| Message ID | 20230426082031.1299149-8-yukuai1@huaweicloud.com (mailing list archive) |
|---|---|
| State | New, archived |
| Delegated to: | Song Liu |
| Headers | show |
| Series | limit the number of plugged bio | expand |
Hi Kuai There is a limitation of the memory in your test. But for most situations, customers should not set this. Can this change introduce a performance regression against other situations? Best Regards Xiao On Wed, Apr 26, 2023 at 4:24 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > From: Yu Kuai <yukuai3@huawei.com> > > bio can be added to plug infinitely, and following writeback test can > trigger huge amount of plugged bio: > > Test script: > modprobe brd rd_nr=4 rd_size=10485760 > mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean > echo 0 > /proc/sys/vm/dirty_background_ratio > echo 60 > /proc/sys/vm/dirty_ratio > fio -filename=/dev/md0 -ioengine=libaio -rw=write -bs=4k -numjobs=1 -iodepth=128 -name=test > > Test result: > Monitor /sys/block/md0/inflight will found that inflight keep increasing > until fio finish writing, after running for about 2 minutes: > > [root@fedora ~]# cat /sys/block/md0/inflight > 0 4474191 > > Fix the problem by limiting the number of plugged bio based on the number > of copies for original bio. > > Signed-off-by: Yu Kuai <yukuai3@huawei.com> > --- > drivers/md/raid1-10.c | 9 ++++++++- > drivers/md/raid1.c | 2 +- > drivers/md/raid10.c | 2 +- > 3 files changed, 10 insertions(+), 3 deletions(-) > > diff --git a/drivers/md/raid1-10.c b/drivers/md/raid1-10.c > index 98d678b7df3f..35fb80aa37aa 100644 > --- a/drivers/md/raid1-10.c > +++ b/drivers/md/raid1-10.c > @@ -21,6 +21,7 @@ > #define IO_MADE_GOOD ((struct bio *)2) > > #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) > +#define MAX_PLUG_BIO 32 > > /* for managing resync I/O pages */ > struct resync_pages { > @@ -31,6 +32,7 @@ struct resync_pages { > struct raid1_plug_cb { > struct blk_plug_cb cb; > struct bio_list pending; > + unsigned int count; > }; > > static void rbio_pool_free(void *rbio, void *data) > @@ -127,7 +129,7 @@ static inline void md_submit_write(struct bio *bio) > } > > static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, > - blk_plug_cb_fn unplug) > + blk_plug_cb_fn unplug, int copies) > { > struct raid1_plug_cb *plug = NULL; > struct blk_plug_cb *cb; > @@ -147,6 +149,11 @@ static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, > > plug = container_of(cb, struct raid1_plug_cb, cb); > bio_list_add(&plug->pending, bio); > + if (++plug->count / MAX_PLUG_BIO >= copies) { > + list_del(&cb->list); > + cb->callback(cb, false); > + } > + > > return true; > } > diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c > index 639e09cecf01..c6066408a913 100644 > --- a/drivers/md/raid1.c > +++ b/drivers/md/raid1.c > @@ -1562,7 +1562,7 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio, > r1_bio->sector); > /* flush_pending_writes() needs access to the rdev so...*/ > mbio->bi_bdev = (void *)rdev; > - if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug)) { > + if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) { > spin_lock_irqsave(&conf->device_lock, flags); > bio_list_add(&conf->pending_bio_list, mbio); > spin_unlock_irqrestore(&conf->device_lock, flags); > diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c > index bd9e655ca408..7135cfaf75db 100644 > --- a/drivers/md/raid10.c > +++ b/drivers/md/raid10.c > @@ -1306,7 +1306,7 @@ static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, > > atomic_inc(&r10_bio->remaining); > > - if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug)) { > + if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) { > spin_lock_irqsave(&conf->device_lock, flags); > bio_list_add(&conf->pending_bio_list, mbio); > spin_unlock_irqrestore(&conf->device_lock, flags); > -- > 2.39.2 >
Hi, 在 2023/05/29 10:08, Xiao Ni 写道: > Hi Kuai > > There is a limitation of the memory in your test. But for most > situations, customers should not set this. Can this change introduce a > performance regression against other situations? Noted that this limitation is just to triggered writeback as soon as possible in the test, and it's 100% sure real situations can trigger dirty pages write back asynchronously and continue to produce new dirty pages. If a lot of bio is not plugged, then it's the same as before; if a lot of bio is plugged, noted that before this patchset, these bio will spent quite a long time in plug, and hence I think performance should be better. Thanks, Kuai > > Best Regards > Xiao > > On Wed, Apr 26, 2023 at 4:24 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: >> >> From: Yu Kuai <yukuai3@huawei.com> >> >> bio can be added to plug infinitely, and following writeback test can >> trigger huge amount of plugged bio: >> >> Test script: >> modprobe brd rd_nr=4 rd_size=10485760 >> mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean >> echo 0 > /proc/sys/vm/dirty_background_ratio >> echo 60 > /proc/sys/vm/dirty_ratio >> fio -filename=/dev/md0 -ioengine=libaio -rw=write -bs=4k -numjobs=1 -iodepth=128 -name=test >> >> Test result: >> Monitor /sys/block/md0/inflight will found that inflight keep increasing >> until fio finish writing, after running for about 2 minutes: >> >> [root@fedora ~]# cat /sys/block/md0/inflight >> 0 4474191 >> >> Fix the problem by limiting the number of plugged bio based on the number >> of copies for original bio. >> >> Signed-off-by: Yu Kuai <yukuai3@huawei.com> >> --- >> drivers/md/raid1-10.c | 9 ++++++++- >> drivers/md/raid1.c | 2 +- >> drivers/md/raid10.c | 2 +- >> 3 files changed, 10 insertions(+), 3 deletions(-) >> >> diff --git a/drivers/md/raid1-10.c b/drivers/md/raid1-10.c >> index 98d678b7df3f..35fb80aa37aa 100644 >> --- a/drivers/md/raid1-10.c >> +++ b/drivers/md/raid1-10.c >> @@ -21,6 +21,7 @@ >> #define IO_MADE_GOOD ((struct bio *)2) >> >> #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) >> +#define MAX_PLUG_BIO 32 >> >> /* for managing resync I/O pages */ >> struct resync_pages { >> @@ -31,6 +32,7 @@ struct resync_pages { >> struct raid1_plug_cb { >> struct blk_plug_cb cb; >> struct bio_list pending; >> + unsigned int count; >> }; >> >> static void rbio_pool_free(void *rbio, void *data) >> @@ -127,7 +129,7 @@ static inline void md_submit_write(struct bio *bio) >> } >> >> static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, >> - blk_plug_cb_fn unplug) >> + blk_plug_cb_fn unplug, int copies) >> { >> struct raid1_plug_cb *plug = NULL; >> struct blk_plug_cb *cb; >> @@ -147,6 +149,11 @@ static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, >> >> plug = container_of(cb, struct raid1_plug_cb, cb); >> bio_list_add(&plug->pending, bio); >> + if (++plug->count / MAX_PLUG_BIO >= copies) { >> + list_del(&cb->list); >> + cb->callback(cb, false); >> + } >> + >> >> return true; >> } >> diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c >> index 639e09cecf01..c6066408a913 100644 >> --- a/drivers/md/raid1.c >> +++ b/drivers/md/raid1.c >> @@ -1562,7 +1562,7 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio, >> r1_bio->sector); >> /* flush_pending_writes() needs access to the rdev so...*/ >> mbio->bi_bdev = (void *)rdev; >> - if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug)) { >> + if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) { >> spin_lock_irqsave(&conf->device_lock, flags); >> bio_list_add(&conf->pending_bio_list, mbio); >> spin_unlock_irqrestore(&conf->device_lock, flags); >> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c >> index bd9e655ca408..7135cfaf75db 100644 >> --- a/drivers/md/raid10.c >> +++ b/drivers/md/raid10.c >> @@ -1306,7 +1306,7 @@ static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, >> >> atomic_inc(&r10_bio->remaining); >> >> - if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug)) { >> + if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) { >> spin_lock_irqsave(&conf->device_lock, flags); >> bio_list_add(&conf->pending_bio_list, mbio); >> spin_unlock_irqrestore(&conf->device_lock, flags); >> -- >> 2.39.2 >> > > . >
On Mon, May 29, 2023 at 10:20 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > Hi, > > 在 2023/05/29 10:08, Xiao Ni 写道: > > Hi Kuai > > > > There is a limitation of the memory in your test. But for most > > situations, customers should not set this. Can this change introduce a > > performance regression against other situations? > > Noted that this limitation is just to triggered writeback as soon as > possible in the test, and it's 100% sure real situations can trigger > dirty pages write back asynchronously and continue to produce new dirty > pages. Hi I'm confused here. If we want to trigger write back quickly, it needs to set these two values with a smaller number, rather than 0 and 60. Right? > > If a lot of bio is not plugged, then it's the same as before; if a lot > of bio is plugged, noted that before this patchset, these bio will spent > quite a long time in plug, and hence I think performance should be > better. Hmm, it depends on if it's sequential or not? If it's a big io request, can it miss the merge opportunity? Regards Xiao > > Thanks, > Kuai > > > > Best Regards > > Xiao > > > > On Wed, Apr 26, 2023 at 4:24 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >> > >> From: Yu Kuai <yukuai3@huawei.com> > >> > >> bio can be added to plug infinitely, and following writeback test can > >> trigger huge amount of plugged bio: > >> > >> Test script: > >> modprobe brd rd_nr=4 rd_size=10485760 > >> mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean > >> echo 0 > /proc/sys/vm/dirty_background_ratio > >> echo 60 > /proc/sys/vm/dirty_ratio > >> fio -filename=/dev/md0 -ioengine=libaio -rw=write -bs=4k -numjobs=1 -iodepth=128 -name=test > >> > >> Test result: > >> Monitor /sys/block/md0/inflight will found that inflight keep increasing > >> until fio finish writing, after running for about 2 minutes: > >> > >> [root@fedora ~]# cat /sys/block/md0/inflight > >> 0 4474191 > >> > >> Fix the problem by limiting the number of plugged bio based on the number > >> of copies for original bio. > >> > >> Signed-off-by: Yu Kuai <yukuai3@huawei.com> > >> --- > >> drivers/md/raid1-10.c | 9 ++++++++- > >> drivers/md/raid1.c | 2 +- > >> drivers/md/raid10.c | 2 +- > >> 3 files changed, 10 insertions(+), 3 deletions(-) > >> > >> diff --git a/drivers/md/raid1-10.c b/drivers/md/raid1-10.c > >> index 98d678b7df3f..35fb80aa37aa 100644 > >> --- a/drivers/md/raid1-10.c > >> +++ b/drivers/md/raid1-10.c > >> @@ -21,6 +21,7 @@ > >> #define IO_MADE_GOOD ((struct bio *)2) > >> > >> #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) > >> +#define MAX_PLUG_BIO 32 > >> > >> /* for managing resync I/O pages */ > >> struct resync_pages { > >> @@ -31,6 +32,7 @@ struct resync_pages { > >> struct raid1_plug_cb { > >> struct blk_plug_cb cb; > >> struct bio_list pending; > >> + unsigned int count; > >> }; > >> > >> static void rbio_pool_free(void *rbio, void *data) > >> @@ -127,7 +129,7 @@ static inline void md_submit_write(struct bio *bio) > >> } > >> > >> static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, > >> - blk_plug_cb_fn unplug) > >> + blk_plug_cb_fn unplug, int copies) > >> { > >> struct raid1_plug_cb *plug = NULL; > >> struct blk_plug_cb *cb; > >> @@ -147,6 +149,11 @@ static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, > >> > >> plug = container_of(cb, struct raid1_plug_cb, cb); > >> bio_list_add(&plug->pending, bio); > >> + if (++plug->count / MAX_PLUG_BIO >= copies) { > >> + list_del(&cb->list); > >> + cb->callback(cb, false); > >> + } > >> + > >> > >> return true; > >> } > >> diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c > >> index 639e09cecf01..c6066408a913 100644 > >> --- a/drivers/md/raid1.c > >> +++ b/drivers/md/raid1.c > >> @@ -1562,7 +1562,7 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio, > >> r1_bio->sector); > >> /* flush_pending_writes() needs access to the rdev so...*/ > >> mbio->bi_bdev = (void *)rdev; > >> - if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug)) { > >> + if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) { > >> spin_lock_irqsave(&conf->device_lock, flags); > >> bio_list_add(&conf->pending_bio_list, mbio); > >> spin_unlock_irqrestore(&conf->device_lock, flags); > >> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c > >> index bd9e655ca408..7135cfaf75db 100644 > >> --- a/drivers/md/raid10.c > >> +++ b/drivers/md/raid10.c > >> @@ -1306,7 +1306,7 @@ static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, > >> > >> atomic_inc(&r10_bio->remaining); > >> > >> - if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug)) { > >> + if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) { > >> spin_lock_irqsave(&conf->device_lock, flags); > >> bio_list_add(&conf->pending_bio_list, mbio); > >> spin_unlock_irqrestore(&conf->device_lock, flags); > >> -- > >> 2.39.2 > >> > > > > . > > >
Hi, 在 2023/05/29 11:10, Xiao Ni 写道: > On Mon, May 29, 2023 at 10:20 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >> >> Hi, >> >> 在 2023/05/29 10:08, Xiao Ni 写道: >>> Hi Kuai >>> >>> There is a limitation of the memory in your test. But for most >>> situations, customers should not set this. Can this change introduce a >>> performance regression against other situations? >> >> Noted that this limitation is just to triggered writeback as soon as >> possible in the test, and it's 100% sure real situations can trigger >> dirty pages write back asynchronously and continue to produce new dirty >> pages. > > Hi > > I'm confused here. If we want to trigger write back quickly, it needs > to set these two values with a smaller number, rather than 0 and 60. > Right? 60 is not required, I'll remove this setting. 0 just means write back if there are any dirty pages. >> >> If a lot of bio is not plugged, then it's the same as before; if a lot >> of bio is plugged, noted that before this patchset, these bio will spent >> quite a long time in plug, and hence I think performance should be >> better. > > Hmm, it depends on if it's sequential or not? If it's a big io > request, can it miss the merge opportunity? The bio will still be merged to underlying disks' rq(if it's rq based), underlying disk won't flush plug untill the number of request exceed threshold. Thanks, Kuai > > Regards > Xiao > >> >> Thanks, >> Kuai >>> >>> Best Regards >>> Xiao >>> >>> On Wed, Apr 26, 2023 at 4:24 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>>> >>>> From: Yu Kuai <yukuai3@huawei.com> >>>> >>>> bio can be added to plug infinitely, and following writeback test can >>>> trigger huge amount of plugged bio: >>>> >>>> Test script: >>>> modprobe brd rd_nr=4 rd_size=10485760 >>>> mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean >>>> echo 0 > /proc/sys/vm/dirty_background_ratio >>>> echo 60 > /proc/sys/vm/dirty_ratio >>>> fio -filename=/dev/md0 -ioengine=libaio -rw=write -bs=4k -numjobs=1 -iodepth=128 -name=test >>>> >>>> Test result: >>>> Monitor /sys/block/md0/inflight will found that inflight keep increasing >>>> until fio finish writing, after running for about 2 minutes: >>>> >>>> [root@fedora ~]# cat /sys/block/md0/inflight >>>> 0 4474191 >>>> >>>> Fix the problem by limiting the number of plugged bio based on the number >>>> of copies for original bio. >>>> >>>> Signed-off-by: Yu Kuai <yukuai3@huawei.com> >>>> --- >>>> drivers/md/raid1-10.c | 9 ++++++++- >>>> drivers/md/raid1.c | 2 +- >>>> drivers/md/raid10.c | 2 +- >>>> 3 files changed, 10 insertions(+), 3 deletions(-) >>>> >>>> diff --git a/drivers/md/raid1-10.c b/drivers/md/raid1-10.c >>>> index 98d678b7df3f..35fb80aa37aa 100644 >>>> --- a/drivers/md/raid1-10.c >>>> +++ b/drivers/md/raid1-10.c >>>> @@ -21,6 +21,7 @@ >>>> #define IO_MADE_GOOD ((struct bio *)2) >>>> >>>> #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) >>>> +#define MAX_PLUG_BIO 32 >>>> >>>> /* for managing resync I/O pages */ >>>> struct resync_pages { >>>> @@ -31,6 +32,7 @@ struct resync_pages { >>>> struct raid1_plug_cb { >>>> struct blk_plug_cb cb; >>>> struct bio_list pending; >>>> + unsigned int count; >>>> }; >>>> >>>> static void rbio_pool_free(void *rbio, void *data) >>>> @@ -127,7 +129,7 @@ static inline void md_submit_write(struct bio *bio) >>>> } >>>> >>>> static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, >>>> - blk_plug_cb_fn unplug) >>>> + blk_plug_cb_fn unplug, int copies) >>>> { >>>> struct raid1_plug_cb *plug = NULL; >>>> struct blk_plug_cb *cb; >>>> @@ -147,6 +149,11 @@ static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, >>>> >>>> plug = container_of(cb, struct raid1_plug_cb, cb); >>>> bio_list_add(&plug->pending, bio); >>>> + if (++plug->count / MAX_PLUG_BIO >= copies) { >>>> + list_del(&cb->list); >>>> + cb->callback(cb, false); >>>> + } >>>> + >>>> >>>> return true; >>>> } >>>> diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c >>>> index 639e09cecf01..c6066408a913 100644 >>>> --- a/drivers/md/raid1.c >>>> +++ b/drivers/md/raid1.c >>>> @@ -1562,7 +1562,7 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio, >>>> r1_bio->sector); >>>> /* flush_pending_writes() needs access to the rdev so...*/ >>>> mbio->bi_bdev = (void *)rdev; >>>> - if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug)) { >>>> + if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) { >>>> spin_lock_irqsave(&conf->device_lock, flags); >>>> bio_list_add(&conf->pending_bio_list, mbio); >>>> spin_unlock_irqrestore(&conf->device_lock, flags); >>>> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c >>>> index bd9e655ca408..7135cfaf75db 100644 >>>> --- a/drivers/md/raid10.c >>>> +++ b/drivers/md/raid10.c >>>> @@ -1306,7 +1306,7 @@ static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, >>>> >>>> atomic_inc(&r10_bio->remaining); >>>> >>>> - if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug)) { >>>> + if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) { >>>> spin_lock_irqsave(&conf->device_lock, flags); >>>> bio_list_add(&conf->pending_bio_list, mbio); >>>> spin_unlock_irqrestore(&conf->device_lock, flags); >>>> -- >>>> 2.39.2 >>>> >>> >>> . >>> >> > > . >
On Mon, May 29, 2023 at 11:18 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > Hi, > > 在 2023/05/29 11:10, Xiao Ni 写道: > > On Mon, May 29, 2023 at 10:20 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >> > >> Hi, > >> > >> 在 2023/05/29 10:08, Xiao Ni 写道: > >>> Hi Kuai > >>> > >>> There is a limitation of the memory in your test. But for most > >>> situations, customers should not set this. Can this change introduce a > >>> performance regression against other situations? > >> > >> Noted that this limitation is just to triggered writeback as soon as > >> possible in the test, and it's 100% sure real situations can trigger > >> dirty pages write back asynchronously and continue to produce new dirty > >> pages. > > > > Hi > > > > I'm confused here. If we want to trigger write back quickly, it needs > > to set these two values with a smaller number, rather than 0 and 60. > > Right? > > 60 is not required, I'll remove this setting. > > 0 just means write back if there are any dirty pages. Hi Kuai Does 0 mean disabling write back? I tried to find the doc that describes the meaning when setting dirty_background_ration to 0, but I didn't find it. In https://www.kernel.org/doc/html/next/admin-guide/sysctl/vm.html it doesn't describe this. But it says something like this Note: dirty_background_bytes is the counterpart of dirty_background_ratio. Only one of them may be specified at a time. When one sysctl is written it is immediately taken into account to evaluate the dirty memory limits and the other appears as 0 when read. Maybe you can specify dirty_background_ratio to 1 if you want to trigger write back ASAP. > >> > >> If a lot of bio is not plugged, then it's the same as before; if a lot > >> of bio is plugged, noted that before this patchset, these bio will spent > >> quite a long time in plug, and hence I think performance should be > >> better. > > > > Hmm, it depends on if it's sequential or not? If it's a big io > > request, can it miss the merge opportunity? > > The bio will still be merged to underlying disks' rq(if it's rq based), > underlying disk won't flush plug untill the number of request exceed > threshold. Thanks for this. Regards Xiao > > Thanks, > Kuai > > > > Regards > > Xiao > > > >> > >> Thanks, > >> Kuai > >>> > >>> Best Regards > >>> Xiao > >>> > >>> On Wed, Apr 26, 2023 at 4:24 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >>>> > >>>> From: Yu Kuai <yukuai3@huawei.com> > >>>> > >>>> bio can be added to plug infinitely, and following writeback test can > >>>> trigger huge amount of plugged bio: > >>>> > >>>> Test script: > >>>> modprobe brd rd_nr=4 rd_size=10485760 > >>>> mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean > >>>> echo 0 > /proc/sys/vm/dirty_background_ratio > >>>> echo 60 > /proc/sys/vm/dirty_ratio > >>>> fio -filename=/dev/md0 -ioengine=libaio -rw=write -bs=4k -numjobs=1 -iodepth=128 -name=test > >>>> > >>>> Test result: > >>>> Monitor /sys/block/md0/inflight will found that inflight keep increasing > >>>> until fio finish writing, after running for about 2 minutes: > >>>> > >>>> [root@fedora ~]# cat /sys/block/md0/inflight > >>>> 0 4474191 > >>>> > >>>> Fix the problem by limiting the number of plugged bio based on the number > >>>> of copies for original bio. > >>>> > >>>> Signed-off-by: Yu Kuai <yukuai3@huawei.com> > >>>> --- > >>>> drivers/md/raid1-10.c | 9 ++++++++- > >>>> drivers/md/raid1.c | 2 +- > >>>> drivers/md/raid10.c | 2 +- > >>>> 3 files changed, 10 insertions(+), 3 deletions(-) > >>>> > >>>> diff --git a/drivers/md/raid1-10.c b/drivers/md/raid1-10.c > >>>> index 98d678b7df3f..35fb80aa37aa 100644 > >>>> --- a/drivers/md/raid1-10.c > >>>> +++ b/drivers/md/raid1-10.c > >>>> @@ -21,6 +21,7 @@ > >>>> #define IO_MADE_GOOD ((struct bio *)2) > >>>> > >>>> #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) > >>>> +#define MAX_PLUG_BIO 32 > >>>> > >>>> /* for managing resync I/O pages */ > >>>> struct resync_pages { > >>>> @@ -31,6 +32,7 @@ struct resync_pages { > >>>> struct raid1_plug_cb { > >>>> struct blk_plug_cb cb; > >>>> struct bio_list pending; > >>>> + unsigned int count; > >>>> }; > >>>> > >>>> static void rbio_pool_free(void *rbio, void *data) > >>>> @@ -127,7 +129,7 @@ static inline void md_submit_write(struct bio *bio) > >>>> } > >>>> > >>>> static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, > >>>> - blk_plug_cb_fn unplug) > >>>> + blk_plug_cb_fn unplug, int copies) > >>>> { > >>>> struct raid1_plug_cb *plug = NULL; > >>>> struct blk_plug_cb *cb; > >>>> @@ -147,6 +149,11 @@ static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, > >>>> > >>>> plug = container_of(cb, struct raid1_plug_cb, cb); > >>>> bio_list_add(&plug->pending, bio); > >>>> + if (++plug->count / MAX_PLUG_BIO >= copies) { > >>>> + list_del(&cb->list); > >>>> + cb->callback(cb, false); > >>>> + } > >>>> + > >>>> > >>>> return true; > >>>> } > >>>> diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c > >>>> index 639e09cecf01..c6066408a913 100644 > >>>> --- a/drivers/md/raid1.c > >>>> +++ b/drivers/md/raid1.c > >>>> @@ -1562,7 +1562,7 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio, > >>>> r1_bio->sector); > >>>> /* flush_pending_writes() needs access to the rdev so...*/ > >>>> mbio->bi_bdev = (void *)rdev; > >>>> - if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug)) { > >>>> + if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) { > >>>> spin_lock_irqsave(&conf->device_lock, flags); > >>>> bio_list_add(&conf->pending_bio_list, mbio); > >>>> spin_unlock_irqrestore(&conf->device_lock, flags); > >>>> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c > >>>> index bd9e655ca408..7135cfaf75db 100644 > >>>> --- a/drivers/md/raid10.c > >>>> +++ b/drivers/md/raid10.c > >>>> @@ -1306,7 +1306,7 @@ static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, > >>>> > >>>> atomic_inc(&r10_bio->remaining); > >>>> > >>>> - if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug)) { > >>>> + if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) { > >>>> spin_lock_irqsave(&conf->device_lock, flags); > >>>> bio_list_add(&conf->pending_bio_list, mbio); > >>>> spin_unlock_irqrestore(&conf->device_lock, flags); > >>>> -- > >>>> 2.39.2 > >>>> > >>> > >>> . > >>> > >> > > > > . > > >
Hi, 在 2023/05/29 15:57, Xiao Ni 写道: > On Mon, May 29, 2023 at 11:18 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >> >> Hi, >> >> 在 2023/05/29 11:10, Xiao Ni 写道: >>> On Mon, May 29, 2023 at 10:20 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>>> >>>> Hi, >>>> >>>> 在 2023/05/29 10:08, Xiao Ni 写道: >>>>> Hi Kuai >>>>> >>>>> There is a limitation of the memory in your test. But for most >>>>> situations, customers should not set this. Can this change introduce a >>>>> performance regression against other situations? >>>> >>>> Noted that this limitation is just to triggered writeback as soon as >>>> possible in the test, and it's 100% sure real situations can trigger >>>> dirty pages write back asynchronously and continue to produce new dirty >>>> pages. >>> >>> Hi >>> >>> I'm confused here. If we want to trigger write back quickly, it needs >>> to set these two values with a smaller number, rather than 0 and 60. >>> Right? >> >> 60 is not required, I'll remove this setting. >> >> 0 just means write back if there are any dirty pages. > > Hi Kuai > > Does 0 mean disabling write back? I tried to find the doc that > describes the meaning when setting dirty_background_ratio to 0, but I > didn't find it. > In https://www.kernel.org/doc/html/next/admin-guide/sysctl/vm.html it > doesn't describe this. But it says something like this > > Note: > dirty_background_bytes is the counterpart of dirty_background_ratio. Only > one of them may be specified at a time. When one sysctl is written it is > immediately taken into account to evaluate the dirty memory limits and the > other appears as 0 when read. > > Maybe you can specify dirty_background_ratio to 1 if you want to > trigger write back ASAP. The purpose here is to trigger write back ASAP, I'm not an expert here, but based on test result, 0 obviously doesn't mean disable write back. Set dirty_background_bytes to a value, dirty_background_ratio will be set to 0 together, which means dirty_background_ratio is disabled. However, change dirty_background_ratio from default value to 0, will end up both dirty_background_ratio and dirty_background_bytes to be 0, and based on following related code, I think 0 just means write back if there are any dirty pages. domain_dirty_limits: bg_bytes = dirty_background_bytes -> 0 bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100 -> 0 if (bg_bytes) bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); else bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE; -> 0 dtc->bg_thresh = bg_thresh; -> 0 balance_dirty_pages nr_reclaimable = global_node_page_state(NR_FILE_DIRTY); if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh && !writeback_in_progress(wb)) wb_start_background_writeback(wb); -> writeback ASAP Thanks, Kuai > >>>> >>>> If a lot of bio is not plugged, then it's the same as before; if a lot >>>> of bio is plugged, noted that before this patchset, these bio will spent >>>> quite a long time in plug, and hence I think performance should be >>>> better. >>> >>> Hmm, it depends on if it's sequential or not? If it's a big io >>> request, can it miss the merge opportunity? >> >> The bio will still be merged to underlying disks' rq(if it's rq based), >> underlying disk won't flush plug untill the number of request exceed >> threshold. > > Thanks for this. > > Regards > Xiao >> >> Thanks, >> Kuai >>> >>> Regards >>> Xiao >>> >>>> >>>> Thanks, >>>> Kuai >>>>> >>>>> Best Regards >>>>> Xiao >>>>> >>>>> On Wed, Apr 26, 2023 at 4:24 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>>>>> >>>>>> From: Yu Kuai <yukuai3@huawei.com> >>>>>> >>>>>> bio can be added to plug infinitely, and following writeback test can >>>>>> trigger huge amount of plugged bio: >>>>>> >>>>>> Test script: >>>>>> modprobe brd rd_nr=4 rd_size=10485760 >>>>>> mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean >>>>>> echo 0 > /proc/sys/vm/dirty_background_ratio >>>>>> echo 60 > /proc/sys/vm/dirty_ratio >>>>>> fio -filename=/dev/md0 -ioengine=libaio -rw=write -bs=4k -numjobs=1 -iodepth=128 -name=test >>>>>> >>>>>> Test result: >>>>>> Monitor /sys/block/md0/inflight will found that inflight keep increasing >>>>>> until fio finish writing, after running for about 2 minutes: >>>>>> >>>>>> [root@fedora ~]# cat /sys/block/md0/inflight >>>>>> 0 4474191 >>>>>> >>>>>> Fix the problem by limiting the number of plugged bio based on the number >>>>>> of copies for original bio. >>>>>> >>>>>> Signed-off-by: Yu Kuai <yukuai3@huawei.com> >>>>>> --- >>>>>> drivers/md/raid1-10.c | 9 ++++++++- >>>>>> drivers/md/raid1.c | 2 +- >>>>>> drivers/md/raid10.c | 2 +- >>>>>> 3 files changed, 10 insertions(+), 3 deletions(-) >>>>>> >>>>>> diff --git a/drivers/md/raid1-10.c b/drivers/md/raid1-10.c >>>>>> index 98d678b7df3f..35fb80aa37aa 100644 >>>>>> --- a/drivers/md/raid1-10.c >>>>>> +++ b/drivers/md/raid1-10.c >>>>>> @@ -21,6 +21,7 @@ >>>>>> #define IO_MADE_GOOD ((struct bio *)2) >>>>>> >>>>>> #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) >>>>>> +#define MAX_PLUG_BIO 32 >>>>>> >>>>>> /* for managing resync I/O pages */ >>>>>> struct resync_pages { >>>>>> @@ -31,6 +32,7 @@ struct resync_pages { >>>>>> struct raid1_plug_cb { >>>>>> struct blk_plug_cb cb; >>>>>> struct bio_list pending; >>>>>> + unsigned int count; >>>>>> }; >>>>>> >>>>>> static void rbio_pool_free(void *rbio, void *data) >>>>>> @@ -127,7 +129,7 @@ static inline void md_submit_write(struct bio *bio) >>>>>> } >>>>>> >>>>>> static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, >>>>>> - blk_plug_cb_fn unplug) >>>>>> + blk_plug_cb_fn unplug, int copies) >>>>>> { >>>>>> struct raid1_plug_cb *plug = NULL; >>>>>> struct blk_plug_cb *cb; >>>>>> @@ -147,6 +149,11 @@ static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, >>>>>> >>>>>> plug = container_of(cb, struct raid1_plug_cb, cb); >>>>>> bio_list_add(&plug->pending, bio); >>>>>> + if (++plug->count / MAX_PLUG_BIO >= copies) { >>>>>> + list_del(&cb->list); >>>>>> + cb->callback(cb, false); >>>>>> + } >>>>>> + >>>>>> >>>>>> return true; >>>>>> } >>>>>> diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c >>>>>> index 639e09cecf01..c6066408a913 100644 >>>>>> --- a/drivers/md/raid1.c >>>>>> +++ b/drivers/md/raid1.c >>>>>> @@ -1562,7 +1562,7 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio, >>>>>> r1_bio->sector); >>>>>> /* flush_pending_writes() needs access to the rdev so...*/ >>>>>> mbio->bi_bdev = (void *)rdev; >>>>>> - if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug)) { >>>>>> + if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) { >>>>>> spin_lock_irqsave(&conf->device_lock, flags); >>>>>> bio_list_add(&conf->pending_bio_list, mbio); >>>>>> spin_unlock_irqrestore(&conf->device_lock, flags); >>>>>> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c >>>>>> index bd9e655ca408..7135cfaf75db 100644 >>>>>> --- a/drivers/md/raid10.c >>>>>> +++ b/drivers/md/raid10.c >>>>>> @@ -1306,7 +1306,7 @@ static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, >>>>>> >>>>>> atomic_inc(&r10_bio->remaining); >>>>>> >>>>>> - if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug)) { >>>>>> + if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) { >>>>>> spin_lock_irqsave(&conf->device_lock, flags); >>>>>> bio_list_add(&conf->pending_bio_list, mbio); >>>>>> spin_unlock_irqrestore(&conf->device_lock, flags); >>>>>> -- >>>>>> 2.39.2 >>>>>> >>>>> >>>>> . >>>>> >>>> >>> >>> . >>> >> > > . >
On Mon, May 29, 2023 at 4:50 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > Hi, > > 在 2023/05/29 15:57, Xiao Ni 写道: > > On Mon, May 29, 2023 at 11:18 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >> > >> Hi, > >> > >> 在 2023/05/29 11:10, Xiao Ni 写道: > >>> On Mon, May 29, 2023 at 10:20 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >>>> > >>>> Hi, > >>>> > >>>> 在 2023/05/29 10:08, Xiao Ni 写道: > >>>>> Hi Kuai > >>>>> > >>>>> There is a limitation of the memory in your test. But for most > >>>>> situations, customers should not set this. Can this change introduce a > >>>>> performance regression against other situations? > >>>> > >>>> Noted that this limitation is just to triggered writeback as soon as > >>>> possible in the test, and it's 100% sure real situations can trigger > >>>> dirty pages write back asynchronously and continue to produce new dirty > >>>> pages. > >>> > >>> Hi > >>> > >>> I'm confused here. If we want to trigger write back quickly, it needs > >>> to set these two values with a smaller number, rather than 0 and 60. > >>> Right? > >> > >> 60 is not required, I'll remove this setting. > >> > >> 0 just means write back if there are any dirty pages. > > > > Hi Kuai > > > > Does 0 mean disabling write back? I tried to find the doc that > > describes the meaning when setting dirty_background_ratio to 0, but I > > didn't find it. > > In https://www.kernel.org/doc/html/next/admin-guide/sysctl/vm.html it > > doesn't describe this. But it says something like this > > > > Note: > > dirty_background_bytes is the counterpart of dirty_background_ratio. Only > > one of them may be specified at a time. When one sysctl is written it is > > immediately taken into account to evaluate the dirty memory limits and the > > other appears as 0 when read. > > > > Maybe you can specify dirty_background_ratio to 1 if you want to > > trigger write back ASAP. > > The purpose here is to trigger write back ASAP, I'm not an expert here, > but based on test result, 0 obviously doesn't mean disable write back. > > Set dirty_background_bytes to a value, dirty_background_ratio will be > set to 0 together, which means dirty_background_ratio is disabled. > However, change dirty_background_ratio from default value to 0, will end > up both dirty_background_ratio and dirty_background_bytes to be 0, and > based on following related code, I think 0 just means write back if > there are any dirty pages. > > domain_dirty_limits: > bg_bytes = dirty_background_bytes -> 0 > bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100 -> 0 > > if (bg_bytes) > bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); > else > bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE; -> 0 > > dtc->bg_thresh = bg_thresh; -> 0 > > balance_dirty_pages > nr_reclaimable = global_node_page_state(NR_FILE_DIRTY); > if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh && > !writeback_in_progress(wb)) > wb_start_background_writeback(wb); -> writeback ASAP > > Thanks, > Kuai Hi Kuai I'm not an expert about this either. Thanks for all your patches, I can study more things too. But I still have some questions. I did a test in my environment something like this: modprobe brd rd_nr=4 rd_size=10485760 mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean echo 0 > /proc/sys/vm/dirty_background_ratio fio -filename=/dev/md0 -ioengine=libaio -rw=write -thread -bs=1k-8k -numjobs=1 -iodepth=128 --runtime=10 -name=xxx It will cause OOM and the system hangs modprobe brd rd_nr=4 rd_size=10485760 mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean echo 1 > /proc/sys/vm/dirty_background_ratio (THIS is the only different place) fio -filename=/dev/md0 -ioengine=libaio -rw=write -thread -bs=1k-8k -numjobs=1 -iodepth=128 --runtime=10 -name=xxx It can finish successfully. The value of dirty_background_ration is 1 here means it flushes ASAP So your method should be the opposite way as you designed. All the memory can't be flushed in time, so it uses all memory very soon and the memory runs out and the system hangs. The reason I'm looking at the test is that do we really need this change. Because in the real world, most customers don't disable write back. Anyway, it depends on Song's decision and thanks for your patches again. I'll review V3 and try to do some performance tests. Best Regards Xiao > > > >>>> > >>>> If a lot of bio is not plugged, then it's the same as before; if a lot > >>>> of bio is plugged, noted that before this patchset, these bio will spent > >>>> quite a long time in plug, and hence I think performance should be > >>>> better. > >>> > >>> Hmm, it depends on if it's sequential or not? If it's a big io > >>> request, can it miss the merge opportunity? > >> > >> The bio will still be merged to underlying disks' rq(if it's rq based), > >> underlying disk won't flush plug untill the number of request exceed > >> threshold. > > > > Thanks for this. > > > > Regards > > Xiao > >> > >> Thanks, > >> Kuai > >>> > >>> Regards > >>> Xiao > >>> > >>>> > >>>> Thanks, > >>>> Kuai > >>>>> > >>>>> Best Regards > >>>>> Xiao > >>>>> > >>>>> On Wed, Apr 26, 2023 at 4:24 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >>>>>> > >>>>>> From: Yu Kuai <yukuai3@huawei.com> > >>>>>> > >>>>>> bio can be added to plug infinitely, and following writeback test can > >>>>>> trigger huge amount of plugged bio: > >>>>>> > >>>>>> Test script: > >>>>>> modprobe brd rd_nr=4 rd_size=10485760 > >>>>>> mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean > >>>>>> echo 0 > /proc/sys/vm/dirty_background_ratio > >>>>>> echo 60 > /proc/sys/vm/dirty_ratio > >>>>>> fio -filename=/dev/md0 -ioengine=libaio -rw=write -bs=4k -numjobs=1 -iodepth=128 -name=test > >>>>>> > >>>>>> Test result: > >>>>>> Monitor /sys/block/md0/inflight will found that inflight keep increasing > >>>>>> until fio finish writing, after running for about 2 minutes: > >>>>>> > >>>>>> [root@fedora ~]# cat /sys/block/md0/inflight > >>>>>> 0 4474191 > >>>>>> > >>>>>> Fix the problem by limiting the number of plugged bio based on the number > >>>>>> of copies for original bio. > >>>>>> > >>>>>> Signed-off-by: Yu Kuai <yukuai3@huawei.com> > >>>>>> --- > >>>>>> drivers/md/raid1-10.c | 9 ++++++++- > >>>>>> drivers/md/raid1.c | 2 +- > >>>>>> drivers/md/raid10.c | 2 +- > >>>>>> 3 files changed, 10 insertions(+), 3 deletions(-) > >>>>>> > >>>>>> diff --git a/drivers/md/raid1-10.c b/drivers/md/raid1-10.c > >>>>>> index 98d678b7df3f..35fb80aa37aa 100644 > >>>>>> --- a/drivers/md/raid1-10.c > >>>>>> +++ b/drivers/md/raid1-10.c > >>>>>> @@ -21,6 +21,7 @@ > >>>>>> #define IO_MADE_GOOD ((struct bio *)2) > >>>>>> > >>>>>> #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) > >>>>>> +#define MAX_PLUG_BIO 32 > >>>>>> > >>>>>> /* for managing resync I/O pages */ > >>>>>> struct resync_pages { > >>>>>> @@ -31,6 +32,7 @@ struct resync_pages { > >>>>>> struct raid1_plug_cb { > >>>>>> struct blk_plug_cb cb; > >>>>>> struct bio_list pending; > >>>>>> + unsigned int count; > >>>>>> }; > >>>>>> > >>>>>> static void rbio_pool_free(void *rbio, void *data) > >>>>>> @@ -127,7 +129,7 @@ static inline void md_submit_write(struct bio *bio) > >>>>>> } > >>>>>> > >>>>>> static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, > >>>>>> - blk_plug_cb_fn unplug) > >>>>>> + blk_plug_cb_fn unplug, int copies) > >>>>>> { > >>>>>> struct raid1_plug_cb *plug = NULL; > >>>>>> struct blk_plug_cb *cb; > >>>>>> @@ -147,6 +149,11 @@ static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, > >>>>>> > >>>>>> plug = container_of(cb, struct raid1_plug_cb, cb); > >>>>>> bio_list_add(&plug->pending, bio); > >>>>>> + if (++plug->count / MAX_PLUG_BIO >= copies) { > >>>>>> + list_del(&cb->list); > >>>>>> + cb->callback(cb, false); > >>>>>> + } > >>>>>> + > >>>>>> > >>>>>> return true; > >>>>>> } > >>>>>> diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c > >>>>>> index 639e09cecf01..c6066408a913 100644 > >>>>>> --- a/drivers/md/raid1.c > >>>>>> +++ b/drivers/md/raid1.c > >>>>>> @@ -1562,7 +1562,7 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio, > >>>>>> r1_bio->sector); > >>>>>> /* flush_pending_writes() needs access to the rdev so...*/ > >>>>>> mbio->bi_bdev = (void *)rdev; > >>>>>> - if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug)) { > >>>>>> + if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) { > >>>>>> spin_lock_irqsave(&conf->device_lock, flags); > >>>>>> bio_list_add(&conf->pending_bio_list, mbio); > >>>>>> spin_unlock_irqrestore(&conf->device_lock, flags); > >>>>>> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c > >>>>>> index bd9e655ca408..7135cfaf75db 100644 > >>>>>> --- a/drivers/md/raid10.c > >>>>>> +++ b/drivers/md/raid10.c > >>>>>> @@ -1306,7 +1306,7 @@ static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, > >>>>>> > >>>>>> atomic_inc(&r10_bio->remaining); > >>>>>> > >>>>>> - if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug)) { > >>>>>> + if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) { > >>>>>> spin_lock_irqsave(&conf->device_lock, flags); > >>>>>> bio_list_add(&conf->pending_bio_list, mbio); > >>>>>> spin_unlock_irqrestore(&conf->device_lock, flags); > >>>>>> -- > >>>>>> 2.39.2 > >>>>>> > >>>>> > >>>>> . > >>>>> > >>>> > >>> > >>> . > >>> > >> > > > > . > > >
Hi, 在 2023/05/30 8:58, Xiao Ni 写道: > On Mon, May 29, 2023 at 4:50 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: >> >> Hi, >> >> 在 2023/05/29 15:57, Xiao Ni 写道: >>> On Mon, May 29, 2023 at 11:18 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>>> >>>> Hi, >>>> >>>> 在 2023/05/29 11:10, Xiao Ni 写道: >>>>> On Mon, May 29, 2023 at 10:20 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>>>>> >>>>>> Hi, >>>>>> >>>>>> 在 2023/05/29 10:08, Xiao Ni 写道: >>>>>>> Hi Kuai >>>>>>> >>>>>>> There is a limitation of the memory in your test. But for most >>>>>>> situations, customers should not set this. Can this change introduce a >>>>>>> performance regression against other situations? >>>>>> >>>>>> Noted that this limitation is just to triggered writeback as soon as >>>>>> possible in the test, and it's 100% sure real situations can trigger >>>>>> dirty pages write back asynchronously and continue to produce new dirty >>>>>> pages. >>>>> >>>>> Hi >>>>> >>>>> I'm confused here. If we want to trigger write back quickly, it needs >>>>> to set these two values with a smaller number, rather than 0 and 60. >>>>> Right? >>>> >>>> 60 is not required, I'll remove this setting. >>>> >>>> 0 just means write back if there are any dirty pages. >>> >>> Hi Kuai >>> >>> Does 0 mean disabling write back? I tried to find the doc that >>> describes the meaning when setting dirty_background_ratio to 0, but I >>> didn't find it. >>> In https://www.kernel.org/doc/html/next/admin-guide/sysctl/vm.html it >>> doesn't describe this. But it says something like this >>> >>> Note: >>> dirty_background_bytes is the counterpart of dirty_background_ratio. Only >>> one of them may be specified at a time. When one sysctl is written it is >>> immediately taken into account to evaluate the dirty memory limits and the >>> other appears as 0 when read. >>> >>> Maybe you can specify dirty_background_ratio to 1 if you want to >>> trigger write back ASAP. >> >> The purpose here is to trigger write back ASAP, I'm not an expert here, >> but based on test result, 0 obviously doesn't mean disable write back. >> >> Set dirty_background_bytes to a value, dirty_background_ratio will be >> set to 0 together, which means dirty_background_ratio is disabled. >> However, change dirty_background_ratio from default value to 0, will end >> up both dirty_background_ratio and dirty_background_bytes to be 0, and >> based on following related code, I think 0 just means write back if >> there are any dirty pages. >> >> domain_dirty_limits: >> bg_bytes = dirty_background_bytes -> 0 >> bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100 -> 0 >> >> if (bg_bytes) >> bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); >> else >> bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE; -> 0 >> >> dtc->bg_thresh = bg_thresh; -> 0 >> >> balance_dirty_pages >> nr_reclaimable = global_node_page_state(NR_FILE_DIRTY); >> if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh && >> !writeback_in_progress(wb)) >> wb_start_background_writeback(wb); -> writeback ASAP >> >> Thanks, >> Kuai > > Hi Kuai > > I'm not an expert about this either. Thanks for all your patches, I > can study more things too. But I still have some questions. > > I did a test in my environment something like this: > modprobe brd rd_nr=4 rd_size=10485760 > mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean > echo 0 > /proc/sys/vm/dirty_background_ratio > fio -filename=/dev/md0 -ioengine=libaio -rw=write -thread -bs=1k-8k > -numjobs=1 -iodepth=128 --runtime=10 -name=xxx > It will cause OOM and the system hangs OOM means you trigger this problem... Plug hold lots of bios and cost lots of memory, it's not that write back is disabled, you can verify this by monitor md inflight, noted that don't use too much memory for ramdisk(rd_nr * rd_size) in the test so that OOM won't be triggered. Have you tried to test with this patchset? > > modprobe brd rd_nr=4 rd_size=10485760 > mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean > echo 1 > /proc/sys/vm/dirty_background_ratio (THIS is the only different place) > fio -filename=/dev/md0 -ioengine=libaio -rw=write -thread -bs=1k-8k > -numjobs=1 -iodepth=128 --runtime=10 -name=xxx > It can finish successfully. The value of dirty_background_ration is 1 > here means it flushes ASAP This really doesn't mean flushes ASAP, our test report this problem in the real test that doesn't modify dirty_background_ratio. I guess somewhere triggers io_scheduler(), probably background thread think dirty pages doesn't match threshold, but I'm not sure for now. Thanks, Kuai > > So your method should be the opposite way as you designed. All the > memory can't be flushed in time, so it uses all memory very soon and > the memory runs out and the system hangs. The reason I'm looking at > the test is that do we really need this change. Because in the real > world, most customers don't disable write back. Anyway, it depends on > Song's decision and thanks for your patches again. I'll review V3 and > try to do some performance tests. > > Best Regards > Xiao
On Tue, May 30, 2023 at 9:20 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > Hi, > > 在 2023/05/30 8:58, Xiao Ni 写道: > > On Mon, May 29, 2023 at 4:50 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >> > >> Hi, > >> > >> 在 2023/05/29 15:57, Xiao Ni 写道: > >>> On Mon, May 29, 2023 at 11:18 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >>>> > >>>> Hi, > >>>> > >>>> 在 2023/05/29 11:10, Xiao Ni 写道: > >>>>> On Mon, May 29, 2023 at 10:20 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >>>>>> > >>>>>> Hi, > >>>>>> > >>>>>> 在 2023/05/29 10:08, Xiao Ni 写道: > >>>>>>> Hi Kuai > >>>>>>> > >>>>>>> There is a limitation of the memory in your test. But for most > >>>>>>> situations, customers should not set this. Can this change introduce a > >>>>>>> performance regression against other situations? > >>>>>> > >>>>>> Noted that this limitation is just to triggered writeback as soon as > >>>>>> possible in the test, and it's 100% sure real situations can trigger > >>>>>> dirty pages write back asynchronously and continue to produce new dirty > >>>>>> pages. > >>>>> > >>>>> Hi > >>>>> > >>>>> I'm confused here. If we want to trigger write back quickly, it needs > >>>>> to set these two values with a smaller number, rather than 0 and 60. > >>>>> Right? > >>>> > >>>> 60 is not required, I'll remove this setting. > >>>> > >>>> 0 just means write back if there are any dirty pages. > >>> > >>> Hi Kuai > >>> > >>> Does 0 mean disabling write back? I tried to find the doc that > >>> describes the meaning when setting dirty_background_ratio to 0, but I > >>> didn't find it. > >>> In https://www.kernel.org/doc/html/next/admin-guide/sysctl/vm.html it > >>> doesn't describe this. But it says something like this > >>> > >>> Note: > >>> dirty_background_bytes is the counterpart of dirty_background_ratio. Only > >>> one of them may be specified at a time. When one sysctl is written it is > >>> immediately taken into account to evaluate the dirty memory limits and the > >>> other appears as 0 when read. > >>> > >>> Maybe you can specify dirty_background_ratio to 1 if you want to > >>> trigger write back ASAP. > >> > >> The purpose here is to trigger write back ASAP, I'm not an expert here, > >> but based on test result, 0 obviously doesn't mean disable write back. > >> > >> Set dirty_background_bytes to a value, dirty_background_ratio will be > >> set to 0 together, which means dirty_background_ratio is disabled. > >> However, change dirty_background_ratio from default value to 0, will end > >> up both dirty_background_ratio and dirty_background_bytes to be 0, and > >> based on following related code, I think 0 just means write back if > >> there are any dirty pages. > >> > >> domain_dirty_limits: > >> bg_bytes = dirty_background_bytes -> 0 > >> bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100 -> 0 > >> > >> if (bg_bytes) > >> bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); > >> else > >> bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE; -> 0 > >> > >> dtc->bg_thresh = bg_thresh; -> 0 > >> > >> balance_dirty_pages > >> nr_reclaimable = global_node_page_state(NR_FILE_DIRTY); > >> if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh && > >> !writeback_in_progress(wb)) > >> wb_start_background_writeback(wb); -> writeback ASAP > >> > >> Thanks, > >> Kuai > > > > Hi Kuai > > > > I'm not an expert about this either. Thanks for all your patches, I > > can study more things too. But I still have some questions. > > > > I did a test in my environment something like this: > > modprobe brd rd_nr=4 rd_size=10485760 > > mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean > > echo 0 > /proc/sys/vm/dirty_background_ratio > > fio -filename=/dev/md0 -ioengine=libaio -rw=write -thread -bs=1k-8k > > -numjobs=1 -iodepth=128 --runtime=10 -name=xxx > > It will cause OOM and the system hangs > > OOM means you trigger this problem... Plug hold lots of bios and cost > lots of memory, it's not that write back is disabled, you can verify > this by monitor md inflight, noted that don't use too much memory for > ramdisk(rd_nr * rd_size) in the test so that OOM won't be triggered. > > Have you tried to test with this patchset? Yes, I know I have reproduced this problem. I'll have the v3 patchest. > > > > > modprobe brd rd_nr=4 rd_size=10485760 > > mdadm -CR /dev/md0 -l10 -n4 /dev/ram[0123] --assume-clean > > echo 1 > /proc/sys/vm/dirty_background_ratio (THIS is the only different place) > > fio -filename=/dev/md0 -ioengine=libaio -rw=write -thread -bs=1k-8k > > -numjobs=1 -iodepth=128 --runtime=10 -name=xxx > > It can finish successfully. The value of dirty_background_ration is 1 > > here means it flushes ASAP > > This really doesn't mean flushes ASAP, our test report this problem in > the real test that doesn't modify dirty_background_ratio. I guess > somewhere triggers io_scheduler(), probably background thread think > dirty pages doesn't match threshold, but I'm not sure for now. Thanks for notifying me of this. Regards Xiao > > Thanks, > Kuai > > > > So your method should be the opposite way as you designed. All the > > memory can't be flushed in time, so it uses all memory very soon and > > the memory runs out and the system hangs. The reason I'm looking at > > the test is that do we really need this change. Because in the real > > world, most customers don't disable write back. Anyway, it depends on > > Song's decision and thanks for your patches again. I'll review V3 and > > try to do some performance tests. > > > > Best Regards > > Xiao >
diff --git a/drivers/md/raid1-10.c b/drivers/md/raid1-10.c index 98d678b7df3f..35fb80aa37aa 100644 --- a/drivers/md/raid1-10.c +++ b/drivers/md/raid1-10.c @@ -21,6 +21,7 @@ #define IO_MADE_GOOD ((struct bio *)2) #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) +#define MAX_PLUG_BIO 32 /* for managing resync I/O pages */ struct resync_pages { @@ -31,6 +32,7 @@ struct resync_pages { struct raid1_plug_cb { struct blk_plug_cb cb; struct bio_list pending; + unsigned int count; }; static void rbio_pool_free(void *rbio, void *data) @@ -127,7 +129,7 @@ static inline void md_submit_write(struct bio *bio) } static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, - blk_plug_cb_fn unplug) + blk_plug_cb_fn unplug, int copies) { struct raid1_plug_cb *plug = NULL; struct blk_plug_cb *cb; @@ -147,6 +149,11 @@ static inline bool md_add_bio_to_plug(struct mddev *mddev, struct bio *bio, plug = container_of(cb, struct raid1_plug_cb, cb); bio_list_add(&plug->pending, bio); + if (++plug->count / MAX_PLUG_BIO >= copies) { + list_del(&cb->list); + cb->callback(cb, false); + } + return true; } diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c index 639e09cecf01..c6066408a913 100644 --- a/drivers/md/raid1.c +++ b/drivers/md/raid1.c @@ -1562,7 +1562,7 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio, r1_bio->sector); /* flush_pending_writes() needs access to the rdev so...*/ mbio->bi_bdev = (void *)rdev; - if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug)) { + if (!md_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) { spin_lock_irqsave(&conf->device_lock, flags); bio_list_add(&conf->pending_bio_list, mbio); spin_unlock_irqrestore(&conf->device_lock, flags); diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c index bd9e655ca408..7135cfaf75db 100644 --- a/drivers/md/raid10.c +++ b/drivers/md/raid10.c @@ -1306,7 +1306,7 @@ static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, atomic_inc(&r10_bio->remaining); - if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug)) { + if (!md_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) { spin_lock_irqsave(&conf->device_lock, flags); bio_list_add(&conf->pending_bio_list, mbio); spin_unlock_irqrestore(&conf->device_lock, flags);