- 09 Oct, 2012 40 commits
-
-
Shaohua Li authored
The x86 implementation of atomic_dec_if_positive is quite generic, so make it available to all architectures. This is needed for "swap: add a simple detector for inappropriate swapin readahead". [akpm@linux-foundation.org: do the "#define foo foo" trick in the conventional manner] Signed-off-by:
Shaohua Li <shli@fusionio.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: "David S. Miller" <davem@davemloft.net> Cc: Rik van Riel <riel@redhat.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Michal Simek <monstr@monstr.eu> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Minchan Kim authored
If race between allocation and isolation in memory-hotplug offline happens, some pages could be in MIGRATE_MOVABLE of free_list although the pageblock's migratetype of the page is MIGRATE_ISOLATE. The race could be detected by get_freepage_migratetype in __test_page_isolated_in_pageblock. If it is detected, now EBUSY gets bubbled all the way up and the hotplug operations fails. But better idea is instead of returning and failing memory-hotremove, move the free page to the correct list at the time it is detected. It could enhance memory-hotremove operation success ratio although the race is really rare. Suggested by Mel Gorman. [akpm@linux-foundation.org: small cleanup] Signed-off-by:
Minchan Kim <minchan@kernel.org> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by:
Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Acked-by:
Mel Gorman <mgorman@suse.de> Cc: Xishi Qiu <qiuxishi@huawei.com> Cc: Wen Congyang <wency@cn.fujitsu.com> Signed-off-by:
-
Minchan Kim authored
Like below, memory-hotplug makes race between page-isolation and page-allocation so it can hit BUG_ON in __offline_isolated_pages. CPU A CPU B start_isolate_page_range set_migratetype_isolate spin_lock_irqsave(zone->lock) free_hot_cold_page(Page A) /* without zone->lock */ migratetype = get_pageblock_migratetype(Page A); /* * Page could be moved into MIGRATE_MOVABLE * of per_cpu_pages */ list_add_tail(&page->lru, &pcp->lists[migratetype]); set_pageblock_isolate move_freepages_block drain_all_pages /* Page A could be in MIGRATE_MOVABLE of free_list. */ check_pages_isolated __test_page_isolated_in_pageblock /* * We can't catch freed page which * is free_list[MIGRATE_MOVABLE] */ if (PageBuddy(page A)) pfn += 1 << page_order(page A); /* So, Page A could be allocated */ __offline_isolated_pages /* * BUG_ON hit or offline page * which is used by someone */ BUG_ON(!PageBuddy(page A)); This patch checks page's migratetype in freelist in __test_page_isolated_in_pageblock. So now __test_page_isolated_in_pageblock can check the page caused by above race and can fail of memory offlining. Signed-off-by:
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by:
-
Minchan Kim authored
The page allocator caches the pageblock information in page->private while it is in the PCP freelists but this is overwritten with the order of the page when freed to the buddy allocator. This patch stores the migratetype of the page in the page->index field so that it is available at all times when the page remain in free_list. This patch adds a new call site in __free_pages_ok so it might be overhead a bit but it's for high order allocation. So I believe damage isn't hurt. Signed-off-by:
-
Minchan Kim authored
The page allocator uses set_page_private and page_private for handling migratetype when it frees page. Let's replace them with [set|get] _freepage_migratetype to make it more clear. Signed-off-by:
-
Bartlomiej Zolnierkiewicz authored
* Add ALLOC_CMA alloc flag and pass it to [__]zone_watermark_ok() (from Minchan Kim). * During watermark check decrease available free pages number by free CMA pages number if necessary (unmovable allocations cannot use pages from CMA areas). Signed-off-by:
Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Signed-off-by:
Kyungmin Park <kyungmin.park@samsung.com> Cc: Marek Szyprowski <m.szyprowski@samsung.com> Cc: Michal Nazarewicz <mina86@mina86.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Hugh Dickins <hughd@google.com> Signed-off-by:
-
Bartlomiej Zolnierkiewicz authored
Add NR_FREE_CMA_PAGES counter to be later used for checking watermark in __zone_watermark_ok(). For simplicity and to avoid #ifdef hell make this counter always available (not only when CONFIG_CMA=y). [akpm@linux-foundation.org: use conventional migratetype naming] Signed-off-by:
-
Bartlomiej Zolnierkiewicz authored
Isolated free pages shouldn't be accounted to NR_FREE_PAGES counter. Fix it by properly decreasing/increasing NR_FREE_PAGES counter in set_migratetype_isolate()/unset_migratetype_isolate() and removing counter adjustment for isolated pages from free_one_page() and split_free_page(). Signed-off-by:
-
Bartlomiej Zolnierkiewicz authored
page->private gets re-used in __free_one_page() to store page order (so trace_mm_page_pcpu_drain() may print order instead of migratetype) thus migratetype value must be cached locally. Fixes regression introduced in commit a7016235 ("mm: fix migratetype bug which slowed swapping"). This caused incorrect data to be attached to the mm_page_pcpu_drain trace event. [akpm@linux-foundation.org: add comment] Cc: Marek Szyprowski <m.szyprowski@samsung.com> Cc: Michal Nazarewicz <mina86@mina86.com> Acked-by:
KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by:
-
Minchan Kim authored
Drop clean cache pages instead of migration during alloc_contig_range() to minimise allocation latency by reducing the amount of migration that is necessary. It's useful for CMA because latency of migration is more important than evicting the background process's working set. In addition, as pages are reclaimed then fewer free pages for migration targets are required so it avoids memory reclaiming to get free pages, which is a contributory factor to increased latency. I measured elapsed time of __alloc_contig_migrate_range() which migrates 10M in 40M movable zone in QEMU machine. Before - 146ms, After - 7ms [akpm@linux-foundation.org: fix nommu build] Signed-off-by:
Michal Nazarewicz <mina86@mina86.com> Cc: Rik van Riel <riel@redhat.com> Tested-by:
-
Andrea Arcangeli authored
The variable must be static especially given the variable name. s/RCU/SRCU/ over a few comments. Signed-off-by:
Andrea Arcangeli <aarcange@redhat.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Sagi Grimberg <sagig@mellanox.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Haggai Eran <haggaie@mellanox.com> Signed-off-by:
-
Xishi Qiu authored
online_pages() does build_all_zonelists() and zone_pcp_update(), I think offline_pages() should do it too. When the zone has no memory to allocate, remove it from other nodes' zonelists. zone_batchsize() depends on zone's present pages, if zone's present pages are changed, zone's pcp should be updated. Signed-off-by:
Xishi Qiu <qiuxishi@huawei.com> Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Signed-off-by:
-
Michel Lespinasse authored
During mremap(), the destination VMA is generally placed after the original vma in rmap traversal order: in move_vma(), we always have new_pgoff >= vma->vm_pgoff, and as a result new_vma->vm_pgoff >= vma->vm_pgoff unless vma_merge() merged the new vma with an adjacent one. When the destination VMA is placed after the original in rmap traversal order, we can avoid taking the rmap locks in move_ptes(). Essentially, this reintroduces the optimization that had been disabled in "mm anon rmap: remove anon_vma_moveto_tail". The difference is that we don't try to impose the rmap traversal order; instead we just rely on things being in the desired order in the common case and fall back to taking locks in the uncommon case. Also we skip the i_mmap_mutex in addition to the anon_vma lock: in both cases, the vmas are traversed in increasing vm_pgoff order with ties resolved in tree insertion order. Signed-off-by:
Michel Lespinasse <walken@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Daniel Santos <daniel.santos@pobox.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by:
-
Michel Lespinasse authored
anon_vma_clone() expects new_vma->vm_{start,end,pgoff} to be correctly set so that the new vma can be indexed on the anon interval tree. copy_vma() was failing to do that, which broke mremap(). Signed-off-by:Sasha Levin <levinsasha928@gmail.com> Signed-off-by:
-
Michel Lespinasse authored
Add a CONFIG_DEBUG_VM_RB build option for the previously existing DEBUG_MM_RB code. Now that Andi Kleen modified it to avoid using recursive algorithms, we can expose it a bit more. Also extend this code to validate_mm() after stack expansion, and to check that the vma's start and last pgoffs have not changed since the nodes were inserted on the anon vma interval tree (as it is important that the nodes be reindexed after each such update). Signed-off-by:
-
Michel Lespinasse authored
In file and anon rmap, we use interval trees to find potentially relevant vmas and then call vma_address() to find the virtual address the given page might be found at in these vmas. vma_address() used to include a check that the returned address falls within the limits of the vma, but this check isn't necessary now that we always use interval trees in rmap: the interval tree just doesn't return any vmas which this check would find to be irrelevant. As a result, we can replace the use of -EFAULT error code (which then needed to be checked in every call site) with a VM_BUG_ON(). Signed-off-by:
-
Michel Lespinasse authored
When a large VMA (anon or private file mapping) is first touched, which will populate its anon_vma field, and then split into many regions through the use of mprotect(), the original anon_vma ends up linking all of the vmas on a linked list. This can cause rmap to become inefficient, as we have to walk potentially thousands of irrelevent vmas before finding the one a given anon page might fall into. By replacing the same_anon_vma linked list with an interval tree (where each avc's interval is determined by its vma's start and last pgoffs), we can make rmap efficient for this use case again. While the change is large, all of its pieces are fairly simple. Most places that were walking the same_anon_vma list were looking for a known pgoff, so they can just use the anon_vma_interval_tree_foreach() interval tree iterator instead. The exception here is ksm, where the page's index is not known. It would probably be possible to rework ksm so that the index would be known, but for now I have decided to keep things simple and just walk the entirety of the interval tree there. When updating vma's that already have an anon_vma assigned, we must take care to re-index the corresponding avc's on their interval tree. This is done through the use of anon_vma_interval_tree_pre_update_vma() and anon_vma_interval_tree_post_update_vma(), which remove the avc's from their interval tree before the update and re-insert them after the update. The anon_vma stays locked during the update, so there is no chance that rmap would miss the vmas that are being updated. Signed-off-by:
-
Michel Lespinasse authored
mremap() had a clever optimization where move_ptes() did not take the anon_vma lock to avoid a race with anon rmap users such as page migration. Instead, the avc's were ordered in such a way that the origin vma was always visited by rmap before the destination. This ordering and the use of page table locks rmap usage safe. However, we want to replace the use of linked lists in anon rmap with an interval tree, and this will make it harder to impose such ordering as the interval tree will always be sorted by the avc->vma->vm_pgoff value. For now, let's replace the anon_vma_moveto_tail() ordering function with proper anon_vma locking in move_ptes(). Once we have the anon interval tree in place, we will re-introduce an optimization to avoid taking these locks in the most common cases. Signed-off-by:
-
Michel Lespinasse authored
Update the generic interval tree code that was introduced in "mm: replace vma prio_tree with an interval tree". Changes: - fixed 'endpoing' typo noticed by Andrew Morton - replaced include/linux/interval_tree_tmpl.h, which was used as a template (including it automatically defined the interval tree functions) with include/linux/interval_tree_generic.h, which only defines a preprocessor macro INTERVAL_TREE_DEFINE(), which itself defines the interval tree functions when invoked. Now that is a very long macro which is unfortunate, but it does make the usage sites (lib/interval_tree.c and mm/interval_tree.c) a bit nicer than previously. - make use of RB_DECLARE_CALLBACKS() in the INTERVAL_TREE_DEFINE() macro, instead of duplicating that code in the interval tree template. - replaced vma_interval_tree_add(), which was actually handling the nonlinear and interval tree cases, with vma_interval_tree_insert_after() which handles only the interval tree case and has an API that is more consistent with the other interval tree handling functions. The nonlinear case is now handled explicitly in kernel/fork.c dup_mmap(). Signed-off-by:
-
Michel Lespinasse authored
Provide rb_insert_augmented() and rb_erase_augmented() through a new rbtree_augmented.h include file. rb_erase_augmented() is defined there as an __always_inline function, in order to allow inlining of augmented rbtree callbacks into it. Since this generates a relatively large function, each augmented rbtree user should make sure to have a single call site. Signed-off-by:
-
Michel Lespinasse authored
After both prio_tree users have been converted to use red-black trees, there is no need to keep around the prio tree library anymore. Signed-off-by:
-
Michel Lespinasse authored
kmemleak uses a tree where each node represents an allocated memory object in order to quickly find out what object a given address is part of. However, the objects don't overlap, so rbtrees are a better choice than prio tree for this use. They are both faster and have lower memory overhead. Tested by booting a kernel with kmemleak enabled, loading the kmemleak_test module, and looking for the expected messages. Signed-off-by:
Catalin Marinas <catalin.marinas@arm.com> Tested-by:
-
Michel Lespinasse authored
Implement an interval tree as a replacement for the VMA prio_tree. The algorithms are similar to lib/interval_tree.c; however that code can't be directly reused as the interval endpoints are not explicitly stored in the VMA. So instead, the common algorithm is moved into a template and the details (node type, how to get interval endpoints from the node, etc) are filled in using the C preprocessor. Once the interval tree functions are available, using them as a replacement to the VMA prio tree is a relatively simple, mechanical job. Signed-off-by:
-
Michel Lespinasse authored
Patch 1 implements support for interval trees, on top of the augmented rbtree API. It also adds synthetic tests to compare the performance of interval trees vs prio trees. Short answers is that interval trees are slightly faster (~25%) on insert/erase, and much faster (~2.4 - 3x) on search. It is debatable how realistic the synthetic test is, and I have not made such measurements yet, but my impression is that interval trees would still come out faster. Patch 2 uses a preprocessor template to make the interval tree generic, and uses it as a replacement for the vma prio_tree. Patch 3 takes the other prio_tree user, kmemleak, and converts it to use a basic rbtree. We don't actually need the augmented rbtree support here because the intervals are always non-overlapping. Patch 4 removes the now-unused prio tree library. Patch 5 proposes an additional optimization to rb_erase_augmented, now providing it as an inline function so that the augmented callbacks can be inlined in. This provides an additional 5-10% performance improvement for the interval tree insert/erase benchmark. There is a maintainance cost as it exposes augmented rbtree users to some of the rbtree library internals; however I think this cost shouldn't be too high as I expect the augmented rbtree will always have much less users than the base rbtree. I should probably add a quick summary of why I think it makes sense to replace prio trees with augmented rbtree based interval trees now. One of the drivers is that we need augmented rbtrees for Rik's vma gap finding code, and once you have them, it just makes sense to use them for interval trees as well, as this is the simpler and more well known algorithm. prio trees, in comparison, seem *too* clever: they impose an additional 'heap' constraint on the tree, which they use to guarantee a faster worst-case complexity of O(k+log N) for stabbing queries in a well-balanced prio tree, vs O(k*log N) for interval trees (where k=number of matches, N=number of intervals). Now this sounds great, but in practice prio trees don't realize this theorical benefit. First, the additional constraint makes them harder to update, so that the kernel implementation has to simplify things by balancing them like a radix tree, which is not always ideal. Second, the fact that there are both index and heap properties makes both tree manipulation and search more complex, which results in a higher multiplicative time constant. As it turns out, the simple interval tree algorithm ends up running faster than the more clever prio tree. This patch: Add two test modules: - prio_tree_test measures the performance of lib/prio_tree.c, both for insertion/removal and for stabbing searches - interval_tree_test measures the performance of a library of equivalent functionality, built using the augmented rbtree support. In order to support the second test module, lib/interval_tree.c is introduced. It is kept separate from the interval_tree_test main file for two reasons: first we don't want to provide an unfair advantage over prio_tree_test by having everything in a single compilation unit, and second there is the possibility that the interval tree functionality could get some non-test users in kernel over time. Signed-off-by:
-
Michel Lespinasse authored
As proposed by Peter Zijlstra, this makes it easier to define the augmented rbtree callbacks. Signed-off-by:
-
Michel Lespinasse authored
convert arch/x86/mm/pat_rbtree.c to the proposed augmented rbtree api and remove the old augmented rbtree implementation. Signed-off-by:
Rik van Riel <riel@redhat.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: David Woodhouse <dwmw2@infradead.org> Signed-off-by:
-
Michel Lespinasse authored
Introduce new augmented rbtree APIs that allow minimal recalculation of augmented node information. A new callback is added to the rbtree insertion and erase rebalancing functions, to be called on each tree rotations. Such rotations preserve the subtree's root augmented value, but require recalculation of the one child that was previously located at the subtree root. In the insertion case, the handcoded search phase must be updated to maintain the augmented information on insertion, and then the rbtree coloring/rebalancing algorithms keep it up to date. In the erase case, things are more complicated since it is library code that manipulates the rbtree in order to remove internal nodes. This requires a couple additional callbacks to copy a subtree's augmented value when a new root is stitched in, and to recompute augmented values down the ancestry path when a node is removed from the tree. In order to preserve maximum speed for the non-augmented case, we provide two versions of each tree manipulation function. rb_insert_augmented() is the augmented equivalent of rb_insert_color(), and rb_erase_augmented() is the augmented equivalent of rb_erase(). Signed-off-by:
-
Michel Lespinasse authored
Small test to measure the performance of augmented rbtrees. Signed-off-by:
-
Michel Lespinasse authored
Various minor optimizations in rb_erase(): - Avoid multiple loading of node->__rb_parent_color when computing parent and color information (possibly not in close sequence, as there might be further branches in the algorithm) - In the 1-child subcase of case 1, copy the __rb_parent_color field from the erased node to the child instead of recomputing it from the desired parent and color - When searching for the erased node's successor, differentiate between cases 2 and 3 based on whether any left links were followed. This avoids a condition later down. - In case 3, keep a pointer to the erased node's right child so we don't have to refetch it later to adjust its parent. - In the no-childs subcase of cases 2 and 3, place the rebalance assigment last so that the compiler can remove the following if(rebalance) test. Also, added some comments to illustrate cases 2 and 3. Signed-off-by:
-
Michel Lespinasse authored
An interesting observation for rb_erase() is that when a node has exactly one child, the node must be black and the child must be red. An interesting consequence is that removing such a node can be done by simply replacing it with its child and making the child black, which we can do efficiently in rb_erase(). __rb_erase_color() then only needs to handle the no-childs case and can be modified accordingly. Signed-off-by:
-
Michel Lespinasse authored
In rb_erase, move the easy case (node to erase has no more than 1 child) first. I feel the code reads easier that way. Signed-off-by:
-
Michel Lespinasse authored
Add __rb_change_child() as an inline helper function to replace code that would otherwise be duplicated 4 times in the source. No changes to binary size or speed. Signed-off-by:
-
Michel Lespinasse authored
Just a small fix to make sparse happy. Signed-off-by:
Fengguang Wu <wfg@linux.intel.com> Acked-by:
-
Michel Lespinasse authored
When looking to fetch a node's sibling, we went through a sequence of: - check if node is the parent's left child - if it is, then fetch the parent's right child This can be replaced with: - fetch the parent's right child as an assumed sibling - check that node is NOT the fetched child This avoids fetching the parent's left child when node is actually that child. Saves a bit on code size, though it doesn't seem to make a large difference in speed. Signed-off-by:
-
Michel Lespinasse authored
Set comment and indentation style to be consistent with linux coding style and the rest of the file, as suggested by Peter Zijlstra Signed-off-by:
David Woodhouse <David.Woodhouse@intel.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Daniel Santos <daniel.santos@pobox.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by:
-
Michel Lespinasse authored
In __rb_erase_color(), we often already have pointers to the nodes being rotated and/or know what their colors must be, so we can generate more efficient code than the generic __rb_rotate_left() and __rb_rotate_right() functions. Also when the current node is red or when flipping the sibling's color, the parent is already known so we can use the more efficient rb_set_parent_color() function to set the desired color. Signed-off-by:
-
Michel Lespinasse authored
In __rb_erase_color(), we have to select one of 3 cases depending on the color on the 'other' node children. If both children are black, we flip a few node colors and iterate. Otherwise, we do either one or two tree rotations, depending on the color of the 'other' child opposite to 'node', and then we are done. The corresponding logic had duplicate checks for the color of the 'other' child opposite to 'node'. It was checking it first to determine if both children are black, and then to determine how many tree rotations are required. Rearrange the logic to avoid that extra check. Signed-off-by:
-
Michel Lespinasse authored
In __rb_erase_color(), we were always setting a node to black after exiting the main loop. And in one case, after fixing up the tree to satisfy all rbtree invariants, we were setting the current node to root just to guarantee a loop exit, at which point the root would be set to black. However this is not necessary, as the root of an rbtree is already known to be black. The only case where the color flip is required is when we exit the loop due to the current node being red, and it's easiest to just do the flip at that point instead of doing it after the loop. [adrian.hunter@intel.com: perf tools: fix build for another rbtree.c change] Signed-off-by:
Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@intel.com> Signed-off-by:
-
Michel Lespinasse authored
- Use the newly introduced rb_set_parent_color() function to flip the color of nodes whose parent is already known. - Optimize rb_parent() when the node is known to be red - there is no need to mask out the color in that case. - Flipping gparent's color to red requires us to fetch its rb_parent_color field, so we can reuse it as the parent value for the next loop iteration. - Do not use __rb_rotate_left() and __rb_rotate_right() to handle tree rotations: we already have pointers to all relevant nodes, and know their colors (either because we want to adjust it, or because we've tested it, or we can deduce it as black due to the node proximity to a known red node). So we can generate more efficient code by making use of the node pointers we already have, and setting both the parent and color attributes for nodes all at once. Also in Case 2, some node attributes don't have to be set because we know another tree rotation (Case 3) will always follow and override them. Signed-off-by:
-
Michel Lespinasse authored
The root node of an rbtree must always be black. However, rb_insert_color() only needs to maintain this invariant when it has been broken - that is, when it exits the loop due to the current (red) node being the root. In all other cases (exiting after tree rotations, or exiting due to an existing black parent) the invariant is already satisfied, so there is no need to adjust the root node color. Signed-off-by:
-
