| Date | Commit message (Collapse) |
|---|
|
I'm seeing ENOBUFS on a RAM-starved system, and slowing the
sender down enough for the receiver to drain the buffers seems
to work. ENOMEM and ETOOMANYREFS could be in the same boat
as ENOBUFS.
Watching for POLLOUT events via select/poll/epoll_wait doesn't
seem to work, since the kernel can already sleep (or return
EAGAIN) for cases where POLLOUT would work.
|
|
I'm not sure why, but I noticed the one of my latest restarts of
public-inbox-httpd wasn't loading the Inline::C .so for Gcf2 nor
Spawn. I also can't reproduce the problem as both .so files are
loaded fine on a restart with zero config changes.
In any case, some extra, automatic diagnostics for build errors
won't hurt, as no extra noise is introduced for successful builds.
This will also make future development of C code more convenient,
hopefully.
|
|
Apparently this feature is only in Perl 5.12+, and we're
still on Perl 5.10.
|
|
We'll be using this to allow the "git clone" process hierarchy
to be killed via Ctrl-C. This also fixes a long-standing bug
in error reporting for the Inline::C version, because we're
actually testing for errors, now!
n.b. strlen(3) is officially async-signal-safe as of
POSIX.1-2016, but I can't think of a reason any previous
implementation prior to that wouldn't be.
|
|
We continue to unblock SIGCHLD unconditionally, but also
any signals not blocked by the parent (wq_worker).
This will allow Ctrl-C (SIGINT) to stop "git clone" and allow
git-clone cleanup to be performed and other long-running
processes when pi_fork_exec supports setpgid(2). This won't
affect existing daemons on systems with signalfd(2) or
EVFILT_SIGNAL at all, since those run with signals blocked
anyways.
|
|
There'll probably be more things which work on both GNU and
*BSD systems which we don't need separate strings for.
|
|
Keeping track of non-standard FDs gets tricky, so make it easier
by relying on st_dev/st_ino mapping in the transmitted objects.
We'll keep using numbers for the standard FDs since we need to
be able to easily redirect them in the producer (main daemon)
process for (gzip|bzip2|xz) if writing to a compressed mbox.
|
|
Mainly around fork() calls, but some nearby places as well.
|
|
It doesn't appear Perl (as of 5.32.x) has any internal
optimization for splitting on a single-byte, so give it
a regexp instead of letting it compile and discard a
new one every single time.
|
|
We don't need the result of query_prepare (for augmenting or
mass unlinking) until we're ready to deduplicate and write
results to the filesystem. This ought to let us hide some of
the cost of Xapian searches on multi-device/core systems for
extremely expensive searches.
|
|
With 4 dedicated workers, this seems to provide a 100-120%
speedup on a 4 core machine when writing thousands of search
results to a Maildir or mbox. This also sets us up for
high-latency IMAP destinations in the future.
This opens the door to more speedup opportunities such
as optimizing dedupe locking and other ways to reduce
contention.
This change is fairly complex and convoluted, unfortunately.
Further work may allow us to simplify it and even improve
performance.
|
|
We'll ensure our {send,recv}_cmd4 implementations are
consistent w.r.t. non-blocking and interrupted sockets.
We'll also support receiving messages without FDs associated
so we don't have to send dummy FDs to keep receivers from
reporting EOF.
|
|
It's easier to make the code more generic by transferring
all four FDs (std(in|out|err) + socket) instead of omitting
stdin.
We'll be reading from stdin on some imports, and possibly
outputting to stdout, so omitting stdin now would needlessly
complicate things.
The differences with IO::FDPass "1" code paths and the "4"
code paths used by Inline::C and Socket::MsgHdr are far too
much to support and test at the moment.
|
|
Actually, sending 4 FDs will be useful for lei internal xsearch
work once we start accepting input from stdin. It won't be used
with the lightweight lei(1) client, however.
For WWW (eventually), a single FD may be enough.
|
|
For another step in in syscall reduction, we'll support
transferring 3 FDs and a buffer with a single sendmsg/recvmsg
syscall using Socket::MsgHdr if available.
Beyond script/lei itself, this will be used for internal IPC
between search backends (perhaps with SOCK_SEQPACKET). There's
a chance this could make it to the public-facing daemons, too.
This adds an optional dependency on the Socket::MsgHdr package,
available as libsocket-msghdr-perl on Debian-based distros
(but not CentOS 7.x and FreeBSD 11.x, at least).
Our Inline::C version in PublicInbox::Spawn remains the last
choice for script/lei due to the high startup time, and
IO::FDPass remains supported for non-Debian distros.
Since the socket name prefix changes from 3 to 4, we'll also
take this opportunity to make the argv+env buffer transfer less
error-prone by relying on argc instead of designated delimiters.
|
|
While our recv_3fds() implementation is more efficient
syscall-wise, loading Inline takes nearly 50ms on my machine
even after Inline::C memoizes the build. The current ~20ms in
the fast path is barely acceptable to me, and 50ms would be
unusable.
Eventually, script/lei may invoke tcc(1) or cc(1) directly in
the fast path, but it needs @INC for the slow path, at least.
We'll encode the number of FDs into the socket name allow
parallel installations, for now.
|
|
We'll always be transferring stdin, stdout, and stderr together
for lei. Perhaps I lack imagination or foresight, but I can't
think of a reason to send more or less FDs.
|
|
IO::FDPass may be an extra installation burden I don't want to
impose on users. We only support Linux and *BSDs, however.
|
|
To get rid of the ugly $PublicInbox::DS::in_loop localization
in MboxReader, we'll distinguish between ->CLOSE and ->DESTROY
with ProcessPipe.
If we end up closing via ->DESTROY, we'll assume the caller will
want to deal with $? asynchronously via the event loop (or not
even care about $?).
If we hit ->CLOSE directly, we'll assume the caller called
close() and wants to check $? synchronously.
Note: wantarray doesn't seem to propagate into tied methods,
otherwise I'd be relying on that.
|
|
Using "make update-copyrights" after setting GNULIB_PATH in my
config.mak
|
|
It seems like a more logical place for it, but we'll favor the
newly-added xsys_e() in tests for BAIL_OUT use.
|
|
fileno(DIRHANDLE) only works on Perl 5.22+, so we need to use
dirfd(3) ourselves from Inline::C (or rely on chattr(1) being
installed).
While we're at it, rename `set_nodatacow' to `nodatacow_fd'
for consistency with `nodatacow_dir'.
|
|
v?fork failures seems to be the cause of locks not getting
released in -watch. Ensure lock release doesn't get skipped
in ->done for both v1 and v2 inboxes. We also need to do
everything we can to ensure DB handles, pipes and processes
get released even in the face of failure.
While we're at it, make failures around `git update-server-info'
non-fatal, since smart HTTP seems more popular anyways.
v2 changes:
- spawn: show failing command
- ensure waitpid is synchronous for inotify events
- teardown all fast-import processes on exception,
not just the failing one
- beef up lock_release error handling
- release lock on fast-import spawn failure
|
|
SQLite and Xapian files are written randomly, thus they become
fragmented under btrfs with copy-on-write. This leads to
noticeable performance problems (and probably ENOSPC) as these
files get big.
lore/git (v2, <1GB) indexes around 20% faster with this on an
ancient SSD. lore/lkml seems to be taking forever and I'll
probably cancel it to save wear on my SSD.
Unfortunately, disabling CoW also means disabling checksumming
(and compression), so we'll be careful to only set the No_COW
attribute on regeneratable data. We want to keep CoW (and
checksums+compression) on git storage because current ref
storage is neither checksummed nor compressed, and git streams
pack output.
|
|
We no longer use writev(2) in pi_fork_exec to emit errors.
|
|
parent.pm is smaller than base.pm, and we'll also move
towards relying on `-w' (or not) to toggle process-wide
warnings during development.
|
|
Making the RLIMITS list a function doesn't allow constant
folding, so just make it an array accessible to other modules.
|
|
Subprocess we spawn may want to use SIGCHLD for themselves.
This also ensures we restore default signal handlers
in the pure Perl version.
|
|
Older versions of Inline (e.g. 0.53 in CentOS 7) did not accept
the `directory' parameter, so use conditional assignment to set
a default value on $ENV{PERL_INLINE_DIRECTORY}, instead.
|
|
Despite several memory reductions and pure Perl performance
improvements, Inline::C spawn() still gives us a noticeable
performance boost.
More user-oriented command-line programs are likely coming,
setting PERL_INLINE_DIRECTORY is annoying to users, and so is
is poor performance. So allow users to opt-in to using our
Inline::C code once by creating a `~/.cache/public-inbox/inline-c'
directory.
XDG_CACHE_HOME is respected to override the location of ~/.cache
independent of HOME, according to
https://specifications.freedesktop.org/basedir-spec/0.6/ar01s03.html
v2: use "/nonexistent" if HOME is undefined, since that's
the home of the "nobody" user on both FreeBSD and Debian.
|
|
Both the C and pure Perl implementions of `pi_fork_exec'
returns `-1' on error, not `undef'.
|
|
I didn't wait until September to do it, this year!
|
|
Commit 9f5a583694396f84 ("spawn (and thus popen_rd) die on failure")
was incomplete in that it only removed error checking for spawn
failures for non-(vfork|fork) calls, but the actual (vfork|fork)
PID result could still be undef.
Fixes: 9f5a583694396f84 ("spawn (and thus popen_rd) die on failure")
|
|
Most spawn and popen_rd callers die on failure to spawn,
anyways, and some are missing checks entirely. This saves
us a bunch of verbose error-checking code in callers.
This also makes popen_rd more consistent, since it already
dies on pipe creation failures.
|
|
There's a bunch of leftover "require" and "use" statements we no
longer need and can get rid of, along with some excessive
imports via "use".
IO::Handle usage isn't always obvious, so add comments
describing why a package loads it. Along the same lines,
document the tmpdir support as the reason we depend on
File::Temp 0.19, even though every Perl 5.10.1+ user has it.
While we're at it, favor "use" over "require", since it it gives
us extra compile-time checking.
|
|
Since vfork always shares memory between the child and parent,
we can propagate errors to the parent errno using shared memory
instead of just dumping to stderr and hoping somebody sees it.
|
|
This simplifies our admin module a bit and allows solver to be
used with v1 inboxes using git versions prior to v1.8.5 (but
still >= git v1.8.0).
|
|
We can save callers the trouble of {-hold} and {-dev_null}
refs as well as the trouble of calling fileno().
|
|
We can use "use" to get the namespace into the "BEGIN" phase of
the interpreter. While we're at it, use \&coderef syntax
explicitly instead of globbing everything.
|
|
It's unnecessary code which I'm not sure we ever used. In
retrospect, completely clearing the environment doesn't make
sense for the processes we spawn. We don't need to clobber
individual environment variables in our code, either
(and if we did for tests, we can use 'local').
|
|
This makes the subroutine behave more like which(1) command
and will make using spawn() in tests easier.
|
|
|
|
Instead, the O_NONBLOCK flag is set by PublicInbox::HTTPD::Async;
and we won't be setting it elsewhere.
|
|
Noticed while testing on FreeBSD 11.2 amd64 with the optional
Inline::C extension using clang 6.0.0. The end result on
FreeBSD was spawning processes failed badly and things were
immediately unusable with this enabled.
av_len is a misleading API, and I failed to read the API
comments in perl:/av.c which state:
> Note that, unlike what the name implies, it returns
> the highest index in the array, so to get the size of
> the array you need to use "av_len(av) + 1".
> This is unlike "sv_len", which returns what you would expect.
If this bug affected anybody, it would've only affected users
using both the optional Inline::C module AND set the
PERL_INLINE_DIRECTORY environment variable.
That said, I've never seen any evidence of it on Debian
GNU/Linux + gcc on any x86 variant. That includes full 64-bit
systems, a full 32-bit system, a 64-bit system with 32-bit
userspace, across multiple gcc versions since 2016.
|
|
Our high-level config already treats single limits as a
soft==hard limit for limiters; so stop handling that redundant
in the low-level spawn() sub.
|
|
This allows users to configure RLIMIT_{CORE,CPU,DATA} using
our "limiter" config directive when spawning external processes.
|
|
cgit (and most other CGI executables) is not typically installed
for use via $PATH, so we'll need to support absolute paths to
run it.
|
|
We'll be spawning cgit and git-diff, which can take gigantic
amounts of CPU time and/or heap given the right (ermm... wrong)
input. Limit the damage that large/expensive diffs can cause.
|
|
Using update-copyrights from gnulib
While we're at it, use the SPDX identifier for AGPL-3.0+ to
ease mechanical processing.
|
|
fork failures are unfortunately common when Xapian has
gigabytes and gigabytes mmapped.
|