that this
is not an issue. there is no reason to be allocating new dtv space at
each load_library; instead it could happen after all new libraries
have been loaded but before they are committed. such changes may be
made later, but this commit fixes the present regression.

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
registration happens.

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.

commit 9d44b646 inadvertently
contained leftover logic from a previous approach to the fallback
signaling loop. it had no adverse effect, since j was always nonzero
if the loop body was reachable, but it makes no sense to be there with
the current approach to avoid signaling self.

addressing &out[k].sa was arguably undefined, despite &out[k] being
defined the slot one past the end of an array, since the member access
.sa is intervening between the [] operator and the & operator.

the backindex stored by getaddrinfo to allow freeaddrinfo to perform
partial-free wrongly used the address result index, rather than the
output slot index, and thus was only valid when they were equal
(nservs==1).

patch based on report with proposed fix by Markus Wichmann.

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 84d061d5 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
joiner waiting.

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
permissible.

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 b8742f32 and
40bae2d3, 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.

this eliminates some ugly hacks that were repurposing the start
function and start argument fields in the pthread structure for timer
use, and the need to longjmp out of a signal handler.

__dl_thread_cleanup is called from the context of an exiting thread
that is not in a consistent state valid for calling application code.
since commit c9f415d7, it's possible
(and supported usage) for the allocator to have been replaced by the
application, so __dl_thread_cleanup can no longer call free. instead,
reuse the message buffer as a linked-list pointer, and queue it to be
freed the next time any dynamic linker error message is generated.

the way gets was implemented in terms of fgets, it used the location
of the null termination to determine where to find and remove the
newline, if any. an embedded null byte prevented this from working.

this also fixes a one-byte buffer overflow, whereby when gets read an
N-byte line (not counting newline), it would store two null
terminators for a total of N+2 bytes. it's unlikely that anyone would
care that a function whose use is pretty much inherently a buffer
overflow writes too much, but it could break the only possible correct
uses of this function, in conjunction with input of known format from
a trusted/same-privilege-domain source, where the buffer length may
have been selected to exactly match a line length contract.

there seems to be no correct way to implement gets in terms of a
single call to fgets or scanf, and using multiple calls would require
explicit locking, so we might as well just write the logic out
explicitly character-at-a-time. this isn't fast, but nobody cares if a
catastrophically unsafe function that's so bad it was removed from the
C language is fast.

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.

fdopendir is specified to fail with EBADF if the file descriptor
passed is not open for reading. while O_PATH is an extension and
arguably exempt from this requirement, it's used, albeit incompletely,
to implement O_SEARCH, and fdopendir should fail when passed an
O_SEARCH file descriptor.

the new check is performed after fstat so that we don't have to
consider the possibility that the fd is invalid.

an alternate solution would be attempting to pre-fill the buffer using
getdents, which would fail with EBADF for us, but that seems more
complex and error-prone and involves either code duplication or
refactoring, so the simple fix with an additional inexpensive syscall
is what I've made for now.

Some packages call gettext to format a message to be sent to perror.
If the currently set user locale points to a non-existent .mo file,
open via __map_file in dcngettext will set errno to ENOENT.

Maintainer's notes: Non-modification of errno is a documented part of
the interface contract for the GNU version of this function and likely
other versions. The issue being fixed here seems to be a regression
from commit 1b52863e, which enabled
setting of errno from __map_file.

commit a6054e3c removed the argument,
making it a constraint violation to pass one. caught by cparser/firm;
other compilers seem to ignore it.

apparently some distros use this form, and it seems to be supported in
the gcc build system.

commit 84d061d5 attempted to do this
already, but omitted from pthread_key_create.c the weak definition of
__pthread_key_delete_synccall, so that the definition provided by
pthread_key_delete.c was always pulled in.

based on patch by Markus Wichmann, but with a weak alias rather than
weak reference for consistency/policy about dependence on tooling
features.

fallback to /etc/shadow should happen only when the entry is not found
in the TCB shadow. otherwise transient errors or permission errors can
cause inconsistent results.

this case is specified as success with a null result, rather than an
error, and errno is not to be set on success.

this reverts commit c0ed5a20, which
was based on a mistaken reading of POSIX due to inconsistency between
the description (which requires return upon interruption by a signal)
and the errors list (which wrongly lists EINTR as "may fail").

since the previously-introduced behavior was a workaround for an old
kernel bug to ensure safety of correct programs that were not hardened
against the bug, an effort has been made to preserve it for programs
which do not use interrupting signal handlers. the stage for this was
set in commit a63c0104, which makes
the futex __timedwait backend suppress EINTR if it's seen when no
interrupting signal handlers have been installed.

based loosely on a patch submitted by Orivej Desh, but with
unnecessary additional changes removed.

the resolution of Austin Group issue #1132 changes the requirement to
fail so that it only applies when the set argument (new mask) is
non-null. this change was made for consistency with the description,
which specified "if set is a null pointer, the value of the argument
how is not significant".

prior to linux 2.6.22, futex wait could fail with EINTR even for
non-interrupting (SA_RESTART) signals. this was no problem provided
the caller simply restarted the wait, but sem_[timed]wait is required
by POSIX to return when interrupted by a signal. commit
a113434c introduced this behavior, and
commit c0ed5a20 reverted it based on a
mistaken belief that it was not required. this belief stems from a bug
in the specification: the description requires the function to return
when interrupted, but the errors section marks EINTR as a "may fail"
condition rather than a "shall fail" one.

since there does seem to be significant value in the change made in
commit c0ed5a20, making it so that
programs that call sem_wait without checking for EINTR don't silently
make forward progress without obtaining the semaphore or treat it as a
fatal error and abort, add a behind-the-scenes mechanism in the
__timedwait backend to suppress EINTR in programs that have never
installed interrupting signal handlers, and have sigaction track and
report this state. this way the semaphore code is not cluttered by
workarounds and can be updated (to be done in next commit) to reflect
the high-level logic for conforming behavior.

these changes are based loosely on a patch by Markus Wichmann, with
the main changes being atomic update to flag object and moving the
workaround from sem_timedwait to the __timedwait futex backend.

it's not clear whether this is required, but it seems arguable that it
should happen. for example aio_suspend is supposed to return
immediately if any of the operations has "completed", which includes
ending with an error status asynchonously and might also be
interpreted to include doing so synchronously.

the map structures in particular are permanent once created, and thus
a large number of aio function calls with invalid file descriptors
could exhaust memory, whereas, assuming normal resource limits, only a
very small number of entries ever need to be allocated. check validity
of the fd before allocating anything new, so that allocation of large
amounts of memory is only possible when resource limits have been
increased and a large number of files are actually open.

this change also improves error reporting for bad file descriptors to
happen at the time the aio submission call is made, as opposed to
asynchronously.

since commit c9f415d7, it has been
possible that the allocator is application-provided code, which cannot
necessarily run safely on io thread stacks, and which should not be
able to see the existence of io threads, since they are an
implementation detail.

instead of having the io thread request and possibly allocate its
queue (and the map structures leading to it), make the submitting
thread responsible for this, and pass the queue pointer into the io
thread via its args structure. this eliminates the only early error
case in io threads, making it no longer necessary to pass an error
status back to the submitting thread via the args structure.