With the 3.2-rc kernel, IOMMU 2M pages in KVM works. But when I tried
to use IOMMU 1GB pages in KVM, I encountered an oops and the 1GB page
failed to be used.
The root cause is that 1GB page allocation calls gup_huge_pud() while 2M
page calls gup_huge_pmd. If compound pages are used and the page is a
tail page, gup_huge_pmd() increases _mapcount to record tail page are
mapped while gup_huge_pud does not do that.
So when the mapped page is relesed, it will result in kernel oops
because the page is not marked mapped.
This patch add tail process for compound page in 1GB huge page which
keeps the same process as 2M page.
Reproduce like:
1. Add grub boot option: hugepagesz=1G hugepages=8
2. mount -t hugetlbfs -o pagesize=1G hugetlbfs /dev/hugepages
3. qemu-kvm -m 2048 -hda os-kvm.img -cpu kvm64 -smp 4 -mem-path /dev/hugepages
-net none -device pci-assign,host=07:00.1
kernel BUG at mm/swap.c:114!
invalid opcode: 0000 [#1] SMP
Call Trace:
put_page+0x15/0x37
kvm_release_pfn_clean+0x31/0x36
kvm_iommu_put_pages+0x94/0xb1
kvm_iommu_unmap_memslots+0x80/0xb6
kvm_assign_device+0xba/0x117
kvm_vm_ioctl_assigned_device+0x301/0xa47
kvm_vm_ioctl+0x36c/0x3a2
do_vfs_ioctl+0x49e/0x4e4
sys_ioctl+0x5a/0x7c
system_call_fastpath+0x16/0x1b
RIP put_compound_page+0xd4/0x168
Signed-off-by: Youquan Song <youquan.song@intel.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This avoids duplicating the function in every arch gup_fast.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: David Miller <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Michel while working on the working set estimation code, noticed that
calling get_page_unless_zero() on a random pfn_to_page(random_pfn)
wasn't safe, if the pfn ended up being a tail page of a transparent
hugepage under splitting by __split_huge_page_refcount().
He then found the problem could also theoretically materialize with
page_cache_get_speculative() during the speculative radix tree lookups
that uses get_page_unless_zero() in SMP if the radix tree page is freed
and reallocated and get_user_pages is called on it before
page_cache_get_speculative has a chance to call get_page_unless_zero().
So the best way to fix the problem is to keep page_tail->_count zero at
all times. This will guarantee that get_page_unless_zero() can never
succeed on any tail page. page_tail->_mapcount is guaranteed zero and
is unused for all tail pages of a compound page, so we can simply
account the tail page references there and transfer them to
tail_page->_count in __split_huge_page_refcount() (in addition to the
head_page->_mapcount).
While debugging this s/_count/_mapcount/ change I also noticed get_page is
called by direct-io.c on pages returned by get_user_pages. That wasn't
entirely safe because the two atomic_inc in get_page weren't atomic. As
opposed to other get_user_page users like secondary-MMU page fault to
establish the shadow pagetables would never call any superflous get_page
after get_user_page returns. It's safer to make get_page universally safe
for tail pages and to use get_page_foll() within follow_page (inside
get_user_pages()). get_page_foll() is safe to do the refcounting for tail
pages without taking any locks because it is run within PT lock protected
critical sections (PT lock for pte and page_table_lock for
pmd_trans_huge).
The standard get_page() as invoked by direct-io instead will now take
the compound_lock but still only for tail pages. The direct-io paths
are usually I/O bound and the compound_lock is per THP so very
finegrined, so there's no risk of scalability issues with it. A simple
direct-io benchmarks with all lockdep prove locking and spinlock
debugging infrastructure enabled shows identical performance and no
overhead. So it's worth it. Ideally direct-io should stop calling
get_page() on pages returned by get_user_pages(). The spinlock in
get_page() is already optimized away for no-THP builds but doing
get_page() on tail pages returned by GUP is generally a rare operation
and usually only run in I/O paths.
This new refcounting on page_tail->_mapcount in addition to avoiding new
RCU critical sections will also allow the working set estimation code to
work without any further complexity associated to the tail page
refcounting with THP.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Michel Lespinasse <walken@google.com>
Reviewed-by: Michel Lespinasse <walken@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: <stable@kernel.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
For GRU and EPT, we need gup-fast to set referenced bit too (this is why
it's correct to return 0 when shadow_access_mask is zero, it requires
gup-fast to set the referenced bit). qemu-kvm access already sets the
young bit in the pte if it isn't zero-copy, if it's zero copy or a shadow
paging EPT minor fault we relay on gup-fast to signal the page is in
use...
We also need to check the young bits on the secondary pagetables for NPT
and not nested shadow mmu as the data may never get accessed again by the
primary pte.
Without this closer accuracy, we'd have to remove the heuristic that
avoids collapsing hugepages in hugepage virtual regions that have not even
a single subpage in use.
->test_young is full backwards compatible with GRU and other usages that
don't have young bits in pagetables set by the hardware and that should
nuke the secondary mmu mappings when ->clear_flush_young runs just like
EPT does.
Removing the heuristic that checks the young bit in
khugepaged/collapse_huge_page completely isn't so bad either probably but
I thought it was worth it and this makes it reliable.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Force gup_fast to take the slow path and block if the pmd is splitting,
not only if it's none.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Alter compound get_page/put_page to keep references on subpages too, in
order to allow __split_huge_page_refcount to split an hugepage even while
subpages have been pinned by one of the get_user_pages() variants.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Fix minor spelling error in comment. No code change.
Signed-off-by: Andy Shevchenko <ext-andriy.shevchenko@nokia.com>
LKML-Reference: <201002022238.o12McDiF018720@imap1.linux-foundation.org>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
* 'perfcounters-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (49 commits)
perfcounter: Handle some IO return values
perf_counter: Push perf_sample_data through the swcounter code
perf_counter tools: Define and use our own u64, s64 etc. definitions
perf_counter: Close race in perf_lock_task_context()
perf_counter, x86: Improve interactions with fast-gup
perf_counter: Simplify and fix task migration counting
perf_counter tools: Add a data file header
perf_counter: Update userspace callchain sampling uses
perf_counter: Make callchain samples extensible
perf report: Filter to parent set by default
perf_counter tools: Handle lost events
perf_counter: Add event overlow handling
fs: Provide empty .set_page_dirty() aop for anon inodes
perf_counter: tools: Makefile tweaks for 64-bit powerpc
perf_counter: powerpc: Add processor back-end for MPC7450 family
perf_counter: powerpc: Make powerpc perf_counter code safe for 32-bit kernels
perf_counter: powerpc: Change how processor-specific back-ends get selected
perf_counter: powerpc: Use unsigned long for register and constraint values
perf_counter: powerpc: Enable use of software counters on 32-bit powerpc
perf_counter tools: Add and use isprint()
...
It's really not right to use 'access_ok()', since that is meant for the
normal "get_user()" and "copy_from/to_user()" accesses, which are done
through the TLB, rather than through the page tables.
Why? access_ok() does both too few, and too many checks. Too many,
because it is meant for regular kernel accesses that will not honor the
'user' bit in the page tables, and because it honors the USER_DS vs
KERNEL_DS distinction that we shouldn't care about in GUP. And too few,
because it doesn't do the 'canonical' check on the address on x86-64,
since the TLB will do that for us.
So instead of using a function that isn't meant for this, and does
something else and much more complicated, just do the real rules: we
don't want the range to overflow, and on x86-64, we want it to be a
canonical low address (on 32-bit, all addresses are canonical).
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Improve a few details in perfcounter call-chain recording that
makes use of fast-GUP:
- Use ACCESS_ONCE() to observe the pte value. ptes are fundamentally
racy and can be changed on another CPU, so we have to be careful
about how we access them. The PAE branch is already careful with
read-barriers - but the non-PAE and 64-bit side needs an
ACCESS_ONCE() to make sure the pte value is observed only once.
- make the checks a bit stricter so that we can feed it any kind of
cra^H^H^H user-space input ;-)
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Introduce a gup_fast() variant which is usable from IRQ/NMI context.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
CC: Nick Piggin <npiggin@suse.de>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
While better than get_user_pages(), the usage of gupf(),
especially the return values and the fact that it can
potentially only partially pin the range, warranted some
documentation.
Signed-off-by: Andy Grover <andy.grover@oracle.com>
Cc: npiggin@suse.de
Cc: akpm@linux-foundation.org
LKML-Reference: <1239320729-3262-1-git-send-email-andy.grover@oracle.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Since pte_flags() is much cheaper than pte_val() in some virtualized
environments (namely, Xen), use the former whereever possible.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Cc: "Nick Piggin" <npiggin@suse.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Alexey Dobriyan reported trouble with LTP with the new fast-gup code,
and Johannes Weiner debugged it to non-page-aligned addresses, where the
new get_user_pages_fast() code would do all the wrong things, including
just traversing past the end of the requested area due to 'addr' never
matching 'end' exactly.
This is not a pretty fix, and we may actually want to move the alignment
into generic code, leaving just the core code per-arch, but Alexey
verified that the vmsplice01 LTP test doesn't crash with this.
Reported-and-tested-by: Alexey Dobriyan <adobriyan@gmail.com>
Debugged-by: Johannes Weiner <hannes@saeurebad.de>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Implement get_user_pages_fast without locking in the fastpath on x86.
Do an optimistic lockless pagetable walk, without taking mmap_sem or any
page table locks or even mmap_sem. Page table existence is guaranteed by
turning interrupts off (combined with the fact that we're always looking
up the current mm, means we can do the lockless page table walk within the
constraints of the TLB shootdown design). Basically we can do this
lockless pagetable walk in a similar manner to the way the CPU's pagetable
walker does not have to take any locks to find present ptes.
This patch (combined with the subsequent ones to convert direct IO to use
it) was found to give about 10% performance improvement on a 2 socket 8
core Intel Xeon system running an OLTP workload on DB2 v9.5
"To test the effects of the patch, an OLTP workload was run on an IBM
x3850 M2 server with 2 processors (quad-core Intel Xeon processors at
2.93 GHz) using IBM DB2 v9.5 running Linux 2.6.24rc7 kernel. Comparing
runs with and without the patch resulted in an overall performance
benefit of ~9.8%. Correspondingly, oprofiles showed that samples from
__up_read and __down_read routines that is seen during thread contention
for system resources was reduced from 2.8% down to .05%. Monitoring the
/proc/vmstat output from the patched run showed that the counter for
fast_gup contained a very high number while the fast_gup_slow value was
zero."
(fast_gup is the old name for get_user_pages_fast, fast_gup_slow is a
counter we had for the number of times the slowpath was invoked).
The main reason for the improvement is that DB2 has multiple threads each
issuing direct-IO. Direct-IO uses get_user_pages, and thus the threads
contend the mmap_sem cacheline, and can also contend on page table locks.
I would anticipate larger performance gains on larger systems, however I
think DB2 uses an adaptive mix of threads and processes, so it could be
that thread contention remains pretty constant as machine size increases.
In which case, we stuck with "only" a 10% gain.
The downside of using get_user_pages_fast is that if there is not a pte
with the correct permissions for the access, we end up falling back to
get_user_pages and so the get_user_pages_fast is a bit of extra work.
However this should not be the common case in most performance critical
code.
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: Kconfig fix]
[akpm@linux-foundation.org: Makefile fix/cleanup]
[akpm@linux-foundation.org: warning fix]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Dave Kleikamp <shaggy@austin.ibm.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Dave Kleikamp <shaggy@austin.ibm.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Zach Brown <zach.brown@oracle.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Youquan Song
Andrea Arcangeli
Andy Shevchenko
Linus Torvalds
Ingo Molnar
Peter Zijlstra
Andy Grover
Harvey Harrison
Jan Beulich
Nick Piggin