It's hard to disable ASLR on current macOS releases, for testing with
-DEXEC_BACKEND. You could already set the environment variable
PG_SHMEM_ADDR to something not likely to collide with mappings created
earlier in process startup. Let's also provide a default value that
works on current releases and architectures, for developer convenience.
As noted in the pre-existing comment, this is a horrible hack, but
-DEXEC_BACKEND is only used by Unix-based PostgreSQL developers for
testing some otherwise Windows-only code paths, so it seems excusable.
Back-patch to all supported branches.
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/20210806032944.m4tz7j2w47mant26%40alap3.anarazel.de
Also "make reformat-dat-files".
The only change worthy of note is that pgindent messed up the formatting
of launcher.c's struct LogicalRepWorkerId, which led me to notice that
that struct wasn't used at all anymore, so I just took it out.
In the error messages referring to the user right "Lock pages in
memory", this is a term from the Windows OS, so it should be
translated in accordance with the OS localization. Refactor the error
messages so this is easier and clearer. Also fix the capitalization
to match the existing capitalization in the OS.
After de8feb1f3a, some warnings remained
that were only visible when using GCC on Windows. Fix those as well.
Note that the ecpg test source files don't use the full pg_config.h,
so we can't use pg_funcptr_t there but have to do it the long way.
Up to now, only ECONNRESET (and EPIPE, in most but not quite all places)
received special treatment in our error handling logic. This patch
changes things so that related error codes such as ECONNABORTED are
also recognized as indicating that the connection's dead and unlikely
to come back.
We continue to think, however, that only ECONNRESET and EPIPE should be
reported as probable server crashes; the other cases indicate network
connectivity problems but prove little about the server's state. Thus,
there's no change in the error message texts that are output for such
cases. The key practical effect is that errcode_for_socket_access()
will report ERRCODE_CONNECTION_FAILURE rather than
ERRCODE_INTERNAL_ERROR for a network failure. It's expected that this
will fix buildfarm member lorikeet's failures since commit 32a9c0bdf,
as that seems to be due to not treating ECONNABORTED equivalently to
ECONNRESET.
The set of errnos treated this way now includes ECONNABORTED, EHOSTDOWN,
EHOSTUNREACH, ENETDOWN, ENETRESET, and ENETUNREACH. Several of these
were second-class citizens in terms of their handling in places like
get_errno_symbol(), so upgrade the infrastructure where necessary.
As committed, this patch assumes that all these symbols are defined
everywhere. POSIX specifies all of them except EHOSTDOWN, but that
seems to exist on all platforms of interest; we'll see what the
buildfarm says about that.
Probably this should be back-patched, but let's see what the buildfarm
thinks of it first.
Fujii Masao and Tom Lane
Discussion: https://postgr.es/m/2621622.1602184554@sss.pgh.pa.us
Since its inception, our Windows signal emulation code has worked by
running a main signal thread that just watches for incoming signal
requests, and then spawns a new thread to handle each such request.
That design is meant for servers in which requests can take substantial
effort to process, and it's worth parallelizing the handling of
requests. But those assumptions are just bogus for our signal code.
It's not much more than pg_queue_signal(), which is cheap and can't
parallelize at all, plus we don't really expect lots of signals to
arrive at the same backend at once. More importantly, this approach
creates failure modes that we could do without: either inability to
spawn a new thread or inability to create a new pipe handle will risk
loss of signals. Hence, dispense with the separate per-signal threads
and just service each request in-line in the main signal thread. This
should be a bit faster (for the normal case of one signal at a time)
as well as more robust.
Patch by me; thanks to Andrew Dunstan for testing and Amit Kapila
for review.
Discussion: https://postgr.es/m/4412.1575748586@sss.pgh.pa.us
pg_signal_dispatch_thread() responded to the client (signal sender)
and disconnected the pipe before actually setting the shared variables
that make the signal visible to the backend process's main thread.
In the worst case, it seems, effective delivery of the signal could be
postponed for as long as the machine has any other work to do.
To fix, just move the pg_queue_signal() call so that we do it before
responding to the client. This essentially makes pgkill() synchronous,
which is a stronger guarantee than we have on Unix. That may be
overkill, but on the other hand we have not seen comparable timing bugs
on any Unix platform.
While at it, add some comments to this sadly underdocumented code.
Problem diagnosis and fix by Amit Kapila; I just added the comments.
Back-patch to all supported versions, as it appears that this can cause
visible NOTIFY timing oddities on all of them, and there might be
other misbehavior due to slow delivery of other signals.
Discussion: https://postgr.es/m/32745.1575303812@sss.pgh.pa.us
When maintaining or merging patches, one of the most common sources
for conflicts are the list of objects in makefiles. Especially when
the split across lines has been changed on both sides, which is
somewhat common due to attempting to stay below 80 columns, those
conflicts are unnecessarily laborious to resolve.
By splitting, and alphabetically sorting, OBJS style lines into one
object per line, conflicts should be less frequent, and easier to
resolve when they still occur.
Author: Andres Freund
Discussion: https://postgr.es/m/20191029200901.vww4idgcxv74cwes@alap3.anarazel.de
This approach provides a much tighter binding between a data directory
and the associated SysV shared memory block (and SysV or named-POSIX
semaphores, if we're using those). Key collisions are still possible,
but only between data directories stored on different filesystems,
so the situation should be negligible in practice. More importantly,
restarting the postmaster with a different port number no longer
risks failing to identify a relevant shared memory block, even when
postmaster.pid has been removed. A standalone backend is likewise
much more certain to detect conflicting leftover backends.
(In the longer term, we might now think about deprecating the port as
a cluster-wide value, so that one postmaster could support sockets
with varying port numbers. But that's for another day.)
The hazards fixed here apply only on Unix systems; our Windows code
paths already use identifiers derived from the data directory path
name rather than the port.
src/test/recovery/t/017_shm.pl, which intends to test key-collision
cases, has been substantially rewritten since it can no longer use
two postmasters with identical port numbers to trigger the case.
Instead, use Perl's IPC::SharedMem module to create a conflicting
shmem segment directly. The test script will be skipped if that
module is not available. (This means that some older buildfarm
members won't run it, but I don't think that that results in any
meaningful coverage loss.)
Patch by me; thanks to Noah Misch and Peter Eisentraut for discussion
and review.
Discussion: https://postgr.es/m/16908.1557521200@sss.pgh.pa.us
There's a very old race condition in our code to see whether a pre-existing
shared memory segment is still in use by a conflicting postmaster: it's
possible for the other postmaster to remove the segment in between our
shmctl() and shmat() calls. It's a narrow window, and there's no risk
unless both postmasters are using the same port number, but that's possible
during parallelized "make check" tests. (Note that while the TAP tests
take some pains to choose a randomized port number, pg_regress doesn't.)
If it does happen, we treated that as an unexpected case and errored out.
To fix, allow EINVAL to be treated as segment-not-present, and the same
for EIDRM on Linux. AFAICS, the considerations here are basically
identical to the checks for acceptable shmctl() failures, so I documented
and coded it that way.
While at it, adjust PGSharedMemoryAttach's API to remove its undocumented
dependency on UsedShmemSegAddr in favor of passing the attach address
explicitly. This makes it easier to be sure we're using a null shmaddr
when probing for segment conflicts (thus avoiding questions about what
EINVAL means). I don't think there was a bug there, but it required
fragile assumptions about the state of UsedShmemSegAddr during
PGSharedMemoryIsInUse.
Commit c09850992 may have made this failure more probable by applying
the conflicting-segment tests more often. Hence, back-patch to all
supported branches, as that was.
Discussion: https://postgr.es/m/22224.1557340366@sss.pgh.pa.us
postmaster startup scrutinizes any shared memory segment recorded in
postmaster.pid, exiting if that segment matches the current data
directory and has an attached process. When the postmaster.pid file was
missing, a starting postmaster used weaker checks. Change to use the
same checks in both scenarios. This increases the chance of a startup
failure, in lieu of data corruption, if the DBA does "kill -9 `head -n1
postmaster.pid` && rm postmaster.pid && pg_ctl -w start". A postmaster
will no longer stop if shmat() of an old segment fails with EACCES. A
postmaster will no longer recycle segments pertaining to other data
directories. That's good for production, but it's bad for integration
tests that crash a postmaster and immediately delete its data directory.
Such a test now leaks a segment indefinitely. No "make check-world"
test does that. win32_shmem.c already avoided all these problems. In
9.6 and later, enhance PostgresNode to facilitate testing. Back-patch
to 9.4 (all supported versions).
Reviewed (in earlier versions) by Daniel Gustafsson and Kyotaro HORIGUCHI.
Discussion: https://postgr.es/m/20190408064141.GA2016666@rfd.leadboat.com
We've long had reports of intermittent "could not reattach to shared
memory" errors on Windows. Buildfarm member dory fails that way when
PGSharedMemoryReAttach() execution overlaps with creation of a thread
for the process's "default thread pool". Fix that by providing a second
region to receive asynchronous allocations that would otherwise intrude
into UsedShmemSegAddr. In pgwin32_ReserveSharedMemoryRegion(), stop
trying to free reservations landing at incorrect addresses; the caller's
next step has been to terminate the affected process. Back-patch to 9.4
(all supported versions).
Reviewed by Tom Lane. He also did much of the prerequisite research;
see commit bcbf2346d6.
Discussion: https://postgr.es/m/20190402135442.GA1173872@rfd.leadboat.com
postmaster startup scrutinizes any shared memory segment recorded in
postmaster.pid, exiting if that segment matches the current data
directory and has an attached process. When the postmaster.pid file was
missing, a starting postmaster used weaker checks. Change to use the
same checks in both scenarios. This increases the chance of a startup
failure, in lieu of data corruption, if the DBA does "kill -9 `head -n1
postmaster.pid` && rm postmaster.pid && pg_ctl -w start". A postmaster
will no longer recycle segments pertaining to other data directories.
That's good for production, but it's bad for integration tests that
crash a postmaster and immediately delete its data directory. Such a
test now leaks a segment indefinitely. No "make check-world" test does
that. win32_shmem.c already avoided all these problems. In 9.6 and
later, enhance PostgresNode to facilitate testing. Back-patch to 9.4
(all supported versions).
Reviewed by Daniel Gustafsson and Kyotaro HORIGUCHI.
Discussion: https://postgr.es/m/20130911033341.GD225735@tornado.leadboat.com
elog.c has long had a private strerror wrapper that handles assorted
possible failures or deficiencies of the platform's strerror. On Windows,
it also knows how to translate Winsock error codes, which the native
strerror does not. Move all this code into src/port/strerror.c and
define strerror() as a macro that invokes it, so that both our frontend
and backend code will have all of this behavior.
I believe this constitutes an actual bug fix on Windows, since AFAICS
our frontend code did not report Winsock error codes properly before this.
However, the main point is to lay the groundwork for implementing %m
in src/port/snprintf.c: the behavior we want %m to have is this one,
not the native strerror's.
Note that this throws away the prior use of src/port/strerror.c,
which was to implement strerror() on platforms lacking it. That's
been dead code for nigh twenty years now, since strerror() was
already required by C89.
We should likewise cause strerror_r to use this behavior, but
I'll tackle that separately.
Patch by me, reviewed by Michael Paquier
Discussion: https://postgr.es/m/2975.1526862605@sss.pgh.pa.us
In a USE_UNNAMED_SEMAPHORES build, the default on Linux and FreeBSD
since commit ecb0d20a, we have an array of sem_t objects. This
turned out to reduce performance compared to the previous default
USE_SYSV_SEMAPHORES on an 8 socket system. Testing showed that the
lost performance could be regained by padding the array elements so
that they have their own cache lines. This matches what we do for
similar hot arrays (see LWLockPadded, WALInsertLockPadded).
Back-patch to 10, where unnamed semaphores were adopted as the default
semaphore interface on those operating systems.
Author: Thomas Munro
Reviewed-by: Andres Freund
Reported-by: Mithun Cy
Tested-by: Mithun Cy, Tom Lane, Thomas Munro
Discussion: https://postgr.es/m/CAD__OugYDM3O%2BdyZnnZSbJprSfsGFJcQ1R%3De59T3hcLmDug4_w%40mail.gmail.com
Revert the map/unmap dance I tried in commit 73042b8d1; that helps
not at all.
Instead, speculate that the unwanted allocation is being done on
another thread, and thus timing variations explain the apparent
unpredictability. Temporarily add a 1-second sleep before the
VirtualFree call, in hopes that any such other threads will
quiesce and not jog our elbow.
This is obviously not a desirable long-term fix, but as a means of
investigation it seems useful.
Discussion: https://postgr.es/m/25495.1524517820@sss.pgh.pa.us
The idea here is to get Windows' userspace infrastructure to allocate
whatever space it needs for MapViewOfFileEx() before we release the
locked-down space that we want to map the shared memory block into.
This is a fairly brute-force attempt, and would likely (for example)
fail with large shared memory on 32-bit Windows. We could perhaps
ameliorate that by mapping only part of the shared memory block in
this way, but for the moment I just want to see if this approach
will fix dory's problem.
Discussion: https://postgr.es/m/25495.1524517820@sss.pgh.pa.us
Rather than elog'ing immediately, push the map data into a preallocated
StringInfo. Perhaps this will prevent some of the mid-operation
allocations that are evidently happening now.
Discussion: https://postgr.es/m/25495.1524517820@sss.pgh.pa.us
This code is evidently allocating memory and thus confusing matters
even more. Let's see whether we can learn anything with
just VirtualQuery.
Discussion: https://postgr.es/m/25495.1524517820@sss.pgh.pa.us
This morning's results from buildfarm member dory make it pretty
clear that something is getting mapped into the just-freed space,
but not what that something is. Replace my minimalistic probes
with a full dump of the process address space and module space,
based on Noah's work at
<20170403065106.GA2624300%40tornado.leadboat.com>
This is all (probably) to get reverted once we have fixed the
problem, but for now we need information.
Discussion: https://postgr.es/m/25495.1524517820@sss.pgh.pa.us
After some thought about the info captured so far, it seems possible
that MapViewOfFileEx is itself causing some DLL to get loaded into
the space just freed by VirtualFree. The previous commit here didn't
capture enough info to really prove the case for that, so let's add
one more VirtualQuery in between those steps. Also, be sure to
capture the post-Map state before we emit any log entries, just in
case elog() is invoking some code not previously loaded.
Discussion: https://postgr.es/m/25495.1524517820@sss.pgh.pa.us
Add some debug printouts focused on the idea that MapViewOfFileEx might
be rounding its virtual memory allocation up more than we expect (and,
in particular, more than VirtualAllocEx does).
Once we've seen what this reports in one of the failures on buildfarm
members dory or jacana, we might revert this ... or perhaps just
decrease the log level.
Discussion: https://postgr.es/m/25495.1524517820@sss.pgh.pa.us
Thomas Munro
Tom Lane
Peter Eisentraut
Bruce Momjian
Andres Freund
Michael Paquier
Noah Misch
Magnus Hagander
Peter Eisentraut