Eric Wong [Thu, 16 Apr 2020 00:29:38 +0000 (00:29 +0000)]
t/httpd-corner: improve reliability and diagnostics
The graceful-shutdown-on-PUT test is unreliable because we can't
rely on a FIFO as we do with the GET tests. So increase the
delay to 100ms since that seems enough on my system even with
CONFIG_HZ=100.
Add a timeout and backtrace to the $check_self sub to help with
further diagnostics while we're at it, too.
It would be nice if there were a portable syscall tracing
mechanism we could attach to the -httpd process to make the test
more determistic...
I've observed FreeBSD 11.2 read(2) having one of three
behaviors after a failed write(2) on a socket:
1) returning number of bytes read
2) failing with ECONNRESET
3) returning with EOF
1) is the most common, and I've only seen 1) on Linux. It may
be possible to use SO_LINGER or shutdown(2) to ensure 1) always
happens, but SO_LINGER behavior seems inconsistent across OSes,
especially with non-blocking sockets.
Since these tests are corner-cases where we're dealing with
broken/malicious clients, lets continue spending the least
amount of syscalls protecting ourselves in the daemon and
instead make the client-side test code tolerate more socket
implementations.
Eric Wong [Sat, 11 Apr 2020 10:53:28 +0000 (10:53 +0000)]
dskqxs: ignore EV_SET errors on EVFILT_WRITE
Just like the EPOLL_CTL_ADD emulation path, the EPOLL_CTL_MOD
and EPOLL_CTL_DEL emulation paths can fail if attempting to
install an EVFILT_WRITE for a read-only pipe.
I've only observed this on the EPOLL_CTL_DEL emulation path, but
I suspect it could happen on the EPOLL_CTL_MOD path as well.
Increasing the amount of read-only pipes we rely on with altid
exports via sqlite3 made this old bug more apparent and
reproducible while looping the test suite.
This may be adjusted in the future to deal with write-only
pipes, but we currently don't have any of those watched by
kqueue.
Eric Wong [Tue, 7 Apr 2020 09:35:14 +0000 (04:35 -0500)]
doc: add technical/whyperl
Some people don't like Perl; but it exists, there's no
avoiding it with everything that depends on it. And
nearly all code still works unmodified after 20 years.
Eric Wong [Tue, 7 Apr 2020 21:55:53 +0000 (21:55 +0000)]
tests: document run_mode=1 as not implemented
It was implemented at some point, but it was more things to
support and the worst of both worlds: both unrealistic compared
to real-world use and slower than run_mode=2.
Eric Wong [Mon, 6 Apr 2020 08:32:52 +0000 (08:32 +0000)]
view: do not redundantly obfuscate addresses
We shouldn't rerun the address obfuscator on data we've
already run through. Instead, run through the unescaped
text part and substitute the UTF-8 "\x{2022}" substitution
before it hits HTML escaping
Eric Wong [Mon, 6 Apr 2020 20:23:01 +0000 (20:23 +0000)]
portability: constants for NetBSD
NetBSD implements O_CLOEXEC, so let us use it to avoid
inadvertant FD sharing. It also has the same value for SIGWINCH
as Linux and the other BSDs we support.
Eric Wong [Sun, 5 Apr 2020 07:53:49 +0000 (07:53 +0000)]
git: reduce stat buffer storage overhead
The stat() array is a whopping 480 bytes (on x86-64, Perl 5.28),
while the new packed representation of two 64-bit doubles as a
scalar is "only" 56 bytes. This can add up when there's many
inboxes. Just use a string comparison on the packed
representation.
Some 32-bit Perl builds (IIRC OpenBSD) lack quad support, so
doubles were chosen for pack() portability.
Eric Wong [Sun, 5 Apr 2020 07:53:48 +0000 (07:53 +0000)]
mbox: halve ->getline "context switches"
We don't need to take extra trips through the event loop for a
single message (in the common case of Message-IDs being unique).
In fact, holding the body reference left behind by Email::Simple
could be harmful to memory usage, though in practice it's not a
big problem since code paths which use Email::MIME take far more.
Eric Wong [Sun, 5 Apr 2020 07:53:47 +0000 (07:53 +0000)]
release large (non ref) scalars using `undef $sv'
Using `undef EXPR' like a function call actually frees the heap
memory associated with the scalar, whereas `$sv = undef' or
`$sv = ""' will hold the buffer around until $sv goes out
of scope.
The `sv_set_undef' documentation in the perlapi(1) manpage
explicitly states this:
The perl equivalent is "$sv = undef;". Note that it doesn't
free any string buffer, unlike "undef $sv".
And I've confirmed by reading Dump() output from Devel::Peek.
We'll also inline the old index_body sub in SearchIdx.pm to make
the scope of the scalar more obvious.
This change saves several hundred kB RSS on both -index and
-httpd when hitting large emails with thousands of lines.
Eric Wong [Sat, 4 Apr 2020 23:51:41 +0000 (23:51 +0000)]
xt/msgtime_cmp: fix false positives from msgtime change
commit d857e7dc0d816b635a7ead09c3273f8c2d2434be
("msgtime: assume +0000 if TZ missing when using Date::Parse")
introduced a behavior change which was causes false positives
when compared to the old code.
Update the "old" implementation to match this overdue behavior
change.
Eric Wong [Sun, 5 Apr 2020 01:28:49 +0000 (01:28 +0000)]
wwwstatic: set "Vary: Accept-Encoding" in static gzip response
We don't want to confuse intermediate caches into serving
gzipped content to any clients which can't handle it. It
probably doesn't matter in practice, though, since every HTTP
client seems to handle "Content-Encoding: gzip" regardless of
whether it was requested or not, though I could expect some
nc/socat/telnet/s_client users being annoyed.
This also matches the behavior of Plack::Middleware::Deflater
and other deflater implementations.
Eric Wong [Sat, 4 Apr 2020 08:03:18 +0000 (08:03 +0000)]
viewdiff: reduce sub parameter count
We're slowly moving towards doing all of our output buffering
into a single buffer, so passing that around on the stack as
a dedicated parameter is confusing.
Eric Wong [Sat, 4 Apr 2020 08:03:17 +0000 (08:03 +0000)]
view: dedupe_subject: allow "0" as a valid Subject
While rare in practice (even by spammers), A single "0" could
theoretically be the entire contents of a Subject line. So
use the Perl 5.10+ defined-or operator to improve correctness
of subject deduplication.
These seem mostly harmless since Perl will just truncate the
match and start a new one on a newline boundary in our case.
The only downside is we'd end up with redundant <span> tags in
HTML.
Limiting the number of line matched ourselves with `{1,$NUM}'
doesn't seem prudent since lines vary in length, so we continue
to defer the job of limiting matches to the Perl regexp engine.
I've noticed this warning in practice on 100K+ line patches to
locale data.
Eric Wong [Wed, 1 Apr 2020 06:16:20 +0000 (06:16 +0000)]
mid: add $MID_EXTRACT regexp for export
This allows us to consistently enforce the same Message-ID
extraction rules everywhere and makes it easier for us to
make changes in the future.
Update scripts/ssoma-replay, as well, but don't rely on
PublicInbox::* modules in that since it's legacy and
public-inbox was never a dependency of ssoma.
Eric Wong [Wed, 1 Apr 2020 06:16:18 +0000 (06:16 +0000)]
smsg: inline _extract_mid functionality
No need to keep an extra sub which isn't called anywhere else,
and the mid_clean call is redundant since mid_mime already
plucks the msgid out of the angle brackets.
Eric Wong [Tue, 31 Mar 2020 08:49:36 +0000 (08:49 +0000)]
v2writable: index Message-IDs w/ spaces properly
Message-IDs can apparently contain spaces and other weird
characters. Ensure we pass those properly to shard subprocesses
when importing messages in parallel mode.
Our NNTP request parser does not deal with spaces in the
Message-ID, yet, and I don't expect most NNTP clients to,
either. Nor does the Net::NNTP client handle them in responses.
Eric Wong [Mon, 30 Mar 2020 19:42:54 +0000 (19:42 +0000)]
t/multi-mid: allow test to run w/o Xapian
While the v1 inbox in this test is created without Xapian,
the v2 inbox in this test defaults to having Xapian enabled
regardless of whether it's installed or not.
Fixes: c7acdfe78bda5bf3 ("v2: SDBM-based multi Message-ID queue")
git-cat-file(1) may return less than the $BIN_DETECT value for
some blobs, so ensure we repopulate the values in $ctx for
retries in that case, otherwise we'll lose `$ctx->{-res}' and
die when attempting to use `undef' as an array ref.
Eric Wong [Mon, 30 Mar 2020 05:18:42 +0000 (00:18 -0500)]
wwwstream::oneshot => html_oneshot
And use Exporter to make our life easier, since WwwAltId was
using a non-existent PublicInbox::WwwResponse namespace in error
paths which doesn't get noticed by `perl -c' or exercised by
tests on normal systems.
Fixes: 6512b1245ebc6fe3 ("www: add endpoint to retrieve altid dumps")
Eric Wong [Sat, 28 Mar 2020 00:56:04 +0000 (00:56 +0000)]
index: support --compact / -c on command-line
It's more convenient to specify `-c' / `--compact' on the
command-line when reindexing than it is to invoke
public-inbox-compact(1) separately.
This is especially convenient in low-space situations when
public-inbox-index is operating on multiple inboxes
sequentially, as compaction can happen immediately after
indexing each inbox, instead of waiting until all inboxes are
indexed.
Eric Wong [Sat, 28 Mar 2020 00:56:03 +0000 (00:56 +0000)]
searchidxshard: ensure we set indexlevel on shard[0]
For sharded v2 repositories with few-enough messages, it is
possible for shard[0] to go unused and never trigger the
->commit_txn_lazy to set the indexlevel field in Xapian
metadata.
So set it immediately at initialization and avoid this case.
While we're at it, avoid triggering needless pwrite syscalls
from ->set_metadata by checking with ->get_metadata, first.
Eric Wong [Thu, 26 Mar 2020 08:21:29 +0000 (08:21 +0000)]
wwwaltid: inform users to use POST instead of GET
Seeing the example config linkified, some users may inevitably
try to following it in a browser with a GET request. Provide
a helpful message to inform users to use POST instead of
attempting to treat /$INBOX/$ALTID.sql.gz as a Message-Id.
Eric Wong [Sat, 21 Mar 2020 02:03:54 +0000 (02:03 +0000)]
www: add endpoint to retrieve altid dumps
This ensures all our indexed data, including data from altid
searches (e.g. "gmane:$ARTNUM") is retrievable.
It uses a "POST" request to avoid wasting cycles when invoked by
crawlers, since it could potentially be several megabytes of
data not indexable by search engines.
Eric Wong [Sat, 21 Mar 2020 02:03:53 +0000 (02:03 +0000)]
altid: warn about non-word prefixes
We only support searching on prefixes matching /\A\w+\z/ because
Xapian requires ':' to delimit the prefix and splits on spaces
without quotes.
I've also verified Xapian supports multibyte UTF-8 characters,
underscores, and bare numbers as search prefixes, so there's
no need to restrict it beyond what Perl's UTF-8 aware \w
character class offers.
Eric Wong [Sat, 21 Mar 2020 02:03:52 +0000 (02:03 +0000)]
wwwtext: show thread endpoint w/ indexlevel=basic
And show contact info when there's no indexing, at all.
Installations where Xapian is too expensive can still support
threading since it only depends on SQLite, so we need to inform
users of what's available.
Eric Wong [Sat, 21 Mar 2020 02:03:51 +0000 (02:03 +0000)]
search: clobber -user_pfx on query parser initialization
While we don't currently reinitialize the query parser for
the lifetime of a PublicInbox::Search object and have no plans
to, it's incorrect to be appending to an existing array in
case we reininitialize the query parser in the future.
Eric Wong [Sat, 21 Mar 2020 02:03:49 +0000 (02:03 +0000)]
mbox: need_gzip uses WwwStream::oneshot
This makes the error page more consistent.
Not that it really matters since Compress::Raw::Zlib and
IO::Compress packages have been distributed with Perl since
5.10.x. Of course, zlib itself is also a dependency of git.
Eric Wong [Sat, 21 Mar 2020 02:03:46 +0000 (02:03 +0000)]
wwwstream: introduce oneshot API to avoid ->getline
The ->getline API is only useful for limiting memory use when
streaming responses containing multiple emails or log messages.
However it's unnecessary complexity and overhead for callers
(PublicInbox::HTTP) when there's only a single message.
Eric Wong [Sat, 21 Mar 2020 02:03:45 +0000 (02:03 +0000)]
gzipfilter: lazy allocate the deflate context
zlib contexts are memory-intensive, particularly when used for
compression. Since the gzip filter may be sitting in a limiter
queue for a long period, delay the allocation we actually have
data to translate, and not a moment sooner.
Eric Wong [Sat, 21 Mar 2020 02:03:44 +0000 (02:03 +0000)]
qspawn: reinstate filter support, add gzip filter
We'll be supporting gzipped from sqlite3(1) dumps
for altid files in future commits.
In the future (and if we survive), we may replace
Plack::Middleware::Deflater with our own GzipFilter to work
better with asynchronous responses without relying on
memory-intensive anonymous subs.
Eric Wong [Fri, 20 Mar 2020 08:18:21 +0000 (08:18 +0000)]
v2: SDBM-based multi Message-ID queue
This lets us store author and committer times for deferred
indexing messages with ambiguous Message-IDs. This allows
us to reproducibly reindex messages with the git commit
and author times when a rare message lacks Received and/or
Date headers while having ambiguous Message-IDs.
Eric Wong [Fri, 20 Mar 2020 08:18:18 +0000 (08:18 +0000)]
*idx: pass $smsg in more places instead of many args
We can pass blessed PublicInbox::Smsg objects to internal
indexing APIs instead of having long parameter lists in some
places. The end goal is to avoid parsing redundant information
each step of the way and hopefully make things more
understandable.
Eric Wong [Fri, 20 Mar 2020 08:18:17 +0000 (08:18 +0000)]
overidx: parse_references: less error-prone args
Favor `$smsg->{mid}' instead of `$mid0' to reduce parameters
down-the-line, but favor passing the Email::MIME::Header object
around instead of relying on the bloat-prone `$smsg->{mime}'
and calling ->header_obj on it.
Since the introduction of over.sqlite3, SearchMsg is not tied to
our search functionality in any way, so stop confusing ourselves
and future hackers by just calling it "PublicInbox::Smsg".
Eric Wong [Fri, 20 Mar 2020 08:18:14 +0000 (08:18 +0000)]
v2writable: preserve timestamps from import
While v2 indexing is triggered immediately after writing the
commit to the git repository, there may be a gap between when
PublicInbox::Import generates a timestamp and when
PublicInbox::SearchIdx sees the message. So follow the mirror
indexing behavior and take the to-be-indexed (time|date)stamps
directly from the git commit.
Eric Wong [Fri, 20 Mar 2020 08:18:13 +0000 (08:18 +0000)]
index: use git commit times on missing Date/Received
When indexing messages without Date: and/or Received: headers,
fall back to using timestamps originally recorded by git in the
commit object. This allows git mirrors to preserve the import
datestamp and timestamp of a message according to what was fed
into git, instead of blindly falling back to the current time.
Eric Wong [Sat, 21 Mar 2020 05:24:32 +0000 (05:24 +0000)]
t/www_listing: avoid 'once' warnings
We reach into the WwwListing package directly to retrieve
that JSON encoder/decoder object, and we can't rely on `use'
since WwwListing loading may fail if Plack is missing.
Eric Wong [Thu, 19 Mar 2020 08:32:56 +0000 (03:32 -0500)]
viewdiff: favor `qr' to precompile regexps
We can also avoid `o' regexp modifier, since it isn't
recommended by Perl upstream, anymore (although we don't
have any bugs or unintended behavior because of it).
Eric Wong [Thu, 19 Mar 2020 08:32:55 +0000 (03:32 -0500)]
daemon: do more immortal allocations up front
Doing immortal allocations late can cause those allocations
to end up in places where it fragments the heap. So do more
things up front for long-lived daemons.
Eric Wong [Thu, 19 Mar 2020 08:32:53 +0000 (03:32 -0500)]
wwwlisting: avoid lazy loading JSON module
We already lazy-load WwwListing for the CGI script, and
hiding another layer of lazy-loading makes things difficult
to do WWW->preload.
We want long-lived processes to do all long-lived allocations up
front to avoid fragmentation in the allocator, but we'll still
support short-lived processes by lazy-loading individual modules
in the PublicInbox::* namespace.
Mixing up allocation lifetimes (e.g. doing immortal allocations
while a large amount of space is taken by short-lived objects)
will cause fragmentation in any allocator which favors large
contiguous regions for performance reasons. This includes any
malloc implementation which relies on sbrk() for the primary
heap, including glibc malloc.
Eric Wong [Thu, 19 Mar 2020 08:32:51 +0000 (03:32 -0500)]
www: update ->preload for newer modules
We'll also avoid explicitly loading standard library modules
like POSIX and Digest::SHA, here; instead we load our own
modules and let those load whatever non-PublicInbox:: modules
they need.
Eric Wong [Thu, 19 Mar 2020 07:51:53 +0000 (07:51 +0000)]
doc: standards: add references to RFC 5322 (and RFC 822)
RFC 5322 is the latest one in this line, but much documentation
and even command-line options in other programs (e.g. git) refer
to RFC 2822 or even RFC 822.
Eric Wong [Tue, 17 Mar 2020 06:52:17 +0000 (06:52 +0000)]
http: fix RFC conformance w.r.t. message length
We need to favor "Transfer-Encoding: chunked" over the value of
the Content-Length header. We should also reject bogus,
duplicate and/or unreasonable values for both these, since they
can trigger unexpected behavior when combined with other HTTP
parsers in proxies such as varnish, nginx, haproxy, etc...
Eric Wong [Tue, 3 Mar 2020 05:03:11 +0000 (05:03 +0000)]
INSTALL: refer to the proper Debian version
Debian 10.0 was released July 2019, so update our documentation
to reflect that. While we're at it, fixup a broken footnote
reference for Inline::C, too.
Eric Wong [Sat, 7 Mar 2020 12:15:23 +0000 (12:15 +0000)]
daemon: remove unused $parent_pipe variable
We can just create a ParentPipe and let PublicInbox::DS
manage its life cycle. While we're at it, favor `\&coderef'
over `*coderef' so we're explicit about it being a code ref
and not some other ref type.
Eric Wong [Tue, 25 Feb 2020 09:23:03 +0000 (09:23 +0000)]
msgtime: assume +0000 if TZ missing when using Date::Parse
Some old emails don't have timezone offsets, since our
Date::Parse code path takes a liberal interpretation of dates,
fallback to using "+0000" as the timezone offset since it's
closer to the actual date of the message than whatever the
current date is.
Eric Wong [Wed, 26 Feb 2020 10:21:12 +0000 (10:21 +0000)]
import: drop '<' and '>' characters in addresses
Some strange "From:" lines will cause Email::Address::XS to
leave '<' (and presumably '>') in the address which
git-fast-import won't accept even if quoted. Workaround this
problem by deleting '<' and '>' the same way we delete them for
the ident name.
Eric Wong [Wed, 26 Feb 2020 00:44:05 +0000 (00:44 +0000)]
searchview: improve naming and simplify hash override
`%over' could be confused for the overview SQLite DB
instance, so call it `%override', instead. There's
also no need to write a loop to override a hash when
the language can do it for us.
Eric Wong [Mon, 24 Feb 2020 08:08:22 +0000 (08:08 +0000)]
v2writable: lookup_content => content_exists
It only needs to return a boolean, since none of the current
callers care about the return value. Thus avoid a hash table
assignment and use of `$smsg->{mime}', here.
Eric Wong [Mon, 24 Feb 2020 07:33:27 +0000 (07:33 +0000)]
viewdiff: remove optional CR handling
The only caller of `flush_diff' is `add_text_body', and that
already did CRLF conversion on the text part. The regexps in
SolverGit still need to preserve CR, however, since that
actually applies patches (instead of rendering them), and we
need to preserve CRLF patches for CRLF files.
Eric Wong [Mon, 24 Feb 2020 07:33:26 +0000 (07:33 +0000)]
hval: ascii_html: drop CRLF => LF conversion
Instead, we add CRLF conversion to the only remaining place
which needs it, ViewVCS. This save many redundant ops in in
many places.
The only other place where this mattered was in
View::add_text_body, but we already started doing CRLF
conversions when we added diff parsing and link generation for
ViewVCS. Otherwise, all other places we used this was for
header viewing and Email::MIME doesn't preserve CRLF in headers.