Message ID | 20240220153059.11233-2-xni@redhat.com (mailing list archive) |
---|---|
State | Superseded, archived |
Delegated to: | Mike Snitzer |
Headers | show |
Series | Fix regression bugs | expand |
Hi, 在 2024/02/20 23:30, Xiao Ni 写道: > MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch > commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). > Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") > dmraid stopped sync thread directy by calling md_reap_sync_thread. > After this patch dmraid stops sync thread asynchronously as md does. > This is right. Now the dmraid stop process is like this: > > 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. > stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING > is cleared > 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the > root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) > 3. md thread calls md_check_recovery (This is the place to reap sync > thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync > thread) > 4. raid_dtr stops/free struct mddev and release dmraid related resources > > dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear > this bit when stopping the dmraid before stopping sync thread. > > But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is > cleared before stopping sync thread. It's the reason stop_sync_thread only > wakes up task. If the task isn't running, it still needs to wake up sync > thread too. > > This deadlock can be reproduced 100% by these commands: > modprobe brd rd_size=34816 rd_nr=5 > while [ 1 ]; do > vgcreate test_vg /dev/ram* > lvcreate --type raid5 -L 16M -n test_lv test_vg > lvconvert -y --stripes 4 /dev/test_vg/test_lv > vgremove test_vg -ff > sleep 1 > done > > Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") > Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") > Signed-off-by: Xiao Ni <xni@redhat.com> I'm not sure about this change, I think MD_RECOVERY_WAIT is hacky and really breaks how sync_thread is working, it should just go away soon, once we make sure sync_thread can't be registered before pers->start() is done. Thanks, Kuai > --- > drivers/md/dm-raid.c | 2 ++ > drivers/md/md.c | 1 + > 2 files changed, 3 insertions(+) > > diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c > index eb009d6bb03a..325767c1140f 100644 > --- a/drivers/md/dm-raid.c > +++ b/drivers/md/dm-raid.c > @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) > struct raid_set *rs = ti->private; > > if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { > + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) > + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); > /* Writes have to be stopped before suspending to avoid deadlocks. */ > if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) > md_stop_writes(&rs->md); > diff --git a/drivers/md/md.c b/drivers/md/md.c > index 2266358d8074..54790261254d 100644 > --- a/drivers/md/md.c > +++ b/drivers/md/md.c > @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) > * never happen > */ > md_wakeup_thread_directly(mddev->sync_thread); > + md_wakeup_thread(mddev->sync_thread); > if (work_pending(&mddev->sync_work)) > flush_work(&mddev->sync_work); > >
Hi, 在 2024/02/20 23:30, Xiao Ni 写道: > MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch > commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). > Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") > dmraid stopped sync thread directy by calling md_reap_sync_thread. > After this patch dmraid stops sync thread asynchronously as md does. > This is right. Now the dmraid stop process is like this: > > 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. > stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING > is cleared > 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the > root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) > 3. md thread calls md_check_recovery (This is the place to reap sync > thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync > thread) > 4. raid_dtr stops/free struct mddev and release dmraid related resources > > dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear > this bit when stopping the dmraid before stopping sync thread. > > But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is > cleared before stopping sync thread. It's the reason stop_sync_thread only > wakes up task. If the task isn't running, it still needs to wake up sync > thread too. > > This deadlock can be reproduced 100% by these commands: > modprobe brd rd_size=34816 rd_nr=5 > while [ 1 ]; do > vgcreate test_vg /dev/ram* > lvcreate --type raid5 -L 16M -n test_lv test_vg > lvconvert -y --stripes 4 /dev/test_vg/test_lv > vgremove test_vg -ff > sleep 1 > done > > Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") > Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") > Signed-off-by: Xiao Ni <xni@redhat.com> > --- > drivers/md/dm-raid.c | 2 ++ > drivers/md/md.c | 1 + > 2 files changed, 3 insertions(+) > > diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c > index eb009d6bb03a..325767c1140f 100644 > --- a/drivers/md/dm-raid.c > +++ b/drivers/md/dm-raid.c > @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) > struct raid_set *rs = ti->private; > > if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { > + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) > + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); Notice that 'MD_RECOVERY_WAIT' will never be cleared, hence sync_thread will never make progress until table reload for dm-raid. And other than stopping dm-raid, raid_postsuspend() call also be called by ioctl to suspend dm-raid, hence this change is wrong. I think we can never clear 'MD_RECOVERY_FROZEN' in this case so that 'MD_RECOVERY_WAIT' can be removed, or use '!MD_RECOVERY_WAIT' as a condition to register new sync_thread. Thanks, Kuai > /* Writes have to be stopped before suspending to avoid deadlocks. */ > if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) > md_stop_writes(&rs->md); > diff --git a/drivers/md/md.c b/drivers/md/md.c > index 2266358d8074..54790261254d 100644 > --- a/drivers/md/md.c > +++ b/drivers/md/md.c > @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) > * never happen > */ > md_wakeup_thread_directly(mddev->sync_thread); > + md_wakeup_thread(mddev->sync_thread); > if (work_pending(&mddev->sync_work)) > flush_work(&mddev->sync_work); > >
On Fri, Feb 23, 2024 at 11:32 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > Hi, > > 在 2024/02/20 23:30, Xiao Ni 写道: > > MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch > > commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). > > Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") > > dmraid stopped sync thread directy by calling md_reap_sync_thread. > > After this patch dmraid stops sync thread asynchronously as md does. > > This is right. Now the dmraid stop process is like this: > > > > 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. > > stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING > > is cleared > > 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the > > root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) > > 3. md thread calls md_check_recovery (This is the place to reap sync > > thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync > > thread) > > 4. raid_dtr stops/free struct mddev and release dmraid related resources > > > > dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear > > this bit when stopping the dmraid before stopping sync thread. > > > > But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is > > cleared before stopping sync thread. It's the reason stop_sync_thread only > > wakes up task. If the task isn't running, it still needs to wake up sync > > thread too. > > > > This deadlock can be reproduced 100% by these commands: > > modprobe brd rd_size=34816 rd_nr=5 > > while [ 1 ]; do > > vgcreate test_vg /dev/ram* > > lvcreate --type raid5 -L 16M -n test_lv test_vg > > lvconvert -y --stripes 4 /dev/test_vg/test_lv > > vgremove test_vg -ff > > sleep 1 > > done > > > > Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") > > Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") > > Signed-off-by: Xiao Ni <xni@redhat.com> > > I'm not sure about this change, I think MD_RECOVERY_WAIT is hacky and > really breaks how sync_thread is working, it should just go away soon, > once we make sure sync_thread can't be registered before pers->start() > is done. Hi Kuai I just want to get to a stable state without changing any existing logic. After fixing these regressions, we can consider other changes. (e.g. remove MD_RECOVERY_WAIT. allow sync thread start/stop when array is suspend. ) I talked with Heinz yesterday, for dm-raid, it can't allow any io to happen including sync thread when array is suspended. Regards Xiao > > Thanks, > Kuai > > --- > > drivers/md/dm-raid.c | 2 ++ > > drivers/md/md.c | 1 + > > 2 files changed, 3 insertions(+) > > > > diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c > > index eb009d6bb03a..325767c1140f 100644 > > --- a/drivers/md/dm-raid.c > > +++ b/drivers/md/dm-raid.c > > @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) > > struct raid_set *rs = ti->private; > > > > if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { > > + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) > > + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); > > /* Writes have to be stopped before suspending to avoid deadlocks. */ > > if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) > > md_stop_writes(&rs->md); > > diff --git a/drivers/md/md.c b/drivers/md/md.c > > index 2266358d8074..54790261254d 100644 > > --- a/drivers/md/md.c > > +++ b/drivers/md/md.c > > @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) > > * never happen > > */ > > md_wakeup_thread_directly(mddev->sync_thread); > > + md_wakeup_thread(mddev->sync_thread); > > if (work_pending(&mddev->sync_work)) > > flush_work(&mddev->sync_work); > > > > >
On Fri, Feb 23, 2024 at 6:31 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > Hi, > > 在 2024/02/20 23:30, Xiao Ni 写道: > > MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch > > commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). > > Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") > > dmraid stopped sync thread directy by calling md_reap_sync_thread. > > After this patch dmraid stops sync thread asynchronously as md does. > > This is right. Now the dmraid stop process is like this: > > > > 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. > > stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING > > is cleared > > 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the > > root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) > > 3. md thread calls md_check_recovery (This is the place to reap sync > > thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync > > thread) > > 4. raid_dtr stops/free struct mddev and release dmraid related resources > > > > dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear > > this bit when stopping the dmraid before stopping sync thread. > > > > But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is > > cleared before stopping sync thread. It's the reason stop_sync_thread only > > wakes up task. If the task isn't running, it still needs to wake up sync > > thread too. > > > > This deadlock can be reproduced 100% by these commands: > > modprobe brd rd_size=34816 rd_nr=5 > > while [ 1 ]; do > > vgcreate test_vg /dev/ram* > > lvcreate --type raid5 -L 16M -n test_lv test_vg > > lvconvert -y --stripes 4 /dev/test_vg/test_lv > > vgremove test_vg -ff > > sleep 1 > > done > > > > Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") > > Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") > > Signed-off-by: Xiao Ni <xni@redhat.com> > > --- > > drivers/md/dm-raid.c | 2 ++ > > drivers/md/md.c | 1 + > > 2 files changed, 3 insertions(+) > > > > diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c > > index eb009d6bb03a..325767c1140f 100644 > > --- a/drivers/md/dm-raid.c > > +++ b/drivers/md/dm-raid.c > > @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) > > struct raid_set *rs = ti->private; > > > > if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { > > + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) > > + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); > > Notice that 'MD_RECOVERY_WAIT' will never be cleared, hence sync_thread > will never make progress until table reload for dm-raid. Hi Kuai After this patch, it indeed fix the problem mentioned in this patch. So it can be cleared before stopping sync thread. I don't know why you say it never be cleared. > > And other than stopping dm-raid, raid_postsuspend() call also be called > by ioctl to suspend dm-raid, hence this change is wrong. From code review, raid_postsuspend is used to stop sync thread and suspend array. Maybe I don't understand right. If I'm right, it needs to clear MD_RECOVERY_WAIT before stopping sync thread. > > I think we can never clear 'MD_RECOVERY_FROZEN' in this case so that > 'MD_RECOVERY_WAIT' can be removed, or use '!MD_RECOVERY_WAIT' as a > condition to register new sync_thread. I don't understand you here. From debug, there are three reloads during dmraidd reshape. In the second reload, it doesn't want to start sync thread. It's the reason that it sets MD_RECOVERY_WAIT because dmraid is still not ready. In the third reload, it stops the mddev which is created in the second reload and create a new mddev. During this process, MD_RECOVERY_WAIT always works until suspend mddev which is created in the second mddev. It has no relationship with MD_RECOVERY_FROZEN. Regards Xiao > > Thanks, > Kuai > > /* Writes have to be stopped before suspending to avoid deadlocks. */ > > if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) > > md_stop_writes(&rs->md); > > diff --git a/drivers/md/md.c b/drivers/md/md.c > > index 2266358d8074..54790261254d 100644 > > --- a/drivers/md/md.c > > +++ b/drivers/md/md.c > > @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) > > * never happen > > */ > > md_wakeup_thread_directly(mddev->sync_thread); > > + md_wakeup_thread(mddev->sync_thread); > > if (work_pending(&mddev->sync_work)) > > flush_work(&mddev->sync_work); > > > > >
Hi, 在 2024/02/23 21:20, Xiao Ni 写道: > On Fri, Feb 23, 2024 at 11:32 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >> >> Hi, >> >> 在 2024/02/20 23:30, Xiao Ni 写道: >>> MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch >>> commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). >>> Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") >>> dmraid stopped sync thread directy by calling md_reap_sync_thread. >>> After this patch dmraid stops sync thread asynchronously as md does. >>> This is right. Now the dmraid stop process is like this: >>> >>> 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. >>> stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING >>> is cleared >>> 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the >>> root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) >>> 3. md thread calls md_check_recovery (This is the place to reap sync >>> thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync >>> thread) >>> 4. raid_dtr stops/free struct mddev and release dmraid related resources >>> >>> dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear >>> this bit when stopping the dmraid before stopping sync thread. >>> >>> But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is >>> cleared before stopping sync thread. It's the reason stop_sync_thread only >>> wakes up task. If the task isn't running, it still needs to wake up sync >>> thread too. >>> >>> This deadlock can be reproduced 100% by these commands: >>> modprobe brd rd_size=34816 rd_nr=5 >>> while [ 1 ]; do >>> vgcreate test_vg /dev/ram* >>> lvcreate --type raid5 -L 16M -n test_lv test_vg >>> lvconvert -y --stripes 4 /dev/test_vg/test_lv >>> vgremove test_vg -ff >>> sleep 1 >>> done >>> >>> Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") >>> Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") >>> Signed-off-by: Xiao Ni <xni@redhat.com> >> >> I'm not sure about this change, I think MD_RECOVERY_WAIT is hacky and >> really breaks how sync_thread is working, it should just go away soon, >> once we make sure sync_thread can't be registered before pers->start() >> is done. > > Hi Kuai > > I just want to get to a stable state without changing any existing > logic. After fixing these regressions, we can consider other changes. > (e.g. remove MD_RECOVERY_WAIT. allow sync thread start/stop when array > is suspend. ) I talked with Heinz yesterday, for dm-raid, it can't > allow any io to happen including sync thread when array is suspended. So, are you still thinking that my patchset will allow this for dm-raid? I already explain a lot why patch 1 from my set is okay, and why the set doesn't introduce any behaviour change like you said in this patch 0: "Kuai's patch set breaks some rules". The only thing that will change is that for md/raid, sync_thrad can start for suspended array, however, I don't think this will be a problem because sync_thread can be running for suspended array already, and 'MD_RECOVERY_FROZEN' is already used to prevent sync_thread to start. Thanks, Kuai > > Regards > Xiao >> >> Thanks, >> Kuai >>> --- >>> drivers/md/dm-raid.c | 2 ++ >>> drivers/md/md.c | 1 + >>> 2 files changed, 3 insertions(+) >>> >>> diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c >>> index eb009d6bb03a..325767c1140f 100644 >>> --- a/drivers/md/dm-raid.c >>> +++ b/drivers/md/dm-raid.c >>> @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) >>> struct raid_set *rs = ti->private; >>> >>> if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { >>> + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) >>> + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); >>> /* Writes have to be stopped before suspending to avoid deadlocks. */ >>> if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) >>> md_stop_writes(&rs->md); >>> diff --git a/drivers/md/md.c b/drivers/md/md.c >>> index 2266358d8074..54790261254d 100644 >>> --- a/drivers/md/md.c >>> +++ b/drivers/md/md.c >>> @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) >>> * never happen >>> */ >>> md_wakeup_thread_directly(mddev->sync_thread); >>> + md_wakeup_thread(mddev->sync_thread); >>> if (work_pending(&mddev->sync_work)) >>> flush_work(&mddev->sync_work); >>> >>> >> > > . >
On Mon, Feb 26, 2024 at 9:31 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > Hi, > > 在 2024/02/23 21:20, Xiao Ni 写道: > > On Fri, Feb 23, 2024 at 11:32 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >> > >> Hi, > >> > >> 在 2024/02/20 23:30, Xiao Ni 写道: > >>> MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch > >>> commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). > >>> Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") > >>> dmraid stopped sync thread directy by calling md_reap_sync_thread. > >>> After this patch dmraid stops sync thread asynchronously as md does. > >>> This is right. Now the dmraid stop process is like this: > >>> > >>> 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. > >>> stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING > >>> is cleared > >>> 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the > >>> root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) > >>> 3. md thread calls md_check_recovery (This is the place to reap sync > >>> thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync > >>> thread) > >>> 4. raid_dtr stops/free struct mddev and release dmraid related resources > >>> > >>> dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear > >>> this bit when stopping the dmraid before stopping sync thread. > >>> > >>> But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is > >>> cleared before stopping sync thread. It's the reason stop_sync_thread only > >>> wakes up task. If the task isn't running, it still needs to wake up sync > >>> thread too. > >>> > >>> This deadlock can be reproduced 100% by these commands: > >>> modprobe brd rd_size=34816 rd_nr=5 > >>> while [ 1 ]; do > >>> vgcreate test_vg /dev/ram* > >>> lvcreate --type raid5 -L 16M -n test_lv test_vg > >>> lvconvert -y --stripes 4 /dev/test_vg/test_lv > >>> vgremove test_vg -ff > >>> sleep 1 > >>> done > >>> > >>> Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") > >>> Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") > >>> Signed-off-by: Xiao Ni <xni@redhat.com> > >> > >> I'm not sure about this change, I think MD_RECOVERY_WAIT is hacky and > >> really breaks how sync_thread is working, it should just go away soon, > >> once we make sure sync_thread can't be registered before pers->start() > >> is done. > > > > Hi Kuai > > > > I just want to get to a stable state without changing any existing > > logic. After fixing these regressions, we can consider other changes. > > (e.g. remove MD_RECOVERY_WAIT. allow sync thread start/stop when array > > is suspend. ) I talked with Heinz yesterday, for dm-raid, it can't > > allow any io to happen including sync thread when array is suspended. > Hi Kuai > So, are you still thinking that my patchset will allow this for dm-raid? > > I already explain a lot why patch 1 from my set is okay, and why the set > doesn't introduce any behaviour change like you said in this patch 0: I'll read all your patches to understand you well. But as I mentioned many times too, we're fixing regression problems. It's better for us to fix them with the smallest change. As you can see, in my patch set, we can fix these regression problems with small changes (Sorry I didn't notice your patch set has some changes which are the same with mine). So why don't we need such a big change to fix the regression problems? Now with my patch set I can reproduce a problem by lvm2 test suit which happens in 6.6 too. It means with this patch set we can back to a state same with 6.6. > > "Kuai's patch set breaks some rules". > > The only thing that will change is that for md/raid, sync_thrad can > start for suspended array, however, I don't think this will be a problem > because sync_thread can be running for suspended array already, and > 'MD_RECOVERY_FROZEN' is already used to prevent sync_thread to start. We can't allow sync thread happen for dmraid when it's suspended. Because it needs to switch table when suspended. It's a base design. If it can happen now. We should fix this. Best Regards Xiao > > Thanks, > Kuai > > > > > Regards > > Xiao > >> > >> Thanks, > >> Kuai > >>> --- > >>> drivers/md/dm-raid.c | 2 ++ > >>> drivers/md/md.c | 1 + > >>> 2 files changed, 3 insertions(+) > >>> > >>> diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c > >>> index eb009d6bb03a..325767c1140f 100644 > >>> --- a/drivers/md/dm-raid.c > >>> +++ b/drivers/md/dm-raid.c > >>> @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) > >>> struct raid_set *rs = ti->private; > >>> > >>> if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { > >>> + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) > >>> + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); > >>> /* Writes have to be stopped before suspending to avoid deadlocks. */ > >>> if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) > >>> md_stop_writes(&rs->md); > >>> diff --git a/drivers/md/md.c b/drivers/md/md.c > >>> index 2266358d8074..54790261254d 100644 > >>> --- a/drivers/md/md.c > >>> +++ b/drivers/md/md.c > >>> @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) > >>> * never happen > >>> */ > >>> md_wakeup_thread_directly(mddev->sync_thread); > >>> + md_wakeup_thread(mddev->sync_thread); > >>> if (work_pending(&mddev->sync_work)) > >>> flush_work(&mddev->sync_work); > >>> > >>> > >> > > > > . > > >
Hi, 在 2024/02/26 13:12, Xiao Ni 写道: > On Mon, Feb 26, 2024 at 9:31 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >> >> Hi, >> >> 在 2024/02/23 21:20, Xiao Ni 写道: >>> On Fri, Feb 23, 2024 at 11:32 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>>> >>>> Hi, >>>> >>>> 在 2024/02/20 23:30, Xiao Ni 写道: >>>>> MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch >>>>> commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). >>>>> Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") >>>>> dmraid stopped sync thread directy by calling md_reap_sync_thread. >>>>> After this patch dmraid stops sync thread asynchronously as md does. >>>>> This is right. Now the dmraid stop process is like this: >>>>> >>>>> 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. >>>>> stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING >>>>> is cleared >>>>> 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the >>>>> root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) >>>>> 3. md thread calls md_check_recovery (This is the place to reap sync >>>>> thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync >>>>> thread) >>>>> 4. raid_dtr stops/free struct mddev and release dmraid related resources >>>>> >>>>> dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear >>>>> this bit when stopping the dmraid before stopping sync thread. >>>>> >>>>> But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is >>>>> cleared before stopping sync thread. It's the reason stop_sync_thread only >>>>> wakes up task. If the task isn't running, it still needs to wake up sync >>>>> thread too. >>>>> >>>>> This deadlock can be reproduced 100% by these commands: >>>>> modprobe brd rd_size=34816 rd_nr=5 >>>>> while [ 1 ]; do >>>>> vgcreate test_vg /dev/ram* >>>>> lvcreate --type raid5 -L 16M -n test_lv test_vg >>>>> lvconvert -y --stripes 4 /dev/test_vg/test_lv >>>>> vgremove test_vg -ff >>>>> sleep 1 >>>>> done >>>>> >>>>> Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") >>>>> Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") >>>>> Signed-off-by: Xiao Ni <xni@redhat.com> >>>> >>>> I'm not sure about this change, I think MD_RECOVERY_WAIT is hacky and >>>> really breaks how sync_thread is working, it should just go away soon, >>>> once we make sure sync_thread can't be registered before pers->start() >>>> is done. >>> >>> Hi Kuai >>> >>> I just want to get to a stable state without changing any existing >>> logic. After fixing these regressions, we can consider other changes. >>> (e.g. remove MD_RECOVERY_WAIT. allow sync thread start/stop when array >>> is suspend. ) I talked with Heinz yesterday, for dm-raid, it can't >>> allow any io to happen including sync thread when array is suspended. >> > > Hi Kuai > >> So, are you still thinking that my patchset will allow this for dm-raid? >> >> I already explain a lot why patch 1 from my set is okay, and why the set >> doesn't introduce any behaviour change like you said in this patch 0: > > > I'll read all your patches to understand you well. > > But as I mentioned many times too, we're fixing regression problems. > It's better for us to fix them with the smallest change. As you can > see, in my patch set, we can fix these regression problems with small > changes (Sorry I didn't notice your patch set has some changes which > are the same with mine). So why don't we need such a big change to > fix the regression problems? Now with my patch set I can reproduce a > problem by lvm2 test suit which happens in 6.6 too. It means with this > patch set we can back to a state same with 6.6. For complexity, I agree that we can go back to the same state with v6.6, and then fix other problems on the top of that. I don't have preference for this, I'll post my patchset anyhow. But other than the test suite, you still need to make sure nothing is broken from the big picture. For example, in the following 3 cases, can MD_RECOVERY_WAIT be cleared as expected, and new sync_thread will not start after a reload? 1) ioctl suspend ioctl reload ioctl resume 2) ioctl reload ioctl resume 3) ioctl suspend // without a reload. ioctl resume > > > >> >> "Kuai's patch set breaks some rules". >> >> The only thing that will change is that for md/raid, sync_thrad can >> start for suspended array, however, I don't think this will be a problem >> because sync_thread can be running for suspended array already, and >> 'MD_RECOVERY_FROZEN' is already used to prevent sync_thread to start. > > We can't allow sync thread happen for dmraid when it's suspended. > Because it needs to switch table when suspended. It's a base design. > If it can happen now. We should fix this. Yes, and my patchset also fix this, and other problems related to sync_thread by managing sync_thread the same as md/raid. BTW, on the top my patchset, I already made some change locally to expand MD_RECOVERY_FROZEN and remove MD_RECOVERY_WAIT, if you're going to review my patchset, you can take a look at following change later, I already tested the change. Thanks, Kuai diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c index e1b3b1917627..a0c8a5b92aab 100644 --- a/drivers/md/dm-raid.c +++ b/drivers/md/dm-raid.c @@ -213,6 +213,7 @@ struct raid_dev { #define RT_FLAG_RS_IN_SYNC 6 #define RT_FLAG_RS_RESYNCING 7 #define RT_FLAG_RS_GROW 8 +#define RT_FLAG_WAIT_RELOAD 9 /* Array elements of 64 bit needed for rebuild/failed disk bits */ #define DISKS_ARRAY_ELEMS ((MAX_RAID_DEVICES + (sizeof(uint64_t) * 8 - 1)) / sizeof(uint64_t) / 8) @@ -3728,6 +3729,9 @@ static int raid_message(struct dm_target *ti, unsigned int argc, char **argv, if (test_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) return -EBUSY; + if (test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) + return -EINVAL; + if (!strcasecmp(argv[0], "frozen")) { ret = mddev_lock(mddev); if (ret) @@ -3809,21 +3813,25 @@ static void raid_presuspend(struct dm_target *ti) struct raid_set *rs = ti->private; struct mddev *mddev = &rs->md; - mddev_lock_nointr(mddev); - md_frozen_sync_thread(mddev); - mddev_unlock(mddev); + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) { + mddev_lock_nointr(mddev); + md_frozen_sync_thread(mddev); + mddev_unlock(mddev); - if (mddev->pers && mddev->pers->prepare_suspend) - mddev->pers->prepare_suspend(mddev); + if (mddev->pers && mddev->pers->prepare_suspend) + mddev->pers->prepare_suspend(mddev); + } } static void raid_presuspend_undo(struct dm_target *ti) { struct raid_set *rs = ti->private; - mddev_lock_nointr(&rs->md); - md_unfrozen_sync_thread(&rs->md); - mddev_unlock(&rs->md); + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) { + mddev_lock_nointr(&rs->md); + md_unfrozen_sync_thread(&rs->md); + mddev_unlock(&rs->md); + } } static void raid_postsuspend(struct dm_target *ti) @@ -3958,9 +3966,6 @@ static int rs_start_reshape(struct raid_set *rs) struct mddev *mddev = &rs->md; struct md_personality *pers = mddev->pers; - /* Don't allow the sync thread to work until the table gets reloaded. */ - set_bit(MD_RECOVERY_WAIT, &mddev->recovery); - r = rs_setup_reshape(rs); if (r) return r; @@ -4055,6 +4060,7 @@ static int raid_preresume(struct dm_target *ti) /* Initiate a reshape. */ rs_set_rdev_sectors(rs); mddev_lock_nointr(mddev); + set_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags); r = rs_start_reshape(rs); mddev_unlock(mddev); if (r) @@ -4089,7 +4095,8 @@ static void raid_resume(struct dm_target *ti) mddev_lock_nointr(mddev); mddev->ro = 0; mddev->in_sync = 0; - md_unfrozen_sync_thread(mddev); + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) + md_unfrozen_sync_thread(mddev); mddev_unlock(mddev); } } diff --git a/drivers/md/md.c b/drivers/md/md.c index 81e6f49a14fc..595b1fbdce20 100644 --- a/drivers/md/md.c +++ b/drivers/md/md.c @@ -6064,7 +6064,8 @@ int md_run(struct mddev *mddev) pr_warn("True protection against single-disk failure might be compromised.\n"); } - mddev->recovery = 0; + /* dm-raid expect sync_thread to be frozen until resume */ + mddev->recovery &= BIT_ULL_MASK(MD_RECOVERY_FROZEN); /* may be over-ridden by personality */ mddev->resync_max_sectors = mddev->dev_sectors; @@ -6237,10 +6238,8 @@ int md_start(struct mddev *mddev) int ret = 0; if (mddev->pers->start) { - set_bit(MD_RECOVERY_WAIT, &mddev->recovery); ret = mddev->pers->start(mddev); - clear_bit(MD_RECOVERY_WAIT, &mddev->recovery); - md_wakeup_thread(mddev->sync_thread); + WARN_ON_ONCE(test_bit(MD_RECOVERY_RUNNING, &mddev->recovery)); } return ret; } @@ -8827,8 +8825,7 @@ void md_do_sync(struct md_thread *thread) if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) goto skip; - if (test_bit(MD_RECOVERY_WAIT, &mddev->recovery) || - !md_is_rdwr(mddev)) {/* never try to sync a read-only array */ + if (!md_is_rdwr(mddev)) {/* never try to sync a read-only array */ set_bit(MD_RECOVERY_INTR, &mddev->recovery); goto skip; } diff --git a/drivers/md/md.h b/drivers/md/md.h index c17b7e68c533..eb00cdadc9c5 100644 --- a/drivers/md/md.h +++ b/drivers/md/md.h @@ -554,7 +554,6 @@ enum recovery_flags { MD_RECOVERY_RESHAPE, /* A reshape is happening */ MD_RECOVERY_FROZEN, /* User request to abort, and not restart, any action */ MD_RECOVERY_ERROR, /* sync-action interrupted because io-error */ - MD_RECOVERY_WAIT, /* waiting for pers->start() to finish */ MD_RESYNCING_REMOTE, /* remote node is running resync thread */ }; diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c index acce2868e491..530a7f0b120b 100644 --- a/drivers/md/raid5.c +++ b/drivers/md/raid5.c @@ -5919,7 +5919,6 @@ static int add_all_stripe_bios(struct r5conf *conf, static bool reshape_disabled(struct mddev *mddev) { return !md_is_rdwr(mddev) || - test_bit(MD_RECOVERY_WAIT, &mddev->recovery) || test_bit(MD_RECOVERY_FROZEN, &mddev->recovery); } > > Best Regards > Xiao >> >> Thanks, >> Kuai >> >>> >>> Regards >>> Xiao >>>> >>>> Thanks, >>>> Kuai >>>>> --- >>>>> drivers/md/dm-raid.c | 2 ++ >>>>> drivers/md/md.c | 1 + >>>>> 2 files changed, 3 insertions(+) >>>>> >>>>> diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c >>>>> index eb009d6bb03a..325767c1140f 100644 >>>>> --- a/drivers/md/dm-raid.c >>>>> +++ b/drivers/md/dm-raid.c >>>>> @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) >>>>> struct raid_set *rs = ti->private; >>>>> >>>>> if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { >>>>> + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) >>>>> + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); >>>>> /* Writes have to be stopped before suspending to avoid deadlocks. */ >>>>> if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) >>>>> md_stop_writes(&rs->md); >>>>> diff --git a/drivers/md/md.c b/drivers/md/md.c >>>>> index 2266358d8074..54790261254d 100644 >>>>> --- a/drivers/md/md.c >>>>> +++ b/drivers/md/md.c >>>>> @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) >>>>> * never happen >>>>> */ >>>>> md_wakeup_thread_directly(mddev->sync_thread); >>>>> + md_wakeup_thread(mddev->sync_thread); >>>>> if (work_pending(&mddev->sync_work)) >>>>> flush_work(&mddev->sync_work); >>>>> >>>>> >>>> >>> >>> . >>> >> > > . >
On Mon, Feb 26, 2024 at 5:36 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: > > Hi, > > 在 2024/02/26 13:12, Xiao Ni 写道: > > On Mon, Feb 26, 2024 at 9:31 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >> > >> Hi, > >> > >> 在 2024/02/23 21:20, Xiao Ni 写道: > >>> On Fri, Feb 23, 2024 at 11:32 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: > >>>> > >>>> Hi, > >>>> > >>>> 在 2024/02/20 23:30, Xiao Ni 写道: > >>>>> MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch > >>>>> commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). > >>>>> Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") > >>>>> dmraid stopped sync thread directy by calling md_reap_sync_thread. > >>>>> After this patch dmraid stops sync thread asynchronously as md does. > >>>>> This is right. Now the dmraid stop process is like this: > >>>>> > >>>>> 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. > >>>>> stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING > >>>>> is cleared > >>>>> 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the > >>>>> root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) > >>>>> 3. md thread calls md_check_recovery (This is the place to reap sync > >>>>> thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync > >>>>> thread) > >>>>> 4. raid_dtr stops/free struct mddev and release dmraid related resources > >>>>> > >>>>> dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear > >>>>> this bit when stopping the dmraid before stopping sync thread. > >>>>> > >>>>> But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is > >>>>> cleared before stopping sync thread. It's the reason stop_sync_thread only > >>>>> wakes up task. If the task isn't running, it still needs to wake up sync > >>>>> thread too. > >>>>> > >>>>> This deadlock can be reproduced 100% by these commands: > >>>>> modprobe brd rd_size=34816 rd_nr=5 > >>>>> while [ 1 ]; do > >>>>> vgcreate test_vg /dev/ram* > >>>>> lvcreate --type raid5 -L 16M -n test_lv test_vg > >>>>> lvconvert -y --stripes 4 /dev/test_vg/test_lv > >>>>> vgremove test_vg -ff > >>>>> sleep 1 > >>>>> done > >>>>> > >>>>> Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") > >>>>> Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") > >>>>> Signed-off-by: Xiao Ni <xni@redhat.com> > >>>> > >>>> I'm not sure about this change, I think MD_RECOVERY_WAIT is hacky and > >>>> really breaks how sync_thread is working, it should just go away soon, > >>>> once we make sure sync_thread can't be registered before pers->start() > >>>> is done. > >>> > >>> Hi Kuai > >>> > >>> I just want to get to a stable state without changing any existing > >>> logic. After fixing these regressions, we can consider other changes. > >>> (e.g. remove MD_RECOVERY_WAIT. allow sync thread start/stop when array > >>> is suspend. ) I talked with Heinz yesterday, for dm-raid, it can't > >>> allow any io to happen including sync thread when array is suspended. > >> > > > > Hi Kuai > > > >> So, are you still thinking that my patchset will allow this for dm-raid? > >> > >> I already explain a lot why patch 1 from my set is okay, and why the set > >> doesn't introduce any behaviour change like you said in this patch 0: > > > > > > I'll read all your patches to understand you well. > > > > But as I mentioned many times too, we're fixing regression problems. > > It's better for us to fix them with the smallest change. As you can > > see, in my patch set, we can fix these regression problems with small > > changes (Sorry I didn't notice your patch set has some changes which > > are the same with mine). So why don't we need such a big change to > > fix the regression problems? Now with my patch set I can reproduce a > > problem by lvm2 test suit which happens in 6.6 too. It means with this > > patch set we can back to a state same with 6.6. Hi Kuai > > For complexity, I agree that we can go back to the same state with v6.6, > and then fix other problems on the top of that. I don't have preference > for this, I'll post my patchset anyhow. But other than the test suite, > you still need to make sure nothing is broken from the big picture. Thanks very much for this. I'm glad that we can reach an agreement :) What's the preference you mean? The branch with my patch set? I'll send a formal patch set later. > > For example, in the following 3 cases, can MD_RECOVERY_WAIT be cleared > as expected, and new sync_thread will not start after a reload? > > 1) > ioctl suspend > ioctl reload > ioctl resume > > 2) > ioctl reload > ioctl resume > > 3) > ioctl suspend > // without a reload. > ioctl resume Are they all dmsetup message commands? MD_RECOVERY_WAIT is used to delay reshape to start. The sync thread (reshape) will start once MD_RECOVERY_WAIT is cleared. I did tests with lvm commands. For the three cases mentioned above, I need to check. > > > > > > > > >> > >> "Kuai's patch set breaks some rules". > >> > >> The only thing that will change is that for md/raid, sync_thrad can > >> start for suspended array, however, I don't think this will be a problem > >> because sync_thread can be running for suspended array already, and > >> 'MD_RECOVERY_FROZEN' is already used to prevent sync_thread to start. > > > > We can't allow sync thread happen for dmraid when it's suspended. > > Because it needs to switch table when suspended. It's a base design. > > If it can happen now. We should fix this. > > Yes, and my patchset also fix this, and other problems related to > sync_thread by managing sync_thread the same as md/raid. > > BTW, on the top my patchset, I already made some change locally to > expand MD_RECOVERY_FROZEN and remove MD_RECOVERY_WAIT, if you're going > to review my patchset, you can take a look at following change later, > I already tested the change. We can work on this in the future to fix problems one by one. Please try to make one patch set to resolve one problem. It's easy for review. Best Regards Xiao > > Thanks, > Kuai > > diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c > index e1b3b1917627..a0c8a5b92aab 100644 > --- a/drivers/md/dm-raid.c > +++ b/drivers/md/dm-raid.c > @@ -213,6 +213,7 @@ struct raid_dev { > #define RT_FLAG_RS_IN_SYNC 6 > #define RT_FLAG_RS_RESYNCING 7 > #define RT_FLAG_RS_GROW 8 > +#define RT_FLAG_WAIT_RELOAD 9 > > /* Array elements of 64 bit needed for rebuild/failed disk bits */ > #define DISKS_ARRAY_ELEMS ((MAX_RAID_DEVICES + (sizeof(uint64_t) * 8 - > 1)) / sizeof(uint64_t) / 8) > @@ -3728,6 +3729,9 @@ static int raid_message(struct dm_target *ti, > unsigned int argc, char **argv, > if (test_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) > return -EBUSY; > > + if (test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) > + return -EINVAL; > + > if (!strcasecmp(argv[0], "frozen")) { > ret = mddev_lock(mddev); > if (ret) > @@ -3809,21 +3813,25 @@ static void raid_presuspend(struct dm_target *ti) > struct raid_set *rs = ti->private; > struct mddev *mddev = &rs->md; > > - mddev_lock_nointr(mddev); > - md_frozen_sync_thread(mddev); > - mddev_unlock(mddev); > + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) { > + mddev_lock_nointr(mddev); > + md_frozen_sync_thread(mddev); > + mddev_unlock(mddev); > > - if (mddev->pers && mddev->pers->prepare_suspend) > - mddev->pers->prepare_suspend(mddev); > + if (mddev->pers && mddev->pers->prepare_suspend) > + mddev->pers->prepare_suspend(mddev); > + } > } > > static void raid_presuspend_undo(struct dm_target *ti) > { > struct raid_set *rs = ti->private; > > - mddev_lock_nointr(&rs->md); > - md_unfrozen_sync_thread(&rs->md); > - mddev_unlock(&rs->md); > + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) { > + mddev_lock_nointr(&rs->md); > + md_unfrozen_sync_thread(&rs->md); > + mddev_unlock(&rs->md); > + } > } > > static void raid_postsuspend(struct dm_target *ti) > @@ -3958,9 +3966,6 @@ static int rs_start_reshape(struct raid_set *rs) > struct mddev *mddev = &rs->md; > struct md_personality *pers = mddev->pers; > > - /* Don't allow the sync thread to work until the table gets > reloaded. */ > - set_bit(MD_RECOVERY_WAIT, &mddev->recovery); > - > r = rs_setup_reshape(rs); > if (r) > return r; > @@ -4055,6 +4060,7 @@ static int raid_preresume(struct dm_target *ti) > /* Initiate a reshape. */ > rs_set_rdev_sectors(rs); > mddev_lock_nointr(mddev); > + set_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags); > r = rs_start_reshape(rs); > mddev_unlock(mddev); > if (r) > @@ -4089,7 +4095,8 @@ static void raid_resume(struct dm_target *ti) > mddev_lock_nointr(mddev); > mddev->ro = 0; > mddev->in_sync = 0; > - md_unfrozen_sync_thread(mddev); > + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) > + md_unfrozen_sync_thread(mddev); > mddev_unlock(mddev); > } > } > diff --git a/drivers/md/md.c b/drivers/md/md.c > index 81e6f49a14fc..595b1fbdce20 100644 > --- a/drivers/md/md.c > +++ b/drivers/md/md.c > @@ -6064,7 +6064,8 @@ int md_run(struct mddev *mddev) > pr_warn("True protection against single-disk > failure might be compromised.\n"); > } > > - mddev->recovery = 0; > + /* dm-raid expect sync_thread to be frozen until resume */ > + mddev->recovery &= BIT_ULL_MASK(MD_RECOVERY_FROZEN); > /* may be over-ridden by personality */ > mddev->resync_max_sectors = mddev->dev_sectors; > > @@ -6237,10 +6238,8 @@ int md_start(struct mddev *mddev) > int ret = 0; > > if (mddev->pers->start) { > - set_bit(MD_RECOVERY_WAIT, &mddev->recovery); > ret = mddev->pers->start(mddev); > - clear_bit(MD_RECOVERY_WAIT, &mddev->recovery); > - md_wakeup_thread(mddev->sync_thread); > + WARN_ON_ONCE(test_bit(MD_RECOVERY_RUNNING, > &mddev->recovery)); > } > return ret; > } > @@ -8827,8 +8825,7 @@ void md_do_sync(struct md_thread *thread) > if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) > goto skip; > > - if (test_bit(MD_RECOVERY_WAIT, &mddev->recovery) || > - !md_is_rdwr(mddev)) {/* never try to sync a read-only array */ > + if (!md_is_rdwr(mddev)) {/* never try to sync a read-only array */ > set_bit(MD_RECOVERY_INTR, &mddev->recovery); > goto skip; > } > diff --git a/drivers/md/md.h b/drivers/md/md.h > index c17b7e68c533..eb00cdadc9c5 100644 > --- a/drivers/md/md.h > +++ b/drivers/md/md.h > @@ -554,7 +554,6 @@ enum recovery_flags { > MD_RECOVERY_RESHAPE, /* A reshape is happening */ > MD_RECOVERY_FROZEN, /* User request to abort, and not > restart, any action */ > MD_RECOVERY_ERROR, /* sync-action interrupted because > io-error */ > - MD_RECOVERY_WAIT, /* waiting for pers->start() to finish */ > MD_RESYNCING_REMOTE, /* remote node is running resync thread */ > }; > > diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c > index acce2868e491..530a7f0b120b 100644 > --- a/drivers/md/raid5.c > +++ b/drivers/md/raid5.c > @@ -5919,7 +5919,6 @@ static int add_all_stripe_bios(struct r5conf *conf, > static bool reshape_disabled(struct mddev *mddev) > { > return !md_is_rdwr(mddev) || > - test_bit(MD_RECOVERY_WAIT, &mddev->recovery) || > test_bit(MD_RECOVERY_FROZEN, &mddev->recovery); > } > > > > > Best Regards > > Xiao > >> > >> Thanks, > >> Kuai > >> > >>> > >>> Regards > >>> Xiao > >>>> > >>>> Thanks, > >>>> Kuai > >>>>> --- > >>>>> drivers/md/dm-raid.c | 2 ++ > >>>>> drivers/md/md.c | 1 + > >>>>> 2 files changed, 3 insertions(+) > >>>>> > >>>>> diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c > >>>>> index eb009d6bb03a..325767c1140f 100644 > >>>>> --- a/drivers/md/dm-raid.c > >>>>> +++ b/drivers/md/dm-raid.c > >>>>> @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) > >>>>> struct raid_set *rs = ti->private; > >>>>> > >>>>> if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { > >>>>> + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) > >>>>> + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); > >>>>> /* Writes have to be stopped before suspending to avoid deadlocks. */ > >>>>> if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) > >>>>> md_stop_writes(&rs->md); > >>>>> diff --git a/drivers/md/md.c b/drivers/md/md.c > >>>>> index 2266358d8074..54790261254d 100644 > >>>>> --- a/drivers/md/md.c > >>>>> +++ b/drivers/md/md.c > >>>>> @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) > >>>>> * never happen > >>>>> */ > >>>>> md_wakeup_thread_directly(mddev->sync_thread); > >>>>> + md_wakeup_thread(mddev->sync_thread); > >>>>> if (work_pending(&mddev->sync_work)) > >>>>> flush_work(&mddev->sync_work); > >>>>> > >>>>> > >>>> > >>> > >>> . > >>> > >> > > > > . > > >
Hi, 在 2024/02/27 15:16, Xiao Ni 写道: > On Mon, Feb 26, 2024 at 5:36 PM Yu Kuai <yukuai1@huaweicloud.com> wrote: >> >> Hi, >> >> 在 2024/02/26 13:12, Xiao Ni 写道: >>> On Mon, Feb 26, 2024 at 9:31 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>>> >>>> Hi, >>>> >>>> 在 2024/02/23 21:20, Xiao Ni 写道: >>>>> On Fri, Feb 23, 2024 at 11:32 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>>>>> >>>>>> Hi, >>>>>> >>>>>> 在 2024/02/20 23:30, Xiao Ni 写道: >>>>>>> MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch >>>>>>> commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). >>>>>>> Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") >>>>>>> dmraid stopped sync thread directy by calling md_reap_sync_thread. >>>>>>> After this patch dmraid stops sync thread asynchronously as md does. >>>>>>> This is right. Now the dmraid stop process is like this: >>>>>>> >>>>>>> 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. >>>>>>> stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING >>>>>>> is cleared >>>>>>> 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the >>>>>>> root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) >>>>>>> 3. md thread calls md_check_recovery (This is the place to reap sync >>>>>>> thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync >>>>>>> thread) >>>>>>> 4. raid_dtr stops/free struct mddev and release dmraid related resources >>>>>>> >>>>>>> dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear >>>>>>> this bit when stopping the dmraid before stopping sync thread. >>>>>>> >>>>>>> But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is >>>>>>> cleared before stopping sync thread. It's the reason stop_sync_thread only >>>>>>> wakes up task. If the task isn't running, it still needs to wake up sync >>>>>>> thread too. >>>>>>> >>>>>>> This deadlock can be reproduced 100% by these commands: >>>>>>> modprobe brd rd_size=34816 rd_nr=5 >>>>>>> while [ 1 ]; do >>>>>>> vgcreate test_vg /dev/ram* >>>>>>> lvcreate --type raid5 -L 16M -n test_lv test_vg >>>>>>> lvconvert -y --stripes 4 /dev/test_vg/test_lv >>>>>>> vgremove test_vg -ff >>>>>>> sleep 1 >>>>>>> done >>>>>>> >>>>>>> Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") >>>>>>> Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") >>>>>>> Signed-off-by: Xiao Ni <xni@redhat.com> >>>>>> >>>>>> I'm not sure about this change, I think MD_RECOVERY_WAIT is hacky and >>>>>> really breaks how sync_thread is working, it should just go away soon, >>>>>> once we make sure sync_thread can't be registered before pers->start() >>>>>> is done. >>>>> >>>>> Hi Kuai >>>>> >>>>> I just want to get to a stable state without changing any existing >>>>> logic. After fixing these regressions, we can consider other changes. >>>>> (e.g. remove MD_RECOVERY_WAIT. allow sync thread start/stop when array >>>>> is suspend. ) I talked with Heinz yesterday, for dm-raid, it can't >>>>> allow any io to happen including sync thread when array is suspended. >>>> >>> >>> Hi Kuai >>> >>>> So, are you still thinking that my patchset will allow this for dm-raid? >>>> >>>> I already explain a lot why patch 1 from my set is okay, and why the set >>>> doesn't introduce any behaviour change like you said in this patch 0: >>> >>> >>> I'll read all your patches to understand you well. >>> >>> But as I mentioned many times too, we're fixing regression problems. >>> It's better for us to fix them with the smallest change. As you can >>> see, in my patch set, we can fix these regression problems with small >>> changes (Sorry I didn't notice your patch set has some changes which >>> are the same with mine). So why don't we need such a big change to >>> fix the regression problems? Now with my patch set I can reproduce a >>> problem by lvm2 test suit which happens in 6.6 too. It means with this >>> patch set we can back to a state same with 6.6. > > Hi Kuai > >> >> For complexity, I agree that we can go back to the same state with v6.6, >> and then fix other problems on the top of that. I don't have preference >> for this, I'll post my patchset anyhow. But other than the test suite, >> you still need to make sure nothing is broken from the big picture. > > Thanks very much for this. I'm glad that we can reach an agreement :) > What's the preference you mean? The branch with my patch set? I'll > send a formal patch set later. > I don't have preference... That will depend on dm-raid maintainer and Song. >> >> For example, in the following 3 cases, can MD_RECOVERY_WAIT be cleared >> as expected, and new sync_thread will not start after a reload? >> >> 1) >> ioctl suspend >> ioctl reload >> ioctl resume >> >> 2) >> ioctl reload >> ioctl resume >> >> 3) >> ioctl suspend >> // without a reload. >> ioctl resume > > Are they all dmsetup message commands? MD_RECOVERY_WAIT is used to > delay reshape to start. The sync thread (reshape) will start once > MD_RECOVERY_WAIT is cleared. I did tests with lvm commands. For the > three cases mentioned above, I need to check. They are dm ioctl, and you can use dmsetup suspend/reload/resume cmd as well. From what I see(with some debug), once MD_RECOVERY_WAIT is set, dm-raid will never clear it, it relies on a reload to allocate a new raid_set without the flag, and later dm_swap_table() from resume will replace this new raid_set with the old one. That's why I think this patch can work for the above case 1) and 2), however, in case 3), MD_RECOVERY_WAIT is cleared without a reload and start reshape this way can corrupt data. And by the way, for case 1) and 2) the race window between clear MD_RECOVERY_WAIT and mddev_suspend(), reshape can still start concurrently, so this still is not a complete fix. Thanks, Kuai >> >>> >>> >>> >>>> >>>> "Kuai's patch set breaks some rules". >>>> >>>> The only thing that will change is that for md/raid, sync_thrad can >>>> start for suspended array, however, I don't think this will be a problem >>>> because sync_thread can be running for suspended array already, and >>>> 'MD_RECOVERY_FROZEN' is already used to prevent sync_thread to start. >>> >>> We can't allow sync thread happen for dmraid when it's suspended. >>> Because it needs to switch table when suspended. It's a base design. >>> If it can happen now. We should fix this. >> >> Yes, and my patchset also fix this, and other problems related to >> sync_thread by managing sync_thread the same as md/raid. >> >> BTW, on the top my patchset, I already made some change locally to >> expand MD_RECOVERY_FROZEN and remove MD_RECOVERY_WAIT, if you're going >> to review my patchset, you can take a look at following change later, >> I already tested the change. > > We can work on this in the future to fix problems one by one. Please > try to make one patch set to resolve one problem. It's easy for > review. > > Best Regards > Xiao > >> >> Thanks, >> Kuai >> >> diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c >> index e1b3b1917627..a0c8a5b92aab 100644 >> --- a/drivers/md/dm-raid.c >> +++ b/drivers/md/dm-raid.c >> @@ -213,6 +213,7 @@ struct raid_dev { >> #define RT_FLAG_RS_IN_SYNC 6 >> #define RT_FLAG_RS_RESYNCING 7 >> #define RT_FLAG_RS_GROW 8 >> +#define RT_FLAG_WAIT_RELOAD 9 >> >> /* Array elements of 64 bit needed for rebuild/failed disk bits */ >> #define DISKS_ARRAY_ELEMS ((MAX_RAID_DEVICES + (sizeof(uint64_t) * 8 - >> 1)) / sizeof(uint64_t) / 8) >> @@ -3728,6 +3729,9 @@ static int raid_message(struct dm_target *ti, >> unsigned int argc, char **argv, >> if (test_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) >> return -EBUSY; >> >> + if (test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) >> + return -EINVAL; >> + >> if (!strcasecmp(argv[0], "frozen")) { >> ret = mddev_lock(mddev); >> if (ret) >> @@ -3809,21 +3813,25 @@ static void raid_presuspend(struct dm_target *ti) >> struct raid_set *rs = ti->private; >> struct mddev *mddev = &rs->md; >> >> - mddev_lock_nointr(mddev); >> - md_frozen_sync_thread(mddev); >> - mddev_unlock(mddev); >> + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) { >> + mddev_lock_nointr(mddev); >> + md_frozen_sync_thread(mddev); >> + mddev_unlock(mddev); >> >> - if (mddev->pers && mddev->pers->prepare_suspend) >> - mddev->pers->prepare_suspend(mddev); >> + if (mddev->pers && mddev->pers->prepare_suspend) >> + mddev->pers->prepare_suspend(mddev); >> + } >> } >> >> static void raid_presuspend_undo(struct dm_target *ti) >> { >> struct raid_set *rs = ti->private; >> >> - mddev_lock_nointr(&rs->md); >> - md_unfrozen_sync_thread(&rs->md); >> - mddev_unlock(&rs->md); >> + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) { >> + mddev_lock_nointr(&rs->md); >> + md_unfrozen_sync_thread(&rs->md); >> + mddev_unlock(&rs->md); >> + } >> } >> >> static void raid_postsuspend(struct dm_target *ti) >> @@ -3958,9 +3966,6 @@ static int rs_start_reshape(struct raid_set *rs) >> struct mddev *mddev = &rs->md; >> struct md_personality *pers = mddev->pers; >> >> - /* Don't allow the sync thread to work until the table gets >> reloaded. */ >> - set_bit(MD_RECOVERY_WAIT, &mddev->recovery); >> - >> r = rs_setup_reshape(rs); >> if (r) >> return r; >> @@ -4055,6 +4060,7 @@ static int raid_preresume(struct dm_target *ti) >> /* Initiate a reshape. */ >> rs_set_rdev_sectors(rs); >> mddev_lock_nointr(mddev); >> + set_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags); >> r = rs_start_reshape(rs); >> mddev_unlock(mddev); >> if (r) >> @@ -4089,7 +4095,8 @@ static void raid_resume(struct dm_target *ti) >> mddev_lock_nointr(mddev); >> mddev->ro = 0; >> mddev->in_sync = 0; >> - md_unfrozen_sync_thread(mddev); >> + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) >> + md_unfrozen_sync_thread(mddev); >> mddev_unlock(mddev); >> } >> } >> diff --git a/drivers/md/md.c b/drivers/md/md.c >> index 81e6f49a14fc..595b1fbdce20 100644 >> --- a/drivers/md/md.c >> +++ b/drivers/md/md.c >> @@ -6064,7 +6064,8 @@ int md_run(struct mddev *mddev) >> pr_warn("True protection against single-disk >> failure might be compromised.\n"); >> } >> >> - mddev->recovery = 0; >> + /* dm-raid expect sync_thread to be frozen until resume */ >> + mddev->recovery &= BIT_ULL_MASK(MD_RECOVERY_FROZEN); >> /* may be over-ridden by personality */ >> mddev->resync_max_sectors = mddev->dev_sectors; >> >> @@ -6237,10 +6238,8 @@ int md_start(struct mddev *mddev) >> int ret = 0; >> >> if (mddev->pers->start) { >> - set_bit(MD_RECOVERY_WAIT, &mddev->recovery); >> ret = mddev->pers->start(mddev); >> - clear_bit(MD_RECOVERY_WAIT, &mddev->recovery); >> - md_wakeup_thread(mddev->sync_thread); >> + WARN_ON_ONCE(test_bit(MD_RECOVERY_RUNNING, >> &mddev->recovery)); >> } >> return ret; >> } >> @@ -8827,8 +8825,7 @@ void md_do_sync(struct md_thread *thread) >> if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) >> goto skip; >> >> - if (test_bit(MD_RECOVERY_WAIT, &mddev->recovery) || >> - !md_is_rdwr(mddev)) {/* never try to sync a read-only array */ >> + if (!md_is_rdwr(mddev)) {/* never try to sync a read-only array */ >> set_bit(MD_RECOVERY_INTR, &mddev->recovery); >> goto skip; >> } >> diff --git a/drivers/md/md.h b/drivers/md/md.h >> index c17b7e68c533..eb00cdadc9c5 100644 >> --- a/drivers/md/md.h >> +++ b/drivers/md/md.h >> @@ -554,7 +554,6 @@ enum recovery_flags { >> MD_RECOVERY_RESHAPE, /* A reshape is happening */ >> MD_RECOVERY_FROZEN, /* User request to abort, and not >> restart, any action */ >> MD_RECOVERY_ERROR, /* sync-action interrupted because >> io-error */ >> - MD_RECOVERY_WAIT, /* waiting for pers->start() to finish */ >> MD_RESYNCING_REMOTE, /* remote node is running resync thread */ >> }; >> >> diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c >> index acce2868e491..530a7f0b120b 100644 >> --- a/drivers/md/raid5.c >> +++ b/drivers/md/raid5.c >> @@ -5919,7 +5919,6 @@ static int add_all_stripe_bios(struct r5conf *conf, >> static bool reshape_disabled(struct mddev *mddev) >> { >> return !md_is_rdwr(mddev) || >> - test_bit(MD_RECOVERY_WAIT, &mddev->recovery) || >> test_bit(MD_RECOVERY_FROZEN, &mddev->recovery); >> } >> >>> >>> Best Regards >>> Xiao >>>> >>>> Thanks, >>>> Kuai >>>> >>>>> >>>>> Regards >>>>> Xiao >>>>>> >>>>>> Thanks, >>>>>> Kuai >>>>>>> --- >>>>>>> drivers/md/dm-raid.c | 2 ++ >>>>>>> drivers/md/md.c | 1 + >>>>>>> 2 files changed, 3 insertions(+) >>>>>>> >>>>>>> diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c >>>>>>> index eb009d6bb03a..325767c1140f 100644 >>>>>>> --- a/drivers/md/dm-raid.c >>>>>>> +++ b/drivers/md/dm-raid.c >>>>>>> @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) >>>>>>> struct raid_set *rs = ti->private; >>>>>>> >>>>>>> if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { >>>>>>> + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) >>>>>>> + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); >>>>>>> /* Writes have to be stopped before suspending to avoid deadlocks. */ >>>>>>> if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) >>>>>>> md_stop_writes(&rs->md); >>>>>>> diff --git a/drivers/md/md.c b/drivers/md/md.c >>>>>>> index 2266358d8074..54790261254d 100644 >>>>>>> --- a/drivers/md/md.c >>>>>>> +++ b/drivers/md/md.c >>>>>>> @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) >>>>>>> * never happen >>>>>>> */ >>>>>>> md_wakeup_thread_directly(mddev->sync_thread); >>>>>>> + md_wakeup_thread(mddev->sync_thread); >>>>>>> if (work_pending(&mddev->sync_work)) >>>>>>> flush_work(&mddev->sync_work); >>>>>>> >>>>>>> >>>>>> >>>>> >>>>> . >>>>> >>>> >>> >>> . >>> >> > > . >
在 2024/2/26 下午5:36, Yu Kuai 写道: > Hi, > > 在 2024/02/26 13:12, Xiao Ni 写道: >> On Mon, Feb 26, 2024 at 9:31 AM Yu Kuai <yukuai1@huaweicloud.com> wrote: >>> >>> Hi, >>> >>> 在 2024/02/23 21:20, Xiao Ni 写道: >>>> On Fri, Feb 23, 2024 at 11:32 AM Yu Kuai <yukuai1@huaweicloud.com> >>>> wrote: >>>>> >>>>> Hi, >>>>> >>>>> 在 2024/02/20 23:30, Xiao Ni 写道: >>>>>> MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch >>>>>> commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). >>>>>> Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") >>>>>> dmraid stopped sync thread directy by calling md_reap_sync_thread. >>>>>> After this patch dmraid stops sync thread asynchronously as md does. >>>>>> This is right. Now the dmraid stop process is like this: >>>>>> >>>>>> 1. >>>>>> raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. >>>>>> stop_sync_thread sets MD_RECOVERY_INTR and wait until >>>>>> MD_RECOVERY_RUNNING >>>>>> is cleared >>>>>> 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the >>>>>> root cause for this deadlock. We hope md_do_sync can set >>>>>> MD_RECOVERY_DONE) >>>>>> 3. md thread calls md_check_recovery (This is the place to reap sync >>>>>> thread. Because MD_RECOVERY_DONE is not set. md thread can't reap >>>>>> sync >>>>>> thread) >>>>>> 4. raid_dtr stops/free struct mddev and release dmraid related >>>>>> resources >>>>>> >>>>>> dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs >>>>>> to clear >>>>>> this bit when stopping the dmraid before stopping sync thread. >>>>>> >>>>>> But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is >>>>>> cleared before stopping sync thread. It's the reason >>>>>> stop_sync_thread only >>>>>> wakes up task. If the task isn't running, it still needs to wake >>>>>> up sync >>>>>> thread too. >>>>>> >>>>>> This deadlock can be reproduced 100% by these commands: >>>>>> modprobe brd rd_size=34816 rd_nr=5 >>>>>> while [ 1 ]; do >>>>>> vgcreate test_vg /dev/ram* >>>>>> lvcreate --type raid5 -L 16M -n test_lv test_vg >>>>>> lvconvert -y --stripes 4 /dev/test_vg/test_lv >>>>>> vgremove test_vg -ff >>>>>> sleep 1 >>>>>> done >>>>>> >>>>>> Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") >>>>>> Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") >>>>>> Signed-off-by: Xiao Ni <xni@redhat.com> >>>>> >>>>> I'm not sure about this change, I think MD_RECOVERY_WAIT is hacky and >>>>> really breaks how sync_thread is working, it should just go away >>>>> soon, >>>>> once we make sure sync_thread can't be registered before >>>>> pers->start() >>>>> is done. >>>> >>>> Hi Kuai >>>> >>>> I just want to get to a stable state without changing any existing >>>> logic. After fixing these regressions, we can consider other changes. >>>> (e.g. remove MD_RECOVERY_WAIT. allow sync thread start/stop when array >>>> is suspend. ) I talked with Heinz yesterday, for dm-raid, it can't >>>> allow any io to happen including sync thread when array is suspended. >>> >> >> Hi Kuai >> >>> So, are you still thinking that my patchset will allow this for >>> dm-raid? >>> >>> I already explain a lot why patch 1 from my set is okay, and why the >>> set >>> doesn't introduce any behaviour change like you said in this patch 0: >> >> >> I'll read all your patches to understand you well. >> >> But as I mentioned many times too, we're fixing regression problems. >> It's better for us to fix them with the smallest change. As you can >> see, in my patch set, we can fix these regression problems with small >> changes (Sorry I didn't notice your patch set has some changes which >> are the same with mine). So why don't we need such a big change to >> fix the regression problems? Now with my patch set I can reproduce a >> problem by lvm2 test suit which happens in 6.6 too. It means with this >> patch set we can back to a state same with 6.6. > > For complexity, I agree that we can go back to the same state with v6.6, > and then fix other problems on the top of that. I don't have preference > for this, I'll post my patchset anyhow. But other than the test suite, > you still need to make sure nothing is broken from the big picture. > > For example, in the following 3 cases, can MD_RECOVERY_WAIT be cleared > as expected, and new sync_thread will not start after a reload? > > 1) > ioctl suspend > ioctl reload > ioctl resume > > 2) > ioctl reload > ioctl resume > > 3) > ioctl suspend > // without a reload. > ioctl resume Hi Kuai Is it the place that you think clearing MD_RECOVERY_WAIT is not safe? For the above three cases, I didn't do tests with dmsetup commands. Because lvm2 tests use lv commands, so I only checked the logs which I use printk to follow the functions calling sequence. l can make sure MD_RECOVERY_WAIT can be cleared when running lvconvert command. I'm not familiar with dmsetup commands. Could you give some examples fot the three above cases? I can do test myself here. Best Regards Xiao > >> >> >> >>> >>> "Kuai's patch set breaks some rules". >>> >>> The only thing that will change is that for md/raid, sync_thrad can >>> start for suspended array, however, I don't think this will be a >>> problem >>> because sync_thread can be running for suspended array already, and >>> 'MD_RECOVERY_FROZEN' is already used to prevent sync_thread to start. >> >> We can't allow sync thread happen for dmraid when it's suspended. >> Because it needs to switch table when suspended. It's a base design. >> If it can happen now. We should fix this. > > Yes, and my patchset also fix this, and other problems related to > sync_thread by managing sync_thread the same as md/raid. > > BTW, on the top my patchset, I already made some change locally to > expand MD_RECOVERY_FROZEN and remove MD_RECOVERY_WAIT, if you're going > to review my patchset, you can take a look at following change later, > I already tested the change. > > Thanks, > Kuai > > diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c > index e1b3b1917627..a0c8a5b92aab 100644 > --- a/drivers/md/dm-raid.c > +++ b/drivers/md/dm-raid.c > @@ -213,6 +213,7 @@ struct raid_dev { > #define RT_FLAG_RS_IN_SYNC 6 > #define RT_FLAG_RS_RESYNCING 7 > #define RT_FLAG_RS_GROW 8 > +#define RT_FLAG_WAIT_RELOAD 9 > > /* Array elements of 64 bit needed for rebuild/failed disk bits */ > #define DISKS_ARRAY_ELEMS ((MAX_RAID_DEVICES + (sizeof(uint64_t) * 8 > - 1)) / sizeof(uint64_t) / 8) > @@ -3728,6 +3729,9 @@ static int raid_message(struct dm_target *ti, > unsigned int argc, char **argv, > if (test_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) > return -EBUSY; > > + if (test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) > + return -EINVAL; > + > if (!strcasecmp(argv[0], "frozen")) { > ret = mddev_lock(mddev); > if (ret) > @@ -3809,21 +3813,25 @@ static void raid_presuspend(struct dm_target *ti) > struct raid_set *rs = ti->private; > struct mddev *mddev = &rs->md; > > - mddev_lock_nointr(mddev); > - md_frozen_sync_thread(mddev); > - mddev_unlock(mddev); > + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) { > + mddev_lock_nointr(mddev); > + md_frozen_sync_thread(mddev); > + mddev_unlock(mddev); > > - if (mddev->pers && mddev->pers->prepare_suspend) > - mddev->pers->prepare_suspend(mddev); > + if (mddev->pers && mddev->pers->prepare_suspend) > + mddev->pers->prepare_suspend(mddev); > + } > } > > static void raid_presuspend_undo(struct dm_target *ti) > { > struct raid_set *rs = ti->private; > > - mddev_lock_nointr(&rs->md); > - md_unfrozen_sync_thread(&rs->md); > - mddev_unlock(&rs->md); > + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) { > + mddev_lock_nointr(&rs->md); > + md_unfrozen_sync_thread(&rs->md); > + mddev_unlock(&rs->md); > + } > } > > static void raid_postsuspend(struct dm_target *ti) > @@ -3958,9 +3966,6 @@ static int rs_start_reshape(struct raid_set *rs) > struct mddev *mddev = &rs->md; > struct md_personality *pers = mddev->pers; > > - /* Don't allow the sync thread to work until the table gets > reloaded. */ > - set_bit(MD_RECOVERY_WAIT, &mddev->recovery); > - > r = rs_setup_reshape(rs); > if (r) > return r; > @@ -4055,6 +4060,7 @@ static int raid_preresume(struct dm_target *ti) > /* Initiate a reshape. */ > rs_set_rdev_sectors(rs); > mddev_lock_nointr(mddev); > + set_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags); > r = rs_start_reshape(rs); > mddev_unlock(mddev); > if (r) > @@ -4089,7 +4095,8 @@ static void raid_resume(struct dm_target *ti) > mddev_lock_nointr(mddev); > mddev->ro = 0; > mddev->in_sync = 0; > - md_unfrozen_sync_thread(mddev); > + if (!test_bit(RT_FLAG_WAIT_RELOAD, &rs->runtime_flags)) > + md_unfrozen_sync_thread(mddev); > mddev_unlock(mddev); > } > } > diff --git a/drivers/md/md.c b/drivers/md/md.c > index 81e6f49a14fc..595b1fbdce20 100644 > --- a/drivers/md/md.c > +++ b/drivers/md/md.c > @@ -6064,7 +6064,8 @@ int md_run(struct mddev *mddev) > pr_warn("True protection against single-disk > failure might be compromised.\n"); > } > > - mddev->recovery = 0; > + /* dm-raid expect sync_thread to be frozen until resume */ > + mddev->recovery &= BIT_ULL_MASK(MD_RECOVERY_FROZEN); > /* may be over-ridden by personality */ > mddev->resync_max_sectors = mddev->dev_sectors; > > @@ -6237,10 +6238,8 @@ int md_start(struct mddev *mddev) > int ret = 0; > > if (mddev->pers->start) { > - set_bit(MD_RECOVERY_WAIT, &mddev->recovery); > ret = mddev->pers->start(mddev); > - clear_bit(MD_RECOVERY_WAIT, &mddev->recovery); > - md_wakeup_thread(mddev->sync_thread); > + WARN_ON_ONCE(test_bit(MD_RECOVERY_RUNNING, > &mddev->recovery)); > } > return ret; > } > @@ -8827,8 +8825,7 @@ void md_do_sync(struct md_thread *thread) > if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) > goto skip; > > - if (test_bit(MD_RECOVERY_WAIT, &mddev->recovery) || > - !md_is_rdwr(mddev)) {/* never try to sync a read-only > array */ > + if (!md_is_rdwr(mddev)) {/* never try to sync a read-only > array */ > set_bit(MD_RECOVERY_INTR, &mddev->recovery); > goto skip; > } > diff --git a/drivers/md/md.h b/drivers/md/md.h > index c17b7e68c533..eb00cdadc9c5 100644 > --- a/drivers/md/md.h > +++ b/drivers/md/md.h > @@ -554,7 +554,6 @@ enum recovery_flags { > MD_RECOVERY_RESHAPE, /* A reshape is happening */ > MD_RECOVERY_FROZEN, /* User request to abort, and not > restart, any action */ > MD_RECOVERY_ERROR, /* sync-action interrupted because > io-error */ > - MD_RECOVERY_WAIT, /* waiting for pers->start() to finish */ > MD_RESYNCING_REMOTE, /* remote node is running resync > thread */ > }; > > diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c > index acce2868e491..530a7f0b120b 100644 > --- a/drivers/md/raid5.c > +++ b/drivers/md/raid5.c > @@ -5919,7 +5919,6 @@ static int add_all_stripe_bios(struct r5conf *conf, > static bool reshape_disabled(struct mddev *mddev) > { > return !md_is_rdwr(mddev) || > - test_bit(MD_RECOVERY_WAIT, &mddev->recovery) || > test_bit(MD_RECOVERY_FROZEN, &mddev->recovery); > } > >> >> Best Regards >> Xiao >>> >>> Thanks, >>> Kuai >>> >>>> >>>> Regards >>>> Xiao >>>>> >>>>> Thanks, >>>>> Kuai >>>>>> --- >>>>>> drivers/md/dm-raid.c | 2 ++ >>>>>> drivers/md/md.c | 1 + >>>>>> 2 files changed, 3 insertions(+) >>>>>> >>>>>> diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c >>>>>> index eb009d6bb03a..325767c1140f 100644 >>>>>> --- a/drivers/md/dm-raid.c >>>>>> +++ b/drivers/md/dm-raid.c >>>>>> @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct >>>>>> dm_target *ti) >>>>>> struct raid_set *rs = ti->private; >>>>>> >>>>>> if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, >>>>>> &rs->runtime_flags)) { >>>>>> + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) >>>>>> + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); >>>>>> /* Writes have to be stopped before suspending to >>>>>> avoid deadlocks. */ >>>>>> if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) >>>>>> md_stop_writes(&rs->md); >>>>>> diff --git a/drivers/md/md.c b/drivers/md/md.c >>>>>> index 2266358d8074..54790261254d 100644 >>>>>> --- a/drivers/md/md.c >>>>>> +++ b/drivers/md/md.c >>>>>> @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev >>>>>> *mddev, bool locked, bool check_seq) >>>>>> * never happen >>>>>> */ >>>>>> md_wakeup_thread_directly(mddev->sync_thread); >>>>>> + md_wakeup_thread(mddev->sync_thread); >>>>>> if (work_pending(&mddev->sync_work)) >>>>>> flush_work(&mddev->sync_work); >>>>>> >>>>>> >>>>> >>>> >>>> . >>>> >>> >> >> . >> >
diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c index eb009d6bb03a..325767c1140f 100644 --- a/drivers/md/dm-raid.c +++ b/drivers/md/dm-raid.c @@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti) struct raid_set *rs = ti->private; if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) { + if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery)) + clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery); /* Writes have to be stopped before suspending to avoid deadlocks. */ if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery)) md_stop_writes(&rs->md); diff --git a/drivers/md/md.c b/drivers/md/md.c index 2266358d8074..54790261254d 100644 --- a/drivers/md/md.c +++ b/drivers/md/md.c @@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq) * never happen */ md_wakeup_thread_directly(mddev->sync_thread); + md_wakeup_thread(mddev->sync_thread); if (work_pending(&mddev->sync_work)) flush_work(&mddev->sync_work);
MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock"). Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread") dmraid stopped sync thread directy by calling md_reap_sync_thread. After this patch dmraid stops sync thread asynchronously as md does. This is right. Now the dmraid stop process is like this: 1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread. stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING is cleared 2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE) 3. md thread calls md_check_recovery (This is the place to reap sync thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync thread) 4. raid_dtr stops/free struct mddev and release dmraid related resources dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear this bit when stopping the dmraid before stopping sync thread. But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is cleared before stopping sync thread. It's the reason stop_sync_thread only wakes up task. If the task isn't running, it still needs to wake up sync thread too. This deadlock can be reproduced 100% by these commands: modprobe brd rd_size=34816 rd_nr=5 while [ 1 ]; do vgcreate test_vg /dev/ram* lvcreate --type raid5 -L 16M -n test_lv test_vg lvconvert -y --stripes 4 /dev/test_vg/test_lv vgremove test_vg -ff sleep 1 done Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock") Fixes: f52f5c71f3d4 ("md: fix stopping sync thread") Signed-off-by: Xiao Ni <xni@redhat.com> --- drivers/md/dm-raid.c | 2 ++ drivers/md/md.c | 1 + 2 files changed, 3 insertions(+)