Eric Wong [Fri, 11 Apr 2025 18:35:51 +0000 (18:35 +0000)]
linkify: give <a> tags to gemini:// URLs
While I still think serving the Gemini text and protocol is
unnecessary bloat on our part, there are web browsers with
gemini:// support, so we might as well support it.
Eric Wong [Fri, 11 Apr 2025 01:38:34 +0000 (01:38 +0000)]
http: send `100 Continue' responses for uploads
curl(1) defaults to sending a `Expect: 100-continue' header and
waits 1 second for a `100' response when using the `-T' switch
to upload via PUT. We'll unconditionally respond with a `100'
response to save curl users 1 second. Expecting curl users to
to disable the Expect: header via `-HExpect:' on every invocation
seems unreasonable, so our behavior should be tailored to curl
default behavior as much as possible.
We can't delegate the decision to send 100 or not to the
underlying Plack app (e.g. PublicInbox::WWW) since PSGI specs
don't provide a convenient callback interface for dealing with
trickled env->{'psgi.input'} reads. Of course, having to
synchronously wait on env->{'psgi.input'}->read would be
unacceptable since that would prevent the server process from
servicing other concurrent clients.
Eric Wong [Tue, 8 Apr 2025 20:49:58 +0000 (20:49 +0000)]
daemon: support backlog= listen parameter
For -netd instances using multiple listeners, it's easiest for me
to configure all ListenStream directives in a single
systemd.socket(5) unit. Unfortunately, this seems to force all
listeners to use the same Backlog= value, and juggling
multiple systemd.socket(5) units for a single systemd service
seems tedious.
So support setting the backlog= parameter on the public-inbox-*d
command-line on a per-listener basis; allowing us to override
the backlog value of inherited listeners and also to set the
backlog for listeners we bind ourselves. This makes it possible
for non-systemd users to easily configure per-listener backlogs,
as well.
Eric Wong [Tue, 8 Apr 2025 20:49:57 +0000 (20:49 +0000)]
daemon: use unpack_sockaddr_* for clarity
unpack_sockaddr_un and unpack_sockaddr_in have been around since
Perl 5.002 in the mid-1990s, so be explicit and use them instead
of relying on wantarray caller context. We'll also omit the
$host check for ($port == 0) since zero is not a valid TCP port.
Eric Wong [Sat, 5 Apr 2025 17:44:13 +0000 (17:44 +0000)]
http: refuse to deal with >4GB chunks in uploads
The `hex' perlop will return an NV (typically 64-bit double) on
UV (unsigned int) overflow and warns on larger values. While
64-bit integer builds of 32-bit perl (e.g. Debian i386) can
handle 64-bit numbers, there are builds of perl which still use
32-bit integers nowadays (e.g. OpenBSD 7.x i386).
It's unlikely we'll ever see chunks even close to 4GB, so just
cap it at 8 hex characters and drop clients which send larger
amounts.
Eric Wong [Thu, 3 Apr 2025 08:46:19 +0000 (08:46 +0000)]
t/httpd-https: test SSL session reuse
Reusing SSL sessions is one avenue to improve performance on
high-latency networks. For now, we can support the built-in
session cache of OpenSSL. For multi-process setups, using
using Cache::FastMmap will likely be the way to go...
Eric Wong [Thu, 3 Apr 2025 08:46:17 +0000 (08:46 +0000)]
ds: close: always call SSL_close on TLS sockets
Instead of relying on the explicit ->shutdn API, just
use the overloaded ->close so $env->{'psgix.io'} callers
can use it directly. This makes it easier to ensure
proper shutdown so SSL session reuse can be a possiblity.
Eric Wong [Wed, 2 Apr 2025 19:00:15 +0000 (19:00 +0000)]
http: support trailers for chunked requests
While HTTP/1.1 trailers are not widely supported by other
software at the moment, they may be someday. Trailers are
useful for things such as calculating checksums and signatures
on streaming uploads of unseekable data. So add support for
them in case they're needed someday, perhaps for allow
uploading mail or ActivityPub messages over HTTP.
Eric Wong [Wed, 2 Apr 2025 19:00:14 +0000 (19:00 +0000)]
http: fix and test Trailer: rejection
We need to check for the existence of Trailers after successful
parsing. I actually intend to support HTTP trailers, and I
noticed this while working on adding support for them.
Eric Wong [Wed, 2 Apr 2025 19:00:13 +0000 (19:00 +0000)]
t/httpd-corner: modernize test
We'll rely on autodie more. We'll also declare internal
subroutines as `my' scalars to ensure they can't be accessed by
other tests when we reuse test processes in `check-run'. We'll
also set avoid setting TCP_NODELAY for cases where it's not
necessary to save a handfull of cycles.
More corner case tests will be added soon to ensure we can
handle HTTPS termination for Varnish (or other caches).
Eric Wong [Fri, 28 Mar 2025 10:34:10 +0000 (10:34 +0000)]
daemon: allow per-listener servername= and serverport=
Being able to specify per-listener servername= with the
`--listen' CLI arg is helpful for running an Tor .onion NNTP
server with the same config file (and thus
publicinbox.nntpserver) as a public-facing NNTP endpoint.
For HTTP, servername= and serverport= allow overriding the
SERVER_NAME and SERVER_PORT PSGI environment variables,
respectively. These are useful for generating full URLs
for clients which don't send the HTTP `Host:' header.
These currently have no effect aside from wasting memory for
POP3 and IMAP listeners. serverport= isn't used by NNTP,
either.
Eric Wong [Fri, 28 Mar 2025 10:34:09 +0000 (10:34 +0000)]
www: omit SERVER_PORT for standard port redirects
For HTTP requests without a `Host:' header, we can omit
the server port component of the URL if using port 443
for `https' or 80 for plain `http'. We'll also consistently
use SCRIPT_NAME in all of our tests since Plack requires
it.
Eric Wong [Thu, 27 Mar 2025 23:20:46 +0000 (23:20 +0000)]
nntp: avoid repeated rand() calls
We only need to generate the secret salt once, so initialize
it early to avoid potentially expensive `rand' ops in repeated
calls to wildmat2re. We'll also stringify it early to hopefully
improve CoW sharing and reduce fragmentation.
Eric Wong [Thu, 27 Mar 2025 23:20:45 +0000 (23:20 +0000)]
nntp: avoid modifying $_[0] in RE conversions
I've been getting occasional t/nntp.t warnings about
uninitialized variables which I'm not able to reproduce
reliably. My current theory is that modifying $_[0] may get
wonky when it's happening across several layers of the call
stack, so stop doing it since it's unlikely to yield any
real world speedups and only made the code more difficult
to understand, here.
Eric Wong [Thu, 27 Mar 2025 23:20:44 +0000 (23:20 +0000)]
nntp: avoid uninitialized vars from bogus LIST args
We can only use the `list_' prefix in NNTP.pm for subcommands
which we intend to dispatch from client requests. So prefix
the list_*_i subs with a `_' to prevent `args_ok' from firing
on a non-prototyped subroutine.
I noticed this problem in the previous change to reduce code
duplication between LIST subcommands.
Eric Wong [Thu, 27 Mar 2025 23:20:42 +0000 (23:20 +0000)]
tests: get rid of most Data::Dumper usage
`explain' from Test::More already uses Data::Dumper, so
it's unnecessary to set it up and call it ourselves.
This saves a bunch of wonky code in t/nntp.t, as well.
Eric Wong [Wed, 26 Mar 2025 20:13:03 +0000 (20:13 +0000)]
content_digest_dbg: more compact display w/ git_quote
Make WWW /$MSGID/d/ endpoints and `lei mail-diff' output
less verbose by getting rid of Data::Dumper in favor of
a lightly-modified git_quote output. The extra [] pairs
from Data::Dumper were unnecessary noise and the quoting
used by git_quote should be more familiar to git users.
Unfortunately, git's escaping of NUL bytes to `\000' is
a bit noisy, so we'll undo that by substituting to `\0'.
Eric Wong [Wed, 26 Mar 2025 03:35:02 +0000 (03:35 +0000)]
search: improve xap_helper mset error reporting
For user errors generating bad queries, dumping exception
messages from Xapian::QueryParser can be helpful. So
unconditionally pass the stderr FD as a regular file to
xap_helper processes for both lei and WWW search users.
Eric Wong [Wed, 26 Mar 2025 03:35:01 +0000 (03:35 +0000)]
treewide: use git_quote for diagnostic dumps
git_quote output is probably less alien to most users than
the Perl-specific Data::Dumper, so use it for any messages
which hit stdout/stderr or syslog.
Eric Wong [Wed, 26 Mar 2025 03:34:59 +0000 (03:34 +0000)]
solver: use git_quote on filenames for diagnostics
git_quote output makes more sense than relying on \Q..\E from
Perl or even Data::Dumper since these filenames are from git
output and more users are likely familiar with git-style quoting
than anything Perl-specific.
Eric Wong [Fri, 21 Mar 2025 22:22:30 +0000 (22:22 +0000)]
limiter: refactor to reduce code duplication
PlackLimiter, Qspawn, and ViewVCS all have roughly the same
code around our base Limiter package, so put everything around
a new Limiter->may_start subroutine. PlackLimiter loses some
stats as a result but that's logged anyways and I doubt the
customizable error message was worth the effort.
We now have ckhup and 499 (client disconnect) handling for all
PSGI uses of Limiter, as well.
t/qspawn.t changes were required since the original ->finalize
logic now relies on on_destroy; but none of the existing PSGI
code using Qspawn required changes.
Eric Wong [Thu, 20 Mar 2025 00:05:35 +0000 (00:05 +0000)]
plack_limiter: PSGI middleware to limit concurrency
While processing several concurrent requests within the same
worker process is helpful to exploit parallelism in git blob
lookups and smooth out delays; excessive parallelism is harmful
since it allows too much memory to be allocated at once for zlib
buffers and such.
While PublicInbox::WWW already uses the limiter for certain
expensive endpoints (e.g. /s/ and anything using Qspawn); some
long-running endpoints with many inexpensive steps (e.g. /T/,
/t/, /d/, *.atom, *.mbox.gz, etc.) can end up using a large
amount of memory for gzip buffers despite being fair to other
responses and being able to stream >500 messages/sec on 2010-era
hardware.
So give sysadmins an option to balance between smoothing out
delays in blob retrieval and memory usage required to compress
and spew out chunks of potentially large multi-email responses.
Eric Wong [Tue, 18 Mar 2025 08:30:27 +0000 (08:30 +0000)]
lg2: disable strict hash verification
Unlike git(1), libgit2 verifies the SHA-(1|256) of objects it
reads by default. This verification results in a large (nearly
100% w/ SHA1DC) performance penalty for us. Since our libgit2
code only reads (and never writes objects), just follow git(1)
and skip verification for normal reads.
This brings our libgit2-based Gcf2 batch loop performance closer
to that of the `git cat-file --batch-command' as shown in the
new xt/lg2_cmp.t developer test. However, Gcf2Client still uses
a more verbose (but more flexible) input format and the Perl
gcf2_loop still incurs normal Perl method dispatch overheads.
Eric Wong [Fri, 14 Mar 2025 09:22:04 +0000 (09:22 +0000)]
daemon: define %TLS_ONLY hash
Defining TLS-only protocols only once makes it easier to support
new protocols in the future since we can rely on only updating
this new hash instead of having to update regexps in other
places.
Eric Wong [Thu, 13 Mar 2025 20:35:04 +0000 (20:35 +0000)]
www: disable legacy encoded Message-IDs for non-v1
For a while in the very early days of the v1 format, we
supported SHA-1 checksums of the Message-ID in the URL. This
never affected v2 inboxes nor extindex, and those URLs would've
been broken if run through public-inbox-convert, anyways.
Eric Wong [Thu, 6 Mar 2025 20:34:42 +0000 (20:34 +0000)]
XapHelper.pm: fix QP_FLAGS initialization
We can't use $PublicInbox::Search::QP_FLAGS until after
load_xapian is called. Failure to set the correct query parser
flags was causing failures in
`PI_NO_CXX=1 TEST_DAEMON_XH=-X0 prove -bw t/imapd.t'
since phrase parsing was broken with the Perl bindings
XapHelper implementation.
Now, the `check-xh0' and `check-xh1' targets both pass with
PI_NO_CXX=1 set to disable the C++ xap_helper.
Fixes: fa6a7919 (xap_helper: enable FLAG_PURE_NOT in external process, 2025-02-23)
Eric Wong [Wed, 5 Mar 2025 23:26:47 +0000 (23:26 +0000)]
search: unoverload {relevance} in options
Take the lead from xap_helper and split out handling of {asc}
and {sort_col} fields while keeping the {relevance} field as a
boolean. As with xap_helper, {sort_col} < 0 means
Xapian::BoolWeight will be used for docid ordering.
Eric Wong [Wed, 5 Mar 2025 23:26:46 +0000 (23:26 +0000)]
search: use common do_enquire for MiscSearch
We can consistently use a {sort_col} search option as in
xap_helper to allow do_enquire to be more generic across
searches. This lets us avoid the need for a specialized
implementation in MiscSearch.
Eric Wong [Wed, 5 Mar 2025 07:18:36 +0000 (07:18 +0000)]
lei: use C++ xap_helper if available
The C++ version of xap_helper allows `thread:' subqueries,
non-fragile parsing of git approxidate with `d:', `dt:' and
`rt:' prefixes, and possibly more features in the future not
available in the SWIG or XS Xapian bindings.
There's no point in lei supporting the Perl version of XapHelper
since lei isn't expected to deal with abusive clients making
expensive queries, so lei will only use the C++ version.
We spawn xap_helper in every lei worker process to guarantee
resource availability without having to resort to preforking.
While pre-forking and on-demand thread||process creation was
considered, I decided having a lingering xap_helper process
probably wasn't worth the startup time improvement and instead
prefer to minimize idle memory use.
The whole implementation is a bit strange since lei support was
an after-thought for xap_helper and we wrap the async_mset API
to make it synchronous once again to reduce the code impact for
code shared with public-facing daemons.
Finally, test_lei in TestCommon gets tweaked to avoid repeatedly
rebuilding in a new directory with every test_lei use in our test
suite.
Eric Wong [Wed, 5 Mar 2025 07:18:35 +0000 (07:18 +0000)]
xap_helper.h: share ThreadFieldProcessor across DBs
Unlike notmuch, we don't actually use the QueryParser from
initialization and can stop storing it in the class. Instead,
we rely on the global `cur_srch' at runtime for both which has
both the Xapian::Databaase and Xapian::QueryParser objects. So
stop carrying a copy of the ThreadFieldProcessor object for
every DB connection we have since some public non-extindex-users
may have many DBs and we can probably save some memory this way.
Eric Wong [Wed, 5 Mar 2025 07:18:34 +0000 (07:18 +0000)]
search: avoid `git rev-parse' if using C++ xap_helper
The C++ xap_helper can work with the Xapian query parser API to
run run git_date_parse(), so there's no need to do janky string
substitutions as we do for the Perl bindings.
Eric Wong [Wed, 5 Mar 2025 07:18:33 +0000 (07:18 +0000)]
xap_helper: use libgit2 git_date_parse in C++ impl
Using proper Xapian RangeProcessor and FieldProcessor subclasses
via the Xapian API means our `d:', `dt:' and `rt:' prefixes can
use git `approxidate' interpretation rules inside quoted
subqueries with the `thread:' prefix.
The only incompatibility is the `rt:' field which no longer
takes seconds with <5 characters (e.g. `rt:0..' as was in
t/xap_helper.t), but that'd be broken anyways in real-world use
which is run through `git rev-parse --since='.
Eric Wong [Wed, 5 Mar 2025 07:18:32 +0000 (07:18 +0000)]
search: make `d:' search prefix consistently date-only
While `d:' previously treated YYYYMMDD and YYYY-MM-DD as
low-precision by assuming 00:00:00 for the HH:MM:SS portion, it
would not do so for dates passed to git-rev-parse (e.g.
"last.year") since the HH:MM:SS of the current time would be
used by git. So stop remapping `d:' in queries to `dt:' and
continue using `d:' in the YYYYMMDD column.
This does break things like `d:2.hours.ago..' by causing too
many (or too few) results to be returned due to lack of
precision, but I expect small time ranges of less than one
day to be of limited use.
`dt:' remains a higher-precision field for searching on both
date and time from the Date: header, while `d:' is now always
date-only.
Eric Wong [Wed, 5 Mar 2025 00:45:36 +0000 (00:45 +0000)]
listener: don't set listen backlog on inherited sockets
By using the listen(2) backlog as-is when inheriting (from
systemd or similar), we can give the sysadmin more control on
controlling overload on a per-listener basis. For systemd
users, this means the `Backlog=' parameter in systemd.socket(5)
can be respected and configured to give certain sockets a
smaller backlog (perhaps combined with with per-listener
`multi-accept' parameter on sockets with the standard (huge)
backlog).
For sockets we create, continue to use INT_MAX and let the
kernel clamp it to whatever system-wide limit there is
(e.g. `net.core.somaxconn' sysctl on Linux).
Eric Wong [Thu, 27 Feb 2025 00:08:18 +0000 (00:08 +0000)]
imap: support external xap_helper process
Moving Xapian requests to an external process prevents expensive
searches from affecting non-search IMAP commands or any other
non-search requests within a -netd process if IMAP is sharing a
process with other public-facing protocols.
Eric Wong [Thu, 27 Feb 2025 00:08:17 +0000 (00:08 +0000)]
daemon: slightly simplify xap_helper spawning
Avoid typing the fully-qualified global $XHC variable twice in
spawn_xh. We'll also delay loading of XhcMset until successful
spawn and improve the error message in case of error. There's
also no need to conditionally local-ize
$PublicInbox::Search::XHC, so do it unconditionally to reduce
branches.
Eric Wong [Thu, 27 Feb 2025 00:08:16 +0000 (00:08 +0000)]
search: remove xhc_start_maybe
We start xap_helper processes early in both public-facing
daemons to ensure maximum sharing. lei may be different
since lingering processes likely sit idle for long periods
and we don't want to tie up RAM or swap space for idle
processes.
Eric Wong [Sun, 23 Feb 2025 13:13:35 +0000 (13:13 +0000)]
xh_thread_fp: optimize OR query generation
For Xapian >= 1.4.10 users, using `|=' will reduce allocations.
`|=' still works for Xapian 1.3.2 .. 1.4.9 users, and I'm not
sure if Xapian 1.2.x is relevant anymore, especially for our
new C++-only features.
While operator overloading is often confusing and frustrating to
me when reading someone else's code, the optimization seems worth
it since (AFAIK) there's no other way to get the allocation
reduction.
cf. Olly in xapian-discuss <20250222043050.GA17282@survex.com>
Eric Wong [Sun, 23 Feb 2025 13:13:34 +0000 (13:13 +0000)]
xap_helper: avoid temporary std::set in thread fp
Instead of creating a temporary std::set to store THREADIDs,
just inject the Xapian::Query::OP_OR operations directly into
the query we return.
Unlike notmuch(1), we can do this without risking redundant
query ops from repeated THREADIDs since use since we use a
column instead of a term for THREADID. Column use allowed
us to use .set_collapse_key to deduplicate the intermediate
mset, already.
Eric Wong [Sun, 23 Feb 2025 13:13:33 +0000 (13:13 +0000)]
xap_helper: enable FLAG_PURE_NOT in external process
Since public-facing WWW uses an external process with
a non-blocking socket, clients will only see 503 errors
if the search is overloaded. While allowing pure NOT
queries is expensive, there are already many possible
ways of triggering expensive queries so it's probably not
a problem to allow one more.
Eric Wong [Thu, 20 Feb 2025 22:14:30 +0000 (22:14 +0000)]
search: index References: for thread:GHOST-MSGID
To search for messages in a thread with a ghost msgid,
we need to be able to search against msgids in the
References: header since (by definition) ghosts don't
show up as any Message-ID: we've indexed.
This should make our implementation of `thread:MSGID'
queries equivalent in capability to `thread:THREADID'
of notmuch.
Eric Wong [Thu, 20 Feb 2025 22:14:29 +0000 (22:14 +0000)]
xap_helper: support thread:{SUBQUERY} via C++
Stealing the idea and code from notmuch to perform subqueries
within search. One major internal difference from notmuch is we
store THREADID as numeric column in Xapian whereas notmuch
stores a boolean term. The use of a column lets us use
set_collapse_key to deduplicate results within Xapian itself.
The other difference from notmuch is we avoid exposing the
numeric THREADID since they're unstable and not reproducible in
mirrors, thus we also support `thread:MSGID' instead of
`thread:THREADID' in brace-less queries.
Eric Wong [Thu, 20 Feb 2025 22:14:28 +0000 (22:14 +0000)]
xap_helper: switch C++ implementation to AGPL-3
GPL-2 approxidate code won't work with the XS/SWIG version, so
it looks like we'll keep calling `git rev-parse' in both
versions for the time being. Meanwhile, it's more valuable to
be able to take GPL-3+ code from notmuch for thread:{} query
parsing.
Eric Wong [Wed, 19 Feb 2025 10:10:32 +0000 (10:10 +0000)]
searchidx: don't index Base-85 w/ CRLF endings
I encountered a false positive search result from a CRLF message
with a Base-85 patch in it. It turns out our Base-85 filtering
code didn't account for the possibility of "\r" showing up in
patch messages, so just ignore all trailing spaces (not just
horizontal spaces) in index_diff().
While we're at it, exclude horizontal whitespace and CR
consistently from Base-85-looking quoted text in
index_body_text(), too, since I'm sure there's messages with
CRCRLF in the wild, too...
Eric Wong [Wed, 19 Feb 2025 12:39:26 +0000 (12:39 +0000)]
cfgwr: fix non-libgit2 case
libgit2 isn't installed on all my test machines, and the
addition of libgit2 support was careless in that it failed to
test the non-libgit2 case thoroughly :x
Fixes: a8073f6c (lg2: use cfgwr_commit to write to configs using libgit2, 2025-02-15)
Eric Wong [Sat, 15 Feb 2025 11:10:11 +0000 (11:10 +0000)]
lg2: use cfgwr_commit to write to configs using libgit2
libgit2 allows us to avoid process spawning overhead when
writing to the config file. While the performance difference is
unlikely to matter in real-world use, it should improve
maintainability of tests by allowing us to write tests with
`public-inbox-init' with a smaller performance penalty (as
opposed to using `PublicInbox::IO::write_file' to setup
configs).
Eric Wong [Sat, 15 Feb 2025 11:10:10 +0000 (11:10 +0000)]
rename Gcf2 -> Lg2 for general libgit2 use
We'll be exploring libgit2 for writing (and possibly reading) config
files, so the `gcf2' (git-cat-file-2) name isn't appropriate
anymore for the package name. We'll still keep `gcf2_' for
subroutine names related to git-cat-file-like behavior.
For testing with TEST_RUN_MODE=0 (which spawns new processes
instead of reusing them), we can't rely on packages being loaded
by subprocesses influencing the main Perl process. Thus we must
load PublicInbox::Config explicitly to prevent failures with
this run mode.
Eric Wong [Sat, 15 Feb 2025 11:10:06 +0000 (11:10 +0000)]
import: clobber $? if global init.defaultBranch fails
We don't want $? influencing further error checking downstream,
such as thw _wq_done_wait call in public-inbox-clone when
git-config(1) invocations which set $?=0 are replaced with
in-process libgit2 config function calls.
Eric Wong [Sat, 15 Feb 2025 11:10:04 +0000 (11:10 +0000)]
favor run_wait() over CORE::system()
run_wait() can use vfork(2) which saves memory in large
processes, such as public-inbox-extindex when dealing with giant
messages. While vfork is unlikely to matter for real-world uses
of public-inbox-edit, PublicInbox::Spawn is a sunk cost treewide
our `make check-run' test target avoids spawning new Perl
processes for most things in script/*, so there can be a small
savings for testing.
Eric Wong [Fri, 14 Feb 2025 09:38:55 +0000 (09:38 +0000)]
imap_searchqp: avoid occasional P::RD hint spew
It turns out Inline::C::Parser::RecDescent increments $::RD_HINT
when Inline::C rebuilds are necessary, thus causing verbose log
spew on bogus IMAP search queries in our own use of
Parse::RecDescent. So, local-ize $::RD_HINT as we do with other
$::RD_* vars to avoid Inline::C influencing our IMAP query
parsing.
Triggering the increment of $::RD_HINT via
Inline::C::Parser::RecDescent may be achieved by forcing an
Inline::C rebuild, such as clearing the contents of
~/.cache/public-inbox/inline-c/* or
$ENV{PERL_INLINE_DIRECTORY}/* (but leaving one of those
top-level directories to trigger Inline::C rebuilds. This bug
mainly manifested in t/imap_searchqp.t under the `make
check-run' test target which reuses existing processes to speed
up tests, though it could manifest in real-world -imapd (and
-netd) usage if the process spawning the daemon built Inline::C
code.
Eric Wong [Tue, 11 Feb 2025 03:55:36 +0000 (03:55 +0000)]
daemon: make xap_helper socket non-blocking
A non-blocking socket for public-facing searches is ideal on
overloaded servers which receive more searches than it can
handle in a short time span. A non-blocking socket here
prevents a sendmsg(2) call from blocking the main event loop
and ensures non-search requests can still be processed even
if all xap_helper subprocesses are overwhelmed.
$search_xh_pid is hoisted out to TestCommon for use with the new
slow-xh-search test and renamed $find_xh_pid to avoid confusion
with Xapian search functionality. We'll also drop a redundant
call to find the xap_helper PID in psgi_v2.t while transitioning
to $find_xh_pid.
Eric Wong [Tue, 11 Feb 2025 03:55:34 +0000 (03:55 +0000)]
search: async_mset: always run callback on exceptions
To produce consistent error behavior, ensure we always run
the user-supplied callback on exceptions when using in-process
Xapian (as opposed to the external xap_helper process).
Eric Wong [Mon, 10 Feb 2025 21:09:34 +0000 (21:09 +0000)]
git_http_backend: input_prepare: die on unrecoverable errors
We shouldn't be guarding against errors at this level to return
500 errors, but rather we can rely on existing eval wrappers in
the stack (e.g. Plack::Util::run_app and Qspawn->parse_hdr_done)
to capture errors and return the HTTP 500 status.
Eric Wong [Mon, 10 Feb 2025 21:09:32 +0000 (21:09 +0000)]
msgmap: use v5.12
No unicode_string dependencies, here, so we can use v5.12 and
perhaps to speed up loading one day if everything drops `strict'
in favor of v5.12 (or higher).
Eric Wong [Mon, 10 Feb 2025 21:09:30 +0000 (21:09 +0000)]
multi_git: remove redundant ops
read_all already returns an array in array context so the
`split' op is unnecessary, and we don't need to use
`join("", ...)' on arrays that are destined for `print'
or `say'.
Eric Wong [Mon, 10 Feb 2025 21:09:29 +0000 (21:09 +0000)]
devel/sysdefs-list: explain why we don't autodie, here
Some distros split out core modules like autodie, and
sysdefs-list has higher portability requirements than the rest
of our code since it's used for uncommon platforms.
Eric Wong [Sat, 8 Feb 2025 03:26:38 +0000 (03:26 +0000)]
qspawn: drop redundant check against limiter->{max}
We only need to guard against excessive parallelism on entry
in one place during Qspawn->start. Redundant guards are
confusing and make bugs harder-to-spot.
Eric Wong [Sat, 8 Feb 2025 03:26:37 +0000 (03:26 +0000)]
viewvcs: -codeblob limiter w/ depth for solver
By adding the `publicinboxlimiter.<name>.depth' parameter for
all limiters, we can have queue overflow detection and reject
requests with HTTP 503 error codes for all limiter uses, not
just the custom queue in ViewVCS as before. Then we can just
use a normal limiter in ViewVCS and get rid of the old custom
queue.
Eric Wong [Sat, 8 Feb 2025 03:26:36 +0000 (03:26 +0000)]
qspawn: use limiter->new default
No need to hardcode `32' here since PublicInbox::Limiter already
uses that value already (and includes a helpful comment about
using the same value as git-daemon).
Eric Wong [Sat, 8 Feb 2025 03:26:35 +0000 (03:26 +0000)]
daemon: check connections before solving codeblobs
The /$MSGID/s/ codeblob reconstruction endpoint is extremely
expensive and one of the few which do a large amount of upfront
processing before being able to write to an HTTP client and
detect if they're still alive (unlike other expensive
endpoints).
For TCP connections, use the TCP_INFO via getsockopt(2) if
supported or attempt to use poll(2) + the FIONREAD ioctl(2) to
check the connection before possibly invoking
git-(apply|ls-files|update-index) dozens or even hundreds of
times to reconstruct a blob.
Eric Wong [Sat, 8 Feb 2025 03:26:34 +0000 (03:26 +0000)]
git_http_backend: fix 32 default connection limit
Using any configurable limiter actually defaults to `1', which
isn't intended for `-httpbackend'. So add a $max_default
parameter to ensure we can set a per-limiter default value for
default limiters such as `-httpbackend'.
Fixes: 0a083241 (git_http_backend: use default limiter from Qspawn, 2025-01-28)
Eric Wong [Sat, 8 Feb 2025 03:26:33 +0000 (03:26 +0000)]
syscall: `use bytes' throughout
All *nix syscalls operate on the byte-level with no knowledge of
encodings, so ensure all `substr' and `length' calculations use
the bytes(3perl) pragma.
Eric Wong [Tue, 28 Jan 2025 08:31:12 +0000 (08:31 +0000)]
www: configurable "-httpbackend" named limiter
The default limiter being hard-coded to 32 is excessive for
smaller systems. With dozens or hundreds of inboxes,
configuring all inboxes to point to a user-defined limiter
is tedious, as well. So give WWW admins the opportunity
to configure the limiter for all git-http-backend(1) processes
with one `publicinbox.-httpbackend.max' knob.
Eric Wong [Tue, 28 Jan 2025 08:31:11 +0000 (08:31 +0000)]
limiter: ignore unparseable rlimit
Instead of blindly accepting whatever bogus values the config
gave us, ignore it and use the `W:' prefix to be consistent
with the rest of our codebase.
Eric Wong [Fri, 24 Jan 2025 09:18:56 +0000 (09:18 +0000)]
viewvcs: handle exceptions in on_destroy cb
We need to account for File::Temp->newdir or autodie::open
croaking on us in case the system is out of space or memory.
These resource errors are already handled in our normal code
path. However, this on_destroy codepath can be triggered
when the solver request needs to be queued due to business.
Thus we explicitly call the {-wcb} to ensure the socket
isn't forgotten.
Eric Wong [Sat, 18 Jan 2025 01:26:15 +0000 (01:26 +0000)]
mda: use read_all for error handling
We already imported PublicInbox::IO for this script and read_all
provides more safety in case default PerlIO semantics change for
read-in-full behavior.
Eric Wong [Sat, 18 Jan 2025 01:26:13 +0000 (01:26 +0000)]
config: config_fh_parse: hardcode FS/RS
We no longer need the flexibility to handle non-NUL-delimited
output from `git config -l', so simplify callers and allow
a theoretically sufficiently-advanced Perl implementation
optimize more easily.
Followup-to: 21146412 (config: drop scalar ref support from internal API, 2023-09-24)
Eric Wong [Sat, 18 Jan 2025 01:26:11 +0000 (01:26 +0000)]
treewide: use autodie for seek+sysseek
The underlying lseek(2) syscall won't fail due to HW errors,
only due to usage errors (ESPIPE, EINVAL); so don't waste
code on error checking ourselves and just let autodie check
things during development.