HWpoison: further fixes and cleanups
Subsystems (3)
v2 cover HWpoison: further fixes and cleanups
The important bit of this patchset is patch#1, which is a fix to take off HWPoison pages off a buddy freelist since it can lead us to having HWPoison pages back in the game without no one noticing it.
So fix it (we did that already for soft_offline_page [1]).
The other patches are clean-ups and not that important, so if anything, consider patch#1 for inclusion.
[1] https://patchwork.kernel.org/cover/11704083/
Thanks
Oscar Salvador (5):
mm,hwpoison: Take free pages off the buddy freelists mm,hwpoison: Refactor madvise_inject_error mm,hwpoison: Drain pcplists before bailing out for non-buddy
zero-refcount page
mm,hwpoison: Drop unneeded pcplist draining mm,hwpoison: Remove stale code
mm/madvise.c | 34 ++++++++++++------------------ mm/memory-failure.c | 50 +++++++++++++++++++++++++++++++++------------ 2 files changed, 50 insertions(+), 34 deletions(-)
Signature · 1 line
2.26.2
PATCH v2 1/5 mm,hwpoison: Take free pages off the buddy freelists
The crux of the matter is that historically we left poisoned pages in the buddy system because we have some checks in place when allocating a page that a gatekeeper for poisoned pages. Unfortunately, we do have other users (e.g: compaction [1]) that scan buddy freelists and try to get a page from there without checking whether the page is HWPoison.
As I stated already, I think it is fundamentally wrong to keep HWPoison pages within the buddy systems, checks in place or not.
Let us fix this we same way we did for soft_offline [2], and take the page off the buddy freelist, so it is completely unreachable.
Note that this is fairly simple to trigger, as we only need to poison free buddy pages (madvise MADV_HWPOISON) and then we need to run some sort of memory stress system.
Just for a matter of reference, I put a dump_page in compaction_alloc to trigger for HWPoison patches:
kernel: page:0000000012b2982b refcount:1 mapcount:0 mapping:0000000000000000 index:0x1 pfn:0x1d5db
kernel: flags: 0xfffffc0800000(hwpoison)
kernel: raw: 000fffffc0800000 ffffea00007573c8 ffffc90000857de0 0000000000000000
kernel: raw: 0000000000000001 0000000000000000 00000001ffffffff 0000000000000000
kernel: page dumped because: compaction_alloc
kernel: CPU: 4 PID: 123 Comm: kcompactd0 Tainted: G E 5.9.0-rc2-mm1-1-default+ #5 kernel: Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 kernel: Call Trace: kernel: dump_stack+0x6d/0x8b kernel: compaction_alloc+0xb2/0xc0 kernel: migrate_pages+0x2a6/0x12a0 kernel: ? isolate_freepages+0xc80/0xc80 kernel: ? __ClearPageMovable+0xb0/0xb0 kernel: compact_zone+0x5eb/0x11c0 kernel: ? finish_task_switch+0x74/0x300 kernel: ? lock_timer_base+0xa8/0x170 kernel: proactive_compact_node+0x89/0xf0 kernel: ? kcompactd+0x2d0/0x3a0 kernel: kcompactd+0x2d0/0x3a0 kernel: ? finish_wait+0x80/0x80 kernel: ? kcompactd_do_work+0x350/0x350 kernel: kthread+0x118/0x130 kernel: ? kthread_associate_blkcg+0xa0/0xa0 kernel: ret_from_fork+0x22/0x30
After that, if e.g: someone faults in the page, that someone will get killed unexpectedly.
While we are at it, I also changed the action result for such cases. I think that MF_DELAYED is a bit misleading, because in case we could contain the page and take it off the buddy, such a page is not to be used again unless it is unpoison, so we "fixed" the situation.
So unless I am missing something, I strongly think that we should report MF_RECOVERED.
[1] https://lore.kernel.org/linux-mm/20190826104144.GA7849@linux/T/#u
[2] https://patchwork.kernel.org/patch/11694847/
--- mm/memory-failure.c | 28 ++++++++++++++++++++++------ 1 file changed, 22 insertions(+), 6 deletions(-)
mm/memory-failure.c +22 -6
@@ -1323,8 +1323,9 @@ int memory_failure(unsigned long pfn, int flags)struct page *hpage;struct page *orig_head;struct dev_pagemap *pgmap; -int res;unsigned long page_flags; +int res =0 ; +bool retry = true;if (!sysctl_memory_failure_recovery)panic ("Memory failure on page %lx" , pfn);
@@ -1364,10 +1365,21 @@ int memory_failure(unsigned long pfn, int flags)* In fact it's dangerous to directly bump up page count from 0, * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. */ +try_again:if (!(flags &MF_COUNT_INCREASED ) && !get_hwpoison_page (p)) {if (is_free_buddy_page (p)) { -action_result (pfn,MF_MSG_BUDDY ,MF_DELAYED ); -return 0 ; +if (take_page_off_buddy (p)) { +action_result (pfn,MF_MSG_BUDDY ,MF_RECOVERED ); +return 0 ; + }else { +/* We lost the race, try again */ +if (retry) { + retry = false; + goto try_again; + } +action_result (pfn,MF_MSG_BUDDY ,MF_IGNORED ); +return -EBUSY ; + } }else {action_result (pfn,MF_MSG_KERNEL_HIGH_ORDER ,MF_IGNORED );return -EBUSY ;
@@ -1393,11 +1405,15 @@ int memory_failure(unsigned long pfn, int flags)shake_page (p,0 );/* shake_page could have turned it free. */ if (!PageLRU (p) &&is_free_buddy_page (p)) { +if (!take_page_off_buddy (p)) + res = -EBUSY ; +if (flags &MF_COUNT_INCREASED ) -action_result (pfn,MF_MSG_BUDDY ,MF_DELAYED ); +action_result (pfn,MF_MSG_BUDDY , res ?MF_IGNORED :MF_RECOVERED );else -action_result (pfn,MF_MSG_BUDDY_2ND ,MF_DELAYED ); -return 0 ; +action_result (pfn,MF_MSG_BUDDY_2ND , res ?MF_IGNORED :MF_RECOVERED ); + +return res; }lock_page (p);
Signature · 1 line
2.26.2
Re: [PATCH v2 1/5] mm,hwpoison: Take free pages off the buddy freelists
Quoted 46 lines · Oscar Salvador
The crux of the matter is that historically we left poisoned pages in the buddy system because we have some checks in place when allocating a page that a gatekeeper for poisoned pages. Unfortunately, we do have other users (e.g: compaction [1]) that scan buddy freelists and try to get a page from there without checking whether the page is HWPoison.
As I stated already, I think it is fundamentally wrong to keep HWPoison pages within the buddy systems, checks in place or not.
Let us fix this we same way we did for
soft_offline[2], and take the page off the buddy freelist, so it is completely unreachable.Note that this is fairly simple to trigger, as we only need to poison free buddy pages (madvise
MADV_HWPOISON) and then we need to run some sort of memory stress system.Just for a matter of reference, I put a
dump_pageincompaction_allocto trigger for HWPoison patches:kernel: page:0000000012b2982b refcount:1 mapcount:0 mapping:0000000000000000 index:0x1 pfn:0x1d5db
kernel: flags: 0xfffffc0800000(hwpoison)
kernel: raw:000fffffc0800000ffffea00007573c8ffffc90000857de00000000000000000
kernel: raw: 0000000000000001 000000000000000000000001ffffffff0000000000000000
kernel: page dumped because:compaction_allockernel: CPU: 4 PID: 123 Comm: kcompactd0 Tainted: G E 5.9.0-rc2-mm1-1-default+ #5 kernel: Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 kernel: Call Trace: kernel: dump_stack+0x6d/0x8b kernel: compaction_alloc+0xb2/0xc0 kernel: migrate_pages+0x2a6/0x12a0 kernel: ? isolate_freepages+0xc80/0xc80 kernel: ? __ClearPageMovable+0xb0/0xb0 kernel: compact_zone+0x5eb/0x11c0 kernel: ? finish_task_switch+0x74/0x300 kernel: ? lock_timer_base+0xa8/0x170 kernel: proactive_compact_node+0x89/0xf0 kernel: ? kcompactd+0x2d0/0x3a0 kernel: kcompactd+0x2d0/0x3a0 kernel: ? finish_wait+0x80/0x80 kernel: ? kcompactd_do_work+0x350/0x350 kernel: kthread+0x118/0x130 kernel: ? kthread_associate_blkcg+0xa0/0xa0 kernel: ret_from_fork+0x22/0x30
Quoted 7 lines
After that, if e.g: someone faults in the page, that someone will get killed unexpectedly.
While we are at it, I also changed the action result for such cases. I think that
MF_DELAYEDis a bit misleading, because in case we could contain the page and take it off the buddy, such a page is not to be used again unless it is unpoison, so we "fixed" the situation.
Thanks for pointing out this.
Originally (before the reported issue related to compaction is recognized), this case is categorized as MF_DELAYED because the error page is to be removed from buddy later when page allocator finds it in subsequent allocation time (not in memory_failure() time).
But this patch changes the situation, and the removal from buddy is done immediately in memory_failure(), so ...
Quoted 3 lines
So unless I am missing something, I strongly think that we should report
MF_RECOVERED.
this change looks correct to me.
Quoted 33 lines
[1] https://lore.kernel.org/linux-mm/20190826104144.GA7849@linux/T/#u
[2] https://patchwork.kernel.org/patch/11694847/Signed-off-by Oscar Salvador <[email protected]>--- mm/memory-failure.c | 28 ++++++++++++++++++++++------ 1 file changed, 22 insertions(+), 6 deletions(-)mm/memory-failure.c +3
-3@@ -1323,8 +1323,9 @@ int memory_failure(unsigned long pfn, int flags)struct page *hpage;struct page *orig_head;struct dev_pagemap *pgmap; -int res;unsigned long page_flags; +int res =0 ; +bool retry = true;if (!sysctl_memory_failure_recovery)panic ("Memory failure on page %lx" , pfn);@@ -1364,10 +1365,21 @@ int memory_failure(unsigned long pfn, int flags)* In fact it's dangerous to directly bump up page count from 0, * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. */ +try_again:if (!(flags &MF_COUNT_INCREASED ) && !get_hwpoison_page (p)) {if (is_free_buddy_page (p)) { -action_result (pfn,MF_MSG_BUDDY ,MF_DELAYED ); -return 0 ;
mm/memory-failure.c int memory_failure(unsigned long pfn, int flags) · line 1380
Quoted 10 lines
mm/memory-failure.c +10
-0+if (take_page_off_buddy (p)) { +action_result (pfn,MF_MSG_BUDDY ,MF_RECOVERED ); +return 0 ; + }else { +/* We lost the race, try again */ +if (retry) { + retry = false; + goto try_again; + } +action_result (pfn,MF_MSG_BUDDY ,MF_IGNORED );
According to include/linux/mm.h MF_IGNORED means "Error: cannot be handled",
/*
* Error handlers for various types of pages.
*/
enum mf_result {
MF_IGNORED, /* Error: cannot be handled */
MF_FAILED, /* Error: handling failed */
MF_DELAYED, /* Will be handled later */
MF_RECOVERED, /* Successfully recovered */
};
This case is an occasional failure, so maybe MF_FAILED seems better to me.
Quoted 11 lines
mm/memory-failure.c +4
-0+return -EBUSY ; + } }else {action_result (pfn,MF_MSG_KERNEL_HIGH_ORDER ,MF_IGNORED );return -EBUSY ;@@ -1393,11 +1405,15 @@ int memory_failure(unsigned long pfn, int flags)shake_page (p,0 );/* shake_page could have turned it free. */ if (!PageLRU (p) &&is_free_buddy_page (p)) { +if (!take_page_off_buddy (p)) + res = -EBUSY ;
mm/memory-failure.c int memory_failure(unsigned long pfn, int flags) · line 1416
Quoted 10 lines
mm/memory-failure.c +5
-3+if (flags &MF_COUNT_INCREASED ) -action_result (pfn,MF_MSG_BUDDY ,MF_DELAYED ); +action_result (pfn,MF_MSG_BUDDY , res ?MF_IGNORED :MF_RECOVERED );else -action_result (pfn,MF_MSG_BUDDY_2ND ,MF_DELAYED ); -return 0 ; +action_result (pfn,MF_MSG_BUDDY_2ND , res ?MF_IGNORED :MF_RECOVERED ); + +return res;
Althouth this code will be removed in 5/5, you can use MF_FAILED in failure case too.
Thanks,
Naoya Horiguchi
PATCH v2 2/5 mm,hwpoison: Refactor madvise_inject_error
In reply to Oscar SalvadorMake a proper if-else condition for {hard,soft}-offline.
--- mm/madvise.c | 30 +++++++++++++----------------- 1 file changed, 13 insertions(+), 17 deletions(-)
mm/madvise.c +13 -17
@@ -885,7 +885,6 @@ static long madvise_remove(struct vm_area_struct *vma,static int madvise_inject_error (int behavior,unsigned long start,unsigned long end) { -struct page *page;struct zone *zone;unsigned long size;
@@ -895,6 +894,7 @@ static int madvise_inject_error(int behavior,for (; start < end; start += size) {unsigned long pfn; +struct page *page;int ret; ret =get_user_pages_fast (start,1 ,0 , &page);
@@ -911,25 +911,21 @@ static int madvise_inject_error(int behavior,if (behavior ==MADV_SOFT_OFFLINE ) {pr_info ("Soft offlining pfn %#lx at process virtual address %#lx\n" , - pfn, start); - + pfn, start); ret =soft_offline_page (pfn,MF_COUNT_INCREASED ); -if (ret) -return ret; -continue ; + }else { +pr_info ("Injecting memory failure for pfn %#lx at process virtual address %#lx\n" , + pfn, start); +/* +* Drop the page reference taken by get_user_pages_fast(). In +* the absence of MF_COUNT_INCREASED the memory_failure() +* routine is responsible for pinning the page to prevent it +* from being released back to the page allocator. +*/ +put_page (page); + ret =memory_failure (pfn,0 ); } -pr_info ("Injecting memory failure for pfn %#lx at process virtual address %#lx\n" , - pfn, start); - -/* -* Drop the page reference taken by get_user_pages_fast(). In -* the absence of MF_COUNT_INCREASED the memory_failure() -* routine is responsible for pinning the page to prevent it -* from being released back to the page allocator. -*/ -put_page (page); - ret =memory_failure (pfn,0 );if (ret)return ret; }
Signature · 1 line
2.26.2
PATCH v2 3/5 mm,hwpoison: Drain pcplists before bailing out for non-buddy zero-refcount page
In reply to Oscar SalvadorA page with 0-refcount and !PageBuddy could perfectly be a pcppage. Currently, we bail out with an error if we encounter such a page, meaning that we do not handle pcppages neither from hard-offline nor from soft-offline path.
Fix this by draining pcplists whenever we find this kind of page and retry the check again.
It might be that pcplists have been spilled into the buddy allocator and so we can handle it.
--- mm/memory-failure.c | 24 ++++++++++++++++++++++-- 1 file changed, 22 insertions(+), 2 deletions(-)
mm/memory-failure.c +22 -2
@@ -950,13 +950,13 @@ static int page_action(struct page_state *ps, struct page *p, }/** -* get_hwpoison_page() - Get refcount for memory error handling: +* __get_hwpoison_page() - Get refcount for memory error handling: * @page: raw error page (hit by memory error) * * Return: return 0 if failed to grab the refcount, otherwise true (some * non-zero value.) */ -static int get_hwpoison_page (struct page *page) +static int __get_hwpoison_page (struct page *page) {struct page *head =compound_head (page);
@@ -986,6 +986,26 @@ static int get_hwpoison_page(struct page *page)return 0 ; } +static int get_hwpoison_page (struct page *p) +{ +int ret; +bool drained = false; + +retry: + ret =__get_hwpoison_page (p); +if (!ret && !is_free_buddy_page (p) && !page_count (p) && !drained) { +/* +* The page might be in a pcplist, so try to drain those +* and see if we are lucky. +*/ +drain_all_pages (page_zone (p)); + drained = true; + goto retry; + } + +return ret; +} +/* * Do all that is necessary to remove user space mappings. Unmap * the pages and send SIGBUS to the processes if the data was dirty.
Signature · 1 line
2.26.2
PATCH v2 4/5 mm,hwpoison: Drop unneeded pcplist draining
In reply to Oscar Salvadormemory_failure and soft_offline_path paths now drain pcplists by calling get_hwpoison_page.memory_failure flags the page as HWPoison before, so that page cannot longer go into a pcplist, and soft_offline_page only flags a page as HWPoison if 1) we took the page off a buddy freelist 2) the page was in-use and we migrated it 3) was a clean pagecache.
Because of that, a page cannot longer be poisoned and be in a pcplist.
--- mm/madvise.c | 4 ---- 1 file changed, 4 deletions(-)
mm/madvise.c +0 -4
@@ -930,10 +930,6 @@ static int madvise_inject_error(int behavior, return ret; } -/* Ensure that all poisoned pages are removed from per-cpu lists */ -for_each_populated_zone (zone) -drain_all_pages (zone); -return 0 ; }#endif
Signature · 1 line
2.26.2
PATCH v2 5/5 mm,hwpoison: Remove stale code
In reply to Oscar SalvadorCurrently we call shake_page and then check whether the page is Buddy because shake_page calls drain_all_pages, which sends pcp-pages back to the buddy freelists, so we could have a chance to handle free pages.
get_hwpoison_page already calls drain_all_pages, and we do call get_hwpoison_page right before coming here, so we should be on the safe side.
--- mm/memory-failure.c | 12 ------------ 1 file changed, 12 deletions(-)
mm/memory-failure.c +0 -12
@@ -1423,18 +1423,6 @@ int memory_failure(unsigned long pfn, int flags)* walked by the page reclaim code, however that's not a big loss. */ shake_page (p,0 ); -/* shake_page could have turned it free. */ -if (!PageLRU (p) &&is_free_buddy_page (p)) { -if (!take_page_off_buddy (p)) - res = -EBUSY ; - -if (flags &MF_COUNT_INCREASED ) -action_result (pfn,MF_MSG_BUDDY , res ?MF_IGNORED :MF_RECOVERED ); -else -action_result (pfn,MF_MSG_BUDDY_2ND , res ?MF_IGNORED :MF_RECOVERED ); - -return res; - }lock_page (p);
Signature · 1 line
2.26.2
Quoted 12 lines · Oscar Salvador
The important bit of this patchset is patch#1, which is a fix to take off HWPoison pages off a buddy freelist since it can lead us to having HWPoison pages back in the game without no one noticing it.
So fix it (we did that already forsoft_offline_page[1]).The other patches are clean-ups and not that important, so if anything, consider patch#1 for inclusion.
[1] https://patchwork.kernel.org/cover/11704083/
Thanks
Quoted 7 lines
Oscar Salvador (5):
mm,hwpoison: Take free pages off the buddy freelists mm,hwpoison: Refactormadvise_inject_errormm,hwpoison: Drain pcplists before bailing out for non-buddy
zero-refcount page
mm,hwpoison: Drop unneeded pcplist draining mm,hwpoison: Remove stale code
Naoya, any insight?
Thanks