Re: [PATCH v3] mm, memcg: fix reclaim deadlock with writeback
On Thu 13-12-18 17:04:00, Johannes Weiner wrote: [...] > Acked-by: Johannes Weiner Thanks! > Just one nit: > > > @@ -2993,6 +2993,17 @@ static vm_fault_t __do_fault(struct vm_fault *vmf) > > struct vm_area_struct *vma = vmf->vma; > > vm_fault_t ret; > > > > + /* > > +* Preallocate pte before we take page_lock because this might lead to > > +* deadlocks for memcg reclaim which waits for pages under writeback. > > +*/ > > + if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { > > + vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm, > > vmf->address); > > + if (!vmf->prealloc_pte) > > + return VM_FAULT_OOM; > > + smp_wmb(); /* See comment in __pte_alloc() */ > > + } > > Could you be more specific in the deadlock comment? git blame will > work fine for a while, but it becomes a pain to find corresponding > patches after stuff gets moved around for years. > > In particular the race diagram between reclaim with a page lock held > and the fs doing SetPageWriteback batches before kicking off IO would > be useful directly in the code, IMO. This? diff --git a/mm/memory.c b/mm/memory.c index bb78e90a9b70..ece221e4da6d 100644 --- a/mm/memory.c +++ b/mm/memory.c @@ -2995,7 +2995,18 @@ static vm_fault_t __do_fault(struct vm_fault *vmf) /* * Preallocate pte before we take page_lock because this might lead to -* deadlocks for memcg reclaim which waits for pages under writeback. +* deadlocks for memcg reclaim which waits for pages under writeback: +* lock_page(A) +* SetPageWriteback(A) +* unlock_page(A) +* lock_page(B) +* lock_page(B) +* pte_alloc_pne +* shrink_page_list +* wait_on_page_writeback(A) +* SetPageWriteback(B) +* unlock_page(B) +* # flush A, B to clear the writeback */ if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm, vmf->address); -- Michal Hocko SUSE Labs
Re: [PATCH v3] mm, memcg: fix reclaim deadlock with writeback
On Thu, Dec 13, 2018 at 10:22:21AM +0100, Michal Hocko wrote: > From: Michal Hocko > > Liu Bo has experienced a deadlock between memcg (legacy) reclaim and the > ext4 writeback > task1: > [] wait_on_page_bit+0x82/0xa0 > [] shrink_page_list+0x907/0x960 > [] shrink_inactive_list+0x2c7/0x680 > [] shrink_node_memcg+0x404/0x830 > [] shrink_node+0xd8/0x300 > [] do_try_to_free_pages+0x10d/0x330 > [] try_to_free_mem_cgroup_pages+0xd5/0x1b0 > [] try_charge+0x14d/0x720 > [] memcg_kmem_charge_memcg+0x3c/0xa0 > [] memcg_kmem_charge+0x7e/0xd0 > [] __alloc_pages_nodemask+0x178/0x260 > [] alloc_pages_current+0x95/0x140 > [] pte_alloc_one+0x17/0x40 > [] __pte_alloc+0x1e/0x110 > [] alloc_set_pte+0x5fe/0xc20 > [] do_fault+0x103/0x970 > [] handle_mm_fault+0x61e/0xd10 > [] __do_page_fault+0x252/0x4d0 > [] do_page_fault+0x30/0x80 > [] page_fault+0x28/0x30 > [] 0x > > task2: > [] __lock_page+0x86/0xa0 > [] mpage_prepare_extent_to_map+0x2e7/0x310 [ext4] > [] ext4_writepages+0x479/0xd60 > [] do_writepages+0x1e/0x30 > [] __writeback_single_inode+0x45/0x320 > [] writeback_sb_inodes+0x272/0x600 > [] __writeback_inodes_wb+0x92/0xc0 > [] wb_writeback+0x268/0x300 > [] wb_workfn+0xb4/0x390 > [] process_one_work+0x189/0x420 > [] worker_thread+0x4e/0x4b0 > [] kthread+0xe6/0x100 > [] ret_from_fork+0x41/0x50 > [] 0x > > He adds > : task1 is waiting for the PageWriteback bit of the page that task2 has > : collected in mpd->io_submit->io_bio, and tasks2 is waiting for the LOCKED > : bit the page which tasks1 has locked. > > More precisely task1 is handling a page fault and it has a page locked > while it charges a new page table to a memcg. That in turn hits a memory > limit reclaim and the memcg reclaim for legacy controller is waiting on > the writeback but that is never going to finish because the writeback > itself is waiting for the page locked in the #PF path. So this is > essentially ABBA deadlock: > lock_page(A) > SetPageWriteback(A) > unlock_page(A) > lock_page(B) > lock_page(B) > pte_alloc_pne > shrink_page_list > wait_on_page_writeback(A) > SetPageWriteback(B) > unlock_page(B) > > # flush A, B to clear the writeback > > This accumulating of more pages to flush is used by several filesystems > to generate a more optimal IO patterns. > > Waiting for the writeback in legacy memcg controller is a workaround > for pre-mature OOM killer invocations because there is no dirty IO > throttling available for the controller. There is no easy way around > that unfortunately. Therefore fix this specific issue by pre-allocating > the page table outside of the page lock. We have that handy > infrastructure for that already so simply reuse the fault-around pattern > which already does this. > > There are probably other hidden __GFP_ACCOUNT | GFP_KERNEL allocations > from under a fs page locked but they should be really rare. I am not > aware of a better solution unfortunately. > Thanks for the update. Looks good to me. Reviewed-by: Liu Bo thanks, -liubo > Reported-and-Debugged-by: Liu Bo > Cc: stable > Fixes: c3b94f44fcb0 ("memcg: further prevent OOM with too many dirty pages") > Signed-off-by: Michal Hocko > --- > mm/memory.c | 11 +++ > 1 file changed, 11 insertions(+) > > diff --git a/mm/memory.c b/mm/memory.c > index 4ad2d293ddc2..bb78e90a9b70 100644 > --- a/mm/memory.c > +++ b/mm/memory.c > @@ -2993,6 +2993,17 @@ static vm_fault_t __do_fault(struct vm_fault *vmf) > struct vm_area_struct *vma = vmf->vma; > vm_fault_t ret; > > + /* > + * Preallocate pte before we take page_lock because this might lead to > + * deadlocks for memcg reclaim which waits for pages under writeback. > + */ > + if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { > + vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm, > vmf->address); > + if (!vmf->prealloc_pte) > + return VM_FAULT_OOM; > + smp_wmb(); /* See comment in __pte_alloc() */ > + } > + > ret = vma->vm_ops->fault(vmf); > if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY | > VM_FAULT_DONE_COW))) > -- > 2.19.2
Re: [PATCH v3] mm, memcg: fix reclaim deadlock with writeback
On Thu, Dec 13, 2018 at 10:22:21AM +0100, Michal Hocko wrote: > From: Michal Hocko > > Liu Bo has experienced a deadlock between memcg (legacy) reclaim and the > ext4 writeback > task1: > [] wait_on_page_bit+0x82/0xa0 > [] shrink_page_list+0x907/0x960 > [] shrink_inactive_list+0x2c7/0x680 > [] shrink_node_memcg+0x404/0x830 > [] shrink_node+0xd8/0x300 > [] do_try_to_free_pages+0x10d/0x330 > [] try_to_free_mem_cgroup_pages+0xd5/0x1b0 > [] try_charge+0x14d/0x720 > [] memcg_kmem_charge_memcg+0x3c/0xa0 > [] memcg_kmem_charge+0x7e/0xd0 > [] __alloc_pages_nodemask+0x178/0x260 > [] alloc_pages_current+0x95/0x140 > [] pte_alloc_one+0x17/0x40 > [] __pte_alloc+0x1e/0x110 > [] alloc_set_pte+0x5fe/0xc20 > [] do_fault+0x103/0x970 > [] handle_mm_fault+0x61e/0xd10 > [] __do_page_fault+0x252/0x4d0 > [] do_page_fault+0x30/0x80 > [] page_fault+0x28/0x30 > [] 0x > > task2: > [] __lock_page+0x86/0xa0 > [] mpage_prepare_extent_to_map+0x2e7/0x310 [ext4] > [] ext4_writepages+0x479/0xd60 > [] do_writepages+0x1e/0x30 > [] __writeback_single_inode+0x45/0x320 > [] writeback_sb_inodes+0x272/0x600 > [] __writeback_inodes_wb+0x92/0xc0 > [] wb_writeback+0x268/0x300 > [] wb_workfn+0xb4/0x390 > [] process_one_work+0x189/0x420 > [] worker_thread+0x4e/0x4b0 > [] kthread+0xe6/0x100 > [] ret_from_fork+0x41/0x50 > [] 0x > > He adds > : task1 is waiting for the PageWriteback bit of the page that task2 has > : collected in mpd->io_submit->io_bio, and tasks2 is waiting for the LOCKED > : bit the page which tasks1 has locked. > > More precisely task1 is handling a page fault and it has a page locked > while it charges a new page table to a memcg. That in turn hits a memory > limit reclaim and the memcg reclaim for legacy controller is waiting on > the writeback but that is never going to finish because the writeback > itself is waiting for the page locked in the #PF path. So this is > essentially ABBA deadlock: > lock_page(A) > SetPageWriteback(A) > unlock_page(A) > lock_page(B) > lock_page(B) > pte_alloc_pne > shrink_page_list > wait_on_page_writeback(A) > SetPageWriteback(B) > unlock_page(B) > > # flush A, B to clear the writeback > > This accumulating of more pages to flush is used by several filesystems > to generate a more optimal IO patterns. > > Waiting for the writeback in legacy memcg controller is a workaround > for pre-mature OOM killer invocations because there is no dirty IO > throttling available for the controller. There is no easy way around > that unfortunately. Therefore fix this specific issue by pre-allocating > the page table outside of the page lock. We have that handy > infrastructure for that already so simply reuse the fault-around pattern > which already does this. > > There are probably other hidden __GFP_ACCOUNT | GFP_KERNEL allocations > from under a fs page locked but they should be really rare. I am not > aware of a better solution unfortunately. > > Reported-and-Debugged-by: Liu Bo > Cc: stable > Fixes: c3b94f44fcb0 ("memcg: further prevent OOM with too many dirty pages") > Signed-off-by: Michal Hocko Acked-by: Johannes Weiner Just one nit: > @@ -2993,6 +2993,17 @@ static vm_fault_t __do_fault(struct vm_fault *vmf) > struct vm_area_struct *vma = vmf->vma; > vm_fault_t ret; > > + /* > + * Preallocate pte before we take page_lock because this might lead to > + * deadlocks for memcg reclaim which waits for pages under writeback. > + */ > + if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { > + vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm, > vmf->address); > + if (!vmf->prealloc_pte) > + return VM_FAULT_OOM; > + smp_wmb(); /* See comment in __pte_alloc() */ > + } Could you be more specific in the deadlock comment? git blame will work fine for a while, but it becomes a pain to find corresponding patches after stuff gets moved around for years. In particular the race diagram between reclaim with a page lock held and the fs doing SetPageWriteback batches before kicking off IO would be useful directly in the code, IMO.
Re: [PATCH v3] mm, memcg: fix reclaim deadlock with writeback
On Thu 13-12-18 13:41:47, Kirill A. Shutemov wrote: > On Thu, Dec 13, 2018 at 10:22:21AM +0100, Michal Hocko wrote: > > From: Michal Hocko > > > > Liu Bo has experienced a deadlock between memcg (legacy) reclaim and the > > ext4 writeback > > task1: > > [] wait_on_page_bit+0x82/0xa0 > > [] shrink_page_list+0x907/0x960 > > [] shrink_inactive_list+0x2c7/0x680 > > [] shrink_node_memcg+0x404/0x830 > > [] shrink_node+0xd8/0x300 > > [] do_try_to_free_pages+0x10d/0x330 > > [] try_to_free_mem_cgroup_pages+0xd5/0x1b0 > > [] try_charge+0x14d/0x720 > > [] memcg_kmem_charge_memcg+0x3c/0xa0 > > [] memcg_kmem_charge+0x7e/0xd0 > > [] __alloc_pages_nodemask+0x178/0x260 > > [] alloc_pages_current+0x95/0x140 > > [] pte_alloc_one+0x17/0x40 > > [] __pte_alloc+0x1e/0x110 > > [] alloc_set_pte+0x5fe/0xc20 > > [] do_fault+0x103/0x970 > > [] handle_mm_fault+0x61e/0xd10 > > [] __do_page_fault+0x252/0x4d0 > > [] do_page_fault+0x30/0x80 > > [] page_fault+0x28/0x30 > > [] 0x > > > > task2: > > [] __lock_page+0x86/0xa0 > > [] mpage_prepare_extent_to_map+0x2e7/0x310 [ext4] > > [] ext4_writepages+0x479/0xd60 > > [] do_writepages+0x1e/0x30 > > [] __writeback_single_inode+0x45/0x320 > > [] writeback_sb_inodes+0x272/0x600 > > [] __writeback_inodes_wb+0x92/0xc0 > > [] wb_writeback+0x268/0x300 > > [] wb_workfn+0xb4/0x390 > > [] process_one_work+0x189/0x420 > > [] worker_thread+0x4e/0x4b0 > > [] kthread+0xe6/0x100 > > [] ret_from_fork+0x41/0x50 > > [] 0x > > > > He adds > > : task1 is waiting for the PageWriteback bit of the page that task2 has > > : collected in mpd->io_submit->io_bio, and tasks2 is waiting for the LOCKED > > : bit the page which tasks1 has locked. > > > > More precisely task1 is handling a page fault and it has a page locked > > while it charges a new page table to a memcg. That in turn hits a memory > > limit reclaim and the memcg reclaim for legacy controller is waiting on > > the writeback but that is never going to finish because the writeback > > itself is waiting for the page locked in the #PF path. So this is > > essentially ABBA deadlock: > > lock_page(A) > > SetPageWriteback(A) > > unlock_page(A) > > lock_page(B) > > lock_page(B) > > pte_alloc_pne > > shrink_page_list > > wait_on_page_writeback(A) > > SetPageWriteback(B) > > unlock_page(B) > > > > # flush A, B to clear the writeback > > > > This accumulating of more pages to flush is used by several filesystems > > to generate a more optimal IO patterns. > > > > Waiting for the writeback in legacy memcg controller is a workaround > > for pre-mature OOM killer invocations because there is no dirty IO > > throttling available for the controller. There is no easy way around > > that unfortunately. Therefore fix this specific issue by pre-allocating > > the page table outside of the page lock. We have that handy > > infrastructure for that already so simply reuse the fault-around pattern > > which already does this. > > > > There are probably other hidden __GFP_ACCOUNT | GFP_KERNEL allocations > > from under a fs page locked but they should be really rare. I am not > > aware of a better solution unfortunately. > > > > Reported-and-Debugged-by: Liu Bo > > Cc: stable > > Fixes: c3b94f44fcb0 ("memcg: further prevent OOM with too many dirty pages") > > Signed-off-by: Michal Hocko > > Acked-by: Kirill A. Shutemov Thanks! > Will you take care about converting vmf_insert_* to use the pre-allocated > page table? I can try but I would appreciate if somebody more familiar with the code could do that. I am busy as hell and I do not want to promis something I will likely not get to soon. -- Michal Hocko SUSE Labs
Re: [PATCH v3] mm, memcg: fix reclaim deadlock with writeback
On Thu, Dec 13, 2018 at 10:22:21AM +0100, Michal Hocko wrote: > From: Michal Hocko > > Liu Bo has experienced a deadlock between memcg (legacy) reclaim and the > ext4 writeback > task1: > [] wait_on_page_bit+0x82/0xa0 > [] shrink_page_list+0x907/0x960 > [] shrink_inactive_list+0x2c7/0x680 > [] shrink_node_memcg+0x404/0x830 > [] shrink_node+0xd8/0x300 > [] do_try_to_free_pages+0x10d/0x330 > [] try_to_free_mem_cgroup_pages+0xd5/0x1b0 > [] try_charge+0x14d/0x720 > [] memcg_kmem_charge_memcg+0x3c/0xa0 > [] memcg_kmem_charge+0x7e/0xd0 > [] __alloc_pages_nodemask+0x178/0x260 > [] alloc_pages_current+0x95/0x140 > [] pte_alloc_one+0x17/0x40 > [] __pte_alloc+0x1e/0x110 > [] alloc_set_pte+0x5fe/0xc20 > [] do_fault+0x103/0x970 > [] handle_mm_fault+0x61e/0xd10 > [] __do_page_fault+0x252/0x4d0 > [] do_page_fault+0x30/0x80 > [] page_fault+0x28/0x30 > [] 0x > > task2: > [] __lock_page+0x86/0xa0 > [] mpage_prepare_extent_to_map+0x2e7/0x310 [ext4] > [] ext4_writepages+0x479/0xd60 > [] do_writepages+0x1e/0x30 > [] __writeback_single_inode+0x45/0x320 > [] writeback_sb_inodes+0x272/0x600 > [] __writeback_inodes_wb+0x92/0xc0 > [] wb_writeback+0x268/0x300 > [] wb_workfn+0xb4/0x390 > [] process_one_work+0x189/0x420 > [] worker_thread+0x4e/0x4b0 > [] kthread+0xe6/0x100 > [] ret_from_fork+0x41/0x50 > [] 0x > > He adds > : task1 is waiting for the PageWriteback bit of the page that task2 has > : collected in mpd->io_submit->io_bio, and tasks2 is waiting for the LOCKED > : bit the page which tasks1 has locked. > > More precisely task1 is handling a page fault and it has a page locked > while it charges a new page table to a memcg. That in turn hits a memory > limit reclaim and the memcg reclaim for legacy controller is waiting on > the writeback but that is never going to finish because the writeback > itself is waiting for the page locked in the #PF path. So this is > essentially ABBA deadlock: > lock_page(A) > SetPageWriteback(A) > unlock_page(A) > lock_page(B) > lock_page(B) > pte_alloc_pne > shrink_page_list > wait_on_page_writeback(A) > SetPageWriteback(B) > unlock_page(B) > > # flush A, B to clear the writeback > > This accumulating of more pages to flush is used by several filesystems > to generate a more optimal IO patterns. > > Waiting for the writeback in legacy memcg controller is a workaround > for pre-mature OOM killer invocations because there is no dirty IO > throttling available for the controller. There is no easy way around > that unfortunately. Therefore fix this specific issue by pre-allocating > the page table outside of the page lock. We have that handy > infrastructure for that already so simply reuse the fault-around pattern > which already does this. > > There are probably other hidden __GFP_ACCOUNT | GFP_KERNEL allocations > from under a fs page locked but they should be really rare. I am not > aware of a better solution unfortunately. > > Reported-and-Debugged-by: Liu Bo > Cc: stable > Fixes: c3b94f44fcb0 ("memcg: further prevent OOM with too many dirty pages") > Signed-off-by: Michal Hocko Acked-by: Kirill A. Shutemov Will you take care about converting vmf_insert_* to use the pre-allocated page table? -- Kirill A. Shutemov