When a write activity is reported on a channel, it is important to keep this
information for the stream because it take part on the analyzers' triggering.
When some data are written, the flag CF_WRITE_PARTIAL is set. It participates to
the task's timeout updates and to the stream's waking. It is also used in
CF_MASK_ANALYSER mask to trigger channels anaylzers. In the past, it was cleared
by process_stream. Because of a bug (fixed in commit 95fad5ba4 ["BUG/MAJOR:
stream-int: don't re-arm recv if send fails"]), It is now cleared before each
send and in stream_int_notify. So it is possible to loss this information when
process_stream is called, preventing analyzers to be called, and possibly
leading to a stalled stream.
Today, this happens in HTTP2 when you call the stat page or when you use the
cache filter. In fact, this happens when the response is sent by an applet. In
HTTP1, everything seems to work as expected.
To fix the problem, we need to make the difference between the write activity
reported to lower layers and the one reported to the stream. So the flag
CF_WRITE_EVENT has been added to notify the stream of the write activity on a
channel. It is set when a send succedded and reset by process_stream. It is also
used in CF_MASK_ANALYSER. finally, it is checked in stream_int_notify to wake up
a stream and in channel_check_timeouts.
This bug is probably present in 1.7 but it seems to have no effect. So for now,
no needs to backport it.
There's no point having the channel marked writable as these functions
only extract data from the channel. The code was retrieved from their
ci/co ancestors.
For HTTP/2 we'll need some buffer-only equivalent functions to some of
the ones applying to channels and still squatting the bi_* / bo_*
namespace. Since these names have kept being misleading for quite some
time now and are really getting annoying, it's time to rename them. This
commit will use "ci/co" as the prefix (for "channel in", "channel out")
instead of "bi/bo". The following ones were renamed :
bi_getblk_nc, bi_getline_nc, bi_putblk, bi_putchr,
bo_getblk, bo_getblk_nc, bo_getline, bo_getline_nc, bo_inject,
bi_putchk, bi_putstr, bo_getchr, bo_skip, bi_swpbuf
When an entity tries to get a buffer, if it cannot be allocted, for example
because the number of buffers which may be allocated per process is limited,
this entity is added in a list (called <buffer_wq>) and wait for an available
buffer.
Historically, the <buffer_wq> list was logically attached to streams because it
were the only entities likely to be added in it. Now, applets can also be
waiting for a free buffer. And with filters, we could imagine to have more other
entities waiting for a buffer. So it make sense to have a generic list.
Anyway, with the current design there is a bug. When an applet failed to get a
buffer, it will wait. But we add the stream attached to the applet in
<buffer_wq>, instead of the applet itself. So when a buffer is available, we
wake up the stream and not the waiting applet. So, it is possible to have
waiting applets and never awakened.
So, now, <buffer_wq> is independant from streams. And we really add the waiting
entity in <buffer_wq>. To be generic, the entity is responsible to define the
callback used to awaken it.
In addition, applets will still request an input buffer when they become
active. But they will not be sleeped anymore if no buffer are available. So this
is the responsibility to the applet I/O handler to check if this buffer is
allocated or not. This way, an applet can decide if this buffer is required or
not and can do additional processing if not.
[wt: backport to 1.7 and 1.6]
In 1.4-dev3, commit 31971e5 ("[MEDIUM] add support for infinite forwarding")
made it possible to configure the lower layer to forward data indefinitely
by setting the forward size to CHN_INFINITE_FORWARD (4GB-1). By then larger
chunk sizes were not supported so there was no confusion in the usage of the
function.
Since 1.5 we support 64-bit content-lengths and chunk sizes and the function
has grown to support 64-bit arguments, though it still limits a single pass
to 32-bit quantities (what fit in the channel's to_forward field). The issue
now becomes that a 4GB-1 content-length can be confused with infinite
forwarding (in fact it's 4GB-1+what was already in the buffer). It causes a
visible effect when transferring this exact size because the transfer rate
is lower than with other sizes due in part to the disabling of the Nagle
algorithm on the sendto() call.
In theory with keep-alive it should prevent a second request from being
processed after such a transfer, but since the analysers are still present,
the forwarding analyser properly counts down the remaining size to transfer
and ultimately the transaction gets correctly reset so there is no visible
effect.
Since the root cause of the issue is an API problem (lack of distinction
between a real valid length and a magic value), this patch modifies the API
to have a new dedicated function called channel_forward_forever() to program
a permanent forwarding. The existing function __channel_forward() was modified
to properly take care of the requested sizes and ensure it 1) never overflows
and 2) never reaches CHN_INFINITE_FORWARD by accident.
It is worth noting that the function used to have a bug causing a 2GB
forward to be scheduled if it was called with less data than what is present
in buf->i. Fortunately this bug couldn't be triggered with existing code.
This fix should be backported to 1.6 and 1.5. While it also theorically
affects 1.4, it's better not to backport it there, as the risk of breaking
large object transfers due to significant API differences is high, compared
to the fact that the largest supported objects (4GB-1) are just slower to
transfer.
Unfortunately, commit 169c470 ("BUG/MEDIUM: channel: fix miscalculation of
available buffer space (3rd try)") was still not enough to completely
address the issue. It fell into an integer comparison trap. Contrary to
expectations, chn->to_forward may also have the sign bit set when
forwarding regular data having a large content-length, resulting in
an incomplete check of the result and of the reserve because the with
to_forward very large, to_forward+o could become very small and also
the reserve could become positive again and make channel_recv_limit()
return a negative value.
One way to reproduce this situation is to transfer a large file (> 2GB)
with http-keep-alive or http-server-close, without splicing, and ensure
that the server uses content-length instead of chunks. The transfer
should stall very early after the first buffer has been transferred
to the client.
This fix now properly checks 1) for an overflow caused by summing o and
to_forward, and 2) for o+to_forward being smaller or larger than maxrw
before performing the subtract, so that all sensitive operations are
properly performed on 33-bit arithmetics.
The code was subjected again to a series of tests using inject+httpterm
scanning a wide range of object sizes (+10MB after each new request) :
$ printf "new page 1\nget 127.0.0.1:8002 / s=%%s0m\n" | \
inject64 -o 1 -u 1 -f /dev/stdin
With previous fix, the transfer would suddenly stop when reaching 2GB :
hits ^hits hits/s ^h/s bytes kB/s last errs tout htime sdht ptime
203 1 2 1 216816173354 2710202 3144892 0 0 685.0 0.0 685.0
205 2 2 2 219257283186 2706880 2441109 0 0 679.5 6.5 679.5
205 0 2 0 219257283186 2673836 0 0 0 0.0 0.0 0.0
205 0 2 0 219257283186 2641622 0 0 0 0.0 0.0 0.0
205 0 2 0 219257283186 2610174 0 0 0 0.0 0.0 0.0
Now it's fine even past 4 GB.
Many thanks to Vedran Furac for reporting this issue early with a common
access pattern helping to troubleshoot this.
This fix must be backported to 1.6 and 1.5 where the commit above was
already backported.
This function returns non-zero if the channel is congested with data in
transit waiting for leaving, indicating to the caller that it should wait
for the reserve to be released before starting to process new data in
case it needs the ability to append data. This is meant to be used while
waiting for a clean response buffer before processing a request.
Latest fix 8a32106 ("BUG/MEDIUM: channel: fix miscalculation of available
buffer space (2nd try)") did happen to fix some observable issues but not
all of them in fact, some corner cases still remained and at least one user
reported a busy loop that appeared possible, though not easily reproducible
under experimental conditions.
The remaining issue is that we still consider min(i, to_fwd) as the number
of bytes in transit, but in fact <i> is not relevant here. Indeed, what
matters is that we can read everything we want at once provided that at
the end, <i> cannot be larger than <size-maxrw> (if it was not already).
This is visible in two cases :
- let's have i=o=max/2 and to_fwd=0. Then i+o >= max indicates that the
buffer is already full, while it is not since once <o> is forwarded,
some space remains.
- when to_fwd is much larger than i, it's obvious that we can fill the
buffer.
The only relevant part in fact is o + to_fwd. to_fwd will ensure that at
least this many bytes will be moved from <i> to <o> hence will leave the
buffer, whatever the number of rounds it takes.
Interestingly, the fix applied here ensures that channel_recv_max() will
now equal (size - maxrw - i + to_fwd), which is indeed what remains
available below maxrw after to_fwd bytes are forwarded from i to o and
leave the buffer.
Additionally, the latest fix made it possible to meet an integer overflow
that was not caught by the range test when forwarding in TCP or tunnel
mode due to to_forward being added to an existing value, causing the
buffer size to be limited when it should not have been, resulting in 2
to 3 recv() calls when a single one was enough. The first one was limited
to the unreserved buffer size, the second one to the size of the reserve
minus 1, and the last one to the last byte. Eg with a 2kB buffer :
recvfrom(22, "HTTP/1.1 200\r\nConnection: close\r"..., 1024, 0, NULL, NULL) = 1024
recvfrom(22, "23456789.123456789.123456789.123"..., 1023, 0, NULL, NULL) = 1023
recvfrom(22, "5", 1, 0, NULL, NULL) = 1
This bug is still present in 1.6 and 1.5 so the fix should be backported
there.
The condition to poll for receive as implemented in channel_may_recv()
is still incorrect. If buf->o is null and buf->i is slightly larger than
chn->to_forward and at least as large as buf->size - maxrewrite, then
reading will be disabled. It may slightly delay some data delivery by
having first to forward pending bytes, but may also cause some random
issues with analysers that wait for some data before starting to forward
what they correctly parsed. For instance, a body analyser may be prevented
from seeing the data that only fits in the reserve.
This bug may also prevent an applet's chk_rcv() function from being called
when part of a buffer is released. It is possible (though not verified)
that this participated to some peers frozen session issues some people
have been facing.
This fix should be backported to 1.6 and 1.5 to ensure better coherency
with channel_recv_limit().
Commit 9c06ee4 ("BUG/MEDIUM: channel: don't schedule data in transit for
leaving until connected") took care of an issue involving POST in conjunction
with http-send-name-header, where we absolutely never want to touch the
reserve until we're sure not to touch the buffer contents anymore, which
is indicated by the output stream-interface being connected.
But channel_may_recv() was not equipped with such a test, so in some
situations it might decide that it is possible to poll for reads, and
later channel_recv_limit() will decide it's not possible to read,
causing a loop. So we must add a similar test there.
Since the fix above was backported to 1.6 and 1.5, this fix must as well.
Commit 999f643 ("BUG/MEDIUM: channel: fix miscalculation of available buffer
space.") introduced a bug which made output data to be ignored when computing
the remaining room in a buffer. The problem is that channel_may_recv()
properly considers them and may declare that the FD may be polled for read
events, but once the even strikes, channel_recv_limit() called before recv()
says the opposite. In 1.6 and later this case is automatically caught by
polling loop detection at the connection level and is harmless. But the
backport in 1.5 ends up with a busy polling loop as soon as it becomes
possible to have a buffer with this conflict. In order to reproduce it, it
is necessary to have less than [maxrewrite] bytes available in a buffer, no
forwarding enabled (end of transfer) and [buf->o >= maxrewrite - free space].
Since this heavily depends on socket buffers, it will randomly strike users.
On 1.5 with 8kB buffers it was possible to reproduce it with httpterm using
the following command line :
$ (printf "GET /?s=675000 HTTP/1.0\r\n\r\n"; sleep 60) | \
nc6 --rcvbuf-size 1 --send-only 127.0.0.1 8002
This bug is only medium in 1.6 and later but is major in the 1.5 backport,
so it must be backported there.
Thanks to Nenad Merdanovic and Janusz Dziemidowicz for reporting this issue
with enough elements to help understand it.
The function channel_recv_limit() relies on channel_reserved() which
itself relies on channel_in_transit(). Individually they're OK but
combined they're doing the wrong thing.
The problem is that we refrain from filling buffers while to_forward
is even much larger than the buffer because of a semantic issue along
the call chain. This is particularly visible when offloading SSL on
moderately large files (1 MB), though it is also visible on clear text.
Twice the number of recv() calls are made compared to what is needed,
and the typical performance drops by 15-20% in SSL in 1.6 and later,
and no directly measurable drop in 1.5 except when using strace.
There's no need for all these intermediate functions, so let's get
rid of them and reimplement channel_recv_limit() from scratch in a
safer way.
This fix needs to be backported to 1.6 and 1.5 (at least). Note that in
1.5 the function is called buffer_recv_limit() and it may differ a bit.
With HTTP/2, we'll have to support multiplexed streams. A stream is in
fact the largest part of what we currently call a session, it has buffers,
logs, etc.
In order to catch any error, this commit removes any reference to the
struct session and tries to rename most "session" occurrences in function
names to "stream" and "sess" to "strm" when that's related to a session.
The files stream.{c,h} were added and session.{c,h} removed.
The session will be reintroduced later and a few parts of the stream
will progressively be moved overthere. It will more or less contain
only what we need in an embryonic session.
Sample fetch functions and converters will have to change a bit so
that they'll use an L5 (session) instead of what's currently called
"L4" which is in fact L6 for now.
Once all changes are completed, we should see approximately this :
L7 - http_txn
L6 - stream
L5 - session
L4 - connection | applet
There will be at most one http_txn per stream, and a same session will
possibly be referenced by multiple streams. A connection will point to
a session and to a stream. The session will hold all the information
we need to keep even when we don't yet have a stream.
Some more cleanup is needed because some code was already far from
being clean. The server queue management still refers to sessions at
many places while comments talk about connections. This will have to
be cleaned up once we have a server-side connection pool manager.
Stream flags "SN_*" still need to be renamed, it doesn't seem like
any of them will need to move to the session.
The purpose of these two macros will be to pass via the session to
find the relevant stream interfaces so that we don't need to store
the ->cons nor ->prod pointers anymore. Currently they're only defined
so that all references could be removed.
Note that many places need a second pass of clean up so that we don't
have any chn_prod(&s->req) anymore and only &s->si[0] instead, and
conversely for the 3 other cases.
We now have functions to retrieve one block and one line from
either the input or the output part of a buffer. They return
up to two (pointer,length) values in case the buffer wraps.
These modifications are done for resolving cross-dependent
includes in the upcoming LUA code.
<proto/channel.h> misses <types/channel.h>.
<types/acl.h> doesn't use <types/session.h> because the session
is already declared in the file as undefined pointer.
appsession.c misses <unistd.h> to use "write()".
Declare undefined pointer "struct session" for <types/proxy.h>
and <types/queue.h>. These includes dont need the detail of this
struct.
It applies to the channel and it doesn't erase outgoing data, only
pending unread data, which is strictly equivalent to what recv()
does with MSG_TRUNC, so that new name is more accurate and intuitive.
This name more accurately reminds that it applies to a channel and not
to a buffer, and that what is returned may be used as a max number of
bytes to pass to recv().
This applies to the channel, not the buffer, so let's fix this name.
Warning, the function's name happens to be the same as the old one
which was mistakenly used during 1.5.
This function's name was poorly chosen and is confusing to the point of
being suspiciously used at some places. The operations it does always
consider the ability to forward pending input data before receiving new
data. This is not obvious at all, especially at some places where it was
used when consuming outgoing data to know if the buffer has any chance
to ever get the missing data. The code needs to be re-audited with that
in mind. Care must be taken with existing code since the polarity of the
function was switched with the renaming.
channel_reserved is confusingly named. It is used to know whether or
not the rewrite area is left intact for situations where we want to
ensure we can use it before proceeding. Let's rename it to fix this
confusion.
Option http-send-name-header is still hurting. If a POST request has to be
redispatched when this option is used, and the next server's name is larger
than the initial one, and the POST body fills the buffer, it becomes
impossible to rewrite the server's name in the buffer when redispatching.
In 1.4, this is worse, the process may crash because of a negative size
computation for the memmove().
The only solution to fix this is to refrain from eating the reserve before
we're certain that we won't modify the buffer anymore. And the condition for
that is that the connection is established.
This patch introduces "channel_may_send()" which helps to detect whether it's
safe to eat the reserve or not. This condition is used by channel_in_transit()
introduced by recent patches.
This patch series must be backported into 1.5, and a simpler version must be
backported into 1.4 where fixing the bug is much easier since there were no
channels by then. Note that in 1.4 the severity is major.
This function returns the amount of bytes in transit in a channel's buffer,
which is the amount of outgoing data plus the amount of incoming data bound
to the forward limit.
We know that all incoming data are going to be purged if to_forward
is greater than them, not only if greater than the buffer size. This
buf has no direct impact on this version, but it participates to some
bugs affecting http-send-name-header since 1.4. This fix will have to
be backported down to 1.4 albeit in a different form.
The buffer_max_len() function is subject to an integer overflow in this
calculus :
int ret = global.tune.maxrewrite - chn->to_forward - chn->buf->o;
- chn->to_forward may be up to 2^31 - 1
- chn->buf->o may be up to chn->buf->size
- global.tune.maxrewrite is by definition smaller than chn->buf->size
Thus here we can subtract (2^31 + buf->o) (highly negative) from something
slightly positive, and result in ret being larger than expected.
Fortunately in 1.5 and 1.6, this is only used by bi_avail() which itself
is used by applets which do not set high values for to_forward so this
problem does not happen there. However in 1.4 the equivalent computation
was used to limit the size of a read and can result in a read overflow
when combined with the nasty http-send-name-header feature.
This fix must be backported to 1.5 and 1.4.
bi_swpbuf() swaps the buffer passed in argument with the one attached to
the channel, but only if this last one is empty. The idea is to avoid a
copy when buffers can simply be swapped.
Till now we'd consider a buffer full even if it had size==0 due to pointing
to buf.size. Now we change this : if buf_wanted is present, it means that we
have already tried to allocate a buffer but failed. Thus the buffer must be
considered full so that we stop trying to poll for reads on it. Otherwise if
it's empty, it's buf_empty and we report !full since we may allocate it on
the fly.
Channels are now created with a valid pointer to a buffer before the
buffer is allocated. This buffer is a global one called "buf_empty" and
of size zero. Thus it prevents any activity from being performed on
the buffer and still ensures that chn->buf may always be dereferenced.
b_free() also resets the buffer to &buf_empty, and was split into
b_drop() which does not reset the buffer.
We'll soon need to be able to switch buffers without touching the
channel, so let's move buffer initialization out of channel_init().
We had the same in compressoin.c.
We store the time stamp of last read in the channel in order to
be able to measure some bit rate and pause lengths. We only use
16 bits which were unused for this. We don't need more, as it
allows us to measure with a millisecond precision for up to 65s.
These ones are only reset during transfers. There is a low but non-null
risk that a first full read causes the previous value to be reused and
immediately to immediately set the CF_STREAMER flag. The impact is only
to increase earlier than expected the SSL record size and to use splice().
This bug was already present in 1.4, so a backport is possible.
Since 1.5-dev12 and commit 3bf1b2b8 (MAJOR: channel: stop relying on
BF_FULL to take action), the HTTP parser switched to channel_full()
instead of BF_FULL to decide whether a buffer had enough room to start
parsing a request or response. The problem is that channel_full()
intentionally ignores outgoing data, so a corner case exists where a
large response might still be left in a response buffer with just a
few bytes left (much less than the reserve), enough to accept a second
response past the last data, but not enough to permit the HTTP processor
to add some headers. Since all the processing relies on this space being
available, we can get some random crashes when clients pipeline requests.
The analysis of a core from haproxy configured with 20480 bytes buffers
shows this : with enough "luck", when sending back the response for the
first request, the client is slow, the TCP window is congested, the socket
buffers are full, and haproxy's buffer fills up. We still have 20230 bytes
of response data in a 20480 response buffer. The second request is sent to
the server which returns 214 bytes which fit in the small 250 bytes left
in this buffer. And the buffer arrangement makes it possible to escape all
the controls in http_wait_for_response() :
|<------ response buffer = 20480 bytes ------>|
[ 2/2 | 3 | 4 | 1/2 ]
^ start of circular buffer
1/2 = beginning of previous response (18240)
2/2 = end of previous response (1990)
3 = current response (214)
4 = free space (36)
- channel_full() returns false (20230 bytes are going to leave)
- the response headers does not wrap at the end of the buffer
- the remaining linear room after the headers is larger than the
reserve, because it's the previous response which wraps :
=> response is processed
Header rewriting causes it to reach 260 bytes, 10 bytes larger than what
the buffer could hold. So all computations during header addition are
wrong and lead to the corruption we've observed.
All the conditions are very hard to meet (which explains why it took
almost one year for this bug to show up) and are almost impossible to
reproduce on purpose on a test platform. But the bug is clearly there.
This issue was reported by Dinko Korunic who kindly devoted a lot of
time to provide countless traces and cores, and to experiment with
troubleshooting patches to knock the bug down. Thanks Dinko!
No backport is needed, but all 1.5-dev versions between dev12 and dev18
included must be upgraded. A workaround consists in setting option
forceclose to prevent pipelined requests from being processed.
Hijackers were functions designed to inject data into channels in the
distant past. They became unused around 1.3.16, and since there has
not been any user of this mechanism to date, it's uncertain whether
the mechanism still works (and it's not really useful anymore). So
better remove it as well as the pointer it uses in the channel struct.
Most calls to channel_forward() are performed with short byte counts and
are already optimized in channel_forward() taking just a few instructions.
Thus it's a waste of CPU cycles to call a function for this, let's just
inline the short byte count case and fall back to the common one for
remaining situations.
Doing so has increased the chunked encoding parser's performance by 12% !
With this commit, we now separate the channel from the buffer. This will
allow us to replace buffers on the fly without touching the channel. Since
nobody is supposed to keep a reference to a buffer anymore, doing so is not
a problem and will also permit some copy-less data manipulation.
Interestingly, these changes have shown a 2% performance increase on some
workloads, probably due to a better cache placement of data.
This is a massive rename of most functions which should make use of the
word "channel" instead of the word "buffer" in their names.
In concerns the following ones (new names) :
unsigned long long channel_forward(struct channel *buf, unsigned long long bytes);
static inline void channel_init(struct channel *buf)
static inline int channel_input_closed(struct channel *buf)
static inline int channel_output_closed(struct channel *buf)
static inline void channel_check_timeouts(struct channel *b)
static inline void channel_erase(struct channel *buf)
static inline void channel_shutr_now(struct channel *buf)
static inline void channel_shutw_now(struct channel *buf)
static inline void channel_abort(struct channel *buf)
static inline void channel_stop_hijacker(struct channel *buf)
static inline void channel_auto_connect(struct channel *buf)
static inline void channel_dont_connect(struct channel *buf)
static inline void channel_auto_close(struct channel *buf)
static inline void channel_dont_close(struct channel *buf)
static inline void channel_auto_read(struct channel *buf)
static inline void channel_dont_read(struct channel *buf)
unsigned long long channel_forward(struct channel *buf, unsigned long long bytes)
Some functions provided by channel.[ch] have kept their "buffer" name because
they are really designed to act on the buffer according to some information
gathered from the channel. They have been moved together to the same place in
the file for better readability but they were not changed at all.
The "buffer" memory pool was also renamed "channel".
