This commit adds logic to retry our poll call when waiting for the
connection to complete, in the event that we are interrupted by a
signal.
Additionally we do some simple bookkeeping to keep track of the overall
timeout specified by the user.
Fixes#1206
* Implement redisSetTcpUserTimeout to set socket option TCP_USER_TIMEOUT
* Documentation for redisSetTcpUserTimeout and some more undocumented functions
Documentation for redisReconnect() and the setters of socket options:
* redisKeepAlive()
* redisEnableKeepAliveWithInterval()
* redisSetTcpUserTimeout()
Improve coverage (#734)
* Remove duplicate tests
- double covered by:
"Can parse RESP3 doubles"
- bool covered via:
"Can parse RESP3 bool"
* Make (connect) timeout in test config general
* Set error string in Unix connect with invalid timeout
Restructure testcase since redisConnectWithTimeout() and
redisConnectUnixWithTimeout() now behaves similar.
* Use quiet flag in lcov/genhtml instead of piping to /dev/null
* Fixup of redisCommandArgv test case
* Update test case to match what it covers
Use new test case info text since the previous one seemed copy&pasted.
The sought coverage was the handling of the parent-chaining
for a double object, which the test case now focuses on.
Co-authored-by: Ariel <ashtul@gmail.com>
Fix ProtocolError
This commit attempts to fix hiredis such that a recoverable write error
will be retried rather than throwing a hard error.
Since our read/write functions are now behind function pointers, we
specify semantically that a return value of < 0 is a hard error, 0 a
recoverable error, and > 0 a success.
Our default `redisNetRead` function was already doing something similar
so this also improves code consistency.
Resolves#961
Co-authored-by: Maksim Tuleika <maksim.tuleika@appcast.io>
Use a windows specific keepalive function.
While it is possible to toggle `TCP_KEEPALIVE` in windows via
setsockopt, you have to use `WSAIoctl` to set the interval.
Since `WSAIoctl` can actually do all of this in one call (toggle the
option, and set the corresponding interval), just use that in Windows
and avoid the call to `setsockopt` alltogether.
Fixes: #1100
* Adds an indirection to every allocation/deallocation to allow users to
plug in ones of their choosing (use custom functions, jemalloc, etc).
* Gracefully handle OOM everywhere in hiredis. This should make it possible
for users of the library to have more flexibility in how they handle such situations.
* Changes `redisReaderTask->elements` from an `int` to a `long long` to prevent
a possible overflow when transferring the task elements into a `redisReply`.
* Adds a configurable `max elements` member to `redisReader` that defaults to
2^32 - 1. This can be set to "unlimited" by setting the value to zero.
Unit tests in Windows and a Windows timeout fix
This commit gets our unit tests compiling and running on Windows as well as removes a duplicated `timeval` -> `DWORD` conversion logic in sockcompat.c
There are minor differences in behavior between Linux and Windows to note:
1. In Windows, opening a non-existent hangs forever in WSAPoll whereas
it correctly returns with a "Connection refused" error on Linux.
For that reason, I simply skip this test in Windows.
It may be related to this known issue:
https://daniel.haxx.se/blog/2012/10/10/wsapoll-is-broken/
2. Timeouts are handled slightly differently in Windows and Linux.
In Linux, we intentionally set REDIS_ERR_IO for connection
timeouts whereas in Windows we set REDIS_ERR_TIMEOUT. It may be
prudent to fix this discrepancy although there are almost certainly
users relying on the current behavior.
The recv/send calls are more portable than read/write, since unlike the
latter, the former work with Windows sockets.
We also check for EWOULDBLOCK instead of EAGAIN. On most Unices, EAGAIN
and EWOULDBLBOCK are the same thing. However, on Windows they are
different, and send/recv are expected to give EWOULDBLOCK for
non-blocking sockets.
The redisFD type should be equal to the system native socket file
desciptor type (for POSIX, this is a plain int).
We also introduce the REDIS_INVALID_FD value, which maps to -1 on POSIX
systems.
This makes hiredis.c free from system calls related to socket I/O. This
is also makes the treatment of raw socket connections more similar to
the SSL backend.
free(NULL) is a valid NOP. Most of the hiredis free functions behave the
same way. redisReaderFree is updated to also be NULL-safe.
There is one redundant NULL check at sds.c:1036, but it's left as is
since sds is imported from upstream.
Signed-off-by: Justin Brewer <jzb0012@auburn.edu>
freeaddrinfo is not required by POSIX to be NULL-safe. OpenBSD will
SIGSEGV. NetBSD will assert. FreeBSD up to 11.1 will SIGSEGV, while in
future versions, it will be a silent NOP [1].
Commit d4b715f0aa ("Fix potential race in 'invalid timeout' tests")
added a code path to _redisContextConnectTcp which calls
freeaddrinfo(NULL), triggering the segfault. Put a NULL check around the
call to freeaddrinfo.
[1] - e916723903
Signed-off-by: Justin Brewer <jzb0012@auburn.edu>
Since _GNU_SOURCE is now guaranteed to be unset, it is no longer
necessary to support the GNU-specific version of strerror_r.
Drop __redis_strerror_r from the header, and call strerror_r directly.
This breaks any external users of this macro, but they shouldn't have
been using it in the first place.
Signed-off-by: Justin Brewer <jzb0012@auburn.edu>
strerror_r and addrinfo require _POSIX_C_SOURCE >= 200112L, which is
implied by _XOPEN_SOURCE >= 600. With the removal of AF_LOCAL usage,
the only non-standard features being used are the TCP_KEEP* socket
flags. _DARWIN_C_SOURCE is required to expose TCP_KEEPALIVE.
Fall back to using _XOPEN_SOURCE 600 for all platforms, and
additionally define _DARWIN_C_SOURCE for Darwin.
Signed-off-by: Justin Brewer <jzb0012@auburn.edu>
AF_LOCAL is the old, non-standardized name for AF_UNIX. Just use
AF_UNIX, rather than wrestling with platform specifics of AF_LOCAL
definitions.
Signed-off-by: Justin Brewer <jzb0012@auburn.edu>
It's possible for the call to connect() to succeed on the very first
try, in which case the logic for checking for invalid timeout fields is
never executed. When this happens, the tests fail because they expect a
REDIS_ERR_IO but no such failure has occurred.
Tests aside, this is a potential source of irritating and hard-to-find
intermittent bugs.
This patch forces the validation to occur early so that we get
predictable behavior whenever an invalid timeout is specified.
Not all code using hiredis can compile using '-std=c99', and/or not all users are able to easily make that change to the build process of various open-source projects, so it is more pragmatic to choose a different identifier that does not impose this requirement.
Originally implemented by @abedra as part of #306.
In case a write or read times out, we force an error state, because we
can't guarantuee that the next read will get the right data.
Instead we need to reconnect to have a clean-state connection, which is
now easily possible with this method.
The strerror_r API has two flavors depending on system options.
The bad flavor uses a static buffer for returning results, so if
you save the pointer from strerror_r, the string you're referencing
becomes useless if anybody else calls strerror_r again
The good flavor does what you expect: it writes the error to your buffer.
This commit uses strerror_r directly if it's a good version or copies
the static buffer into our private buffer if it's a bad version.
Thanks to gemorin for explaining the problem and drafting a fix.
Fixes#239
Signed-off-by: Chris Lamb <chris@chris-lamb.co.uk>
[Instead of checking for "not solaris" we feature detect
for availability of what we want, then remove the system
that advertises compatability but doesn't actually provide it
(given our assumptions about what we're guarding).]
Closes#254
