Eric Wong [Sun, 5 Jul 2020 23:27:41 +0000 (23:27 +0000)]
remove unused/redundant zlib-related imports
Z_FINISH is the default for Compress::Raw::Zlib::Deflate->flush,
anyways, so there's no reason to import it. And none of C::R::Z
is needed in WwwText now that gzf_maybe handles it all.
Eric Wong [Sun, 5 Jul 2020 23:27:40 +0000 (23:27 +0000)]
www: start making gzipfilter the parent response class
Virtually all of our responses are going to be gzipped, anyways.
This will allow us to utilize zlib as a buffering layer and
share common code for async blob retrieval responses.
To streamline this and allow GzipFilter to be a parent class,
we'll replace the NoopFilter with a similar CompressNoop class
which emulates the two Compress::Raw::Zlib::Deflate methods we
use.
This drops a bunch of redundant code and will hopefully make
upcoming WwwStream changes easier to reason about.
Eric Wong [Sun, 5 Jul 2020 23:27:39 +0000 (23:27 +0000)]
wwwstream: reduce blob fetch paths for ->getline
This will make it easier to support asynchronous blob
retrievals. The `$ctx->{nr}' counter is no longer implicitly
supplied since many users didn't care for it, so stack overhead
is slightly reduced.
Eric Wong [Sun, 5 Jul 2020 23:27:38 +0000 (23:27 +0000)]
wwwstream: reduce object graph depth
Like with WwwAtomStream and MboxGz, we can bless the existing
$ctx object directly to avoid allocating a new hashref. We'll
also switch from "->" to "::" to reduce stack utilization.
Eric Wong [Sun, 5 Jul 2020 23:27:37 +0000 (23:27 +0000)]
wwwatomstream: support async blob fetch
This allows -httpd to handle other requests while waiting
for git to retrieve and decode blobs. We'll also break
apart t/psgi_v2.t further to ensure tests run against
-httpd in addition to generic PSGI testing.
Using xt/httpd-async-stream.t to test against clones of meta@public-inbox.org
shows a 10-12% performance improvement with the following env:
TEST_JOBS=1000 TEST_CURL_OPT=--compressed TEST_ENDPOINT=new.atom
Eric Wong [Sun, 5 Jul 2020 23:27:32 +0000 (23:27 +0000)]
mbox: async blob fetch for "single message" raw mboxrd
This restores gzip-by-default behavior for /$INBOX/$MSGID/raw
endpoints for all indexed inboxes. Unindexed v1 inboxes will
remain uncompressed, for now.
Eric Wong [Sun, 5 Jul 2020 23:27:30 +0000 (23:27 +0000)]
mboxgz: do asynchronous git blob retrievals
This lets the -httpd worker process make better use of time
instead of waiting for git-cat-file to respond. With 4 jobs in
the new test case against a clone of
<https://public-inbox.org/meta/>, a speedup of 10-12% is shown.
Even a single job shows a 2-5% improvement on an SSD.
Eric Wong [Sun, 5 Jul 2020 23:27:29 +0000 (23:27 +0000)]
stop auto-loading Plack::Middleware::Deflater
Instead of gzipping some (mbox.gz, manifest.js.gz) responses and
leaving P::M::D to do the rest, we gzip everything ourselves,
now, so P::M::D is redundant.
Eric Wong [Sun, 5 Jul 2020 23:27:27 +0000 (23:27 +0000)]
wwwstatic: support gzipped directory listings
This will allow others to mimic our award-winning homepage
design without needing to rely on Plack::Middleware::Deflater
or varnish to compress responses.
Eric Wong [Sun, 5 Jul 2020 23:27:25 +0000 (23:27 +0000)]
{gzip,noop}filter: ->zmore returns undef, always
This simplifies callers, as witnessed by the change to
WwwListing. It adds overhead to NoopFilter, but NoopFilter
should see little use as nearly all HTTP clients request gzip.
Eric Wong [Sun, 5 Jul 2020 23:27:24 +0000 (23:27 +0000)]
gzipfilter: replace Compress::Raw::Deflate usages
The new ->zmore and ->zflush APIs make it possible to replace
existing verbose usages of Compress::Raw::Deflate and simplify
buffering logic for streaming large gzipped data.
One potentially user visible change is we now break the mbox.gz
response on zlib failures, instead of silently continuing onto
the next message. zlib only seems to fail on OOM, which should
be rare; so it's ideal we drop the connection anyways.
Eric Wong [Sun, 5 Jul 2020 23:27:18 +0000 (23:27 +0000)]
wwwstream: oneshot: perform gzip without middleware
Plack::Middleware::Deflater forces us to use a memory-intensive
closure. Instead, work towards building compressed strings in
memory to reduce the overhead of buffering large HTML output.
Eric Wong [Sun, 5 Jul 2020 23:27:17 +0000 (23:27 +0000)]
gzipfilter: minor cleanups
We currently don't use bytes::length in ->write, so there's no
need to `use bytes'. Favor `//=' to describe the intent of the
conditional assignment since the C::R::Z::Deflate object is
always truthy. Also use the local $gz variable to avoid
unnecessary {gz} hash lookups.
Eric Wong [Mon, 6 Jul 2020 06:11:02 +0000 (06:11 +0000)]
git_async_cat: unref pipes on EOF from git->cleanup
We avoided a managed circular reference in 10ee3548084c125f
but introduced a pipe FD leak, instead. So handle the EOF
we get when the "git cat-file --batch" process exits and
closes its stdout FD.
v2: remove ->close entirely. PublicInbox::Git->cleanup
handles all cleanup. This prevents us from inadvertantly
deleting the {async_cat} field associated with a different
pipe than the one GAC is monitoring.
Eric Wong [Sat, 4 Jul 2020 20:25:25 +0000 (20:25 +0000)]
t/import: test for nasty characters
Spammers may send emails with nasty characters which can throw
off git-fast-import. Users with non-existent or weaker spam
filters may be susceptible to corruption in the fast-import
stream as a result.
This was actually quietly fixed in git on 2020-06-01 by
commit 9ab886546cc89f37819e1ef09cb49fd9325b3a41
("smsg: introduce ->populate method"), but no test case
was created.
Eric Wong [Sat, 4 Jul 2020 21:33:13 +0000 (21:33 +0000)]
watch: don't burn CPU on IDLE failures
Network connections fail and need to be detected sooner rather
than later during IDLE to avoid backtrace floods. In case the
IDLE process dies completely, don't respawn right away, either,
to avoid entering a respawn loop.
Eric Wong [Wed, 1 Jul 2020 21:06:18 +0000 (21:06 +0000)]
tests: add use/require statements for TEST_RUN_MODE=0
The default (and fast) TEST_RUN_MODE=2 preloads most modules,
but TEST_RUN_MODE=0 is more realistic and can catch some
problems which may show up in real-world use.
Eric Wong [Wed, 1 Jul 2020 21:06:17 +0000 (21:06 +0000)]
watch: retry signals to kill IDLE and polling processes
To ensure reliable signal delivery in Perl, it seems we need to
repeatedly signal processes which aren't using signalfd (or
EVFILT_SIGNAL) with our event loop.
Eric Wong [Thu, 2 Jul 2020 03:32:56 +0000 (03:32 +0000)]
inboxidle: avoid per-inbox anonymous subs
Anonymous subs cost over 5K each on x86-64. So prefer the
less-recommended-but-still-documented way of using
Linux::Inotify2::watch to register watchers.
This also updates FakeInotify to detect modifications correctly
when used on systems with neither IO::KQueue nor
Linux::Inotify2.
Eric Wong [Mon, 29 Jun 2020 10:34:21 +0000 (10:34 +0000)]
watch: make waitpid() synchronous for Maildir scans
Maildir scanning still happens in the main process. Scanning
dozens of Maildirs is still time-consuming and monopolizes the
event loop during WatchMaildir::event_step. This can cause
cause zombies to accumulate before Sigfd::event_step triggers
DS::reap_pids.
Eric Wong [Mon, 29 Jun 2020 10:34:19 +0000 (10:34 +0000)]
watch: ensure SIGCHLD works in forked children
In case our git or spam checker subprocesses spawn
subprocesses of their own. We'll also ensure signal
handlers are properly setup before unblocking them.
Eric Wong [Mon, 29 Jun 2020 10:34:17 +0000 (10:34 +0000)]
watch: check for duplicates in ->over before spamcheck
It's cheaper to check for duplicates than run `spamc'
repeatedly when rechecking. We already do this for
v1 with by using the "ls" command with fast-import,
but v2 requires checking against over.sqlite3.
Eric Wong [Sat, 27 Jun 2020 10:04:00 +0000 (10:04 +0000)]
watch: simplify internal structures
We won't be attempting to reuse Mail::IMAPConnections used to
check authentication info, for now, so stop storing
$self->{mics}.
We can also combine $poll initialization for IMAP and NNTP
to avoid data structure duplication. Furthermore, rely on
autovivification to create {idle_pids} and {poll_pids}.
Eric Wong [Sat, 27 Jun 2020 10:03:59 +0000 (10:03 +0000)]
imaptracker: use flock(2) around writes
SQLite only issues non-blocking F_SETLK ops (not F_SETLKW) and
retries failures using a configurable busy_timeout. SQLite's
busy loop sleeps for a millisecond and retries the lock until
the configured busy_timeout is hit.
Trying to set ->sqlite_busy_timeout to larger values (e.g. 30000
milliseconds) still leads to failure when running the new stress
test with 8 processes with TMPDIR on a 7200 RPM HDD.
Inspection of SQLite source reveals there's no built-in way to
use F_SETLKW, so tack on the existing flock(2) support we use to
synchronize git + SQLite + Xapian for inbox writing. We use
flock(2) instead of POSIX fcntl(2) locks since Perl doesn't
provide a way to manipulate "struct flock" portably.
Eric Wong [Sat, 27 Jun 2020 10:03:58 +0000 (10:03 +0000)]
watch: support ~/.netrc via Net::Netrc
While git-credential-netrc exists in git.git contrib/, it may
not be widely known or installed. Net::Netrc is already a
standard part of most (if not all) Perl installations, so use it
directly if available.
Eric Wong [Sat, 27 Jun 2020 10:03:57 +0000 (10:03 +0000)]
watch: use our own "git credential" wrapper
Git.pm may not be installed on some systems; or some users have
multiple Perl installations and Git.pm is not available to the
Perl running -watch. Accomodate both those types of users by
providing our own "git credential" wrapper.
Eric Wong [Sat, 27 Jun 2020 10:03:56 +0000 (10:03 +0000)]
watch: enable autoflush for STDOUT and STDERR
In case output is redirected to a pipe, ensure stdout and stderr
are always unbuffered, as -watch may go long periods without
any output to fill up buffers.
Eric Wong [Sat, 27 Jun 2020 10:03:55 +0000 (10:03 +0000)]
watch: show user-specified URL consistently.
Since we use the non-ref scalar URL in many error messages,
favor keeping the unblessed URL in the long-lived process.
This avoids showing "snews://" to users who've specified
"nntps://" URLs, since "nntps" is IANA-registered nowadays and
what we show in our documentation, while "snews" was just a
draft the URI package picked up decades ago.
Existing use of the $ENV{TAIL} relied on parsing --std{out,err},
which was only usable for read-only daemons. However, -watch
doesn't use PublicInbox::Daemon code(*), so attempt to figure
out redirects.
(*) -watch won't able to run as a daemon in cases when
git-credential prompts for IMAP/NNTP passwords.
PublicInbox::Daemon is also designed for read-only
parallelism where all worker processes are the same.
Any subprocesses spawned by -watch are to do specific
tasks for a particular set of inboxes.
Eric Wong [Sat, 27 Jun 2020 10:03:52 +0000 (10:03 +0000)]
watch: just use ->urlmatch
We may just modify PublicInbox::Config->urlmatch in the future
to support git <1.8.5, but I wonder if there's enough users on
git <1.8.5 to justify it.
Eric Wong [Sat, 27 Jun 2020 10:03:50 +0000 (10:03 +0000)]
watch: support multiple watch: directives per-inbox
Some users will find it useful to merge several Maildir or
IMAP mailboxes into one public-inbox. Let them do it, since
we've always supported multi-address inboxes.
Eric Wong [Sat, 27 Jun 2020 10:03:49 +0000 (10:03 +0000)]
watch: imap: be quiet about disconnecting on quit
If ->idle_done was handled successfully, we can just
let normal ->DESTROY disconnect and avoid ugly backtraces
when a user hits Ctrl-C to take down the process group.
Eric Wong [Sat, 27 Jun 2020 10:03:48 +0000 (10:03 +0000)]
watch: support imap.fetchBatchSize parameter
IMAP allows retrieving multiple messages with a single command,
and Mail::IMAPClient supports that. Unfortunately, it means we
slurp multiple messages into memory at once. This option allows
users to trade off memory usage to reduce network round-trips.
Ideally, we'd support pipelining; but AFAIK no widely installed
Perl IMAP library supports it.
Eric Wong [Sat, 27 Jun 2020 10:03:47 +0000 (10:03 +0000)]
watch: avoid long transaction to IMAPTracker
With different polling intervals, multiple processes may
simultaneously write to IMAPtracker. This ought to reduce
SQLite busy waiting and contention issues when importing
many inboxes in parallel.
Eric Wong [Sat, 27 Jun 2020 10:03:46 +0000 (10:03 +0000)]
imaptracker: drop {dbname} field
It's not used anywhere since the IMAPTracker object doesn't
disconnect and reconnect. If we ever need the filename,
{dbh}->sqlite_db_filename may be used.
Eric Wong [Sat, 27 Jun 2020 10:03:45 +0000 (10:03 +0000)]
imaptracker: add {url} field to reduce args
Passing a $url parameter to every function was error-prone
and having {url} field for a short-lived object is appropriate.
This matches the version of IMAPTracker posted by
Eric W. Biederman on 2020-05-15 at:
https://public-inbox.org/meta/87ftc0c3r4.fsf_-_@x220.int.ebiederm.org/
The version I originally imported was based on the one
posted on 2019-10-09:
https://public-inbox.org/meta/874l0i9vhc.fsf_-_@x220.int.ebiederm.org/
Eric Wong [Sat, 27 Jun 2020 10:03:42 +0000 (10:03 +0000)]
watch: stop importers before forking
This fixes cases where watch is handling both Maildirs and IMAP
connections. While we're at it, close open directories in the
IMAP children to save FDs.
Eric Wong [Sat, 27 Jun 2020 10:03:40 +0000 (10:03 +0000)]
watch: support IMAP polling
Not all IMAP servers support IDLE, and IDLE may be prohibitively
expensive for some IMAP servers with many inboxes. So allow
configuring a imap.$IMAP_URL.pollInterval=SECONDS to poll
mailboxes.
We'll also need to poll for NNTP servers in the future.
Eric Wong [Sat, 27 Jun 2020 10:03:39 +0000 (10:03 +0000)]
watch: wire up IMAP IDLE reapers to DS
We can avoid synchronous `waitpid(-1, 0)' and save a process
when simultaneously watching Maildirs.
One DS bug is fixed: ->Reset needs to clear the DS $in_loop flag
in forked children so dwaitpid() fails and allows git processes
to be reaped synchronously. TestCommon also calls DS->Reset
when spawning new processes, since t/imapd.t uses DS->EventLoop
while waiting on -watch to write.
Eric Wong [Sat, 27 Jun 2020 10:03:38 +0000 (10:03 +0000)]
ds: remove fields.pm usage
Since the removal of pseudo-hash support in Perl 5.10, the
"fields" module no longer provides the space or speed benefits
it did in 5.8. It also does not allow for compile-time checks,
only run-time checks.
To me, the extra developer overhead in maintaining "use fields"
args has become a hassle. None of our non-DS-related code uses
fields.pm, nor do any of our current dependencies. In fact,
Danga::Socket (which DS was originally forked from) and its
subclasses are the only fields.pm users I've ever encountered in
the wild. Removing fields may make our code more approachable
to other Perl hackers.
So stop using fields.pm and locked hashes, but continue to
document what fields do for non-trivial classes.
Eric Wong [Sat, 27 Jun 2020 10:03:37 +0000 (10:03 +0000)]
watch: use signalfd for Maildir watching
We can get rid of the janky wannabe
self-using-a-directory-instead-of-pipe thing we needed to
workaround Filesys::Notify::Simple being blocking.
For existing Maildir users, this should be more robust and
immune to missed wakeups for signalfd and kqueue-enabled
systems; as well as being immune to BOFHs clearing $TMPDIR
and preventing notifications from firing.
The IMAP IDLE code still uses normal Perl signals, so it's still
vulnerable to missed wakeups. That will be addressed in future
commits.
Eric Wong [Sat, 27 Jun 2020 10:03:35 +0000 (10:03 +0000)]
kqnotify|fake_inotify: detect Maildir write ops
We need to detect link(2) and rename(2) in other apps
writing to the Maildir.
We'll be removing the Filesys::Notify::Simple from -watch
in favor of using IO::KQueue or Linux::Inotify2 directly.
Ensure non-inotify emulations can support everything we
expect for Maildir writers.
Eric Wong [Sat, 27 Jun 2020 10:03:34 +0000 (10:03 +0000)]
watch: preliminary IMAP support
Only servers with IDLE are supported, for now. Polling will
be needed since users may need to watch many inboxes with
a few active connections due to IMAP server limitations.
Eric Wong [Sat, 27 Jun 2020 10:03:32 +0000 (10:03 +0000)]
watchmaildir: fix check for spam vs ham inbox conflicts
The old check was ineffective since we process the spam folder
config before ham inboxes; and would only fail when attempting
to treat the scalar "watchspam" string as an array ref.
Eric Wong [Sat, 27 Jun 2020 10:03:30 +0000 (10:03 +0000)]
imaptracker: use ~/.local/share/public-inbox/imap.sqlite3
Respect XDG_DATA_HOME to avoid cluttering ~/.public-inbox/.
Existing users of ~/.public-inbox/imap.sqlite3 will remain
supported, but the preference for new data is to use
~/.local/share and other paths standardized by XDG.
Apparently, using $1 from an octet string still results in a
multi-byte string. Thus we need to perform utf8::encode after
the regexp character match to ensure wide characters don't get
passed to encode_base64.
This fixes a bug in which caused -watch to crash when using
PublicInbox::Filter::SubjectTag to remove "[list prefix]"
tags from Subject: lines.
I only found this bug because the proposed -watch updates for
NNTP/IMAP support introduced a possible bug which triggered a
full rescan of old archives:
Eric Wong [Fri, 26 Jun 2020 23:26:18 +0000 (23:26 +0000)]
linkify: support imap, imaps, news, and snews URIs
Since we'll have an IMAP server released soon, maybe imaps://
and imap:// URLs can become popular.
news:// is defined with nntp:// in RFC 5538, and we can at least
support the news:// form in rendered HTML. snews:// may appear
in old mail archives, too, so we'll attempt to support it in
case clients do.
Eric Wong [Thu, 25 Jun 2020 10:09:36 +0000 (10:09 +0000)]
imap: always send EXISTS on uo2m_extend
Clients which are NOT in an IDLE state still need to be
notified of message existence. Unlike the EXPUNGE response,
untagged EXISTS responses seem to be allowed at any time
according to RFC 3501.
We'll also perform uo2m_extend on the NOOP command, since
NOOP is the recommended command for message polling.
Eric Wong [Tue, 23 Jun 2020 18:09:06 +0000 (18:09 +0000)]
ds: flush_write: early return on closed socket
This quiets warnings from IMAP::fetch_blob (called via
long_response) failing to access `$self->{ibx}->git'
because ->{ibx} gets deleted by IMAP::close.
Eric Wong [Tue, 23 Jun 2020 23:21:12 +0000 (23:21 +0000)]
git_async_cat: remove circular reference
While this circular reference was carefully managed to not leak
memory; it was still triggering a warning at -imapd/-nntpd
shutdown due to the EPOLL_CTL_DEL op failing after the $Epoll FD
gets closed.
So remove the circular reference by providing a ref to `undef',
instead.
Eric Wong [Wed, 24 Jun 2020 18:45:07 +0000 (18:45 +0000)]
lock: reduce inotify wakeups
We can reduce the amount of platform-specific code by always
relying on IN_MODIFY/NOTE_WRITE notifications from lock release.
This reduces the number of times our read-only daemons will
need to wake up when -watch sees no-op message changes
(e.g. replied, seen, recent flag changes).
Eric Wong [Tue, 23 Jun 2020 20:51:49 +0000 (20:51 +0000)]
testcommon: DS->Reset when using fork-only subprocess
This fixes a bug on FreeBSD 11 here -nntpd + TEST_RUN_MODE=2
(default) was occasionally causing failures in t/v2writable.t
due to the kqueue descriptor being auto-closed by the OS on fork.
Eric Wong [Sun, 21 Jun 2020 00:21:33 +0000 (00:21 +0000)]
init: add --skip-artnum parameter
For archivists with only newer mail archives, this option allows
reserving reserve NNTP article numbers for yet-to-be-archived
old messages. Indexers will need to be updated to support this
feature in future commits.
-V1 inboxes will now be initialized with SQLite and Xapian
support if this option is used, or if --indexlevel= is
specified.
Eric Wong [Sun, 21 Jun 2020 00:21:31 +0000 (00:21 +0000)]
init: add -j / --jobs parameter
On a powerful (by my standards) machine with 16GB RAM and an
7200 RPM HDD marketed for "enterprise" use, indexing a 8.1G (in
git) LKML snapshot from Sep 2019 did not finish after 7 days
with the default number (3) of Xapian shards (`--jobs=4') and
`--batch-size=10m'.
Indexing starts off fast, but progressively get slower as
contents of the inbox (including Xapian + SQLite DBs) could no
longer be cached by the kernel. Once the on-disk size
increased, HDD seek contention between the Xapian shard workers
slowed the process down to a crawl.
With a single shard, it still took around 3.5 days to index on
the HDD. That's not good, but it's far better than not
finishing after 7 days. So allow unfortunate HDD users to
easily specify a single shard on public-inbox-init.
For reference, a freshly TRIM-ed low-end TLC SSD on the SATA II
bus on the same machine indexes that same snapshot of LKML in
~7 hours with 3 shards and the same 10m batch size. In the past,
a higher-end consumer grade MLC SSDs on similar hardware indexed
a similarly sized-data set in ~4 hours.
Eric Wong [Tue, 16 Jun 2020 22:31:22 +0000 (22:31 +0000)]
nntp: support slow blob retrievals
Having `git cat-file' as a separate process naturally lends
itself to asynchronous dispatch. Our event loop for -nntpd no
longer blocks on slow git storage.
Pipelining in -imapd was tricky and bugs were exposed by
mbsync(1). Update t/nntpd.t to support pipelining ARTICLE
requests to ensure we don't have the same problems -imapd
did during development.