- 15 Mar, 2016 40 commits
-
-
Vlastimil Babka authored
The following patch will need to declare array of struct trace_print_flags in a header. To prevent this header from pulling in all of RCU through trace_events.h, move the struct trace_print_flags{_64} definitions to the new lightweight tracepoint-defs.h header. Signed-off-by:Vlastimil Babka <vbabka@suse.cz> Acked-by:
David Rientjes <rientjes@google.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Mel Gorman authored
In the generic read paths the kernel looks up a page in the page cache and if it's up to date, it is used. If not, the page lock is acquired to wait for IO to complete and then check the page. If multiple processes are waiting on IO, they all serialise against the lock and duplicate the checks. This is unnecessary. The page lock in itself does not give any guarantees to the callers about the page state as it can be immediately truncated or reclaimed after the page is unlocked. It's sufficient to wait_on_page_locked and then continue if the page is up to date on wakeup. It is possible that a truncated but up-to-date page is returned but the reference taken during read prevents it disappearing underneath the caller and the data is still valid if PageUptodate. The overall impact is small as even if processes serialise on the lock, the lock section is tiny once the IO is complete. Profiles indicated that unlock_page and friends are generally a tiny portion of a read-intensive workload. An artificial test was created that had instances of dd access a cache-cold file on an ext4 filesystem and measure how long the read took. paralleldd 4.4.0 4.4.0 vanilla avoidlock Amean Elapsd-1 5.28 ( 0.00%) 5.15 ( 2.50%) Amean Elapsd-4 5.29 ( 0.00%) 5.17 ( 2.12%) Amean Elapsd-7 5.28 ( 0.00%) 5.18 ( 1.78%) Amean Elapsd-12 5.20 ( 0.00%) 5.33 ( -2.50%) Amean Elapsd-21 5.14 ( 0.00%) 5.21 ( -1.41%) Amean Elapsd-30 5.30 ( 0.00%) 5.12 ( 3.38%) Amean Elapsd-48 5.78 ( 0.00%) 5.42 ( 6.21%) Amean Elapsd-79 6.78 ( 0.00%) 6.62 ( 2.46%) Amean Elapsd-110 9.09 ( 0.00%) 8.99 ( 1.15%) Amean Elapsd-128 10.60 ( 0.00%) 10.43 ( 1.66%) The impact is small but intuitively, it makes sense to avoid unnecessary calls to lock_page. Signed-off-by:Mel Gorman <mgorman@techsingularity.net> Reviewed-by:
Jan Kara <jack@suse.cz> Cc: Hugh Dickins <hughd@google.com> Signed-off-by:
-
Mel Gorman authored
do_read_cache_page and __read_cache_page duplicate page filler code when filling the page for the first time. This patch simply removes the duplicate logic. Signed-off-by:
-
Andreas Ziegler authored
Commit 4b94ffdc ("x86, mm: introduce vmem_altmap to augment vmemmap_populate()"), introduced the to_vmem_altmap() function. The comments in this function contain two typos (one misspelling of the Kconfig option CONFIG_SPARSEMEM_VMEMMAP, and one missing letter 'n'), let's fix them up. Signed-off-by:
Andreas Ziegler <andreas.ziegler@fau.de> Signed-off-by:
-
Christian Borntraeger authored
Since commit 031bc574 ("mm/debug-pagealloc: make debug-pagealloc boottime configurable") CONFIG_DEBUG_PAGEALLOC is by default not adding any page debugging. This resulted in several unnoticed bugs, e.g. https://lkml.kernel.org/g/<569F5E29.3090107@de.ibm.com> or https://lkml.kernel.org/g/<56A20F30.4050705@de.ibm.com> as this behaviour change was not even documented in Kconfig. Let's provide a new Kconfig symbol that allows to change the default back to enabled, e.g. for debug kernels. This also makes the change obvious to kernel packagers. Let's also change the Kconfig description for CONFIG_DEBUG_PAGEALLOC, to indicate that there are two stages of overhead. Signed-off-by:
Christian Borntraeger <borntraeger@de.ibm.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Signed-off-by:
-
Andrey Ryabinin authored
Calculation of dirty_ratelimit sometimes is not correct. E.g. initial values of dirty_ratelimit == INIT_BW and step == 0, lead to the following result: UBSAN: Undefined behaviour in ../mm/page-writeback.c:1286:7 shift exponent 25600 is too large for 64-bit type 'long unsigned int' The fix is straightforward - make step 0 if the shift exponent is too big. Signed-off-by:
Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Tejun Heo <tj@kernel.org> Cc: Andy Shevchenko <andy.shevchenko@gmail.com> Signed-off-by:
-
Andrew Morton authored
This function is getting full of weird tricks to avoid word-wrapping. Use a goto to eliminate a tab stop then use the new space Cc: Taku Izumi <izumi.taku@jp.fujitsu.com> Signed-off-by:
-
Taku Izumi authored
This patch extends existing "kernelcore" option and introduces kernelcore=mirror option. By specifying "mirror" instead of specifying the amount of memory, non-mirrored (non-reliable) region will be arranged into ZONE_MOVABLE. [akpm@linux-foundation.org: fix build with CONFIG_HAVE_MEMBLOCK_NODE_MAP=n] Signed-off-by:
Taku Izumi <izumi.taku@jp.fujitsu.com> Tested-by:
Sudeep Holla <sudeep.holla@arm.com> Cc: Tony Luck <tony.luck@intel.com> Cc: Xishi Qiu <qiuxishi@huawei.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Steve Capper <steve.capper@linaro.org> Signed-off-by:
-
Taku Izumi authored
Xeon E7 v3 based systems supports Address Range Mirroring and UEFI BIOS complied with UEFI spec 2.5 can notify which ranges are mirrored (reliable) via EFI memory map. Now Linux kernel utilize its information and allocates boot time memory from reliable region. My requirement is: - allocate kernel memory from mirrored region - allocate user memory from non-mirrored region In order to meet my requirement, ZONE_MOVABLE is useful. By arranging non-mirrored range into ZONE_MOVABLE, mirrored memory is used for kernel allocations. My idea is to extend existing "kernelcore" option and introduces kernelcore=mirror option. By specifying "mirror" instead of specifying the amount of memory, non-mirrored region will be arranged into ZONE_MOVABLE. Earlier discussions are at: https://lkml.org/lkml/2015/10/9/24 https://lkml.org/lkml/2015/10/15/9 https://lkml.org/lkml/2015/11/27/18 https://lkml.org/lkml/2015/12/8/836 For example, suppose 2-nodes system with the following memory range: node 0 [mem 0x0000000000001000-0x000000109fffffff] node 1 [mem 0x00000010a0000000-0x000000209fffffff] and the following ranges are marked as reliable (mirrored): [0x0000000000000000-0x0000000100000000] [0x0000000100000000-0x0000000180000000] [0x0000000800000000-0x0000000880000000] [0x00000010a0000000-0x0000001120000000] [0x00000017a0000000-0x0000001820000000] If you specify kernelcore=mirror, ZONE_NORMAL and ZONE_MOVABLE are arranged like bellow: - node 0: ZONE_NORMAL : [0x0000000100000000-0x00000010a0000000] ZONE_MOVABLE: [0x0000000180000000-0x00000010a0000000] - node 1: ZONE_NORMAL : [0x00000010a0000000-0x00000020a0000000] ZONE_MOVABLE: [0x0000001120000000-0x00000020a0000000] In overlapped range, pages to be ZONE_MOVABLE in ZONE_NORMAL are treated as absent pages, and vice versa. This patch (of 2): Currently each zone's zone_start_pfn is calculated at free_area_init_core(). However zone's range is fixed at the time when invoking zone_spanned_pages_in_node(). This patch changes how each zone->zone_start_pfn is calculated in zone_spanned_pages_in_node(). Signed-off-by:
-
Andrew Morton authored
Signed-off-by:
-
Joonsoo Kim authored
SLUB already has a redzone debugging feature. But it is only positioned at the end of object (aka right redzone) so it cannot catch left oob. Although current object's right redzone acts as left redzone of next object, first object in a slab cannot take advantage of this effect. This patch explicitly adds a left red zone to each object to detect left oob more precisely. Background: Someone complained to me that left OOB doesn't catch even if KASAN is enabled which does page allocation debugging. That page is out of our control so it would be allocated when left OOB happens and, in this case, we can't find OOB. Moreover, SLUB debugging feature can be enabled without page allocator debugging and, in this case, we will miss that OOB. Before trying to implement, I expected that changes would be too complex, but, it doesn't look that complex to me now. Almost changes are applied to debug specific functions so I feel okay. Signed-off-by:
Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by:
-
Laura Abbott authored
When debug options are enabled, cmpxchg on the page is disabled. This is because the page must be locked to ensure there are no false positives when performing consistency checks. Some debug options such as poisoning and red zoning only act on the object itself. There is no need to protect other CPUs from modification on only the object. Allow cmpxchg to happen with poisoning and red zoning are set on a slab. Credit to Mathias Krause for the original work which inspired this series Signed-off-by:
Laura Abbott <labbott@fedoraproject.org> Acked-by:
Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <js1304@gmail.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mathias Krause <minipli@googlemail.com> Signed-off-by:
-
Laura Abbott authored
SLAB_DEBUG_FREE allows expensive consistency checks at free to be turned on or off. Expand its use to be able to turn off all consistency checks. This gives a nice speed up if you only want features such as poisoning or tracing. Credit to Mathias Krause for the original work which inspired this series Signed-off-by:
-
Laura Abbott authored
Since commit 19c7ff9e ("slub: Take node lock during object free checks") check_object has been incorrectly returning success as it follows the out label which just returns the node. Thanks to refactoring, the out and fail paths are now basically the same. Combine the two into one and just use a single label. Credit to Mathias Krause for the original work which inspired this series Signed-off-by:
-
Laura Abbott authored
This series takes the suggestion of Christoph Lameter and only focuses on optimizing the slow path where the debug processing runs. The two main optimizations in this series are letting the consistency checks be skipped and relaxing the cmpxchg restrictions when we are not doing consistency checks. With hackbench -g 20 -l 1000 averaged over 100 runs: Before slub_debug=P mean 15.607 variance .086 stdev .294 After slub_debug=P mean 10.836 variance .155 stdev .394 This still isn't as fast as what is in grsecurity unfortunately so there's still work to be done. Profiling ___slab_alloc shows that 25-50% of time is spent in deactivate_slab. I haven't looked too closely to see if this is something that can be optimized. My plan for now is to focus on getting all of this merged (if appropriate) before digging in to another task. This patch (of 4): Currently, free_debug_processing has a comment "Keep node_lock to preserve integrity until the object is actually freed". In actuallity, the lock is dropped immediately in __slab_free. Rather than wait until __slab_free and potentially throw off the unlikely marking, just drop the lock in __slab_free. This also lets free_debug_processing take its own copy of the spinlock flags rather than trying to share the ones from __slab_free. Since there is no use for the node afterwards, change the return type of free_debug_processing to return an int like alloc_debug_processing. Credit to Mathias Krause for the original work which inspired this series [akpm@linux-foundation.org: fix build] Signed-off-by:
-
Joonsoo Kim authored
Current implementation of pfmemalloc handling in SLAB has some problems. 1) pfmemalloc_active is set to true when there is just one or more pfmemalloc slabs in the system, but it is cleared when there is no pfmemalloc slab in one arbitrary kmem_cache. So, pfmemalloc_active could be wrongly cleared. 2) Search to partial and free list doesn't happen when non-pfmemalloc object are not found in cpu cache. Instead, allocating new slab happens and it is not optimal. 3) Even after sk_memalloc_socks() is disabled, cpu cache would keep pfmemalloc objects tagged with SLAB_OBJ_PFMEMALLOC. It isn't cleared if sk_memalloc_socks() is disabled so it could cause problem. 4) If cpu cache is filled with pfmemalloc objects, it would cause slow down non-pfmemalloc allocation. To me, current pointer tagging approach looks complex and fragile so this patch re-implement whole thing instead of fixing problems one by one. Design principle for new implementation is that 1) Don't disrupt non-pfmemalloc allocation in fast path even if sk_memalloc_socks() is enabled. It's more likely case than pfmemalloc allocation. 2) Ensure that pfmemalloc slab is used only for pfmemalloc allocation. 3) Don't consider performance of pfmemalloc allocation in memory deficiency state. As a result, all pfmemalloc alloc/free in memory tight state will be handled in slow-path. If there is non-pfmemalloc free object, it will be returned first even for pfmemalloc user in fast-path so that performance of pfmemalloc user isn't affected in normal case and pfmemalloc objects will be kept as long as possible. Signed-off-by:
-
Joonsoo Kim authored
Returing values by reference is bad practice. Instead, just use function return value. Signed-off-by:
-
Joonsoo Kim authored
SLAB needs an array to manage freed objects in a slab. It is only used if some objects are freed so we can use free object itself as this array. This requires additional branch in somewhat critical lock path to check if it is first freed object or not but that's all we need. Benefits is that we can save extra memory usage and reduce some computational overhead by allocating a management array when new slab is created. Code change is rather complex than what we can expect from the idea, in order to handle debugging feature efficiently. If you want to see core idea only, please remove '#if DEBUG' block in the patch. Although this idea can apply to all caches whose size is larger than management array size, it isn't applied to caches which have a constructor. If such cache's object is used for management array, constructor should be called for it before that object is returned to user. I guess that overhead overwhelm benefit in that case so this idea doesn't applied to them at least now. For summary, from now on, slab management type is determined by following logic. 1) if management array size is smaller than object size and no ctor, it becomes OBJFREELIST_SLAB. 2) if management array size is smaller than leftover, it becomes NORMAL_SLAB which uses leftover as a array. 3) if OFF_SLAB help to save memory than way 4), it becomes OFF_SLAB. It allocate a management array from the other cache so memory waste happens. 4) others become NORMAL_SLAB. It uses dedicated internal memory in a slab as a management array so it causes memory waste. In my system, without enabling CONFIG_DEBUG_SLAB, Almost caches become OBJFREELIST_SLAB and NORMAL_SLAB (using leftover) which doesn't waste memory. Following is the result of number of caches with specific slab management type. TOTAL = OBJFREELIST + NORMAL(leftover) + NORMAL + OFF /Before/ 126 = 0 + 60 + 25 + 41 /After/ 126 = 97 + 12 + 15 + 2 Result shows that number of caches that doesn't waste memory increase from 60 to 109. I did some benchmarking and it looks that benefit are more than loss. Kmalloc: Repeatedly allocate then free test /Before/ [ 0.286809] 1. Kmalloc: Repeatedly allocate then free test [ 1.143674] 100000 times kmalloc(32) -> 116 cycles kfree -> 78 cycles [ 1.441726] 100000 times kmalloc(64) -> 121 cycles kfree -> 80 cycles [ 1.815734] 100000 times kmalloc(128) -> 168 cycles kfree -> 85 cycles [ 2.380709] 100000 times kmalloc(256) -> 287 cycles kfree -> 95 cycles [ 3.101153] 100000 times kmalloc(512) -> 370 cycles kfree -> 117 cycles [ 3.942432] 100000 times kmalloc(1024) -> 413 cycles kfree -> 156 cycles [ 5.227396] 100000 times kmalloc(2048) -> 622 cycles kfree -> 248 cycles [ 7.519793] 100000 times kmalloc(4096) -> 1102 cycles kfree -> 452 cycles /After/ [ 1.205313] 100000 times kmalloc(32) -> 117 cycles kfree -> 78 cycles [ 1.510526] 100000 times kmalloc(64) -> 124 cycles kfree -> 81 cycles [ 1.827382] 100000 times kmalloc(128) -> 130 cycles kfree -> 84 cycles [ 2.226073] 100000 times kmalloc(256) -> 177 cycles kfree -> 92 cycles [ 2.814747] 100000 times kmalloc(512) -> 286 cycles kfree -> 112 cycles [ 3.532952] 100000 times kmalloc(1024) -> 344 cycles kfree -> 141 cycles [ 4.608777] 100000 times kmalloc(2048) -> 519 cycles kfree -> 210 cycles [ 6.350105] 100000 times kmalloc(4096) -> 789 cycles kfree -> 391 cycles In fact, I tested another idea implementing OBJFREELIST_SLAB with extendable linked array through another freed object. It can remove memory waste completely but it causes more computational overhead in critical lock path and it seems that overhead outweigh benefit. So, this patch doesn't include it. Signed-off-by:
-
Joonsoo Kim authored
cache_init_objs() will be changed in following patch and current form doesn't fit well for that change. So, before doing it, this patch separates debugging initialization. This would cause two loop iteration when debugging is enabled, but, this overhead seems too light than debug feature itself so effect may not be visible. This patch will greatly simplify changes in cache_init_objs() in following patch. Signed-off-by:
-
Joonsoo Kim authored
Slab list should be fixed up after object is detached from the slab and this happens at two places. They do exactly same thing. They will be changed in the following patch, so, to reduce code duplication, this patch factor out them and make it common function. Signed-off-by:
-
Joonsoo Kim authored
To become an off slab, there are some constraints to avoid bootstrapping problem and recursive call. This can be avoided differently by simply checking that corresponding kmalloc cache is ready and it's not a off slab. It would be more robust because static size checking can be affected by cache size change or architecture type but dynamic checking isn't. One check 'freelist_cache->size > cachep->size / 2' is added to check benefit of choosing off slab, because, now, there is no size constraint which ensures enough advantage when selecting off slab. Signed-off-by:
-
Joonsoo Kim authored
We can fail to setup off slab in some conditions. Even in this case, debug pagealloc increases cache size to PAGE_SIZE in advance and it is waste because debug pagealloc cannot work for it when it isn't the off slab. To improve this situation, this patch checks first that this cache with increased size is suitable for off slab. It actually increases cache size when it is suitable for off-slab, so possible waste is removed. Signed-off-by:
-
Joonsoo Kim authored
Current cache type determination code is open-code and looks not understandable. Following patch will introduce one more cache type and it would make code more complex. So, before it happens, this patch abstracts these codes. Signed-off-by:
-
Joonsoo Kim authored
Finding suitable OFF_SLAB candidate is more related to aligned cache size rather than original size. Same reasoning can be applied to the debug pagealloc candidate. So, this patch moves up alignment fixup to proper position. From that point, size is aligned so we can remove some alignment fixups. Signed-off-by:
-
Joonsoo Kim authored
Currently, the freelist is at the front of slab page. This requires extra space to meet object alignment requirement. If we put the freelist at the end of a slab page, objects could start at page boundary and will be at correct alignment. This is possible because freelist has no alignment constraint itself. This gives us two benefits: It removes extra memory space for the freelist alignment and remove complex calculation at cache initialization step. I can't think notable drawback here. I mentioned that this would reduce extra memory space, but, this benefit is rather theoretical because it can be applied to very few cases. Following is the example cache type that can get benefit from this change. size align num before after 32 8 124 4100 4092 64 8 63 4103 4095 88 8 46 4102 4094 272 8 15 4103 4095 408 8 10 4098 4090 32 16 124 4108 4092 64 16 63 4111 4095 32 32 124 4124 4092 64 32 63 4127 4095 96 32 42 4106 4074 before means whole size for objects and aligned freelist before applying patch and after shows the result of this patch. Since before is more than 4096, number of object should decrease and memory waste happens. Anyway, this patch removes complex calculation so looks beneficial to me. [akpm@linux-foundation.org: fix kerneldoc] Signed-off-by: -
Joonsoo Kim authored
Now, we don't use object status buffer in any setup. Remove it. Signed-off-by:
-
Joonsoo Kim authored
DEBUG_SLAB_LEAK is a debug option. It's current implementation requires status buffer so we need more memory to use it. And, it cause kmem_cache initialization step more complex. To remove this extra memory usage and to simplify initialization step, this patch implement this feature with another way. When user requests to get slab object owner information, it marks that getting information is started. And then, all free objects in caches are flushed to corresponding slab page. Now, we can distinguish all freed object so we can know all allocated objects, too. After collecting slab object owner information on allocated objects, mark is checked that there is no free during the processing. If true, we can be sure that our information is correct so information is returned to user. Although this way is rather complex, it has two important benefits mentioned above. So, I think it is worth changing. There is one drawback that it takes more time to get slab object owner information but it is just a debug option so it doesn't matter at all. To help review, this patch implements new way only. Following patch will remove useless code. Signed-off-by:
-
Joonsoo Kim authored
Currently, open code for checking DEBUG_PAGEALLOC cache is spread to some sites. It makes code unreadable and hard to change. This patch cleans up this code. The following patch will change the criteria for DEBUG_PAGEALLOC cache so this clean-up will help it, too. [akpm@linux-foundation.org: fix build with CONFIG_DEBUG_PAGEALLOC=n] Signed-off-by:
-
Joonsoo Kim authored
debug_pagealloc debugging is related to SLAB_POISON flag rather than FORCED_DEBUG option, although FORCED_DEBUG option will enable SLAB_POISON. Fix it. Signed-off-by:
-
Joonsoo Kim authored
Signed-off-by:
-
Joonsoo Kim authored
Some of "#if DEBUG" are for reporting slab implementation bug rather than user usecase bug. It's not really needed because slab is stable for a quite long time and it makes code too dirty. This patch remove it. Signed-off-by:
-
Joonsoo Kim authored
It is obsolete so remove it. Signed-off-by:
-
Joonsoo Kim authored
This patchset implements a new freed object management way, that is, OBJFREELIST_SLAB. Purpose of it is to reduce memory overhead in SLAB. SLAB needs a array to manage freed objects in a slab. If there is leftover after objects are packed into a slab, we can use it as a management array, and, in this case, there is no memory waste. But, in the other cases, we need to allocate extra memory for a management array or utilize dedicated internal memory in a slab for it. Both cases causes memory waste so it's not good. With this patchset, freed object itself can be used for a management array. So, memory waste could be reduced. Detailed idea and numbers are described in last patch's commit description. Please refer it. In fact, I tested another idea implementing OBJFREELIST_SLAB with extendable linked array through another freed object. It can remove memory waste completely but it causes more computational overhead in critical lock path and it seems that overhead outweigh benefit. So, this patchset doesn't include it. I will attach prototype just for a reference. This patch (of 16): We use freelist_idx_t type for free object management whose size would be smaller than size of unsigned int. Fix it. Signed-off-by:
-
Jesper Dangaard Brouer authored
Fix up trivial spelling errors, noticed while reading the code. Signed-off-by:
Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Signed-off-by:
-
Jesper Dangaard Brouer authored
This patch introduce a new API call kfree_bulk() for bulk freeing memory objects not bound to a single kmem_cache. Christoph pointed out that it is possible to implement freeing of objects, without knowing the kmem_cache pointer as that information is available from the object's page->slab_cache. Proposing to remove the kmem_cache argument from the bulk free API. Jesper demonstrated that these extra steps per object comes at a performance cost. It is only in the case CONFIG_MEMCG_KMEM is compiled in and activated runtime that these steps are done anyhow. The extra cost is most visible for SLAB allocator, because the SLUB allocator does the page lookup (virt_to_head_page()) anyhow. Thus, the conclusion was to keep the kmem_cache free bulk API with a kmem_cache pointer, but we can still implement a kfree_bulk() API fairly easily. Simply by handling if kmem_cache_free_bulk() gets called with a kmem_cache NULL pointer. This does increase the code size a bit, but implementing a separate kfree_bulk() call would likely increase code size even more. Below benchmarks cost of alloc+free (obj size 256 bytes) on CPU i7-4790K @ 4.00GHz, no PREEMPT and CONFIG_MEMCG_KMEM=y. Code size increase for SLAB: add/remove: 0/0 grow/shrink: 1/0 up/down: 74/0 (74) function old new delta kmem_cache_free_bulk 660 734 +74 SLAB fastpath: 87 cycles(tsc) 21.814 sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 103 cycles 25.878 ns - 41 cycles 10.498 ns - 81 cycles 20.312 ns 2 - 94 cycles 23.673 ns - 26 cycles 6.682 ns - 42 cycles 10.649 ns 3 - 92 cycles 23.181 ns - 21 cycles 5.325 ns - 39 cycles 9.950 ns 4 - 90 cycles 22.727 ns - 18 cycles 4.673 ns - 26 cycles 6.693 ns 8 - 89 cycles 22.270 ns - 14 cycles 3.664 ns - 23 cycles 5.835 ns 16 - 88 cycles 22.038 ns - 14 cycles 3.503 ns - 22 cycles 5.543 ns 30 - 89 cycles 22.284 ns - 13 cycles 3.310 ns - 20 cycles 5.197 ns 32 - 88 cycles 22.249 ns - 13 cycles 3.420 ns - 20 cycles 5.166 ns 34 - 88 cycles 22.224 ns - 14 cycles 3.643 ns - 20 cycles 5.170 ns 48 - 88 cycles 22.088 ns - 14 cycles 3.507 ns - 20 cycles 5.203 ns 64 - 88 cycles 22.063 ns - 13 cycles 3.428 ns - 20 cycles 5.152 ns 128 - 89 cycles 22.483 ns - 15 cycles 3.891 ns - 23 cycles 5.885 ns 158 - 89 cycles 22.381 ns - 15 cycles 3.779 ns - 22 cycles 5.548 ns 250 - 91 cycles 22.798 ns - 16 cycles 4.152 ns - 23 cycles 5.967 ns SLAB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 130 cycles(tsc) 32.684 ns (step:0) 1 - 148 cycles 37.220 ns - 66 cycles 16.622 ns - 66 cycles 16.583 ns 2 - 141 cycles 35.510 ns - 51 cycles 12.820 ns - 58 cycles 14.625 ns 3 - 140 cycles 35.017 ns - 37 cycles 9.326 ns - 33 cycles 8.474 ns 4 - 137 cycles 34.507 ns - 31 cycles 7.888 ns - 33 cycles 8.300 ns 8 - 140 cycles 35.069 ns - 25 cycles 6.461 ns - 25 cycles 6.436 ns 16 - 138 cycles 34.542 ns - 23 cycles 5.945 ns - 22 cycles 5.670 ns 30 - 136 cycles 34.227 ns - 22 cycles 5.502 ns - 22 cycles 5.587 ns 32 - 136 cycles 34.253 ns - 21 cycles 5.475 ns - 21 cycles 5.324 ns 34 - 136 cycles 34.254 ns - 21 cycles 5.448 ns - 20 cycles 5.194 ns 48 - 136 cycles 34.075 ns - 21 cycles 5.458 ns - 21 cycles 5.367 ns 64 - 135 cycles 33.994 ns - 21 cycles 5.350 ns - 21 cycles 5.259 ns 128 - 137 cycles 34.446 ns - 23 cycles 5.816 ns - 22 cycles 5.688 ns 158 - 137 cycles 34.379 ns - 22 cycles 5.727 ns - 22 cycles 5.602 ns 250 - 138 cycles 34.755 ns - 24 cycles 6.093 ns - 23 cycles 5.986 ns Code size increase for SLUB: function old new delta kmem_cache_free_bulk 717 799 +82 SLUB benchmark: SLUB fastpath: 46 cycles(tsc) 11.691 ns (step:0) sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 61 cycles 15.486 ns - 53 cycles 13.364 ns - 57 cycles 14.464 ns 2 - 54 cycles 13.703 ns - 32 cycles 8.110 ns - 33 cycles 8.482 ns 3 - 53 cycles 13.272 ns - 25 cycles 6.362 ns - 27 cycles 6.947 ns 4 - 51 cycles 12.994 ns - 24 cycles 6.087 ns - 24 cycles 6.078 ns 8 - 50 cycles 12.576 ns - 21 cycles 5.354 ns - 22 cycles 5.513 ns 16 - 49 cycles 12.368 ns - 20 cycles 5.054 ns - 20 cycles 5.042 ns 30 - 49 cycles 12.273 ns - 18 cycles 4.748 ns - 19 cycles 4.758 ns 32 - 49 cycles 12.401 ns - 19 cycles 4.821 ns - 19 cycles 4.810 ns 34 - 98 cycles 24.519 ns - 24 cycles 6.154 ns - 24 cycles 6.157 ns 48 - 83 cycles 20.833 ns - 21 cycles 5.446 ns - 21 cycles 5.429 ns 64 - 75 cycles 18.891 ns - 20 cycles 5.247 ns - 20 cycles 5.238 ns 128 - 93 cycles 23.271 ns - 27 cycles 6.856 ns - 27 cycles 6.823 ns 158 - 102 cycles 25.581 ns - 30 cycles 7.714 ns - 30 cycles 7.695 ns 250 - 107 cycles 26.917 ns - 38 cycles 9.514 ns - 38 cycles 9.506 ns SLUB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 71 cycles(tsc) 17.897 ns (step:0) 1 - 85 cycles 21.484 ns - 78 cycles 19.569 ns - 75 cycles 18.938 ns 2 - 81 cycles 20.363 ns - 45 cycles 11.258 ns - 44 cycles 11.076 ns 3 - 78 cycles 19.709 ns - 33 cycles 8.354 ns - 32 cycles 8.044 ns 4 - 77 cycles 19.430 ns - 28 cycles 7.216 ns - 28 cycles 7.003 ns 8 - 101 cycles 25.288 ns - 23 cycles 5.849 ns - 23 cycles 5.787 ns 16 - 76 cycles 19.148 ns - 20 cycles 5.162 ns - 20 cycles 5.081 ns 30 - 76 cycles 19.067 ns - 19 cycles 4.868 ns - 19 cycles 4.821 ns 32 - 76 cycles 19.052 ns - 19 cycles 4.857 ns - 19 cycles 4.815 ns 34 - 121 cycles 30.291 ns - 25 cycles 6.333 ns - 25 cycles 6.268 ns 48 - 108 cycles 27.111 ns - 21 cycles 5.498 ns - 21 cycles 5.458 ns 64 - 100 cycles 25.164 ns - 20 cycles 5.242 ns - 20 cycles 5.229 ns 128 - 155 cycles 38.976 ns - 27 cycles 6.886 ns - 27 cycles 6.892 ns 158 - 132 cycles 33.034 ns - 30 cycles 7.711 ns - 30 cycles 7.728 ns 250 - 130 cycles 32.612 ns - 38 cycles 9.560 ns - 38 cycles 9.549 ns Signed-off-by:
-
Jesper Dangaard Brouer authored
This patch implements the free side of bulk API for the SLAB allocator kmem_cache_free_bulk(), and concludes the implementation of optimized bulk API for SLAB allocator. Benchmarked[1] cost of alloc+free (obj size 256 bytes) on CPU i7-4790K @ 4.00GHz, with no debug options, no PREEMPT and CONFIG_MEMCG_KMEM=y but no active user of kmemcg. SLAB single alloc+free cost: 87 cycles(tsc) 21.814 ns with this optimized config. bulk- Current fallback - optimized SLAB bulk 1 - 102 cycles(tsc) 25.747 ns - 41 cycles(tsc) 10.490 ns - improved 59.8% 2 - 94 cycles(tsc) 23.546 ns - 26 cycles(tsc) 6.567 ns - improved 72.3% 3 - 92 cycles(tsc) 23.127 ns - 20 cycles(tsc) 5.244 ns - improved 78.3% 4 - 90 cycles(tsc) 22.663 ns - 18 cycles(tsc) 4.588 ns - improved 80.0% 8 - 88 cycles(tsc) 22.242 ns - 14 cycles(tsc) 3.656 ns - improved 84.1% 16 - 88 cycles(tsc) 22.010 ns - 13 cycles(tsc) 3.480 ns - improved 85.2% 30 - 89 cycles(tsc) 22.305 ns - 13 cycles(tsc) 3.303 ns - improved 85.4% 32 - 89 cycles(tsc) 22.277 ns - 13 cycles(tsc) 3.309 ns - improved 85.4% 34 - 88 cycles(tsc) 22.246 ns - 13 cycles(tsc) 3.294 ns - improved 85.2% 48 - 88 cycles(tsc) 22.121 ns - 13 cycles(tsc) 3.492 ns - improved 85.2% 64 - 88 cycles(tsc) 22.052 ns - 13 cycles(tsc) 3.411 ns - improved 85.2% 128 - 89 cycles(tsc) 22.452 ns - 15 cycles(tsc) 3.841 ns - improved 83.1% 158 - 89 cycles(tsc) 22.403 ns - 14 cycles(tsc) 3.746 ns - improved 84.3% 250 - 91 cycles(tsc) 22.775 ns - 16 cycles(tsc) 4.111 ns - improved 82.4% Notice it is not recommended to do very large bulk operation with this bulk API, because local IRQs are disabled in this period. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.cSigned-off-by:
-
Jesper Dangaard Brouer authored
Move the call to cache_alloc_debugcheck_after() outside the IRQ disabled section in kmem_cache_alloc_bulk(). When CONFIG_DEBUG_SLAB is disabled the compiler should remove this code. Signed-off-by:
-
Jesper Dangaard Brouer authored
This patch implements the alloc side of bulk API for the SLAB allocator. Further optimization are still possible by changing the call to __do_cache_alloc() into something that can return multiple objects. This optimization is left for later, given end results already show in the area of 80% speedup. Signed-off-by:
-
Jesper Dangaard Brouer authored
Reviewers notice that the order in slab_post_alloc_hook() of kmemcheck_slab_alloc() and kmemleak_alloc_recursive() gets swapped compared to slab.c / SLAB allocator. Also notice memset now occurs before calling kmemcheck_slab_alloc() and kmemleak_alloc_recursive(). I assume this reordering of kmemcheck, kmemleak and memset is okay because this is the order they are used by the SLUB allocator. This patch completes the sharing of alloc_hook's between SLUB and SLAB. Signed-off-by:
-
Jesper Dangaard Brouer authored
In the SLAB allocator kmemcheck_slab_alloc() is guarded against being called in case the object is NULL. In SLUB allocator this NULL pointer invocation can happen, which seems like an oversight. Move the NULL pointer check into kmemcheck code (kmemcheck_slab_alloc) so the check gets moved out of the fastpath, when not compiled with CONFIG_KMEMCHECK. This is a step towards sharing post_alloc_hook between SLUB and SLAB, because slab_post_alloc_hook() does not perform this check before calling kmemcheck_slab_alloc(). Signed-off-by:
-
