linux-toradex.git/mm/memory.c, branch v4.4.85

mm, mprotect: flush TLB if potentially racing with a parallel reclaim leaving stale TLB entries

2017-08-11T16:08:50+00:00

commit 3ea277194daaeaa84ce75180ec7c7a2075027a68 upstream.

Stable note for 4.4: The upstream patch patches madvise(MADV_FREE) but 4.4
	does not have support for that feature. The changelog is left
	as-is but the hunk related to madvise is omitted from the backport.

Nadav Amit identified a theoritical race between page reclaim and
mprotect due to TLB flushes being batched outside of the PTL being held.

He described the race as follows:

        CPU0                            CPU1
        ----                            ----
                                        user accesses memory using RW PTE
                                        [PTE now cached in TLB]
        try_to_unmap_one()
        ==> ptep_get_and_clear()
        ==> set_tlb_ubc_flush_pending()
                                        mprotect(addr, PROT_READ)
                                        ==> change_pte_range()
                                        ==> [ PTE non-present - no flush ]

                                        user writes using cached RW PTE
        ...

        try_to_unmap_flush()

The same type of race exists for reads when protecting for PROT_NONE and
also exists for operations that can leave an old TLB entry behind such
as munmap, mremap and madvise.

For some operations like mprotect, it's not necessarily a data integrity
issue but it is a correctness issue as there is a window where an
mprotect that limits access still allows access.  For munmap, it's
potentially a data integrity issue although the race is massive as an
munmap, mmap and return to userspace must all complete between the
window when reclaim drops the PTL and flushes the TLB.  However, it's
theoritically possible so handle this issue by flushing the mm if
reclaim is potentially currently batching TLB flushes.

Other instances where a flush is required for a present pte should be ok
as either the page lock is held preventing parallel reclaim or a page
reference count is elevated preventing a parallel free leading to
corruption.  In the case of page_mkclean there isn't an obvious path
that userspace could take advantage of without using the operations that
are guarded by this patch.  Other users such as gup as a race with
reclaim looks just at PTEs.  huge page variants should be ok as they
don't race with reclaim.  mincore only looks at PTEs.  userfault also
should be ok as if a parallel reclaim takes place, it will either fault
the page back in or read some of the data before the flush occurs
triggering a fault.

Note that a variant of this patch was acked by Andy Lutomirski but this
was for the x86 parts on top of his PCID work which didn't make the 4.13
merge window as expected.  His ack is dropped from this version and
there will be a follow-on patch on top of PCID that will include his
ack.

[akpm@linux-foundation.org: tweak comments]
[akpm@linux-foundation.org: fix spello]
Link: http://lkml.kernel.org/r/20170717155523.emckq2esjro6hf3z@suse.de
Reported-by: Nadav Amit 
Signed-off-by: Mel Gorman 
Cc: Andy Lutomirski 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds 
Signed-off-by: Greg Kroah-Hartman

mm: larger stack guard gap, between vmas

2017-06-26T05:13:11+00:00

commit 1be7107fbe18eed3e319a6c3e83c78254b693acb upstream.

Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.

This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.

Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.

One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications.  For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).

Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.

Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.

Original-patch-by: Oleg Nesterov 
Original-patch-by: Michal Hocko 
Signed-off-by: Hugh Dickins 
Acked-by: Michal Hocko 
Tested-by: Helge Deller  # parisc
Signed-off-by: Linus Torvalds 
[wt: backport to 4.11: adjust context]
[wt: backport to 4.9: adjust context ; kernel doc was not in admin-guide]
[wt: backport to 4.4: adjust context ; drop ppc hugetlb_radix changes]
Signed-off-by: Willy Tarreau 
[gkh: minor build fixes for 4.4]
Signed-off-by: Greg Kroah-Hartman

numa: fix /proc//numa_maps for THP

2016-05-04T21:48:49+00:00

commit 28093f9f34cedeaea0f481c58446d9dac6dd620f upstream.

In gather_pte_stats() a THP pmd is cast into a pte, which is wrong
because the layouts may differ depending on the architecture.  On s390
this will lead to inaccurate numa_maps accounting in /proc because of
misguided pte_present() and pte_dirty() checks on the fake pte.

On other architectures pte_present() and pte_dirty() may work by chance,
but there may be an issue with direct-access (dax) mappings w/o
underlying struct pages when HAVE_PTE_SPECIAL is set and THP is
available.  In vm_normal_page() the fake pte will be checked with
pte_special() and because there is no "special" bit in a pmd, this will
always return false and the VM_PFNMAP | VM_MIXEDMAP checking will be
skipped.  On dax mappings w/o struct pages, an invalid struct page
pointer would then be returned that can crash the kernel.

This patch fixes the numa_maps THP handling by introducing new "_pmd"
variants of the can_gather_numa_stats() and vm_normal_page() functions.

Signed-off-by: Gerald Schaefer 
Cc: Naoya Horiguchi 
Cc: "Kirill A . Shutemov" 
Cc: Konstantin Khlebnikov 
Cc: Michal Hocko 
Cc: Vlastimil Babka 
Cc: Jerome Marchand 
Cc: Johannes Weiner 
Cc: Dave Hansen 
Cc: Mel Gorman 
Cc: Dan Williams 
Cc: Martin Schwidefsky 
Cc: Heiko Carstens 
Cc: Michael Holzheu 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds 
Signed-off-by: Greg Kroah-Hartman

mm: thp: fix SMP race condition between THP page fault and MADV_DONTNEED

2016-03-03T23:07:10+00:00

commit ad33bb04b2a6cee6c1f99fabb15cddbf93ff0433 upstream.

pmd_trans_unstable()/pmd_none_or_trans_huge_or_clear_bad() were
introduced to locklessy (but atomically) detect when a pmd is a regular
(stable) pmd or when the pmd is unstable and can infinitely transition
from pmd_none() and pmd_trans_huge() from under us, while only holding
the mmap_sem for reading (for writing not).

While holding the mmap_sem only for reading, MADV_DONTNEED can run from
under us and so before we can assume the pmd to be a regular stable pmd
we need to compare it against pmd_none() and pmd_trans_huge() in an
atomic way, with pmd_trans_unstable().  The old pmd_trans_huge() left a
tiny window for a race.

Useful applications are unlikely to notice the difference as doing
MADV_DONTNEED concurrently with a page fault would lead to undefined
behavior.

[akpm@linux-foundation.org: tidy up comment grammar/layout]
Signed-off-by: Andrea Arcangeli 
Reported-by: Kirill A. Shutemov 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds 
Signed-off-by: Greg Kroah-Hartman

mm, dax: fix DAX deadlocks (COW fault)

2015-11-19T00:54:36+00:00

DAX handling of COW faults has wrong locking sequence:
	dax_fault does i_mmap_lock_read
	do_cow_fault does i_mmap_unlock_write

Ross's commit[1] missed a fix[2] that Kirill added to Matthew's
commit[3].

Original COW locking logic was introduced by Matthew here[4].

This should be applied to v4.3 as well.

[1] 0f90cc6609c7 mm, dax: fix DAX deadlocks
[2] 52a2b53ffde6 mm, dax: use i_mmap_unlock_write() in do_cow_fault()
[3] 843172978bb9 dax: fix race between simultaneous faults
[4] 2e4cdab0584f mm: allow page fault handlers to perform the COW

Cc: 
Cc: Boaz Harrosh 
Cc: Alexander Viro 
Cc: Dave Chinner 
Cc: Jan Kara 
Cc: "Kirill A. Shutemov" 
Cc: Matthew Wilcox 
Acked-by: Ross Zwisler 
Signed-off-by: Yigal Korman 
Signed-off-by: Dan Williams

mm, dax: fix DAX deadlocks

2015-10-16T18:42:28+00:00

The following two locking commits in the DAX code:

commit 843172978bb9 ("dax: fix race between simultaneous faults")
commit 46c043ede471 ("mm: take i_mmap_lock in unmap_mapping_range() for DAX")

introduced a number of deadlocks and other issues which need to be fixed
for the v4.3 kernel.  The list of issues in DAX after these commits
(some newly introduced by the commits, some preexisting) can be found
here:

  https://lkml.org/lkml/2015/9/25/602 (Subject: "Re: [PATCH] dax: fix deadlock in __dax_fault").

This undoes most of the changes introduced by those two commits,
essentially returning us to the DAX locking scheme that was used in
v4.2.

Signed-off-by: Ross Zwisler 
Cc: Alexander Viro 
Cc: Dan Williams 
Tested-by: Dave Chinner 
Cc: Jan Kara 
Cc: "Kirill A. Shutemov" 
Cc: Matthew Wilcox 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

mm: use vma_is_anonymous() in create_huge_pmd() and wp_huge_pmd()

2015-09-10T20:29:01+00:00

Let's use helper rather than direct check of vma->vm_ops to distinguish
anonymous VMA.

Signed-off-by: Kirill A. Shutemov 
Reviewed-by: Oleg Nesterov 
Cc: "H. Peter Anvin" 
Cc: Andy Lutomirski 
Cc: Dave Hansen 
Cc: Ingo Molnar 
Cc: Minchan Kim 
Cc: Thomas Gleixner 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

mm, dax: use i_mmap_unlock_write() in do_cow_fault()

2015-09-08T22:35:28+00:00

__dax_fault() takes i_mmap_lock for write. Let's pair it with write
unlock on do_cow_fault() side.

Signed-off-by: Kirill A. Shutemov 
Acked-by: Matthew Wilcox 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

mm: take i_mmap_lock in unmap_mapping_range() for DAX

2015-09-08T22:35:28+00:00

DAX is not so special: we need i_mmap_lock to protect mapping->i_mmap.

__dax_pmd_fault() uses unmap_mapping_range() shoot out zero page from
all mappings.  We need to drop i_mmap_lock there to avoid lock deadlock.

Re-aquiring the lock should be fine since we check i_size after the
point.

Signed-off-by: Kirill A. Shutemov 
Cc: Matthew Wilcox 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

dax: fix race between simultaneous faults

2015-09-08T22:35:28+00:00

If two threads write-fault on the same hole at the same time, the winner
of the race will return to userspace and complete their store, only to
have the loser overwrite their store with zeroes.  Fix this for now by
taking the i_mmap_sem for write instead of read, and do so outside the
call to get_block().  Now the loser of the race will see the block has
already been zeroed, and will not zero it again.

This severely limits our scalability.  I have ideas for improving it, but
those can wait for a later patch.

Signed-off-by: Matthew Wilcox 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds