aboutsummaryrefslogtreecommitdiff
path: root/mm/ksm.c
AgeCommit message (Collapse)Author
2011-10-31oom: fix race while temporarily setting current's oom_score_adjDavid Rientjes
test_set_oom_score_adj() was introduced in 72788c385604 ("oom: replace PF_OOM_ORIGIN with toggling oom_score_adj") to temporarily elevate current's oom_score_adj for ksm and swapoff without requiring an additional per-process flag. Using that function to both set oom_score_adj to OOM_SCORE_ADJ_MAX and then reinstate the previous value is racy since it's possible that userspace can set the value to something else itself before the old value is reinstated. That results in userspace setting current's oom_score_adj to a different value and then the kernel immediately setting it back to its previous value without notification. To fix this, a new compare_swap_oom_score_adj() function is introduced with the same semantics as the compare and swap CAS instruction, or CMPXCHG on x86. It is used to reinstate the previous value of oom_score_adj if and only if the present value is the same as the old value. Signed-off-by: David Rientjes <rientjes@google.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Ying Han <yinghan@google.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-06-15ksm: fix NULL pointer dereference in scan_get_next_rmap_item()Hugh Dickins
Andrea Righi reported a case where an exiting task can race against ksmd::scan_get_next_rmap_item (http://lkml.org/lkml/2011/6/1/742) easily triggering a NULL pointer dereference in ksmd. ksm_scan.mm_slot == &ksm_mm_head with only one registered mm CPU 1 (__ksm_exit) CPU 2 (scan_get_next_rmap_item) list_empty() is false lock slot == &ksm_mm_head list_del(slot->mm_list) (list now empty) unlock lock slot = list_entry(slot->mm_list.next) (list is empty, so slot is still ksm_mm_head) unlock slot->mm == NULL ... Oops Close this race by revalidating that the new slot is not simply the list head again. Andrea's test case: #include <stdio.h> #include <stdlib.h> #include <unistd.h> #include <sys/mman.h> #define BUFSIZE getpagesize() int main(int argc, char **argv) { void *ptr; if (posix_memalign(&ptr, getpagesize(), BUFSIZE) < 0) { perror("posix_memalign"); exit(1); } if (madvise(ptr, BUFSIZE, MADV_MERGEABLE) < 0) { perror("madvise"); exit(1); } *(char *)NULL = 0; return 0; } Reported-by: Andrea Righi <andrea@betterlinux.com> Tested-by: Andrea Righi <andrea@betterlinux.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Chris Wright <chrisw@sous-sol.org> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-05-25oom: replace PF_OOM_ORIGIN with toggling oom_score_adjDavid Rientjes
There's a kernel-wide shortage of per-process flags, so it's always helpful to trim one when possible without incurring a significant penalty. It's even more important when you're planning on adding a per- process flag yourself, which I plan to do shortly for transparent hugepages. PF_OOM_ORIGIN is used by ksm and swapoff to prefer current since it has a tendency to allocate large amounts of memory and should be preferred for killing over other tasks. We'd rather immediately kill the task making the errant syscall rather than penalizing an innocent task. This patch removes PF_OOM_ORIGIN since its behavior is equivalent to setting the process's oom_score_adj to OOM_SCORE_ADJ_MAX. The process's old oom_score_adj is stored and then set to OOM_SCORE_ADJ_MAX during the time it used to have PF_OOM_ORIGIN. The old value is then reinstated when the process should no longer be considered a high priority for oom killing. Signed-off-by: David Rientjes <rientjes@google.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Cc: Hugh Dickins <hughd@google.com> Cc: Izik Eidus <ieidus@redhat.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-31Fix common misspellingsLucas De Marchi
Fixes generated by 'codespell' and manually reviewed. Signed-off-by: Lucas De Marchi <lucas.demarchi@profusion.mobi>
2011-03-22mm: rename drop_anon_vma() to put_anon_vma()Peter Zijlstra
The normal code pattern used in the kernel is: get/put. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Hugh Dickins <hughd@google.com> Reviewed-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>
2011-01-13ksm: drain pagevecs to lruHugh Dickins
It was hard to explain the page counts which were causing new LTP tests of KSM to fail: we need to drain the per-cpu pagevecs to LRU occasionally. Signed-off-by: Hugh Dickins <hughd@google.com> Reported-by: CAI Qian <caiqian@redhat.com> Cc:Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13thp: add compound_trans_head() helperAndrea Arcangeli
Cleanup some code with common compound_trans_head helper. Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <jweiner@redhat.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Avi Kivity <avi@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13thp: KSM on THPAndrea Arcangeli
This makes KSM full operational with THP pages. Subpages are scanned while the hugepage is still in place and delivering max cpu performance, and only if there's a match and we're going to deduplicate memory, the single hugepages with the subpage match is split. There will be no false sharing between ksmd and khugepaged. khugepaged won't collapse 2m virtual regions with KSM pages inside. ksmd also should only split pages when the checksum matches and we're likely to split an hugepage for some long living ksm page (usual ksm heuristic to avoid sharing pages that get de-cowed). Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13thp: freeze khugepaged and ksmdAndrea Arcangeli
It's unclear why schedule friendly kernel threads can't be taken away by the CPU through the scheduler itself. It's safer to stop them as they can trigger memory allocation, if kswapd also freezes itself to avoid generating I/O they have too. 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>
2011-01-13thp: skip transhuge pages in ksm for nowAndrea Arcangeli
Skip transhuge pages in ksm for now. Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Reviewed-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>
2011-01-13thp: ksm: free swap when swapcache page is replacedHugh Dickins
When a swapcache page is replaced by a ksm page, it's best to free that swap immediately. Reported-by: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Hugh Dickins <hughd@google.com> 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>
2010-12-02ksm: annotate ksm_thread_mutex is no deadlock sourceKOSAKI Motohiro
commit 62b61f611e ("ksm: memory hotremove migration only") caused the following new lockdep warning. ======================================================= [ INFO: possible circular locking dependency detected ] ------------------------------------------------------- bash/1621 is trying to acquire lock: ((memory_chain).rwsem){.+.+.+}, at: [<ffffffff81079339>] __blocking_notifier_call_chain+0x69/0xc0 but task is already holding lock: (ksm_thread_mutex){+.+.+.}, at: [<ffffffff8113a3aa>] ksm_memory_callback+0x3a/0xc0 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #1 (ksm_thread_mutex){+.+.+.}: [<ffffffff8108b70a>] lock_acquire+0xaa/0x140 [<ffffffff81505d74>] __mutex_lock_common+0x44/0x3f0 [<ffffffff81506228>] mutex_lock_nested+0x48/0x60 [<ffffffff8113a3aa>] ksm_memory_callback+0x3a/0xc0 [<ffffffff8150c21c>] notifier_call_chain+0x8c/0xe0 [<ffffffff8107934e>] __blocking_notifier_call_chain+0x7e/0xc0 [<ffffffff810793a6>] blocking_notifier_call_chain+0x16/0x20 [<ffffffff813afbfb>] memory_notify+0x1b/0x20 [<ffffffff81141b7c>] remove_memory+0x1cc/0x5f0 [<ffffffff813af53d>] memory_block_change_state+0xfd/0x1a0 [<ffffffff813afd62>] store_mem_state+0xe2/0xf0 [<ffffffff813a0bb0>] sysdev_store+0x20/0x30 [<ffffffff811bc116>] sysfs_write_file+0xe6/0x170 [<ffffffff8114f398>] vfs_write+0xc8/0x190 [<ffffffff8114fc14>] sys_write+0x54/0x90 [<ffffffff810028b2>] system_call_fastpath+0x16/0x1b -> #0 ((memory_chain).rwsem){.+.+.+}: [<ffffffff8108b5ba>] __lock_acquire+0x155a/0x1600 [<ffffffff8108b70a>] lock_acquire+0xaa/0x140 [<ffffffff81506601>] down_read+0x51/0xa0 [<ffffffff81079339>] __blocking_notifier_call_chain+0x69/0xc0 [<ffffffff810793a6>] blocking_notifier_call_chain+0x16/0x20 [<ffffffff813afbfb>] memory_notify+0x1b/0x20 [<ffffffff81141f1e>] remove_memory+0x56e/0x5f0 [<ffffffff813af53d>] memory_block_change_state+0xfd/0x1a0 [<ffffffff813afd62>] store_mem_state+0xe2/0xf0 [<ffffffff813a0bb0>] sysdev_store+0x20/0x30 [<ffffffff811bc116>] sysfs_write_file+0xe6/0x170 [<ffffffff8114f398>] vfs_write+0xc8/0x190 [<ffffffff8114fc14>] sys_write+0x54/0x90 [<ffffffff810028b2>] system_call_fastpath+0x16/0x1b But it's a false positive. Both memory_chain.rwsem and ksm_thread_mutex have an outer lock (mem_hotplug_mutex). So they cannot deadlock. Thus, This patch annotate ksm_thread_mutex is not deadlock source. [akpm@linux-foundation.org: update comment, from Hugh] Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-04ksm: fix bad user data when swappingHugh Dickins
Building under memory pressure, with KSM on 2.6.36-rc5, collapsed with an internal compiler error: typically indicating an error in swapping. Perhaps there's a timing issue which makes it now more likely, perhaps it's just a long time since I tried for so long: this bug goes back to KSM swapping in 2.6.33. Notice how reuse_swap_page() allows an exclusive page to be reused, but only does SetPageDirty if it can delete it from swap cache right then - if it's currently under Writeback, it has to be left in cache and we don't SetPageDirty, but the page can be reused. Fine, the dirty bit will get set in the pte; but notice how zap_pte_range() does not bother to transfer pte_dirty to page_dirty when unmapping a PageAnon. If KSM chooses to share such a page, it will look like a clean copy of swapcache, and not be written out to swap when its memory is needed; then stale data read back from swap when it's needed again. We could fix this in reuse_swap_page() (or even refuse to reuse a page under writeback), but it's more honest to fix my oversight in KSM's write_protect_page(). Several days of testing on three machines confirms that this fixes the issue they showed. Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: stable@kernel.org Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-09-09mm: fix swapin race conditionAndrea Arcangeli
The pte_same check is reliable only if the swap entry remains pinned (by the page lock on swapcache). We've also to ensure the swapcache isn't removed before we take the lock as try_to_free_swap won't care about the page pin. One of the possible impacts of this patch is that a KSM-shared page can point to the anon_vma of another process, which could exit before the page is freed. This can leave a page with a pointer to a recycled anon_vma object, or worse, a pointer to something that is no longer an anon_vma. [riel@redhat.com: changelog help] Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Acked-by: Hugh Dickins <hughd@google.com> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-08-09ksm: cleanup for mm_slots_hashLai Jiangshan
Use compile-allocated memory instead of dynamic allocated memory for mm_slots_hash. Use hash_ptr() instead divisions for bucket calculation. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Izik Eidus <ieidus@redhat.com> Cc: Avi Kivity <avi@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-08-09mm: extend KSM refcounts to the anon_vma rootRik van Riel
KSM reference counts can cause an anon_vma to exist after the processe it belongs to have already exited. Because the anon_vma lock now lives in the root anon_vma, we need to ensure that the root anon_vma stays around until after all the "child" anon_vmas have been freed. The obvious way to do this is to have a "child" anon_vma take a reference to the root in anon_vma_fork. When the anon_vma is freed at munmap or process exit, we drop the refcount in anon_vma_unlink and possibly free the root anon_vma. The KSM anon_vma reference count function also needs to be modified to deal with the possibility of freeing 2 levels of anon_vma. The easiest way to do this is to break out the KSM magic and make it generic. When compiling without CONFIG_KSM, this code is compiled out. Signed-off-by: Rik van Riel <riel@redhat.com> Tested-by: Larry Woodman <lwoodman@redhat.com> Acked-by: Larry Woodman <lwoodman@redhat.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Tested-by: Dave Young <hidave.darkstar@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-08-09mm: always lock the root (oldest) anon_vmaRik van Riel
Always (and only) lock the root (oldest) anon_vma whenever we do something in an anon_vma. The recently introduced anon_vma scalability is due to the rmap code scanning only the VMAs that need to be scanned. Many common operations still took the anon_vma lock on the root anon_vma, so always taking that lock is not expected to introduce any scalability issues. However, always taking the same lock does mean we only need to take one lock, which means rmap_walk on pages from any anon_vma in the vma is excluded from occurring during an munmap, expand_stack or other operation that needs to exclude rmap_walk and similar functions. Also add the proper locking to vma_adjust. Signed-off-by: Rik van Riel <riel@redhat.com> Tested-by: Larry Woodman <lwoodman@redhat.com> Acked-by: Larry Woodman <lwoodman@redhat.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-08-09mm: change direct call of spin_lock(anon_vma->lock) to inline functionRik van Riel
Subsitute a direct call of spin_lock(anon_vma->lock) with an inline function doing exactly the same. This makes it easier to do the substitution to the root anon_vma lock in a following patch. We will deal with the handful of special locks (nested, dec_and_lock, etc) separately. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mel@csn.ul.ie> Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Tested-by: Larry Woodman <lwoodman@redhat.com> Acked-by: Larry Woodman <lwoodman@redhat.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-25mm: migration: share the anon_vma ref counts between KSM and page migrationMel Gorman
For clarity of review, KSM and page migration have separate refcounts on the anon_vma. While clear, this is a waste of memory. This patch gets KSM and page migration to share their toys in a spirit of harmony. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reviewed-by: Christoph Lameter <cl@linux-foundation.org> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-04-24ksm: check for ERR_PTR from follow_page()Dan Carpenter
The follow_page() function can potentially return -EFAULT so I added checks for this. Also I silenced an uninitialized variable warning on my version of gcc (version 4.3.2). Signed-off-by: Dan Carpenter <error27@gmail.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-24mm/ksm.c is doing an unneeded _notify in write_protect_page.Robin Holt
ksm.c's write_protect_page implements a lockless means of verifying a page does not have any users of the page which are not accounted for via other kernel tracking means. It does this by removing the writable pte with TLB flushes, checking the page_count against the total known users, and then using set_pte_at_notify to make it a read-only entry. An unneeded mmu_notifier callout is made in the case where the known users does not match the page_count. In that event, we are inserting the identical pte and there is no need for the set_pte_at_notify, but rather the simpler set_pte_at suffices. Signed-off-by: Robin Holt <holt@sgi.com> Acked-by: Izik Eidus <ieidus@redhat.com> Acked-by: Andrea Arcangeli <aarcange@redhat.com> Acked-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06mm: change anon_vma linking to fix multi-process server scalability issueRik van Riel
The old anon_vma code can lead to scalability issues with heavily forking workloads. Specifically, each anon_vma will be shared between the parent process and all its child processes. In a workload with 1000 child processes and a VMA with 1000 anonymous pages per process that get COWed, this leads to a system with a million anonymous pages in the same anon_vma, each of which is mapped in just one of the 1000 processes. However, the current rmap code needs to walk them all, leading to O(N) scanning complexity for each page. This can result in systems where one CPU is walking the page tables of 1000 processes in page_referenced_one, while all other CPUs are stuck on the anon_vma lock. This leads to catastrophic failure for a benchmark like AIM7, where the total number of processes can reach in the tens of thousands. Real workloads are still a factor 10 less process intensive than AIM7, but they are catching up. This patch changes the way anon_vmas and VMAs are linked, which allows us to associate multiple anon_vmas with a VMA. At fork time, each child process gets its own anon_vmas, in which its COWed pages will be instantiated. The parents' anon_vma is also linked to the VMA, because non-COWed pages could be present in any of the children. This reduces rmap scanning complexity to O(1) for the pages of the 1000 child processes, with O(N) complexity for at most 1/N pages in the system. This reduces the average scanning cost in heavily forking workloads from O(N) to 2. The only real complexity in this patch stems from the fact that linking a VMA to anon_vmas now involves memory allocations. This means vma_adjust can fail, if it needs to attach a VMA to anon_vma structures. This in turn means error handling needs to be added to the calling functions. A second source of complexity is that, because there can be multiple anon_vmas, the anon_vma linking in vma_adjust can no longer be done under "the" anon_vma lock. To prevent the rmap code from walking up an incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h to make sure it is impossible to compile a kernel that needs both symbolic values for the same bitflag. Some test results: Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test box with 16GB RAM and not quite enough IO), the system ends up running >99% in system time, with every CPU on the same anon_vma lock in the pageout code. With these changes, AIM7 hits the cross-over point around 29.7k users. This happens with ~99% IO wait time, there never seems to be any spike in system time. The anon_vma lock contention appears to be resolved. [akpm@linux-foundation.org: cleanups] Signed-off-by: Rik van Riel <riel@redhat.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Larry Woodman <lwoodman@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: remove unswappable max_kernel_pagesHugh Dickins
Now that ksm pages are swappable, and the known holes plugged, remove mention of unswappable kernel pages from KSM documentation and comments. Remove the totalram_pages/4 initialization of max_kernel_pages. In fact, remove max_kernel_pages altogether - we can reinstate it if removal turns out to break someone's script; but if we later want to limit KSM's memory usage, limiting the stable nodes would not be an effective approach. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: memory hotremove migration onlyHugh Dickins
The previous patch enables page migration of ksm pages, but that soon gets into trouble: not surprising, since we're using the ksm page lock to lock operations on its stable_node, but page migration switches the page whose lock is to be used for that. Another layer of locking would fix it, but do we need that yet? Do we actually need page migration of ksm pages? Yes, memory hotremove needs to offline sections of memory: and since we stopped allocating ksm pages with GFP_HIGHUSER, they will tend to be GFP_HIGHUSER_MOVABLE candidates for migration. But KSM is currently unconscious of NUMA issues, happily merging pages from different NUMA nodes: at present the rule must be, not to use MADV_MERGEABLE where you care about NUMA. So no, NUMA page migration of ksm pages does not make sense yet. So, to complete support for ksm swapping we need to make hotremove safe. ksm_memory_callback() take ksm_thread_mutex when MEM_GOING_OFFLINE and release it when MEM_OFFLINE or MEM_CANCEL_OFFLINE. But if mapped pages are freed before migration reaches them, stable_nodes may be left still pointing to struct pages which have been removed from the system: the stable_node needs to identify a page by pfn rather than page pointer, then it can safely prune them when MEM_OFFLINE. And make NUMA migration skip PageKsm pages where it skips PageReserved. But it's only when we reach unmap_and_move() that the page lock is taken and we can be sure that raised pagecount has prevented a PageAnon from being upgraded: so add offlining arg to migrate_pages(), to migrate ksm page when offlining (has sufficient locking) but reject it otherwise. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: rmap_walk to remove_migation_ptesHugh Dickins
A side-effect of making ksm pages swappable is that they have to be placed on the LRUs: which then exposes them to isolate_lru_page() and hence to page migration. Add rmap_walk() for remove_migration_ptes() to use: rmap_walk_anon() and rmap_walk_file() in rmap.c, but rmap_walk_ksm() in ksm.c. Perhaps some consolidation with existing code is possible, but don't attempt that yet (try_to_unmap needs to handle nonlinears, but migration pte removal does not). rmap_walk() is sadly less general than it appears: rmap_walk_anon(), like remove_anon_migration_ptes() which it replaces, avoids calling page_lock_anon_vma(), because that includes a page_mapped() test which fails when all migration ptes are in place. That was valid when NUMA page migration was introduced (holding mmap_sem provided the missing guarantee that anon_vma's slab had not already been destroyed), but I believe not valid in the memory hotremove case added since. For now do the same as before, and consider the best way to fix that unlikely race later on. When fixed, we can probably use rmap_walk() on hwpoisoned ksm pages too: for now, they remain among hwpoison's various exceptions (its PageKsm test comes before the page is locked, but its page_lock_anon_vma fails safely if an anon gets upgraded). Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: share anon page without allocatingHugh Dickins
When ksm pages were unswappable, it made no sense to include them in mem cgroup accounting; but now that they are swappable (although I see no strict logical connection) the principle of least surprise implies that they should be accounted (with the usual dissatisfaction, that a shared page is accounted to only one of the cgroups using it). This patch was intended to add mem cgroup accounting where necessary; but turned inside out, it now avoids allocating a ksm page, instead upgrading an anon page to ksm - which brings its existing mem cgroup accounting with it. Thus mem cgroups don't appear in the patch at all. This upgrade from PageAnon to PageKsm takes place under page lock (via a somewhat hacky NULL kpage interface), and audit showed only one place which needed to cope with the race - page_referenced() is sometimes used without page lock, so page_lock_anon_vma() needs an ACCESS_ONCE() to be sure of getting anon_vma and flags together (no problem if the page goes ksm an instant after, the integrity of that anon_vma list is unaffected). Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: take keyhole reference to pageHugh Dickins
There's a lamentable flaw in KSM swapping: the stable_node holds a reference to the ksm page, so the page to be freed cannot actually be freed until ksmd works its way around to removing the last rmap_item from its stable_node. Which in some configurations may take minutes: not quite responsive enough for memory reclaim. And we don't want to twist KSM and its locking more tightly into the rest of mm. What a pity. But although the stable_node needs to hold a pointer to the ksm page, does it actually need to raise the reference count of that page? No. It would need to do so if struct pages were ordinary kmalloc'ed objects; but they are more stable than that, and reused in particular ways according to particular rules. Access to stable_node from its pointer in struct page is no problem, so long as we never free a stable_node before the ksm page itself has been freed. Access to struct page from its pointer in stable_node: reintroduce get_ksm_page(), and let that peep out through its keyhole (the stable_node pointer to ksm page), to see if that struct page still holds the right key to open it (the ksm page mapping pointer back to this stable_node). This relies upon the established way in which free_hot_cold_page() sets an anon (including ksm) page->mapping to NULL; and relies upon no other user of a struct page to put something which looks like the original stable_node pointer (with two low bits also set) into page->mapping. It also needs get_page_unless_zero() technique pioneered by speculative pagecache; and uses rcu_read_lock() to keep the guarantees that gives. There are several drivers which put pointers of their own into page-> mapping; but none of those could coincide with our stable_node pointers, since KSM won't free a stable_node until it sees that the page has gone. The only problem case found is the pagetable spinlock USE_SPLIT_PTLOCKS places in struct page (my own abuse): to accommodate GENERIC_LOCKBREAK's break_lock on 32-bit, that spans both page->private and page->mapping. Since break_lock is only 0 or 1, again no confusion for get_ksm_page(). But what of DEBUG_SPINLOCK on 64-bit bigendian? When owner_cpu is 3 (matching PageKsm low bits), it might see 0xdead4ead00000003 in page-> mapping, which might coincide? We could get around that by... but a better answer is to suppress USE_SPLIT_PTLOCKS when DEBUG_SPINLOCK or DEBUG_LOCK_ALLOC, to stop bloating sizeof(struct page) in their case - already proposed in an earlier mm/Kconfig patch. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: hold anon_vma in rmap_itemHugh Dickins
For full functionality, page_referenced_one() and try_to_unmap_one() need to know the vma: to pass vma down to arch-dependent flushes, or to observe VM_LOCKED or VM_EXEC. But KSM keeps no record of vma: nor can it, since vmas get split and merged without its knowledge. Instead, note page's anon_vma in its rmap_item when adding to stable tree: all the vmas which might map that page are listed by its anon_vma. page_referenced_ksm() and try_to_unmap_ksm() then traverse the anon_vma, first to find the probable vma, that which matches rmap_item's mm; but if that is not enough to locate all instances, traverse again to try the others. This catches those occasions when fork has duplicated a pte of a ksm page, but ksmd has not yet come around to assign it an rmap_item. But each rmap_item in the stable tree which refers to an anon_vma needs to take a reference to it. Andrea's anon_vma design cleverly avoided a reference count (an anon_vma was free when its list of vmas was empty), but KSM now needs to add that. Is a 32-bit count sufficient? I believe so - the anon_vma is only free when both count is 0 and list is empty. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: let shared pages be swappableHugh Dickins
Initial implementation for swapping out KSM's shared pages: add page_referenced_ksm() and try_to_unmap_ksm(), which rmap.c calls when faced with a PageKsm page. Most of what's needed can be got from the rmap_items listed from the stable_node of the ksm page, without discovering the actual vma: so in this patch just fake up a struct vma for page_referenced_one() or try_to_unmap_one(), then refine that in the next patch. Add VM_NONLINEAR to ksm_madvise()'s list of exclusions: it has always been implicit there (being only set with VM_SHARED, already excluded), but let's make it explicit, to help justify the lack of nonlinear unmap. Rely on the page lock to protect against concurrent modifications to that page's node of the stable tree. The awkward part is not swapout but swapin: do_swap_page() and page_add_anon_rmap() now have to allow for new possibilities - perhaps a ksm page still in swapcache, perhaps a swapcache page associated with one location in one anon_vma now needed for another location or anon_vma. (And the vma might even be no longer VM_MERGEABLE when that happens.) ksm_might_need_to_copy() checks for that case, and supplies a duplicate page when necessary, simply leaving it to a subsequent pass of ksmd to rediscover the identity and merge them back into one ksm page. Disappointingly primitive: but the alternative would have to accumulate unswappable info about the swapped out ksm pages, limiting swappability. Remove page_add_ksm_rmap(): page_add_anon_rmap() now has to allow for the particular case it was handling, so just use it instead. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: fix mlockfreed to munlockedHugh Dickins
When KSM merges an mlocked page, it has been forgetting to munlock it: that's been left to free_page_mlock(), which reports it in /proc/vmstat as unevictable_pgs_mlockfreed instead of unevictable_pgs_munlocked (and whinges "Page flag mlocked set for process" in mmotm, whereas mainline is silently forgiving). Call munlock_vma_page() to fix that. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@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>
2009-12-15ksm: stable_node point to page and backHugh Dickins
Add a pointer to the ksm page into struct stable_node, holding a reference to the page while the node exists. Put a pointer to the stable_node into the ksm page's ->mapping. Then we don't need get_ksm_page() while traversing the stable tree: the page to compare against is sure to be present and correct, even if it's no longer visible through any of its existing rmap_items. And we can handle the forked ksm page case more efficiently: no need to memcmp our way through the tree to find its match. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: separate stable_nodeHugh Dickins
Though we still do well to keep rmap_items in the unstable tree without a separate tree_item at the node, for several reasons it becomes awkward to keep rmap_items in the stable tree without a separate stable_node: lack of space in the nicely-sized rmap_item, the need for an anchor as rmap_items are removed, the need for a node even when temporarily no rmap_items are attached to it. So declare struct stable_node (rb_node to place it in the tree and hlist_head for the rmap_items hanging off it), and convert stable tree handling to use it: without yet taking advantage of it. Note how one stable_tree_insert() of a node now has _two_ stable_tree_append()s of the two rmap_items being merged. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: singly-linked rmap_listHugh Dickins
Free up a pointer in struct rmap_item, by making the mm_slot's rmap_list a singly-linked list: we always traverse that list sequentially, and we don't even lose any prefetches (but should consider adding a few later). Name it rmap_list throughout. Do we need to free up that pointer? Not immediately, and in the end, we could continue to avoid it with a union; but having done the conversion, let's keep it this way, since there's no downside, and maybe we'll want more in future (struct rmap_item is a cache-friendly 32 bytes on 32-bit and 64 bytes on 64-bit, so we shall want to avoid expanding it). Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: cleanup some function argumentsHugh Dickins
Cleanup: make argument names more consistent from cmp_and_merge_page() down to replace_page(), so that it's easier to follow the rmap_item's page and the matching tree_page and the merged kpage through that code. In some places, e.g. break_cow(), pass rmap_item instead of separate mm and address. cmp_and_merge_page() initialize tree_page to NULL, to avoid a "may be used uninitialized" warning seen in one config by Anil SB. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: remove redundancies when merging pageHugh Dickins
There is no need for replace_page() to calculate a write-protected prot vm_page_prot must already be write-protected for an anonymous page (see mm/memory.c do_anonymous_page() for similar reliance on vm_page_prot). There is no need for try_to_merge_one_page() to get_page and put_page on newpage and oldpage: in every case we already hold a reference to each of them. But some instinct makes me move try_to_merge_one_page()'s unlock_page of oldpage down after replace_page(): that doesn't increase contention on the ksm page, and makes thinking about the transition easier. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15ksm: three remove_rmap_item_from_tree cleanupsHugh Dickins
1. remove_rmap_item_from_tree() is called as a precaution from various places: don't dirty the rmap_item cacheline unnecessarily, just mask the flags out of the address when they have been set. 2. First get_next_rmap_item() removes an unstable rmap_item from its tree, then shortly afterwards cmp_and_merge_page() removes a stable rmap_item from its tree: it's easier just to do both at once (but definitely keep the BUG_ON(age > 1) which guards against a future omission). 3. When cmp_and_merge_page() moves an rmap_item from unstable to stable tree, it does its own rb_erase() and accounting: that's better expressed by remove_rmap_item_from_tree(). Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-11-09ksm: cond_resched in unstable treeHugh Dickins
KSM needs a cond_resched() for CONFIG_PREEMPT_NONE, in its unbounded search of the unstable tree. The stable tree cases already have one, and originally there was one down inside get_user_pages(); but I missed it when I converted to follow_page() instead. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-10-08ksm: more on default valuesHugh Dickins
Adjust the max_kernel_pages default to a quarter of totalram_pages, instead of nr_free_buffer_pages() / 4: the KSM pages themselves come from highmem, and even on a 16GB PAE machine, 4GB of KSM pages would only be pinning 32MB of lowmem with their rmap_items, so no need for the more obscure calculation (nor for its own special init function). There is no way for the user to switch KSM on if CONFIG_SYSFS is not enabled, so in that case default run to KSM_RUN_MERGE. Update KSM Documentation and Kconfig to reflect the new defaults. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-24ksm: change default values to better fit into mainline kernelIzik Eidus
Now that ksm is in mainline it is better to change the default values to better fit to most of the users. This patch change the ksm default values to be: ksm_thread_pages_to_scan = 100 (instead of 200) ksm_thread_sleep_millisecs = 20 (like before) ksm_run = KSM_RUN_STOP (instead of KSM_RUN_MERGE - meaning ksm is disabled by default) ksm_max_kernel_pages = nr_free_buffer_pages / 4 (instead of 2046) The important aspect of this patch is: it disables ksm by default, and sets the number of the kernel_pages that can be allocated to be a reasonable number. Signed-off-by: Izik Eidus <ieidus@redhat.com> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: unmerge is an origin of OOMsHugh Dickins
Just as the swapoff system call allocates many pages of RAM to various processes, perhaps triggering OOM, so "echo 2 >/sys/kernel/mm/ksm/run" (unmerge) is liable to allocate many pages of RAM to various processes, perhaps triggering OOM; and each is normally run from a modest admin process (swapoff or shell), easily repeated until it succeeds. So treat unmerge_and_remove_all_rmap_items() in the same way that we treat try_to_unuse(): generalize PF_SWAPOFF to PF_OOM_ORIGIN, and bracket both with that, to ask the OOM killer to kill them first, to prevent them from spawning more and more OOM kills. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: clean up obsolete referencesHugh Dickins
A few cleanups, given the munlock fix: the comment on ksm_test_exit() no longer applies, and it can be made private to ksm.c; there's no more reference to mmu_gather or tlb.h, and mmap.c doesn't need ksm.h. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: sysfs and defaultsHugh Dickins
At present KSM is just a waste of space if you don't have CONFIG_SYSFS=y to provide the /sys/kernel/mm/ksm files to tune and activate it. Make KSM depend on SYSFS? Could do, but it might be better to provide some defaults so that KSM works out-of-the-box, ready for testers to madvise MADV_MERGEABLE, even without SYSFS. Though anyone serious is likely to want to retune the numbers to their taste once they have experience; and whether these settings ever reach 2.6.32 can be discussed along the way. Save 1kB from tiny kernels by #ifdef'ing the SYSFS side of it. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: fix deadlock with munlock in exit_mmapAndrea Arcangeli
Rawhide users have reported hang at startup when cryptsetup is run: the same problem can be simply reproduced by running a program int main() { mlockall(MCL_CURRENT | MCL_FUTURE); return 0; } The problem is that exit_mmap() applies munlock_vma_pages_all() to clean up VM_LOCKED areas, and its current implementation (stupidly) tries to fault in absent pages, for example where PROT_NONE prevented them being faulted in when mlocking. Whereas the "ksm: fix oom deadlock" patch, knowing there's a race by which KSM might try to fault in pages after exit_mmap() had finally zapped the range, backs out of such faults doing nothing when its ksm_test_exit() notices mm_users 0. So revert that part of "ksm: fix oom deadlock" which moved the ksm_exit() call from before exit_mmap() to the middle of exit_mmap(); and remove those ksm_test_exit() checks from the page fault paths, so allowing the munlocking to proceed without interference. ksm_exit, if there are rmap_items still chained on this mm slot, takes mmap_sem write side: so preventing KSM from working on an mm while exit_mmap runs. And KSM will bail out as soon as it notices that mm_users is already zero, thanks to its internal ksm_test_exit checks. So that when a task is killed by OOM killer or the user, KSM will not indefinitely prevent it from running exit_mmap to release its memory. This does break a part of what "ksm: fix oom deadlock" was trying to achieve. When unmerging KSM (echo 2 >/sys/kernel/mm/ksm), and even when ksmd itself has to cancel a KSM page, it is possible that the first OOM-kill victim would be the KSM process being faulted: then its memory won't be freed until a second victim has been selected (freeing memory for the unmerging fault to complete). But the OOM killer is already liable to kill a second victim once the intended victim's p->mm goes to NULL: so there's not much point in rejecting this KSM patch before fixing that OOM behaviour. It is very much more important to allow KSM users to boot up, than to haggle over an unlikely and poorly supported OOM case. We also intend to fix munlocking to not fault pages: at which point this patch _could_ be reverted; though that would be controversial, so we hope to find a better solution. Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Acked-by: Justin M. Forbes <jforbes@redhat.com> Acked-for-now-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: fix oom deadlockHugh Dickins
There's a now-obvious deadlock in KSM's out-of-memory handling: imagine ksmd or KSM_RUN_UNMERGE handling, holding ksm_thread_mutex, trying to allocate a page to break KSM in an mm which becomes the OOM victim (quite likely in the unmerge case): it's killed and goes to exit, and hangs there waiting to acquire ksm_thread_mutex. Clearly we must not require ksm_thread_mutex in __ksm_exit, simple though that made everything else: perhaps use mmap_sem somehow? And part of the answer lies in the comments on unmerge_ksm_pages: __ksm_exit should also leave all the rmap_item removal to ksmd. But there's a fundamental problem, that KSM relies upon mmap_sem to guarantee the consistency of the mm it's dealing with, yet exit_mmap tears down an mm without taking mmap_sem. And bumping mm_users won't help at all, that just ensures that the pages the OOM killer assumes are on their way to being freed will not be freed. The best answer seems to be, to move the ksm_exit callout from just before exit_mmap, to the middle of exit_mmap: after the mm's pages have been freed (if the mmu_gather is flushed), but before its page tables and vma structures have been freed; and down_write,up_write mmap_sem there to serialize with KSM's own reliance on mmap_sem. But KSM then needs to be careful, whenever it downs mmap_sem, to check that the mm is not already exiting: there's a danger of using find_vma on a layout that's being torn apart, or writing into page tables which have been freed for reuse; and even do_anonymous_page and __do_fault need to check they're not being called by break_ksm to reinstate a pte after zap_pte_range has zapped that page table. Though it might be clearer to add an exiting flag, set while holding mmap_sem in __ksm_exit, that wouldn't cover the issue of reinstating a zapped pte. All we need is to check whether mm_users is 0 - but must remember that ksmd may detect that before __ksm_exit is reached. So, ksm_test_exit(mm) added to comment such checks on mm->mm_users. __ksm_exit now has to leave clearing up the rmap_items to ksmd, that needs ksm_thread_mutex; but shift the exiting mm just after the ksm_scan cursor so that it will soon be dealt with. __ksm_enter raise mm_count to hold the mm_struct, ksmd's exit processing (exactly like its processing when it finds all VM_MERGEABLEs unmapped) mmdrop it, similar procedure for KSM_RUN_UNMERGE (which has stopped ksmd). But also give __ksm_exit a fast path: when there's no complication (no rmap_items attached to mm and it's not at the ksm_scan cursor), it can safely do all the exiting work itself. This is not just an optimization: when ksmd is not running, the raised mm_count would otherwise leak mm_structs. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: distribute remove_mm_from_listsHugh Dickins
Do some housekeeping in ksm.c, to help make the next patch easier to understand: remove the function remove_mm_from_lists, distributing its code to its callsites scan_get_next_rmap_item and __ksm_exit. That turns out to be a win in scan_get_next_rmap_item: move its remove_trailing_rmap_items and cursor advancement up, and it becomes simpler than before. __ksm_exit becomes messier, but will change again; and moving its remove_trailing_rmap_items up lets us strengthen the unstable tree item's age condition in remove_rmap_item_from_tree. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: fix endless loop on oomHugh Dickins
break_ksm has been looping endlessly ignoring VM_FAULT_OOM: that should only be a problem for ksmd when a memory control group imposes limits (normally the OOM killer will kill others with an mm until it succeeds); but in general (especially for MADV_UNMERGEABLE and KSM_RUN_UNMERGE) we do need to route the error (or kill) back to the caller (or sighandling). Test signal_pending in unmerge_ksm_pages, which could be a lengthy procedure if it has to spill into swap: returning -ERESTARTSYS so that trivial signals will restart but fatals will terminate (is that right? we do different things in different places in mm, none exactly this). unmerge_and_remove_all_rmap_items was forgetting to lock when going down the mm_list: fix that. Whether it's successful or not, reset ksm_scan cursor to head; but only if it's successful, reset seqnr (shown in full_scans) - page counts will have gone down to zero. This patch leaves a significant OOM deadlock, but it's a good step on the way, and that deadlock is fixed in a subsequent patch. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: five little cleanupsHugh Dickins
1. We don't use __break_cow entry point now: merge it into break_cow. 2. remove_all_slot_rmap_items is just a special case of remove_trailing_rmap_items: use the latter instead. 3. Extend comment on unmerge_ksm_pages and rmap_items. 4. try_to_merge_two_pages should use try_to_merge_with_ksm_page instead of duplicating its code; and so swap them around. 5. Comment on cmp_and_merge_page described last year's: update it. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: keep quiet while list emptyHugh Dickins
ksm_scan_thread already sleeps in wait_event_interruptible until setting ksm_run activates it; but if there's nothing on its list to look at, i.e. nobody has yet said madvise MADV_MERGEABLE, it's a shame to be clocking up system time and full_scans: ksmd_should_run added to check that too. And move the mutex_lock out around it: the new counts showed that when ksm_run is stopped, a little work often got done afterwards, because it had been read before taking the mutex. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: break cow once unsharedHugh Dickins
We kept agreeing not to bother about the unswappable shared KSM pages which later become unshared by others: observation suggests they're not a significant proportion. But they are disadvantageous, and it is easier to break COW to replace them by swappable pages, than offer statistics to show that they don't matter; then we can stop worrying about them. Doing this in ksm_do_scan, they don't go through cmp_and_merge_page on this pass: give them a good chance of getting into the unstable tree on the next pass, or back into the stable, by computing checksum now. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22ksm: pages_unshared and pages_volatileHugh Dickins
The pages_shared and pages_sharing counts give a good picture of how successful KSM is at sharing; but no clue to how much wasted work it's doing to get there. Add pages_unshared (count of unique pages waiting in the unstable tree, hoping to find a mate) and pages_volatile. pages_volatile is harder to define. It includes those pages changing too fast to get into the unstable tree, but also whatever other edge conditions prevent a page getting into the trees: a high value may deserve investigation. Don't try to calculate it from the various conditions: it's the total of rmap_items less those accounted for. Also show full_scans: the number of completed scans of everything registered in the mm list. The locking for all these counts is simply ksm_thread_mutex. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Izik Eidus <ieidus@redhat.com> Acked-by: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>