Hi,
The following piece of code in ll_rw_block() aims to limit the number of
locked buffers by making processes throttle on IO if the number of on
flight requests is bigger than a high watermaker. IO will only start
again if we're under a low watermark.
if (atomic_read(&queued_sectors) >= high_queued_sectors) {
run_task_queue(&tq_disk);
wait_event(blk_buffers_wait,
atomic_read(&queued_sectors) < low_queued_sectors);
}
However, if submit_bh() is used to queue IO (which is used by ->readpage()
for ext2, for example), no throttling happens.
It looks like ll_rw_block() users (writes, metadata reads) can be starved
by submit_bh() (data reads).
If I'm not missing something, the watermark check should be moved to
submit_bh().
On Thu, Feb 22 2001, Marcelo Tosatti wrote:
> The following piece of code in ll_rw_block() aims to limit the number of
> locked buffers by making processes throttle on IO if the number of on
> flight requests is bigger than a high watermaker. IO will only start
> again if we're under a low watermark.
>
> if (atomic_read(&queued_sectors) >= high_queued_sectors) {
> run_task_queue(&tq_disk);
> wait_event(blk_buffers_wait,
> atomic_read(&queued_sectors) < low_queued_sectors);
> }
>
>
> However, if submit_bh() is used to queue IO (which is used by ->readpage()
> for ext2, for example), no throttling happens.
>
> It looks like ll_rw_block() users (writes, metadata reads) can be starved
> by submit_bh() (data reads).
>
> If I'm not missing something, the watermark check should be moved to
> submit_bh().
We might as well put it there, the idea was to not lock this one
buffer either but I doubt this would make any different in reality :-)
Linus, could you apply?
--- /opt/kernel/linux-2.4.2/drivers/block/ll_rw_blk.c Thu Feb 22 14:55:22 2001
+++ drivers/block/ll_rw_blk.c Thu Feb 22 14:53:07 2001
@@ -957,6 +959,20 @@
if (!test_bit(BH_Lock, &bh->b_state))
BUG();
+ /*
+ * don't lock any more buffers if we are above the high
+ * water mark. instead start I/O on the queued stuff.
+ */
+ if (atomic_read(&queued_sectors) >= high_queued_sectors) {
+ run_task_queue(&tq_disk);
+ if (rw == READA) {
+ bh->b_end_io(bh, test_bit(BH_Uptodate, &bh->b_state));
+ return;
+ }
+ wait_event(blk_buffers_wait,
+ atomic_read(&queued_sectors) < low_queued_sectors);
+ }
+
set_bit(BH_Req, &bh->b_state);
/*
@@ -1057,16 +1073,6 @@
for (i = 0; i < nr; i++) {
struct buffer_head *bh = bhs[i];
-
- /*
- * don't lock any more buffers if we are above the high
- * water mark. instead start I/O on the queued stuff.
- */
- if (atomic_read(&queued_sectors) >= high_queued_sectors) {
- run_task_queue(&tq_disk);
- wait_event(blk_buffers_wait,
- atomic_read(&queued_sectors) < low_queued_sectors);
- }
/* Only one thread can actually submit the I/O. */
if (test_and_set_bit(BH_Lock, &bh->b_state))
--
Jens Axboe
On Thu, 22 Feb 2001, Jens Axboe wrote:
> On Thu, Feb 22 2001, Marcelo Tosatti wrote:
> > The following piece of code in ll_rw_block() aims to limit the number of
> > locked buffers by making processes throttle on IO if the number of on
> > flight requests is bigger than a high watermaker. IO will only start
> > again if we're under a low watermark.
> >
> > if (atomic_read(&queued_sectors) >= high_queued_sectors) {
> > run_task_queue(&tq_disk);
> > wait_event(blk_buffers_wait,
> > atomic_read(&queued_sectors) < low_queued_sectors);
> > }
> >
> >
> > However, if submit_bh() is used to queue IO (which is used by ->readpage()
> > for ext2, for example), no throttling happens.
> >
> > It looks like ll_rw_block() users (writes, metadata reads) can be starved
> > by submit_bh() (data reads).
> >
> > If I'm not missing something, the watermark check should be moved to
> > submit_bh().
>
> We might as well put it there, the idea was to not lock this one
> buffer either but I doubt this would make any different in reality :-)
I'd prefer for this check to be a per-queue one.
Right now a slow device (like a floppy) would artifically throttle a fast
one, if I read the above right. So instead of moving it down the
call-chain, I'd rather remove the check completely as it looks wrong to
me.
Now, if people want throttling, I'd much rather see that done per-queue.
(There's another level of throttling that migth make sense: right now the
swap-out code has this "nr_async_pages" throttling which is very different
from the queue throttling. It might make sense to move that _VM_-level
throttling to writepage too - so that syncing of dirty mmap's will not
cause an overload of pages in flight. This was one of the reasons I
changed the semantics of write-page - so that shared mappings could do
that kind of smoothing too).
Linus
On Thu, Feb 22, 2001 at 10:59:20AM -0800, Linus Torvalds wrote:
> I'd prefer for this check to be a per-queue one.
I'm running this in my tree since a few weeks, however I never had the courage
to post it publically because I didn't benchmarked it carefully yet and I
prefer to finish another thing first. This is actually based on the code I had
in my blkdev tree after I merged last time with Jens the 512K I/O requests and
elevator fixes. I think it won't generate bad numbers and it was running fine
on a 32way SMP (though I didn't stressed the I/O subsystem much there) but
please don't include until somebody benchmarks it carefully with dbench and
tiotest. (it still applys cleanly against 2.4.2)
diff -urN 2.4.2pre3/drivers/block/DAC960.c x/drivers/block/DAC960.c
--- 2.4.2pre3/drivers/block/DAC960.c Fri Feb 9 18:35:16 2001
+++ x/drivers/block/DAC960.c Fri Feb 9 19:00:58 2001
@@ -2827,7 +2827,6 @@
static inline void DAC960_ProcessCompletedBuffer(BufferHeader_T *BufferHeader,
boolean SuccessfulIO)
{
- blk_finished_io(BufferHeader->b_size >> 9);
BufferHeader->b_end_io(BufferHeader, SuccessfulIO);
}
diff -urN 2.4.2pre3/drivers/block/cciss.c x/drivers/block/cciss.c
--- 2.4.2pre3/drivers/block/cciss.c Fri Feb 9 18:35:17 2001
+++ x/drivers/block/cciss.c Fri Feb 9 19:00:59 2001
@@ -1072,7 +1072,6 @@
{
xbh = bh->b_reqnext;
bh->b_reqnext = NULL;
- blk_finished_io(bh->b_size >> 9);
bh->b_end_io(bh, status);
bh = xbh;
}
diff -urN 2.4.2pre3/drivers/block/cpqarray.c x/drivers/block/cpqarray.c
--- 2.4.2pre3/drivers/block/cpqarray.c Fri Feb 9 18:35:17 2001
+++ x/drivers/block/cpqarray.c Fri Feb 9 19:00:59 2001
@@ -1031,7 +1031,6 @@
xbh = bh->b_reqnext;
bh->b_reqnext = NULL;
- blk_finished_io(bh->b_size >> 9);
bh->b_end_io(bh, ok);
bh = xbh;
diff -urN 2.4.2pre3/drivers/block/ll_rw_blk.c x/drivers/block/ll_rw_blk.c
--- 2.4.2pre3/drivers/block/ll_rw_blk.c Fri Feb 9 18:35:17 2001
+++ x/drivers/block/ll_rw_blk.c Fri Feb 9 19:06:11 2001
@@ -118,17 +118,10 @@
int * max_sectors[MAX_BLKDEV];
/*
- * queued sectors for all devices, used to make sure we don't fill all
- * of memory with locked buffers
- */
-atomic_t queued_sectors;
-
-/*
* high and low watermark for above
*/
-static int high_queued_sectors, low_queued_sectors;
+int max_in_flight_sectors;
static int batch_requests, queue_nr_requests;
-static DECLARE_WAIT_QUEUE_HEAD(blk_buffers_wait);
static inline int get_max_sectors(kdev_t dev)
{
@@ -397,11 +390,17 @@
rq = kmem_cache_alloc(request_cachep, SLAB_KERNEL);
memset(rq, 0, sizeof(struct request));
rq->rq_status = RQ_INACTIVE;
- list_add(&rq->table, &q->request_freelist[i & 1]);
+ list_add(&rq->table, &q->request_freelist[i < (queue_nr_requests >> 1) ? READ : WRITE]);
}
- init_waitqueue_head(&q->wait_for_request);
+ init_waitqueue_head(&q->wait_for_request[READ]);
+ init_waitqueue_head(&q->wait_for_request[WRITE]);
+
+ init_waitqueue_head(&q->in_flight_wait);
+ q->in_flight_sectors = 0;
+#if 0
spin_lock_init(&q->queue_lock);
+#endif
}
static int __make_request(request_queue_t * q, int rw, struct buffer_head * bh);
@@ -491,9 +490,9 @@
register struct request *rq;
DECLARE_WAITQUEUE(wait, current);
- add_wait_queue_exclusive(&q->wait_for_request, &wait);
+ add_wait_queue_exclusive(&q->wait_for_request[rw], &wait);
for (;;) {
- __set_current_state(TASK_UNINTERRUPTIBLE);
+ set_current_state(TASK_UNINTERRUPTIBLE);
spin_lock_irq(&io_request_lock);
rq = get_request(q, rw);
spin_unlock_irq(&io_request_lock);
@@ -502,8 +501,8 @@
generic_unplug_device(q);
schedule();
}
- remove_wait_queue(&q->wait_for_request, &wait);
- current->state = TASK_RUNNING;
+ remove_wait_queue(&q->wait_for_request[rw], &wait);
+ __set_current_state(TASK_RUNNING);
return rq;
}
@@ -590,12 +589,26 @@
list_add(&req->queue, insert_here);
}
-void inline blk_refill_freelist(request_queue_t *q, int rw)
+static inline void blk_refill_freelist(request_queue_t *q, int rw)
{
- if (q->pending_free[rw]) {
- list_splice(&q->pending_freelist[rw], &q->request_freelist[rw]);
- INIT_LIST_HEAD(&q->pending_freelist[rw]);
- q->pending_free[rw] = 0;
+ list_splice(&q->pending_freelist[rw], &q->request_freelist[rw]);
+ INIT_LIST_HEAD(&q->pending_freelist[rw]);
+ q->pending_free[rw] = 0;
+}
+
+static inline void blk_finished_io(request_queue_t * q, struct request * req)
+{
+ int nsects = req->in_flight_sectors;
+
+ req->in_flight_sectors = 0;
+ q->in_flight_sectors -= nsects;
+ if (q->in_flight_sectors < 0) {
+ printk("blkdev: in_flight_sectors < 0\n");
+ q->in_flight_sectors = 0;
+ }
+ if (waitqueue_active(&q->in_flight_wait) &&
+ q->in_flight_sectors < max_in_flight_sectors) {
+ wake_up(&(q)->in_flight_wait);
}
}
@@ -612,26 +625,27 @@
/*
* Request may not have originated from ll_rw_blk. if not,
- * asumme it has free buffers and check waiters
+ * assumme it has free buffers and check waiters
*/
if (q) {
- /*
- * we've released enough buffers to start I/O again
- */
- if (waitqueue_active(&blk_buffers_wait)
- && atomic_read(&queued_sectors) < low_queued_sectors)
- wake_up(&blk_buffers_wait);
-
- /*
- * Add to pending free list and batch wakeups
- */
- list_add(&req->table, &q->pending_freelist[rw]);
-
- if (++q->pending_free[rw] >= batch_requests) {
- int wake_up = q->pending_free[rw];
- blk_refill_freelist(q, rw);
- wake_up_nr(&q->wait_for_request, wake_up);
+ if (!list_empty(&q->request_freelist[rw])) {
+ if (q->pending_free[rw])
+ BUG();
+ list_add(&req->table, &q->request_freelist[rw]);
+ wake_up(&q->wait_for_request[rw]);
+ } else {
+ /*
+ * Add to pending free list and batch wakeups
+ */
+ list_add(&req->table, &q->pending_freelist[rw]);
+ if (++q->pending_free[rw] >= batch_requests) {
+ int wake_up = q->pending_free[rw];
+ blk_refill_freelist(q, rw);
+ wake_up_nr(&q->wait_for_request[rw], wake_up);
+ }
}
+
+ blk_finished_io(q, req);
}
}
@@ -694,6 +708,12 @@
attempt_merge(q, blkdev_entry_to_request(prev), max_sectors, max_segments);
}
+#define blk_started_io(q, req, nsects) \
+do { \
+ (q)->in_flight_sectors += (nsects); \
+ (req)->in_flight_sectors += (nsects); \
+} while(0)
+
static int __make_request(request_queue_t * q, int rw,
struct buffer_head * bh)
{
@@ -753,6 +773,27 @@
*/
spin_lock_irq(&io_request_lock);
+ /*
+ * don't lock any more buffers if we are above the max
+ * water mark. instead start I/O on the queued stuff.
+ */
+ if (q->in_flight_sectors >= max_in_flight_sectors) {
+ DECLARE_WAITQUEUE(wait, current);
+
+ add_wait_queue(&q->in_flight_wait, &wait);
+ for (;;) {
+ __set_current_state(TASK_UNINTERRUPTIBLE);
+ __generic_unplug_device(q);
+ spin_unlock_irq(&io_request_lock);
+ schedule();
+ spin_lock_irq(&io_request_lock);
+ if (q->in_flight_sectors < max_in_flight_sectors)
+ break;
+ }
+ remove_wait_queue(&q->in_flight_wait, &wait);
+ __set_current_state(TASK_RUNNING);
+ }
+
insert_here = head->prev;
if (list_empty(head)) {
q->plug_device_fn(q, bh->b_rdev); /* is atomic */
@@ -771,7 +812,7 @@
req->bhtail->b_reqnext = bh;
req->bhtail = bh;
req->nr_sectors = req->hard_nr_sectors += count;
- blk_started_io(count);
+ blk_started_io(q, req, count);
drive_stat_acct(req->rq_dev, req->cmd, count, 0);
attempt_back_merge(q, req, max_sectors, max_segments);
goto out;
@@ -786,7 +827,7 @@
req->current_nr_sectors = count;
req->sector = req->hard_sector = sector;
req->nr_sectors = req->hard_nr_sectors += count;
- blk_started_io(count);
+ blk_started_io(q, req, count);
drive_stat_acct(req->rq_dev, req->cmd, count, 0);
attempt_front_merge(q, head, req, max_sectors, max_segments);
goto out;
@@ -841,7 +882,7 @@
req->bh = bh;
req->bhtail = bh;
req->rq_dev = bh->b_rdev;
- blk_started_io(count);
+ blk_started_io(q, req, count);
add_request(q, req, insert_here);
out:
if (freereq)
@@ -1059,16 +1100,6 @@
for (i = 0; i < nr; i++) {
struct buffer_head *bh = bhs[i];
- /*
- * don't lock any more buffers if we are above the high
- * water mark. instead start I/O on the queued stuff.
- */
- if (atomic_read(&queued_sectors) >= high_queued_sectors) {
- run_task_queue(&tq_disk);
- wait_event(blk_buffers_wait,
- atomic_read(&queued_sectors) < low_queued_sectors);
- }
-
/* Only one thread can actually submit the I/O. */
if (test_and_set_bit(BH_Lock, &bh->b_state))
continue;
@@ -1143,7 +1174,6 @@
if ((bh = req->bh) != NULL) {
nsect = bh->b_size >> 9;
- blk_finished_io(nsect);
req->bh = bh->b_reqnext;
bh->b_reqnext = NULL;
bh->b_end_io(bh, uptodate);
@@ -1173,8 +1203,6 @@
blkdev_release_request(req);
}
-#define MB(kb) ((kb) << 10)
-
int __init blk_dev_init(void)
{
struct blk_dev_struct *dev;
@@ -1194,30 +1222,19 @@
memset(max_readahead, 0, sizeof(max_readahead));
memset(max_sectors, 0, sizeof(max_sectors));
- atomic_set(&queued_sectors, 0);
total_ram = nr_free_pages() << (PAGE_SHIFT - 10);
- /*
- * Try to keep 128MB max hysteris. If not possible,
- * use half of RAM
- */
- high_queued_sectors = (total_ram * 2) / 3;
- low_queued_sectors = high_queued_sectors / 3;
- if (high_queued_sectors - low_queued_sectors > MB(128))
- low_queued_sectors = high_queued_sectors - MB(128);
-
+ max_in_flight_sectors = total_ram / 3;
/*
* make it sectors (512b)
*/
- high_queued_sectors <<= 1;
- low_queued_sectors <<= 1;
+ max_in_flight_sectors <<= 1;
/*
* Scale free request slots per queue too
*/
- total_ram = (total_ram + MB(32) - 1) & ~(MB(32) - 1);
- if ((queue_nr_requests = total_ram >> 9) > QUEUE_NR_REQUESTS)
+ if ((queue_nr_requests = total_ram >> 5) > QUEUE_NR_REQUESTS)
queue_nr_requests = QUEUE_NR_REQUESTS;
/*
@@ -1225,11 +1242,12 @@
*/
if ((batch_requests = queue_nr_requests >> 3) > 32)
batch_requests = 32;
+ if (!batch_requests)
+ batch_requests = 1;
- printk("block: queued sectors max/low %dkB/%dkB, %d slots per queue\n",
- high_queued_sectors / 2,
- low_queued_sectors / 2,
- queue_nr_requests);
+ printk("block: queued sectors max %dkB, %d slots per queue, %d batch\n",
+ max_in_flight_sectors / 2,
+ queue_nr_requests, batch_requests);
#ifdef CONFIG_AMIGA_Z2RAM
z2_init();
@@ -1350,4 +1368,3 @@
EXPORT_SYMBOL(generic_make_request);
EXPORT_SYMBOL(blkdev_release_request);
EXPORT_SYMBOL(generic_unplug_device);
-EXPORT_SYMBOL(queued_sectors);
diff -urN 2.4.2pre3/drivers/scsi/scsi_lib.c x/drivers/scsi/scsi_lib.c
--- 2.4.2pre3/drivers/scsi/scsi_lib.c Fri Feb 9 18:35:34 2001
+++ x/drivers/scsi/scsi_lib.c Fri Feb 9 19:00:59 2001
@@ -375,7 +375,6 @@
do {
if ((bh = req->bh) != NULL) {
nsect = bh->b_size >> 9;
- blk_finished_io(nsect);
req->bh = bh->b_reqnext;
req->nr_sectors -= nsect;
req->sector += nsect;
diff -urN 2.4.2pre3/fs/buffer.c x/fs/buffer.c
--- 2.4.2pre3/fs/buffer.c Fri Feb 9 18:35:44 2001
+++ x/fs/buffer.c Fri Feb 9 19:00:59 2001
@@ -628,7 +628,7 @@
to do in order to release the ramdisk memory is to destroy dirty buffers.
These are two special cases. Normal usage imply the device driver
- to issue a sync on the device (without waiting I/O completation) and
+ to issue a sync on the device (without waiting I/O completion) and
then an invalidate_buffers call that doesn't trash dirty buffers. */
void __invalidate_buffers(kdev_t dev, int destroy_dirty_buffers)
{
@@ -1027,12 +1027,13 @@
write_unlock(&hash_table_lock);
spin_unlock(&lru_list_lock);
refill_freelist(size);
+ /* FIXME: getblk should fail if there's no enough memory */
goto repeat;
}
/* -1 -> no need to flush
0 -> async flush
- 1 -> sync flush (wait for I/O completation) */
+ 1 -> sync flush (wait for I/O completion) */
int balance_dirty_state(kdev_t dev)
{
unsigned long dirty, tot, hard_dirty_limit, soft_dirty_limit;
@@ -1431,6 +1432,7 @@
{
struct buffer_head *bh, *head, *tail;
+ /* FIXME: create_buffers should fail if there's no enough memory */
head = create_buffers(page, blocksize, 1);
if (page->buffers)
BUG();
@@ -2367,11 +2369,9 @@
spin_lock(&free_list[index].lock);
tmp = bh;
do {
- struct buffer_head *p = tmp;
-
- tmp = tmp->b_this_page;
- if (buffer_busy(p))
+ if (buffer_busy(tmp))
goto busy_buffer_page;
+ tmp = tmp->b_this_page;
} while (tmp != bh);
spin_lock(&unused_list_lock);
diff -urN 2.4.2pre3/include/linux/blkdev.h x/include/linux/blkdev.h
--- 2.4.2pre3/include/linux/blkdev.h Fri Feb 9 18:34:13 2001
+++ x/include/linux/blkdev.h Fri Feb 9 21:28:15 2001
@@ -35,7 +35,7 @@
int errors;
unsigned long sector;
unsigned long nr_sectors;
- unsigned long hard_sector, hard_nr_sectors;
+ unsigned long hard_sector, hard_nr_sectors, in_flight_sectors;
unsigned int nr_segments;
unsigned int nr_hw_segments;
unsigned long current_nr_sectors;
@@ -77,6 +77,7 @@
struct list_head request_freelist[2];
struct list_head pending_freelist[2];
int pending_free[2];
+ int in_flight_sectors;
/*
* Together with queue_head for cacheline sharing
@@ -112,16 +113,20 @@
*/
char head_active;
+ wait_queue_head_t in_flight_wait;
+
/*
- * Is meant to protect the queue in the future instead of
- * io_request_lock
+ * Tasks wait here for free request
*/
- spinlock_t queue_lock;
+ wait_queue_head_t wait_for_request[2];
+#if 0
/*
- * Tasks wait here for free request
+ * Is meant to protect the queue in the future instead of
+ * io_request_lock
*/
- wait_queue_head_t wait_for_request;
+ spinlock_t queue_lock;
+#endif
};
struct blk_dev_struct {
@@ -177,7 +182,7 @@
extern int * max_segments[MAX_BLKDEV];
-extern atomic_t queued_sectors;
+extern int max_in_flight_sectors;
#define MAX_SEGMENTS 128
#define MAX_SECTORS (MAX_SEGMENTS*8)
@@ -205,15 +210,5 @@
else
return 512;
}
-
-#define blk_finished_io(nsects) \
- atomic_sub(nsects, &queued_sectors); \
- if (atomic_read(&queued_sectors) < 0) { \
- printk("block: queued_sectors < 0\n"); \
- atomic_set(&queued_sectors, 0); \
- }
-
-#define blk_started_io(nsects) \
- atomic_add(nsects, &queued_sectors);
#endif
Andrea
On Thu, 22 Feb 2001, Linus Torvalds wrote:
>
>
> On Thu, 22 Feb 2001, Jens Axboe wrote:
>
> > On Thu, Feb 22 2001, Marcelo Tosatti wrote:
> > > The following piece of code in ll_rw_block() aims to limit the number of
> > > locked buffers by making processes throttle on IO if the number of on
> > > flight requests is bigger than a high watermaker. IO will only start
> > > again if we're under a low watermark.
> > >
> > > if (atomic_read(&queued_sectors) >= high_queued_sectors) {
> > > run_task_queue(&tq_disk);
> > > wait_event(blk_buffers_wait,
> > > atomic_read(&queued_sectors) < low_queued_sectors);
> > > }
> > >
> > >
> > > However, if submit_bh() is used to queue IO (which is used by ->readpage()
> > > for ext2, for example), no throttling happens.
> > >
> > > It looks like ll_rw_block() users (writes, metadata reads) can be starved
> > > by submit_bh() (data reads).
> > >
> > > If I'm not missing something, the watermark check should be moved to
> > > submit_bh().
> >
> > We might as well put it there, the idea was to not lock this one
> > buffer either but I doubt this would make any different in reality :-)
>
> I'd prefer for this check to be a per-queue one.
>
> Right now a slow device (like a floppy) would artifically throttle a fast
> one, if I read the above right. So instead of moving it down the
> call-chain, I'd rather remove the check completely as it looks wrong to
> me.
>
> Now, if people want throttling, I'd much rather see that done per-queue.
>
> (There's another level of throttling that migth make sense: right now the
> swap-out code has this "nr_async_pages" throttling which is very different
> from the queue throttling. It might make sense to move that _VM_-level
> throttling to writepage too - so that syncing of dirty mmap's will not
> cause an overload of pages in flight. This was one of the reasons I
> changed the semantics of write-page - so that shared mappings could do
> that kind of smoothing too).
And what about write() and read() if you do throttling with nr_async_pages ?
The current scheme inside the block-layer throttles write()'s based on the
number of locked buffers in _main memory_.
On Thu, 22 Feb 2001, Andrea Arcangeli wrote:
> On Thu, Feb 22, 2001 at 10:59:20AM -0800, Linus Torvalds wrote:
> > I'd prefer for this check to be a per-queue one.
>
> I'm running this in my tree since a few weeks, however I never had the courage
> to post it publically because I didn't benchmarked it carefully yet and I
> prefer to finish another thing first.
You want to throttle IO if the amount of on flight data is higher than
a given percentage of _main memory_.
As far as I can see, your patch avoids each individual queue from being
bigger than the high watermark (which is a percentage of main
memory). However, you do not avoid multiple queues together from being
bigger than the high watermark.
On Thu, Feb 22, 2001 at 06:40:48PM -0200, Marcelo Tosatti wrote:
> You want to throttle IO if the amount of on flight data is higher than
> a given percentage of _main memory_.
>
> As far as I can see, your patch avoids each individual queue from being
> bigger than the high watermark (which is a percentage of main
> memory). However, you do not avoid multiple queues together from being
> bigger than the high watermark.
I of course see what you mean and I considered but I tend to believe that's a
minor issue and that most machines will be happier without the global unplug
even if without the global limit.
The only reason we added the limit of I/O in flight is to be allowed to have an
huge number of requests so we can do very large reordering and merges in the
elevator with seeking I/O (4k large IO request) _but_ still we don't have to
wait to lock in ram giga of pages before starting the I/O if the I/O was
contigous. We absolutely need such a sanity limit, it would be absolutely
unsane to wait kupdate to submit 10G of ram to a single harddisk before
unplugging on a 30G machine.
It doesn't need to be exactly "if we unplug not exactly after 1/3 of the global
ram is locked then performance sucks or the machine crashes or task gets
killed". As Jens noticed sync_page_buffers will unplug the queue at some point
if we're low on ram anyways.
The limit just says "unplug after a rasonable limit, after it doesn't matter
anymore to try to delay requests for this harddisk, not matter if there are
still I/O requests available".
However if you have houndred of different queues doing I/O at the same time it
may make a difference, but probably with tons of harddisks you'll also have
tons of ram... In theory we could put a global limit on top of the the
per-queue one. Or we could at least upper bound the total_ram / 3.
Note that 2.4.0 as well doesn't enforce a global limit of packets in flight.
(while in 2.2 the limit is global as it has a shared pool of I/O requests).
Andrea
On Thu, Feb 22, 2001 at 11:57:00PM +0100, Andrea Arcangeli wrote:
> unsane to wait kupdate to submit 10G of ram to a single harddisk before
> unplugging on a 30G machine.
actually kupdate will unplug itself the queue but in theory it can grow the
queue still up to such level after the I/O started. I think we'd better
add an high limit on the in flight I/O watermark.
Andrea
On Thu, 22 Feb 2001, Andrea Arcangeli wrote:
<snip>
> However if you have houndred of different queues doing I/O at the same
> time it may make a difference, but probably with tons of harddisks
> you'll also have tons of ram... In theory we could put a global limit
> on top of the the per-queue one. Or we could at least upper bound the
> total_ram / 3.
The global limit on top of the per-queue limit sounds good.
Since you're talking about the "total_ram / 3" hardcoded value... it
should be /proc tunable IMO. (Andi Kleen already suggested this)
On Thu, Feb 22, 2001 at 07:44:11PM -0200, Marcelo Tosatti wrote:
> The global limit on top of the per-queue limit sounds good.
Probably.
> Since you're talking about the "total_ram / 3" hardcoded value... it
> should be /proc tunable IMO. (Andi Kleen already suggested this)
Yes, IIRC Andi also proposed that a few weeks ago.
Andrea
On Thu, Feb 22 2001, Andrea Arcangeli wrote:
> On Thu, Feb 22, 2001 at 10:59:20AM -0800, Linus Torvalds wrote:
> > I'd prefer for this check to be a per-queue one.
>
> I'm running this in my tree since a few weeks, however I never had the courage
> to post it publically because I didn't benchmarked it carefully yet and I
> prefer to finish another thing first. This is actually based on the code I had
> in my blkdev tree after I merged last time with Jens the 512K I/O requests and
> elevator fixes. I think it won't generate bad numbers and it was running fine
> on a 32way SMP (though I didn't stressed the I/O subsystem much there) but
> please don't include until somebody benchmarks it carefully with dbench and
> tiotest. (it still applys cleanly against 2.4.2)
Thinking about this a bit, I have to agree with you and Linus. It
is possible to find pathetic cases where the per-queue limit suffers
compared to the global one, but in reality I don't think it's worth
it. And the per-queue limits saves us the atomic updates since it's
done under the io_request_lock (or queue later, still fine) so that's
a win too.
I have had rw wait queues before, was removed when I did the request
stealing which is now gone again. I'm not even sure it's worth it
now, Marcelo and I discussed it last week and I did some tests that
showed nothing remarkable. But it's mainly for free, so we might
as well do it.
Any reason why you don't have a lower wake-up limit for the queue?
Do you mind if I do some testing with this patch and fold it in,
possibly?
--
Jens Axboe
On Sun, Feb 25, 2001 at 06:34:01PM +0100, Jens Axboe wrote:
> Any reason why you don't have a lower wake-up limit for the queue?
The watermark diff looked too high (it's 128M in current Linus's tree), but it's
probably a good idea to resurrect it with a max difference of a few full sized
requests (1/2mbytes).
> Do you mind if I do some testing with this patch and fold it in,
> possibly?
Go ahead, thanks,
Andrea