WaitLatch, WaitLatchOrSocket, and WaitEventSetWait now taken an
additional wait_event_info parameter; legal values are defined in
pgstat.h. This makes it possible to uniquely identify every point in
the core code where we are waiting for a latch; extensions can pass
WAIT_EXTENSION.
Because latches were the major wait primitive not previously covered
by this patch, it is now possible to see information in
pg_stat_activity on a large number of important wait events not
previously addressed, such as ClientRead, ClientWrite, and SyncRep.
Unfortunately, many of the wait events added by this patch will fail
to appear in pg_stat_activity because they're only used in background
processes which don't currently appear in pg_stat_activity. We should
fix this either by creating a separate view for such information, or
else by deciding to include them in pg_stat_activity after all.
Michael Paquier and Robert Haas, reviewed by Alexander Korotkov and
Thomas Munro.
Commit ac1d794 ("Make idle backends exit if the postmaster dies.")
introduced a regression on, at least, large linux systems. Constantly
adding the same postmaster_alive_fds to the OSs internal datastructures
for implementing poll/select can cause significant contention; leading
to a performance regression of nearly 3x in one example.
This can be avoided by using e.g. linux' epoll, which avoids having to
add/remove file descriptors to the wait datastructures at a high rate.
Unfortunately the current latch interface makes it hard to allocate any
persistent per-backend resources.
Replace, with a backward compatibility layer, WaitLatchOrSocket with a
new WaitEventSet API. Users can allocate such a Set across multiple
calls, and add more than one file-descriptor to wait on. The latter has
been added because there's upcoming postgres features where that will be
helpful.
In addition to the previously existing poll(2), select(2),
WaitForMultipleObjects() implementations also provide an epoll_wait(2)
based implementation to address the aforementioned performance
problem. Epoll is only available on linux, but that is the most likely
OS for machines large enough (four sockets) to reproduce the problem.
To actually address the aforementioned regression, create and use a
long-lived WaitEventSet for FE/BE communication. There are additional
places that would benefit from a long-lived set, but that's a task for
another day.
Thanks to Amit Kapila, who helped make the windows code I blindly wrote
actually work.
Reported-By: Dmitry Vasilyev Discussion:
CAB-SwXZh44_2ybvS5Z67p_CDz=XFn4hNAD=CnMEF+QqkXwFrGg@mail.gmail.com20160114143931.GG10941@awork2.anarazel.de
Letting backends continue to run if the postmaster has exited prevents
PostgreSQL from being restarted, which in many environments is
catastrophic. Worse, if some other backend crashes, we no longer have
any protection against shared memory corruption. So, arrange for them
to exit instead. We don't want to expend many cycles on this, but
including postmaster death in the set of things that we wait for when
a backend is idle seems cheap enough.
Rajeev Rastogi and Robert Haas
While postgres' use of SSL renegotiation is a good idea in theory, it
turned out to not work well in practice. The specification and openssl's
implementation of it have lead to several security issues. Postgres' use
of renegotiation also had its share of bugs.
Additionally OpenSSL has a bunch of bugs around renegotiation, reported
and open for years, that regularly lead to connections breaking with
obscure error messages. We tried increasingly complex workarounds to get
around these bugs, but we didn't find anything complete.
Since these connection breakages often lead to hard to debug problems,
e.g. spuriously failing base backups and significant latency spikes when
synchronous replication is used, we have decided to change the default
setting for ssl renegotiation to 0 (disabled) in the released
backbranches and remove it entirely in 9.5 and master.
Author: Andres Freund
Discussion: 20150624144148.GQ4797@alap3.anarazel.de
Backpatch: 9.5 and master, 9.0-9.4 get a different patch
The client socket is always in non-blocking mode, and if we actually want
blocking behaviour, we emulate it by sleeping and retrying. But we have
retry loops at different layers for reads and writes, which was confusing.
To simplify, remove all the sleeping and retrying code from the lower
levels, from be_tls_read and secure_raw_read and secure_raw_write, and put
all the logic in secure_read() and secure_write().
Up to now it was impossible to terminate a backend that was trying to
send/recv data to/from the client when the socket's buffer was already
full/empty. While the send/recv calls itself might have gotten
interrupted by signals on some platforms, we just immediately retried.
That could lead to situations where a backend couldn't be terminated ,
after a client died without the connection being closed, because it
was blocked in send/recv.
The problem was far more likely to be hit when sending data than when
reading. That's because while reading a command from the client, and
during authentication, we processed interrupts immediately . That
primarily left COPY FROM STDIN as being problematic for recv.
Change things so that that we process 'die' events immediately when
the appropriate signal arrives. We can't sensibly react to query
cancels at that point, because we might loose sync with the client as
we could be in the middle of writing a message.
We don't interrupt writes if the write buffer isn't full, as indicated
by write() returning EWOULDBLOCK, as that would lead to fewer error
messages reaching clients.
Per discussion with Kyotaro HORIGUCHI and Heikki Linnakangas
Discussion: 20140927191243.GD5423@alap3.anarazel.de
Up to now large swathes of backend code ran inside signal handlers
while reading commands from the client, to allow for speedy reaction to
asynchronous events. Most prominently shared invalidation and NOTIFY
handling. That means that complex code like the starting/stopping of
transactions is run in signal handlers... The required code was
fragile and verbose, and is likely to contain bugs.
That approach also severely limited what could be done while
communicating with the client. As the read might be from within
openssl it wasn't safely possible to trigger an error, e.g. to cancel
a backend in idle-in-transaction state. We did that in some cases,
namely fatal errors, nonetheless.
Now that FE/BE communication in the backend employs non-blocking
sockets and latches to block, we can quite simply interrupt reads from
signal handlers by setting the latch. That allows us to signal an
interrupted read, which is supposed to be retried after returning from
within the ssl library.
As signal handlers now only need to set the latch to guarantee timely
interrupt processing, remove a fair amount of complicated & fragile
code from async.c and sinval.c.
We could now actually start to process some kinds of interrupts, like
sinval ones, more often that before, but that seems better done
separately.
This work will hopefully allow to handle cases like being blocked by
sending data, interrupting idle transactions and similar to be
implemented without too much effort. In addition to allowing getting
rid of ImmediateInterruptOK, that is.
Author: Andres Freund
Reviewed-By: Heikki Linnakangas
This allows to introduce more elaborate handling of interrupts while
reading from a socket. Currently some interrupt handlers have to do
significant work from inside signal handlers, and it's very hard to
correctly write code to do so. Generic signal handler limitations,
combined with the fact that we can't safely jump out of a signal
handler while reading from the client have prohibited implementation
of features like timeouts for idle-in-transaction.
Additionally we use the latch code to wait in a couple places where we
previously only had waiting code on windows as other platforms just
busy looped.
This can increase the number of systemcalls happening during FE/BE
communication. Benchmarks so far indicate that the impact isn't very
high, and there's room for optimization in the latch code. The chance
of cleaning up the usage of latches gives us, seem to outweigh the
risk of small performance regressions.
This commit theoretically can't used without the next patch in the
series, as WaitLatchOrSocket is not defined to be fully signal
safe. As we already do that in some cases though, it seems better to
keep the commits separate, so they're easier to understand.
Author: Andres Freund
Reviewed-By: Heikki Linnakangas
This refactoring is in preparation for adding support for other SSL
implementations, with no user-visible effects. There are now two #defines,
USE_OPENSSL which is defined when building with OpenSSL, and USE_SSL which
is defined when building with any SSL implementation. Currently, OpenSSL is
the only implementation so the two #defines go together, but USE_SSL is
supposed to be used for implementation-independent code.
The libpq SSL code is changed to use a custom BIO, which does all the raw
I/O, like we've been doing in the backend for a long time. That makes it
possible to use MSG_NOSIGNAL to block SIGPIPE when using SSL, which avoids
a couple of syscall for each send(). Probably doesn't make much performance
difference in practice - the SSL encryption is expensive enough to mask the
effect - but it was a natural result of this refactoring.
Based on a patch by Martijn van Oosterhout from 2006. Briefly reviewed by
Alvaro Herrera, Andreas Karlsson, Jeff Janes.
A number of issues were identified by the Coverity scanner and are
addressed in this patch. None of these appear to be security issues
and many are mostly cosmetic changes.
Short comments for each of the changes follows.
Correct the semi-colon placement in be-secure.c regarding SSL retries.
Remove a useless comparison-to-NULL in proc.c (value is dereferenced
prior to this check and therefore can't be NULL).
Add checking of chmod() return values to initdb.
Fix a couple minor memory leaks in initdb.
Fix memory leak in pg_ctl- involves free'ing the config file contents.
Use an int to capture fgetc() return instead of an enum in pg_dump.
Fix minor memory leaks in pg_dump.
(note minor change to convertOperatorReference()'s API)
Check fclose()/remove() return codes in psql.
Check fstat(), find_my_exec() return codes in psql.
Various ECPG memory leak fixes.
Check find_my_exec() return in ECPG.
Explicitly ignore pqFlush return in libpq error-path.
Change PQfnumber() to avoid doing an strdup() when no changes required.
Remove a few useless check-against-NULL's (value deref'd beforehand).
Check rmtree(), malloc() results in pg_regress.
Also check get_alternative_expectfile() return in pg_regress.
Commit 820f08cabd claimed to make the server
and libpq handle SSL protocol versions identically, but actually the server
was still accepting SSL v3 protocol while libpq wasn't. Per discussion,
SSL v3 is obsolete, and there's no good reason to continue to accept it.
So make the code really equivalent on both sides. The behavior now is
that we use the highest mutually-supported TLS protocol version.
Marko Kreen, some comment-smithing by me
This sets up ECDH key exchange, when compiling against OpenSSL that
supports EC. Then the ECDHE-RSA and ECDHE-ECDSA cipher suites can be
used for SSL connections. The latter one means that EC keys are now
usable.
The reason for EC key exchange is that it's faster than DHE and it
allows to go to higher security levels where RSA will be horribly slow.
There is also new GUC option ssl_ecdh_curve that specifies the curve
name used for ECDH. It defaults to "prime256v1", which is the most
common curve in use in HTTPS.
From: Marko Kreen <markokr@gmail.com>
Reviewed-by: Adrian Klaver <adrian.klaver@gmail.com>
By default, OpenSSL (and SSL/TLS in general) lets the client cipher
order take priority. This is OK for browsers where the ciphers were
tuned, but few PostgreSQL client libraries make the cipher order
configurable. So it makes sense to have the cipher order in
postgresql.conf take priority over client defaults.
This patch adds the setting "ssl_prefer_server_ciphers" that can be
turned on so that server cipher order is preferred. Per discussion,
this now defaults to on.
From: Marko Kreen <markokr@gmail.com>
Reviewed-by: Adrian Klaver <adrian.klaver@gmail.com>
Current OpenSSL code includes a BIO_clear_retry_flags() step in the
sock_write() function. Either we failed to copy the code correctly, or
they added this since we copied it. In any case, lack of the clear step
appears to be the cause of the server lockup after connection loss reported
in bug #8647 from Valentine Gogichashvili. Assume that this is correct
coding for all OpenSSL versions, and hence back-patch to all supported
branches.
Diagnosis and patch by Alexander Kukushkin.
These functions must be careful that they return the intended value of
errno to their callers. There were several scenarios where this might
not happen:
1. The recent SSL renegotiation patch added a hunk of code that would
execute after setting errno. In the first place, it's doubtful that we
should consider renegotiation to be successfully completed after a failure,
and in the second, there's no real guarantee that the called OpenSSL
routines wouldn't clobber errno. Fix by not executing that hunk except
during success exit.
2. errno was left in an unknown state in case of an unrecognized return
code from SSL_get_error(). While this is a "can't happen" case, it seems
like a good idea to be sure we know what would happen, so reset errno to
ECONNRESET in such cases. (The corresponding code in libpq's fe-secure.c
already did this.)
3. There was an (undocumented) assumption that client_read_ended() wouldn't
change errno. While true in the current state of the code, this seems less
than future-proof. Add explicit saving/restoring of errno to make sure
that changes in the called functions won't break things.
I see no need to back-patch, since #1 is new code and the other two issues
are mostly hypothetical.
Per discussion with Amit Kapila.
The existing renegotiation code was home for several bugs: it might
erroneously report that renegotiation had failed; it might try to
execute another renegotiation while the previous one was pending; it
failed to terminate the connection if the renegotiation never actually
took place; if a renegotiation was started, the byte count was reset,
even if the renegotiation wasn't completed (this isn't good from a
security perspective because it means continuing to use a session that
should be considered compromised due to volume of data transferred.)
The new code is structured to avoid these pitfalls: renegotiation is
started a little earlier than the limit has expired; the handshake
sequence is retried until it has actually returned successfully, and no
more than that, but if it fails too many times, the connection is
closed. The byte count is reset only when the renegotiation has
succeeded, and if the renegotiation byte count limit expires, the
connection is terminated.
This commit only touches the master branch, because some of the changes
are controversial. If everything goes well, a back-patch might be
considered.
Per discussion started by message
20130710212017.GB4941@eldon.alvh.no-ip.org
The Solaris Studio compiler warns about these instances, unlike more
mainstream compilers such as gcc. But manual inspection showed that
the code is clearly not reachable, and we hope no worthy compiler will
complain about removing this code.
Both libpq and the backend would truncate a common name extracted from a
certificate at 32 bytes. Replace that fixed-size buffer with dynamically
allocated string so that there is no hard limit. While at it, remove the
code for extracting peer_dn, which we weren't using for anything; and
don't bother to store peer_cn longer than we need it in libpq.
This limit was not so terribly unreasonable when the code was written,
because we weren't using the result for anything critical, just logging it.
But now that there are options for checking the common name against the
server host name (in libpq) or using it as the user's name (in the server),
this could result in undesirable failures. In the worst case it even seems
possible to spoof a server name or user name, if the correct name is
exactly 32 bytes and the attacker can persuade a trusted CA to issue a
certificate in which that string is a prefix of the certificate's common
name. (To exploit this for a server name, he'd also have to send the
connection astray via phony DNS data or some such.) The case that this is
a realistic security threat is a bit thin, but nonetheless we'll treat it
as one.
Back-patch to 8.4. Older releases contain the faulty code, but it's not
a security problem because the common name wasn't used for anything
interesting.
Reported and patched by Heikki Linnakangas
Security: CVE-2012-0867
This allows changing the location of the files that were previously
hard-coded to server.crt, server.key, root.crt, root.crl.
server.crt and server.key continue to be the default settings and are
thus required to be present by default if SSL is enabled. But the
settings for the server-side CA and CRL are now empty by default, and
if they are set, the files are required to be present. This replaces
the previous behavior of ignoring the functionality if the files were
not found.
This disables an entirely unnecessary "sanity check" that causes failures
in nonblocking mode, because OpenSSL complains if we move or compact the
write buffer. The only actual requirement is that we not modify pending
data once we've attempted to send it, which we don't. Per testing and
research by Martin Pihlak, though this fix is a lot simpler than his patch.
I put the same change into the backend, although it's less clear whether
it's necessary there. We do use nonblock mode in some situations in
streaming replication, so seems best to keep the same behavior in the
backend as in libpq.
Back-patch to all supported releases.
requests for client certs. This lets a client with a keystore select the
appropriate client certificate to send. In particular, this is necessary
to get Java clients to work in all but the most trivial configurations.
Per discussion of bug #5468.
Craig Ringer
how often we do SSL session key renegotiation. Can be set to
0 to disable renegotiation completely, which is required if
a broken SSL library is used (broken patches to CVE-2009-3555
a known cause) or when using a client library that can't do
renegotiation.
This includes two new kinds of postmaster processes, walsenders and
walreceiver. Walreceiver is responsible for connecting to the primary server
and streaming WAL to disk, while walsender runs in the primary server and
streams WAL from disk to the client.
Documentation still needs work, but the basics are there. We will probably
pull the replication section to a new chapter later on, as well as the
sections describing file-based replication. But let's do that as a separate
patch, so that it's easier to see what has been added/changed. This patch
also adds a new section to the chapter about FE/BE protocol, documenting the
protocol used by walsender/walreceivxer.
Bump catalog version because of two new functions,
pg_last_xlog_receive_location() and pg_last_xlog_replay_location(), for
monitoring the progress of replication.
Fujii Masao, with additional hacking by me
at least in some Windows versions, these functions are capable of returning
a failure indication without setting errno. That puts us into an infinite
loop if the previous value happened to be EINTR. Per report from Brendan
Hill.
Back-patch to 8.2. We could take it further back, but since this is only
known to be an issue on Windows and we don't support Windows before 8.2,
it does not seem worth the trouble.
attacks where an attacker would put <attack>\0<propername> in the field and
trick the validation code that the certificate was for <attack>.
This is a very low risk attack since it reuqires the attacker to trick the
CA into issuing a certificate with an incorrect field, and the common
PostgreSQL deployments are with private CAs, and not external ones. Also,
default mode in 8.4 does not do any name validation, and is thus also not
vulnerable - but the higher security modes are.
Backpatch all the way. Even though versions 8.3.x and before didn't have
certificate name validation support, they still exposed this field for
the user to perform the validation in the application code, and there
is no way to detect this problem through that API.
Security: CVE-2009-4034
to the documented API value. The previous code got it right as
it's implemented, but accepted too much/too little compared to
the API documentation.
Per comment from Zdenek Kotala.