Released version 2.5-dev11 with the following main changes :
- DEV: coccinelle: Add strcmp.cocci
- CLEANUP: Apply strcmp.cocci
- CI: Add `permissions` to GitHub Actions
- CI: Clean up formatting in GitHub Action definitions
- MINOR: add ::1 to predefined LOCALHOST acl
- CLEANUP: assorted typo fixes in the code and comments
- CLEANUP: Consistently `unsigned int` for bitfields
- MEDIUM: resolvers: lower-case labels when converting from/to DNS names
- MEDIUM: resolvers: replace bogus resolv_hostname_cmp() with memcmp()
- MINOR: jwt: Empty the certificate tree during deinit
- MINOR: jwt: jwt_verify returns negative values in case of error
- MINOR: jwt: Do not rely on enum order anymore
- BUG/MEDIUM: stream: Keep FLT_END analyzers if a stream detects a channel error
- MINOR: httpclient/cli: access should be only done from expert mode
- DOC: management: doc about the CLI httpclient
- BUG/MEDIUM: tcpcheck: Properly catch early HTTP parsing errors
- BUG/MAJOR: dns: tcp session can remain attached to a list after a free
- BUG/MAJOR: dns: attempt to lock globaly for msg waiter list instead of use barrier
- CLEANUP: dns: always detach the appctx from the dns session on release
- DEBUG: dns: add a few more BUG_ON at sensitive places
- BUG/MAJOR: resolvers: add other missing references during resolution removal
- CLEANUP: resolvers: do not export resolv_purge_resolution_answer_records()
- BUILD: resolvers: avoid a possible warning on null-deref
- BUG/MEDIUM: resolvers: always check a valid item in query_list
- CLEANUP: always initialize the answer_list
- CLEANUP: resolvers: simplify resolv_link_resolution() regarding requesters
- CLEANUP: resolvers: replace all LIST_DELETE with LIST_DEL_INIT
- MEDIUM: resolvers: use a kill list to preserve the list consistency
- MEDIUM: resolvers: remove the last occurrences of the "safe" argument
- BUG/MEDIUM: checks: fix the starting thread for external checks
- MEDIUM: resolvers: replace the answer_list with a (flat) tree
- MEDIUM: resolvers: hash the records before inserting them into the tree
- BUG/MAJOR: buf: fix varint API post- vs pre- increment
- OPTIM: resolvers: move the eb32 node before the data in the answer_item
- MINOR: list: add new macro LIST_INLIST_ATOMIC()
- OPTIM: dns: use an atomic check for the list membership
- BUG/MINOR: task: do not set TASK_F_USR1 for no reason
- BUG/MINOR: mux-h2: do not prevent from sending a final GOAWAY frame
- MINOR: connection: add a new CO_FL_WANT_DRAIN flag to force drain on close
- MINOR: mux-h2: perform a full cycle shutdown+drain on close
- CLEANUP: resolvers: get rid of single-iteration loop in resolv_get_ip_from_response()
- MINOR: quic: Increase the size of handshake RX UDP datagrams
- BUG/MEDIUM: lua: fix memory leaks with realloc() on non-glibc systems
- MINOR: memprof: report the delta between alloc and free on realloc()
- MINOR: memprof: add one pointer size to the size of allocations
- BUILD: fix compilation on NetBSD
- MINOR: backend: add traces for idle connections reuse
- BUG/MINOR: backend: fix improper insert in avail tree for always reuse
- MINOR: backend: improve perf with tcp proxies skipping idle conns
- MINOR: connection: remove unneeded memset 0 for idle conns
Remove the zeroing of an idle connection node on remove from a tree.
This is not needed and should improve slightly the performance of idle
connection usage. Besides, it breaks the memory poisoning feature.
Skip the hash connection calcul when reuse must not be used in
connect_server() : this is the case for TCP proxies. This should result
in slightly better performance when using this use-case.
In connect_server(), if http-reuse always is set, the backend connection
is inserted into the available tree as soon as created. However, the
hash connection field is only set later at the end of the function.
This seems to have no impact as the hash connection field is always
position before a lookup. However, this is not a proper usage of ebmb
API. Fix this by setting the hash connection field before the insertion
into the avail tree.
This must be backported up to 2.4.
Add traces in connect_server() to debug idle connection reuse. These
are attached to stream trace module, as it's already in use in
backend.c with the macro TRACE_SOURCE.
Use include file <sys/time.h> to fix compilation error with timeval in
some files. This is as reported as 'man 7 system_data_types'. The build
error is reported on NetBSD 9.2.
This should be backported up to 2.2.
The current model causes an issue when trying to spot memory leaks,
because malloc(0) or realloc(0) do not count as allocations since we only
account for the application-usable size. This is the problem that made
issue #1406 not to appear as a leak.
What we're doing now is to account for one extra pointer (the one that
memory allocators usually place before the returned area), so that a
malloc(0) will properly account for 4 or 8 bytes. We don't need something
exact, we just need something non-zero so that a realloc(X) followed by a
realloc(0) without a free() gives a small non-zero result.
It was verified that the results are stable including in the presence
of lots of malloc/realloc/free as happens when stressing Lua.
It would make sense to backport this to 2.4 as it helps in bug reports.
realloc() calls are painful to analyse because they have two non-zero
columns and trying to spot a leaking one requires a bit of scripting.
Let's simply append the delta at the end of the line when alloc and
free are non-nul.
It would be useful to backport this to 2.4 to help with bug reports.
In issue #1406, Lev Petrushchak reported a nasty memory leak on Alpine
since haproxy 2.4 when using Lua, that memory profiling didn't detect.
After inspecting the code and Lua's code, it appeared that Lua's default
allocator does an explicit free() on size zero, while since 2.4 commit
d36c7fa5e ("MINOR: lua: simplify hlua_alloc() to only rely on realloc()"),
haproxy only calls realloc(ptr,0) that performs a free() on glibc but not
on other systems as it's not required by POSIX...
This patch reinstalls the explicit test for nsize==0 to call free().
Thanks to Lev for the very documented report, and to Tim for the links
to a musl thread on the same subject that confirms the diagnostic.
This must be backported to 2.4.
Some browsers may send Initial packets with sizes greater than 1252 bytes
(QUIC_INITIAL_IPV4_MTU). Let us increase this size limit up to 2048 bytes.
Also use this size for "max_udp_payload_size" transport parameter to limit
the size of the datagrams we want to receive.
In issue 1424 Coverity reports that the loop increment is unreachable,
which is true, the list_for_each_entry() was replaced with a for loop,
but it was already not needed and was instead used as a convenient
construct for a single iteration lookup. Let's get rid of all this
now and replace the loop with an "if" statement.
While in H1 we can usually close quickly, in H2 a client might be sending
window updates or anything while we're sending a GOAWAY and the pending
data in the socket buffers at the moment the close() is performed on the
socket results in the output data being lost and an RST being emitted.
One example where this happens easily is with h2spec, which randomly
reports connection resets when waiting for a GOAWAY while haproxy sends
it, as seen in issue #1422. With h2spec it's not window updates that are
causing this but the fact that h2spec has to upload the payload that
comes with invalid frames to accommodate various implementations, and
does that in two different segments. When haproxy aborts on the invalid
frame header, the payload was not yet received and causes an RST to
be sent.
Here we're dealing with this two ways:
- we perform a shutdown(WR) on the connection to forcefully push pending
data on a front connection after the xprt is shut and closed ;
- we drain pending data
- then we close
This totally solves the issue with h2spec, and the extra cost is very
low, especially if we consider that H2 connections are not set up and
torn down often. This issue was never observed with regular clients,
most likely because this pattern does not happen in regular traffic.
After more testing it could make sense to backport this, at least to
avoid reporting errors on h2spec tests.
Sometimes we'd like to do our best to drain pending data before closing
in order to save the peer from risking to receive an RST on close.
This adds a new connection flag CO_FL_WANT_DRAIN that is used to
trigger a call to conn_ctrl_drain() from conn_ctrl_close(), and the
sock_drain() function ignores fd_recv_ready() if this flag is set,
in order to catch latest data. It's not used for now.
Some checks were added by commit 9a3d3fcb5 ("BUG/MAJOR: mux-h2: Don't try
to send data if we know it is no longer possible") to make sure we don't
loop forever trying to send data that cannot leave. But one of the
conditions there is not correct, the one relying on H2_CS_ERROR2. Indeed,
this state indicates that the error code was serialized into the mux
buffer, and since the test is placed before trying to send the data to
the socket, if the connection states only contains a GOAWAY frame, it
may refrain from sending and may close without sending anything. It's
not dramatic, as GOAWAY reports connection errors in situations where
delivery is not even certain, but it's cleaner to make sure the error
is properly sent, and it avoids upsetting h2spec, as seen in github
issue #1422.
Given that the patch above was backported as far as 1.8, this patch will
also have to be backported that far.
Thanks to Ilya for reporting this one.
This applicationn specific flag was added in 2.4-dev by commit 6fa8bcdc7
("MINOR: task: add an application specific flag to the state: TASK_F_USR1")
to help preserve a the idle connections status across wakeup calls. While
the code to do this was OK for tasklets, it was wrong for tasks, as in an
effort not to lose it when setting the RUNNING flag (that tasklets don't
have), it ended up being inconditionally set. It just happens that for now
no regular tasks use it, only tasklets.
This fix makes sure we always atomically perform (state & flags | running)
there, using a CAS. It also does it for tasklets because it was possible
to lose some such flags if set by another thread, even though this should
not happen with current code. In order to make the code more readable (and
avoid the previous mistake of repeated flags in the bit field), a new
TASK_PERSISTENT aggregate was declared in task.h for this.
In practice the CAS is cheap here because task states are stable or
convergent so the loop will almost never be taken.
This should be backported to 2.4.
The crash that was fixed by commit 7045590d8 ("BUG/MAJOR: dns: attempt
to lock globaly for msg waiter list instead of use barrier") was now
completely analysed and confirmed to be partially a result of the
debugging code added to LIST_INLIST(), which was looking at both
pointers and their reciprocals, and that, if used in a concurrent
context, could perfectly return false if a neighbor was being added or
removed while the current one didn't change, allowing the LIST_APPEND
to fail.
As the LIST API was not designed to be used in a concurrent context,
we should not rely on LIST_INLIST() but on the newly introduced
LIST_INLIST_ATOMIC().
This patch simply reverts the commit above to switch to the new test,
saving a lock during potentially long operations. It was verified that
the check doesn't fail anymore.
It is unsure what the performance impact of the fix above could be in
some contexts. If any performance regression is observed, it could make
sense to backport this patch, along with the previous commit introducing
the LIST_INLIST_ATOMIC() macro.
This macro is similar to LIST_INLIST() except that it is guaranteed to
perform the test atomically, so that even if LIST_INLIST() is intrumented
with debugging code to perform extra consistency checks, it will not fail
when used in the context of barriers and atomic ops.
perf top shows that we spend a lot of time trying to read item->type in
the lookup loop, because the node is placed after the very long name,
so when the node is found, no data is in the cache yet. Let's simply
move the node upper in the struct. This results in the CPU usage of
resolv_validate_dns_response() to drop by 4 points.
A bogus test in b_get_varint(), b_put_varint(), b_peek_varint() shifts
the end of the buffer by one byte. Since the bug is the same in the read
and write functions, the buffer contents remain compatible, which explains
why this bug was not detected earlier. But if the buffer ends on an
aligned address or page, it can result in a one-byte overflow which will
typically cause a crash or an inconsistent behavior.
This API is only used by rings (e.g. for traces and boot messages) and
by DNS responses, so the probability to hit it is extremely low, but a
crash on boot was observed.
This must be backported to 2.2.
We're using an XXH32() on the record to insert it into or look it up from
the tree. This way we don't change the rest of the code, the comparisons
are still made on all fields and the next node is visited on mismatch. This
also allows to continue to use roundrobin between identical nodes.
Just doing this is sufficient to see the CPU usage go down from ~60-70% to
4% at ~2k DNS requests per second for farm with 300 servers. A larger
config with 12 backends of 2000 servers each shows ~8-9% CPU for 6-10000
DNS requests per second.
It would probably be possible to go further with multiple levels of indexing
but it's not worth it, and it's important to remember that tree nodes take
space (the struct answer_list went back from 576 to 600 bytes).
With SRV records, a huge amount of time is spent looking for records
by walking long lists. It is possible to reduce this by indexing values
in trees instead. However the whole code relies a lot on the list
ordering, and even implements some round-robin on it to distribute IP
addresses to servers.
This patch starts carefully by replacing the list with a an eb32 tree
that is still used like a list, with a constant key 0. Since ebtrees
preserve insertion order for duplicates, the tree walk visits the nodes
in the exact same order it did with the lists. This allows to implement
the required infrastructure without changing the behavior.
When cleaning up the code to remove most explicit task masks in commit
beeabf531 ("MINOR: task: provide 3 task_new_* wrappers to simplify the
API"), a mistake was done with the external checks where the call does
task_new_on(1) instead of task_new_on(0) due to the confusion with the
previous mask 1.
No backport is needed as that's only 2.5-dev.
This one was used to indicate whether the callee had to follow particularly
safe code path when removing resolutions. Since the code now uses a kill
list, this is not needed anymore.
When scanning resolution.curr it's possible to try to free some
resolutions which will themselves result in freeing other ones. If
one of these other ones is exactly the next one in the list, the list
walk visits deleted nodes and causes memory corruption, double-frees
and so on. The approach taken using the "safe" argument to some
functions seems to work but it's extremely brittle as it is required
to carefully check all call paths from process_ressolvers() and pass
the argument to 1 there to refrain from deleting entries, so the bug
is very likely to come back after some tiny changes to this code.
A variant was tried, checking at various places that the current task
corresponds to process_resolvers() but this is also quite brittle even
though a bit less.
This patch uses another approach which consists in carefully unlinking
elements from the list and deferring their removal by placing it in a
kill list instead of deleting them synchronously. The real benefit here
is that the complexity only has to be placed where the complications
are.
A thread-local list is fed with elements to be deleted before scanning
the resolutions, and it's flushed at the end by picking the first one
until the list is empty. This way we never dereference the next element
and do not care about its presence or not in the list. One function,
resolv_unlink_resolution(), is exported and used outside, so it had to
be modified to use this list as well. Internal code has to use
_resolv_unlink_resolution() instead.
The code as it is uses crossed lists between many elements, and at
many places the code relies on list iterators or emptiness checks,
which does not work with only LIST_DELETE. Further, it is quite
difficult to place debugging code and checks in the current situation,
and gdb is helpless.
This code replaces all LIST_DELETE calls with LIST_DEL_INIT so that
it becomes possible to trust the lists.
This function allocates requesters by hand for each and every type. This
is complex and error-prone, and it doesn't even initialize the list part,
leaving dangling pointers that complicate debugging.
This patch introduces a new function resolv_get_requester() that either
returns the current pointer if valid or tries to allocate a new one and
links it to its destination. Then it makes use of it in the function
above to clean it up quite a bit. This allows to remove complicated but
unneeded tests.
Similar to the previous patch, the answer's list was only initialized the
first time it was added to a list, leading to bogus outdated pointer to
appear when debugging code is added around it to watch it. Let's make
sure it's always initialized upon allocation.
The query_list is physically stored in the struct resolution itself,
so we have a list that contains a list to items stored in itself (and
there is a single item). But the list is first initialized in
resolv_validate_dns_response(), while it's scanned in
resolv_process_responses() later after calling the former. First,
this results in crashes as soon as the code is instrumented a little
bit for debugging, as elements from a previous incarnation can appear.
But in addition to this, the presence of an element is checked by
verifying that the return of LIST_NEXT() is not NULL, while it may
never be NULL even for an empty list, resulting in bugs or crashes
if the number of responses does not match the list's contents. This
is easily triggered by testing for the list non-emptiness outside of
the function.
Let's make sure the list is always correct, i.e. it's initialized to
an empty list when the structure is allocated, elements are checked by
first verifying the list is not empty, they are deleted once checked,
and in any case at end so that there are no dangling pointers.
This should be backported, but only as long as the patch fits without
modifications, as adaptations can be risky there given that bugs tend
to hide each other.
Depending on the code that precedes the loop, gcc may emit this warning:
src/resolvers.c: In function 'resolv_process_responses':
src/resolvers.c:1009:11: warning: potential null pointer dereference [-Wnull-dereference]
1009 | if (query->type != DNS_RTYPE_SRV && flags & DNS_FLAG_TRUNCATED) {
| ~~~~~^~~~~~
However after carefully checking, r_res->header.qdcount it exclusively 1
when reaching this place, which forces the for() loop to enter for at
least one iteration, and <query> to be set. Thus there's no code path
leading to a null deref. It's possibly just because the assignment is
too far and the compiler cannot figure that the condition is always OK.
Let's just mark it to please the compiler.
This code is dangerous enough that we certainly don't want external code
to ever approach it, let's not export unnecessary functions like this one.
It was made static and a comment was added about its purpose.
There is a fundamental design bug in the resolvers code which is that
a list of active resolutions is being walked to try to delete outdated
entries, and that the code responsible for removing them also removes
other elements, including the next one which will be visited by the
list iterator. This randomly causes a use-after-free condition leading
to crashes, infinite loops and various other issues such as random memory
corruption.
A first fix for the memory fix for this was brought by commit 0efc0993e
("BUG/MEDIUM: resolvers: Don't release resolution from a requester
callbacks"). While preparing for more fixes, some code was factored by
commit 11c6c3965 ("MINOR: resolvers: Clean server in a dedicated function
when removing a SRV item"), which inadvertently passed "0" as the "safe"
argument all the time, missing one case of removal protection, instead
of always using "safe". This patch reintroduces the correct argument.
This must be backported with all fixes above.
Cc: Christopher Faulet <cfaulet@haproxy.com>
A few places have been caught triggering late bugs recently, always cases
of use-after-free because a freed element was still found in one of the
lists. This patch adds a few checks for such elements in dns_session_free()
before the final pool_free() and dns_session_io_handler() before adding
elements to lists to make sure they remain consistent. They do not trigger
anymore now.
When dns_session_release() calls dns_session_free(), it was shown that
it might still be attached there:
Program terminated with signal SIGSEGV, Segmentation fault.
#0 0x00000000006437d7 in dns_session_free (ds=0x7f895439e810) at src/dns.c:768
768 BUG_ON(!LIST_ISEMPTY(&ds->ring.waiters));
[Current thread is 1 (Thread 0x7f895bbe2700 (LWP 31792))]
(gdb) bt
#0 0x00000000006437d7 in dns_session_free (ds=0x7f895439e810) at src/dns.c:768
#1 0x0000000000643ab8 in dns_session_release (appctx=0x7f89545a4ff0) at src/dns.c:805
#2 0x000000000062e35a in si_applet_release (si=0x7f89545a5550) at include/haproxy/stream_interface.h:236
#3 0x000000000063150f in stream_int_shutw_applet (si=0x7f89545a5550) at src/stream_interface.c:1697
#4 0x0000000000640ab8 in si_shutw (si=0x7f89545a5550) at include/haproxy/stream_interface.h:437
#5 0x0000000000643103 in dns_session_io_handler (appctx=0x7f89545a4ff0) at src/dns.c:725
#6 0x00000000006d776f in task_run_applet (t=0x7f89545a5100, context=0x7f89545a4ff0, state=81924) at src/applet.c:90
#7 0x000000000068b82b in run_tasks_from_lists (budgets=0x7f895bbbf5c0) at src/task.c:611
#8 0x000000000068c258 in process_runnable_tasks () at src/task.c:850
#9 0x0000000000621e61 in run_poll_loop () at src/haproxy.c:2636
#10 0x0000000000622328 in run_thread_poll_loop (data=0x8d7440 <ha_thread_info+64>) at src/haproxy.c:2807
#11 0x00007f895c54a06b in start_thread () from /lib64/libpthread.so.0
#12 0x00007f895bf3772f in clone () from /lib64/libc.so.6
(gdb) p &ds->ring.waiters
$1 = (struct list *) 0x7f895439e8a8
(gdb) p ds->ring.waiters
$2 = {
n = 0x7f89545a5078,
p = 0x7f89545a5078
}
(gdb) p ds->ring.waiters->n
$3 = (struct list *) 0x7f89545a5078
(gdb) p *ds->ring.waiters->n
$4 = {
n = 0x7f895439e8a8,
p = 0x7f895439e8a8
}
Let's always detach it before freeing so that it remains possible to
check the dns_session's ring before releasing it, and possibly catch
bugs.
The barrier is insufficient here to protect the waiters list as we can
definitely catch situations where ds->waiter shows an inconsistency
whereby the element is not attached when entering the "if" block and
is already attached when attaching it later.
This patch uses a larger lock to maintain consistency. Without it the
code would crash in 30-180 minutes under heavy stress, always showing
the same problem (ds->waiter->n->p != &ds->waiter). Now it seems to
always resist, suggesting that this was indeed the problem.
This will have to be backported to 2.4.
Using tcp, after a session release and free, the session can remain
attached to the list of sessions with a response message waiting for
a commit (ds->waiter). This results to a use after free of this
session.
Also, on some error path and after free, a session could remain attached
to the lists of available idle/free sessions (ds->list).
This patch ensure to remove the session from those external lists
before a free.
This patch should be backported to all version including
the dns over tcp (2.4)
When an HTTP response is parsed, early parsing errors are not properly
handled. When this error is reported by the multiplexer, nothing is copied
into the input buffer. The HTX message remains empty but the
HTX_FL_PARSING_ERROR flag is set. In addition CS_FL_EOI is set on the
conn-stream. This last flag must be handled to prevent subscription for
receive events. Otherwise, in the best case, a L7 timeout error is
reported. But a transient loop is also possible if a shutdown is received
because the multiplexer notifies the check of the event while the check
never handles it and waits for more data.
Now, if CS_FL_EOI flag is set on the conn-stream, expect rules are
evaluated. Any error must be handled there.
Thanks to @kazeburo for his valuable report.
This patch should fix the issue #1420. It must be backported at least to
2.4. On 2.3 and 2.2, there is no loop but the wrong error is reported (empty
response instead of invalid one). Thus it may also be backported as far as
2.2.
If a channel error (READ_ERRO|READ_TIMEOUT|WRITE_ERROR|WRITE_TIMEOUT) is
detected by the stream, in process_stream(), FLT_END analyers must be
preserved. It is important to be sure to ends filter analysis and be able to
release the stream.
First, filters may release some ressources when FLT_END analyzers are
called. Then, the CF_FL_ANALYZE flag is used to sync end of analysis for the
request and the response. If FLT_END analyzer is ignored on a channel, this
may block the other side and freeze the stream.
This patch must be backported to all stable versions
Replace the test based on the enum value of the algorithm by an explicit
switch statement in case someone reorders it for some reason (while
still managing not to break the regtest).
In order for all the error return values to be distributed on the same
side (instead of surrounding the success error code), the return values
for errors other than a simple verification failure are switched to
negative values. This way the result of the jwt_verify converter can be
compared strictly to 1 as well relative to 0 (any <= 0 return value is
an error).
The documentation was also modified to discourage conversion of the
return value into a boolean (which would definitely not work).
resolv_hostname_cmp() is bogus, it is applied on labels and not plain
names, but doesn't make any distinction between length prefixes and
characters, so it compares the labels lengths via tolower() as well.
The only reason for which it doesn't break is because labels cannot
be larger than 63 bytes, and that none of the common encoding systems
have upper case letters in the lower 63 bytes, that could be turned
into a different value via tolower().
Now that all labels are stored in lower case, we don't need to burn
CPU cycles in tolower() at run time and can use memcmp() instead of
resolv_hostname_cmp(). This results in a ~22% lower CPU usage on large
farms using SRV records:
before:
18.33% haproxy [.] resolv_validate_dns_response
10.58% haproxy [.] process_resolvers
10.28% haproxy [.] resolv_hostname_cmp
7.50% libc-2.30.so [.] tolower
46.69% total
after:
24.73% haproxy [.] resolv_validate_dns_response
7.78% libc-2.30.so [.] __memcmp_avx2_movbe
3.65% haproxy [.] process_resolvers
36.16% total
The whole code relies on performing case-insensitive comparison on
lookups, which is extremely inefficient.
Let's make sure that all labels to be looked up or sent are first
converted to lower case. Doing so is also the opportunity to eliminate
an inefficient memcpy() in resolv_dn_label_to_str() that essentially
runs over a few unaligned bytes at once. As a side note, that call
was dangerous because it relied on a sign-extended size taken from a
string that had to be sanitized first.
This is tagged medium because while this is 100% safe, it may cause
visible changes on the wire at the packet level and trigger bugs in
test programs.
