MINOR: dynbuf: make the buffer wait queue per thread
The buffer wait queue used to be global historically but this doest not
make any sense anymore given that the most common use case is to have
thread-local pools. Thus there's no point waking up waiters of other
threads after releasing an entry, as they won't benefit from it.
Let's move the queue head to the thread_info structure and use
ti->buffer_wq from now on.
diff --git a/include/haproxy/channel.h b/include/haproxy/channel.h
index de51d04..812808c 100644
--- a/include/haproxy/channel.h
+++ b/include/haproxy/channel.h
@@ -852,7 +852,7 @@
return 1;
if (!MT_LIST_ADDED(&wait->list))
- MT_LIST_ADDQ(&buffer_wq, &wait->list);
+ MT_LIST_ADDQ(&ti->buffer_wq, &wait->list);
return 0;
}
diff --git a/include/haproxy/dynbuf.h b/include/haproxy/dynbuf.h
index 2a7aa47..c26ad21 100644
--- a/include/haproxy/dynbuf.h
+++ b/include/haproxy/dynbuf.h
@@ -35,7 +35,6 @@
#include <haproxy/pool.h>
extern struct pool_head *pool_head_buffer;
-extern struct mt_list buffer_wq;
int init_buffer();
void buffer_dump(FILE *o, struct buffer *b, int from, int to);
@@ -192,13 +191,13 @@
* passing a buffer to oneself in case of failed allocations (e.g. need two
* buffers, get one, fail, release it and wake up self again). In case of
* normal buffer release where it is expected that the caller is not waiting
- * for a buffer, NULL is fine.
+ * for a buffer, NULL is fine. It will wake waiters on the current thread only.
*/
void __offer_buffer(void *from, unsigned int threshold);
static inline void offer_buffers(void *from, unsigned int threshold)
{
- if (!MT_LIST_ISEMPTY(&buffer_wq))
+ if (!MT_LIST_ISEMPTY(&ti->buffer_wq))
__offer_buffer(from, threshold);
}
diff --git a/include/haproxy/tinfo-t.h b/include/haproxy/tinfo-t.h
index f3badbd..6dc0458 100644
--- a/include/haproxy/tinfo-t.h
+++ b/include/haproxy/tinfo-t.h
@@ -45,6 +45,8 @@
#ifdef CONFIG_HAP_LOCAL_POOLS
struct list pool_lru_head; /* oldest objects */
#endif
+ struct mt_list buffer_wq; /* buffer waiters */
+
/* pad to cache line (64B) */
char __pad[0]; /* unused except to check remaining room */
char __end[0] __attribute__((aligned(64)));
diff --git a/src/check.c b/src/check.c
index a09c67a..ff4effa 100644
--- a/src/check.c
+++ b/src/check.c
@@ -1019,7 +1019,7 @@
unlikely((buf = b_alloc_margin(bptr, 0)) == NULL)) {
check->buf_wait.target = check;
check->buf_wait.wakeup_cb = check_buf_available;
- MT_LIST_ADDQ(&buffer_wq, &check->buf_wait.list);
+ MT_LIST_ADDQ(&ti->buffer_wq, &check->buf_wait.list);
}
return buf;
}
diff --git a/src/dynbuf.c b/src/dynbuf.c
index ad2cde9..395fa8a 100644
--- a/src/dynbuf.c
+++ b/src/dynbuf.c
@@ -22,18 +22,20 @@
struct pool_head *pool_head_buffer;
-/* list of objects waiting for at least one buffer */
-struct mt_list buffer_wq = LIST_HEAD_INIT(buffer_wq);
-
/* perform minimal intializations, report 0 in case of error, 1 if OK. */
int init_buffer()
{
void *buffer;
+ int thr;
pool_head_buffer = create_pool("buffer", global.tune.bufsize, MEM_F_SHARED|MEM_F_EXACT);
if (!pool_head_buffer)
return 0;
+ for (thr = 0; thr < MAX_THREADS; thr++)
+ MT_LIST_INIT(&ha_thread_info[thr].buffer_wq);
+
+
/* The reserved buffer is what we leave behind us. Thus we always need
* at least one extra buffer in minavail otherwise we'll end up waking
* up tasks with no memory available, causing a lot of useless wakeups.
@@ -112,7 +114,7 @@
*/
avail = pool_head_buffer->allocated - pool_head_buffer->used - global.tune.reserved_bufs / 2;
- mt_list_for_each_entry_safe(wait, &buffer_wq, list, elt1, elt2) {
+ mt_list_for_each_entry_safe(wait, &ti->buffer_wq, list, elt1, elt2) {
if (avail <= threshold)
break;
diff --git a/src/flt_spoe.c b/src/flt_spoe.c
index b9471d9..27a15b6 100644
--- a/src/flt_spoe.c
+++ b/src/flt_spoe.c
@@ -2828,7 +2828,7 @@
if (b_alloc_margin(buf, global.tune.reserved_bufs))
return 1;
- MT_LIST_ADDQ(&buffer_wq, &buffer_wait->list);
+ MT_LIST_ADDQ(&ti->buffer_wq, &buffer_wait->list);
return 0;
}
diff --git a/src/mux_fcgi.c b/src/mux_fcgi.c
index a2c1c86..2ff7aa9 100644
--- a/src/mux_fcgi.c
+++ b/src/mux_fcgi.c
@@ -608,7 +608,7 @@
unlikely((buf = b_alloc_margin(bptr, 0)) == NULL)) {
fconn->buf_wait.target = fconn;
fconn->buf_wait.wakeup_cb = fcgi_buf_available;
- MT_LIST_ADDQ(&buffer_wq, &fconn->buf_wait.list);
+ MT_LIST_ADDQ(&ti->buffer_wq, &fconn->buf_wait.list);
}
return buf;
}
diff --git a/src/mux_h1.c b/src/mux_h1.c
index 6333935..077f4ea 100644
--- a/src/mux_h1.c
+++ b/src/mux_h1.c
@@ -452,7 +452,7 @@
unlikely((buf = b_alloc_margin(bptr, 0)) == NULL)) {
h1c->buf_wait.target = h1c;
h1c->buf_wait.wakeup_cb = h1_buf_available;
- MT_LIST_ADDQ(&buffer_wq, &h1c->buf_wait.list);
+ MT_LIST_ADDQ(&ti->buffer_wq, &h1c->buf_wait.list);
}
return buf;
}
diff --git a/src/mux_h2.c b/src/mux_h2.c
index 21c14b8..cfa5f8c 100644
--- a/src/mux_h2.c
+++ b/src/mux_h2.c
@@ -810,7 +810,7 @@
unlikely((buf = b_alloc_margin(bptr, 0)) == NULL)) {
h2c->buf_wait.target = h2c;
h2c->buf_wait.wakeup_cb = h2_buf_available;
- MT_LIST_ADDQ(&buffer_wq, &h2c->buf_wait.list);
+ MT_LIST_ADDQ(&ti->buffer_wq, &h2c->buf_wait.list);
}
return buf;
}
diff --git a/src/stream.c b/src/stream.c
index 9394ac0..d747211 100644
--- a/src/stream.c
+++ b/src/stream.c
@@ -773,7 +773,7 @@
if (b_alloc_margin(&s->res.buf, 0))
return 1;
- MT_LIST_ADDQ(&buffer_wq, &s->buffer_wait.list);
+ MT_LIST_ADDQ(&ti->buffer_wq, &s->buffer_wait.list);
return 0;
}