NFS: introduce writeback wait queue [Kernel]

Prev: 2.6.31 regression sis5513 PIO Mode 0 hang
Next: [PATCH] NFS: introduce writeback wait queue

From: Wu Fengguang on 3 Oct 2009 23:10

Hi Jens,

This is a bug fix for 2.6.32. Maybe other not block-queue based
filesystems will have similar issues ..

Thanks,
Fengguang

On Sun, Oct 04, 2009 at 11:01:53AM +0800, Wu Fengguang wrote:
> The generic writeback routines are departing from congestion_wait()
> in preference of get_request_wait(), aka. to wait on the block queues.
>
> Introduce the missing writeback wait queue for NFS, otherwise its
> writeback pages will grow out of control.
>
> The SYNC writes can use the full queue space (2*nfs_congestion_kb), while
> the ASYNC writes can only use half queue space. This way SYNC writes won't
> be blocked by the ASYNC ones at all.
>
> We'll be waiting inside the NFS_INO_FLUSHING lock, hence also be
> blocking possible dirtiers. This should not be a bit problem.
> And we should be able to obsolete the NFS_INO_FLUSHING with more
> general writeback improvements in long term.
>
> CC: Jens Axboe <jens.axboe(a)oracle.com>
> CC: Trond Myklebust <Trond.Myklebust(a)netapp.com>
> Signed-off-by: Wu Fengguang <fengguang.wu(a)intel.com>
> ---
> fs/nfs/client.c | 2
> fs/nfs/write.c | 73 +++++++++++++++++++++++++++++-------
> include/linux/nfs_fs_sb.h | 1
> 3 files changed, 62 insertions(+), 14 deletions(-)
>
> --- linux.orig/fs/nfs/write.c 2009-10-04 08:47:16.000000000 +0800
> +++ linux/fs/nfs/write.c 2009-10-04 10:55:32.000000000 +0800
> @@ -189,24 +189,58 @@ static int wb_priority(struct writeback_
>
> int nfs_congestion_kb;
>
> +/*
> + * SYNC requests will be blocked on NFS_SYNC_*_THRESH
> + * ASYNC requests will be blocked on NFS_CONGESTION_*_THRESH
> + */
> +#define NFS_SYNC_WAIT_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-11))
> +#define NFS_SYNC_WAKEUP_THRESH \
> + (NFS_SYNC_WAIT_THRESH - (NFS_SYNC_WAIT_THRESH >> 2))
> +
> #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
> #define NFS_CONGESTION_OFF_THRESH \
> (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
>
> -static int nfs_set_page_writeback(struct page *page)
> +static void nfs_writeback_wait(struct page *page, struct writeback_control *wbc)
> {
> - int ret = test_set_page_writeback(page);
> + struct inode *inode = page->mapping->host;
> + struct nfs_server *nfss = NFS_SERVER(inode);
> + int is_sync = wbc->sync_mode == WB_SYNC_NONE;
> + DEFINE_WAIT(wait);
>
> - if (!ret) {
> - struct inode *inode = page->mapping->host;
> - struct nfs_server *nfss = NFS_SERVER(inode);
> + if (atomic_long_inc_return(&nfss->writeback) < NFS_CONGESTION_ON_THRESH)
> + return;
>
> - if (atomic_long_inc_return(&nfss->writeback) >
> - NFS_CONGESTION_ON_THRESH) {
> - set_bdi_congested(&nfss->backing_dev_info,
> - BLK_RW_ASYNC);
> - }
> + set_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
> +
> + if (is_sync && atomic_long_read(&nfss->writeback) <
> + NFS_SYNC_WAIT_THRESH)
> + return;
> +
> + for (;;) {
> + prepare_to_wait_exclusive(&nfss->writeback_wait[is_sync], &wait,
> + TASK_UNINTERRUPTIBLE);
> +
> + io_schedule();
> +
> + finish_wait(&nfss->writeback_wait[is_sync], &wait);
> +
> + if (atomic_long_read(&nfss->writeback) <
> + NFS_CONGESTION_OFF_THRESH)
> + break;
> + if (is_sync && atomic_long_read(&nfss->writeback) <
> + NFS_SYNC_WAKEUP_THRESH)
> + break;
> }
> +}
> +
> +static int nfs_set_page_writeback(struct page *page, struct writeback_control *wbc)
> +{
> + int ret = test_set_page_writeback(page);
> +
> + if (!ret)
> + nfs_writeback_wait(page, wbc);
> +
> return ret;
> }
>
> @@ -216,8 +250,18 @@ static void nfs_end_page_writeback(struc
> struct nfs_server *nfss = NFS_SERVER(inode);
>
> end_page_writeback(page);
> - if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
> - clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
> +
> + if (atomic_long_dec_return(&nfss->writeback) < NFS_SYNC_WAKEUP_THRESH) {
> + if (waitqueue_active(&nfss->writeback_wait[1]))
> + wake_up(&nfss->writeback_wait[1]);
> + if (atomic_long_read(&nfss->writeback) <
> + NFS_CONGESTION_OFF_THRESH) {
> + clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
> + if (waitqueue_active(&nfss->writeback_wait[0]))
> + wake_up(&nfss->writeback_wait[0]);
> + }
> + }
> +
> }
>
> static struct nfs_page *nfs_find_and_lock_request(struct page *page)
> @@ -254,6 +298,7 @@ static struct nfs_page *nfs_find_and_loc
> * May return an error if the user signalled nfs_wait_on_request().
> */
> static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
> + struct writeback_control *wbc,
> struct page *page)
> {
> struct nfs_page *req;
> @@ -266,7 +311,7 @@ static int nfs_page_async_flush(struct n
> if (IS_ERR(req))
> goto out;
>
> - ret = nfs_set_page_writeback(page);
> + ret = nfs_set_page_writeback(page, wbc);
> BUG_ON(ret != 0);
> BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
>
> @@ -286,7 +331,7 @@ static int nfs_do_writepage(struct page
> nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
>
> nfs_pageio_cond_complete(pgio, page->index);
> - return nfs_page_async_flush(pgio, page);
> + return nfs_page_async_flush(pgio, wbc, page);
> }
>
> /*
> --- linux.orig/include/linux/nfs_fs_sb.h 2009-10-04 09:31:25.000000000 +0800
> +++ linux/include/linux/nfs_fs_sb.h 2009-10-04 09:58:11.000000000 +0800
> @@ -108,6 +108,7 @@ struct nfs_server {
> struct nfs_iostats * io_stats; /* I/O statistics */
> struct backing_dev_info backing_dev_info;
> atomic_long_t writeback; /* number of writeback pages */
> + wait_queue_head_t writeback_wait[2];
> int flags; /* various flags */
> unsigned int caps; /* server capabilities */
> unsigned int rsize; /* read size */
> --- linux.orig/fs/nfs/client.c 2009-10-04 09:59:46.000000000 +0800
> +++ linux/fs/nfs/client.c 2009-10-04 10:00:55.000000000 +0800
> @@ -991,6 +991,8 @@ static struct nfs_server *nfs_alloc_serv
> INIT_LIST_HEAD(&server->master_link);
>
> atomic_set(&server->active, 0);
> + init_waitqueue_head(&server->writeback_wait[BLK_RW_SYNC]);
> + init_waitqueue_head(&server->writeback_wait[BLK_RW_ASYNC]);
>
> server->io_stats = nfs_alloc_iostats();
> if (!server->io_stats) {
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

From: Wu Fengguang on 5 Oct 2009 03:50

On Mon, Oct 05, 2009 at 03:35:51PM +0800, Wu Fengguang wrote:
> Trond, I see this trace on linux-next. There are no more dirty pages
> when `cp' aborts after filling up the partition:
>
> cp: writing `/mnt/test/zero3': No space left on device
>
> I noticed that since then nr_writeback is decreased very slowly
> (~100 pages per second). Looks like an interesting behavior.

In the mean time, there are constant 7-8MB/s writes in the NFS server.
The network flow is much smaller ~400K/s. How can I debug this issue?

Thanks,
Fengguang

> nr_writeback nr_dirty nr_unstable
> 41230 36284 8764
> 41230 37307 7755
> 40009 42812 3818
> 32619 42812 11198
> 32314 42812 11503
> 31894 42812 11862
> 31832 42812 11871
> 31770 42812 11871
> 31721 42812 11871
> 31653 42812 11871
> 40789 33754 11871
> 40713 33754 11871
> 40638 33754 11871
> 40566 33754 11871
> 43901 30313 11871
> 74164 0 11871
> 74062 0 11871
> 73978 0 11871
> 73904 0 11871
> 73858 0 11871
> 73798 0 11871
> 73688 0 11871
> 73580 0 11871
> 73477 0 11871
>
> Thanks,
> Fengguang
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

From: Wu Fengguang on 5 Oct 2009 03:50

On Mon, Oct 05, 2009 at 03:10:26PM +0800, Wu Fengguang wrote:
> Hi all,
>
> This version makes two standalone functions for easier reuse.
>
> Before patch, nr_writeback is near 1G on my 2GB laptop:
>
> nr_writeback nr_dirty nr_unstable
> 203994 2 154469
> 203994 2 154469

Sorry, I cannot reproduce the above trace on linux-next. Maybe it's
one of my private patches' fault.

Trond, I see this trace on linux-next. There are no more dirty pages
when `cp' aborts after filling up the partition:

cp: writing `/mnt/test/zero3': No space left on device

I noticed that since then nr_writeback is decreased very slowly
(~100 pages per second). Looks like an interesting behavior.

nr_writeback nr_dirty nr_unstable
41230 36284 8764
41230 37307 7755
40009 42812 3818
32619 42812 11198
32314 42812 11503
31894 42812 11862
31832 42812 11871
31770 42812 11871
31721 42812 11871
31653 42812 11871
40789 33754 11871
40713 33754 11871
40638 33754 11871
40566 33754 11871
43901 30313 11871
74164 0 11871
74062 0 11871
73978 0 11871
73904 0 11871
73858 0 11871
73798 0 11871
73688 0 11871
73580 0 11871
73477 0 11871

Thanks,
Fengguang

> After patch, nr_writeback is limited to nfs_congestion_kb=42MB.
>
> nr_writeback nr_dirty nr_unstable
> 11180 34195 11754
> 9865 36821 8234
> 10137 36695 9338
>
> One minor problem I noticed is, NFS writeback is not very smooth.
> This per 0.1s sampled trace shows that it can sometimes stuck for
> up to 0.5s:
>
> nr_writeback nr_dirty nr_unstable
> 11055 37408 9599
> 10311 37315 10529
> 10869 35920 11459
> 10869 35920 11459
> 10869 35920 11459
> 10869 35920 11459
> 10869 35920 11459
> 10838 35891 10042
> 10466 35891 10414
> 10900 34744 11437
> 10249 34744 12088
> 10249 34744 12088
> 10249 34744 12088
> 10249 34744 12088
> 10249 34744 12088
> 10249 34744 12088
> 10133 34743 10663
> 10505 34743 11035
> 10970 34991 11345
> 10691 34991 11593
> 10691 34991 11593
> 10691 34991 11593
> 10691 34991 11593
> 10691 34991 11593
>
> Trond, I guess nr_writeback/nr_unstable are decreased in async RPC
> "complete" events. It is understandable that nr_dirty can sometimes
> stuck on local waits, but the "local determined" nr_dirty and "remote
> determined" nr_writeback/nr_unstable tend to stuck at the same time?
> Did I miss something (that could be obvious to you)?
>
> Thanks,
> Fengguang
> ---
> Subject: NFS: introduce writeback wait queue
>
> The generic writeback routines are departing from congestion_wait()
> in preferance of get_request_wait(), aka. waiting on the block queues.
>
> Introduce the missing writeback wait queue for NFS, otherwise its
> writeback pages will grow out of control.
>
> CC: Jens Axboe <jens.axboe(a)oracle.com>
> CC: Chris Mason <chris.mason(a)oracle.com>
> CC: Trond Myklebust <Trond.Myklebust(a)netapp.com>
> Signed-off-by: Wu Fengguang <fengguang.wu(a)intel.com>
> ---
>
> fs/nfs/client.c | 2
> fs/nfs/write.c | 86 ++++++++++++++++++++++++++++--------
> include/linux/nfs_fs_sb.h | 1
> 3 files changed, 72 insertions(+), 17 deletions(-)
>
> --- linux.orig/fs/nfs/write.c 2009-10-05 13:27:20.000000000 +0800
> +++ linux/fs/nfs/write.c 2009-10-05 14:48:39.000000000 +0800
> @@ -189,24 +189,72 @@ static int wb_priority(struct writeback_
>
> int nfs_congestion_kb;
>
> -#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
> -#define NFS_CONGESTION_OFF_THRESH \
> - (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
> +/*
> + * SYNC requests will be blocked on (2*limit) and wakeup on (2*limit - limit/8)
> + * ASYNC requests will be blocked on (limit) and wakeup on (limit - limit/8)
> + * In this way SYNC writes will never be blocked by ASYNC ones.
> + */
>
> -static int nfs_set_page_writeback(struct page *page)
> +static void nfs_writeback_wait(atomic_long_t *nr, long limit, int is_sync,
> + struct backing_dev_info *bdi,
> + wait_queue_head_t *wqh)
> {
> - int ret = test_set_page_writeback(page);
> + DEFINE_WAIT(wait);
> + int hard_limit = limit * 2;
>
> - if (!ret) {
> - struct inode *inode = page->mapping->host;
> - struct nfs_server *nfss = NFS_SERVER(inode);
> + if (atomic_long_read(nr) <= limit)
> + return;
> +
> + set_bdi_congested(bdi, BLK_RW_ASYNC);
>
> - if (atomic_long_inc_return(&nfss->writeback) >
> - NFS_CONGESTION_ON_THRESH) {
> - set_bdi_congested(&nfss->backing_dev_info,
> - BLK_RW_ASYNC);
> + if (is_sync && atomic_long_read(nr) <= hard_limit)
> + return;
> +
> + for (;;) {
> + prepare_to_wait(&wqh[is_sync], &wait, TASK_UNINTERRUPTIBLE);
> +
> + io_schedule();
> +
> + if (atomic_long_read(nr) <= limit - limit/8)
> + break;
> + if (is_sync && atomic_long_read(nr) <= hard_limit - limit/8)
> + break;
> + }
> + finish_wait(&wqh[is_sync], &wait);
> +}
> +
> +static void nfs_writeback_wakeup(long nr, long limit,
> + struct backing_dev_info *bdi,
> + wait_queue_head_t *wqh)
> +{
> + int hard_limit = limit * 2;
> +
> + if (nr < hard_limit - limit/8) {
> + if (waitqueue_active(&wqh[BLK_RW_SYNC]))
> + wake_up(&wqh[BLK_RW_SYNC]);
> + if (nr < limit - limit/8) {
> + clear_bdi_congested(bdi, BLK_RW_ASYNC);
> + if (waitqueue_active(&wqh[BLK_RW_ASYNC]))
> + wake_up(&wqh[BLK_RW_ASYNC]);
> }
> }
> +}
> +
> +static int nfs_set_page_writeback(struct page *page,
> + struct writeback_control *wbc)
> +{
> + struct inode *inode = page->mapping->host;
> + struct nfs_server *nfss = NFS_SERVER(inode);
> + int ret = test_set_page_writeback(page);
> +
> + if (!ret) {
> + atomic_long_inc(&nfss->writeback);
> + nfs_writeback_wait(&nfss->writeback,
> + nfs_congestion_kb >> (PAGE_SHIFT-10),
> + wbc->sync_mode == WB_SYNC_ALL,
> + &nfss->backing_dev_info,
> + nfss->writeback_wait);
> + }
> return ret;
> }
>
> @@ -216,8 +264,11 @@ static void nfs_end_page_writeback(struc
> struct nfs_server *nfss = NFS_SERVER(inode);
>
> end_page_writeback(page);
> - if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
> - clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
> +
> + nfs_writeback_wakeup(atomic_long_dec_return(&nfss->writeback),
> + nfs_congestion_kb >> (PAGE_SHIFT-10),
> + &nfss->backing_dev_info,
> + nfss->writeback_wait);
> }
>
> static struct nfs_page *nfs_find_and_lock_request(struct page *page)
> @@ -254,7 +305,8 @@ static struct nfs_page *nfs_find_and_loc
> * May return an error if the user signalled nfs_wait_on_request().
> */
> static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
> - struct page *page)
> + struct page *page,
> + struct writeback_control *wbc)
> {
> struct nfs_page *req;
> int ret = 0;
> @@ -266,7 +318,7 @@ static int nfs_page_async_flush(struct n
> if (IS_ERR(req))
> goto out;
>
> - ret = nfs_set_page_writeback(page);
> + ret = nfs_set_page_writeback(page, wbc);
> BUG_ON(ret != 0);
> BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
>
> @@ -286,7 +338,7 @@ static int nfs_do_writepage(struct page
> nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
>
> nfs_pageio_cond_complete(pgio, page->index);
> - return nfs_page_async_flush(pgio, page);
> + return nfs_page_async_flush(pgio, page, wbc);
> }
>
> /*
> --- linux.orig/include/linux/nfs_fs_sb.h 2009-10-05 13:27:20.000000000 +0800
> +++ linux/include/linux/nfs_fs_sb.h 2009-10-05 13:28:31.000000000 +0800
> @@ -108,6 +108,7 @@ struct nfs_server {
> struct nfs_iostats * io_stats; /* I/O statistics */
> struct backing_dev_info backing_dev_info;
> atomic_long_t writeback; /* number of writeback pages */
> + wait_queue_head_t writeback_wait[2];
> int flags; /* various flags */
> unsigned int caps; /* server capabilities */
> unsigned int rsize; /* read size */
> --- linux.orig/fs/nfs/client.c 2009-10-05 13:27:20.000000000 +0800
> +++ linux/fs/nfs/client.c 2009-10-05 13:28:31.000000000 +0800
> @@ -991,6 +991,8 @@ static struct nfs_server *nfs_alloc_serv
> INIT_LIST_HEAD(&server->master_link);
>
> atomic_set(&server->active, 0);
> + init_waitqueue_head(&server->writeback_wait[BLK_RW_SYNC]);
> + init_waitqueue_head(&server->writeback_wait[BLK_RW_ASYNC]);
>
> server->io_stats = nfs_alloc_iostats();
> if (!server->io_stats) {
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

From: Myklebust, Trond on 5 Oct 2009 07:10

On Oct 5, 2009, at 3:40, "Wu Fengguang" <fengguang.wu(a)intel.com> wrote:

> On Mon, Oct 05, 2009 at 03:35:51PM +0800, Wu Fengguang wrote:
>> Trond, I see this trace on linux-next. There are no more dirty pages
>> when `cp' aborts after filling up the partition:
>>
>> cp: writing `/mnt/test/zero3': No space left on device
>>
>> I noticed that since then nr_writeback is decreased very slowly
>> (~100 pages per second). Looks like an interesting behavior.
>
> In the mean time, there are constant 7-8MB/s writes in the NFS server.
> The network flow is much smaller ~400K/s. How can I debug this issue?

Hi Fengguang

This is deliberate behaviour. When asynchronous writes start recieving
errors, then we switch to synchronous write mode until the error
condition clears.

The reason is for doing so is firstly because some filesystems (XFS)
perform very poorly under ENOSPC, and so it takes forever to write
back pages (we don't want to cancel all writebacks for temporary
conditions like ENOSPC). It also allows us to deliver the errors more
promptly to the application.

Cheers
Trond

>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

From: Jens Axboe on 5 Oct 2009 07:10

On Sun, Oct 04 2009, Wu Fengguang wrote:
> Hi Jens,
>
> This is a bug fix for 2.6.32. Maybe other not block-queue based
> filesystems will have similar issues ..

Not that I'm aware of, the NFS use is fairly special. Given that this is
purely in the realm of nfs/, I'll let Trond decide how to include and
push this (when a final patch is agreed upon).

Thanks for looking into this!

--
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

| Next | Last
Pages: 1 2
Prev: 2.6.31 regression sis5513 PIO Mode 0 hang
Next: [PATCH] NFS: introduce writeback wait queue