If prepare_write fails with AOP_TRUNCATED_PAGE, or if commit_write fails, then
we may have failed the write operation despite prepare_write having
instantiated blocks past i_size.  Fix this, and consolidate the trimming into
one place.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Adrian Bunk <bunk@kernel.org>

static functions mustn't be exported.

Signed-off-by: Adrian Bunk <bunk@kernel.org>

tmpfs was misconverted to __GFP_ZERO in 2.6.11.  There's an unusual case in
which shmem_getpage receives the page from its caller instead of allocating.
We must cover this case by clear_highpage before SetPageUptodate, as before.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Adrian Bunk <bunk@kernel.org>

The lats commit causes the wrong return value.
is_hugepage_only_range() is a boolean, so we should return
-EINVAL rather than 1.

Also - we can use "mm" instead of looking up "current->mm" again.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Adrian Bunk <bunk@kernel.org>

Unlike mmap(), the codepath for brk() creates a vma without first checking
that it doesn't touch a region exclusively reserved for hugepages.  On
powerpc, this can allow it to create a normal page vma in a hugepage
region, causing oopses and other badness.

Add a test to prevent this.  With this patch, brk() will simply fail if it
attempts to move the break into a hugepage reserved region.

Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Adrian Bunk <bunk@kernel.org>

When expanding the stack, we don't currently check if the VMA will cross
into an area of the address space that is reserved for hugetlb pages.
Subsequent faults on the expanded portion of such a VMA will confuse the
low-level MMU code, resulting in an OOPS.  Check for this.

Signed-off-by: Adam Litke <agl@us.ibm.com>
Signed-off-by: Adrian Bunk <bunk@kernel.org>

sys_madvise has down_write of mmap_sem, then madvise_remove calls
vmtruncate_range which takes i_mutex and i_alloc_sem: no, we can
easily devise deadlocks from that ordering.

madvise_remove drop mmap_sem while calling vmtruncate_range: luckily,
since madvise_remove doesn't split or merge vmas, it's easy to handle
this case with a NULL prev, without restructuring sys_madvise.  (Though
sad to retake mmap_sem when it's unlikely to be needed, and certainly
down_read is sufficient for MADV_REMOVE, unlike the other madvices.)

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>

shmem_truncate_range has its own truncate_inode_pages_range, to free any
pages racily instantiated while it was in progress: a SHMEM_PAGEIN flag
is set when this might have happened.  But holepunching gets no chance
to clear that flag at the start of vmtruncate_range, so it's always set
(unless a truncate came just before), so holepunch almost always does
this second truncate_inode_pages_range.

shmem holepunch has unlikely swap<->file races hereabouts whatever we do
(without a fuller rework than is fit for this release): I was going to
skip the second truncate in the punch_hole case, but Miklos points out
that would make holepunch correctness more vulnerable to swapoff.  So
keep the second truncate, but follow it by an unmap_mapping_range to
eliminate the disconnected pages (freed from pagecache while still
mapped in userspace) that it might have left behind.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>

Miklos Szeredi observes that during truncation of shmem page directories,
info->lock is released to improve latency (after lowering i_size and
next_index to exclude races); but this is quite wrong for holepunching,
which receives no such protection from i_size or next_index, and is left
vulnerable to races with shmem_unuse, shmem_getpage and shmem_writepage.

Hold info->lock throughout when holepunching?  No, any user could prevent
rescheduling for far too long.  Instead take info->lock just when needed:
in shmem_free_swp when removing the swap entries, and whenever removing
a directory page from the level above.  But so long as we remove before
scanning, we can safely skip taking the lock at the lower levels, except
at misaligned start and end of the hole.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>

Miklos Szeredi observes BUG_ON(!entry) in shmem_writepage() triggered
in rare circumstances, because shmem_truncate_range() erroneously
removes partially truncated directory pages at the end of the range:
later reclaim on pages pointing to these removed directories triggers
the BUG.  Indeed, and it can also cause data loss beyond the hole.

Fix this as in the patch proposed by Miklos, but distinguish between
"limit" (how far we need to search: ignore truncation's next_index
optimization in the holepunch case - if there are races it's more
consistent to act on the whole range specified) and "upper_limit"
(how far we can free directory pages: generally we must be careful
to keep partially punched pages, but can relax at end of file -
i_size being held stable by i_mutex).

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>