The STG_REGISTER init level is used to register known keywords and
protocol stacks. It must be called earlier because some of the init
code already relies on it to be known. For example, "haproxy -vv"
for now is constrained to start very late only because of this.
This patch moves it between STG_LOCK and STG_ALLOC, which is fine as
it's used for static registration.
Now -dM will set POOL_DBG_POISON for consistency with the rest of the
pool debugging options. As such now we only check for the new flag,
which allows the default value to be preset.
This option used to allow to store a marker at the end of the area, which
was used as a canary and detection against wrong freeing while the object
is used, and as a pointer to the last pool_free() caller when back in cache.
Now that we can compute the offsets at runtime, let's check it at run time
and continue the code simplification.
This option used to allow to store a pointer to the caller of the last
pool_alloc() or pool_free() at the end of the area. Now that we can
compute the offsets at runtime, let's check it at run time and continue
the code simplification. In __pool_alloc() we now always calculate the
return address (which is quite cheap), and the POOL_DEBUG_TRACE_CALLER()
calls are conditionned on the status of debugging option.
This macro is build-time dependent and is almost unused, yet where it
cannot easily be avoided. Now that we store the distinction between
pool->size and pool->alloc_sz, we don't need to maintain it and we
can instead compute it on the fly when creating a pool. This is what
this patch does. The variables are for now pretty static, but this is
sufficient to kill the macro and will allow to set them more dynamically.
The allocated size is the visible size plus the extra storage. Since
for now we can store up to two extra elements (mark and tracer), it's
convenient because now we know that the mark is always stored at
->size, and the tracer is always before ->alloc_sz.
Like previous patches, this replaces the build-time code paths that were
conditionned by CONFIG_HAP_POOLS with runtime paths conditionned by
!POOL_DBG_NO_CACHE. One trivial test had to be added in the hot path in
__pool_alloc() to refrain from calling pool_get_from_cache(), and another
one in __pool_free() to avoid calling pool_put_to_cache().
All cache-specific functions were instrumented with a BUG_ON() to make
sure we never call them with cache disabled. Additionally the cache[]
array was not initialized (remains NULL) so that we can later drop it
if not needed. It's particularly huge and should be turned to dynamic
with a pointer to a per-thread area where all the objects are located.
This will solve the memory usage issue and will improve locality, or
even help better deal with NUMA machines once each thread uses its own
arena.
There were very few functions left that were specific to global pools,
and even the checks they used to participate to are not directly on the
most critical path so they can suffer an extra "if".
What's done now is that pool_releasable() always returns 0 when global
pools are disabled (like the one before) so that pool_evict_last_items()
never tries to place evicted objects there. As such there will never be
any object in the free list. However pool_refill_local_from_shared() is
bypassed when global pools are disabled so that we even avoid the atomic
loads from this function.
The default global setting is still adjusted based on the original
CONFIG_NO_GLOBAL_POOLS that is set depending on threads and the allocator.
The global executable only grew by 1.1kB by keeping this code enabled,
and the code is simplified and will later support runtime options.
The test to decide whether or not to enforce integrity checks on cached
objects is now enabled at runtime and conditionned by this new debugging
flag. While previously it was not a concern to inflate the code size by
keeping the two functions static, they were moved to pool.c to limit the
impact. In pool_get_from_cache(), the fast code path remains fast by
having both flags tested at once to open a slower branch when either
POOL_DBG_COLD_FIRST or POOL_DBG_INTEGRITY are set.
When enabling pools integrity checks, we usually prefer to allocate cold
objects first in order to maximize the time the objects spend in the
cache. In order to make this configurable at runtime, let's introduce
a new debugging flag to control this allocation order. It is currently
preset by the DEBUG_POOL_INTEGRITY build-time setting.
This test used to appear at a single location in create_pool() to
enable a check on the pool name or unconditionally merge similarly
sized pools.
This patch introduces POOL_DBG_DONT_MERGE and conditions the test on
this new runtime flag, that is preset according to the aforementioned
debugging option.
The fail-alloc test used to be enabled/disabled at build time using
the DEBUG_FAIL_ALLOC macro, but it happens that the cost of the test
is quite cheap and that it can be enabled as one of the pool_debugging
options.
This patch thus introduces the first POOL_DBG_FAIL_ALLOC option, whose
default value depends on DEBUG_FAIL_ALLOC. The mem_should_fail() function
is now always built, but it was made static since it's never used outside.
This read-mostly variable will be used at runtime to enable/disable
certain pool-debugging features and will be set by the command-line
parser. A future option -dP will take a number of debugging features
as arguments to configure this variable's contents.
The poisonning performed on pool_free() used to help a little bit with
use-after-free detection, but usually did more harm than good in that
it was never possible to perform post-mortem analysis on released
objects once poisonning was enabled on allocation. Now that there is
a dedicated DEBUG_POOL_INTEGRITY, let's get rid of this annoyance
which is not even documented in the management manual.
There's no point keeping the vars_init_head() call in init() when we
already have a vars_init() registered at the right time to do that,
and it complexifies the boot sequence, so let's move it there.
Let's not use a trash there anymore. The function is called at very
early boot (for "haproxy -vv"), and the need for a trash prevents the
arguments from being parsed earlier. Moreover, the function only uses
a FILE* on output with fprintf(), so there's not even any benefit in
using chunk_printf() on an intermediary variable, emitting the output
directly is both clearer and safer.
REGISTER is meant to only assemble static lists, not to initialize
code that may depend on some elements possibly initialized at this
level. For example the init code currently looks up transport protocols
such as XPRT_RAW and XPRT_SSL which ought to be themselves registered
from at REGISTER stage, and which currently work only because they're
still registered directly from a constructor. INIT is perfectly suited
for this level.
Add the ability to set a "server timeout" on the httpclient with either
the httpclient_set_timeout() API or the timeout argument in a request.
Issue #1470.
In process_stream(), we force the response buffer allocation before any
processing to be able to return an error message. It is important because,
when an error is triggered, the stream is immediately closed. Thus we cannot
wait for the response buffer allocation.
When the allocation fails, the stream analysis is stopped and the expiration
date of the stream's task is updated before exiting process_stream(). But if
the stream was woken up because of a connection or an analysis timeout, the
expiration date remains blocked in the past. This means the stream is woken
up in loop as long as the response buffer is not properly allocated.
Alone, this behavior is already a bug. But because the mechanism to handle
buffer allocation failures is totally broken since a while, this bug becomes
more problematic. Because, most of time, the watchdog will kill HAProxy in
this case because it will detect a spinning loop.
To fix it, at least temporarily, an allocation failure at this stage is now
reported as an error and the processing is aborted. It's not satisfying but
it is better than nothing. If the buffers allocation mechanism is
refactored, this part will be reviewed.
This patch must be backported, probably as far as 2.0. It may be perceived
as a regression, but the actual behavior is probably even worse. And
because it was not reported, it is probably not a common situation.
This one started to randomly fail on me again and I could figure the
problem. It mixes one checked server with one unchecked on in each
backend, and tries to make sure that each checked server receives
exactly one request. But that doesn't work and is entirely time-
dependent because if the check starts before the client, a pure
TCP check is sent to the server, which sees an aborted connection
and makes the whole check fail.
Here what is done is that we make sure that only the second server
and not the first one is checked. The traffic is delivered to all
first servers, and each HTTP server must always receive a valid HTTP
request. In parallel, checks must not fail as they're delivered to
dummy servers. The check doesn't fail anymore, even when started on
a single thread at nice +5 while 8 processes are fighting on the same
core to inject HTTP traffic at 25 Gbps, which used to systematically
make it fail previously.
Since it took more than one hour to fix the "expect" line for the stats
output, I did it using a small script that I pasted into the vtc file
in case it's needed later. The relevance of this test is questionable
once its complexity is factored in. Let's keep it as long as it works
without too much effort.
This function was renderred obsolete by commit a0b5831ee ("MEDIUM: pools:
centralize cache eviction in a common function") which replaced its last
call inside the loop with a single call out of the loop to pool_releasable()
as introduced by commit 91a8e28f9 ("MINOR: pool: add a function to estimate
how many may be released at once"). Let's remove it before it becomes wrong
and used again.
The mem_poison_byte, mem_fail_rate, using_default_allocator and the
pools list are all only set once at boot time and never changed later,
while they're heavily used at run time. Let's optimize their usage from
all threads by marking them read-mostly so that them reside in a shared
cache line.
The recent changes was not complete.
d1c76f24fd
MINOR: quic: do not modify offset node if quic_rx_strm_frm in tree
The frame length and data pointer should incremented after the data
copy. A BUG_ON statement has been added to detect an incorrect decrement
operaiton.
Some variables were only checked via BUG_ON macro. If compiling without
DEBUG_STRICT, this instruction is a noop. Fix this by using an explicit
condition + ABORT_NOW.
This should fix the github issue #1549.
qc_rx_strm_frm_cpy is unsafe because it updates the offset field of the
frame. This is not safe as the frame is inserted in the tree when
calling this function and offset serves as the key node.
To fix this, the API is modified so that qc_rx_strm_frm_cpy does not
update the frame parameter. The caller is responsible to update
offset/length in case of a partial copy.
The impact of this bug is not known. It can only happened with received
STREAM frames out-of-order. This might be triggered with large h3 POST
requests.
In si_cs_recv(), the mux must never set CS_FL_WANT_ROOM flag on the
conn-stream if the input buffer is empty and nothing was copied. It is
important because, there is nothing the app layer can do in this case to
make some room. If this happens, this will most probably lead to a ping-pong
loop between the mux and the stream.
With this BUG_ON(), it will be easier to spot such bugs.
If a parsing error is detected and the corresponding HTX flag is set
(HTX_FL_PARSING_ERROR), we must be sure to always report it to the app
layer. It is especially important when the error occurs during the response
parsing, on the server side. In this case, the RX buffer contains an empty
HTX message to carry the flag. And it remains in this state till the info is
reported to the app layer. This must be done otherwise, on the conn-stream,
the CS_FL_ERR_PENDING flag cannot be switched to CS_FL_ERROR and the
CS_FL_WANT_ROOM flag is always set when h2_rcv_buf() is called. The result
is a ping-pong loop between the mux and the stream.
Note that this patch fixes a bug. But it also reveals a design issue. The
error must not be reported at the HTX level. The error is already carried by
the conn-stream. There is no reason to duplicate it. In addition, it is
errorprone to have an empty HTX message only to report the error to the app
layer.
This patch should fix the issue #1561. It must be backported as far as 2.0
but the bug only affects HAProxy >= 2.4.
After sending some data, we try to wake the H1 stream to resume data
processing at the stream level, except if the output buffer is still
full. However we must also be sure the mux is not blocked because of an
allocation failure on this buffer. Otherwise, it may lead to a ping-pong
loop between the stream and the mux to send more data with an unallocated
output buffer.
Note there is a mechanism to queue buffers allocations when a failure
happens. However this mechanism is totally broken since the filters were
introducted in HAProxy 1.7. And it is worse now with the multiplexers. So
this patch fixes a possible loop needlessly consuming all the CPU. But
buffer allocation failures must remain pretty rare.
This patch must be backported as far as 2.0.
In htx_copy_msg(), if the destination buffer is empty, we perform a raw copy
of the message instead of a copy block per block. But we must be sure the
destianation buffer was really allocated. In other word, to perform a raw
copy, the HTX message must be empty _AND_ it must have some free space
available.
This function is only used to copy an HTTP reply (for instance, an error or
a redirect) in the buffer of the response channel. For now, we are sure the
buffer was allocated because it is a pre-requisite to call stream
analyzers. However, it may be a source of bug in future.
This patch may be backported as far as 2.3.
The qcc instance should be tested as it is implied by a previous test
that it may be NULL. In this case, qc_timeout_task can be stopped.
This should fix github issue #1559.
A bug was uncovered by commit fc5912914 ("MINOR: httpclient: Don't limit
data transfer to 1024 bytes"), it happens that callers of b_xfer() and
b_force_xfer() are expected to check for available room in the target
buffer. Previously it was unlikely to be full but now with full buffer-
sized transfers, it happens more often and in practice it is possible
to crash the process with the debug command "httpclient" on the CLI by
going beyond a the max buffer size. Other call places ought to be
rechecked by now and it might be time to rethink this API if it tends
to generalize.
This must be backported to 2.5.
The url2sa implementation is inconsitent when parsing an IPv4, indeed
url2sa() takes a <ulen> as a parameter where the call to url2ipv4() takes
a null terminated string. Which means url2ipv4 could try to read more
that it is supposed to.
This function is only used from a buffer so it never reach a unallocated
space. It can only cause an issue when used from the httpclient which
uses it with an ist.
This patch fixes the issue by copying everything in the trash and
null-terminated it.
Must be backported in all supported version.
When calling ssl_ocsp_response_print which is used to display an OCSP
response's details when calling the "show ssl ocsp-response" on the CLI,
we use the BIO_read function that copies an OpenSSL BIO into a trash.
The return value was not checked though, which could lead to some
crashes since BIO_read can return a negative value in case of error.
This patch should be backported to 2.5.
When calling the "show ssl ocsp-response" CLI command some OpenSSL
objects need to be created in order to get some information related to
the OCSP response and some of them were not freed.
It should be backported to 2.5.
The b_istput function called to append the last data block to the end of
an OCSP response's detailed output was not checked in
ssl_ocsp_response_print. The ssl_ocsp_response_print return value checks
were added as well since some of them were missing.
This error was raised by Coverity (CID 1469513).
This patch fixes GitHub issue #1541.
It can be backported to 2.5.
The 'dst' optionnal field on a httpclient request can be used to set an
alternative server address in the haproxy address format. Which means it
could be use with unix@, ipv6@ etc.
Should fix issue #1471.
When starting for the 2nd time a request from the same httpclient *hc
context, the flags are not reinitialized and the httpclient will stop
after the first call to the IO handler, because the END flag is always
present.
This patch also add a test before httpclient_start() to ensure we don't
start a client already started.
Must be backported in 2.5.
The idle connection delay calculation before a request is a bit tricky,
especially for multiplexed protocols. It changed between 2.3 and 2.4 by
the integration of the idle delay inside the session itself with these
commits:
dd78921c6 ("MINOR: logs: Use session idle duration when no stream is provided")
7a6c51324 ("MINOR: stream: Always get idle duration from the session")
and by then it was only set by the H1 mux. But over multiple changes, what
used to be a zero idle delay + a request delay for H2 became a bit odd, with
the idle time slipping into the request time measurement. The effect is that,
as reported in GH issue #1395, some H2 request times look huge.
This patch introduces the calculation of the session's idle time on the
H2 mux before creating the stream. This is made possible because the
stream_new() code immediately copies this value into the stream for use
at log time. Thus we don't care about changing something that will be
touched by every single request. The idle time is calculated as documented,
i.e. the delay from the previous request to the current one. This also
means that when a single stream is present on a connection, a part of
the server's response time may appear in the %Ti measurement, but this
reflects the reality since nothing would prevent the client from using
the connection to fetch more objects. In addition this shows how long
it takes a client to find references to objects in an HTML page and
start to fetch them.
A different approach could have consisted in counting from the last time
the connection was left without any request (i.e. really idle), but this
would at least require a documentation change and it's not certain this
would provide a more useful information.
Thanks to Bart Butler and Luke Seelenbinder for reporting enough elements
to diagnose this issue.
This should be backported to 2.4.
Sadly, despite particular care, commit 39a0a1e12 ("MEDIUM: h2/hpack: emit
a Dynamic Table Size Update after settings change") broke H2 when sending
DTSU. A missing negation on the flag caused the DTSU_EMITTED flag to be
lost and the DTSU to be sent again on the next stream, and possibly to
break flow control or a few other internal states.
This will have to be backported wherever the patch above was backported.
Thanks to Yves Lafon for notifying us with elements to reproduce the
issue!
tls_basic_sync_wo_stkt_backend fails once every 200 runs for me. This
seems to be because the startup delay doesn't always allow peers to
perform a simultaneous connect, close and new attempt. With 3s I can't
see it fail anymore. In addition the long "delay 0.2" are still way too
much since we do not really care about the startup order in practice.
Sometimes when sending commands to shut down a server, haproxy complains
that some connections remain, this is because the server-side connection
might not always be completely released at the moment the client leaves
and the operation is emitted. While shutting down server sessions work,
it seems cleaner to just use "option httpclose" which releases the server
earlier and avoids the race.
This can be backported to 2.5.
There's a bug in spoe_release_appctx() which checks the presence of items
in the wrong list rt[tid].agents to run over rt[tid].waiting_queue and
zero their spoe_appctx. The effect is that these contexts are not zeroed
and if spoe_stop_processing() is called, "sa->cur_fpa--" will be applied
to one of these recently freed contexts and will corrupt random memory
locations, as found at least in bugs #1494 and #1525.
This must be backported to all stable versions.
Many thanks to Christian Ruppert from Babiel for exchanging so many
useful traces over the last two months, testing debugging code and
helping set up a similar environment to reproduce it!