The check I/O handler, process_chk_conn and server_warmup are often
present in complex backtraces as they're impacted by locking or I/O
issues. Let's export them so that they resolve cleanly.
This finally adds the long-awaited solution to inspect the run queues
and figure what is eating the CPU or causing latencies. We can even see
the experienced latencies when profiling is enabled. Example on a
saturated process:
> show tasks
Running tasks: 14983 (4 threads)
function places % lat_tot lat_avg
process_stream 4948 33.0 5.840m 70.82ms
h1_io_cb 2535 16.9 - -
main+0x9e670 2508 16.7 2.930m 70.10ms
ssl_sock_io_cb 2499 16.6 - -
si_cs_io_cb 2493 16.6 - -
If a user enables profiling by hand, it makes sense to reset the stats
counters to provide fresh new measurements. Therefore it's worth using
this as the standard method to reset counters.
"show profiling" will now dump the stats collected by the scheduler if
profiling was previously enabled. This will immediately make it obvious
what functions are responsible for others' high latencies or which ones
are suffering from others, and should help spot issues like undesired
wakeups.
Example:
Per-task CPU profiling : on # set profiling tasks {on|auto|off}
Tasks activity:
function calls cpu_tot cpu_avg lat_tot lat_avg
si_cs_io_cb 5569479 23.37s 4.196us - -
h1_io_cb 5558654 13.60s 2.446us - -
process_stream 250841 1.476s 5.882us 3.499s 13.95us
main+0x9e670 198 - - 5.526ms 27.91us
task_run_applet 17 1.509ms 88.77us 205.8us 12.11us
srv_cleanup_idle_connections 12 44.51us 3.708us 25.71us 2.142us
main+0x158c80 9 48.72us 5.413us - -
srv_cleanup_toremove_connections 5 165.1us 33.02us 123.6us 24.72us
Now when the profiling is enabled, the scheduler wlil update per-function
task-level statistics on number of calls, cpu usage and lateny, that could
later be checked using "show profiling". This will immediately make it
obvious what functions are responsible for others' high latencies or which
ones are suffering from others, and should help spot issues like undesired
wakeups. For now the stats are only collected but not reported (though they
are readable from sched_activity[] under gdb).
The new sched_activity structure will be used to collect task-level
activity based on the target function. The principle is to declare a
large enough array to make collisions rare (256 entries), and hash
the function pointer using a reduced XXH to decide where to store the
stats. On first computation an entry is definitely assigned to the
array and it's done atomically. A special entry (0) is used to store
collisions ("others"). The goal is to make it easy and inexpensive for
the scheduler code to use these to store #calls, cpu_time and lat_time
for each task.
In 2.0, commit d2d3348ac ("MINOR: activity: enable automatic profiling
turn on/off") introduced an automatic mode to enable/disable profiling.
The problem is that the automatic mode automatically changes to on/off,
which implied that the forced on/off modes aren't sticky anymore. It's
annoying when debugging because as soon as the load decreases, profiling
stops.
This makes a small change which ought to have been done first, which
consists in having two states for "auto" (auto-on, auto-off) to
distinguish them from the forced states. Setting to "auto" in the config
defaults to "auto-off" as before, and setting it on the CLI switches to
auto but keeps the current operating state.
This is simple enough to be backported to older releases if needed.
When reporting some values in debugging output we often need to have
some condensed, stable-length values. This function prints a duration
from nanosecond to years with at least 4 digits of accuracy using the
most suitable unit, always on 7 chars.
Do not consider reuse connection if available list is not allocated for
the target server. This will prevent a crash when using a standalone
server for an external purpose like socket_tcp/socket_ssl on hlua code.
For the idle/safe lists, they are considered allocated if
srv.max_idle_conns is not null.
Note that the hlua code is currently safe thanks to the additional
checks on proxy http mode and stream reuse policy not never. However,
this might not be sufficient for future code.
This patch should be backported in every branches containing the
following patch :
7f68d815af356fbe1b2e1080a88b9935581c91d2 (2.4 tree)
REORG: backend: simplify conn_backend_get
This reverts commit 62e8aaa1bd5ca96089eaa88487c700c4af4617f4.
While is works extremely well to address SSL handshake floods, it prevents
establishment of new connections during regular traffic above 50-60 Gbps,
because for an unknown reason the queue seems to have ~1.7 active tasks
per connection all the time, which makes no sense as these ought to be
waiting on subscribed events. It might uncover a deeper issue but at least
for now a different solution is needed. cf issue #822.
The test is trivial to run, just start a config with tune.runqueue-depth 10
and inject on 1GB objects with more than 10 connections. Try to connect to
the stats socket, it only works once, then the listeners are not dequeued.
In github issue #822, user @ngaugler reported some performance problems when
dealing with many concurrent SSL connections on restarts, after migrating
from 1.6 to 2.2, indicating a long time required to re-establish connections.
The Run_queue metric in the traces showed an abnormally high number of tasks
in the run queue, likely indicating we were accepting faster than we could
process. And this is indeed one of the differences between 1.6 and 2.2, the
accept I/O loop and the TLS handshakes are totally independent, so much that
they can even run on different threads. In 1.6 the SSL handshake was handled
almost immediately after the accept(), so this was limiting the input rate.
With large maxconn values, as long as there are incoming connections, new
I/Os are scheduled and many of them pass before the handshake, being tagged
for low latency processing.
The result is that handshakes get postponed, and are further postponed as
new connections are accepted. When they are finally able to be processed,
some of them fail as the client is gone, and the client had already queued
new ones. This causes an excess number of apparent connections and total
number of handshakes to be processed, just because we were accepting
connections on a temporarily saturated machine.
The solution is to temporarily pause new incoming connections when the
load already indicates that more tasks are already queued than will be
handled in a poll loop. The difficulty with this usually is to be able
to come back to re-enable the operation, but given that the metric is
the run queue, we just have to queue the global_listener_queue task so
that it gets picked by any thread once the run queues get flushed.
Before this patch, injecting with SSL reneg with 10000 concurrent
connections resulted in 350k tasks in the run queue, and a majority of
handshake timeouts noticed by the client. With the patch, the run queue
fluctuates between 1-3x runqueue-depth, the process is constantly busy, the
accept rate is maximized and clients observe no error anymore.
It would be desirable to backport this patch to 2.3 and 2.2 after some more
testing, provided the accept loop there is compatible.
The allowed chunk size was historically limited to 2GB to avoid risk of
overflow. This restriction is no longer necessary because the chunk size is
immediately stored into a 64bits integer after the parsing. Thus, it is now
possible to raise this limit. However to never fed possibly bogus values
from languages that use floats for their integers, we don't get more than 13
hexa-digit (2^52 - 1). 4 petabytes is probably enough !
This patch should fix the issue #1065. It may be backported as far as
2.1. For the 2.0, the legacy HTTP part must be reviewed. But there is
honestely no reason to do so.
A number of traces could be added or changed to report errors with
TRACE_ERROR. The goal is to be able to enable error tracing only to detect
anomalies.
A number of traces could be added or changed to report errors with
TRACE_ERROR. The goal is to be able to enable error tracing only to detect
anomalies.
In order to announce support for the Extended CONNECT h2 method by
haproxy, always send the ENABLE_CONNECT_PROTOCOL h2 settings. This new
setting has been described in the rfc 8441.
After receiving ENABLE_CONNECT_PROTOCOL, the client is free to use the
Extended CONNECT h2 method. This can notably be useful for the support
of websocket handshake on http/2.
Support for the rfc 8441 Bootstraping WebSockets with HTTP/2
Convert an Extended CONNECT HTTP/2 request into a htx representation.
The htx message uses the GET method with an Upgrade header field to be
fully compatible with the equivalent HTTP/1.1 Upgrade mechanism.
The Extended CONNECT is of the following form :
:method = CONNECT
:protocol = websocket
:scheme = https
:path = /chat
:authority = server.example.com
The new pseudo-header :protocol has been defined and is used to identify
an Extended CONNECT method. Contrary to standard CONNECT, Extended
CONNECT must have :scheme, :path and :authority defined.
Add the header Sec-Websocket-Key when generating a h1 handshake websocket
without this header. This is the case when doing h2-h1 conversion.
The key is randomly generated and base64 encoded. It is stored on the session
side to be able to verify response key and reject it if not valid.
Support for the rfc 8441 Bootstraping WebSockets with HTTP/2
Generate an HTTP/2 Extended CONNECT request from a htx Upgrade message.
This conversion is done when seeing the header Connection: Upgrade. A
CONNECT request is written with the :protocol pseudo-header set from the
Upgrade htx header value.
The protocol is saved in the h2s structure. This is needed on the
response side because the protocol is not present on HTTP/2 response but
is needed if the client side is using HTTP/1.1 with 101 status code.
Support for the rfc 8441 Bootstraping WebSockets with HTTP/2
Convert a 200 status reply from an Extended CONNECT request into a htx
representation. The htx message is set to 101 status code to be fully
compatible with the equivalent HTTP/1.1 Upgrade mechanism.
This conversion is only done if the stream flags H2_SF_EXT_CONNECT_SENT
has been set. This is true if an Extended CONNECT request has already
been seen on the stream.
Besides the 101 status, the additional headers Connection/Upgrade are
added to the htx message. The protocol is set from the value stored in
h2s. Typically it will be extracted from the client request. This is
only used if the client is using h1 as only the HTTP/1.1 101 Response
contains the Upgrade header.
This flag is used to signal that an Extended CONNECT has been sent by
the server mux on the current stream.
This will allow to convert the response to a 101 htx status message.
Add the Sec-Websocket-Accept header on a websocket handshake response.
This header may be missing if a h2 server is used with a h1 client.
The response key is calculated following the rfc6455. For this, the
handshake request key must be stored in the h1 session, as a new field
name ws_key. Note that this is only done if the message has been
prealably identified as a Websocket handshake request.
If a request is identified as a WebSocket handshake, it must contains a
websocket key header or else it can be reject, following the rfc6455.
A new flag H1_MF_UPG_WEBSOCKET is set on such messages. For the request
te be identified as a WebSocket handshake, it must contains the headers:
Connection: upgrade
Upgrade: websocket
This commit is a compagnon of
"MEDIUM: h1: generate WebSocket key on response if needed" and
"MEDIUM: h1: add a WebSocket key on handshake if needed".
Indeed, it ensures that a WebSocket key is added only from a http/2 side
and not for a http/1 bogus peer.
The code dealing with the copy of requests in the L7-buffer and the
retransmits during L7 retries has been moved in the HTTP analysers. The copy
is now performed in the REQ_HTTP_XFER_BODY analyser and the L7 retries is
performed in the RES_WAIT_HTTP analyser. This way, si_cs_recv() and
si_cs_send() don't care of it anymore. It is much more natural to deal with
L7 retry in HTTP analysers.
Some responses must not contain data. Reponses to HEAD requests and 204/304
responses. But there is no warranty that this will be really respected by
the senders or even if it is possible. For instance, the method may be
rewritten by an http-request rule (HEAD->GET). Thus, it is not really
possible to always strip these data from the response at the receive
stage. And the response may be emitted by an applet or an internal service
not strictly following the spec. All that to say that we may be prepared to
handle payload for bodyless responses on the sending path.
In addition, unlike the HTTP/1, it is not really clear that the trailers is
part of the payload or not. Thus, some clients may expect to have the
trailers, if any, in the response to a HEAD request. For instance, the GRPC
status is placed in a trailer and clients rely on it. But what happens for
204 responses then. Read the following thread for details :
https://lists.w3.org/Archives/Public/ietf-http-wg/2020OctDec/0040.html
So, thanks to previous patches, it is now possible to know on the sending
path if a response must be bodyless or not. So, for such responses, no DATA
frame is emitted, except eventually the last empty one carring the ES
flag. However, the TRAILERS frames are still emitted. The h2s_skip_data()
function is added to take care to remove HTX DATA blocks without emitting
any DATA frame expect the last one, if there is no trailers.
The H2 message flag H2_MSGF_BODYLESS_RSP is now used during the request or
the response parsing to notify the mux that, considering the parsed message,
the response is known to have no body. This happens during HEAD requests
parsing and during 204/304 responses parsing.
On the H2 multiplexer, the equivalent flag is set on H2 streams. Thus the
H2_SF_BODYLESS_RESP flag is set on a H2 stream if the H2_MSGF_BODYLESS_RSP
is found after a HEADERS frame parsing. Conversely, this flag is also set
when a HEADERS frame is emitted for HEAD requests and for 204/304 responses.
The H2_SF_BODYLESS_RESP flag will be used to ignore data payload from the
response but not the trailers.
No connection header must be added by the H1 mux in 1xx messages, including
101. Existing connection headers remains untouched, especially the "Connection:
upgrade" of 101 responses. This patch only avoids to add "Connection: close" or
"Connection: keep-alive" to 1xx responses.
204 and 1xx responses must not have any payload. Now, the H1 mux takes care
of that in last resort. But they also must not have any C-L or T-E
headers. Thus, if found on the sending path, these headers are ignored.
Some responses must not contain data. Reponses to HEAD requests and 204/304
xresponses. But there is no warranty that this will be really respected by
the senders or even if it is possible. For instance, the method may be
rewritten by an http-request rule (HEAD->GET). Thus, it is not really
possible to always strip the payload from the response at the receive
stage. And the response may be emitted by an applet or an internal service
not strictly following the spec. All that to say that we may be prepared to
handle payload for bodyless responses on the sending path.
So, thanks to previous patches, it is now possible to know on the sending
path if a response must be bodyless or not. So, for such responses, no
payload is emitted, all HTX blocks after the EOH are silently removed
(including the trailers).
In HTTP/1, responses to HEAD requests and 204/304 must not have payload. The
H1S_F_BODYLESS_RESP flag is not set on streams that should handle such
responses, on the client side and the server side.
On the client side, this flag is set when a HEAD request is parsed and when
a 204/304 response is emitted. On the server side, this happends when a HEAD
request is emitted or a 204/304 response is parsed.
The EOM block may be removed. The HTX_FL_EOM flags is enough. Most of time,
to know if the end of the message is reached, we just need to have an empty
HTX message with HTX_FL_EOM flag set. It may also be detected when the last
block of a message with HTX_FL_EOM flag is manipulated.
Removing EOM blocks simplifies the HTX message filling. Indeed, there is no
more edge problems when the message ends but there is no more space to write
the EOM block. However, some part are more tricky. Especially the
compression filter or the FCGI mux. The compression filter must finish the
compression on the last DATA block. Before it was performed on the EOM
block, an extra DATA block with the checksum was added. Now, we must detect
the last DATA block to be sure to finish the compression. The FCGI mux on
its part must be sure to reserve the space for the empty STDIN record on the
last DATA block while this record was inserted on the EOM block.
The H2 multiplexer is probably the part that benefits the most from this
change. Indeed, it is now fairly easier to known when to set the ES flag.
The HTX documentaion has been updated accordingly.
Tunnel management between the H1 and H2 multiplexers is a bit blurred. And
the HTX is not enough well defined on this point to make things clear. In
fact, Establishing a tunnel between an H2 client and an H1 server, or the
opposite is buggy because the both multiplexers don't handle the EOM block
the same way when a tunnel is established. In fact, the H2 multiplexer is
pretty strict and add an END_STREAM flag when an EOM block is found, while
the H1 multiplexer is more flexible.
The purpose of this patch is to make the EOM block usage pretty clear and to
fix the HTTP multiplexers to really handle HTTP tunnels in the right
way. Now, an EOM block is used to mark the end of an HTTP message,
semantically speaking. That means it may be followed by tunneled data. Thus,
CONNECT requests are now finished by an EOM block, just after the EOH block.
On the H1 multiplexer side, a tunnel is now only established on the response
path. So a CONNECT request remains in a DONE state waiting for the 2xx
response. On the H2 multiplexer side, a flag is used to know an HTTP tunnel
is requested, to not immediately add the END_STREAM flag on the EOM block.
All these changes are sensitives and not backportable because of recent
changes. The same problem exists on earlier versions and should be
addressed. But it will only be possible with a specific patchset.
This patch relies on the following ones :
* MEDIUM: mux-h1: Properly handle tunnel establishments and aborts
* MEDIUM: mux-h2: Close streams when processing data for an aborted tunnel
* MEDIUM: mux-h2: Block client data on server side waiting tunnel establishment
* MINOR: mux-h2: Add 2 flags to help to properly handle tunnel mode
* MINOR: mux-h1: Split H1C_F_WAIT_OPPOSITE flag to separate input/output sides
* MINOR: mux-h1/mux-fcgi: Don't set TUNNEL mode if payload length is unknown
In the same way than the H2 mux, we now bloc data sending on the server side
if a tunnel is not fully established. In addition, if some data are still
pending for a aborted tunnel, an error is triggered and the server
connection is closed.
To do so, we rely on the H1C_F_WAIT_INPUT flag to bloc the output
processing. This patch contributes to fix the tunnel mode between the H1 and
the H2 muxes.
In the previous patch ("MEDIUM: mux-h2: Block client data on server side
waiting tunnel establishment"), we added a way to block client data for not
fully established tunnel on the server side. This one closes the stream with
an ERR_CANCEL erorr if there are some pending tunneled data while the tunnel
was aborted. This may happen on the client side if a non-empty DATA frame or
an empty DATA frame without the ES flag is received. This may also happen on
the server side if there is a DATA htx block. However in this last case, we
first wait the response is fully forwarded.
This patch contributes to fix the tunnel mode between the H1 and the H2
muxes.
On the server side, when a tunnel is not fully established, we must block
tunneled data, waiting for the server response. It is mandatory because the
server may refuse the tunnel. This happens when a DATA htx block is
processed in tunnel mode (H2_SF_BODY_TUNNEL flag set) but before the
response HEADERS frame is received (H2_SF_HEADERS_RCVD flag no set). In this
case, the H2_SF_BLK_MBUSY flag is set to mark the stream as busy. This flag
is removed when the tunnel is fully established or aborted.
This patch contributes to fix the tunnel mode between the H1 and the H2
muxes.
H2_SF_BODY_TUNNEL and H2_SF_TUNNEL_ABRT flags are added to properly handle
the tunnel mode in the H2 mux. The first one is used to detect tunnel
establishment or fully established tunnel. The second one is used to abort a
tunnel attempt. It is the first commit having as a goal to fix tunnel
establishment between H1 and H2 muxes.
There is a subtlety in h2_rcv_buf(). CS_FL_EOS flag is added on the
conn-stream when ES is received on a tunneled stream. It really reflects the
conn-stream state and is mandatory for next commits.
The H1C_F_WAIT_OPPOSITE flag is now splitted in 2 flags, H1C_F_WAIT_INPUT
and H1C_F_WAIT_OUTPUT, depending on the side is waiting. The change is a
prerequisite to fix the tunnel mode management in HTTP muxes.
H1C_F_WAIT_INPUT must be used to bloc the output side and to wait for an
event from the input side. H1C_F_WAIT_OUTPUT does the opposite. It bloc the
input side and wait for an event from the output side.
Responses with no C-L and T-E headers are no longer switched in TUNNEL mode
and remains in DATA mode instead. The H1 and FCGI muxes are updated
accordingly. This change reflects the real message state. It is not a true
tunnel. Data received are still part of the message.
It is not a bug. However, this message may be backported after some
observation period (at least as far as 2.2).
As stated in the RFC7540, section 8.1.1, the HTTP/2 removes support for the
101 informational status code. Thus a PROTOCOL_ERROR is now returned to the
server if a 101-switching-protocols response is received. Thus, the server
connection is aborted.
This patch may be backported as far as 2.0.
A 101-switching-protocols response must contain a Connection header with the
Upgrade option. And this response must only be received from a server if the
client explicitly requested a protocol upgrade. Thus, the request must also
contain a Connection header with the Upgrade option. If not, a
502-bad-gateway response is returned to the client. This way, a tunnel is
only established if both sides are agree.
It is closer to what the RFC says, but it remains a bit flexible because
there is no check on the Upgrade header itself. However, that's probably
enough to ensure a tunnel is not established when not requested.
This one is not tagged as a bug. But it may be backported, at least to
2.3. It relies on :
* MINOR: htx/http-ana: Save info about Upgrade option in the Connection header
Add an HTX start-line flag and its counterpart into the HTTP message to
track the presence of the Upgrade option into the Connection header. This
way, without parsing the Connection header again, it will be easy to know if
a client asks for a protocol upgrade and if the server agrees to do so. It
will also be easy to perform some conformance checks when a
101-switching-protocols is received.
It is the second part and the most important of the fix.
Since the mux-h1 refactoring, and more specifically since the commit
c4bfa59f1 ("MAJOR: mux-h1: Create the client stream as later as possible"),
the upgrade from a TCP client connection to H1 is broken. Indeed, now the H1
mux is responsible to create the frontend conn-stream once the request
headers are fully received. But, to properly support TCP to H1 upgrades, we
must inherit from the existing conn-stream. To do so, if the conn-stream
already exists when the client H1 connection is created, we create a H1
stream in ST_ATTACHED state, but not ST_READY, and the conn-stream is
attached to it. Because the ST_READY state is not set, no data are xferred
to the data layer when h1_rcv_buf() is called and shutdowns are inhibited
except on client aborts. This way, the request is parsed the same way than
for a classical H1 connection. Once the request headers are fully received
and parsed, the data stream is upgraded and the ST_READY state is set.
A tricky case appears when an H2 upgrade is performed because the H2 preface
is matched. In this case, the conn-stream must be detached and destroyed
before switching to the H2 mux and releasing the current H1 mux. We must
also take care to detach and destroy the conn-stream when a timeout
occurres.
This patch relies on the following series of patches :
* BUG/MEDIUM: stream: Don't immediatly ack the TCP to H1 upgrades
* MEDIUM: http-ana: Do nothing in wait-for-request analyzer if not htx
* MINOR: stream: Add a function to validate TCP to H1 upgrades
* MEDIUM: mux-h1: Add ST_READY state for the H1 connections
* MINOR: mux-h1: Wake up instead of subscribe for reads after H1C creation
* MINOR: mux-h1: Try to wake up data layer first before calling its wake callback
* MINOR: stream-int: Take care of EOS in the SI wake callback function
* BUG/MINOR: stream: Don't update counters when TCP to H2 upgrades are performed
This fix is specific for 2.4. No backport needed.
Instead of switching the stream to HTX mode, the request channel is only
reset (the request buffer is xferred to the mux) and the SF_IGNORE flag is
set on the stream. This flag prevent any processing in case of abort. Once
the upgrade confirmed, the flag is removed, in stream_upgrade_from_cs().
It is only the first part of the fix. The next one ("BUG/MAJOR: mux-h1:
Properly handle TCP to H1 upgrades") is also required. Both rely on the
following series of patches :
* MEDIUM: http-ana: Do nothing in wait-for-request analyzer if not htx
* MINOR: stream: Add a function to validate TCP to H1 upgrades
* MEDIUM: mux-h1: Add ST_READY state for the H1 connections
* MINOR: mux-h1: Wake up instead of subscribe for reads after H1C creation
* MINOR: mux-h1: Try to wake up data layer first before calling its wake callback
* MINOR: stream-int: Take care of EOS in the SI wake callback function
* BUG/MINOR: stream: Don't update counters when TCP to H2 upgrades are performed
This fix is specific for 2.4. No backport needed.
If http_wait_for_request() analyzer is called with a non-htx stream, nothing
is performed and we return immediatly. For now, it is totally unexpected.
But it will be true during TCP to H1 upgrades, once fixed. Indeed, there
will be a transition period during these upgrades. First the mux will be
upgraded and the not the stream, and finally the stream will be upgraded by
the mux once ready. In the meantime, the stream will still be in raw
mode. Nothing will be performed in wait-for-request analyzer because it will
be the mux responsibility to handle errors.
This patch is required to fix the TCP to H1 upgrades.
TCP to H1 upgrades are buggy for now. When such upgrade is performed, a
crash is experienced. The bug is the result of the recent H1 mux
refactoring, and more specifically because of the commit c4bfa59f1 ("MAJOR:
mux-h1: Create the client stream as later as possible"). Indeed, now the H1
mux is responsible to create the frontend conn-stream once the request
headers are fully received. Thus the TCP to H1 upgrade is a problem because
the frontend conn-stream already exists.
To fix the bug, we must keep this conn-stream and the associate stream and
use it in the H1 mux. To do so, the upgrade will be performed in two
steps. First, the mux is upgraded from mux-pt to mux-h1. Then, the mux-h1
performs the stream upgrade, once the request headers are fully received and
parsed. To do so, stream_upgrade_from_cs() must be used. This function set
the SF_HTX flags to switch the stream to HTX mode, it removes the SF_IGNORE
flags and eventually it fills the request channel with some input data.
This patch is required to fix the TCP to H1 upgrades and is intimately
linked with the next commits.
