On Thu, 11 Mar 2021 08:55:30 +0000
HORIGUCHI NAOYA(堀口 直也) <naoya.horiguchi@xxxxxxx> wrote:
> On Wed, Mar 10, 2021 at 02:10:42PM +0800, Aili Yao wrote:
> > On Fri, 5 Mar 2021 15:55:25 +0000
> > "Luck, Tony" <tony.luck@xxxxxxxxx> wrote:
> >
> > > > From the walk, it seems we have got the virtual address, can we just send a SIGBUS with it?
> > >
> > > If the walk wins the race and the pte for the poisoned page is still valid, then yes.
> > >
> > > But we could have:
> > >
> > > CPU1 CPU2
> > > memory_failure sets poison
> > > bit for struct page
> > >
> > >
> > > rmap finds page in task
> > > on CPU2 and sets PTE
> > > to not-valid-poison
> > >
> > > memory_failure returns
> > > early because struct page
> > > already marked as poison
> > >
> > > walk page tables looking
> > > for mapping - don't find it
> > >
> > > -Tony
> >
> > While I don't think there is a race condition, and if you really think the pfn with SIGBUS is not
> > proper, I think following patch maybe one way.
> > I copy your abandon code, and make a little modification, and just now it pass
> > my simple test.
> >
> > And also this is a RFC version, only valid if you think the pfn with SIGBUS is not right.
> >
> > Thanks!
> >
> > From a522ab8856e3a332a2318d57bb19f3c59594d462 Mon Sep 17 00:00:00 2001
> > From: Aili Yao <yaoaili@xxxxxxxxxxxx>
> > Date: Wed, 10 Mar 2021 13:59:18 +0800
> > Subject: [PATCH] x86/mce: fix invalid SIGBUS address
> >
> > walk the current process pte and compare with the pfn;
> > 1. only test for normal page and 2M hugetlb page;
> > 2. 1G hugetlb and transparentHuge is not support currently;
> > 3. May other fails is not recognized, This is a RFC version.
> >
> > ---
> > arch/x86/kernel/cpu/mce/core.c | 83 ++++++++++++++++++++++++++++++++++++++++--
> > 1 file changed, 80 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
> > index db4afc5..65d7ef7 100644
> > --- a/arch/x86/kernel/cpu/mce/core.c
> > +++ b/arch/x86/kernel/cpu/mce/core.c
> > @@ -28,8 +28,12 @@
> > #include <linux/sysfs.h>
> > #include <linux/types.h>
> > #include <linux/slab.h>
> > +#include <linux/hugetlb.h>
> > +#include <linux/swap.h>
> > +#include <linux/swapops.h>
> > #include <linux/init.h>
> > #include <linux/kmod.h>
> > +#include <linux/pagewalk.h>
> > #include <linux/poll.h>
> > #include <linux/nmi.h>
> > #include <linux/cpu.h>
>
> Maybe requiring many dependencies like this implies that you might be better
> to do below in mm/memory-failure.c instead of in this file.
Yes, agree, I will change this, Thanks!
> > @@ -1235,6 +1239,81 @@ static void __mc_scan_banks(struct mce *m, struct pt_regs *regs, struct mce *fin
> > /* mce_clear_state will clear *final, save locally for use later */
> > *m = *final;
> > }
> > +static int mc_pte_entry(pte_t *pte, unsigned long addr, unsigned long next, struct mm_walk *walk)
> > +{
> > + u64 *buff = (u64 *)walk->private;
> > + u64 pfn = buff[0];
> > +
> > + if (!pte_present(*pte) && is_hwpoison_entry(pte_to_swp_entry(*pte)))
> > + goto find;
> > + else if (pte_pfn(*pte) == pfn)
> > + goto find;
> > +
> > + return 0;
> > +find:
> > + buff[0] = addr;
> > + buff[1] = PAGE_SHIFT;
> > + return true;
>
> Returning true means you stop walking when you find the first entry pointing
> to a given pfn. But there could be multiple such entries, so if MCE SRAR is
> triggered by memory access to the larger address in hwpoisoned entries, the
> returned virtual address might be wrong.
Yes, We need to consider multiple posion page entries, I will fix this. Thanks for
you sugguestion!
> > +}
> > +
> > +extern bool is_hugetlb_entry_hwpoisoned(pte_t pte);
> > +
> > +static int mc_hugetlb_range(pte_t *ptep, unsigned long hmask,
> > + unsigned long addr, unsigned long end,
> > + struct mm_walk *walk)
> > +{
> > + u64 *buff = (u64 *)walk->private;
> > + u64 pfn = buff[0];
> > + int shift = PMD_SHIFT;
> > + pte_t pte = huge_ptep_get(ptep);
> > +
> > + if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
> > + goto find;
> > +
> > + if (pte_pfn(*ptep) == pfn)
> > + goto find;
> > +
> > + return 0;
> > +find:
> > + buff[0] = addr;
> > + buff[1] = shift;
> > + return true;
> > +}
> > +
> > +static struct mm_walk_ops walk = {
> > + .pte_entry = mc_pte_entry,
> > + .hugetlb_entry = mc_hugetlb_range
> > +};
> > +
> > +void mc_memory_failure_error(struct task_struct *p, unsigned long pfn)
> > +{
> > + u64 buff[2] = {pfn, 0};
> > + struct page *page;
> > + int ret = -1;
> > +
> > + page = pfn_to_page(pfn);
> > + if (!page)
> > + goto force_sigbus;
> > +
> > + if (is_zone_device_page(page))
> > + goto force_sigbus;
> > +
> > + mmap_read_lock(p->mm);
> > + ret = walk_page_range(p->mm, 0, TASK_SIZE_MAX, &walk, (void *)buff);
> > + mmap_read_unlock(p->mm);
> > +
> > + if (ret && buff[0]) {
> > + pr_err("Memory error may not recovered: %#llx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
> > + buff[0], p->comm, p->pid);
> > + force_sig_mceerr(BUS_MCEERR_AR, (void __user *)buff[0], buff[1]);
> > + } else {
> > +force_sigbus:
> > + pr_err("Memory error may not recovered, pfn: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
> > + pfn, p->comm, p->pid);
> > + force_sig_mceerr(BUS_MCEERR_AR, (void __user *)pfn, PAGE_SHIFT);
> > + }
> > +
> > +}
> >
> > static void kill_me_now(struct callback_head *ch)
> > {
> > @@ -1259,9 +1338,7 @@ static void kill_me_maybe(struct callback_head *cb)
> > }
> >
> > if (p->mce_vaddr != (void __user *)-1l) {
> > - pr_err("Memory error may not recovered: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
> > - p->mce_addr >> PAGE_SHIFT, p->comm, p->pid);
> > - force_sig_mceerr(BUS_MCEERR_AR, p->mce_vaddr, PAGE_SHIFT);
> > + mc_memory_failure_error(current, p->mce_addr >> PAGE_SHIFT);
>
> I guess that p->mce_vaddr stores the virtual address of the error here.
> If so, sending SIGBUS with the address looks enough as we do now, so why
> do you walk page table to find the error virtual address?
I check the code again, yes, I should have placed mc_memory_failure_error in else branch, but it seems p->mce_vaddr is not correctly
initialized and for my test, it has a zero value then code goes into if (p->mce_vaddr != (void __user *)-1l) branch; It seems this is
another issue needing to fix;
And from the p->mce_vaddr, possibly there is a better way to get vaddr, i will dig more about this.
--
Thanks!
Aili Yao
