mm/ksm.c: ignore STABLE_FLAG of rmap_item->address in rmap_walk_ksm()

commit 1105a2f upstream.

In our armv8a server(QDF2400), I noticed lots of WARN_ON caused by
PAGE_SIZE unaligned for rmap_item->address under memory pressure
tests(start 20 guests and run memhog in the host).

  WARNING: CPU: 4 PID: 4641 at virt/kvm/arm/mmu.c:1826 kvm_age_hva_handler+0xc0/0xc8
  CPU: 4 PID: 4641 Comm: memhog Tainted: G        W 4.17.0-rc3+ Nuvoton-Israel#8
  Call trace:
   kvm_age_hva_handler+0xc0/0xc8
   handle_hva_to_gpa+0xa8/0xe0
   kvm_age_hva+0x4c/0xe8
   kvm_mmu_notifier_clear_flush_young+0x54/0x98
   __mmu_notifier_clear_flush_young+0x6c/0xa0
   page_referenced_one+0x154/0x1d8
   rmap_walk_ksm+0x12c/0x1d0
   rmap_walk+0x94/0xa0
   page_referenced+0x194/0x1b0
   shrink_page_list+0x674/0xc28
   shrink_inactive_list+0x26c/0x5b8
   shrink_node_memcg+0x35c/0x620
   shrink_node+0x100/0x430
   do_try_to_free_pages+0xe0/0x3a8
   try_to_free_pages+0xe4/0x230
   __alloc_pages_nodemask+0x564/0xdc0
   alloc_pages_vma+0x90/0x228
   do_anonymous_page+0xc8/0x4d0
   __handle_mm_fault+0x4a0/0x508
   handle_mm_fault+0xf8/0x1b0
   do_page_fault+0x218/0x4b8
   do_translation_fault+0x90/0xa0
   do_mem_abort+0x68/0xf0
   el0_da+0x24/0x28

In rmap_walk_ksm, the rmap_item->address might still have the
STABLE_FLAG, then the start and end in handle_hva_to_gpa might not be
PAGE_SIZE aligned.  Thus it will cause exceptions in handle_hva_to_gpa
on arm64.

This patch fixes it by ignoring (not removing) the low bits of address
when doing rmap_walk_ksm.

IMO, it should be backported to stable tree.  the storm of WARN_ONs is
very easy for me to reproduce.  More than that, I watched a panic (not
reproducible) as follows:

  page:ffff7fe003742d80 count:-4871 mapcount:-2126053375 mapping: (null) index:0x0
  flags: 0x1fffc00000000000()
  raw: 1fffc00000000000 0000000000000000 0000000000000000 ffffecf981470000
  raw: dead000000000100 dead000000000200 ffff8017c001c000 0000000000000000
  page dumped because: nonzero _refcount
  CPU: 29 PID: 18323 Comm: qemu-kvm Tainted: G W 4.14.15-5.hxt.aarch64 Nuvoton-Israel#1
  Hardware name: <snip for confidential issues>
  Call trace:
    dump_backtrace+0x0/0x22c
    show_stack+0x24/0x2c
    dump_stack+0x8c/0xb0
    bad_page+0xf4/0x154
    free_pages_check_bad+0x90/0x9c
    free_pcppages_bulk+0x464/0x518
    free_hot_cold_page+0x22c/0x300
    __put_page+0x54/0x60
    unmap_stage2_range+0x170/0x2b4
    kvm_unmap_hva_handler+0x30/0x40
    handle_hva_to_gpa+0xb0/0xec
    kvm_unmap_hva_range+0x5c/0xd0

I even injected a fault on purpose in kvm_unmap_hva_range by seting
size=size-0x200, the call trace is similar as above.  So I thought the
panic is similarly caused by the root cause of WARN_ON.

Andrea said:

: It looks a straightforward safe fix, on x86 hva_to_gfn_memslot would
: zap those bits and hide the misalignment caused by the low metadata
: bits being erroneously left set in the address, but the arm code
: notices when that's the last page in the memslot and the hva_end is
: getting aligned and the size is below one page.
:
: I think the problem triggers in the addr += PAGE_SIZE of
: unmap_stage2_ptes that never matches end because end is aligned but
: addr is not.
:
: 	} while (pte++, addr += PAGE_SIZE, addr != end);
:
: x86 again only works on hva_start/hva_end after converting it to
: gfn_start/end and that being in pfn units the bits are zapped before
: they risk to cause trouble.

Jia He said:

: I've tested by myself in arm64 server (QDF2400,46 cpus,96G mem) Without
: this patch, the WARN_ON is very easy for reproducing.  After this patch, I
: have run the same benchmarch for a whole day without any WARN_ONs

Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Jia He <[email protected]>
Reviewed-by: Andrea Arcangeli <[email protected]>
Tested-by: Jia He <[email protected]>
Cc: Suzuki K Poulose <[email protected]>
Cc: Minchan Kim <[email protected]>
Cc: Claudio Imbrenda <[email protected]>
Cc: Arvind Yadav <[email protected]>
Cc: Mike Rapoport <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>

Author: Jia He

mm/ksm.c

@@ -199,6 +199,8 @@ struct rmap_item {
 #define SEQNR_MASK	0x0ff	/* low bits of unstable tree seqnr */
 #define UNSTABLE_FLAG	0x100	/* is a node of the unstable tree */
 #define STABLE_FLAG	0x200	/* is listed from the stable tree */
+#define KSM_FLAG_MASK	(SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
+				/* to mask all the flags */
 
 /* The stable and unstable tree heads */
 static struct rb_root one_stable_tree[1] = { RB_ROOT };
@@ -2570,10 +2572,15 @@ void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
 		anon_vma_lock_read(anon_vma);
 		anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
 					       0, ULONG_MAX) {
+			unsigned long addr;
+
 			cond_resched();
 			vma = vmac->vma;
-			if (rmap_item->address < vma->vm_start ||
-			    rmap_item->address >= vma->vm_end)
+
+			/* Ignore the stable/unstable/sqnr flags */
+			addr = rmap_item->address & ~KSM_FLAG_MASK;
+
+			if (addr < vma->vm_start || addr >= vma->vm_end)
 				continue;
 			/*
 			 * Initially we examine only the vma which covers this
@@ -2587,8 +2594,7 @@ void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
 			if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
 				continue;
 
-			if (!rwc->rmap_one(page, vma,
-					rmap_item->address, rwc->arg)) {
+			if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
 				anon_vma_unlock_read(anon_vma);
 				return;
 			}