These calls can take quite some time and leave the thread harmless so
it's better to mark it as such. This makes "show sess" respond way
faster during high loads running on processes build with DEBUG_UAF
since these calls are stressed a lot.
When calling mmap(), in general the system gives us a page but does not
really allocate it until we first dereference it. And it turns out that
this time is much longer than the time to perform the mmap() syscall.
Unfortunately, when running with memory debugging enabled, we mmap/munmap()
each object resulting in lots of such calls and a high contention on the
allocator. And the first accesses to the page being done under the pool
lock is extremely damaging to other threads.
The simple fact of writing a 0 at the beginning of the page after
allocating it and placing the POOL_LINK pointer outside of the lock is
enough to boost the performance by 8x in debug mode and to save the
watchdog from triggering on lock contention. This is what this patch
does.
The malloc and free calls and especially the underlying mmap/munmap()
can occasionally take a huge amount of time and even cause the thread
to sleep. This is visible when haproxy is compiled with DEBUG_UAF which
causes every single pool allocation/free to allocate and release pages.
In this case, when using the locked pools, the watchdog can occasionally
fire under high contention (typically requesting 40000 1M objects in
parallel over 8 threads). Then, "perf top" shows that 50% of the CPU
time is spent in mmap() and munmap(). The reason the watchdog fires is
because some threads spin on the pool lock which is held by other threads
waiting on mmap() or munmap().
This patch modifies this so that the pool lock is released during these
syscalls. Not only this allows other threads to request try to allocate
their data in parallel, but it also considerably reduces the lock
contention.
Note that the locked pools are only used on small architectures where
high thread counts would not make sense, so this will not provide any
benefit in the general case. However it makes the debugging versions
way more stable, which is always appreciated.
In the function stream_int_notify(), when the opposite stream-interface is
blocked because there is no more room into the input buffer, if the flag
CF_WRITE_PARTIAL is set on this buffer, it is unblocked. It is a way to unblock
the reads on the other side because some data was sent.
But it is a problem during the fast-forwarding because only the stream is able
to remove the flag CF_WRITE_PARTIAL. So it is possible to have this flag because
of a previous send while the input buffer of the opposite stream-interface is
now full. In such case, the opposite stream-interface will be woken up for
nothing because its input buffer is full. If the same happens on the opposite
side, we will have a loop consumming all the CPU.
To fix the bug, the opposite side is now only notify if there is some available
room in its input buffer in the function si_cs_send(), so only if some data was
sent.
This patch must be backported to 2.0 and 1.9.
This code should be now used by action to stop at the same time the rules
processing and the possible following processings. And from its side, the return
code ACT_RET_STOP should be used to only stop rules processing.
So concretely, for TCP rules, there is no changes. ACT_RET_STOP and ACT_RET_DONE
are handled the same way. However, for HTTP rules, ACT_RET_STOP should now be
mapped on HTTP_RULE_RES_STOP and ACT_RET_DONE on HTTP_RULE_RES_DONE. So this
way, a action will have the possibilty to stop all processing or only rules
processing.
Note that changes about the TCP is done in this commit but changes about the
HTTP will be done in another one because it will fix a bug in the same time.
This patch must be backported to 2.0 because a bugfix depends on it.
When deciding if we keep an idle connection in the session, check if
the number of connections currently in the session is >= the max allowed,
not >, or we'll keep an extra connection.
This should be backported to 1.9 and 2.0.
Just calling conn_force_unsubscribe() from conn_upgrade_mux_fe() is not
enough, as there may be multiple XPRT involved. Instead, require that
any user of conn_upgrade_mux_fe() unsubscribe itself before calling it.
This should fix upgrading a TCP connection to HTX when using SSL.
This should be backported to 2.0.
The receive limit of an HTX channel must be calculated against the total size of
the HTX message. Otherwise, the buffer may never be seen as full whereas the
receive limit is 0. Indeed, the function channel_htx_full() already takes care
to add a block size to the buffer's reserve (8 bytes). So if the function
channel_htx_recv_limit() also keep a block size free in addition to the buffer's
reserve, it means that at least 2 block size will be kept free but only one will
be taken into account, freezing the stream if the option http-buffer-request is
enabled.
This patch fixes the Github issue #136. It should be backported to 2.0 and
1.9. Thanks jaroslawr (Jarosław Rzeszótko) for his help.
Revert commit fe4abe62c7c5206dff1802f42d17014e198b9141.
The goal was to make sure for health-checks, we would not get sockets in
TIME_WAIT. To do so, we would not call shutdown() if linger_risk is set.
However that is wrong, and that means shutw would never be forwarded to
the server, and thus we could get connection that are never properly closed.
Instead, to fix the original problem as described here :
https://www.mail-archive.com/haproxy@formilux.org/msg34080.html
Just make sure the checks code call cs_shutr() before calling cs_shutw().
If shutr has been called, conn_sock_shutw() will make no attempt to call
shutdown(), as it knows close() will be called.
We should really review and revamp the shutr/shutw code, as described in
github issue #142.
This should be backported to 1.9 and 2.0.
When using a level lower than admin on the master CLI, a \n is output
before the response, this is caused by the response of the "operator" or
"user" that are sent before the actual command.
To fix this problem we introduce the flag APPCTX_CLI_ST1_NOLF which ask
a command response to not be followed by the final \n.
This patch made a special case with the command operator and user
followed by a - so they are not followed by \n.
This patch must be backported to 2.0 and 1.9.
As its name suggest, this function change the value length of a block. But it
also update the HTX message accordingly. It simplifies the HTX API. The function
htx_set_blk_value_len() is still available and must be used with caution because
this one does not update the HTX message. It just updates the HTX block. It
should be considered as an internal function. When possible,
htx_change_blk_value_len() should be used instead.
This function is used to fix a bug affecting the 2.0. So, this patch must be
backported to 2.0.
Server states can be recovered from either a "global" file (all backends)
or a "local" file (per backend).
The way the algorithm to parse the state file was first implemented was good
enough for a low number of backends and servers per backend.
Basically, for each backend the state file (global or local) is opened,
parsed entirely and for each line we check if it contains data related to
a server from the backend we're currently processing.
We must read the file entirely, just in case some lines for the current
backend are stored at the end of the file.
This does not scale at all!
This patch changes the behavior above for the "global" file only. Now,
the global file is read and parsed once and all lines it contains are
stored in a tree, for faster discovery.
This result in way much less fopen, fgets, and strcmp calls, which make
loading of very big state files very quick now.
In commit 86eded6c6 ("CLEANUP: tasks: rename task_remove_from_tasklet_list()
to tasklet_remove_*") which consisted in removing the casts between tasks
and tasklet, I was a bit too fast to believe that we only saw tasklets in
this function since process_runnable_tasks() also uses it with tasks under
a cast. So removing the bookkeeping on task_list_size was not appropriate.
Bah, the joy of casts which hide the real thing...
This patch does two things at once to address this mess once for all:
- it restores the decrement of task_list_size when it's a real task,
but moves it to process_runnable_task() since it's the only place
where it's allowed to call it with a task
- it moves the increment there as well and renames
task_insert_into_tasklet_list() to tasklet_insert_into_tasklet_list()
of obvious consistency reasons.
This way the increment/decrement of task_list_size is made at the only
places where the cast is enforced, so it has less risks to be missed.
The comments on top of these functions were updated to reflect that they
are only supposed to be used with tasklets and that the caller is responsible
for keeping task_list_size up to date if it decides to enforce a task there.
Now we don't have to worry anymore about how these functions work outside
of the scheduler, which is better longterm-wise. Thanks to Christopher for
spotting this mistake.
No backport is needed.
In conn_sock_shutw(), avoid calling shutdown() if linger_risk is set. Not
doing so will result in getting sockets in TIME_WAIT for some time.
This is particularly observable with health checks.
This should be backported to 1.9.
The function really only operates on tasklets, its arguments are always
tasklets cast as tasks to match the function's type, to be cast back to
a struct tasklet. Let's rename it to tasklet_remove_from_tasklet_list(),
take a struct tasklet, and get rid of the undesired task casts.
It's really confusing to call it a task because it's a tasklet and used
in places where tasks and tasklets are used together. Let's rename it
to tasklet to remove this confusion.
The first one, HTX_SL_F_HAS_SCHM, will be used to know the request has an
explicit scheme. So, in H2, it is always true because the pseudo-header
":scheme" is mandatory. In H1, it is only true when an absolute URI is found on
the start-line. The other flags, HTX_SL_F_SCHM_HTTP and HTX_SL_F_SCHM_HTTPS,
will be used to know which scheme the request have. For now, other protocols are
not handled.
The aim of these flags is to pass this information to the backend side in
general, and to the H2 mux in particular. So the multiplexer will have a chance
to use this information to send the right scheme to the server.
In the HTX structure, the field <first> is used to know where to (re)start the
analysis. It may differ from the message's head. It is especially important to
update it to handle 1xx messages, to be sure to restart the analysis on the next
message (another 1xx message or the final one). It is also updated when some
data are forwarded (the headers or part of the body). But this update is an
error and must never be done at the analysis level. It is a bug, because some
sample fetches may be used after the data forwarding (but before the first send
of course). At this stage, if the first block position does not point on the
start-line, most of HTTP sample fetches fail.
So now, when something is forwarding by HTX analyzers, the first block position
is not update anymore.
This issue was reported on Github. See #119. No backport needed.
When channel_full() is called for an HTX stream, we fall back on the HTX
version. This function is called, among other, from tcp_inspect_request(). With
this patch, the inspect delay is respected again.
This patch must be backported to 1.9.
With both I/O and tasks in the same tasklet list, we now have a very
smooth and responsive scheduler, providing a good fairness between I/O
activities. With the lower layers relying on tasklet a lot (I/O wakeup,
subscribe, etc), there may often be a large number of totally autonomous
tasklets doing their business such as forwarding data between two muxes.
But the task scheduler historically refrained from picking tasks from the
priority-ordered run queue to put them into the tasklet list until this
later had less than max_runqueue_depth entries. This was to make sure that
low-latency, high-priority tasks would have an opportunity to be dequeued
before others even if they arrive late. But the counter used for this is
still the tasklet list size, which contains countless I/O events. This
causes an unfairness between unbounded I/Os and bounded tasks, resulting
for example in the CLI responding slower when forwarding 40 Gbps of HTTP
traffic spread over a thousand of connections.
A good solution consists in sticking to the initial intent of
max_runqueue_depth which is to limit the number of tasks in the list
(to maintain fairness between them) and not to limit the number of these
tasks among tasklets. It just turns out that the task_list_size initially
was this task counter and changed over time to be a tasklet list size.
Let's simply refrain from updating it for pure tasklets so that it takes
back its original role of counting real tasks as its name implies. With
this change the CLI becomes instantly responsive under load again.
This patch may possibly be backported to 1.9 though it requires some
careful checks.
The function htx_add_data_before() was removed because it was buggy. The
function htx_move_blk_before() may be used if necessary to do something
equivalent, except it just moves blocks. It doesn't handle the adding.
In an HTX message, it may have 2 available rooms to store a new block. The first
one is between the blocks and their payload. Blocks are added starting from the
end of the buffer and their payloads are added starting from the begining. So
the first free room is between these 2 edges. The second one is at the begining
of the buffer, when we start to wrap to add new payloads. Once we start to use
this one, the other one is ignored until the next defragmentation of the HTX
message.
In theory, there is no problem. But in practice, some lacks in the HTX structure
force us to defragment too often HTX messages to always be in a known state. The
second free room is not tracked as it should do and the first one may be easily
corrupted when rewrites happen.
So to fix the problem and avoid unecessary defragmentation, the HTX structure
has been refactored. The front (the block's position of the first payload before
the blocks) is no more stored. Instead we keep the relative addresses of 3 edges:
* tail_addr : The start address of the free space in front of the the blocks
table
* head_addr : The start address of the free space at the beginning
* end_addr : The end address of the free space at the beginning
Here is the general view of the HTX message now:
head_addr end_addr tail_addr
| | |
V V V
+------------+------------+------------+------------+------------------+
| | | | | |
| PAYLOAD | Free space | PAYLOAD | Free space | Blocks area |
| ==> | 1 | ==> | 2 | <== |
+------------+------------+------------+------------+------------------+
<head_addr> is always lower or equal to <end_addr> and <tail_addr>. <end_addr>
is always lower or equal to <tail_addr>.
In addition;, to simplify everything, the blocks area are now contiguous. It
doesn't wrap anymore. So the head is always the block with the lowest position,
and the tail is always the one with the highest position.
The function htx_add_data_before() is buggy and cannot work. It first add a data
block and then move it before another one, passed in argument. The problem
happens when a defragmentation is done to add the new block. In this case, the
reference is no longer valid, because the blocks are rearranged. So, instead of
moving the new block before the reference, it is moved at the head of the HTX
message.
So this function has been removed. It was only used by the compression filter to
add a last data block before a TLR, EOT or EOM block. Now, the new function
htx_add_last_data() is used. It adds a last data block, after all others and
before any TLR, EOT or EOM block. Then, the next bock is get. It is the first
non-data block after data in the HTX message. The compression loop continues
with it.
This patch must be backported to 1.9.
This function provides an alternate way to leave a critical section run
under thread_isolate(). Currently, a thread may remain in thread_release()
without having the time to notice that the rdv mask was released and taken
again by another thread entering thread_isolate() (often the same that just
released it). This is because threads wait in harmless mode in the loop,
which is compatible with the conditions to enter thread_isolate(). It's
not possible to make them wait with the harmless bit off or we cannot know
when the job is finished for the next thread to start in thread_isolate(),
and if we don't clear the rdv bit when going there, we create another
race on the start point of thread_isolate().
This new synchronous variant of thread_release() makes use of an extra
mask to indicate the threads that want to be synchronously released. In
this case, they will be marked harmless before releasing their sync bit,
and will wait for others to release their bit as well, guaranteeing that
thread_isolate() cannot be started by any of them before they all left
thread_sync_release(). This allows to construct synchronized blocks like
this :
thread_isolate()
/* optionally do something alone here */
thread_sync_release()
/* do something together here */
thread_isolate()
/* optionally do something alone here */
thread_sync_release()
And so on. This is particularly useful during initialization where several
steps have to be respected and no thread must start a step before the
previous one is completed by other threads.
This one must not be placed after any call to thread_release() or it would
risk to block an earlier call to thread_isolate() which the current thread
managed to leave without waiting for others to complete, and end up here
with the thread's harmless bit cleared, blocking others. This might be
improved in the future.
As reported in GH issue #109 and in discourse issue
https://discourse.haproxy.org/t/haproxy-returns-408-or-504-error-when-timeout-client-value-is-every-25d
the time parser doesn't error on overflows nor underflows. This is a
recurring problem which additionally has the bad taste of taking a long
time before hitting the user.
This patch makes parse_time_err() return special error codes for overflows
and underflows, and adds the control in the call places to report suitable
errors depending on the requested unit. In practice, underflows are almost
never returned as the parsing function takes care of rounding values up,
so this might possibly happen on 64-bit overflows returning exactly zero
after rounding though. It is not really possible to cut the patch into
pieces as it changes the function's API, hence all callers.
Tests were run on about every relevant part (cookie maxlife/maxidle,
server inter, stats timeout, timeout*, cli's set timeout command,
tcp-request/response inspect-delay).
When we look up an dictionary entry in the cache used upon transmission
we store the last result in ->prev_lookup of struct dcache_tx so that
to compare it with the subsequent entries to look up and save performances.
When allocating new dictionary entries we store the length of the strings.
May be useful so that not to have to call strlen() too much often at runing
time.
We store pointers to server names dictionary entries in a pre-allocated array of
ebpt_node's (->entries member of struct dcache_tx) to cache those sent to remote
peers. Consequently the ID used to identify the server name dictionary entry is
also used as index for this array. There is no need to implement a lookup by key
for this dictionary cache.
The fd_sets we've been using in the log encoding functions are not portable
and were shown to break at least under Cygwin. This patch gets rid of them
in favor of the new bitmap functions. It was verified with the config below
that the log output was exactly the same before and after the change :
defaults
mode http
option httplog
log stdout local0
timeout client 1s
timeout server 1s
timeout connect 1s
frontend foo
bind :8001
capture request header chars len 255
backend bar
option httpchk "GET" "/" "HTTP/1.0\r\nchars: \x01\x02\x03\x04\x05\x06\x07\x08\x09\x0b\x0c\x0e\x0f\x10\x11\x12\x13\x14\x15\x16\x17\x18\x19\x1a\x1b\x1c\x1d\x1e\x1f\x20\x21\x22\x23\x24\x25\x26\x27\x28\x29\x2a\x2b\x2c\x2d\x2e\x2f\x30\x31\x32\x33\x34\x35\x36\x37\x38\x39\x3a\x3b\x3c\x3d\x3e\x3f\x40\x41\x42\x43\x44\x45\x46\x47\x48\x49\x4a\x4b\x4c\x4d\x4e\x4f\x50\x51\x52\x53\x54\x55\x56\x57\x58\x59\x5a\x5b\x5c\x5d\x5e\x5f\x60\x61\x62\x63\x64\x65\x66\x67\x68\x69\x6a\x6b\x6c\x6d\x6e\x6f\x70\x71\x72\x73\x74\x75\x76\x77\x78\x79\x7a\x7b\x7c\x7d\x7e\x7f\x80\x81\x82\x83\x84\x85\x86\x87\x88\x89\x8a\x8b\x8c\x8d\x8e\x8f\x90\x91\x92\x93\x94\x95\x96\x97\x98\x99\x9a\x9b\x9c\x9d\x9e\x9f\xa0\xa1\xa2\xa3\xa4\xa5\xa6\xa7\xa8\xa9\xaa\xab\xac\xad\xae\xaf\xb0\xb1\xb2\xb3\xb4\xb5\xb6\xb7\xb8\xb9\xba\xbb\xbc\xbd\xbe\xbf\xc0\xc1\xc2\xc3\xc4\xc5\xc6\xc7\xc8\xc9\xca\xcb\xcc\xcd\xce\xcf\xd0\xd1\xd2\xd3\xd4\xd5\xd6\xd7\xd8\xd9\xda\xdb\xdc\xdd\xde\xdf\xe0\xe1\xe2\xe3\xe4\xe5\xe6\xe7\xe8\xe9\xea\xeb\xec\xed\xee\xef\xf0\xf1\xf2\xf3\xf4\xf5\xf6\xf7\xf8\xf9\xfa\xfb\xfc\xfd\xfe\xff"
server foo 127.0.0.1:8001 check
We now have ha_bit_{set,clr,flip,test} to manipulate bitfields made
of arrays of longs. The goal is to get rid of the remaining non-portable
FD_{SET,CLR,ISSET} that still exist at a few places.
Just like we have a synchronous recv() function for the stream interface,
let's have a synchronous send function that we'll be able to call from
different places. For now this only moves the code, nothing more.
We should not update the two directions at once, in fact we should update
the Rx path after recv() and the Tx path after send(). Let's start by
splitting the update function in two for this.
The purpose of making idle-conns switch to SI_ST_CON was to make the
transition detectable and the operation retryable in case of connection
error. Now we have the RDY state for this which is much more suitable
since it indicates a validated connection on which we didn't necessarily
send anything yet. This will still lead to a transition to EST while not
requiring unnatural write polling nor connect timeouts.
The main reason for all the trouble we're facing with stream interface
error or timeout reports during the connection phase is that we currently
can't make the difference between a connection attempt and a validated
connection attempt. It is problematic because we tend to switch early
to SI_ST_EST but can't always do what we want in this state since it's
supposed to be set when we don't need to visit sess_establish() again.
This patch introduces a new state betwen SI_ST_CON and SI_ST_EST, which
is SI_ST_RDY. It indicates that we've verified that the connection is
ready. It's a transient state, like SI_ST_DIS, that cannot persist when
leaving process_stream(). For now it is not set, only verified in various
tests where SI_ST_CON was used or SI_ST_EST depending on the cases.
The stream-int state diagram was minimally updated to reflect the new
state, though it is largely obsolete and would need to be seriously
updated.
The stream interface state checks involving ranges were replaced with
checks on a set of states, already revealing some issues. No issue was
fixed, all was replaced in a one-to-one mapping for easier control. Some
checks involving a strict difference were also replaced with fields to
be clearer. At this stage, the result must be strictly equivalent. A few
tests were also turned to their bit-field equivalent for better readability
or in preparation for upcoming changes.
The test performed in the SPOE filter was swapped so that the closed and
error states are evicted first and that the established vs conn state is
tested second.
At some places we do check for ranges of stream-int states but those
are confusing as states ordering is not well known (e.g. it's not obvious
that CER is between CON and EST). Let's create a bit field from states so
that we can match multiple states at once instead. The new enum si_state_bit
contains SI_SB_* which are state bits instead of state values. The function
si_state_in() indicates if the state in argument is one of those represented
by the bit mask in second argument.
Now that the various handshakes come with their own XPRT, there's no
need for the CONN_FL_SOCK* flags, and the conn_sock_want|stop functions,
so garbage-collect them.
Add a new XPRT that is used when using non-SSL handshakes, such as proxy
protocol or Netscaler, instead of taking care of it in conn_fd_handler().
This XPRT is installed when any of those is used, and it removes itself once
the handshake is done.
This should allow us to remove the distinction between CO_FL_SOCK* and
CO_FL_XPRT*.
Add a new method to xprt_ops, remove_xprt. When called, if the provided
xprt_ctx is the same as the xprt's underlying xprt_ctx, it then uses the
new xprt provided, otherwise it calls the remove_xprt method of the next
xprt.
The goal is to be able to add a temporary xprt, that removes itself from
the chain when it did what it had to do. This will be used to implement
a pseudo-xprt for anything that just requires a handshake (such as the
proxy protocol).
As the SSL code may have different needs than the upper layer, ie it may want
to receive when the upper layer wants to right, instead of directly forwarding
the subscribe to the underlying xprt, handle it ourself. The SSL code will
know remember any subscribe call, and wake the tasklet when it is ready
for more I/O.
