On Tue, 2 Mar 2021 19:39:53 -0800 "Luck, Tony" <tony.luck@xxxxxxxxx> wrote: > On Fri, Feb 26, 2021 at 10:59:15AM +0800, Aili Yao wrote: > > Hi naoya, tony: > > > > > > > > Idea for what we should do next ... Now that x86 is calling memory_failure() > > > > from user context ... maybe parallel calls for the same page should > > > > be blocked until the first caller completes so we can: > > > > a) know that pages are unmapped (if that happens) > > > > b) all get the same success/fail status > > > > > > One memory_failure() call changes the target page's status and > > > affects all mappings to all affected processes, so I think that > > > (ideally) we don't have to block other threads (letting them > > > early return seems fine). Sometimes memory_failure() fails, > > > but even in such case, PG_hwpoison is set on the page and other > > > threads properly get SIGBUSs with this patch, so I think that > > > we can avoid the worst scenario (like system stall by MCE loop). > > > > > I agree with naoya's point, if we block for this issue, Does this change the result > > that the process should be killed? Or is there something other still need to be considered? > > Ok ... no blocking ... I think someone in this thread suggested > scanning the page tables to make sure the poisoned page had been > unmapped. > > There's a walk_page_range() function that does all the work for that. > Just need to supply some callback routines that check whether a > mapping includes the bad PFN and clear the PRESENT bit. > > RFC patch below against v5.12-rc1 > > -Tony > > From 8de23b7f1be00ad38e129690dfe0b1558fad5ff8 Mon Sep 17 00:00:00 2001 > From: Tony Luck <tony.luck@xxxxxxxxx> > Date: Tue, 2 Mar 2021 15:06:33 -0800 > Subject: [PATCH] x86/mce: Handle races between machine checks > > When multiple CPUs hit the same poison memory there is a race. The > first CPU into memory_failure() atomically marks the page as poison > and continues processing to hunt down all the tasks that map this page > so that the virtual addresses can be marked not-present and SIGBUS > sent to the task that did the access. > > Later CPUs get an early return from memory_failure() and may return > to user mode and access the poison again. > > Add a new argument to memory_failure() so that it can indicate when > the race has been lost. Fix kill_me_maybe() to scan page tables in > this case to unmap pages. > + > static void kill_me_now(struct callback_head *ch) > { > force_sig(SIGBUS); > @@ -1257,15 +1304,19 @@ static void kill_me_maybe(struct callback_head *cb) > { > struct task_struct *p = container_of(cb, struct task_struct, mce_kill_me); > int flags = MF_ACTION_REQUIRED; > + int already = 0; > > pr_err("Uncorrected hardware memory error in user-access at %llx", p->mce_addr); > > if (!p->mce_ripv) > flags |= MF_MUST_KILL; > > - if (!memory_failure(p->mce_addr >> PAGE_SHIFT, flags) && > + if (!memory_failure(p->mce_addr >> PAGE_SHIFT, flags, &already) && > !(p->mce_kflags & MCE_IN_KERNEL_COPYIN)) { > - set_mce_nospec(p->mce_addr >> PAGE_SHIFT, p->mce_whole_page); > + if (already) > + walk_page_range(current->mm, 0, TASK_SIZE_MAX, &walk, (void *)(p->mce_addr >> PAGE_SHIFT)); > + else > + set_mce_nospec(p->mce_addr >> PAGE_SHIFT, p->mce_whole_page); > sync_core(); > return; > MF_MUST_KILL = 1 << 2, > MF_SOFT_OFFLINE = 1 << 3, > }; I have one doubt here, when "walk_page_range(current->mm, 0, TASK_SIZE_MAX, &walk, (void *)(p->mce_addr >> PAGE_SHIFT));" is done, so how is the process triggering this error returned if it have taken the wrong data? -- Thanks! Aili Yao