A regression was introduced by commit 2421c6fa7d ("BUG/MEDIUM: stconn: Block
zero-copy forwarding if EOS/ERROR on consumer side"). When zero-copy
forwarding is inuse and the consumer side is shut or in error, we declare it
as blocked and it is woken up. The idea is to handle this state at the
stream-connector level. However this definitly blocks receives on the
producer side. So if the mux is unable to close by itself, but instead wait
the peer to shut, this can lead to a wake up loop. And indeed, with the
passthrough multiplexer this may happen.
To fix the issue and prevent any loop, instead of blocking the zero-copy
forwarding, we now disable it. This way, the stream-connector on producer
side will fallback on classical receives and will be able to handle peer
shutdown properly. In addition, the wakeup of the consumer side was
removed. This will be handled, if necessary, by sc_notify().
This patch should fix the issue #2395. It must be backported to 2.9.
When the producer side (h1 for now) negociates with the consumer side to
perform a zero-copy forwarding, we now consider the consumer side as blocked
if it is closed and this was reported to the SE via a end-of-stream or a
(pending) error.
It is performed before calling ->nego_ff callback function, in se_nego_ff().
This way, all consumer are concerned automatically. The aim of this patch is
to fix an issue with the QUIC mux. Indeed, it is unexpected to send a frame
on an closed stream. This triggers a BUG_ON(). Other muxes are not affected
but it remains useless to try to send data if the stream is closed.
This patch should fix the issue #2372. It must be backported to 2.9.
qcc_app_ops is a set of callbacks used to unify application protocol
running over QUIC. This commit introduces some changes to clarify its
API :
* write simple comment to reflect each callback purpose
* rename decode_qcs to rcv_buf as this name is more common and is
similar to already existing snd_buf
* finalize is moved up as it is used during connection init stage
All these changes are ported to HTTP/3 layer. Also function comments
have been extended to highlight HTTP/3 special characteristics.
This function is similar to the previous one, but this time for QCS
sending buffer.
Previously, each application layer redefine their own version of
mux_get_buf() which was used to allocate <qcs.tx.buf>. Unify it under a
single function renamed qcc_get_stream_txbuf().
Replaces qcs_get_buf() function which naming does not reflect its
purpose. Add a new function qcc_get_stream_rxbuf() which allocate if
needed <qcs.rx.app_buf> and returns the buffer pointer. This function is
reserved for application protocol layer. This buffer is then accessed by
stconn layer.
For other qcs_get_buf() invocation which was used in effect for a local
buffer, replace these by a plain b_alloc().
"set severity-output" is one of these command that changes the appctx
state so the next commands are affected.
Unfortunately the master CLI works with pipelining and server close
mode, which means the connection between the master and the worker is
closed after each response, so for the next command this is a new appctx
state.
To fix the problem, 2 new flags are added ACCESS_MCLI_SEVERITY_STR and
ACCESS_MCLI_SEVERITY_NB which are used to prefix each command sent to
the worker with the right "set severity-output" command.
This patch fixes issue #2350.
It could be backported as far as 2.6.
Before this patch, it was not possible to use a list of patterns, map or a
list of acls, without an existing file. However, it could be handy to just
use an ID, with no file on the disk. It is pretty useful for everyone
managing dynamically these lists. It could also be handy to try to load a
list from a file if it exists without failing if not. This way, it could be
possible to make a cold start without any file (instead of empty file),
dynamically add and del patterns, dump the list to the file periodically to
reuse it on reload (via an external process).
In this patch, we uses some prefixes to be able to use virtual or optional
files.
The default case remains unchanged. regular files are used. A filename, with
no prefix, is used as reference, and it must exist on the disk. With the
prefix "file@", the same is performed. Internally this prefix is
skipped. Thus the same file, with ou without "file@" prefix, references the
same list of patterns.
To use a virtual map, "virt@" prefix must be used. No file is read, even if
the following name looks like a file. It is just an ID. The prefix is part
of ID and must always be used.
To use a optional file, ie a file that may or may not exist on a disk at
startup, "opt@" prefix must be used. If the file exists, its content is
loaded. But HAProxy doesn't complain if not. The prefix is not part of
ID. For a given file, optional files and regular files reference the same
list of patterns.
This patch should fix the issue #2202.
tune.cache.zero-copy-forwarding parameter can now be used to enable or
disable the zero-copy fast-forwarding for the cache applet only. It is
enabled ('on') by default. It can be disabled by setting the parameter to
'off'.
For now, CF_STREAMER and CF_STREAMER_FAST flags are set in sc_conn_recv()
function. The logic is moved in dedicated functions.
First, channel_check_idletimer() function is now responsible to check the
channel's last read date against the idle timer value to be sure the
producer is still streaming data. Otherwise, it removes STREAMER flags.
Then, channel_check_xfer() function is responsible to check amount of data
transferred avec a receive, to eventually update STREAMER flags.
In sc_conn_recv(), we now use these functions.
Zero-copy fast-forwading feature is a quite new and is a bit sensitive.
There is an option to disable it globally. However, all protocols have not
the same maturity. For instance, for the PT multiplexer, there is nothing
really new. The zero-copy fast-forwading is only another name for the kernel
splicing. However, for the QUIC/H3, it is pretty new, not really optimized
and it will evolved. And soon, the support will be added for the cache
applet.
In this context, it is usefull to be able to enable/disable zero-copy
fast-forwading per-protocol and applet. And when it is applicable, on sends
or receives separately. So, instead of having one flag to disable it
globally, there is now a dedicated bitfield, global.tune.no_zero_copy_fwd.
It is possible that a server's addr family is temporarily set to AF_UNSPEC
even if we're certain to be in INET context (ipv4, ipv6).
Indeed, as soon as IP address resolving is involved, srv->addr family will
be set to AF_UNSPEC when the resolution fails (could happen at anytime).
However, _srv_event_hdl_prepare_inetaddr() wrongly assumed that it would
only be called with AF_INET or AF_INET6 families. Because of that, the
function will handle AF_UNSPEC address as an IPV6 address: not only
we could risk reading from an unititialized area, but we would then
propagate false information when publishing the event.
In this patch we make sure to properly handle the AF_UNSPEC family in
both the "prev" and the "next" part for SERVER_INETADDR event and that
every members are explicitly initialized.
This bug was introduced by 6fde37e046 ("MINOR: server/event_hdl: add
SERVER_INETADDR event"), no backport needed.
preferred_address is a transport parameter specify by the server. It
specified both an IPv4 and IPv6 address. These addresses were defined as
plain array in <struct tp_preferred_address>.
Convert these adressees to use the common types in_addr/in6_addr. With
this change, dumping of preferred_address is extended. It now displays
the addresses using inet_ntop() and CID value.
retrieve_qc_conn_from_cid() requires listener as argument whereas it is
unused. This is an artifact from the old architecture where CID trees
where stored on listener instances instead of globally. Remove it to
better reflect this change.
Just like the ->ctl() callback function, used to send commands to mux
connections, the ->sctl() callback function can now be used to send commands
to mux streams. The first command, MUX_SCTL_SID, is a way to request the mux
stream ID.
It will be implemented later for each mux.
Instead of the generic MUX_, we now use MUX_CTL_ prefix for all mux_ctl_type
value. This will avoid any ambiguities with other enums, especially with a
new one that will be added to get information on mux streams.
It is now possible to retrieve the session terminate state, using
"txn.sess_term_state". The sample fetch returns the 2-character session
termation state.
Of course, the result of this sample fetch is volatile. It is subject to
change. It is also most of time useless because no termation state is set
except at the end. It should only be useful in http-after-response rule
sets. It may also be used to customize the logs using a log-format
directive.
This patch should fix the issue #2221.
The code returned by the "status" sample fetch is the one in the HTTP
response at the moment the sample is evaluated. It may be the status code in
the server response or the one of the HAProxy reply in case of error, deny,
redirect...
However, it could be handy to retrieve the status code returned by the
server, when a HTTP response was really received from it. It is the purpose
of the "server_status" sample fetch. The server status code itself is stored
in the HTTP txn.
In previous log backend implementation, we created a pseudo log target
for each declared log server, and we made the log target's address point
to the actual server address to save some time and prevent unecessary
copies.
But this was done without knowing that when FQDN is involved (more broadly
when dns/resolution is involved), the "port" part of server addr should
not be relied upon, and we should explicitly use ->svc_port for that
purpose.
With that in mind and thanks to the previous commit, some changes were
required: we allocate a dedicated addr within the log target when target
is in DGRAM mode. The addr is first initialized with known values and it
is then updated automatically by _srv_set_inetaddr() during runtime.
(the change is atomic so readers don't need to worry about it)
addr from server "log target" (INET/DGRAM mode) is made of the combination
of server's address (lacking the port part) and server's svc_port.
The local variable "event_hdl_async_max_notif_at_once" which was
introduced with the event_hdl API was left as is but with a TODO note
telling that we should make it a global tunable.
Well, we're doing this now. To prepare for upcoming tunables related to
event_hdl API, we add a dedicated struct named event_hdl_tune which is
globally exposed through the event_hdl header file so that it may be used
from everywhere. The struct is automatically initialized in
event_hdl_init() according to defaults.h.
"event_hdl_async_max_notif_at_once" now becomes
"event_hdl_tune.max_events_at_once" with it's dedicated
configuation keyword: "tune.events.max-events-at-once".
We're also taking this opportunity to raise the default value from 10
to 100 since it's seems quite reasonnable given existing async event_hdl
users.
The documentation was updated accordingly.
Implements the customized payload pattern for the master CLI.
The pattern is stored in the stream in char pcli_payload_pat[8].
The principle is basically the same as the CLI one, it looks for '<<'
then stores what's between '<<' and '\n', and look for it to exit the
payload mode.
The CLI payload syntax has some limitation, it can't handle payloads
with empty lines, which is a common problem when uploading a PEM file
over the CLI.
This patch implements a way to customize the ending pattern of the CLI,
so we can't look for other things than empty lines.
A char cli_payload_pat[8] is used in the appctx to store the customized
pattern. The pattern can't be more than 7 characters and can still empty
to match an empty line.
The cli_io_handler() identifies the pattern and stores it, and
cli_parse_request() identifies the end of the payload.
If the customized pattern between "<<" and "\n" is more than 7
characters, it is not considered as a pattern.
This patch only implements the parser for the 'stats socket', another
patch is needed for the 'master CLI'.
Move qc_notify_send() from quic_tx.c to quic_conn.c. Note that it was already
exported from both quic_conn.h and quic_tx.h. Modify this latter header
to fix the duplication.
Such a warning appeared after having added quic_retry.h which includes only
headers for types (quic_cid-t.h, clock-t.h...)
In file included from include/haproxy/quic_retry.h:12,
from src/quic_retry.c:5:
include/haproxy/quic_cid-t.h:26:26: error: field ‘seq_num’ has incomplete type
26 | struct eb64_node seq_num;
Add quic_retry.c new C file for the QUIC retry feature:
quic_saddr_cpy() moved from quic_tx.c,
quic_generate_retry_token_aad() moved from
quic_generate_retry_token() moved from
parse_retry_token() moved from
quic_retry_token_check() moved from
quic_retry_token_check() moved from
This function is in relation with the Initial packet number space which is
more linked to the QUIC TLS specifications. Let's move it to quic_tls.h
to be inlined.
Move quic_path struct from quic_conn-t.h to quic_cc-t.h and rename it to quic_cc_path.
Update the code consequently.
Also some inlined functions in relation with QUIC path to quic_cc.h
Move quic_pkt_type(), quic_saddr_cpy(), quic_write_uint32(), max_available_room(),
max_stream_data_size(), quic_packet_number_length(), quic_packet_number_encode()
and quic_compute_ack_delay_us() to quic_tx.c because only used in this file.
Also move quic_ack_delay_ms() and quic_read_uint32() to quic_tx.c because they
are used only in this file.
Move quic_rx_packet_refinc() and quic_rx_packet_refdec() to quic_rx.h header.
Move qc_el_rx_pkts(), qc_el_rx_pkts_del() and qc_list_qel_rx_pkts() to quic_tls.h
header.
Move quic_cstream struct definition from quic_conn-t.h to quic_tls-t.h.
Its pool is also moved from quic_conn module to quic_tls. Same thing for
quic_cstream_new() and quic_cstream_free().
Fix such building issues:
In file included from src/quic_tx.c:15:
include/haproxy/quic_tx.h:51:23: warning: ‘struct quic_rx_packet’
Do not know why the compiler warns about such missing header inclusions
just now. It should have complained a long time ago during the big QUIC
source code split.
Move quic_cid and quic_connnection_id from quic_conn-t.h to new quic_cid-t.h header.
Move defintions of quic_stateless_reset_token_init(), quic_derive_cid(),
new_quic_cid(), quic_get_cid_tid() and retrieve_qc_conn_from_cid() to quic_cid.c
new C file.
When zero-copy data fast-forwarding is inuse, if the opposite SC is blocked,
there is no reason to try to fast-forward more data. Worst, in some cases,
this can lead to a receive loop of the producer side while the consumer side
is blocked.
No backport needed.
Add an optional argument for "-dt". This argument is interpreted as a
list of several trace statement separated by comma. For each statement,
a specific trace name can be specifed, or none to act on all sources.
Using double-colon separator, it is possible to add specifications on
the wanted level and verbosity.
Add '-dt' haproxy process argument. This will automatically activate all
trace sources on stderr with the error level. This could be useful to
troubleshoot issues such as protocol violations.
In the pat_ref_elt struct, the pattern string is stored outside of the
node element, using a pointer to an strdup(). Not only this needlessly
wastes at least 16-24 bytes per entry (8 for the pointer, 8-16 for the
allocator), it also makes the tree descent less efficient since both
the node and the string have to be visited for each layer (hence at least
two cache lines). Let's use an ebmb storage and place the pattern right
at the end of the pat_ref_elt, making it a variable-sized element instead.
The set-map test below jumps from 173 to 182 kreq/s/core, and the memory
usage drops from 356 MB to 324 MB:
http-request set-map(/dev/null) %[rand(1000000)] 1
This is even more visible with large maps: after loading 16M IP addresses
into a map, the process uses this amount of memory:
- 3.15 GB with haproxy-2.8
- 4.21 GB with haproxy-2.9-dev11
- 3.68 GB with this patch
So that's a net saving of 32 bytes per entry here, which cuts in half the
extra cost of the tree, and loading a large map takes about 20% less time.
Task_drop_running() is used to remove the RUNNING bit and check if
while the task was running it got a new wakeup from itself. Thus
each time task_drop_running() marks itself as a caller, it in fact
removes the previous caller that woke up the task, such as below:
Tasks activity over 10.439 sec till 0.000 sec ago:
function calls cpu_tot cpu_avg lat_tot lat_avg
task_run_applet 57895273 6.396m 6.628us 2.733h 170.0us <- run_tasks_from_lists@src/task.c:658 task_drop_running
Better not mark this function as a caller and keep the original one:
Tasks activity over 13.834 sec till 0.000 sec ago:
function calls cpu_tot cpu_avg lat_tot lat_avg
task_run_applet 62424582 5.825m 5.599us 5.717h 329.7us <- sc_app_chk_rcv_applet@src/stconn.c:952 appctx_wakeup
The mworker mode never had a proper 'hard-stop' (-st) for the reload,
this is a mode which was commonly used with the daemon mode, but it was
never implemented in mworker mode.
This patch fixes the problem by implementing a "hard-reload" command
over the master CLI. It does the same as the "reload" command, but
instead of waiting for the connections to stop in the previous process,
it immediately quits the previous process after binding.
Take the px->server_rules freeing part out of free_proxy() and make it
a dedicated helper function so that it becomes possible to use it from
anywhere.
In this patch we fix the prototype for ipcmp() and ipcpy() functions so
that input pointers that are used exclusively for reads are used as const
pointers. This way, the compiler can safely assume that those variables
won't be altered by the function.
In this patch we add the support for a new SERVER event in the
event_hdl API.
SERVER_INETADDR is implemented as an advanced server event.
It is published each time the server's ip address or port is
about to change. (ie: from the cli, dns, lua...)
SERVER_INETADDR data is an event_hdl_cb_data_server_inetaddr struct
that provides additional info related to the server inet addr change,
but can be casted as a regular event_hdl_cb_data_server struct if
additional info is not needed.
When possible, we try send DATA frame without copying data. To do so, we
swap the input buffer with QCS tx buffer. It is only possible iff:
* There is only one HTX block of data at the beginning of the message
* Amount of data to send is equal to the size of the HTX data block
* The QCS tx buffer is empty
In this case, both buffers are swapped. The frame metadata are written at
the begining of the buffer, before data and where the HTX structure is
stored.
Add a new member <nb_rhttp_conns> in thread_ctx structure. Its purpose
is to count the current number of opened reverse HTTP connections
regarding from their listeners membership.
This patch will be useful to support multi-thread for active reverse
HTTP, in order to select the less loaded thread.
Note that despite access to <nb_rhttp_conns> are only done by the
current thread, atomic operations are used. This is because once
multi-thread support will be added, external threads will also retrieve
values from others.
Previous commit renames 'proto_reverse_connect' module to 'proto_rhttp'.
This commits follows this by replacing various custom prefix by 'rhttp_'
to make the code uniform.
Note that 'reverse_' prefix was kept in connection module. This is
because if a new reversable protocol not based on HTTP is implemented,
it may be necessary to reused the same connection function which are
protocol agnostic.
In Github issue #2128, @jvincze84 explained the complexity of using
external checks in some advanced setups due to the systematic purge of
environment variables, and expressed the desire to preserve the
existing environment. During the discussion an agreement was found
around having an option to "external-check" to do that and that
solution was tested and confirmed to work by user @nyxi.
This patch just cleans this up, implements the option as
"preserve-env" and documents it. The default behavior does not change,
the environment is still purged, unless "preserve-env" is passed. The
choice of not using "import-env" instead was made so that we could
later use it to name specific variables that have to be imported
instead of keeping the whole environment.
The patch is simple enough that it could be backported if needed (and
was in fact tested on 2.6 first).
Here the idea is to collect components' versions and build options. The
main component is haproxy, but the API is made so that any sub-system
can easily add a component there (for example the detailed version of a
device detection lib, or some info about a lib loaded from Lua).
The elements are stored as a pointer to an array of structs and its count
so that it's sufficient to issue this in gdb to list them all at once:
print *post_mortem.components@post_mortem.nb_components
For now we collect name, version, toolchain, toolchain options, build
options and path. Maybe more could be useful in the future.
When debugging a core, it's difficult to match a given gdb thread number
against an internal thread. Let's just store the pthread ID and the stack
pointer in each tinfo. This could help in the future by allowing to just
glance over them and pick the right one depending what info is found
first.
Add missing CO_FL_REVERSED and CO_FL_ACT_REVERSING flag definitions in
conn_show_flags(). These flags were introduced in this release with
reverse HTTP support.
No need to backport
On CONNECTION_CLOSE reception/emission, QUIC connections enter CLOSING
state. At this stage, only CONNECTION_CLOSE can be reemitted and all
other exchanges are stopped.
Previously, on haproxy process stopping, if all QUIC connections were in
CLOSING state, they were released before their closing timer expiration
to not block the process shutdown. However, since a recent commit, the
closing timer has been shorten to a more reasonable delay. It is now
consider viable to respect connections closing state even on process
shutdown. As such, stopping specific code in QUIC connections idle timer
task was removed.
A specific function quic_handle_stopping() was implemented to notify
QUIC connections on shutdown from main() function. It should have been
deleted along the removal in QUIC idle timer task. This patch just does
this.
In 2.3, we started to get a cleaner socket unbinding mechanism with
commit f58b8db47 ("MEDIUM: receivers: add an rx_unbind() method in
the protocols"). This mechanism rightfully refrains from unbinding
when sockets are expected to be transferrable to another worker via
"expose-fd listeners", but this is not compatible with ABNS sockets,
which do not support reuseport, unbinding nor being renamed: in short
they will always prevent a new process from binding.
It turns out that this is not much visible because by pure accident,
GTUNE_SOCKET_TRANSFER is only set in the code dealing with master mode
and deamons, so it's never set in foreground mode nor in tests even if
present on the stats socket. However with master mode, it is now always
set even when not present on the stats socket, and will always conflict.
The only reasonable approach seems to consist in marking these abns
sockets as non-suspendable so that the generic sock_unbind() code can
decide to just unbind them regardless of GTUNE_SOCKET_TRANSFER.
This should carefully be backported as far as 2.4.
add proxy_cfg_ensure_no_log() function (similar to
proxy_cfg_ensure_no_http()) to ensure at the end of proxy parsing that
no log exclusive options are found if the proxy is not in log mode.
"log-balance" directive was recently introduced to configure the
balancing algorithm to use when in a log backend. However, it is
confusing and it causes issues when used in default section.
In this patch, we take another approach: first we remove the
"log-balance" directive, and instead we rely on existing "balance"
directive to configure log load balancing in log backend.
Some algorithms such as roundrobin can be used as-is in a log backend,
and for log-only algorithms, they are implemented as "log-$name" inside
the "backend" directive.
The documentation was updated accordingly.
Make sure lbprm.algo can store 32bits by declaring it as uint32_t
Then, use all 32 available bits to offer 4 extra bits for the BE_LB_NEED
inputs. This will allow new required inputs to be easily added (up to 4
new ones, plus one that wasn't used yet if we keep them exclusive)
This required some cleanup: all ALGO bitfields were rewritten in the
32bits format and the high ones were shifted to make room for the
new BE_LB_NEED bits.
BE_LB_HASH_RND was introduced with 760e81d35 ("MINOR: backend: implement
random-based load balancing") but was never used since. Removing it
to regain an extra slot for future types.
A dummy connect() function previously had to be installed for the log
server so that a reverse-http address could be referenced on a "server"
line, but after the recent rework of the server line parsing, this is
no longer needed, and this is actually annoying as it makes one believe
there is a way to connect outside, which is not true. Let's now get rid
of this function.
The idle timer task may be used to trigger the client handshake timeout.
The hanshake timeout expiration date (qc->hs_expire) is initialized when the
connection is allocated. Obviously, this timeout is taken into an account only
during the handshake by qc_idle_timer_do_rearm() whose job is to rearm the idle timer.
The idle timer expiration date could be initialized only one time, then
never updated until the hanshake completes. But this only works if the
handshake timeout is smaller than the idle timer task timeout. If the handshake
timeout is set greater than the idle timeout, this latter may expire before the
handshake timeout.
This patch may have an impact on the L1/C1 interop tests (with heavy packet loss
or corruption). This is why I guess some implementations with a hanshake timeout
support set a big timeout during this test. This is at least the case for ngtcp2
which sets a 180s hanshake timeout! haproxy will certainly have to proceed the
same way if it wants to have a chance to pass this test as before this handshake
timeout.
Add a new timeout for the handshake, on the frontend side only. Such a hanshake
will be typically used for TLS hanshakes during client connections to TLS/TCP or
QUIC frontends.
Partial sends is an activity, not a full blocking. Thus a read activity must
be reported for non-independent stream. It is especially important for very
congested stream where full sends are uncommon.
This patch must be backported to 2.8.
This patch adds HXT-aware versions of the functions c_data(), ci_data() and
c_empty(). channel_data() function returns the amount of data in the
channel, channel_input_data() returns the amount of input data and
channel_empty() returns true if the channel's buffer is empty. These
functions handles HTX buffers.
In addition, channel_data_limit() function, still HTX-aware, can be used to
get the maximum absolute amount of data that can be copied in a buffer,
independently on data already present in the buffer.
The overhead induced by the HTX format was set to the HTX structure itself
and two HTX blocks. It was set this way to optimize zero-copy during
transfers. This value may (and will) be used at different places. Thus we
now use a macro, called HTX_BUF_OVERHEAD.
The first-send-blocked date was originally designed to save the date of the
first send of a series where some data remain blocked. It was relaxed
recently (3083fd90e "BUG/MEDIUM: stconn: Report a send activity everytime
data were sent") to save the date of the first full blocked send. However,
it is not accurrate.
When all data are sent, the fsb value must be reset to TICK_ETERNITY. When
nothing is sent and if it is not already set, it must be set. But when data
are partially sent, the value must be updated and not reset. Otherwise the
write timeout may be ignored because fsb date is never set.
So, changes brought by the patch above are reverted and
sc_ep_report_blocked_send() was changed to know if some data were sent or
not. This way we are able to update fsb value.
l
This patch must be backported to 2.8.
This global variable was used to avoid using locks on shared_contexts in
the unlikely case of nbthread==1. Since the locks do not do anything
when USE_THREAD is not defined, it will be more beneficial to simply
remove this variable and the systematic test on its value in the shared
context locking functions.
A reference counter on the cache_entry was added in a previous commit.
Its value is atomically increased and decreased via the retain_entry and
release_entry functions.
This is needed because of the latest cache and shared_context
modifications that introduced two separate locks instead of the
preexisting single shctx_lock one.
With the new logic, we have two main blocks competing for the two locks:
- the one in the http_action_req_cache_use that performs a lookup in the
cache tree (locked by the cache lock) and then tries to remove the
corresponding blocks from the shared_context's 'avail' list until the
response is sent to the client by the cache applet,
- the shctx_row_reserve_hot that traverses the 'avail' list and gives
them back to the caller, while removing previous row heads from the
cache tree
Those two blocks require the two locks but one of them would take the
cache lock first, and the other one the shctx_lock first, which would
end in a deadlock without the current patch.
The way this conflict is resolved in this patch is by ensuring that at
least one of those uses works without taking the two locks at the same
time.
The solution found was to keep taking the two locks in the cache_use
case. We first lock the cache to lookup for an entry and we then take
the shctx lock as well to detach the corresponding blocks from the
'avail' list. The subtlety is that between the cache lookup and the
actual locking of the shctx, another thread might have called the
reserve_hot function in which we only take the shctx lock.
In this function we traverse the 'avail' list to remove blocks that are
then given to the caller. If one of those blocks corresponds to a
previous row head, we call the 'free_blocks' callback that used to
delete the cache entry from the tree.
We now avoid deleting directly the cache entries in reserve_hot and we
rather set the cache entries 'complete' param to 0 so that no other
thread tries to work with this entry. This way, when we release the
shctx lock in reserve_hot, the first thread that had performed the cache
lookup and had found an entry that we just gave to another thread will
see that the 'complete' field is 0 and it won't try to work with this
response.
The actual removal of entries from the cache tree will now be performed
in the new 'reserve_finish' callback called at the end of the
shctx_row_reserve_hot function. It will iterate on all the row head that
were inserted in a dedicated list in the 'free_block' callback and
perform the actual delete.
This patch adds a reserve_finish callback that can be defined by the
subsystems that require a shared_context. It is called at the end of
shctx_row_reserve_hot after the shared_context lock is released.
Since a lock on the cache tree was added in the latest cache changes, we
do not need to use the shared_context's lock to lock more than pure
shared_context related data anymore. This already existing lock will now
only cover the 'avail' list from the shared_context. It can then be
changed to a rwlock instead of a spinlock because we might want to only
run through the avail list sometimes.
Apart form changing the type of the shctx lock, the main modification
introduced by this patch is to limit the amount of code covered by the
shctx lock. This lock does not need to cover any code strictly related
to the cache tree anymore.
The "hot" list stored in a shared_context was used to keep a reference
to shared blocks that were currently being used and were thus removed
from the available list (so that they don't get reused for another cache
response). This 'hot' list does not ever need to be shared across
threads since every one of them only works on their current row.
The main need behind this 'hot' list was to detach the corresponding
blocks from the 'avail' list and to have a known list root when calling
list_for_each_entry_from in shctx_row_data_append (for instance).
Since we actually never need to iterate over all members of the 'hot'
list, we can remove it and replace the inc_hot/dec_hot logic by a
detach/reattach one.
Every use of the cache tree was covered by the shctx lock even when no
operations were performed on the shared_context lists (avail and hot).
This patch adds a dedicated RW lock for the cache so that blocks of code
that work on the cache tree only can use this lock instead of the
superseding shctx one. This is useful for operations during which the
concerned blocks are already in the hot list.
When the two locks need to be taken at the same time, in
http_action_req_cache_use and in shctx_row_reserve_hot, the shctx one
must be taken first.
A new parameter needed to be added to the shared_context's free_block
callback prototype so that cache_free_block can take the cache lock and
release it afterwards.
Change the flags used for reversed connection :
* CO_FL_REVERSED is now put after reversal for passive connect. For
active connect, it is delayed when accept is completed after reversal.
* CO_FL_ACT_REVERSING replace the old CO_FL_REVERSED. It is put only for
active connect on reversal and removes once accept is done.
This allows to identify a connection as reversed during its whole
lifetime. This should be useful to extend reverse connect.
Make all the congestion support the maximum congestion control window
set by configuration. There is nothing special to explain. For each
each algo, each time the window is incremented it is also bounded.
Add a new ->max_cwnd member to bind_conf struct to store the maximum
congestion control window value for each QUIC binding.
Modify the "quic-cc-algo" keyword parsing to add an optional parameter
to its value: the maximum congestion window value between parentheses
as follows:
ex: quic-cc-algo cubic(10m)
This value must be bounded, greater than 10k and smaller than 1g.
This bug was introduced with 29b76ca ("BUG/MEDIUM: server/log: "mode log"
after server keyword causes crash ")
Indeed, we cannot safely rely on addr_proto being set when str2sa_range()
returns in parse_server() (even if SRV_PARSE_PARSE_ADDR is set), because
proto lookup might be bypassed when FQDN addresses are involved.
Unfortunately, the above patch wrongly assumed that proto would always
be set when SRV_PARSE_PARSE_ADDR was passed to parse_server() (so when
str2sa_range() was called), resulting in invalid postparsing checks being
performed, which could as well lead to crashes with log backends
("mode log" set) because some postparsing init was skipped as a result of
proto not being set and this wasn't expected later in the init code.
To fix this, we now make use of the previous patch to perform server's
address compatibility checks on hints that are always set when
str2sa_range() succesfully returns.
For log backend, we're also adding a complementary test to check if the
address family is of expected type, else we report an error, plus we're
moving the postinit logic in log api since _srv_check_proxy_mode() is
only meant to check proxy mode compatibility and we were abusing it.
This patch depends on:
- "MINOR: tools: make str2sa_range() directly return type hints"
No backport required unless 29b76ca gets backported.
str2sa_range() already allows the caller to provide <proto> in order to
get a pointer on the protocol matching with the string input thanks to
5fc9328a ("MINOR: tools: make str2sa_range() directly return the protocol")
However, as stated into the commit message, there is a trick:
"we can fail to return a protocol in case the caller
accepts an fqdn for use later. This is what servers do and in this
case it is valid to return no protocol"
In this case, we're unable to return protocol because the protocol lookup
depends on both the [proto type + xprt type] and the [family type] to be
known.
While family type might not be directly resolved when fqdn is involved
(because family type might be discovered using DNS queries), proto type
and xprt type are already known. As such, the caller might be interested
in knowing those address related hints even if the address family type is
not yet resolved and thus the matching protocol cannot be looked up.
Thus in this patch we add the optional net_addr_type (custom type)
argument to str2sa_range to enable the caller to check the protocol type
and transport type when the function succeeds.
It's common to see process_stream() being woken up by wake_expired_tasks
in the profiling output, without knowing which timeout was set to cause
this. By making it possible to record the call places of task_queue()
and task_schedule(), and by making wake_expired_tasks() explicitly not
replace it, we'll be able to know which task_queue() or task_schedule()
was triggered for a given wakeup.
For example below:
process_stream 51200 311.4ms 6.081us 34.59s 675.6us <- run_tasks_from_lists@src/task.c:659 task_queue
process_stream 19227 70.00ms 3.640us 9.813m 30.62ms <- sc_notify@src/stconn.c:1136 task_wakeup
process_stream 6414 102.3ms 15.95us 8.093m 75.70ms <- stream_new@src/stream.c:578 task_wakeup
It's visible that it's the run_tasks_from_lists() which in fact applies
on the task->expire returned by the ->process() function itself.
This is used for tracing and profiling. By permitting to have a NULL
caller, we allow a caller to explicitly pass zero to state that the
current caller must not be replaced. This will soon be used by
wake_expired_tasks() to avoid replacing a caller in the expire loop.
QUIC connections are pushed manually into a dedicated listener queue
when they are ready to be accepted. This happens after handshake
finalization or on 0-RTT packet reception. Listener is then woken up to
dequeue them with listener_accept().
This patch comptabilizes the number of connections currently stored in
the accept queue. If reaching a certain limit, INITIAL packets are
dropped on reception to prevent further QUIC connections allocation.
This should help to preserve system resources.
This limit is automatically derived from the listener backlog. Half of
its value is reserved for handshakes and the other half for accept
queues. By default, backlog is equal to maxconn which guarantee that
there can't be no more than maxconn connections in handshake or waiting
to be accepted.
Implement a limit per listener for concurrent number of QUIC
connections. When reached, INITIAL packets for new connections are
automatically dropped until the number of handshakes is reduced.
The limit value is automatically based on listener backlog, which itself
defaults to maxconn.
This feature is important to ensure CPU and memory resources are not
consume if too many handshakes attempt are started in parallel.
Special care is taken if a connection is released before handshake
completion. In this case, counter must be decremented. This forces to
ensure that member <qc.state> is set early in qc_new_conn() before any
quic_conn_release() invocation.
Accounting is implemented for half open connections which represent QUIC
connections waiting for handshake completion. When reaching a certain
limit, Retry mechanism is automatically activated prior to instantiate
new connections.
The issue with this behavior is that two notions are mixed : QUIC
connection handshake phase and Retry which is mechanism against
amplification attacks. As such, only peer address validation should be
taken into account to activate Retry protection.
This patch chooses to reduce the scope of half_open_conn. Now only
connection waiting to validate the peer address are now accounted for.
Most notably, connections instantiated with a validated Retry token
check are not accounted.
One impact of this patch is that it should prevent to activate Retry
mechanism too early, in particular in case if multiple handshakes are
too slow. Another limitation should be implemented to protect against
this scenario.
When a new QUIC connection is created, server considers peer address as
not yet validated. The server must limit its sending up to 3 times the
content already received. This is a defensive measure to avoid flooding
a remote host victim of address spoofing.
This patch adjust the condition to consider the peer address as
validated. Two conditions are now considered :
* successful handling of a received HANDSHAKE packet. This was already
done before although implemented in a different way.
* validation of a Retry token. This was not considered prior this patch
despite RFC recommandation.
This patch also adjusts how a connection is internally labelled as using
a validated peer address. Before, above conditions were checked via
quic_peer_validated_addr(). Now, a flag QUIC_FL_CONN_PEER_VALIDATED_ADDR
is set to labelled this. It already existed prior this patch but was
only used for quic_cc_conn. This should now be more explicit.
This bug could be reproduced with -dMfail option and detected by libasan.
During the TLS secrets allocations, when failed, quic_tls_ctx_secs_free()
is called. It resets the already initialized secrets. Some were detected
as initialized when not, or with a non initialized length, which leads
to big "memset(0)" detected by libsasan.
Ensure that all the secrets are really initialized with correct lengths.
No need to be backported.
This was no reason not to release as soon as possible the TLS/SSL QUIC connection
context from quic_conn_release() before allocating a "closing connection" connection
(quic_cc_conn struct).
This patch sets the handshake task in heavy task mode when receiving in disorder
CRYPTO data which results in in order bufferized CRYPTO data. This is done
thanks to a non-contiguous buffer and from qc_handle_crypto_frm() after having
potentially bufferized CRYPTO data in this buffer.
qc_treat_rx_crypto_frms() is no more called from qc_treat_rx_pkts() but instead
this is where the task is set in heavy task mode. Consequently,
this is the job of qc_ssl_provide_all_quic_data() to call directly
qc_treat_rx_crypto_frms() to provide the in order bufferized CRYPTO data to the
TLS stack. As this function releases the non-contiguous buffer for the CRYPTO
data, if possible, there is no need to do that from qc_treat_rx_crypto_frms()
anymore.
Add a new pool for the CRYPTO data frames received in order.
Add ->rx.crypto_frms list to each encryption level to store such frames
when they are received in order from qc_handle_crypto_frm().
Also set the handshake task (qc_conn_io_cb()) in heavy task mode from
this function after having received such frames. When this task
detects that it is set in heavy mode, it calls qc_ssl_provide_all_quic_data()
newly implemented function to provide the CRYPTO data to the TLS task.
Modify quic_conn_enc_level_uninit() to release these CRYPTO frames
when releasing the encryption level they are in relation with.
IOBUF_FL_EOI iobuf flag is now set by the producer to notify the consumer
that the end of input was reached. Thanks to this flag, we can remove the
ugly ack in h2_done_ff() to test the opposite SE flags.
Of course, for now, it works and it is good enough. But we must keep in mind
that EOI is always forwarded from the producer side to the consumer side in
this case. But if this change, a new CO_RFL_ flag will have to be added to
instruct the producer if it can forward EOI or not.
In the mux-to-mux data forwarding, we now try, as far as possible to send at
least a buffer. Of course, if the consumer side is congested or if nothing
more can be received, we leave. But the idea is to retry to fast-forward
data if less than a buffer was forwarded. It is only performed for buffer
fast-forwarding, not splicing.
The idea behind this patch is to optimise the forwarding, when a first
forward was performed to complete a buffer with some existing data. In this
case, the amount of data forwarded is artificially limited because we are
using a non-empty buffer. But without this limitation, it is highly probable
that a full buffer could have been sent. And indeed, with H2 client, a
significant improvement was observed during our test.
To do so, .done_fastfwd() callback function must be able to deal with
interim forwards. Especially for the H2 mux, to remove H2_SF_NOTIFIED flags
on the H2S on the last call only. Otherwise, the H2 stream can be blocked by
itself because it is in the send_list. IOBUF_FL_INTERIM_FF iobuf flag is
used to notify the consumer it is not the last call. This flag is then
removed on the last call.
In 2.6-dev1, the method used to decide how many pool entries could be
released at once was revisited to support releases in batches. This was
done with commits 91a8e28f9 ("MINOR: pool: add a function to estimate
how many may be released at once") and 361e31e3f ("MEDIUM: pool: compute
the number of evictable entries once per pool").
The first commit takes care of the possible inconsistency between the
moment the allocated count and the used count are read, but unfortunately
fixed it the wrong way, by adjusting "used" to match "alloc" whenever it
was lower (i.e. almost always). This results in a nasty case which is that
as soon as the allocated value becomes higher than the estimated count of
needed entries, we end up returning pool->minavail, which causes very
small batches to be released, starting from commit 1513c5479 ("MEDIUM:
pools: release cached objects in batches").
The problem was further amplified in 2.9-dev3 with commit 7bf829ace
("MAJOR: pools: move the shared pool's free_list over multiple buckets")
because it now becomes possible for a thread to allocate from one bucket
and release into a few other different ones, causing an accumulation of
entries in that bucket.
The fix is trivial, simply adjust the alloc counter if the used one is
higher, before performing operations.
This must be backported to 2.6.
QUIC connections are accounted inside global sslconns. As with QUIC
actconn, it suffered from a similar issue if an intermediary allocation
failed inside qc_new_conn().
Fix this similarly by moving increment operation inside qc_new_conn().
Increment and error path are now centralized and much easier to
validate.
The consequences are similar to the actconn fix : on memory allocation
global sslconns may wrap, this time blocking any future QUIC or SSL
connections on the process.
This must be backported up to 2.6.
Recent fixes have shown <lra> and <fsb> uses were not prettu clear. So let's
try to improve documentation about these value. Especially when <lra> is
updated and how to used it.
In sc_need_room(), we compute the maximum room that can be requested to
restarted reading to be sure to be able to unblock the SC. At worst when the
buffer is emptied. Here, the buffer reserve is considered but it is an issue.
Counting the reserve can lead to a wicked bug with the H1 multiplexer, when
small amount of data are found at the end of the HTX buffer. In this case,
to not wrap, the H1 mux requests more room. It is an optim to be able to
resync the buffer with the consumer side and to be able to perform zero-copy
transfers. However, if this amount of data is smaller than the reserve and
if the consumer is congested, we fall in a loop because the wrong value is
used to request more room. The H1 mux continues to pretend there is not
enough space in the buffer, while the effective requested value is lower
than the free space in the buffer. While the consumer is congested and does
not consume these data, the is no way to stop the loop.
We can fix the function by removing the buffer reserve from the
computation. But it remains a dangerous decision to apply a max value on
room_needed. It is safer to require the caller must set a correct value. For
now, it is true. But at the end, it is totally unexepected to wait for more
room than an empty buffer can contain.
This patch must be backported to 2.8.
When receive or send expiration date of a stream-connector is retrieved, we
now automatically check if it may expire. If not, TICK_ETERNITY is returned.
The expiration dates of the frontend and backend stream-connectors are used
to compute the stream expiration date. This operation is performed at 2
places: at the end of process_stream() and in sc_notify() if the stream is
not woken up.
With this patch, there is no special changes for process_stream() because it
was already handled. It make thing a little simpler. However, it fixes
sc_notify() by avoiding to erroneously compute an expiration date in
past. This highly reduce the stream wakeups when there is contention on the
consumer side.
The bug was introduced with the commit 8073094bf ("NUG/MEDIUM: stconn:
Always update stream's expiration date after I/O"). It was an error to
unconditionnaly set the stream expiration data, without testing blocking
conditions on both SC.
This patch must be backported to 2.8.
When data are directly forwarded from a mux to the opposite one, we must not
forget to report send activity when data are successfully sent or report a
blocked send with data are blocked. It is important because otherwise, if
the transfer is quite long, longer than the client or server timeout, an
error may be triggered because the write timeout is reached.
H1, H2 and PT muxes are concerned. To fix the issue, The done_fastword()
callback now returns the amount of data consummed. This way it is possible
to update/reset the FSB data accordingly.
No backport needed.
This commit introduces a generic server-side parsing of type-value pair
arguments and allocation of a TLV list via a new keyword called
set-proxy-v2-tlv-fmt.
This allows to 1) forward any TLV type with the help of fc_pp_tlv,
2) generally, send out any TLV type and value via a log format expression.
To have this fully working the connection will need to be updated in
a follow-up commit to actually respect the new server TLV list.
default-server support has also been implemented.
In this patch, we add the possibility to declare on a table definition
("table" in peer section, or "stick-table" in proxy section) that we
want the remote/peer updates on that table to be pushed on a local
haproxy table in addition to the source table.
Consider this example:
|peers mypeers
| peer local 127.0.0.1:3334
| peer clust 127.0.0.1:3333
| table t1.local type string size 10m store server_id,server_key expire 30s
| table t1.clust type string size 10m store server_id,server_key write-to mypeers/t1.local expire 30s
With this setup, we consider haproxy uses t1.local as cache/local table
for read and write operations, and that t1.clust is a remote table
containing datas processed from t1.local and similar tables from other
haproxy peers in a cluster setup. The t1.clust table will be used to
refresh the local/cache one via the "write-to" statement.
What will happen, is that every time haproxy will see entry updates for
the t1.clust table: it will overwrite t1.local table with fresh data and
will update the entry expiration timer. If t1.local entry doesn't exist
yet (key doesn't exist), it will automatically create it. Note that only
types that cannot be used for arithmetic ops will be handled, and this
to prevent processed values from the remote table from interfering with
computations based on values from the local table. (ie: prevent
cumulative counters from growing indefinitely).
"write-to" will only push supported types if they both exist in the source
and the target table. Be careful with server_id and server_key storage
because they are often declared implicitly when referencing a table in
sticking rules but it is required to declare them explicitly for them to
be pushed between a remote and a local table through "write-to" option.
Also note that the "write-to" target table should have the same type as
the source one, and that the key length should be strictly equal,
otherwise haproxy will raise an error due to the tables being
incompatibles. A table that is already being written to cannot be used
as a source table for a "write-to" target.
Thanks to this patch, it will now be possible to use sticking rules in
peer cluster context by using a local table as a local cache which
will be automatically refreshed by one or multiple remote table(s).
This commit depends on:
- "MINOR: stktable: stktable_init() sets err_msg on error"
- "MINOR: stktable: check if a type should be used as-is"
stick table types now have an extra bit named 'as_is' that allows us to
check if such type should be used as-is or if it may be involved in
arithmetic operations such as counters. This can be useful since those
types are not common and may require specific handling.
e.g.: stktable_data_types[data_type].as_is will be set to 1 if the type
cannot be used in arithmetic operations.
Simplify stick and store sticktable proxy rules postparsing by adding
a sticking rule entry resolve (postparsing) function.
This will ease code maintenance.
This new fetcher can be used to extract the list of cookie names from
Cookie request header or from Set-Cookie response header depending on
the stream direction. There is an optional argument that can be used
as the delimiter (which is assumed to be the first character of the
argument) between cookie names. The default delimiter is comma (,).
Note that we will treat the Cookie request header as a semi-colon
separated list of cookies and each Set-Cookie response header as
a single cookie and extract the cookie names accordingly.
Similar to the previous commit which check for maxconn before allocating
a QUIC connection, this patch checks for maxsslconn at the same step.
This is necessary as a QUIC connection cannot run without a SSL context.
This should be backported up to 2.6. It relies on the following patch :
"BUG/MINOR: ssl: use a thread-safe sslconns increment"
Increment actconn and check maxconn limit when a quic_conn is
instantiated. This is necessary because prior to this patch, quic_conn
instances where not counted. Global actconn was only incremented after
the handshake has been completed and the connection structure is
allocated.
The increment is done using increment_actconn() on INITIAL packet
parsing if a new connection is about to be created. If the limit is
reached, the allocation is cancelled and the INITIAL packet is dropped.
The decrement is done under quic_conn_release(). This means that
quic_cc_conn instances are not taken into account. This seems safe
enough because quic_cc_conn are only used for minimal usage.
The counterpart of this change is that maxconn must not be checked a
second time when listener_accept() is done over a QUIC connection. For
this, a new bind_conf flag BC_O_XPRT_MAXCONN is set for listeners when
maxconn is already counted by the lower layer. For the moment, it is
positionned only for QUIC listeners.
Without this patch, haproxy process could suffer from heavy memory/CPU
load if the number of concurrent handshake is high.
This patch is not considered a bug fix per-se. However, it has a major
benefit to protect against too many QUIC handshakes. As such, it should
be backported up to 2.6. For this, it relies on the following patch :
"MINOR: frontend: implement a dedicated actconn increment function"
Each time a new SSL context is allocated, global.sslconns is
incremented. If global.maxsslconn is reached, the allocation is
cancelled.
This procedure was not entirely thread-safe due to the check and
increment operations conducted at different stage. This could lead to
global.maxsslconn slightly exceeded when several threads allocate SSL
context while sslconns is near the limit.
To fix this, use a CAS operation in a do/while loop. This code is
similar to the actconn/maxconn increment for connection.
A new function increment_sslconn() is defined for this operation. For
the moment, only SSL code is using it. However, it is expected that QUIC
will also use it to count QUIC connections as SSL ones.
This should be backported to all stable releases. Note that prior to the
2.6, sslconns was outside of global struct, so this commit should be
slightly adjusted.
When a new frontend connection is instantiated, actconn global counter
is incremented. If global maxconn value is reached, the connection is
cancelled. This ensures that system limit are under control.
Prior to this patch, the atomic check/increment operations were done
directly into listener_accept(). Move them in a dedicated function
increment_actconn() in frontend module. This will be useful when QUIC
connections will be counted in actconn counter.
Now when calling ha_panic() with a thread still under malloc_trim(),
we'll set a new tainted flag to easily report it, and the output
trace will report that this condition happened and will suggest to
use no-memory-trimming to avoid it in the future.
William suggested that since we can detect the presence of Lua in the
stack, let's combine it with stuck detection to set a new pair of flags
indicating a stuck Lua context and a stuck Lua shared context.
Now, executing an infinite loop in a Lua sample fetch function with
yield disabled crashes with tainted=0xe40 if loaded from a lua-load
statement, or tainted=0x640 from a lua-load-per-thread statement.
In addition, at the end of the panic dump, we can check if Lua was
seen stuck and emit recommendations about lua-load-per-thread and
the choice of dependencies depending on the presence of threads
and/or shared context.
This will make it easier to know that the panic function was called,
for the occasional case where the dump crashes and/or the stack is
corrupted and not much exploitable. Now at least it will be sufficient
to check the tainted value to know that someone called ha_panic(), and
it will also be usable to condition extra analysis.
This function allows to safely map proxy mode to corresponding proto_mode
This will allow for easier code maintenance and prevent mixups between
proxy mode and proto mode.
In 9a74a6c ("MAJOR: log: introduce log backends"), a mistake was made:
it was assumed that the proxy mode was already known during server
keyword parsing in parse_server() function, but this is wrong.
Indeed, "mode log" can be declared late in the proxy section. Due to this,
a simple config like this will cause the process to crash:
|backend test
|
| server name 127.0.0.1:8080
| mode log
In order to fix this, we relax some checks in _srv_parse_init() and store
the address protocol from str2sa_range() in server struct, then we set-up
a postparsing function that is to be called after config parsing to
finish the server checks/initialization that depend on the proxy mode
to be known. We achieve this by checking the PR_CAP_LB capability from
the parent proxy to know if we're in such case where the effective proxy
mode is not yet known (it is assumed that other proxies which are implicit
ones don't provide this possibility and thus don't suffer from this
constraint).
Only then, if the capability is not found, we immediately perform the
server checks that depend on the proxy mode, else the check is postponed
and it will automatically be performed during postparsing thanks to the
REGISTER_POST_SERVER_CHECK() hook.
Note that we remove the SRV_PARSE_IN_LOG_BE flag because it was introduced
in the above commit and it is no longer relevant.
No backport needed unless 9a74a6c gets backported.
The two recent commits below each added one flag to h2c but omitted to
update the __APPEND_FLAG macro used by dev/flags so they are not
properly decoded:
3dd963b35 ("BUG/MINOR: mux-h2: fix http-request and http-keep-alive timeouts again")
68d02e5fa ("BUG/MINOR: mux-h2: make up other blocked streams upon removal from list")
This can be backported along with these commits.
Define a new function srv_add_to_avail_list(). This function is used to
centralize connection insertion in available tree. It reuses a BUG_ON()
statement to ensure the connection is not present in the idle list.
Reverse HTTP bind is very specific in that in rely on a server to
initiate connection. All connection settings are defined on the server
line and ignored from the bind line.
Before this patch, most of keywords were silently ignored. This could
result in a configuration from doing unexpected things from the user
point of view. To improve this situation, add a new 'rhttp_ok' field in
bind_kw structure. If not set, the keyword is forbidden on a reverse
bind line and will cause a fatal config error.
For the moment, only the following keywords are usable with reverse bind
'id', 'name' and 'nbconn'.
This change is safe as it's already forbidden to mix reverse and
standard addresses on the same bind line.
Previously, maxconn keyword was reused for a specific usage on reverse
HTTP binds to specify the number of active connect to proceed. To avoid
confusion, introduce a new dedicated keyword 'nbconn' which is specific
to reverse HTTP bind.
This new keyword is forbidden for non-reverse listener. A fatal error is
emitted during config parsing if this rule is not respected. It's safe
because it's also forbidden to mix standard and reverse addresses on the
same bind line.
Internally, nbconn value will be reassigned to 'maxconn' member of
bind_conf structure. This ensures that listener layer will automatically
reenable the preconnect task each time a connection is closed.
Reverse HTTP listeners are very specific and share only a very limited
subset of keywords with other listeners. As such, it is probable
meaningless to mix standard and reverse addresses on the same bind line.
This patch emits a fatal error during configuration parsing if this is
the case.
The number of updates sent at once was limited to not loop too long to emit
updates when the buffer size is huge or when the number of sync tables is
huge. The limit can be configured and is set to 200 by default. However,
this fix introduced a bug. It is impossible to syncrhonize two peers if the
number of tables is higher than this limit. Thus by default, it is not
possible to sync two peers if there are more than 200 tables to sync.
Technically speacking, a teaching process is finished if we loop on all tables
with no new update messages sent. Because we are limited at each call, the loop
is splitted on several calls. However the restart point for the next loop is
always the last table for which we emitted an update message. Thus with more
tables than the limit, the loop never reachs the end point.
Worse, in conjunction with the bug fixed by "BUG/MEDIUM: peers: Be sure to
always refresh recconnect timer in sync task", it is possible to trigger the
watchdog because the applets may be woken up in loop and leave requesting
more room while its buffer is empty.
To fix the issue, restart conditions for a teaching loop were changed. If
the teach process is interrupted, we now save the restart point, called
stop_local_table. It is the last evaluated table on the previous loop. This
restart point is reset when the teach process is finished.
In additionn, the updates_sent variable in peer_send_msgs() was renamed to
updates to avoid ambiguities. Indeed, the variable is incremented, whether
messages were sent or not.
This patch must be backported as far as 2.6.
Stefan Behte reported that since commit f279a2f14 ("BUG/MINOR: mux-h2:
refresh the idle_timer when the mux is empty"), the http-request and
http-keep-alive timeouts don't work anymore on H2. Before this patch,
and since 3e448b9b64 ("BUG/MEDIUM: mux-h2: make sure control frames do
not refresh the idle timeout"), they would only be refreshed after stream
frames were sent (HEADERS or DATA) but the patch above that adds more
refresh points broke these so they don't expire anymore as long as
there's some activity.
We cannot just revert the fix since it also addressed an isse by which
sometimes the timeout would trigger too early and provoque truncated
responses. The right approach here is in fact to only use refresh the
idle timer when the mux buffer was flushed from any such stream frames.
In order to achieve this, we're now setting a flag on the connection
whenever we write a stream frame, and we consider that flag when deciding
to refresh the buffer after it's emptied. This way we'll only clear that
flag once the buffer is empty and there were stream data in it, not if
there were no such stream data. In theory it remains possible to leave
the flag on if some control data is appended after the buffer and it's
never cleared, but in practice it's not a problem as a buffer will always
get sent in large blocks when the window opens. Even a large buffer should
be emptied once in a while as control frames will not fill it as much as
data frames could.
Given the patch above was backported as far as 2.6, this patch should
also be backported as far as 2.6.
tune.rcvbuf.client and tune.rcvbuf.server are not suitable for shared
dgram sockets because they're per connection so their units are not the
same. However, QUIC's listener and log servers are not connected and
take per-thread or per-process traffic where a socket log buffer might
be too small, causing undesirable packet losses and retransmits in the
case of QUIC. This essentially manifests in listener mode with new
connections taking a lot of time to set up under heavy traffic due to
the small queues causing delays. Let's add a few new settings allowing
to set these shared socket sizes on the frontend and backend side (which
reminds that these are per-front/back and not per client/server hence
not per connection).
Instead of speaking of an initialisation stage for each data
fast-forwarding, we now use the negociate term. Thus init_ff/init_fastfwd
functions were renamed nego_ff/nego_fastfwd.
The zero-copy forwarding or the mux-to-mux forwarding is a way to
fast-forward data without using the channels buffers. Data are transferred
from a mux to the other one. The kernel splicing is an optimization of the
zero-copy forwarding. But it can also use normal buffers (but not channels
ones). This way, it could be possible to fast-forward data with muxes not
supporting the kernel splicing (H2 and H3 muxes) but also with applets.
However, this mode can introduce regressions or bugs in future (just like
the kernel splicing). Thus, It could be usefull to disable this optim. To do
so, in configuration, the global tune settting
'tune.disable-zero-copy-forwarding' may be set in a global section or the
'-dZ' command line parameter may be used to start HAProxy. Of course, this
also disables the kernel splicing.
Because channel_is_empty() function does now only check the channel's
buffer, we can remove it and rely on co_data() instead. Of course, all tests
must be inverted.
channel_is_empty() is thus removed.
It is important to split channels and I/O buffers. When data are pushed in
an I/O buffer, we consider them as forwarded. The channel never sees
them. Fast-forwarded data are now handled in the SE only.
When data were sent using the kernel splicing, we tried to send all data
with no restriction. Most of time it is valid. However, because the payload
representation may differ between the producer and the consumer, it is
important to be able to specify how must data to send via the splicing.
Of course, for performance reason, it is important to maximize amount of
data send via splicing at each call. However, on edge-cases, this now can be
limited.
The kernel splicing support was totally remove waiting for the mux-to-mux
fast-forward implementation. So corresponding mux callbacks can be removed
now.
mux-to-mux fast-forwarding will be added. To avoid mix with the splicing and
simplify the commits, the kernel splicing support is removed from the
stconn. CF_KERN_SPLICING flag is removed and the support is no longer tested
in process_stream().
In the stconn part, rcv_pipe() callback function is no longer called.
Reg-tests scripts testing the kernel splicing are temporarly marked as
broken.
To perform the mux-to-mux data fast-forwarding, 4 new callbacks were added
into the mux_ops structure. 2 callbacks will be used from the stconn for
fast-forward data. The 2 other callbacks will be used by the endpoint to
request an iobuf to the opposite endpoint.
* fastfwd() callback function is used by a producer to forward data
* resume_fastfwd() callback function is used by a consumer if some data are
blocked in the iobuf, to resume the data forwarding.
* init_fastfwd() must be used by an endpoint (the producer one), inside the
fastfwd() callback to request an iobuf to the opposite side (the consumer
one).
* done_fastfwd() must be used by an endpoint (the producer one) at the end
of fastfwd() to notify the opposite endpoint (the consumer one) if data
were forwarded or not.
This API is still under development, so it may evolved. Especially when the
fast-forward will be extended to applets.
2 helper functions were also added into the SE api to wrap init_fastfwd()
and done_fastfwd() callback function of the underlying endpoint.
For now, this API is unsed and not implemented at all in muxes.
It is unused for now, but the iobuf structure now owns a pointer to a
buffer. This buffer will be used to perform mux-to-mux fast-forwarding when
splicing is not supported or unusable. This pointer should be filled by an
endpoint to let the opposite one forward data.
Extra fields, in addition to the buffer, are mandatory because the buffer
may already contains some data. the ".offset" field may be used may be used
as the position to start to copy data. Finally, the amount of data copied in
this buffer must be saved in ".data" field.
Some flags are also added to prepare next changes. And helper stconn
fnuctions are updated to also count data in the buffer. For a first
implementation, it is not planned to handle data in the buffer and in the
pipe in same time. But it will be possible to do so.
Instead of talking about kernel splicing at stconn/sedesc level, we now try
to talk about mux-to-mux fast-forwarding. To do so, 2 functions were added
to know if there are fast-forwarded data and to retrieve this amount of
data. Of course, for now, there is only data in a pipe.
In addition, some flags were renamed to reflect this notion. Note the
channel's documentation was not updated yet.
The pipes used to put data when the kernel splicing is in used are moved in
the SE descriptors. For now, it is just a simple remplacement but there is a
major difference with the pipes in the channel. The data are pushed in the
consumer's pipe while it was pushed in the producer's pipe. So it means the
request data are now pushed in the pipe of the backend SE descriptor and
response data are pushed in the pipe of the frontend SE descriptor.
The idea is to hide the pipe from the channel/SC side and to be able to
handle fast-forwading in pipe but also in buffer. To do so, the pipe is
inside a new entity, called iobuf. This entity will be extended.
An interesting issue was met when testing the mux-to-mux forwarding code.
In order to preserve fairness, in h2_snd_buf() if other streams are waiting
in send_list or fctl_list, the stream that is attempting to send also goes
to its list, and will be woken up by h2_process_mux() or h2_send() when
some space is released. But on rare occasions, there are only a few (or
even a single) streams waiting in this list, and these streams are just
quickly removed because of a timeout or a quick h2_detach() that calls
h2s_destroy(). In this case there's no even to wake up the other waiting
stream in its list, and this will possibly resume processing after some
client WINDOW_UPDATE frames or even new streams, so usually it doesn't
last too long and it not much noticeable, reason why it was left that
long. In addition, measures have shown that in heavy network-bound
benchmark, this exact situation happens on less than 1% of the streams
(reached 4% with mux-mux).
The fix here consists in replacing these LIST_DEL_INIT() calls on
h2s->list with a function call that checks if other streams were queued
to the send_list recently, and if so, which also tries to resume them
by calling h2_resume_each_sending_h2s(). The detection of late additions
is made via a new flag on the connection, H2_CF_WAIT_INLIST, which is set
when a stream is queued due to other streams being present, and which is
cleared when this is function is called.
It is particularly difficult to reproduce this case which is particularly
timing-dependent, but in a constrained environment, a test involving 32
conns of 20 streams each, all downloading a 10 MB object previously
showed a limitation of 17 Gbps with lots of idle CPU time, and now
filled the cable at 25 Gbps.
This should be backported to all versions where it applies.
