linux/mm, branch v3.5

Merge branch 'akpm' (Andrew's patch-bomb)

2012-07-17T23:24:09Z

Merge Andrew's remaining patches for 3.5: "Nine fixes" * Merge emailed patches from Andrew Morton : (9 commits) mm: fix lost kswapd wakeup in kswapd_stop() m32r: make memset() global for CONFIG_KERNEL_BZIP2=y m32r: add memcpy() for CONFIG_KERNEL_GZIP=y m32r: consistently use "suffix-$(...)" m32r: fix 'fix breakage from "m32r: use generic ptrace_resume code"' fallout m32r: fix pull clearing RESTORE_SIGMASK into block_sigmask() fallout m32r: remove duplicate definition of PTRACE_O_TRACESYSGOOD mn10300: fix "pull clearing RESTORE_SIGMASK into block_sigmask()" fallout bootmem: make ___alloc_bootmem_node_nopanic() really nopanic

mm: fix lost kswapd wakeup in kswapd_stop()

2012-07-17T23:21:30Z

Offlining memory may block forever, waiting for kswapd() to wake up because kswapd() does not check the event kthread->should_stop before sleeping. The proper pattern, from Documentation/memory-barriers.txt, is: --- waker --- event_indicated = 1; wake_up_process(event_daemon); --- sleeper --- for (;;) { set_current_state(TASK_UNINTERRUPTIBLE); if (event_indicated) break; schedule(); } set_current_state() may be wrapped by: prepare_to_wait(); In the kswapd() case, event_indicated is kthread->should_stop. === offlining memory (waker) === kswapd_stop() kthread_stop() kthread->should_stop = 1 wake_up_process() wait_for_completion() === kswapd_try_to_sleep (sleeper) === kswapd_try_to_sleep() prepare_to_wait() . . schedule() . . finish_wait() The schedule() needs to be protected by a test of kthread->should_stop, which is wrapped by kthread_should_stop(). Reproducer: Do heavy file I/O in background. Do a memory offline/online in a tight loop Signed-off-by: Aaditya Kumar Acked-by: KOSAKI Motohiro Reviewed-by: Minchan Kim Acked-by: Mel Gorman Cc: Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

bootmem: make ___alloc_bootmem_node_nopanic() really nopanic

2012-07-17T23:21:29Z

In reaction to commit 99ab7b19440a ("mm: sparse: fix usemap allocation above node descriptor section") Johannes said: | while backporting the below patch, I realised that your fix busted | f5bf18fa22f8 again. The problem was not a panicking version on | allocation failure but when the usemap size was too large such that | goal + size > limit triggers the BUG_ON in the bootmem allocator. So | we need a version that passes limit ONLY if the usemap is smaller than | the section. after checking the code, the name of ___alloc_bootmem_node_nopanic() does not reflect the fact. Make bootmem really not panic. Hope will kill bootmem sooner. Signed-off-by: Yinghai Lu Cc: Johannes Weiner Cc: [3.3.x, 3.4.x] Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

Merge branch 'fixes-for-linus' of git://git.linaro.org/people/mszyprowski/linux-dma-mapping

2012-07-17T15:43:12Z

Pull CMA and DMA-mapping fixes from Marek Szyprowski: "Another set of minor fixups for recently merged Contiguous Memory Allocator and ARM DMA-mapping changes. Those patches fix mysterious crashes on systems with CMA and Himem enabled as well as some corner cases caused by typical off-by-one bug." * 'fixes-for-linus' of git://git.linaro.org/people/mszyprowski/linux-dma-mapping: ARM: dma-mapping: modify condition check while freeing pages mm: cma: fix condition check when setting global cma area mm: cma: don't replace lowmem pages with highmem

memblock: free allocated memblock_reserved_regions later

2012-07-11T23:04:50Z

memblock_free_reserved_regions() calls memblock_free(), but memblock_free() would double reserved.regions too, so we could free the old range for reserved.regions. Also tj said there is another bug which could be related to this. | I don't think we're saving any noticeable | amount by doing this "free - give it to page allocator - reserve | again" dancing. We should just allocate regions aligned to page | boundaries and free them later when memblock is no longer in use. in that case, when DEBUG_PAGEALLOC, will get panic: memblock_free: [0x0000102febc080-0x0000102febf080] memblock_free_reserved_regions+0x37/0x39 BUG: unable to handle kernel paging request at ffff88102febd948 IP: [] __next_free_mem_range+0x9b/0x155 PGD 4826063 PUD cf67a067 PMD cf7fa067 PTE 800000102febd160 Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC CPU 0 Pid: 0, comm: swapper Not tainted 3.5.0-rc2-next-20120614-sasha #447 RIP: 0010:[] [] __next_free_mem_range+0x9b/0x155 See the discussion at https://lkml.org/lkml/2012/6/13/469 So try to allocate with PAGE_SIZE alignment and free it later. Reported-by: Sasha Levin Acked-by: Tejun Heo Cc: Benjamin Herrenschmidt Signed-off-by: Yinghai Lu Cc: Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

mm: sparse: fix usemap allocation above node descriptor section

2012-07-11T23:04:49Z

After commit f5bf18fa22f8 ("bootmem/sparsemem: remove limit constraint in alloc_bootmem_section"), usemap allocations may easily be placed outside the optimal section that holds the node descriptor, even if there is space available in that section. This results in unnecessary hotplug dependencies that need to have the node unplugged before the section holding the usemap. The reason is that the bootmem allocator doesn't guarantee a linear search starting from the passed allocation goal but may start out at a much higher address absent an upper limit. Fix this by trying the allocation with the limit at the section end, then retry without if that fails. This keeps the fix from f5bf18fa22f8 of not panicking if the allocation does not fit in the section, but still makes sure to try to stay within the section at first. Signed-off-by: Yinghai Lu Signed-off-by: Johannes Weiner Cc: [3.3.x, 3.4.x] Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

mm: sparse: fix section usemap placement calculation

2012-07-11T23:04:49Z

Commit 238305bb4d41 ("mm: remove sparsemem allocation details from the bootmem allocator") introduced a bug in the allocation goal calculation that put section usemaps not in the same section as the node descriptors, creating unnecessary hotplug dependencies between them: node 0 must be removed before remove section 16399 node 1 must be removed before remove section 16399 node 2 must be removed before remove section 16399 node 3 must be removed before remove section 16399 node 4 must be removed before remove section 16399 node 5 must be removed before remove section 16399 node 6 must be removed before remove section 16399 The reason is that it applies PAGE_SECTION_MASK to the physical address of the node descriptor when finding a suitable place to put the usemap, when this mask is actually intended to be used with PFNs. Because the PFN mask is wider, the target address will point beyond the wanted section holding the node descriptor and the node must be offlined before the section holding the usemap can go. Fix this by extending the mask to address width before use. Signed-off-by: Yinghai Lu Signed-off-by: Johannes Weiner Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

shmem: cleanup shmem_add_to_page_cache

2012-07-11T23:04:48Z

shmem_add_to_page_cache() has three callsites, but only one of them wants the radix_tree_preload() (an exceptional entry guarantees that the radix tree node is present in the other cases), and only that site can achieve mem_cgroup_uncharge_cache_page() (PageSwapCache makes it a no-op in the other cases). We did it this way originally to reflect add_to_page_cache_locked(); but it's confusing now, so move the radix_tree preloading and mem_cgroup uncharging to that one caller. Signed-off-by: Hugh Dickins Acked-by: Johannes Weiner Cc: KAMEZAWA Hiroyuki Cc: Michal Hocko Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

shmem: fix negative rss in memcg memory.stat

2012-07-11T23:04:48Z

When adding the page_private checks before calling shmem_replace_page(), I did realize that there is a further race, but thought it too unlikely to need a hurried fix. But independently I've been chasing why a mem cgroup's memory.stat sometimes shows negative rss after all tasks have gone: I expected it to be a stats gathering bug, but actually it's shmem swapping's fault. It's an old surprise, that when you lock_page(lookup_swap_cache(swap)), the page may have been removed from swapcache before getting the lock; or it may have been freed and reused and be back in swapcache; and it can even be using the same swap location as before (page_private same). The swapoff case is already secure against this (swap cannot be reused until the whole area has been swapped off, and a new swapped on); and shmem_getpage_gfp() is protected by shmem_add_to_page_cache()'s check for the expected radix_tree entry - but a little too late. By that time, we might have already decided to shmem_replace_page(): I don't know of a problem from that, but I'd feel more at ease not to do so spuriously. And we have already done mem_cgroup_cache_charge(), on perhaps the wrong mem cgroup: and this charge is not then undone on the error path, because PageSwapCache ends up preventing that. It's this last case which causes the occasional negative rss in memory.stat: the page is charged here as cache, but (sometimes) found to be anon when eventually it's uncharged - and in between, it's an undeserved charge on the wrong memcg. Fix this by adding an earlier check on the radix_tree entry: it's inelegant to descend the tree twice, but swapping is not the fast path, and a better solution would need a pair (try+commit) of memcg calls, and a rework of shmem_replace_page() to keep out of the swapcache. We can use the added shmem_confirm_swap() function to replace the find_get_page+page_cache_release we were already doing on the error path. And add a comment on that -EEXIST: it seems a peculiar errno to be using, but originates from its use in radix_tree_insert(). [It can be surprising to see positive rss left in a memcg's memory.stat after all tasks have gone, since it is supposed to count anonymous but not shmem. Aside from sharing anon pages via fork with a task in some other memcg, it often happens after swapping: because a swap page can't be freed while under writeback, nor while locked. So it's not an error, and these residual pages are easily freed once pressure demands.] Signed-off-by: Hugh Dickins Acked-by: Johannes Weiner Cc: KAMEZAWA Hiroyuki Cc: Michal Hocko Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

tmpfs: revert SEEK_DATA and SEEK_HOLE

2012-07-11T23:04:48Z

Revert 4fb5ef089b28 ("tmpfs: support SEEK_DATA and SEEK_HOLE"). I believe it's correct, and it's been nice to have from rc1 to rc6; but as the original commit said: I don't know who actually uses SEEK_DATA or SEEK_HOLE, and whether it would be of any use to them on tmpfs. This code adds 92 lines and 752 bytes on x86_64 - is that bloat or worthwhile? Nobody asked for it, so I conclude that it's bloat: let's revert tmpfs to the dumb generic support for v3.5. We can always reinstate it later if useful, and anyone needing it in a hurry can just get it out of git. Signed-off-by: Hugh Dickins Cc: Christoph Hellwig Cc: Josef Bacik Cc: Andi Kleen Cc: Andreas Dilger Cc: Dave Chinner Cc: Marco Stornelli Cc: Jeff liu Cc: Chris Mason Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds