haproxy/src/listener.c

1584 lines
45 KiB
C
Raw Normal View History

/*
* Listener management functions.
*
* Copyright 2000-2013 Willy Tarreau <w@1wt.eu>
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*
*/
#include <ctype.h>
#include <errno.h>
#include <stdio.h>
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
#include <haproxy/acl.h>
#include <haproxy/api.h>
#include <haproxy/cfgparse.h>
#include <haproxy/connection.h>
#include <haproxy/errors.h>
#include <haproxy/fd.h>
#include <haproxy/freq_ctr.h>
#include <haproxy/global.h>
#include <haproxy/list.h>
#include <haproxy/listener.h>
#include <haproxy/log.h>
#include <haproxy/protocol-t.h>
#include <haproxy/protocol.h>
#include <haproxy/sample.h>
#include <haproxy/stream.h>
#include <haproxy/task.h>
#include <haproxy/time.h>
#include <haproxy/tools.h>
/* List head of all known bind keywords */
static struct bind_kw_list bind_keywords = {
.list = LIST_HEAD_INIT(bind_keywords.list)
};
/* list of the temporarily limited listeners because of lack of resource */
static struct mt_list global_listener_queue = MT_LIST_HEAD_INIT(global_listener_queue);
static struct task *global_listener_queue_task;
/* listener status for stats */
const char* li_status_st[LI_STATE_COUNT] = {
[LI_STATUS_WAITING] = "WAITING",
[LI_STATUS_OPEN] = "OPEN",
[LI_STATUS_FULL] = "FULL",
};
#if defined(USE_THREAD)
struct accept_queue_ring accept_queue_rings[MAX_THREADS] __attribute__((aligned(64))) = { };
/* dequeue and process a pending connection from the local accept queue (single
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
* consumer). Returns the accepted connection or NULL if none was found.
*/
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
struct connection *accept_queue_pop_sc(struct accept_queue_ring *ring)
{
unsigned int pos, next;
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
struct connection *ptr;
struct connection **e;
pos = ring->head;
if (pos == ring->tail)
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
return NULL;
next = pos + 1;
if (next >= ACCEPT_QUEUE_SIZE)
next = 0;
e = &ring->entry[pos];
/* wait for the producer to update the listener's pointer */
while (1) {
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
ptr = *e;
__ha_barrier_load();
if (ptr)
break;
pl_cpu_relax();
}
/* release the entry */
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
*e = NULL;
__ha_barrier_store();
ring->head = next;
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
return ptr;
}
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
/* tries to push a new accepted connection <conn> into ring <ring>. Returns
* non-zero if it succeeds, or zero if the ring is full. Supports multiple
* producers.
*/
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
int accept_queue_push_mp(struct accept_queue_ring *ring, struct connection *conn)
{
unsigned int pos, next;
pos = ring->tail;
do {
next = pos + 1;
if (next >= ACCEPT_QUEUE_SIZE)
next = 0;
if (next == ring->head)
return 0; // ring full
} while (unlikely(!_HA_ATOMIC_CAS(&ring->tail, &pos, next)));
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
ring->entry[pos] = conn;
__ha_barrier_store();
return 1;
}
/* proceed with accepting new connections. Don't mark it static so that it appears
* in task dumps.
*/
struct task *accept_queue_process(struct task *t, void *context, unsigned int state)
{
struct accept_queue_ring *ring = context;
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
struct connection *conn;
struct listener *li;
unsigned int max_accept;
int ret;
/* if global.tune.maxaccept is -1, then max_accept is UINT_MAX. It
* is not really illimited, but it is probably enough.
*/
max_accept = global.tune.maxaccept ? global.tune.maxaccept : MAX_ACCEPT;
for (; max_accept; max_accept--) {
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
conn = accept_queue_pop_sc(ring);
if (!conn)
break;
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
li = __objt_listener(conn->target);
_HA_ATOMIC_ADD(&li->thr_conn[tid], 1);
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
ret = li->accept(conn);
if (ret <= 0) {
/* connection was terminated by the application */
continue;
}
/* increase the per-process number of cumulated sessions, this
* may only be done once l->accept() has accepted the connection.
*/
if (!(li->options & LI_O_UNLIMITED)) {
HA_ATOMIC_UPDATE_MAX(&global.sps_max,
update_freq_ctr(&global.sess_per_sec, 1));
if (li->bind_conf && li->bind_conf->is_ssl) {
HA_ATOMIC_UPDATE_MAX(&global.ssl_max,
update_freq_ctr(&global.ssl_per_sec, 1));
}
}
}
/* ran out of budget ? Let's come here ASAP */
if (!max_accept)
tasklet_wakeup(ring->tasklet);
return NULL;
}
/* Initializes the accept-queues. Returns 0 on success, otherwise ERR_* flags */
static int accept_queue_init()
{
struct tasklet *t;
int i;
for (i = 0; i < global.nbthread; i++) {
t = tasklet_new();
if (!t) {
ha_alert("Out of memory while initializing accept queue for thread %d\n", i);
return ERR_FATAL|ERR_ABORT;
}
t->tid = i;
t->process = accept_queue_process;
t->context = &accept_queue_rings[i];
accept_queue_rings[i].tasklet = t;
}
return 0;
}
REGISTER_CONFIG_POSTPARSER("multi-threaded accept queue", accept_queue_init);
#endif // USE_THREAD
/* helper to get listener status for stats */
enum li_status get_li_status(struct listener *l)
{
if (!l->maxconn || l->nbconn < l->maxconn) {
if (l->state == LI_LIMITED)
return LI_STATUS_WAITING;
else
return LI_STATUS_OPEN;
}
return LI_STATUS_FULL;
}
/* adjust the listener's state and its proxy's listener counters if needed.
* It must be called under the listener's lock, but uses atomic ops to change
* the proxy's counters so that the proxy lock is not needed.
*/
void listener_set_state(struct listener *l, enum li_state st)
{
struct proxy *px = l->bind_conf->frontend;
if (px) {
/* from state */
switch (l->state) {
case LI_NEW: /* first call */
_HA_ATOMIC_ADD(&px->li_all, 1);
break;
case LI_INIT:
case LI_ASSIGNED:
break;
case LI_PAUSED:
_HA_ATOMIC_SUB(&px->li_paused, 1);
break;
case LI_LISTEN:
_HA_ATOMIC_SUB(&px->li_bound, 1);
break;
case LI_READY:
case LI_FULL:
case LI_LIMITED:
_HA_ATOMIC_SUB(&px->li_ready, 1);
break;
}
/* to state */
switch (st) {
case LI_NEW:
case LI_INIT:
case LI_ASSIGNED:
break;
case LI_PAUSED:
BUG_ON(l->rx.fd == -1);
_HA_ATOMIC_ADD(&px->li_paused, 1);
break;
case LI_LISTEN:
BUG_ON(l->rx.fd == -1);
_HA_ATOMIC_ADD(&px->li_bound, 1);
break;
case LI_READY:
case LI_FULL:
case LI_LIMITED:
BUG_ON(l->rx.fd == -1);
_HA_ATOMIC_ADD(&px->li_ready, 1);
break;
}
}
l->state = st;
}
/* This function adds the specified listener's file descriptor to the polling
* lists if it is in the LI_LISTEN state. The listener enters LI_READY or
* LI_FULL state depending on its number of connections. In daemon mode, we
* also support binding only the relevant processes to their respective
* listeners. We don't do that in debug mode however.
*/
void enable_listener(struct listener *listener)
{
HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
/* If this listener is supposed to be only in the master, close it in
* the workers. Conversely, if it's supposed to be only in the workers
* close it in the master.
*/
if (!!master != !!(listener->rx.flags & RX_F_MWORKER))
do_unbind_listener(listener);
if (listener->state == LI_LISTEN) {
BUG_ON(listener->rx.fd == -1);
if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
(!!master != !!(listener->rx.flags & RX_F_MWORKER) ||
!(proc_mask(listener->rx.settings->bind_proc) & pid_bit))) {
/* we don't want to enable this listener and don't
* want any fd event to reach it.
*/
do_unbind_listener(listener);
}
else if (!listener->maxconn || listener->nbconn < listener->maxconn) {
MEDIUM: listeners: now use the listener's ->enable/disable At each place we used to manipulate the FDs directly we can now call the listener protocol's enable/disable/rx_enable/rx_disable depending on whether the state changes on the listener or the receiver. One exception currently remains in listener_accept() which is a bit special and which should be split into 2 or 3 parts in the various protocol layers. The test of fd_updt in do_unbind_listener() that was added by commit a51885621 ("BUG/MEDIUM: listeners: Don't call fd_stop_recv() if fd_updt is NULL.") could finally be removed since that part is correctly handled in the low-level disable() function. One disable() was added in resume_listener() before switching to LI_FULL because rx_resume() enables polling on the FD for the receiver while we want to disable it if the listener is full. There are different ways to clean this up in the future. One of them could be to consider that TCP receivers only act at the listener level. But in fact it does not translate reality. The reality is that only the receiver is paused and that the listener's state ought not be affected here. Ultimately the resume_listener() function should be split so that the part controlled by the protocols only acts on the receiver, and that the receiver itself notifies the upper listener about the change so that the listener protocol may decide to disable or enable polling. Conversely the listener should automatically update its receiver when they share the same state. Since there is no harm proceeding like this, let's keep this for now.
2020-09-25 18:32:28 +00:00
listener->rx.proto->enable(listener);
listener_set_state(listener, LI_READY);
}
else {
listener_set_state(listener, LI_FULL);
}
}
HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
}
/*
* This function completely stops a listener. It will need to operate under the
* proxy's lock, the protocol's lock, and the listener's lock. The caller is
* responsible for indicating in lpx, lpr, lli whether the respective locks are
* already held (non-zero) or not (zero) so that the function picks the missing
* ones, in this order. The proxy's listeners count is updated and the proxy is
* disabled and woken up after the last one is gone.
*/
void stop_listener(struct listener *l, int lpx, int lpr, int lli)
{
struct proxy *px = l->bind_conf->frontend;
if (l->options & LI_O_NOSTOP) {
/* master-worker sockpairs are never closed but don't count as a
* job.
*/
return;
}
if (!lpx)
HA_RWLOCK_WRLOCK(PROXY_LOCK, &px->lock);
if (!lpr)
HA_SPIN_LOCK(PROTO_LOCK, &proto_lock);
if (!lli)
HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
if (l->state > LI_INIT) {
do_unbind_listener(l);
if (l->state >= LI_ASSIGNED)
__delete_listener(l);
proxy_cond_disable(px);
}
if (!lli)
HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
if (!lpr)
HA_SPIN_UNLOCK(PROTO_LOCK, &proto_lock);
if (!lpx)
HA_RWLOCK_WRUNLOCK(PROXY_LOCK, &px->lock);
}
/* This function adds the specified <listener> to the protocol <proto>. It
* does nothing if the protocol was already added. The listener's state is
* automatically updated from LI_INIT to LI_ASSIGNED. The number of listeners
* for the protocol is updated. This must be called with the proto lock held.
*/
void default_add_listener(struct protocol *proto, struct listener *listener)
{
if (listener->state != LI_INIT)
return;
listener_set_state(listener, LI_ASSIGNED);
listener->rx.proto = proto;
LIST_ADDQ(&proto->receivers, &listener->rx.proto_list);
proto->nb_receivers++;
}
/* default function called to suspend a listener: it simply passes the call to
* the underlying receiver. This is find for most socket-based protocols. This
* must be called under the listener's lock. It will return non-zero on success,
* 0 on failure. If no receiver-level suspend is provided, the operation is
* assumed to succeed.
*/
int default_suspend_listener(struct listener *l)
{
int ret = 1;
if (!l->rx.proto->rx_suspend)
return 1;
ret = l->rx.proto->rx_suspend(&l->rx);
return ret > 0 ? ret : 0;
}
/* Tries to resume a suspended listener, and returns non-zero on success or
* zero on failure. On certain errors, an alert or a warning might be displayed.
* It must be called with the listener's lock held. Depending on the listener's
* state and protocol, a listen() call might be used to resume operations, or a
* call to the receiver's resume() function might be used as well. This is
* suitable as a default function for TCP and UDP. This must be called with the
* listener's lock held.
*/
int default_resume_listener(struct listener *l)
{
int ret = 1;
if (l->state == LI_ASSIGNED) {
char msg[100];
int err;
err = l->rx.proto->listen(l, msg, sizeof(msg));
if (err & ERR_ALERT)
ha_alert("Resuming listener: %s\n", msg);
else if (err & ERR_WARN)
ha_warning("Resuming listener: %s\n", msg);
if (err & (ERR_FATAL | ERR_ABORT)) {
ret = 0;
goto end;
}
}
if (l->state < LI_PAUSED) {
ret = 0;
goto end;
}
if (l->state == LI_PAUSED && l->rx.proto->rx_resume &&
l->rx.proto->rx_resume(&l->rx) <= 0)
ret = 0;
end:
return ret;
}
/* This function tries to temporarily disable a listener, depending on the OS
* capabilities. Linux unbinds the listen socket after a SHUT_RD, and ignores
* SHUT_WR. Solaris refuses either shutdown(). OpenBSD ignores SHUT_RD but
* closes upon SHUT_WR and refuses to rebind. So a common validation path
* involves SHUT_WR && listen && SHUT_RD. In case of success, the FD's polling
* is disabled. It normally returns non-zero, unless an error is reported.
*/
int pause_listener(struct listener *l)
{
struct proxy *px = l->bind_conf->frontend;
int ret = 1;
HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
!(proc_mask(l->rx.settings->bind_proc) & pid_bit))
goto end;
if (l->state <= LI_PAUSED)
goto end;
if (l->rx.proto->suspend)
ret = l->rx.proto->suspend(l);
MT_LIST_DEL(&l->wait_queue);
listener_set_state(l, LI_PAUSED);
if (px && !px->li_ready) {
ha_warning("Paused %s %s.\n", proxy_cap_str(px->cap), px->id);
send_log(px, LOG_WARNING, "Paused %s %s.\n", proxy_cap_str(px->cap), px->id);
}
end:
HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
return ret;
}
/* This function tries to resume a temporarily disabled listener. Paused, full,
* limited and disabled listeners are handled, which means that this function
* may replace enable_listener(). The resulting state will either be LI_READY
* or LI_FULL. 0 is returned in case of failure to resume (eg: dead socket).
* Listeners bound to a different process are not woken up unless we're in
* foreground mode, and are ignored. If the listener was only in the assigned
* state, it's totally rebound. This can happen if a pause() has completely
* stopped it. If the resume fails, 0 is returned and an error might be
* displayed.
*/
BUG/MEDIUM: listener: use a self-locked list for the dequeue lists There is a very difficult to reproduce race in the listener's accept code, which is much easier to reproduce once connection limits are properly enforced. It's an ABBA lock issue : - the following functions take l->lock then lq_lock : disable_listener, pause_listener, listener_full, limit_listener, do_unbind_listener - the following ones take lq_lock then l->lock : resume_listener, dequeue_all_listener This is because __resume_listener() only takes the listener's lock and expects to be called with lq_lock held. The problem can easily happen when listener_full() and limit_listener() are called a lot while in parallel another thread releases sessions for the same listener using listener_release() which in turn calls resume_listener(). This scenario is more prevalent in 2.0-dev since the removal of the accept lock in listener_accept(). However in 1.9 and before, a different but extremely unlikely scenario can happen : thread1 thread2 ............................ enter listener_accept() limit_listener() ............................ long pause before taking the lock session_free() dequeue_all_listeners() lock(lq_lock) [1] ............................ try_lock(l->lock) [2] __resume_listener() spin_lock(l->lock) =>WAIT[2] ............................ accept() l->accept() nbconn==maxconn => listener_full() state==LI_LIMITED => lock(lq_lock) =>DEADLOCK[1]! In practice it is almost impossible to trigger it because it requires to limit both on the listener's maxconn and the frontend's rate limit, at the same time, and to release the listener when the connection rate goes below the limit between poll() returns the FD and the lock is taken (a few nanoseconds). But maybe with threads competing on the same core it has more chances to appear. This patch removes the lq_lock and replaces it with a lockless queue for the listener's wait queue (well, technically speaking a self-locked queue) brought by commit a8434ec14 ("MINOR: lists: Implement locked variations.") and its few subsequent fixes. This relieves us from the need of the lq_lock and removes the deadlock. It also gets rid of the distinction between __resume_listener() and resume_listener() since the only difference was the lq_lock. All listener removals from the list are now unconditional to avoid races on the state. It's worth noting that the list used to never be initialized and that it used to work only thanks to the state tests, so the initialization has now been added. This patch must carefully be backported to 1.9 and very likely 1.8. It is mandatory to be careful about replacing all manipulations of l->wait_queue, global.listener_queue and p->listener_queue.
2019-02-28 09:27:18 +00:00
int resume_listener(struct listener *l)
{
struct proxy *px = l->bind_conf->frontend;
int was_paused = px && px->li_paused;
int ret = 1;
HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
BUG/MAJOR: listener: fix thread safety in resume_listener() resume_listener() can be called from a thread not part of the listener's mask after a curr_conn has gone lower than a proxy's or the process' limit. This results in fd_may_recv() being called unlocked if the listener is bound to only one thread, and quickly locks up. This patch solves this by creating a per-thread work_list dedicated to listeners, and modifying resume_listener() so that it bounces the listener to one of its owning thread's work_list and waking it up. This thread will then call resume_listener() again and will perform the operation on the file descriptor itself. It is important to do it this way so that the listener's state cannot be modified while the listener is being moved, otherwise multiple threads can take conflicting decisions and the listener could be put back into the global queue if the listener was used at the same time. It seems like a slightly simpler approach would be possible if the locked list API would provide the ability to return a locked element. In this case the listener would be immediately requeued in dequeue_all_listeners() without having to go through resume_listener() with its associated lock. This fix must be backported to all versions having the lock-less accept loop, which is as far as 1.8 since deadlock fixes involving this feature had to be backported there. It is expected that the code should not differ too much there. However, previous commit "MINOR: task: introduce work lists" will be needed as well and should not present difficulties either. For 1.8, the commits introducing thread_mask() and LIST_ADDED() will be needed as well, either backporting my_flsl() or switching to my_ffsl() will be OK, and some changes will have to be performed so that the init function is properly called (and maybe the deinit one can be dropped). In order to test for the fix, simply set up a multi-threaded frontend with multiple bind lines each attached to a single thread (reproduced with 16 threads here), set up a very low maxconn value on the frontend, and inject heavy traffic on all listeners in parallel with slightly more connections than the configured limit ( typically +20%) so that it flips very frequently. If the bug is still there, at some point (5-20 seconds) the traffic will go much lower or even stop, either with spinning threads or not.
2019-07-11 08:08:31 +00:00
/* check that another thread didn't to the job in parallel (e.g. at the
* end of listen_accept() while we'd come from dequeue_all_listeners().
*/
if (MT_LIST_ADDED(&l->wait_queue))
BUG/MAJOR: listener: fix thread safety in resume_listener() resume_listener() can be called from a thread not part of the listener's mask after a curr_conn has gone lower than a proxy's or the process' limit. This results in fd_may_recv() being called unlocked if the listener is bound to only one thread, and quickly locks up. This patch solves this by creating a per-thread work_list dedicated to listeners, and modifying resume_listener() so that it bounces the listener to one of its owning thread's work_list and waking it up. This thread will then call resume_listener() again and will perform the operation on the file descriptor itself. It is important to do it this way so that the listener's state cannot be modified while the listener is being moved, otherwise multiple threads can take conflicting decisions and the listener could be put back into the global queue if the listener was used at the same time. It seems like a slightly simpler approach would be possible if the locked list API would provide the ability to return a locked element. In this case the listener would be immediately requeued in dequeue_all_listeners() without having to go through resume_listener() with its associated lock. This fix must be backported to all versions having the lock-less accept loop, which is as far as 1.8 since deadlock fixes involving this feature had to be backported there. It is expected that the code should not differ too much there. However, previous commit "MINOR: task: introduce work lists" will be needed as well and should not present difficulties either. For 1.8, the commits introducing thread_mask() and LIST_ADDED() will be needed as well, either backporting my_flsl() or switching to my_ffsl() will be OK, and some changes will have to be performed so that the init function is properly called (and maybe the deinit one can be dropped). In order to test for the fix, simply set up a multi-threaded frontend with multiple bind lines each attached to a single thread (reproduced with 16 threads here), set up a very low maxconn value on the frontend, and inject heavy traffic on all listeners in parallel with slightly more connections than the configured limit ( typically +20%) so that it flips very frequently. If the bug is still there, at some point (5-20 seconds) the traffic will go much lower or even stop, either with spinning threads or not.
2019-07-11 08:08:31 +00:00
goto end;
if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
!(proc_mask(l->rx.settings->bind_proc) & pid_bit))
goto end;
if (l->state == LI_READY)
goto end;
if (l->rx.proto->resume)
ret = l->rx.proto->resume(l);
if (l->maxconn && l->nbconn >= l->maxconn) {
MEDIUM: listeners: now use the listener's ->enable/disable At each place we used to manipulate the FDs directly we can now call the listener protocol's enable/disable/rx_enable/rx_disable depending on whether the state changes on the listener or the receiver. One exception currently remains in listener_accept() which is a bit special and which should be split into 2 or 3 parts in the various protocol layers. The test of fd_updt in do_unbind_listener() that was added by commit a51885621 ("BUG/MEDIUM: listeners: Don't call fd_stop_recv() if fd_updt is NULL.") could finally be removed since that part is correctly handled in the low-level disable() function. One disable() was added in resume_listener() before switching to LI_FULL because rx_resume() enables polling on the FD for the receiver while we want to disable it if the listener is full. There are different ways to clean this up in the future. One of them could be to consider that TCP receivers only act at the listener level. But in fact it does not translate reality. The reality is that only the receiver is paused and that the listener's state ought not be affected here. Ultimately the resume_listener() function should be split so that the part controlled by the protocols only acts on the receiver, and that the receiver itself notifies the upper listener about the change so that the listener protocol may decide to disable or enable polling. Conversely the listener should automatically update its receiver when they share the same state. Since there is no harm proceeding like this, let's keep this for now.
2020-09-25 18:32:28 +00:00
l->rx.proto->disable(l);
listener_set_state(l, LI_FULL);
goto done;
}
MEDIUM: listeners: now use the listener's ->enable/disable At each place we used to manipulate the FDs directly we can now call the listener protocol's enable/disable/rx_enable/rx_disable depending on whether the state changes on the listener or the receiver. One exception currently remains in listener_accept() which is a bit special and which should be split into 2 or 3 parts in the various protocol layers. The test of fd_updt in do_unbind_listener() that was added by commit a51885621 ("BUG/MEDIUM: listeners: Don't call fd_stop_recv() if fd_updt is NULL.") could finally be removed since that part is correctly handled in the low-level disable() function. One disable() was added in resume_listener() before switching to LI_FULL because rx_resume() enables polling on the FD for the receiver while we want to disable it if the listener is full. There are different ways to clean this up in the future. One of them could be to consider that TCP receivers only act at the listener level. But in fact it does not translate reality. The reality is that only the receiver is paused and that the listener's state ought not be affected here. Ultimately the resume_listener() function should be split so that the part controlled by the protocols only acts on the receiver, and that the receiver itself notifies the upper listener about the change so that the listener protocol may decide to disable or enable polling. Conversely the listener should automatically update its receiver when they share the same state. Since there is no harm proceeding like this, let's keep this for now.
2020-09-25 18:32:28 +00:00
l->rx.proto->enable(l);
listener_set_state(l, LI_READY);
done:
if (was_paused && !px->li_paused) {
ha_warning("Resumed %s %s.\n", proxy_cap_str(px->cap), px->id);
send_log(px, LOG_WARNING, "Resumed %s %s.\n", proxy_cap_str(px->cap), px->id);
}
end:
HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
return ret;
}
REORG/MAJOR: session: rename the "session" entity to "stream" With HTTP/2, we'll have to support multiplexed streams. A stream is in fact the largest part of what we currently call a session, it has buffers, logs, etc. In order to catch any error, this commit removes any reference to the struct session and tries to rename most "session" occurrences in function names to "stream" and "sess" to "strm" when that's related to a session. The files stream.{c,h} were added and session.{c,h} removed. The session will be reintroduced later and a few parts of the stream will progressively be moved overthere. It will more or less contain only what we need in an embryonic session. Sample fetch functions and converters will have to change a bit so that they'll use an L5 (session) instead of what's currently called "L4" which is in fact L6 for now. Once all changes are completed, we should see approximately this : L7 - http_txn L6 - stream L5 - session L4 - connection | applet There will be at most one http_txn per stream, and a same session will possibly be referenced by multiple streams. A connection will point to a session and to a stream. The session will hold all the information we need to keep even when we don't yet have a stream. Some more cleanup is needed because some code was already far from being clean. The server queue management still refers to sessions at many places while comments talk about connections. This will have to be cleaned up once we have a server-side connection pool manager. Stream flags "SN_*" still need to be renamed, it doesn't seem like any of them will need to move to the session.
2015-04-02 22:22:06 +00:00
/* Marks a ready listener as full so that the stream code tries to re-enable
* it upon next close() using resume_listener().
*/
static void listener_full(struct listener *l)
{
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
if (l->state >= LI_READY) {
MT_LIST_DEL(&l->wait_queue);
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
if (l->state != LI_FULL) {
MEDIUM: listeners: now use the listener's ->enable/disable At each place we used to manipulate the FDs directly we can now call the listener protocol's enable/disable/rx_enable/rx_disable depending on whether the state changes on the listener or the receiver. One exception currently remains in listener_accept() which is a bit special and which should be split into 2 or 3 parts in the various protocol layers. The test of fd_updt in do_unbind_listener() that was added by commit a51885621 ("BUG/MEDIUM: listeners: Don't call fd_stop_recv() if fd_updt is NULL.") could finally be removed since that part is correctly handled in the low-level disable() function. One disable() was added in resume_listener() before switching to LI_FULL because rx_resume() enables polling on the FD for the receiver while we want to disable it if the listener is full. There are different ways to clean this up in the future. One of them could be to consider that TCP receivers only act at the listener level. But in fact it does not translate reality. The reality is that only the receiver is paused and that the listener's state ought not be affected here. Ultimately the resume_listener() function should be split so that the part controlled by the protocols only acts on the receiver, and that the receiver itself notifies the upper listener about the change so that the listener protocol may decide to disable or enable polling. Conversely the listener should automatically update its receiver when they share the same state. Since there is no harm proceeding like this, let's keep this for now.
2020-09-25 18:32:28 +00:00
l->rx.proto->disable(l);
listener_set_state(l, LI_FULL);
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
}
}
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
}
/* Marks a ready listener as limited so that we only try to re-enable it when
* resources are free again. It will be queued into the specified queue.
*/
static void limit_listener(struct listener *l, struct mt_list *list)
{
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
if (l->state == LI_READY) {
MINOR: lists: rename some MT_LIST operations to clarify them Initially when mt_lists were added, their purpose was to be used with the scheduler, where anyone may concurrently add the same tasklet, so it sounded natural to implement a check in MT_LIST_ADD{,Q}. Later their usage was extended and MT_LIST_ADD{,Q} started to be used on situations where the element to be added was exclusively owned by the one performing the operation so a conflict was impossible. This became more obvious with the idle connections and the new macro was called MT_LIST_ADDQ_NOCHECK. But this remains confusing and at many places it's not expected that an MT_LIST_ADD could possibly fail, and worse, at some places we start by initializing it before adding (and the test is superflous) so let's rename them to something more conventional to denote the presence of the check or not: MT_LIST_ADD{,Q} : inconditional operation, the caller owns the element, and doesn't care about the element's current state (exactly like LIST_ADD) MT_LIST_TRY_ADD{,Q}: only perform the operation if the element is not already added or in the process of being added. This means that the previously "safe" MT_LIST_ADD{,Q} are not "safe" anymore. This also means that in case of backport mistakes in the future causing this to be overlooked, the slower and safer functions will still be used by default. Note that the missing unchecked MT_LIST_ADD macro was added. The rest of the code will have to be reviewed so that a number of callers of MT_LIST_TRY_ADDQ are changed to MT_LIST_ADDQ to remove the unneeded test.
2020-07-10 06:10:29 +00:00
MT_LIST_TRY_ADDQ(list, &l->wait_queue);
MEDIUM: listeners: now use the listener's ->enable/disable At each place we used to manipulate the FDs directly we can now call the listener protocol's enable/disable/rx_enable/rx_disable depending on whether the state changes on the listener or the receiver. One exception currently remains in listener_accept() which is a bit special and which should be split into 2 or 3 parts in the various protocol layers. The test of fd_updt in do_unbind_listener() that was added by commit a51885621 ("BUG/MEDIUM: listeners: Don't call fd_stop_recv() if fd_updt is NULL.") could finally be removed since that part is correctly handled in the low-level disable() function. One disable() was added in resume_listener() before switching to LI_FULL because rx_resume() enables polling on the FD for the receiver while we want to disable it if the listener is full. There are different ways to clean this up in the future. One of them could be to consider that TCP receivers only act at the listener level. But in fact it does not translate reality. The reality is that only the receiver is paused and that the listener's state ought not be affected here. Ultimately the resume_listener() function should be split so that the part controlled by the protocols only acts on the receiver, and that the receiver itself notifies the upper listener about the change so that the listener protocol may decide to disable or enable polling. Conversely the listener should automatically update its receiver when they share the same state. Since there is no harm proceeding like this, let's keep this for now.
2020-09-25 18:32:28 +00:00
l->rx.proto->disable(l);
listener_set_state(l, LI_LIMITED);
}
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
}
/* Dequeues all listeners waiting for a resource the global wait queue */
void dequeue_all_listeners()
{
BUG/MEDIUM: listener: use a self-locked list for the dequeue lists There is a very difficult to reproduce race in the listener's accept code, which is much easier to reproduce once connection limits are properly enforced. It's an ABBA lock issue : - the following functions take l->lock then lq_lock : disable_listener, pause_listener, listener_full, limit_listener, do_unbind_listener - the following ones take lq_lock then l->lock : resume_listener, dequeue_all_listener This is because __resume_listener() only takes the listener's lock and expects to be called with lq_lock held. The problem can easily happen when listener_full() and limit_listener() are called a lot while in parallel another thread releases sessions for the same listener using listener_release() which in turn calls resume_listener(). This scenario is more prevalent in 2.0-dev since the removal of the accept lock in listener_accept(). However in 1.9 and before, a different but extremely unlikely scenario can happen : thread1 thread2 ............................ enter listener_accept() limit_listener() ............................ long pause before taking the lock session_free() dequeue_all_listeners() lock(lq_lock) [1] ............................ try_lock(l->lock) [2] __resume_listener() spin_lock(l->lock) =>WAIT[2] ............................ accept() l->accept() nbconn==maxconn => listener_full() state==LI_LIMITED => lock(lq_lock) =>DEADLOCK[1]! In practice it is almost impossible to trigger it because it requires to limit both on the listener's maxconn and the frontend's rate limit, at the same time, and to release the listener when the connection rate goes below the limit between poll() returns the FD and the lock is taken (a few nanoseconds). But maybe with threads competing on the same core it has more chances to appear. This patch removes the lq_lock and replaces it with a lockless queue for the listener's wait queue (well, technically speaking a self-locked queue) brought by commit a8434ec14 ("MINOR: lists: Implement locked variations.") and its few subsequent fixes. This relieves us from the need of the lq_lock and removes the deadlock. It also gets rid of the distinction between __resume_listener() and resume_listener() since the only difference was the lq_lock. All listener removals from the list are now unconditional to avoid races on the state. It's worth noting that the list used to never be initialized and that it used to work only thanks to the state tests, so the initialization has now been added. This patch must carefully be backported to 1.9 and very likely 1.8. It is mandatory to be careful about replacing all manipulations of l->wait_queue, global.listener_queue and p->listener_queue.
2019-02-28 09:27:18 +00:00
struct listener *listener;
while ((listener = MT_LIST_POP(&global_listener_queue, struct listener *, wait_queue))) {
/* This cannot fail because the listeners are by definition in
* the LI_LIMITED state.
*/
resume_listener(listener);
}
}
/* Dequeues all listeners waiting for a resource in proxy <px>'s queue */
void dequeue_proxy_listeners(struct proxy *px)
{
struct listener *listener;
while ((listener = MT_LIST_POP(&px->listener_queue, struct listener *, wait_queue))) {
/* This cannot fail because the listeners are by definition in
BUG/MEDIUM: listener: use a self-locked list for the dequeue lists There is a very difficult to reproduce race in the listener's accept code, which is much easier to reproduce once connection limits are properly enforced. It's an ABBA lock issue : - the following functions take l->lock then lq_lock : disable_listener, pause_listener, listener_full, limit_listener, do_unbind_listener - the following ones take lq_lock then l->lock : resume_listener, dequeue_all_listener This is because __resume_listener() only takes the listener's lock and expects to be called with lq_lock held. The problem can easily happen when listener_full() and limit_listener() are called a lot while in parallel another thread releases sessions for the same listener using listener_release() which in turn calls resume_listener(). This scenario is more prevalent in 2.0-dev since the removal of the accept lock in listener_accept(). However in 1.9 and before, a different but extremely unlikely scenario can happen : thread1 thread2 ............................ enter listener_accept() limit_listener() ............................ long pause before taking the lock session_free() dequeue_all_listeners() lock(lq_lock) [1] ............................ try_lock(l->lock) [2] __resume_listener() spin_lock(l->lock) =>WAIT[2] ............................ accept() l->accept() nbconn==maxconn => listener_full() state==LI_LIMITED => lock(lq_lock) =>DEADLOCK[1]! In practice it is almost impossible to trigger it because it requires to limit both on the listener's maxconn and the frontend's rate limit, at the same time, and to release the listener when the connection rate goes below the limit between poll() returns the FD and the lock is taken (a few nanoseconds). But maybe with threads competing on the same core it has more chances to appear. This patch removes the lq_lock and replaces it with a lockless queue for the listener's wait queue (well, technically speaking a self-locked queue) brought by commit a8434ec14 ("MINOR: lists: Implement locked variations.") and its few subsequent fixes. This relieves us from the need of the lq_lock and removes the deadlock. It also gets rid of the distinction between __resume_listener() and resume_listener() since the only difference was the lq_lock. All listener removals from the list are now unconditional to avoid races on the state. It's worth noting that the list used to never be initialized and that it used to work only thanks to the state tests, so the initialization has now been added. This patch must carefully be backported to 1.9 and very likely 1.8. It is mandatory to be careful about replacing all manipulations of l->wait_queue, global.listener_queue and p->listener_queue.
2019-02-28 09:27:18 +00:00
* the LI_LIMITED state.
*/
BUG/MEDIUM: listener: use a self-locked list for the dequeue lists There is a very difficult to reproduce race in the listener's accept code, which is much easier to reproduce once connection limits are properly enforced. It's an ABBA lock issue : - the following functions take l->lock then lq_lock : disable_listener, pause_listener, listener_full, limit_listener, do_unbind_listener - the following ones take lq_lock then l->lock : resume_listener, dequeue_all_listener This is because __resume_listener() only takes the listener's lock and expects to be called with lq_lock held. The problem can easily happen when listener_full() and limit_listener() are called a lot while in parallel another thread releases sessions for the same listener using listener_release() which in turn calls resume_listener(). This scenario is more prevalent in 2.0-dev since the removal of the accept lock in listener_accept(). However in 1.9 and before, a different but extremely unlikely scenario can happen : thread1 thread2 ............................ enter listener_accept() limit_listener() ............................ long pause before taking the lock session_free() dequeue_all_listeners() lock(lq_lock) [1] ............................ try_lock(l->lock) [2] __resume_listener() spin_lock(l->lock) =>WAIT[2] ............................ accept() l->accept() nbconn==maxconn => listener_full() state==LI_LIMITED => lock(lq_lock) =>DEADLOCK[1]! In practice it is almost impossible to trigger it because it requires to limit both on the listener's maxconn and the frontend's rate limit, at the same time, and to release the listener when the connection rate goes below the limit between poll() returns the FD and the lock is taken (a few nanoseconds). But maybe with threads competing on the same core it has more chances to appear. This patch removes the lq_lock and replaces it with a lockless queue for the listener's wait queue (well, technically speaking a self-locked queue) brought by commit a8434ec14 ("MINOR: lists: Implement locked variations.") and its few subsequent fixes. This relieves us from the need of the lq_lock and removes the deadlock. It also gets rid of the distinction between __resume_listener() and resume_listener() since the only difference was the lq_lock. All listener removals from the list are now unconditional to avoid races on the state. It's worth noting that the list used to never be initialized and that it used to work only thanks to the state tests, so the initialization has now been added. This patch must carefully be backported to 1.9 and very likely 1.8. It is mandatory to be careful about replacing all manipulations of l->wait_queue, global.listener_queue and p->listener_queue.
2019-02-28 09:27:18 +00:00
resume_listener(listener);
}
}
/* default function used to unbind a listener. This is for use by standard
* protocols working on top of accepted sockets. The receiver's rx_unbind()
* will automatically be used after the listener is disabled if the socket is
* still bound. This must be used under the listener's lock.
*/
void default_unbind_listener(struct listener *listener)
{
if (listener->state <= LI_ASSIGNED)
goto out_close;
if (listener->rx.fd == -1) {
listener_set_state(listener, LI_ASSIGNED);
goto out_close;
}
if (listener->state >= LI_READY) {
listener->rx.proto->disable(listener);
if (listener->rx.flags & RX_F_BOUND)
listener_set_state(listener, LI_LISTEN);
}
out_close:
if (listener->rx.flags & RX_F_BOUND)
listener->rx.proto->rx_unbind(&listener->rx);
}
/* This function closes the listening socket for the specified listener,
* provided that it's already in a listening state. The protocol's unbind()
* is called to put the listener into LI_ASSIGNED or LI_LISTEN and handle
* the unbinding tasks. The listener enters then the LI_ASSIGNED state if
* the receiver is unbound. Must be called with the lock held.
*/
void do_unbind_listener(struct listener *listener)
{
MT_LIST_DEL(&listener->wait_queue);
if (listener->rx.proto->unbind)
listener->rx.proto->unbind(listener);
/* we may have to downgrade the listener if the rx was closed */
if (!(listener->rx.flags & RX_F_BOUND) && listener->state > LI_ASSIGNED)
listener_set_state(listener, LI_ASSIGNED);
}
/* This function closes the listening socket for the specified listener,
* provided that it's already in a listening state. The listener enters the
* LI_ASSIGNED state, except if the FD is not closed, in which case it may
* remain in LI_LISTEN. This function is intended to be used as a generic
* function for standard protocols.
*/
void unbind_listener(struct listener *listener)
{
HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
do_unbind_listener(listener);
HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
}
/* creates one or multiple listeners for bind_conf <bc> on sockaddr <ss> on port
* range <portl> to <porth>, and possibly attached to fd <fd> (or -1 for auto
* allocation). The address family is taken from ss->ss_family, and the protocol
* passed in <proto> must be usable on this family. The protocol's default iocb
* is automatically preset as the receivers' iocb. The number of jobs and
* listeners is automatically increased by the number of listeners created. It
* returns non-zero on success, zero on error with the error message set in <err>.
*/
int create_listeners(struct bind_conf *bc, const struct sockaddr_storage *ss,
int portl, int porth, int fd, struct protocol *proto, char **err)
{
struct listener *l;
int port;
for (port = portl; port <= porth; port++) {
l = calloc(1, sizeof(*l));
if (!l) {
memprintf(err, "out of memory");
return 0;
}
l->obj_type = OBJ_TYPE_LISTENER;
LIST_ADDQ(&bc->frontend->conf.listeners, &l->by_fe);
LIST_ADDQ(&bc->listeners, &l->by_bind);
l->bind_conf = bc;
l->rx.settings = &bc->settings;
l->rx.owner = l;
l->rx.iocb = proto->default_iocb;
l->rx.fd = fd;
memcpy(&l->rx.addr, ss, sizeof(*ss));
if (proto->fam->set_port)
proto->fam->set_port(&l->rx.addr, port);
MT_LIST_INIT(&l->wait_queue);
listener_set_state(l, LI_INIT);
proto->add(proto, l);
if (fd != -1)
l->rx.flags |= RX_F_INHERITED;
l->extra_counters = NULL;
HA_SPIN_INIT(&l->lock);
_HA_ATOMIC_ADD(&jobs, 1);
_HA_ATOMIC_ADD(&listeners, 1);
}
return 1;
}
/* Delete a listener from its protocol's list of listeners. The listener's
* state is automatically updated from LI_ASSIGNED to LI_INIT. The protocol's
* number of listeners is updated, as well as the global number of listeners
* and jobs. Note that the listener must have previously been unbound. This
* is a low-level function expected to be called with the proto_lock and the
* listener's lock held.
*/
void __delete_listener(struct listener *listener)
{
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
if (listener->state == LI_ASSIGNED) {
listener_set_state(listener, LI_INIT);
LIST_DEL(&listener->rx.proto_list);
listener->rx.proto->nb_receivers--;
_HA_ATOMIC_SUB(&jobs, 1);
_HA_ATOMIC_SUB(&listeners, 1);
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
}
}
/* Delete a listener from its protocol's list of listeners (please check
* __delete_listener() above). The proto_lock and the listener's lock will
* be grabbed in this order.
*/
void delete_listener(struct listener *listener)
{
HA_SPIN_LOCK(PROTO_LOCK, &proto_lock);
HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
__delete_listener(listener);
HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
HA_SPIN_UNLOCK(PROTO_LOCK, &proto_lock);
}
/* Returns a suitable value for a listener's backlog. It uses the listener's,
* otherwise the frontend's backlog, otherwise the listener's maxconn,
* otherwise the frontend's maxconn, otherwise 1024.
*/
int listener_backlog(const struct listener *l)
{
if (l->backlog)
return l->backlog;
if (l->bind_conf->frontend->backlog)
return l->bind_conf->frontend->backlog;
if (l->maxconn)
return l->maxconn;
if (l->bind_conf->frontend->maxconn)
return l->bind_conf->frontend->maxconn;
return 1024;
}
/* This function is called on a read event from a listening socket, corresponding
* to an accept. It tries to accept as many connections as possible, and for each
* calls the listener's accept handler (generally the frontend's accept handler).
*/
void listener_accept(struct listener *l)
{
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
struct connection *cli_conn;
struct proxy *p;
unsigned int max_accept;
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
int next_conn = 0;
int next_feconn = 0;
int next_actconn = 0;
int expire;
int ret;
p = l->bind_conf->frontend;
/* if l->maxaccept is -1, then max_accept is UINT_MAX. It is not really
* illimited, but it is probably enough.
*/
max_accept = l->maxaccept ? l->maxaccept : 1;
if (!(l->options & LI_O_UNLIMITED) && global.sps_lim) {
int max = freq_ctr_remain(&global.sess_per_sec, global.sps_lim, 0);
if (unlikely(!max)) {
/* frontend accept rate limit was reached */
expire = tick_add(now_ms, next_event_delay(&global.sess_per_sec, global.sps_lim, 0));
goto limit_global;
}
if (max_accept > max)
max_accept = max;
}
if (!(l->options & LI_O_UNLIMITED) && global.cps_lim) {
int max = freq_ctr_remain(&global.conn_per_sec, global.cps_lim, 0);
if (unlikely(!max)) {
/* frontend accept rate limit was reached */
expire = tick_add(now_ms, next_event_delay(&global.conn_per_sec, global.cps_lim, 0));
goto limit_global;
}
if (max_accept > max)
max_accept = max;
}
#ifdef USE_OPENSSL
if (!(l->options & LI_O_UNLIMITED) && global.ssl_lim && l->bind_conf && l->bind_conf->is_ssl) {
int max = freq_ctr_remain(&global.ssl_per_sec, global.ssl_lim, 0);
if (unlikely(!max)) {
/* frontend accept rate limit was reached */
expire = tick_add(now_ms, next_event_delay(&global.ssl_per_sec, global.ssl_lim, 0));
goto limit_global;
}
if (max_accept > max)
max_accept = max;
}
#endif
if (p && p->fe_sps_lim) {
int max = freq_ctr_remain(&p->fe_sess_per_sec, p->fe_sps_lim, 0);
if (unlikely(!max)) {
/* frontend accept rate limit was reached */
expire = tick_add(now_ms, next_event_delay(&p->fe_sess_per_sec, p->fe_sps_lim, 0));
goto limit_proxy;
}
if (max_accept > max)
max_accept = max;
}
/* Note: if we fail to allocate a connection because of configured
* limits, we'll schedule a new attempt worst 1 second later in the
* worst case. If we fail due to system limits or temporary resource
* shortage, we try again 100ms later in the worst case.
*/
for (; max_accept; next_conn = next_feconn = next_actconn = 0, max_accept--) {
unsigned int count;
int status;
__decl_thread(unsigned long mask);
/* pre-increase the number of connections without going too far.
* We process the listener, then the proxy, then the process.
* We know which ones to unroll based on the next_xxx value.
*/
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
do {
count = l->nbconn;
if (unlikely(l->maxconn && count >= l->maxconn)) {
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
/* the listener was marked full or another
* thread is going to do it.
*/
next_conn = 0;
listener_full(l);
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
goto end;
}
next_conn = count + 1;
} while (!_HA_ATOMIC_CAS(&l->nbconn, (int *)(&count), next_conn));
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
if (p) {
do {
count = p->feconn;
if (unlikely(count >= p->maxconn)) {
/* the frontend was marked full or another
* thread is going to do it.
*/
next_feconn = 0;
expire = TICK_ETERNITY;
goto limit_proxy;
}
next_feconn = count + 1;
} while (!_HA_ATOMIC_CAS(&p->feconn, &count, next_feconn));
}
if (!(l->options & LI_O_UNLIMITED)) {
do {
count = actconn;
if (unlikely(count >= global.maxconn)) {
/* the process was marked full or another
* thread is going to do it.
*/
next_actconn = 0;
expire = tick_add(now_ms, 1000); /* try again in 1 second */
goto limit_global;
}
next_actconn = count + 1;
} while (!_HA_ATOMIC_CAS(&actconn, (int *)(&count), next_actconn));
}
cli_conn = l->rx.proto->accept_conn(l, &status);
if (!cli_conn) {
switch (status) {
case CO_AC_DONE:
goto end;
case CO_AC_RETRY: /* likely a signal */
_HA_ATOMIC_SUB(&l->nbconn, 1);
if (p)
_HA_ATOMIC_SUB(&p->feconn, 1);
if (!(l->options & LI_O_UNLIMITED))
_HA_ATOMIC_SUB(&actconn, 1);
continue;
case CO_AC_YIELD:
max_accept = 0;
goto end;
default:
goto transient_error;
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
}
}
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
/* The connection was accepted, it must be counted as such */
if (l->counters)
HA_ATOMIC_UPDATE_MAX(&l->counters->conn_max, next_conn);
if (p)
HA_ATOMIC_UPDATE_MAX(&p->fe_counters.conn_max, next_feconn);
proxy_inc_fe_conn_ctr(l, p);
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
if (!(l->options & LI_O_UNLIMITED)) {
count = update_freq_ctr(&global.conn_per_sec, 1);
HA_ATOMIC_UPDATE_MAX(&global.cps_max, count);
}
_HA_ATOMIC_ADD(&activity[tid].accepted, 1);
if (unlikely(cli_conn->handle.fd >= global.maxsock)) {
send_log(p, LOG_EMERG,
"Proxy %s reached the configured maximum connection limit. Please check the global 'maxconn' value.\n",
p->id);
close(cli_conn->handle.fd);
conn_free(cli_conn);
expire = tick_add(now_ms, 1000); /* try again in 1 second */
goto limit_global;
}
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
/* past this point, l->accept() will automatically decrement
* l->nbconn, feconn and actconn once done. Setting next_*conn=0
* allows the error path not to rollback on nbconn. It's more
* convenient than duplicating all exit labels.
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
*/
next_conn = 0;
next_feconn = 0;
next_actconn = 0;
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
#if defined(USE_THREAD)
mask = thread_mask(l->rx.settings->bind_thread) & all_threads_mask;
if (atleast2(mask) && (global.tune.options & GTUNE_LISTENER_MQ) && !stopping) {
struct accept_queue_ring *ring;
unsigned int t, t0, t1, t2;
/* The principle is that we have two running indexes,
* each visiting in turn all threads bound to this
* listener. The connection will be assigned to the one
* with the least connections, and the other one will
* be updated. This provides a good fairness on short
* connections (round robin) and on long ones (conn
* count), without ever missing any idle thread.
*/
/* keep a copy for the final update. thr_idx is composite
* and made of (t2<<16) + t1.
*/
t0 = l->thr_idx;
do {
unsigned long m1, m2;
int q1, q2;
t2 = t1 = t0;
t2 >>= 16;
t1 &= 0xFFFF;
/* t1 walks low to high bits ;
* t2 walks high to low.
*/
m1 = mask >> t1;
m2 = mask & (t2 ? nbits(t2 + 1) : ~0UL);
if (unlikely(!(m1 & 1))) {
m1 &= ~1UL;
if (!m1) {
m1 = mask;
t1 = 0;
}
t1 += my_ffsl(m1) - 1;
}
if (unlikely(!(m2 & (1UL << t2)) || t1 == t2)) {
/* highest bit not set */
if (!m2)
m2 = mask;
t2 = my_flsl(m2) - 1;
}
/* now we have two distinct thread IDs belonging to the mask */
q1 = accept_queue_rings[t1].tail - accept_queue_rings[t1].head + ACCEPT_QUEUE_SIZE;
if (q1 >= ACCEPT_QUEUE_SIZE)
q1 -= ACCEPT_QUEUE_SIZE;
q2 = accept_queue_rings[t2].tail - accept_queue_rings[t2].head + ACCEPT_QUEUE_SIZE;
if (q2 >= ACCEPT_QUEUE_SIZE)
q2 -= ACCEPT_QUEUE_SIZE;
/* we have 3 possibilities now :
* q1 < q2 : t1 is less loaded than t2, so we pick it
* and update t2 (since t1 might still be
* lower than another thread)
* q1 > q2 : t2 is less loaded than t1, so we pick it
* and update t1 (since t2 might still be
* lower than another thread)
* q1 = q2 : both are equally loaded, thus we pick t1
* and update t1 as it will become more loaded
* than t2.
*/
q1 += l->thr_conn[t1];
q2 += l->thr_conn[t2];
if (q1 - q2 < 0) {
t = t1;
t2 = t2 ? t2 - 1 : LONGBITS - 1;
}
else if (q1 - q2 > 0) {
t = t2;
t1++;
if (t1 >= LONGBITS)
t1 = 0;
}
else {
t = t1;
t1++;
if (t1 >= LONGBITS)
t1 = 0;
}
/* new value for thr_idx */
t1 += (t2 << 16);
} while (unlikely(!_HA_ATOMIC_CAS(&l->thr_idx, &t0, t1)));
/* We successfully selected the best thread "t" for this
* connection. We use deferred accepts even if it's the
* local thread because tests show that it's the best
* performing model, likely due to better cache locality
* when processing this loop.
*/
ring = &accept_queue_rings[t];
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
if (accept_queue_push_mp(ring, cli_conn)) {
_HA_ATOMIC_ADD(&activity[t].accq_pushed, 1);
tasklet_wakeup(ring->tasklet);
continue;
}
/* If the ring is full we do a synchronous accept on
* the local thread here.
*/
_HA_ATOMIC_ADD(&activity[t].accq_full, 1);
}
#endif // USE_THREAD
_HA_ATOMIC_ADD(&l->thr_conn[tid], 1);
MEDIUM: listener: allocate the connection before queuing a new connection Till now we would keep a per-thread queue of pending incoming connections for which we would store: - the listener - the accepted FD - the source address - the source address' length And these elements were first used in session_accept_fd() running on the target thread to allocate a connection and duplicate them again. Doing this induces various problems. The first one is that session_accept_fd() may only run on file descriptors and cannot be reused for QUIC. The second issue is that it induces lots of memory copies and that the listerner queue thrashes a lot of cache, consuming 64 bytes per entry. This patch changes this by allocating the connection before queueing it, and by only placing the connection's pointer into the queue. Indeed, the first two calls used to initialize the connection already store all the information above, which can be retrieved from the connection pointer alone. So we just have to pop one pointer from the target thread, and pass it to session_accept_fd() which only needs the FD for the final settings. This starts to make the accept path a bit more transport-agnostic, and saves memory and CPU cycles at the same time (1% connection rate increase was noticed with 4 threads). Thanks to dividing the accept-queue entry size from 64 to 8 bytes, its size could be increased from 256 to 1024 connections while still dividing the overall size by two. No single queue full condition was met. One minor drawback is that connection may be allocated from one thread's pool to be used into another one. But this already happens a lot with connection reuse so there is really nothing new here.
2020-10-14 15:37:17 +00:00
ret = l->accept(cli_conn);
if (unlikely(ret <= 0)) {
REORG/MAJOR: session: rename the "session" entity to "stream" With HTTP/2, we'll have to support multiplexed streams. A stream is in fact the largest part of what we currently call a session, it has buffers, logs, etc. In order to catch any error, this commit removes any reference to the struct session and tries to rename most "session" occurrences in function names to "stream" and "sess" to "strm" when that's related to a session. The files stream.{c,h} were added and session.{c,h} removed. The session will be reintroduced later and a few parts of the stream will progressively be moved overthere. It will more or less contain only what we need in an embryonic session. Sample fetch functions and converters will have to change a bit so that they'll use an L5 (session) instead of what's currently called "L4" which is in fact L6 for now. Once all changes are completed, we should see approximately this : L7 - http_txn L6 - stream L5 - session L4 - connection | applet There will be at most one http_txn per stream, and a same session will possibly be referenced by multiple streams. A connection will point to a session and to a stream. The session will hold all the information we need to keep even when we don't yet have a stream. Some more cleanup is needed because some code was already far from being clean. The server queue management still refers to sessions at many places while comments talk about connections. This will have to be cleaned up once we have a server-side connection pool manager. Stream flags "SN_*" still need to be renamed, it doesn't seem like any of them will need to move to the session.
2015-04-02 22:22:06 +00:00
/* The connection was closed by stream_accept(). Either
* we just have to ignore it (ret == 0) or it's a critical
* error due to a resource shortage, and we must stop the
* listener (ret < 0).
*/
if (ret == 0) /* successful termination */
continue;
goto transient_error;
}
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
/* increase the per-process number of cumulated sessions, this
* may only be done once l->accept() has accepted the connection.
*/
if (!(l->options & LI_O_UNLIMITED)) {
count = update_freq_ctr(&global.sess_per_sec, 1);
HA_ATOMIC_UPDATE_MAX(&global.sps_max, count);
}
#ifdef USE_OPENSSL
if (!(l->options & LI_O_UNLIMITED) && l->bind_conf && l->bind_conf->is_ssl) {
count = update_freq_ctr(&global.ssl_per_sec, 1);
HA_ATOMIC_UPDATE_MAX(&global.ssl_max, count);
}
#endif
ti->flags &= ~TI_FL_STUCK; // this thread is still running
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
} /* end of for (max_accept--) */
end:
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
if (next_conn)
_HA_ATOMIC_SUB(&l->nbconn, 1);
if (p && next_feconn)
_HA_ATOMIC_SUB(&p->feconn, 1);
if (next_actconn)
_HA_ATOMIC_SUB(&actconn, 1);
if ((l->state == LI_FULL && (!l->maxconn || l->nbconn < l->maxconn)) ||
(l->state == LI_LIMITED &&
BUG/MINOR: listener: do not immediately resume on transient error The listener supports a "transient error" situation, which corresponds to those situations where accept fails badly but poll() reports an event. This happens for example when a listener is paused, or on out of FD. The same mechanism is used when facing a maxconn or maxsessrate limitation. When this happens, the listener is disabled for up to 100ms and put back into the global listener queue so that it automatically wakes up again as soon as the conditions change from an existing connection releasing one resource, or the system recovers from a transient issue. The listener_accept() function has a bug in its exit path causing a freshly limited listener to be immediately enabled again because all the conditions are met (connection count < max). It doesn't take into account the fact that the listener might have been queued and must first wait for the timeout to expire before doing so. The impact is that upon certain errors, the faulty process will busy loop on the accept code without sleeping. This is the scenario reported and diagnosed by @hedong0411 in issue #382. This commit fixes it by verifying that the global queue's delay is at least expired before deciding to resume the listener. Another approach could consist in having an extra state like LI_DELAY for situations where only a delay is acceptable, but this would probably not bring anything except more complex code. This issue was introduced with the lock-free listener accept code (commits 3f0d02b and 82c9789a) that were backported to 1.8.20+ and 1.9.7+, so this fix must be backported to the relevant branches.
2019-12-11 14:06:30 +00:00
((!p || p->feconn < p->maxconn) && (actconn < global.maxconn) &&
(!tick_isset(global_listener_queue_task->expire) ||
tick_is_expired(global_listener_queue_task->expire, now_ms))))) {
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
/* at least one thread has to this when quitting */
resume_listener(l);
/* Dequeues all of the listeners waiting for a resource */
dequeue_all_listeners();
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
if (p && !MT_LIST_ISEMPTY(&p->listener_queue) &&
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
(!p->fe_sps_lim || freq_ctr_remain(&p->fe_sess_per_sec, p->fe_sps_lim, 0) > 0))
dequeue_proxy_listeners(p);
MAJOR: listener: do not hold the listener lock in listener_accept() This function used to hold the listener's lock as a way to stay safe against concurrent manipulations, but it turns out this is wrong. First, the lock is held during l->accept(), which itself might indirectly call listener_release(), which, if the listener is marked full, could result in __resume_listener() to be called and the lock being taken twice. In practice it doesn't happen right now because the listener's FULL state cannot change while we're doing this. Second, all the code does is now protected against concurrent accesses. It used not to be the case in the early days of threads : the frequency counters are thread-safe. The rate limiting doesn't require extreme precision. Only the nbconn check is not thread safe. Third, the parts called here will have to be called from different threads without holding this lock, and this becomes a bigger issue if we need to keep this one. This patch does 3 things which need to be addressed at once : 1) it moves the lock to the only 2 functions that were not protected since called form listener_accept() : - limit_listener() - listener_full() 2) it makes sure delete_listener() properly checks its state within the lock. 3) it updates the l->nbconn tracking to make sure that it is always properly reported and accounted for. There is a point of particular care around the situation where the listener's maxconn is reached because the listener has to be marked full before accepting the connection, then resumed if the connection finally gets dropped. It is not possible to perform this change without removing the lock due to the deadlock issue explained above. This patch almost doubles the accept rate in multi-thread on a shared port between 8 threads, and multiplies by 4 the connection rate on a tcp-request connection reject rule.
2019-02-25 18:23:37 +00:00
}
return;
transient_error:
/* pause the listener for up to 100 ms */
expire = tick_add(now_ms, 100);
/* This may be a shared socket that was paused by another process.
* Let's put it to pause in this case.
*/
if (l->rx.proto && l->rx.proto->rx_listening(&l->rx) == 0) {
pause_listener(l);
goto end;
}
limit_global:
/* (re-)queue the listener to the global queue and set it to expire no
* later than <expire> ahead. The listener turns to LI_LIMITED.
*/
limit_listener(l, &global_listener_queue);
task_schedule(global_listener_queue_task, expire);
goto end;
limit_proxy:
/* (re-)queue the listener to the proxy's queue and set it to expire no
* later than <expire> ahead. The listener turns to LI_LIMITED.
*/
limit_listener(l, &p->listener_queue);
BUG/MAJOR: listener: do not schedule a task-less proxy Apparently seamingless commit 0591bf7deb ("MINOR: listener: make the wait paths cleaner and more reliable") caused a nasty regression and revealed a rare race that hits regtest stickiness/lb-services.vtc about 4% of the times for 8 threads. The problem is that when a multi-threaded listener wakes up on an incoming connection, several threads can receive the event, especially when idle. And all of them will race to accept the connections in parallel, adjusting the listener's nbconn and proxy's feconn until one reaches the proxy's limit and declines. At this step the changes are cancelled, the listener is marked "limited", and when the threads exit the function, one of them will unlimit the listener/proxy again so that it can accept incoming connections again. The problem happens when many threads connect to a small peers section because its maxconn is very limited (typically 6 for 2 peers), and it's sometimes possible for enough competing threads to hit the limit and one of them will limit the listener and queue the proxy's task... except that peers do not initialize their proxy task since they do not use rate limiting. Thus the process crashes when doing task_schedule(p->task). Prior to the cleanup patch above, this didn't happen because the error path that was dedicated to only limiting the listener did not call task_schedule(p->task). Given that the proxy's task is optional, and that the expire value passed there is always TICK_ETERNITY, it's sufficient and reasonable to avoid calling this task_schedule() when expire is not set. And for long term safety we can also avoid to do it when the task is not set. A first fix consisted in allocating a task for the peers proxies but it's never used and would eat resources for reason. No backport is needed as this commit was only merged into 2.2.
2020-01-08 18:15:07 +00:00
if (p->task && tick_isset(expire))
task_schedule(p->task, expire);
goto end;
}
/* Notify the listener that a connection initiated from it was released. This
* is used to keep the connection count consistent and to possibly re-open
* listening when it was limited.
*/
void listener_release(struct listener *l)
{
struct proxy *fe = l->bind_conf->frontend;
if (!(l->options & LI_O_UNLIMITED))
_HA_ATOMIC_SUB(&actconn, 1);
if (fe)
_HA_ATOMIC_SUB(&fe->feconn, 1);
_HA_ATOMIC_SUB(&l->nbconn, 1);
_HA_ATOMIC_SUB(&l->thr_conn[tid], 1);
if (l->state == LI_FULL || l->state == LI_LIMITED)
resume_listener(l);
/* Dequeues all of the listeners waiting for a resource */
dequeue_all_listeners();
if (!MT_LIST_ISEMPTY(&fe->listener_queue) &&
(!fe->fe_sps_lim || freq_ctr_remain(&fe->fe_sess_per_sec, fe->fe_sps_lim, 0) > 0))
dequeue_proxy_listeners(fe);
}
BUG/MAJOR: listener: fix thread safety in resume_listener() resume_listener() can be called from a thread not part of the listener's mask after a curr_conn has gone lower than a proxy's or the process' limit. This results in fd_may_recv() being called unlocked if the listener is bound to only one thread, and quickly locks up. This patch solves this by creating a per-thread work_list dedicated to listeners, and modifying resume_listener() so that it bounces the listener to one of its owning thread's work_list and waking it up. This thread will then call resume_listener() again and will perform the operation on the file descriptor itself. It is important to do it this way so that the listener's state cannot be modified while the listener is being moved, otherwise multiple threads can take conflicting decisions and the listener could be put back into the global queue if the listener was used at the same time. It seems like a slightly simpler approach would be possible if the locked list API would provide the ability to return a locked element. In this case the listener would be immediately requeued in dequeue_all_listeners() without having to go through resume_listener() with its associated lock. This fix must be backported to all versions having the lock-less accept loop, which is as far as 1.8 since deadlock fixes involving this feature had to be backported there. It is expected that the code should not differ too much there. However, previous commit "MINOR: task: introduce work lists" will be needed as well and should not present difficulties either. For 1.8, the commits introducing thread_mask() and LIST_ADDED() will be needed as well, either backporting my_flsl() or switching to my_ffsl() will be OK, and some changes will have to be performed so that the init function is properly called (and maybe the deinit one can be dropped). In order to test for the fix, simply set up a multi-threaded frontend with multiple bind lines each attached to a single thread (reproduced with 16 threads here), set up a very low maxconn value on the frontend, and inject heavy traffic on all listeners in parallel with slightly more connections than the configured limit ( typically +20%) so that it flips very frequently. If the bug is still there, at some point (5-20 seconds) the traffic will go much lower or even stop, either with spinning threads or not.
2019-07-11 08:08:31 +00:00
/* Initializes the listener queues. Returns 0 on success, otherwise ERR_* flags */
static int listener_queue_init()
{
global_listener_queue_task = task_new(MAX_THREADS_MASK);
if (!global_listener_queue_task) {
ha_alert("Out of memory when initializing global listener queue\n");
return ERR_FATAL|ERR_ABORT;
}
/* very simple initialization, users will queue the task if needed */
global_listener_queue_task->context = NULL; /* not even a context! */
global_listener_queue_task->process = manage_global_listener_queue;
BUG/MAJOR: listener: fix thread safety in resume_listener() resume_listener() can be called from a thread not part of the listener's mask after a curr_conn has gone lower than a proxy's or the process' limit. This results in fd_may_recv() being called unlocked if the listener is bound to only one thread, and quickly locks up. This patch solves this by creating a per-thread work_list dedicated to listeners, and modifying resume_listener() so that it bounces the listener to one of its owning thread's work_list and waking it up. This thread will then call resume_listener() again and will perform the operation on the file descriptor itself. It is important to do it this way so that the listener's state cannot be modified while the listener is being moved, otherwise multiple threads can take conflicting decisions and the listener could be put back into the global queue if the listener was used at the same time. It seems like a slightly simpler approach would be possible if the locked list API would provide the ability to return a locked element. In this case the listener would be immediately requeued in dequeue_all_listeners() without having to go through resume_listener() with its associated lock. This fix must be backported to all versions having the lock-less accept loop, which is as far as 1.8 since deadlock fixes involving this feature had to be backported there. It is expected that the code should not differ too much there. However, previous commit "MINOR: task: introduce work lists" will be needed as well and should not present difficulties either. For 1.8, the commits introducing thread_mask() and LIST_ADDED() will be needed as well, either backporting my_flsl() or switching to my_ffsl() will be OK, and some changes will have to be performed so that the init function is properly called (and maybe the deinit one can be dropped). In order to test for the fix, simply set up a multi-threaded frontend with multiple bind lines each attached to a single thread (reproduced with 16 threads here), set up a very low maxconn value on the frontend, and inject heavy traffic on all listeners in parallel with slightly more connections than the configured limit ( typically +20%) so that it flips very frequently. If the bug is still there, at some point (5-20 seconds) the traffic will go much lower or even stop, either with spinning threads or not.
2019-07-11 08:08:31 +00:00
return 0;
}
static void listener_queue_deinit()
{
task_destroy(global_listener_queue_task);
global_listener_queue_task = NULL;
BUG/MAJOR: listener: fix thread safety in resume_listener() resume_listener() can be called from a thread not part of the listener's mask after a curr_conn has gone lower than a proxy's or the process' limit. This results in fd_may_recv() being called unlocked if the listener is bound to only one thread, and quickly locks up. This patch solves this by creating a per-thread work_list dedicated to listeners, and modifying resume_listener() so that it bounces the listener to one of its owning thread's work_list and waking it up. This thread will then call resume_listener() again and will perform the operation on the file descriptor itself. It is important to do it this way so that the listener's state cannot be modified while the listener is being moved, otherwise multiple threads can take conflicting decisions and the listener could be put back into the global queue if the listener was used at the same time. It seems like a slightly simpler approach would be possible if the locked list API would provide the ability to return a locked element. In this case the listener would be immediately requeued in dequeue_all_listeners() without having to go through resume_listener() with its associated lock. This fix must be backported to all versions having the lock-less accept loop, which is as far as 1.8 since deadlock fixes involving this feature had to be backported there. It is expected that the code should not differ too much there. However, previous commit "MINOR: task: introduce work lists" will be needed as well and should not present difficulties either. For 1.8, the commits introducing thread_mask() and LIST_ADDED() will be needed as well, either backporting my_flsl() or switching to my_ffsl() will be OK, and some changes will have to be performed so that the init function is properly called (and maybe the deinit one can be dropped). In order to test for the fix, simply set up a multi-threaded frontend with multiple bind lines each attached to a single thread (reproduced with 16 threads here), set up a very low maxconn value on the frontend, and inject heavy traffic on all listeners in parallel with slightly more connections than the configured limit ( typically +20%) so that it flips very frequently. If the bug is still there, at some point (5-20 seconds) the traffic will go much lower or even stop, either with spinning threads or not.
2019-07-11 08:08:31 +00:00
}
REGISTER_CONFIG_POSTPARSER("multi-threaded listener queue", listener_queue_init);
REGISTER_POST_DEINIT(listener_queue_deinit);
/* This is the global management task for listeners. It enables listeners waiting
* for global resources when there are enough free resource, or at least once in
* a while. It is designed to be called as a task. It's exported so that it's easy
* to spot in "show tasks" or "show profiling".
*/
struct task *manage_global_listener_queue(struct task *t, void *context, unsigned int state)
{
/* If there are still too many concurrent connections, let's wait for
* some of them to go away. We don't need to re-arm the timer because
* each of them will scan the queue anyway.
*/
if (unlikely(actconn >= global.maxconn))
goto out;
/* We should periodically try to enable listeners waiting for a global
* resource here, because it is possible, though very unlikely, that
* they have been blocked by a temporary lack of global resource such
* as a file descriptor or memory and that the temporary condition has
* disappeared.
*/
dequeue_all_listeners();
out:
t->expire = TICK_ETERNITY;
task_queue(t);
return t;
}
/*
* Registers the bind keyword list <kwl> as a list of valid keywords for next
* parsing sessions.
*/
void bind_register_keywords(struct bind_kw_list *kwl)
{
LIST_ADDQ(&bind_keywords.list, &kwl->list);
}
/* Return a pointer to the bind keyword <kw>, or NULL if not found. If the
* keyword is found with a NULL ->parse() function, then an attempt is made to
* find one with a valid ->parse() function. This way it is possible to declare
* platform-dependant, known keywords as NULL, then only declare them as valid
* if some options are met. Note that if the requested keyword contains an
* opening parenthesis, everything from this point is ignored.
*/
struct bind_kw *bind_find_kw(const char *kw)
{
int index;
const char *kwend;
struct bind_kw_list *kwl;
struct bind_kw *ret = NULL;
kwend = strchr(kw, '(');
if (!kwend)
kwend = kw + strlen(kw);
list_for_each_entry(kwl, &bind_keywords.list, list) {
for (index = 0; kwl->kw[index].kw != NULL; index++) {
if ((strncmp(kwl->kw[index].kw, kw, kwend - kw) == 0) &&
kwl->kw[index].kw[kwend-kw] == 0) {
if (kwl->kw[index].parse)
return &kwl->kw[index]; /* found it !*/
else
ret = &kwl->kw[index]; /* may be OK */
}
}
}
return ret;
}
/* Dumps all registered "bind" keywords to the <out> string pointer. The
* unsupported keywords are only dumped if their supported form was not
* found.
*/
void bind_dump_kws(char **out)
{
struct bind_kw_list *kwl;
int index;
if (!out)
return;
*out = NULL;
list_for_each_entry(kwl, &bind_keywords.list, list) {
for (index = 0; kwl->kw[index].kw != NULL; index++) {
if (kwl->kw[index].parse ||
bind_find_kw(kwl->kw[index].kw) == &kwl->kw[index]) {
memprintf(out, "%s[%4s] %s%s%s\n", *out ? *out : "",
kwl->scope,
kwl->kw[index].kw,
kwl->kw[index].skip ? " <arg>" : "",
kwl->kw[index].parse ? "" : " (not supported)");
}
}
}
}
/************************************************************************/
/* All supported sample and ACL keywords must be declared here. */
/************************************************************************/
/* set temp integer to the number of connexions to the same listening socket */
static int
smp_fetch_dconn(const struct arg *args, struct sample *smp, const char *kw, void *private)
{
smp->data.type = SMP_T_SINT;
smp->data.u.sint = smp->sess->listener->nbconn;
return 1;
}
/* set temp integer to the id of the socket (listener) */
static int
smp_fetch_so_id(const struct arg *args, struct sample *smp, const char *kw, void *private)
{
smp->data.type = SMP_T_SINT;
smp->data.u.sint = smp->sess->listener->luid;
return 1;
}
static int
smp_fetch_so_name(const struct arg *args, struct sample *smp, const char *kw, void *private)
{
smp->data.u.str.area = smp->sess->listener->name;
if (!smp->data.u.str.area)
return 0;
smp->data.type = SMP_T_STR;
smp->flags = SMP_F_CONST;
smp->data.u.str.data = strlen(smp->data.u.str.area);
return 1;
}
/* parse the "accept-proxy" bind keyword */
static int bind_parse_accept_proxy(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
struct listener *l;
list_for_each_entry(l, &conf->listeners, by_bind)
l->options |= LI_O_ACC_PROXY;
return 0;
}
/* parse the "accept-netscaler-cip" bind keyword */
static int bind_parse_accept_netscaler_cip(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
struct listener *l;
uint32_t val;
if (!*args[cur_arg + 1]) {
memprintf(err, "'%s' : missing value", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
val = atol(args[cur_arg + 1]);
if (val <= 0) {
memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
return ERR_ALERT | ERR_FATAL;
}
list_for_each_entry(l, &conf->listeners, by_bind) {
l->options |= LI_O_ACC_CIP;
conf->ns_cip_magic = val;
}
return 0;
}
/* parse the "backlog" bind keyword */
static int bind_parse_backlog(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
struct listener *l;
int val;
if (!*args[cur_arg + 1]) {
memprintf(err, "'%s' : missing value", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
val = atol(args[cur_arg + 1]);
if (val < 0) {
memprintf(err, "'%s' : invalid value %d, must be > 0", args[cur_arg], val);
return ERR_ALERT | ERR_FATAL;
}
list_for_each_entry(l, &conf->listeners, by_bind)
l->backlog = val;
return 0;
}
/* parse the "id" bind keyword */
static int bind_parse_id(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
struct eb32_node *node;
struct listener *l, *new;
char *error;
if (conf->listeners.n != conf->listeners.p) {
memprintf(err, "'%s' can only be used with a single socket", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
if (!*args[cur_arg + 1]) {
memprintf(err, "'%s' : expects an integer argument", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
new = LIST_NEXT(&conf->listeners, struct listener *, by_bind);
new->luid = strtol(args[cur_arg + 1], &error, 10);
if (*error != '\0') {
memprintf(err, "'%s' : expects an integer argument, found '%s'", args[cur_arg], args[cur_arg + 1]);
return ERR_ALERT | ERR_FATAL;
}
new->conf.id.key = new->luid;
if (new->luid <= 0) {
memprintf(err, "'%s' : custom id has to be > 0", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
node = eb32_lookup(&px->conf.used_listener_id, new->luid);
if (node) {
l = container_of(node, struct listener, conf.id);
memprintf(err, "'%s' : custom id %d already used at %s:%d ('bind %s')",
args[cur_arg], l->luid, l->bind_conf->file, l->bind_conf->line,
l->bind_conf->arg);
return ERR_ALERT | ERR_FATAL;
}
eb32_insert(&px->conf.used_listener_id, &new->conf.id);
return 0;
}
/* parse the "maxconn" bind keyword */
static int bind_parse_maxconn(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
struct listener *l;
int val;
if (!*args[cur_arg + 1]) {
memprintf(err, "'%s' : missing value", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
val = atol(args[cur_arg + 1]);
if (val < 0) {
memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
return ERR_ALERT | ERR_FATAL;
}
list_for_each_entry(l, &conf->listeners, by_bind)
l->maxconn = val;
return 0;
}
/* parse the "name" bind keyword */
static int bind_parse_name(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
struct listener *l;
if (!*args[cur_arg + 1]) {
memprintf(err, "'%s' : missing name", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
list_for_each_entry(l, &conf->listeners, by_bind)
l->name = strdup(args[cur_arg + 1]);
return 0;
}
/* parse the "nice" bind keyword */
static int bind_parse_nice(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
struct listener *l;
int val;
if (!*args[cur_arg + 1]) {
memprintf(err, "'%s' : missing value", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
val = atol(args[cur_arg + 1]);
if (val < -1024 || val > 1024) {
memprintf(err, "'%s' : invalid value %d, allowed range is -1024..1024", args[cur_arg], val);
return ERR_ALERT | ERR_FATAL;
}
list_for_each_entry(l, &conf->listeners, by_bind)
l->nice = val;
return 0;
}
/* parse the "process" bind keyword */
static int bind_parse_process(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
char *slash;
unsigned long proc = 0, thread = 0;
if ((slash = strchr(args[cur_arg + 1], '/')) != NULL)
*slash = 0;
if (parse_process_number(args[cur_arg + 1], &proc, MAX_PROCS, NULL, err)) {
memprintf(err, "'%s' : %s", args[cur_arg], *err);
return ERR_ALERT | ERR_FATAL;
}
if (slash) {
if (parse_process_number(slash+1, &thread, MAX_THREADS, NULL, err)) {
memprintf(err, "'%s' : %s", args[cur_arg], *err);
return ERR_ALERT | ERR_FATAL;
}
*slash = '/';
}
conf->settings.bind_proc |= proc;
conf->settings.bind_thread |= thread;
return 0;
}
/* parse the "proto" bind keyword */
static int bind_parse_proto(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
{
struct ist proto;
if (!*args[cur_arg + 1]) {
memprintf(err, "'%s' : missing value", args[cur_arg]);
return ERR_ALERT | ERR_FATAL;
}
proto = ist(args[cur_arg + 1]);
conf->mux_proto = get_mux_proto(proto);
if (!conf->mux_proto) {
memprintf(err, "'%s' : unknown MUX protocol '%s'", args[cur_arg], args[cur_arg+1]);
return ERR_ALERT | ERR_FATAL;
}
return 0;
}
/* config parser for global "tune.listener.multi-queue", accepts "on" or "off" */
static int cfg_parse_tune_listener_mq(char **args, int section_type, struct proxy *curpx,
struct proxy *defpx, const char *file, int line,
char **err)
{
if (too_many_args(1, args, err, NULL))
return -1;
if (strcmp(args[1], "on") == 0)
global.tune.options |= GTUNE_LISTENER_MQ;
else if (strcmp(args[1], "off") == 0)
global.tune.options &= ~GTUNE_LISTENER_MQ;
else {
memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
return -1;
}
return 0;
}
/* Note: must not be declared <const> as its list will be overwritten.
* Please take care of keeping this list alphabetically sorted.
*/
static struct sample_fetch_kw_list smp_kws = {ILH, {
{ "dst_conn", smp_fetch_dconn, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
{ "so_id", smp_fetch_so_id, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
{ "so_name", smp_fetch_so_name, 0, NULL, SMP_T_STR, SMP_USE_FTEND, },
{ /* END */ },
}};
INITCALL1(STG_REGISTER, sample_register_fetches, &smp_kws);
/* Note: must not be declared <const> as its list will be overwritten.
* Please take care of keeping this list alphabetically sorted.
*/
static struct acl_kw_list acl_kws = {ILH, {
{ /* END */ },
}};
INITCALL1(STG_REGISTER, acl_register_keywords, &acl_kws);
/* Note: must not be declared <const> as its list will be overwritten.
* Please take care of keeping this list alphabetically sorted, doing so helps
* all code contributors.
* Optional keywords are also declared with a NULL ->parse() function so that
* the config parser can report an appropriate error when a known keyword was
* not enabled.
*/
static struct bind_kw_list bind_kws = { "ALL", { }, {
{ "accept-netscaler-cip", bind_parse_accept_netscaler_cip, 1 }, /* enable NetScaler Client IP insertion protocol */
{ "accept-proxy", bind_parse_accept_proxy, 0 }, /* enable PROXY protocol */
{ "backlog", bind_parse_backlog, 1 }, /* set backlog of listening socket */
{ "id", bind_parse_id, 1 }, /* set id of listening socket */
{ "maxconn", bind_parse_maxconn, 1 }, /* set maxconn of listening socket */
{ "name", bind_parse_name, 1 }, /* set name of listening socket */
{ "nice", bind_parse_nice, 1 }, /* set nice of listening socket */
{ "process", bind_parse_process, 1 }, /* set list of allowed process for this socket */
{ "proto", bind_parse_proto, 1 }, /* set the proto to use for all incoming connections */
{ /* END */ },
}};
INITCALL1(STG_REGISTER, bind_register_keywords, &bind_kws);
/* config keyword parsers */
static struct cfg_kw_list cfg_kws = {ILH, {
{ CFG_GLOBAL, "tune.listener.multi-queue", cfg_parse_tune_listener_mq },
{ 0, NULL, NULL }
}};
INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
/*
* Local variables:
* c-indent-level: 8
* c-basic-offset: 8
* End:
*/