- 16 Oct, 2007 40 commits
-
-
Adam Litke authored
The maximum size of the huge page pool can be controlled using the overall size of the hugetlb filesystem (via its 'size' mount option). However in the common case the this will not be set as the pool is traditionally fixed in size at boot time. In order to maintain the expected semantics, we need to prevent the pool expanding by default. This patch introduces a new sysctl controlling dynamic pool resizing. When this is enabled the pool will expand beyond its base size up to the size of the hugetlb filesystem. It is disabled by default. Signed-off-by:
Adam Litke <agl@us.ibm.com> Acked-by:
Andy Whitcroft <apw@shadowen.org> Acked-by:
Dave McCracken <dave.mccracken@oracle.com> Cc: William Irwin <bill.irwin@oracle.com> Cc: David Gibson <david@gibson.dropbear.id.au> Cc: Ken Chen <kenchen@google.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Adam Litke authored
Shared mappings require special handling because the huge pages needed to fully populate the VMA must be reserved at mmap time. If not enough pages are available when making the reservation, allocate all of the shortfall at once from the buddy allocator and add the pages directly to the hugetlb pool. If they cannot be allocated, then fail the mapping. The page surplus is accounted for in the same way as for private mappings; faulted surplus pages will be freed at unmap time. Reserved, surplus pages that have not been used must be freed separately when their reservation has been released. Signed-off-by:
-
Adam Litke authored
Because we overcommit hugepages for MAP_PRIVATE mappings, it is possible that the hugetlb pool will be exhausted or completely reserved when a hugepage is needed to satisfy a page fault. Before killing the process in this situation, try to allocate a hugepage directly from the buddy allocator. The explicitly configured pool size becomes a low watermark. When dynamically grown, the allocated huge pages are accounted as a surplus over the watermark. As huge pages are freed on a node, surplus pages are released to the buddy allocator so that the pool will shrink back to the watermark. Surplus accounting also allows for friendlier explicit pool resizing. When shrinking a pool that is fully in-use, increase the surplus so pages will be returned to the buddy allocator as soon as they are freed. When growing a pool that has a surplus, consume the surplus first and then allocate new pages. Signed-off-by:
Mel Gorman <mel@csn.ul.ie> Acked-by:
-
Adam Litke authored
Dynamic huge page pool resizing. In most real-world scenarios, configuring the size of the hugetlb pool correctly is a difficult task. If too few pages are allocated to the pool, applications using MAP_SHARED may fail to mmap() a hugepage region and applications using MAP_PRIVATE may receive SIGBUS. Isolating too much memory in the hugetlb pool means it is not available for other uses, especially those programs not using huge pages. The obvious answer is to let the hugetlb pool grow and shrink in response to the runtime demand for huge pages. The work Mel Gorman has been doing to establish a memory zone for movable memory allocations makes dynamically resizing the hugetlb pool reliable within the limits of that zone. This patch series implements dynamic pool resizing for private and shared mappings while being careful to maintain existing semantics. Please reply with your comments and feedback; even just to say whether it would be a useful feature to you. Thanks. How it works ============ Upon depletion of the hugetlb pool, rather than reporting an error immediately, first try and allocate the needed huge pages directly from the buddy allocator. Care must be taken to avoid unbounded growth of the hugetlb pool, so the hugetlb filesystem quota is used to limit overall pool size. The real work begins when we decide there is a shortage of huge pages. What happens next depends on whether the pages are for a private or shared mapping. Private mappings are straightforward. At fault time, if alloc_huge_page() fails, we allocate a page from the buddy allocator and increment the source node's surplus_huge_pages counter. When free_huge_page() is called for a page on a node with a surplus, the page is freed directly to the buddy allocator instead of the hugetlb pool. Because shared mappings require all of the pages to be reserved up front, some additional work must be done at mmap() to support them. We determine the reservation shortage and allocate the required number of pages all at once. These pages are then added to the hugetlb pool and marked reserved. Where that is not possible the mmap() will fail. As with private mappings, the appropriate surplus counters are updated. Since reserved huge pages won't necessarily be used by the process, we can't be sure that free_huge_page() will always be called to return surplus pages to the buddy allocator. To prevent the huge page pool from bloating, we must free unused surplus pages when their reservation has ended. Controlling it ============== With the entire patch series applied, pool resizing is off by default so unless specific action is taken, the semantics are unchanged. To take advantage of the flexibility afforded by this patch series one must tolerate a change in semantics. To control hugetlb pool growth, the following techniques can be employed: * A sysctl tunable to enable/disable the feature entirely * The size= mount option for hugetlbfs filesystems to limit pool size Performance =========== When contiguous memory is readily available, it is expected that the cost of dynamicly resizing the pool will be small. This series has been performance tested with 'stream' to measure this cost. Stream (http://www.cs.virginia.edu/stream/) was linked with libhugetlbfs to enable remapping of the text and data/bss segments into huge pages. Stream with small array ----------------------- Baseline: nr_hugepages = 0, No libhugetlbfs segment remapping Preallocated: nr_hugepages = 5, Text and data/bss remapping Dynamic: nr_hugepages = 0, Text and data/bss remapping Rate (MB/s) Function Baseline Preallocated Dynamic Copy: 4695.6266 5942.8371 5982.2287 Scale: 4451.5776 5017.1419 5658.7843 Add: 5815.8849 7927.7827 8119.3552 Triad: 5949.4144 8527.6492 8110.6903 Stream with large array ----------------------- Baseline: nr_hugepages = 0, No libhugetlbfs segment remapping Preallocated: nr_hugepages = 67, Text and data/bss remapping Dynamic: nr_hugepages = 0, Text and data/bss remapping Rate (MB/s) Function Baseline Preallocated Dynamic Copy: 2227.8281 2544.2732 2546.4947 Scale: 2136.3208 2430.7294 2421.2074 Add: 2773.1449 4004.0021 3999.4331 Triad: 2748.4502 3777.0109 3773.4970 * All numbers are averages taken from 10 consecutive runs with a maximum standard deviation of 1.3 percent noted. This patch: Simply move update_and_free_page() so that it can be reused later in this patch series. The implementation is not changed. Signed-off-by:
William Irwin <bill.irwin@oracle.com> Cc: David Gibson <david@gibson.dropbear.id.au> Cc: Ken Chen <kenchen@google.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Signed-off-by:
-
Yasunori Goto authored
This patch is to avoid panic when memory hot-add is executed with sparsemem-vmemmap. Current vmemmap-sparsemem code doesn't support memory hot-add. Vmemmap must be populated when hot-add. This is for 2.6.23-rc2-mm2. Todo: # Even if this patch is applied, the message "[xxxx-xxxx] potential offnode page_structs" is displayed. To allocate memmap on its node, memmap (and pgdat) must be initialized itself like chicken and egg relationship. # vmemmap_unpopulate will be necessary for followings. - For cancel hot-add due to error. - For unplug. Signed-off-by:Yasunori Goto <y-goto@jp.fujitsu.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Christoph Lameter <clameter@sgi.com> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by:
-
KAMEZAWA Hiroyuki authored
Now, arch dependent code around CONFIG_MEMORY_HOTREMOVE is a mess. This patch cleans up them. This is against 2.6.23-rc6-mm1. - fix compile failure on ia64/ CONFIG_MEMORY_HOTPLUG && !CONFIG_MEMORY_HOTREMOVE case. - For !CONFIG_MEMORY_HOTREMOVE, add generic no-op remove_memory(), which returns -EINVAL. - removed remove_pages() only used in powerpc. - removed no-op remove_memory() in i386, sh, sparc64, x86_64. - only powerpc returns -ENOSYS at memory hot remove(no-op). changes it to return -EINVAL. Note: Currently, only ia64 supports CONFIG_MEMORY_HOTREMOVE. I welcome other archs if there are requirements and testers. Signed-off-by:
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by:
-
KAMEZAWA Hiroyuki authored
IA64 memory unplug interface. Signed-off-by:
-
KAMEZAWA Hiroyuki authored
Logic. - set all pages in [start,end) as isolated migration-type. by this, all free pages in the range will be not-for-use. - Migrate all LRU pages in the range. - Test all pages in the range's refcnt is zero or not. Todo: - allocate migration destination page from better area. - confirm page_count(page)== 0 && PageReserved(page) page is safe to be freed.. (I don't like this kind of page but.. - Find out pages which cannot be migrated. - more running tests. - Use reclaim for unplugging other memory type area. Signed-off-by:
-
KAMEZAWA Hiroyuki authored
Implement generic chunk-of-pages isolation method by using page grouping ops. This patch add MIGRATE_ISOLATE to MIGRATE_TYPES. By this - MIGRATE_TYPES increases. - bitmap for migratetype is enlarged. pages of MIGRATE_ISOLATE migratetype will not be allocated even if it is free. By this, you can isolated *freed* pages from users. How-to-free pages is not a purpose of this patch. You may use reclaim and migrate codes to free pages. If start_isolate_page_range(start,end) is called, - migratetype of the range turns to be MIGRATE_ISOLATE if its type is MIGRATE_MOVABLE. (*) this check can be updated if other memory reclaiming works make progress. - MIGRATE_ISOLATE is not on migratetype fallback list. - All free pages and will-be-freed pages are isolated. To check all pages in the range are isolated or not, use test_pages_isolated(), To cancel isolation, use undo_isolate_page_range(). Changes V6 -> V7 - removed unnecessary #ifdef There are HOLES_IN_ZONE handling codes...I'm glad if we can remove them.. Signed-off-by:
-
KAMEZAWA Hiroyuki authored
A clean up patch for "scanning memory resource [start, end)" operation. Now, find_next_system_ram() function is used in memory hotplug, but this interface is not easy to use and codes are complicated. This patch adds walk_memory_resouce(start,len,arg,func) function. The function 'func' is called per valid memory resouce range in [start,pfn). [pbadari@us.ibm.com: Error handling in walk_memory_resource()] Signed-off-by:
Badari Pulavarty <pbadari@us.ibm.com> Signed-off-by:
-
Mel Gorman authored
The statistics patch later needs to know what order a free page is on the free lists. Rather than having special knowledge of page_private() when PageBuddy() is set, this patch places out page_order() in internal.h and adds a VM_BUG_ON to catch using it on non-PageBuddy pages. Signed-off-by:
Christoph Lameter <clameter@sgi.com> Acked-by:
-
Andrew Morton authored
It's a short-lived allocation. Cc: Christoph Lameter <clameter@sgi.com> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by:
-
Christoph Lameter authored
We touch a cacheline in the kmem_cache structure for zeroing to get the size. However, the hot paths in slab_alloc and slab_free do not reference any other fields in kmem_cache, so we may have to just bring in the cacheline for this one access. Add a new field to kmem_cache_cpu that contains the object size. That cacheline must already be used in the hotpaths. So we save one cacheline on every slab_alloc if we zero. We need to update the kmem_cache_cpu object size if an aliasing operation changes the objsize of an non debug slab. Signed-off-by:
-
Christoph Lameter authored
The kmem_cache_cpu structures introduced are currently an array placed in the kmem_cache struct. Meaning the kmem_cache_cpu structures are overwhelmingly on the wrong node for systems with a higher amount of nodes. These are performance critical structures since the per node information has to be touched for every alloc and free in a slab. In order to place the kmem_cache_cpu structure optimally we put an array of pointers to kmem_cache_cpu structs in kmem_cache (similar to SLAB). However, the kmem_cache_cpu structures can now be allocated in a more intelligent way. We would like to put per cpu structures for the same cpu but different slab caches in cachelines together to save space and decrease the cache footprint. However, the slab allocators itself control only allocations per node. We set up a simple per cpu array for every processor with 100 per cpu structures which is usually enough to get them all set up right. If we run out then we fall back to kmalloc_node. This also solves the bootstrap problem since we do not have to use slab allocator functions early in boot to get memory for the small per cpu structures. Pro: - NUMA aware placement improves memory performance - All global structures in struct kmem_cache become readonly - Dense packing of per cpu structures reduces cacheline footprint in SMP and NUMA. - Potential avoidance of exclusive cacheline fetches on the free and alloc hotpath since multiple kmem_cache_cpu structures are in one cacheline. This is particularly important for the kmalloc array. Cons: - Additional reference to one read only cacheline (per cpu array of pointers to kmem_cache_cpu) in both slab_alloc() and slab_free(). [akinobu.mita@gmail.com: fix cpu hotplug offline/online path] Signed-off-by:
-
Christoph Lameter authored
Set c->node to -1 if we allocate from a debug slab instead for SlabDebug which requires access the page struct cacheline. Signed-off-by:
Alexey Dobriyan <adobriyan@sw.ru> Signed-off-by:
-
Christoph Lameter authored
We need the offset from the page struct during slab_alloc and slab_free. In both cases we also reference the cacheline of the kmem_cache_cpu structure. We can therefore move the offset field into the kmem_cache_cpu structure freeing up 16 bits in the page struct. Moving the offset allows an allocation from slab_alloc() without touching the page struct in the hot path. The only thing left in slab_free() that touches the page struct cacheline for per cpu freeing is the checking of SlabDebug(page). The next patch deals with that. Use the available 16 bits to broaden page->inuse. More than 64k objects per slab become possible and we can get rid of the checks for that limitation. No need anymore to shrink the order of slabs if we boot with 2M sized slabs (slub_min_order=9). No need anymore to switch off the offset calculation for very large slabs since the field in the kmem_cache_cpu structure is 32 bits and so the offset field can now handle slab sizes of up to 8GB. Signed-off-by:
-
Christoph Lameter authored
After moving the lockless_freelist to kmem_cache_cpu we no longer need page->lockless_freelist. Restructure the use of the struct page fields in such a way that we never touch the mapping field. This is turn allows us to remove the special casing of SLUB when determining the mapping of a page (needed for corner cases of virtual caches machines that need to flush caches of processors mapping a page). Signed-off-by:
-
Christoph Lameter authored
A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by:
-
Adrian Bunk authored
This patch makes needlessly global code static. Signed-off-by:
Adrian Bunk <bunk@stusta.de> Signed-off-by:
-
Mel Gorman authored
This patch provides fragmentation avoidance statistics via /proc/pagetypeinfo. The information is collected only on request so there is no runtime overhead. The statistics are in three parts: The first part prints information on the size of blocks that pages are being grouped on and looks like Page block order: 10 Pages per block: 1024 The second part is a more detailed version of /proc/buddyinfo and looks like Free pages count per migrate type at order 0 1 2 3 4 5 6 7 8 9 10 Node 0, zone DMA, type Unmovable 0 0 0 0 0 0 0 0 0 0 0 Node 0, zone DMA, type Reclaimable 1 0 0 0 0 0 0 0 0 0 0 Node 0, zone DMA, type Movable 0 0 0 0 0 0 0 0 0 0 0 Node 0, zone DMA, type Reserve 0 4 4 0 0 0 0 1 0 1 0 Node 0, zone Normal, type Unmovable 111 8 4 4 2 3 1 0 0 0 0 Node 0, zone Normal, type Reclaimable 293 89 8 0 0 0 0 0 0 0 0 Node 0, zone Normal, type Movable 1 6 13 9 7 6 3 0 0 0 0 Node 0, zone Normal, type Reserve 0 0 0 0 0 0 0 0 0 0 4 The third part looks like Number of blocks type Unmovable Reclaimable Movable Reserve Node 0, zone DMA 0 1 2 1 Node 0, zone Normal 3 17 94 4 To walk the zones within a node with interrupts disabled, walk_zones_in_node() is introduced and shared between /proc/buddyinfo, /proc/zoneinfo and /proc/pagetypeinfo to reduce code duplication. It seems specific to what vmstat.c requires but could be broken out as a general utility function in mmzone.c if there were other other potential users. Signed-off-by:
-
Mel Gorman authored
Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by:
-
Mel Gorman authored
move_freepages_block() returns the number of blocks moved. This value is used to determine if a block of pages should be stolen for the exclusive use of a migrate type or not. However, the value returned is being used correctly. This patch fixes the calculation to return the number of base pages that have been moved. This should be considered a fix to the patch move-free-pages-between-lists-on-steal.patch Credit to Andy Whitcroft for spotting the problem. Signed-off-by:
-
Mel Gorman authored
Grouping high-order atomic allocations together was intended to allow bursty users of atomic allocations to work such as e1000 in situations where their preallocated buffers were depleted. This did not work in at least one case with a wireless network adapter needing order-1 allocations frequently. To resolve that, the free pages used for min_free_kbytes were moved to separate contiguous blocks with the patch bias-the-location-of-pages-freed-for-min_free_kbytes-in-the-same-max_order_nr_pages-blocks. It is felt that keeping the free pages in the same contiguous blocks should be sufficient for bursty short-lived high-order atomic allocations to succeed, maybe even with the e1000. Even if there is a failure, increasing the value of min_free_kbytes will free pages as contiguous bloks in contrast to the standard buddy allocator which makes no attempt to keep the minimum number of free pages contiguous. This patch backs out grouping high order atomic allocations together to determine if it is really needed or not. If a new report comes in about high-order atomic allocations failing, the feature can be reintroduced to determine if it fixes the problem or not. As a side-effect, this patch reduces by 1 the number of bits required to track the mobility type of pages within a MAX_ORDER_NR_PAGES block. Signed-off-by:
-
Mel Gorman authored
Grouping pages by mobility can be disabled at compile-time. This was considered undesirable by a number of people. However, in the current stack of patches, it is not a simple case of just dropping the configurable patch as it would cause merge conflicts. This patch backs out the configuration option. Signed-off-by:
-
Mel Gorman authored
The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by:
-
Mel Gorman authored
There are problems in the use of SPARSEMEM and pageblock flags that causes problems on ia64. The first part of the problem is that units are incorrect in SECTION_BLOCKFLAGS_BITS computation. This results in a map_section's section_mem_map being treated as part of a bitmap which isn't good. This was evident with an invalid virtual address when mem_init attempted to free bootmem pages while relinquishing control from the bootmem allocator. The second part of the problem occurs because the pageblock flags bitmap is be located with the mem_section. The SECTIONS_PER_ROOT computation using sizeof (mem_section) may not be a power of 2 depending on the size of the bitmap. This renders masks and other such things not power of 2 base. This issue was seen with SPARSEMEM_EXTREME on ia64. This patch moves the bitmap outside of mem_section and uses a pointer instead in the mem_section. The bitmaps are allocated when the section is being initialised. Note that sparse_early_usemap_alloc() does not use alloc_remap() like sparse_early_mem_map_alloc(). The allocation required for the bitmap on x86, the only architecture that uses alloc_remap is typically smaller than a cache line. alloc_remap() pads out allocations to the cache size which would be a needless waste. Credit to Bob Picco for identifying the original problem and effecting a fix for the SECTION_BLOCKFLAGS_BITS calculation. Credit to Andy Whitcroft for devising the best way of allocating the bitmaps only when required for the section. [wli@holomorphy.com: warning fix] Signed-off-by:
Bob Picco <bob.picco@hp.com> Signed-off-by:
-
Mel Gorman authored
MIGRATE_RECLAIMABLE allocations tend to be very bursty in nature like when updatedb starts. It is likely this will occur in situations where MAX_ORDER blocks of pages are not free. This means that updatedb can scatter MIGRATE_RECLAIMABLE pages throughout the address space. This patch is more agressive about stealing blocks of pages for MIGRATE_RECLAIMABLE. Signed-off-by:
-
Mel Gorman authored
This patch chooses blocks with lower PFNs when placing kernel allocations. This is particularly important during fallback in low memory situations to stop unmovable pages being placed throughout the entire address space. Signed-off-by:
-
Mel Gorman authored
Grouping pages by mobility can only successfully operate when there are more MAX_ORDER_NR_PAGES areas than mobility types. When there are insufficient areas, fallbacks cannot be avoided. This has noticeable performance impacts on machines with small amounts of memory in comparison to MAX_ORDER_NR_PAGES. For example, on IA64 with a configuration including huge pages spans 1GiB with MAX_ORDER_NR_PAGES so would need at least 4GiB of RAM before grouping pages by mobility would be useful. In comparison, an x86 would need 16MB. This patch checks the size of vm_total_pages in build_all_zonelists(). If there are not enough areas, mobility is effectivly disabled by considering all allocations as the same type (UNMOVABLE). This is achived via a __read_mostly flag. With this patch, performance is comparable to disabling grouping pages by mobility at compile-time on a test machine with insufficient memory. With this patch, it is reasonable to get rid of grouping pages by mobility a compile-time option. Signed-off-by:
-
Mel Gorman authored
In rare cases, the kernel needs to allocate a high-order block of pages without sleeping. For example, this is the case with e1000 cards configured to use jumbo frames. Migrating or reclaiming pages in this situation is not an option. This patch groups these allocations together as much as possible by adding a new MIGRATE_TYPE. The MIGRATE_HIGHATOMIC type are exactly what they sound like. Care is taken that pages of other migrate types do not use the same blocks as high-order atomic allocations. Signed-off-by:
-
Mel Gorman authored
This patch marks a number of allocations that are either short-lived such as network buffers or are reclaimable such as inode allocations. When something like updatedb is called, long-lived and unmovable kernel allocations tend to be spread throughout the address space which increases fragmentation. This patch groups these allocations together as much as possible by adding a new MIGRATE_TYPE. The MIGRATE_RECLAIMABLE type is for allocations that can be reclaimed on demand, but not moved. i.e. they can be migrated by deleting them and re-reading the information from elsewhere. Signed-off-by:
-
Mel Gorman authored
When a fallback occurs, there will be free pages for one allocation type stored on the list for another. When a large steal occurs, this patch will move all the free pages within one list to the other. [y-goto@jp.fujitsu.com: fix BUG_ON check at move_freepages()] [apw@shadowen.org: Move to using pfn_valid_within()] Signed-off-by:
Andy Whitcroft <andyw@uk.ibm.com> Cc: Bob Picco <bob.picco@hp.com> Signed-off-by:
-
Mel Gorman authored
Per-cpu pages can accidentally cause fragmentation because they are free, but pinned pages in an otherwise contiguous block. When this patch is applied, the per-cpu caches are drained after the direct-reclaim is entered if the requested order is greater than 0. It simply reuses the code used by suspend and hotplug. Signed-off-by:
-
Mel Gorman authored
The grouping mechanism has some memory overhead and a more complex allocation path. This patch allows the strategy to be disabled for small memory systems or if it is known the workload is suffering because of the strategy. It also acts to show where the page groupings strategy interacts with the standard buddy allocator. Signed-off-by:
Joel Schopp <jschopp@austin.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by:
-
Mel Gorman authored
The freelists for each migrate type can slowly become polluted due to the per-cpu list. Consider what happens when the following happens 1. A 2^(MAX_ORDER-1) list is reserved for __GFP_MOVABLE pages 2. An order-0 page is allocated from the newly reserved block 3. The page is freed and placed on the per-cpu list 4. alloc_page() is called with GFP_KERNEL as the gfp_mask 5. The per-cpu list is used to satisfy the allocation This results in a kernel page is in the middle of a migratable region. This patch prevents this leak occuring by storing the MIGRATE_ type of the page in page->private. On allocate, a page will only be returned of the desired type, else more pages will be allocated. This may temporarily allow a per-cpu list to go over the pcp->high limit but it'll be corrected on the next free. Care is taken to preserve the hotness of pages recently freed. The additional code is not measurably slower for the workloads we've tested. Signed-off-by:
-
Mel Gorman authored
This patch adds the core of the fragmentation reduction strategy. It works by grouping pages together based on their ability to migrate or be reclaimed. Basically, it works by breaking the list in zone->free_area list into MIGRATE_TYPES number of lists. Signed-off-by:
-
Mel Gorman authored
Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by:
-
KAMEZAWA Hiroyuki authored
Current ia64 kernel flushes icache by lazy_mmu_prot_update() *after* set_pte(). This is too late. This patch removes lazy_mmu_prot_update and add modfied set_pte() for flushing if necessary. This patch flush icache of a page when new pte has exec bit. && new pte has present bit && new pte is user's page. && (old *ptep is not present || new pte's pfn is not same to old *ptep's ptn) && new pte's page has no Pg_arch_1 bit. Pg_arch_1 is set when a page is cache consistent. I think this condition checks are much easier to understand than considering "Where sync_icache_dcache() should be inserted ?". pte_user() for ia64 was removed by http://lkml.org/lkml/2007/6/12/67 as clean-up. So, I added it again. Signed-off-by:David S. Miller <davem@davemloft.net> Signed-off-by:
-
KAMEZAWA Hiroyuki authored
In migration, a new page should be cache flushed before set_pte() in some archs which have virtually-tagged cache. Signed-off-by:
-
Andrea Arcangeli authored
Swappiness isn't a safe sysctl. Setting it to 0 for example can hang a system. That's a corner case but even setting it to 10 or lower can waste enormous amounts of cpu without making much progress. We've customers who wants to use swappiness but they can't because of the current implementation (if you change it so the system stops swapping it really stops swapping and nothing works sane anymore if you really had to swap something to make progress). This patch from Kurt Garloff makes swappiness safer to use (no more huge cpu usage or hangs with low swappiness values). I think the prev_priority can also be nuked since it wastes 4 bytes per zone (that would be an incremental patch but I wait the nr_scan_[in]active to be nuked first for similar reasons). Clearly somebody at some point noticed how broken that thing was and they had to add min(priority, prev_priority) to give it some reliability, but they didn't go the last mile to nuke prev_priority too. Calculating distress only in function of not-racy priority is correct and sure more than enough without having to add randomness into the equation. Patch is tested on older kernels but it compiles and it's quite simple so... Overall I'm not very satisified by the swappiness tweak, since it doesn't rally do anything with the dirty pagecache that may be inactive. We need another kind of tweak that controls the inactive scan and tunes the can_writepage feature (not yet in mainline despite having submitted it a few times), not only the active one. That new tweak will tell the kernel how hard to scan the inactive list for pure clean pagecache (something the mainline kernel isn't capable of yet). We already have that feature working in all our enterprise kernels with the default reasonable tune, or they can't even run a readonly backup with tar without triggering huge write I/O. I think it should be available also in mainline later. Cc: Nick Piggin <npiggin@suse.de> Signed-off-by:
Kurt Garloff <garloff@suse.de> Signed-off-by:
Andrea Arcangeli <andrea@suse.de> Signed-off-by:
Fengguang Wu <wfg@mail.ustc.edu.cn> Signed-off-by:
-
