For a chunk size exactly equal to bufsize/2, the "max_size" variable could
underflow if a separator is provided. Indeed, "max_size" is first set to
(trash->size - 2 * chunk_size), so to 0. And on the first iteration, the
separator length is removed, making it to underflow.
On 3.3 and lower it is especially an issue with samples larger than
bufsize/2 because data could be written ouside of the buffer. On 3.4 and
3.5, it is not an issue because we fail to retrieve a chunk.
This should be backported to all supported versions.
The proxy used to parse healthcheck sections was not inserted in the proxies
list. So its initialization was not properly finalized. Among other things,
it was an issue for some arguments, like regular expressions.
This proxy is now inserted in the proxies list.
This patch should fix the issue #3440. It must be backported to 3.4.
When encountering an unexpected escape sequence, treat it as a literal
character instead of skipping it.
This is a deliberate choice for two reasons:
- to avoid a desynchronization between the h->data counter and the
buffer content, which would otherwise leave uninitialized memory
("garbage") in the destination buffer;
- to ensure that an invalid configuration string triggers a parsing
error immediately (fail-fast), rather than resulting in a silently
malformed configuration.
Thank to @dirkmueller for having reported this issue.
No need to backport. huf arrived with this current version.
When a stream is freed, hld_strm_task() was unconditionally re-queuing
the user task, even when no streams were left or when no update was
needed. This caused spurious task wakeups, which could incorrectly inflate
connection and request rate counters to come for the -R option implementation.
Only queue the user task if there are remaining streams to process, and
properly update the expiration time.
No need to backport. haload arrived with this current version.
Update the README to cover the new syncing feature: the deployment
section explains the three optional pieces and their relations. We
also explain the special cases of "Get updates" and "Save updates"
when combined with pending/conflicting changes (particularly in
edition mode).
An explanation of the difficulties to set git permission is also
provided.
Every save used to be committed as a bare "update 3.5", which makes the
storage history useless to navigate. The subject now names the branch
and the first commit whose review was touched, followed by "+ N more"
when several, and the body lists all of them one per line:
update 3.5: 6a7b27a0 + 2 more
6a7b27a0
d13aaf05
b12dd0b5
This is what to grep for when hand-editing the storage repository, for
example to locate the change to revert or rebase. A commit touched by
several directives of the same save (state plus notes) is only listed
once.
thttpd forwards the CGI's stderr to the client *ahead* of its stdout:
the socket receives the HTTP status line, then any stderr log line, and
only then the CGI headers, turning the log line into a bogus response
header; on an error path the same mechanism could push garbage in front
of the "Status:" header and corrupt the response entirely.
All diagnostics are already carried by the response itself (die()'s
body, the in-band "warning:" line with git's captured error), so the
duplicated stderr logging brings nothing and only risks breaking the
channel it leaks into: drop it, and state the constraint in a comment
above die() so it doesn't come back. The usage text for a bad command
line is folded into the 500 response body, which is also what a shell
user sees when testing by hand.
The "git commit failed" warning said nothing about the cause, leaving
the admin to guess between a missing identity, an ownership refusal, a
git binary absent from the restricted PATH the web server gives to its
CGIs, etc. run_git() now captures the command's stdout and stderr
through a pipe and the warning line carries git's own message (first
255 bytes, control chars flattened), so the admin directly sees the
cause; as the command runs through /bin/sh, an unfindable git yields
status 127 and the shell's message, reworded as "cannot execute git:
..." to directly point at the typical PATH issue. Capturing also
guarantees that git output can never corrupt the CGI response nor leak
to the client on servers which wire the CGI's stderr to the socket.
When the git commit fails after a save (typically a missing committer
identity in the storage repository, or an ownership/permission issue),
the failure was only logged to stderr, which lands in the web server's
error log at best: the file kept being updated but the history silently
stopped being recorded. Report it as a "warning: git commit failed ..."
line appended to the response, where the page can show it to the user,
on top of the stderr log. Also stop treating a no-op as a failure:
re-pushing identical content stages nothing, so the commit is now
simply skipped when "git diff --cached --quiet" reports no staged
change, instead of letting "git commit" fail on an empty commit.
Notes are append-only on the wire, which makes concurrent edits
conflict-free but leaves no way to revise or clean up a note from the
page: fixing a note requires a hand-edit of the storage file. This adds
the replacement directive that the design had reserved:
<cid> setnotes <hash> <replacement text>
Unlike the other directives, a replacement must carry a token of the
base it was computed from, or it could silently destroy a concurrent
update (one reviewer's append landing between another's read and
replace). The token is the SDBM hash (8 hex chars) of the note blob
the client based its edit on: the server only applies the replacement
if it still matches the stored blob (empty-string hash for a commit
without notes). On mismatch the directive is dropped, the line is left
exactly as found, and a "conflict <cid>" line is emitted in the
response before the resulting lines so that the client can point the
user at what needs manual reconciliation; other directives from the
same POST are still applied, and nothing is written nor committed when
everything conflicted. SDBM is trivially computed on both sides (a
concurrency token, not a security feature, and JS crypto is unavailable
on plain http anyway), and its small multiplier keeps the whole hash
computation exact in awk's double-precision arithmetic, which is
precisely why it was chosen over wider-multiplier hashes.
An empty replacement deletes the notes (and the line if no state is
left), finally allowing obsolete notes to be removed without editing
the file by hand. Replacements are capped to 4000 chars instead of the
500-char append cap, since a coalesced blob may legitimately have grown
beyond a single addition.
This adds the read side of the review persistence CGI: GET
update.cgi?branch=X.Y now returns the current overlay for that branch
as a JSON array of {"cid", "state", "notes"} objects with absent fields
omitted; a missing or empty file yields "[]". The raw storage format
never travels: the notes are unescaped by the parser and JSON-escaped
on output, so the client can JSON.parse() the response and insert the
notes via textContent directly. Unparseable lines or fields are
silently skipped as everywhere else.
Reads are lockless: the atomic rename on the write side guarantees
that the file is always a complete valid version. The response carries
Cache-Control: no-store so that a browser never reuses a stale overlay
on refresh.
The backport review page keeps the human edits (verdict overrides and
notes) only in the loaded DOM: they are lost on reload and never shared
between reviewers. This adds the server side of the shared persistence
design: update.awk, a GNU awk CGI script which stores these edits into
one file per major branch (e.g. "3.5") inside a dedicated git
repository, one line per touched commit:
<commit_id> [state <n|u|w|y>] [notes "<quoted notes>"]
The overlay only ever stores human edits keyed on the commit id; the AI
verdict and explanation stay in the generated HTML. Commit ids are
length-agnostic and matched by symmetric prefix (first match wins), so
the current 8-char pipeline and a future 12-char one both work without
any migration.
POST applies line-oriented directives ("<cid> state <n|u|w|y|revert>",
"<cid> notes <text to append>"), none of which carries a base value:
states are last-write-wins and notes are append-only (capped to 500
chars per push and sanitised so that no newline may ever enter a stored
line), which keeps concurrent edits conflict-free. Broken directives,
fields or lines are silently ignored, never fatal, and lines not being
modified are preserved byte-for-byte so that admin hand-edits survive.
Writers are serialised by a mkdir lock at an obvious place (<repo>/lock)
with PID-gated crash takeover via atomic rename, the file itself is
replaced by an atomic rename from a temp file inside the lock dir, and
every resulting state is committed to git, which acts as the event log
(git blame/log -L provide the full history). A crash at any point leaves
at worst a stale lock (reclaimed on the next write) or a valid but
uncommitted tree (folded into the next commit), never a broken file.
The takeover races are covered: the staleness decision and the takeover
rename are not one atomic operation, so the thief verifies after the
rename, discarding the stolen dir only if it still carries the pid that
was judged dead and renaming it back in place untouched otherwise; the
victim redoes the whole locked cycle from a fresh read when its final
rename fails, since nothing was applied to the branch file yet; a writer
finding its own pid in the lock adopts it as stale; and the release only
removes the lock after checking that it still contains our own pid.
A few awk specifics are worth noting: external commands (git, mkdir,
mv) go through /bin/sh, so everything interpolated into a command line
is shell-quoted (single-quote is escaped and the argument placed inside
single quotes); the -b (bytes) flag keeps all string operations byte-
based regardless of the locale; and a manual argv parsing due to gawk
silently consumes a leading "-r" argument, ignoring ours.
Also note that gawk uses a file cache for getline() and co, which opens
lots of traps so we need to be extremely careful about properly closing
files if we want to check for changes (e.g. lock's pid file).
Finally, writing through a redirection whose target cannot be opened is
a fatal awk error terminating the script without even a response, so the
writes into the (stealable) lock dir are arranged to resist a takeover:
the pid is written through the shell, where a vanished dir is a plain
command failure, and the temp file is opened the very instant the lock
is acquired, its descriptor surviving a later theft. The update.cgi
wrapper considers any >0 return code as a failure and returns a generic
error as it will indicate that the awk script itself couldn't produce
a valid response.
For now, only the POST ("save changes") action is implemented.
The "Get updates" and "Save changes" buttons only existed at the top
right of the page, while a review session ends at the bottom of the
table: with no button left in sight there, it was way too easy to
forget to save the work. Emit a copy of both buttons and of the status
line at the bottom right, sharing the same handlers; the status
message and the save-button graying are applied to both instances at
once so the two spots always tell the same story.
Some server setups leak the CGI's stderr into the response body (e.g.
inetd-style servers where fd 2 is the client socket): since stderr is
unbuffered and stdout is buffered, a git error message then lands
*before* the "OK <n>" line, and the applied-count check failed with a
cryptic "server applied only ? of N changes" although everything had
been applied. Scan the whole response for the lines of interest (the
"OK <n>" count, the "conflict" and "warning" lines) instead of assuming
they come first, dump the raw response to the console when no count is
found at all to ease diagnosis, and display the server's warning lines
(such as the new "git commit failed" one) next to the save status so
that a recording problem is visible from the page instead of being
buried in a server log.
The save handler treated any HTTP 200 as a full success and advanced
the local reference for everything it had sent, only special-casing the
reported conflicts. But the server legitimately drops directives it
cannot parse, and answers "OK <n> directives applied" with what it
really did. The typical case is an outdated update scripton the server
which ignores the whole "setnotes" directive, applies nothing, and
the client still displayed the edit as saved... until the next "Get
updates" reverted it.
Let's count the directives sent, and when the server's applied count
plus the reported conflicts don't add up, believe the server, not
ourselves: advance nothing, keep every edit local (boxes open, save
button lit) and tell the user how many changes were ignored,
suggesting a version mismatch.
The "Save changes" button used to remain active all the time, giving no
hint about whether anything was pending. It is now disabled whenever
nothing differs from the reference: no verdict change, no non-empty
note addition, no note edition differing from its base. It gets
re-evaluated after every action which may change that (verdict clicks,
typing in a note input, opening/cancelling a box, updates and saves),
bailing out at the first pending change so the common case stays cheap.
As a side effect, the button lighting up right after a reload confirms
at a glance that the browser restored unsaved local edits.
An "[edit note]" link now appears next to "[add note]" whenever a line
has shared notes: it presents the whole note blob in the input box for
edition, and the save sends it as a replacement (the "setnotes"
directive, carrying the hash of the blob the edit was based on).
Emptying the box deletes the note. Clicking "add note" first and then
"edit note" merges the reference notes with the pending addition so
nothing typed so far is lost, and a "[cancel]" link aborts an edition
opened by mistake without touching anything.
An open input only disappears once its content is synchronized with the
reference: a successful save, an update proving an exact match, or an
explicit cancel. To that end, "Get updates" first silently closes the
no-op boxes (opened but nothing changed), then after applying the
fetched state it closes the boxes it made redundant (an addition
someone already pushed, an edition matching the current notes), and
re-bases an edition whose base moved, turning it red: red uniformly
means "the reference changed under your edit, review the notes above
against your text before saving". A save refused by the server (the
"conflict <cid>" response lines) turns the input red the same way,
keeping it in edition; states and additions from the same save are
unaffected. After a reload, a browser-restored pending note reopens in
append mode, the only safe assumption since a lost replacement base
cannot be recovered.
This is the write side of the review syncing: a "Save changes" button
next to "Get updates" collects the local edits and pushes them to
update.cgi. An edit is a radio button change differing from the
reference state (so clicking around and coming back to the reference
sends nothing), or a note typed in the per-line input revealed by the
new "[add note]" link under the AI explanation (500 chars max, matching
the server-side cap).
No directive carries a base value: states are last-write-wins and notes
are append-only server-side, so two reviewers saving concurrently
cannot conflict. On success the reference advances to the pushed values
and the note inputs are cleared, so the page is clean without needing a
refetch; on error (server busy or unreachable) everything stays local
and a later click simply retries.
This adds the read side of the review syncing to the generated page: a
"Get updates" button at the top right retrieves the shared state from
update.cgi (reached by a bare relative URL, so it must be in a cgi-bin
directory next to the page) and applies it. Nothing is fetched
automatically, not even at load time: it's up to the user to explicitly
click to resynchronize, and without it (or with the server down) the
page keeps behaving fully standalone as today.
Three states exist per line to make this work. The original state is
the verdict the bot chose, captured at load time and constant. The
reference state is the last known shared state, on top of which the
user's edits sit; it starts equal to the original. The local state is
the DOM itself (the checked radios). Applying a fetched overlay
recomputes every line's reference as "the server's entry if any,
otherwise the bot's verdict", so a removed override properly falls back
to the original; the reference always advances but the displayed state
only moves where the user had no local edit: local edits win, and
re-applying the same overlay twice changes nothing.
The received commit ids are resolved exactly first, then by symmetric
prefix (one id being a prefix of the other, for mixed-length ids), first
line wins. The shared notes land in a dedicated container below the AI
explanation, rendered via innerText (no HTML injection) and replaced
wholesale so the operation stays idempotent; entries for commits absent
from the page are simply ignored.
The whole exchange was tested with the real generated page's scripts
running against a stubbed DOM covering load capture, prefix resolution,
adopt-vs-keep on both changed and disappeared entries, idempotent
re-application, and silent degradation on fetch failure.
The review page will need to exchange its state with the update.cgi
service sitting next to it, and for this it must know which branch it
covers since all branches' pages may share a directory. post-ai.sh
takes a new "-v <version>" argument and emits it as a "branch" JS
variable; when absent the variable is empty and the page will simply
not offer syncing, keeping the output standalone as today.
update-3.0.sh deduces the version from its own name (update-3.0.sh ->
3.0), so that adding a symlink with another version for a new branch
continues to work with no other change, and passes it to post-ai.sh.
The commit id column doesn't need to show more than 8 chars to stay
unambiguous within a single page, and longer ids needlessly widen the
table. Everything that is keyed on the id (the commit link, the row's
name= attribute and the cid[] JS array) keeps the full id produced by
the pipeline, whatever its length, so this is a display-only change
which also gets us ready for a possible future move of the pipeline to
longer ids.
When the page is reloaded, the browser restores the "review" radio
column to the user's last selection (e.g. "All"), but the "review" JS
variable is regenerated to the default first line to review: the
listing then restarts from that line while "All" still appears
selected, and one has to click a random line then "All" again to
really see everything.
Give an id to each line's review radio and resynchronize the variable
from the actually checked radio when the page loads: a restored "All"
(or any restored line) now behaves as selected, and on a fresh load the
checked radio is the generated one so nothing changes.
Several QUIC related keywords were removed in 3.4. The legacy options
were marked as deprecated and scheduled for removal in 3.5. This patch
applies this removal for the upcoming 3.5.
Commit 061754b249 attempted to make it
possible to close file descriptors belonging to other thread groups by
using thread isolation.
The problem is, closing other thread groups' fds usually happens when we
destroy a listener, in which case we hold the listener lock. If any
other thread tries to get that lock while we're waiting for the thread
isolation, then they will deadlock.
This can happen with any type of listener, but it is easier to reproduce
with a suspend/resume loop with ABNS sockets, as a suspend translates to
a close here.
To fix that, instead of using thread isolation, do something similar to
what's done when the fd belongs to our thread group. Increase the tgid
ref counter, so that we're sure nobody will close the fd while we're
dealing with it, then set the FD_MUST_CLOSE bit and set thread_mask to
0.
At this point, if no thread was running on that fd, no one will and we
can safely close it. So just call _fd_delete_orphan() if the
running_mask is 0 and if the FD_MUST_CLOSE bit is still there, otherwise
we can safely assume another thread will take care of it.
This should be backported up to 2.8.
For frontend connections, quic_conn layer is able to reject any new
streams opened after MUX closure. This is necessary as the peer may not
have been notified yet of the closure.
This operation is unnecessary on backend side. This is due to the fact
that only HTTP protocols are currently supported on top of QUIC, with
requests initiated by the client. For requests started before the MUX
closure, either they are already completed or closed early with a
STOP_SENDING emitted during stream shut.
Prior to this patch, spurrious RESET_STREAM could have been emitted on
backend connections after MUX closure as quic_conn stream_max_bidi was
not correctly set. Now reject is only performed for frontend connections
so this should not occured anymore.
This should be backported up to 3.3.
Prior to this patch, shut stream callback only handles write channel
closure. In case of an early closure, a RESET_STREAM would be emitted.
On the frontend side in most cases this is sufficient as read channel is
already closed, as HTTP/3 GET requests has been fully received. However,
this may not be the case for POST requests. Also, on the backend side,
haproxy acts a client. In this case, a stream early closure will
typically happen before receiving the full response. Nothing will be
emitted (RESET_STREAM is unnecessary as write channel is already
closed), thus the server peer will continue to emit.
To fix this situation, the current patch implement read channel closure
on shut if SE_SHR_RESET is set. Callback lclose from app_ops is called
with a new dedicated mode for read channel closure, which will result in
a STOP_SENDING frame generated by H3 and hq transcoders. This will
instruct the peer to stop emission.
This should be backported up to 3.3. Note that this depends on the
following patch :
dde3ee06c30f20091443bdafdda0e0294f7ac26b
MINOR: mux_quic: use separate error code for STOP_SENDING
Prior to this patch, a single error code was registrable at the QCS
level. This code was used both for RESET_STREAM and STOP_SENDING
emission. It was specified via qcc_reset_stream().
This patch extends the API so that now a dedicated error code is
implemented for STOP_SENDING as well. This may be necessary as both
frames can be sent in different context, with a diverging error code.
This patch is required to implement STOP_SENDING emission during shut
callback when read channel is closed.
When stream detach callback is called, the default behavior is to free
the associated QCS instance. However, QCS may be preserved in so-called
detached state if there is remaining data to sent.
This condition is checked via qcs_is_close_local() which ensures that
either FIN or a RESET_STREAM was emitted. However, this does not take
into account a scheduled STOP_SENDING emission, which can happen in case
of request abort for example.
Adjusts qcm_strm_detach() to also take into account STOP_SENDING
emission before freeing or keeping a detached QCS instance. As a
complement, QCS have to be purged after STOP_SENDING emission when
reaching completion.
On frontend side, this bug is probably only visible in case of HTTP/3
POST. When dealing with GET, FIN is most of the time received earlier,
which render STOP_SENDING unnecessary. This issue however has a bigger
impact on the backend side. In case of stream abort, for example on
timeout, the server may be left unnotified and will continue to emit
STREAM data despite QCS closure on haproxy client side.
Note that this fix also has a side effect on backend connection reuse.
Indeed it may increase the rate of QCS in detached state. This may
prevent an idle connection to be reinserted in the server pool, without
any possibility to reinsert it later. In the end this causes a lower
reuse rate. This is an issue which must be addressed in a dedicated
patch. For now, add a COUNT_IF_HOT() to report when such situation
occurs.
This should be backported to all stable releases, after a period of
observation. COUNT_IF_HOT() is unnecessary on 3.2 and below.
This target disables all possible features except poll(). It is meant to
serve as a base for small embedded setups, on top of which one may manually
enable select features. Even threads, traces/h2/fcgi/SPOE are disabled.
The default executable is roughly 80% smaller than with linux-glibc:
$ size haproxy-linux-glibc haproxy-tiny
text data bss dec hex filename
3660924 176964 9868784 13706672 d125b0 haproxy-linux-glibc
2537864 146512 84928 2769304 2a4198 haproxy-tiny
With SSL enabled, the difference shrinks a bit (-77%):
$ size haproxy-linux-glibc-ssl haproxy-tiny-ssl
text data bss dec hex filename
4163373 208788 9873904 14246065 d960b1 haproxy-linux-glibc-ssl
2950852 177732 90048 3218632 311cc8 haproxy-tiny-ssl
There's no point in building these ones anymore when traces are disabled,
nothing relies on them. This brings extra 28kB savings, resulting in 709kB
total savings when disabling traces.
The various trace sources always have the same pattern:
- trace events
- trace source
- trace decoding function
Dropping these when USE_TRACE=0 definitely makes sense. There are two
modes of definition here:
- those designed after mux_h2 which interleave #define and the entry
definition in the event. These ones cannot be removed without a
significant code move to split the #define and usage apart. Instead
here we mark the struct __maybe_unused, so that the compiler will
just not implement it.
- those designed like stream.c where defines are separated. Here we
can simply enclose the events definition inside the USE_TRACE guard
For most of these the static declaration of the trace function was moved
after the events so that the #if defined(USE_TRACE) could be placed between
the two. Nothing else was changed.
This saves another 51 kB of object code when USE_TRACE=0.
It requires essentially a few ifdefs and to add a dummy definition of
h3_trace_header() to completely disable traces in H3. This reduces the
object code by 35 kB.
The function takes a buffer in argument which is the target buffer. The
first calls properly use it but the subsequent ones, probably due to
reused/moved code, directly write into &trace_buf, thus ignoring the
buf argument. Fortunately all call places pass &trace_buf for buf, so
it currently has no impact but could possibly change.
No backport is needed, but it doesn't hurt to backport it if it helps.
By just moving a few definitions, creating two dummy inline functions and
a few ifdefs, we can get rid of the entire trace generation code in the
H2 mux and save ~96 kB. This is what this patch does. Even the trace_h2
struct is removed in this case.
When disabling traces, "conn" isn't used between ctx assignment and its
first usage, and as usual, gcc wrongly believes that a null check in a
shared function implies the checked argument may be NULL where it's used,
leading to this warning:
src/ssl_sock.c: In function 'ssl_sock_handshake.constprop':
src/ssl_sock.c:6049:7: warning: null pointer dereference [-Wnull-dereference]
Assigning ctx after the conn_ctrl_ready() check is sufficient to shut it
up, so let's do this. It should also result in slightly better code.
When traces are disabled, we used to make TRACE() and other macros just
emit a "do { } while (0)" statement, which has the unfortunate limitation
of explicitly marking the arguments as not used. As such, all variables
that are initialized in functions for the sole purpose of being passed
to the trace calls end up emitting warnings about "foo defined but not
used". It is difficult to keep these in a clean state all the time, and
to always think about adding __maybe_unused after each declaration, and
the traces try hard to be developer-friendly in order to gain in adoption.
Let's just remap all macros to __eat_all_args() which will mark all
arguments as used. No code is emitted, the output binary is the same
as with the while(0) stuff, but syntactically speaking the argument is
used and the compiler is happy.
It may be useful to backport this to 3.4 as it's already expected that
some future fixes will trigger build warnings there otherwise. This
commit requires these two ones:
CLEANUP: traces: get rid of a few rare empty args in TRACE calls
MINOR: compiler: add a macro to ignore all arguments
Regularly when disabling features (e.g. traces), some macros that would
make use of some arguments end up not consuming them at all, making the
compiler complain that "variable foo defined but not used".
An elegant way to generically mark arguments as used is to pass them to
a variadic function. However a first argument is needed. So we create a
macro that passes (0, __VA_ARGS__) to an inline function that does nothing
from its arguments, and that's done.
The TRACE macro allows to leave empty args and automatically turns them
into zeroes. However it also limits how we can remap the macro, because
functions do not accept this for example. There are very few places all
over the code where ',,' exists in TRACE calls, so let's explicitly add
the 0 there. It could even make some editors' syntax highlighting happier.
In quic_transport_params_store(), we call qc_early_transport_params_cpy()
if edata_accepted is set, which copies one by one all tx_params into the
locally allocated etps struct, and later after updates we call
qc_early_transport_params_validate() to check if they changed. It turns
out that when USE_TRACE is disabled, gcc 4 to 13 are confused and believe
that one or several of the fields compared in the later function might be
used uninitialized. A careful code inspection proves that this is not the
case. Setting them to zero in the _cpy() function makes the warning
disappear, it's really an issue related to variable propagation it seems,
which can explain why it doesn't happen with traces (code is a bit more
complex). Gcc-13 only emits a warning about a single field, and gcc-14
completely solved it. Playing with consts, __maybe_unused etc has no
effect.
One thing works however, it is to mark the _validate() function noinline.
In this case it is implemented normally and the compiler doesn't put its
nose into the propagation path and doesn't complain.
Such comments are always scary because one may seriously wonder whether
the compiler emits valid code when it says this...
It should be backported to 3.4 which experiences the same warning with
USE_TRACE=0.
A QUIC stream may be aborted to ignore future data read. This also
prepares a STOP_SENDING frame to instruct the peer to close its write
channel.
This capability is exposed via qcc_abort_stream_read() which should be
guarded against multiple invokation for a single stream. This was
checked via QC_SF_TO_STOP_SENDING flag. However, this flag is resetted
once STOP_SENDING frame is emitted. Thus in theory it could be possible
to emit several STOP_SENDING for a single stream.
This patch improves this by using QC_SF_READ_ABORTED flag check. This
flag is set during qcc_abort_stream_read() and never removed even after
STOP_SENDING frame is emitted.
This bug was never encountered in a real situation. However, this patch
is necessary to definitely guarantee that it cannot occur.
This should be backported up to 2.8.
On success, h3_rcv_buf() returns the number of parsed STREAM bytes which
are removed by the caller afterwards. A success value is mandatory so
that the underlying QUIC packet is acknowledged.
When H3 parser detects an error during HEADERS or DATA parsing, the
stream or the connetcion is flagged for closure. If there is remaining
frames, they are simply ignored and h3_rcv_buf() returns the remaining
input buffer size.
However, this value is wrong in case one or several frames were already
parsed before the invalid frame in the same h3_rcv_buf() invokation.
This instructs caller to only remove a subset of the data and parsing is
restarted on a random boundary. Most of the times this generates again a
new final yet invalid error, possibly overwriting a stream error with a
full connection closure.
This patch fixes the return value in case of an error during HEADERS or
DATA parsing by ensuring that total variable is always incremented
instead of being directly assigned.
This must be backported up to 2.8.
http_cookie_merge() function is responsible to add a cookie header and merge
all values from a list. However, it was performed by appending values by
hand, using the pointer to the header value and changing the block
length. This was totally by-passing the HTX API.
For now, there is now bug because the function is called by h2 and quic
muxes when a HTTP message is parsed. And the cookie header is the last one
inserted. The HTX message is never fragmented and data from other blocs
cannot be overwritten. But, it could be an issue if it is called in another
context, from the HTTP analysis for instance.
To fix the issue, the function now relies on htx_replace_blk_value()
function to add the value separator first and then a cookie value.
This patch must be backported as far as 2.6.
