I'm not sure if I'm completely missing something here, but AFAIKS the
reference to the mysterious "COW SMC race" confuses the issue. The
original changelog and mailing list thread didn't help me either.
This SMC race is where the problem was detected, but isn't the general
problem bigger and more obvious: that the new PTE could be picked up at
any time by any TLB while entries for the old PTE exist in other TLBs
before the TLB flush takes effect?
The case where the iTLB and dTLB of a CPU are pointing at different pages
is an interesting one but follows from the general problem.
The other (minor) thing with the comment I think it makes it a bit clearer
to say what the old code was doing (i.e., it avoids the race as opposed to
what?).
References: 4ce072f1fa ("mm: fix a race condition under SMC + COW")
Link: https://lkml.kernel.org/r/20201215121119.351650-1-npiggin@gmail.com
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Suresh Siddha <sbsiddha@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Despite a comment that said that page fault accounting would be charged to
whatever task_struct* was passed into __access_remote_vm(), the tsk
argument was actually unused.
Making page fault accounting actually use this task struct is quite a
project, so there is no point in keeping the tsk argument.
Delete both the comment, and the argument.
[rppt@linux.ibm.com: changelog addition]
Link: https://lkml.kernel.org/r/20201026074137.4147787-1-jhubbard@nvidia.com
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Mike Rapoport <rppt@linux.ibm.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Since commit 70e806e4e6 ("mm: Do early cow for pinned pages during
fork() for ptes") pages under a FOLL_PIN will not be write protected
during COW for fork. This means that pages returned from
pin_user_pages(FOLL_WRITE) should not become write protected while the pin
is active.
However, there is a small race where get_user_pages_fast(FOLL_PIN) can
establish a FOLL_PIN at the same time copy_present_page() is write
protecting it:
CPU 0 CPU 1
get_user_pages_fast()
internal_get_user_pages_fast()
copy_page_range()
pte_alloc_map_lock()
copy_present_page()
atomic_read(has_pinned) == 0
page_maybe_dma_pinned() == false
atomic_set(has_pinned, 1);
gup_pgd_range()
gup_pte_range()
pte_t pte = gup_get_pte(ptep)
pte_access_permitted(pte)
try_grab_compound_head()
pte = pte_wrprotect(pte)
set_pte_at();
pte_unmap_unlock()
// GUP now returns with a write protected page
The first attempt to resolve this by using the write protect caused
problems (and was missing a barrrier), see commit f3c64eda3e ("mm: avoid
early COW write protect games during fork()")
Instead wrap copy_p4d_range() with the write side of a seqcount and check
the read side around gup_pgd_range(). If there is a collision then
get_user_pages_fast() fails and falls back to slow GUP.
Slow GUP is safe against this race because copy_page_range() is only
called while holding the exclusive side of the mmap_lock on the src
mm_struct.
[akpm@linux-foundation.org: coding style fixes]
Link: https://lore.kernel.org/r/CAHk-=wi=iCnYCARbPGjkVJu9eyYeZ13N64tZYLdOB8CP5Q_PLw@mail.gmail.com
Link: https://lkml.kernel.org/r/2-v4-908497cf359a+4782-gup_fork_jgg@nvidia.com
Fixes: f3c64eda3e ("mm: avoid early COW write protect games during fork()")
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Peter Xu <peterx@redhat.com>
Acked-by: "Ahmed S. Darwish" <a.darwish@linutronix.de> [seqcount_t parts]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jann Horn <jannh@google.com>
Cc: Kirill Shutemov <kirill@shutemov.name>
Cc: Kirill Tkhai <ktkhai@virtuozzo.com>
Cc: Leon Romanovsky <leonro@nvidia.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull dma-mapping updates from Christoph Hellwig:
- rework the non-coherent DMA allocator
- move private definitions out of <linux/dma-mapping.h>
- lower CMA_ALIGNMENT (Paul Cercueil)
- remove the omap1 dma address translation in favor of the common code
- make dma-direct aware of multiple dma offset ranges (Jim Quinlan)
- support per-node DMA CMA areas (Barry Song)
- increase the default seg boundary limit (Nicolin Chen)
- misc fixes (Robin Murphy, Thomas Tai, Xu Wang)
- various cleanups
* tag 'dma-mapping-5.10' of git://git.infradead.org/users/hch/dma-mapping: (63 commits)
ARM/ixp4xx: add a missing include of dma-map-ops.h
dma-direct: simplify the DMA_ATTR_NO_KERNEL_MAPPING handling
dma-direct: factor out a dma_direct_alloc_from_pool helper
dma-direct check for highmem pages in dma_direct_alloc_pages
dma-mapping: merge <linux/dma-noncoherent.h> into <linux/dma-map-ops.h>
dma-mapping: move large parts of <linux/dma-direct.h> to kernel/dma
dma-mapping: move dma-debug.h to kernel/dma/
dma-mapping: remove <asm/dma-contiguous.h>
dma-mapping: merge <linux/dma-contiguous.h> into <linux/dma-map-ops.h>
dma-contiguous: remove dma_contiguous_set_default
dma-contiguous: remove dev_set_cma_area
dma-contiguous: remove dma_declare_contiguous
dma-mapping: split <linux/dma-mapping.h>
cma: decrease CMA_ALIGNMENT lower limit to 2
firewire-ohci: use dma_alloc_pages
dma-iommu: implement ->alloc_noncoherent
dma-mapping: add new {alloc,free}_noncoherent dma_map_ops methods
dma-mapping: add a new dma_alloc_pages API
dma-mapping: remove dma_cache_sync
53c700: convert to dma_alloc_noncoherent
...
In commit 70e806e4e6 ("mm: Do early cow for pinned pages during fork()
for ptes") we write-protected the PTE before doing the page pinning
check, in order to avoid a race with concurrent fast-GUP pinning (which
doesn't take the mm semaphore or the page table lock).
That trick doesn't actually work - it doesn't handle memory ordering
properly, and doing so would be prohibitively expensive.
It also isn't really needed. While we're moving in the direction of
allowing and supporting page pinning without marking the pinned area
with MADV_DONTFORK, the fact is that we've never really supported this
kind of odd "concurrent fork() and page pinning", and doing the
serialization on a pte level is just wrong.
We can add serialization with a per-mm sequence counter, so we know how
to solve that race properly, but we'll do that at a more appropriate
time. Right now this just removes the write protect games.
It also turns out that the write protect games actually break on Power,
as reported by Aneesh Kumar:
"Architecture like ppc64 expects set_pte_at to be not used for updating
a valid pte. This is further explained in commit 56eecdb912 ("mm:
Use ptep/pmdp_set_numa() for updating _PAGE_NUMA bit")"
and the code triggered a warning there:
WARNING: CPU: 0 PID: 30613 at arch/powerpc/mm/pgtable.c:185 set_pte_at+0x2a8/0x3a0 arch/powerpc/mm/pgtable.c:185
Call Trace:
copy_present_page mm/memory.c:857 [inline]
copy_present_pte mm/memory.c:899 [inline]
copy_pte_range mm/memory.c:1014 [inline]
copy_pmd_range mm/memory.c:1092 [inline]
copy_pud_range mm/memory.c:1127 [inline]
copy_p4d_range mm/memory.c:1150 [inline]
copy_page_range+0x1f6c/0x2cc0 mm/memory.c:1212
dup_mmap kernel/fork.c:592 [inline]
dup_mm+0x77c/0xab0 kernel/fork.c:1355
copy_mm kernel/fork.c:1411 [inline]
copy_process+0x1f00/0x2740 kernel/fork.c:2070
_do_fork+0xc4/0x10b0 kernel/fork.c:2429
Link: https://lore.kernel.org/lkml/CAHk-=wiWr+gO0Ro4LvnJBMs90OiePNyrE3E+pJvc9PzdBShdmw@mail.gmail.com/
Link: https://lore.kernel.org/linuxppc-dev/20201008092541.398079-1-aneesh.kumar@linux.ibm.com/
Reported-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Tested-by: Leon Romanovsky <leonro@nvidia.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Jan Kara <jack@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Kirill Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Most of dma-debug.h is not required by anything outside of kernel/dma.
Move the four declarations needed by dma-mappin.h or dma-ops providers
into dma-mapping.h and dma-map-ops.h, and move the remainder of the
file to kernel/dma/debug.h.
Signed-off-by: Christoph Hellwig <hch@lst.de>
This allows copy_pte_range() to do early cow if the pages were pinned on
the source mm.
Currently we don't have an accurate way to know whether a page is pinned
or not. The only thing we have is page_maybe_dma_pinned(). However
that's good enough for now. Especially, with the newly added
mm->has_pinned flag to make sure we won't affect processes that never
pinned any pages.
It would be easier if we can do GFP_KERNEL allocation within
copy_one_pte(). Unluckily, we can't because we're with the page table
locks held for both the parent and child processes. So the page
allocation needs to be done outside copy_one_pte().
Some trick is there in copy_present_pte(), majorly the wrprotect trick
to block concurrent fast-gup. Comments in the function should explain
better in place.
Oleg Nesterov reported a (probably harmless) bug during review that we
didn't reset entry.val properly in copy_pte_range() so that potentially
there's chance to call add_swap_count_continuation() multiple times on
the same swp entry. However that should be harmless since even if it
happens, the same function (add_swap_count_continuation()) will return
directly noticing that there're enough space for the swp counter. So
instead of a standalone stable patch, it is touched up in this patch
directly.
Link: https://lore.kernel.org/lkml/20200914143829.GA1424636@nvidia.com/
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This prepares for the future work to trigger early cow on pinned pages
during fork().
No functional change intended.
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 09854ba94c ("mm: do_wp_page() simplification") reorganized all
the code around the page re-use vs copy, but in the process also moved
the final unlock_page() around to after the wp_page_reuse() call.
That normally doesn't matter - but it means that the unlock_page() is
now done after releasing the page table lock. Again, not a big deal,
you'd think.
But it turns out that it's very wrong indeed, because once we've
released the page table lock, we've basically lost our only reference to
the page - the page tables - and it could now be free'd at any time. We
do hold the mmap_sem, so no actual unmap() can happen, but madvise can
come in and a MADV_DONTNEED will zap the page range - and free the page.
So now the page may be free'd just as we're unlocking it, which in turn
will usually trigger a "Bad page state" error in the freeing path. To
make matters more confusing, by the time the debug code prints out the
page state, the unlock has typically completed and everything looks fine
again.
This all doesn't happen in any normal situations, but it does trigger
with the dirtyc0w_child LTP test. And it seems to trigger much more
easily (but not expclusively) on s390 than elsewhere, probably because
s390 doesn't do the "batch pages up for freeing after the TLB flush"
that gives the unlock_page() more time to complete and makes the race
harder to hit.
Fixes: 09854ba94c ("mm: do_wp_page() simplification")
Link: https://lore.kernel.org/lkml/a46e9bbef2ed4e17778f5615e818526ef848d791.camel@redhat.com/
Link: https://lore.kernel.org/linux-mm/c41149a8-211e-390b-af1d-d5eee690fecb@linux.alibaba.com/
Reported-by: Qian Cai <cai@redhat.com>
Reported-by: Alex Shi <alex.shi@linux.alibaba.com>
Bisected-and-analyzed-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Tested-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This completes the split of the non-present and present pte cases by
moving the check for the source pte being present into the single
caller, which also means that we clearly separate out the very different
return value case for a non-present pte.
The present pte case currently always succeeds.
This is a pure code re-organization with no semantic change: the intent
is to make it much easier to add a new return case to the present pte
case for when we do early COW at page table copy time.
This was split out from the previous commit simply to make it easy to
visually see that there were no semantic changes from this code
re-organization.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is a purely mechanical split of the copy_one_pte() function. It's
not immediately obvious when looking at the diff because of the
indentation change, but the way to see what is going on in this commit
is to use the "-w" flag to not show pure whitespace changes, and you see
how the first part of copy_one_pte() is simply lifted out into a
separate function.
And since the non-present case is marked unlikely, don't make the new
function be inlined. Not that gcc really seems to care, since it looks
like it will inline it anyway due to the whole "single callsite for
static function" logic. In fact, code generation with the function
split is almost identical to before. But not marking it inline is the
right thing to do.
This is pure prep-work and cleanup for subsequent changes.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Merge misc fixes from Andrew Morton:
"19 patches.
Subsystems affected by this patch series: MAINTAINERS, ipc, fork,
checkpatch, lib, and mm (memcg, slub, pagemap, madvise, migration,
hugetlb)"
* emailed patches from Andrew Morton <akpm@linux-foundation.org>:
include/linux/log2.h: add missing () around n in roundup_pow_of_two()
mm/khugepaged.c: fix khugepaged's request size in collapse_file
mm/hugetlb: fix a race between hugetlb sysctl handlers
mm/hugetlb: try preferred node first when alloc gigantic page from cma
mm/migrate: preserve soft dirty in remove_migration_pte()
mm/migrate: remove unnecessary is_zone_device_page() check
mm/rmap: fixup copying of soft dirty and uffd ptes
mm/migrate: fixup setting UFFD_WP flag
mm: madvise: fix vma user-after-free
checkpatch: fix the usage of capture group ( ... )
fork: adjust sysctl_max_threads definition to match prototype
ipc: adjust proc_ipc_sem_dointvec definition to match prototype
mm: track page table modifications in __apply_to_page_range()
MAINTAINERS: IA64: mark Status as Odd Fixes only
MAINTAINERS: add LLVM maintainers
MAINTAINERS: update Cavium/Marvell entries
mm: slub: fix conversion of freelist_corrupted()
mm: memcg: fix memcg reclaim soft lockup
memcg: fix use-after-free in uncharge_batch
Merge emailed patches from Peter Xu:
"This is a small series that I picked up from Linus's suggestion to
simplify cow handling (and also make it more strict) by checking
against page refcounts rather than mapcounts.
This makes uffd-wp work again (verified by running upmapsort)"
Note: this is horrendously bad timing, and making this kind of
fundamental vm change after -rc3 is not at all how things should work.
The saving grace is that it really is a a nice simplification:
8 files changed, 29 insertions(+), 120 deletions(-)
The reason for the bad timing is that it turns out that commit
17839856fd ("gup: document and work around 'COW can break either way'
issue" broke not just UFFD functionality (as Peter noticed), but Mikulas
Patocka also reports that it caused issues for strace when running in a
DAX environment with ext4 on a persistent memory setup.
And we can't just revert that commit without re-introducing the original
issue that is a potential security hole, so making COW stricter (and in
the process much simpler) is a step to then undoing the forced COW that
broke other uses.
Link: https://lore.kernel.org/lkml/alpine.LRH.2.02.2009031328040.6929@file01.intranet.prod.int.rdu2.redhat.com/
* emailed patches from Peter Xu <peterx@redhat.com>:
mm: Add PGREUSE counter
mm/gup: Remove enfornced COW mechanism
mm/ksm: Remove reuse_ksm_page()
mm: do_wp_page() simplification
How about we just make sure we're the only possible valid user fo the
page before we bother to reuse it?
Simplify, simplify, simplify.
And get rid of the nasty serialization on the page lock at the same time.
[peterx: add subject prefix]
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Recently we found regression when running will_it_scale/page_fault3 test
on ARM64. Over 70% down for the multi processes cases and over 20% down
for the multi threads cases. It turns out the regression is caused by
commit 89b15332af ("mm: drop mmap_sem before calling
balance_dirty_pages() in write fault").
The test mmaps a memory size file then write to the mapping, this would
make all memory dirty and trigger dirty pages throttle, that upstream
commit would release mmap_sem then retry the page fault. The retried
page fault would see correct PTEs installed then just fall through to
spurious TLB flush. The regression is caused by the excessive spurious
TLB flush. It is fine on x86 since x86's spurious TLB flush is no-op.
We could just skip the spurious TLB flush to mitigate the regression.
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reported-by: Xu Yu <xuyu@linux.alibaba.com>
Debugged-by: Xu Yu <xuyu@linux.alibaba.com>
Tested-by: Xu Yu <xuyu@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Yang Shi <shy828301@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Nicholas Piggin
Christoph Hellwig
John Hubbard
Jason Gunthorpe
Matthew Wilcox (Oracle)
Linus Torvalds
Peter Xu
Randy Dunlap
Yanfei Xu
Joerg Roedel
Yang Shi