- 14 Jan, 2011 40 commits
-
-
Andrea Arcangeli authored
This documents how split_huge_page is safe vs new vma inserctions into the anon_vma that may have already released the anon_vma->lock but not established pmds yet when split_huge_page starts. Signed-off-by:
Andrea Arcangeli <aarcange@redhat.com> Acked-by:
Mel Gorman <mel@csn.ul.ie> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Hugh Dickins authored
If you configure THP in addition to HUGETLB_PAGE on x86_32 without PAE, the p?d-folding works out that munlock_vma_pages_range() can crash to follow_page()'s pud_huge() BUG_ON(flags & FOLL_GET): it needs the same VM_HUGETLB check already there on the pmd_huge() line. Conveniently, openSUSE provides a "blogd" which tests this out at startup! Signed-off-by:
Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by:
-
Andrea Arcangeli authored
Lately I've been working to make KVM use hugepages transparently without the usual restrictions of hugetlbfs. Some of the restrictions I'd like to see removed: 1) hugepages have to be swappable or the guest physical memory remains locked in RAM and can't be paged out to swap 2) if a hugepage allocation fails, regular pages should be allocated instead and mixed in the same vma without any failure and without userland noticing 3) if some task quits and more hugepages become available in the buddy, guest physical memory backed by regular pages should be relocated on hugepages automatically in regions under madvise(MADV_HUGEPAGE) (ideally event driven by waking up the kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes not null) 4) avoidance of reservation and maximization of use of hugepages whenever possible. Reservation (needed to avoid runtime fatal faliures) may be ok for 1 machine with 1 database with 1 database cache with 1 database cache size known at boot time. It's definitely not feasible with a virtualization hypervisor usage like RHEV-H that runs an unknown number of virtual machines with an unknown size of each virtual machine with an unknown amount of pagecache that could be potentially useful in the host for guest not using O_DIRECT (aka cache=off). hugepages in the virtualization hypervisor (and also in the guest!) are much more important than in a regular host not using virtualization, becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24 to 19 in case only the hypervisor uses transparent hugepages, and they decrease the tlb-miss cacheline accesses from 19 to 15 in case both the linux hypervisor and the linux guest both uses this patch (though the guest will limit the addition speedup to anonymous regions only for now...). Even more important is that the tlb miss handler is much slower on a NPT/EPT guest than for a regular shadow paging or no-virtualization scenario. So maximizing the amount of virtual memory cached by the TLB pays off significantly more with NPT/EPT than without (even if there would be no significant speedup in the tlb-miss runtime). The first (and more tedious) part of this work requires allowing the VM to handle anonymous hugepages mixed with regular pages transparently on regular anonymous vmas. This is what this patch tries to achieve in the least intrusive possible way. We want hugepages and hugetlb to be used in a way so that all applications can benefit without changes (as usual we leverage the KVM virtualization design: by improving the Linux VM at large, KVM gets the performance boost too). The most important design choice is: always fallback to 4k allocation if the hugepage allocation fails! This is the _very_ opposite of some large pagecache patches that failed with -EIO back then if a 64k (or similar) allocation failed... Second important decision (to reduce the impact of the feature on the existing pagetable handling code) is that at any time we can split an hugepage into 512 regular pages and it has to be done with an operation that can't fail. This way the reliability of the swapping isn't decreased (no need to allocate memory when we are short on memory to swap) and it's trivial to plug a split_huge_page* one-liner where needed without polluting the VM. Over time we can teach mprotect, mremap and friends to handle pmd_trans_huge natively without calling split_huge_page*. The fact it can't fail isn't just for swap: if split_huge_page would return -ENOMEM (instead of the current void) we'd need to rollback the mprotect from the middle of it (ideally including undoing the split_vma) which would be a big change and in the very wrong direction (it'd likely be simpler not to call split_huge_page at all and to teach mprotect and friends to handle hugepages instead of rolling them back from the middle). In short the very value of split_huge_page is that it can't fail. The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and incremental and it'll just be an "harmless" addition later if this initial part is agreed upon. It also should be noted that locking-wise replacing regular pages with hugepages is going to be very easy if compared to what I'm doing below in split_huge_page, as it will only happen when page_count(page) matches page_mapcount(page) if we can take the PG_lock and mmap_sem in write mode. collapse_huge_page will be a "best effort" that (unlike split_huge_page) can fail at the minimal sign of trouble and we can try again later. collapse_huge_page will be similar to how KSM works and the madvise(MADV_HUGEPAGE) will work similar to madvise(MADV_MERGEABLE). The default I like is that transparent hugepages are used at page fault time. This can be changed with /sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set to three values "always", "madvise", "never" which mean respectively that hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions, or never used. /sys/kernel/mm/transparent_hugepage/defrag instead controls if the hugepage allocation should defrag memory aggressively "always", only inside "madvise" regions, or "never". The pmd_trans_splitting/pmd_trans_huge locking is very solid. The put_page (from get_user_page users that can't use mmu notifier like O_DIRECT) that runs against a __split_huge_page_refcount instead was a pain to serialize in a way that would result always in a coherent page count for both tail and head. I think my locking solution with a compound_lock taken only after the page_first is valid and is still a PageHead should be safe but it surely needs review from SMP race point of view. In short there is no current existing way to serialize the O_DIRECT final put_page against split_huge_page_refcount so I had to invent a new one (O_DIRECT loses knowledge on the mapping status by the time gup_fast returns so...). And I didn't want to impact all gup/gup_fast users for now, maybe if we change the gup interface substantially we can avoid this locking, I admit I didn't think too much about it because changing the gup unpinning interface would be invasive. If we ignored O_DIRECT we could stick to the existing compound refcounting code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM (and any other mmu notifier user) would call it without FOLL_GET (and if FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the current task mmu notifier list yet). But O_DIRECT is fundamental for decent performance of virtualized I/O on fast storage so we can't avoid it to solve the race of put_page against split_huge_page_refcount to achieve a complete hugepage feature for KVM. Swap and oom works fine (well just like with regular pages ;). MMU notifier is handled transparently too, with the exception of the young bit on the pmd, that didn't have a range check but I think KVM will be fine because the whole point of hugepages is that EPT/NPT will also use a huge pmd when they notice gup returns pages with PageCompound set, so they won't care of a range and there's just the pmd young bit to check in that case. NOTE: in some cases if the L2 cache is small, this may slowdown and waste memory during COWs because 4M of memory are accessed in a single fault instead of 8k (the payoff is that after COW the program can run faster). So we might want to switch the copy_huge_page (and clear_huge_page too) to not temporal stores. I also extensively researched ways to avoid this cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k up to 1M (I can send those patches that fully implemented prefault) but I concluded they're not worth it and they add an huge additional complexity and they remove all tlb benefits until the full hugepage has been faulted in, to save a little bit of memory and some cache during app startup, but they still don't improve substantially the cache-trashing during startup if the prefault happens in >4k chunks. One reason is that those 4k pte entries copied are still mapped on a perfectly cache-colored hugepage, so the trashing is the worst one can generate in those copies (cow of 4k page copies aren't so well colored so they trashes less, but again this results in software running faster after the page fault). Those prefault patches allowed things like a pte where post-cow pages were local 4k regular anon pages and the not-yet-cowed pte entries were pointing in the middle of some hugepage mapped read-only. If it doesn't payoff substantially with todays hardware it will payoff even less in the future with larger l2 caches, and the prefault logic would blot the VM a lot. If one is emebdded transparent_hugepage can be disabled during boot with sysfs or with the boot commandline parameter transparent_hugepage=0 (or transparent_hugepage=2 to restrict hugepages inside madvise regions) that will ensure not a single hugepage is allocated at boot time. It is simple enough to just disable transparent hugepage globally and let transparent hugepages be allocated selectively by applications in the MADV_HUGEPAGE region (both at page fault time, and if enabled with the collapse_huge_page too through the kernel daemon). This patch supports only hugepages mapped in the pmd, archs that have smaller hugepages will not fit in this patch alone. Also some archs like power have certain tlb limits that prevents mixing different page size in the same regions so they will not fit in this framework that requires "graceful fallback" to basic PAGE_SIZE in case of physical memory fragmentation. hugetlbfs remains a perfect fit for those because its software limits happen to match the hardware limits. hugetlbfs also remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped to be found not fragmented after a certain system uptime and that would be very expensive to defragment with relocation, so requiring reservation. hugetlbfs is the "reservation way", the point of transparent hugepages is not to have any reservation at all and maximizing the use of cache and hugepages at all times automatically. Some performance result: vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep ages3 memset page fault 1566023 memset tlb miss 453854 memset second tlb miss 453321 random access tlb miss 41635 random access second tlb miss 41658 vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3 memset page fault 1566471 memset tlb miss 453375 memset second tlb miss 453320 random access tlb miss 41636 random access second tlb miss 41637 vmx andrea # ./largepages3 memset page fault 1566642 memset tlb miss 453417 memset second tlb miss 453313 random access tlb miss 41630 random access second tlb miss 41647 vmx andrea # ./largepages3 memset page fault 1566872 memset tlb miss 453418 memset second tlb miss 453315 random access tlb miss 41618 random access second tlb miss 41659 vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage vmx andrea # ./largepages3 memset page fault 2182476 memset tlb miss 460305 memset second tlb miss 460179 random access tlb miss 44483 random access second tlb miss 44186 vmx andrea # ./largepages3 memset page fault 2182791 memset tlb miss 460742 memset second tlb miss 459962 random access tlb miss 43981 random access second tlb miss 43988 ============ #include <stdio.h> #include <stdlib.h> #include <string.h> #include <sys/time.h> #define SIZE (3UL*1024*1024*1024) int main() { char *p = malloc(SIZE), *p2; struct timeval before, after; gettimeofday(&before, NULL); memset(p, 0, SIZE); gettimeofday(&after, NULL); printf("memset page fault %Lu\n", (after.tv_sec-before.tv_sec)*1000000UL + after.tv_usec-before.tv_usec); gettimeofday(&before, NULL); memset(p, 0, SIZE); gettimeofday(&after, NULL); printf("memset tlb miss %Lu\n", (after.tv_sec-before.tv_sec)*1000000UL + after.tv_usec-before.tv_usec); gettimeofday(&before, NULL); memset(p, 0, SIZE); gettimeofday(&after, NULL); printf("memset second tlb miss %Lu\n", (after.tv_sec-before.tv_sec)*1000000UL + after.tv_usec-before.tv_usec); gettimeofday(&before, NULL); for (p2 = p; p2 < p+SIZE; p2 += 4096) *p2 = 0; gettimeofday(&after, NULL); printf("random access tlb miss %Lu\n", (after.tv_sec-before.tv_sec)*1000000UL + after.tv_usec-before.tv_usec); gettimeofday(&before, NULL); for (p2 = p; p2 < p+SIZE; p2 += 4096) *p2 = 0; gettimeofday(&after, NULL); printf("random access second tlb miss %Lu\n", (after.tv_sec-before.tv_sec)*1000000UL + after.tv_usec-before.tv_usec); return 0; } ============ Signed-off-by:Rik van Riel <riel@redhat.com> Signed-off-by:
Johannes Weiner <hannes@cmpxchg.org> Signed-off-by:
-
Andrea Arcangeli authored
Not worth throwing away the precious reserved free memory pool for allocations that can fail gracefully (either through mempool or because they're transhuge allocations later falling back to 4k allocations). Signed-off-by:
KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by:
-
Andrea Arcangeli authored
Transparent hugepage allocations must be allowed not to invoke kswapd or any other kind of indirect reclaim (especially when the defrag sysfs is control disabled). It's unacceptable to swap out anonymous pages (potentially anonymous transparent hugepages) in order to create new transparent hugepages. This is true for the MADV_HUGEPAGE areas too (swapping out a kvm virtual machine and so having it suffer an unbearable slowdown, so another one with guest physical memory marked MADV_HUGEPAGE can run 30% faster if it is running memory intensive workloads, makes no sense). If a transparent hugepage allocation fails the slowdown is minor and there is total fallback, so kswapd should never be asked to swapout memory to allow the high order allocation to succeed. Signed-off-by:
-
Andrea Arcangeli authored
This should work for both hugetlbfs and transparent hugepages. [akpm@linux-foundation.org: bring forward PageTransCompound() addition for bisectability] Signed-off-by:
-
Andrea Arcangeli authored
Move the copy/clear_huge_page functions to common code to share between hugetlb.c and huge_memory.c. Signed-off-by:
-
Andrea Arcangeli authored
Paging logic that splits the page before it is unmapped and added to swap to ensure backwards compatibility with the legacy swap code. Eventually swap should natively pageout the hugepages to increase performance and decrease seeking and fragmentation of swap space. swapoff can just skip over huge pmd as they cannot be part of swap yet. In add_to_swap be careful to split the page only if we got a valid swap entry so we don't split hugepages with a full swap. In theory we could split pages before isolating them during the lru scan, but for khugepaged to be safe, I'm relying on either mmap_sem write mode, or PG_lock taken, so split_huge_page has to run either with mmap_sem read/write mode or PG_lock taken. Calling it from isolate_lru_page would make locking more complicated, in addition to that split_huge_page would deadlock if called by __isolate_lru_page because it has to take the lru lock to add the tail pages. Signed-off-by:
-
Andrea Arcangeli authored
split_huge_page_pmd compat code. Each one of those would need to be expanded to hundred of lines of complex code without a fully reliable split_huge_page_pmd design. Signed-off-by:
-
Andrea Arcangeli authored
This increase the size of the mm struct a bit but it is needed to preallocate one pte for each hugepage so that split_huge_page will not require a fail path. Guarantee of success is a fundamental property of split_huge_page to avoid decrasing swapping reliability and to avoid adding -ENOMEM fail paths that would otherwise force the hugepage-unaware VM code to learn rolling back in the middle of its pte mangling operations (if something we need it to learn handling pmd_trans_huge natively rather being capable of rollback). When split_huge_page runs a pte is needed to succeed the split, to map the newly splitted regular pages with a regular pte. This way all existing VM code remains backwards compatible by just adding a split_huge_page* one liner. The memory waste of those preallocated ptes is negligible and so it is worth it. Signed-off-by:
-
Andrea Arcangeli authored
split_huge_page must transform a compound page to a regular page and needs ClearPageCompound. Signed-off-by:
Christoph Lameter <cl@linux-foundation.org> Acked-by:
-
Andrea Arcangeli authored
Add mmu notifier helpers to handle pmd huge operations. Signed-off-by:
-
Andrea Arcangeli authored
pte alloc routines must wait for split_huge_page if the pmd is not present and not null (i.e. pmd_trans_splitting). The additional branches are optimized away at compile time by pmd_trans_splitting if the config option is off. However we must pass the vma down in order to know the anon_vma lock to wait for. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by:
-
Andrea Arcangeli authored
Force gup_fast to take the slow path and block if the pmd is splitting, not only if it's none. Signed-off-by:
-
Andrea Arcangeli authored
Add needed pmd mangling functions with symmetry with their pte counterparts. pmdp_splitting_flush() is the only new addition on the pmd_ methods and it's needed to serialize the VM against split_huge_page. It simply atomically sets the splitting bit in a similar way pmdp_clear_flush_young atomically clears the accessed bit. pmdp_splitting_flush() also has to flush the tlb to make it effective against gup_fast, but it wouldn't really require to flush the tlb too. Just the tlb flush is the simplest operation we can invoke to serialize pmdp_splitting_flush() against gup_fast. Signed-off-by:
-
Andrea Arcangeli authored
Some are needed to build but not actually used on archs not supporting transparent hugepages. Others like pmdp_clear_flush are used by x86 too. Signed-off-by:
-
Andrea Arcangeli authored
These returns 0 at compile time when the config option is disabled, to allow gcc to eliminate the transparent hugepage function calls at compile time without additional #ifdefs (only the export of those functions have to be visible to gcc but they won't be required at link time and huge_memory.o can be not built at all). _PAGE_BIT_UNUSED1 is never used for pmd, only on pte. Signed-off-by:
-
Andrea Arcangeli authored
Add config option. Signed-off-by:
-
Andrea Arcangeli authored
Warn destroy_compound_page that __split_huge_page_refcount is heavily dependent on its internal behavior. Signed-off-by:
-
Andrea Arcangeli authored
huge_memory.c needs it too when it fallbacks in copying hugepages into regular fragmented pages if hugepage allocation fails during COW. Signed-off-by:
-
Andrea Arcangeli authored
No paravirt version of set_pmd_at/pmd_update/pmd_update_defer. Signed-off-by:
-
Andrea Arcangeli authored
Paravirt ops pmd_update/pmd_update_defer/pmd_set_at. Not all might be necessary (vmware needs pmd_update, Xen needs set_pmd_at, nobody needs pmd_update_defer), but this is to keep full simmetry with pte paravirt ops, which looks cleaner and simpler from a common code POV. Signed-off-by:
-
Andrea Arcangeli authored
Used by paravirt and not paravirt set_pmd_at. Signed-off-by:
-
Andrea Arcangeli authored
Clear compound mapping for anonymous compound pages like it already happens for regular anonymous pages. But crash if mapping is set for any tail page, also the PageAnon check is meaningless for tail pages. This check only makes sense for the head page, for tail page it can only hide bugs and we definitely don't want to hide bugs. Signed-off-by:
-
Andrea Arcangeli authored
Futex code is smarter than most other gup_fast O_DIRECT code and knows about the compound internals. However now doing a put_page(head_page) will not release the pin on the tail page taken by gup-fast, leading to all sort of refcounting bugchecks. Getting a stable head_page is a little tricky. page_head = page is there because if this is not a tail page it's also the page_head. Only in case this is a tail page, compound_head is called, otherwise it's guaranteed unnecessary. And if it's a tail page compound_head has to run atomically inside irq disabled section __get_user_pages_fast before returning. Otherwise ->first_page won't be a stable pointer. Disableing irq before __get_user_page_fast and releasing irq after running compound_head is needed because if __get_user_page_fast returns == 1, it means the huge pmd is established and cannot go away from under us. pmdp_splitting_flush_notify in __split_huge_page_splitting will have to wait for local_irq_enable before the IPI delivery can return. This means __split_huge_page_refcount can't be running from under us, and in turn when we run compound_head(page) we're not reading a dangling pointer from tailpage->first_page. Then after we get to stable head page, we are always safe to call compound_lock and after taking the compound lock on head page we can finally re-check if the page returned by gup-fast is still a tail page. in which case we're set and we didn't need to split the hugepage in order to take a futex on it. Signed-off-by:
-
Andrea Arcangeli authored
After releasing the compound_lock split_huge_page can still run and release the page before put_page_testzero runs. Signed-off-by:
-
Andrea Arcangeli authored
Alter compound get_page/put_page to keep references on subpages too, in order to allow __split_huge_page_refcount to split an hugepage even while subpages have been pinned by one of the get_user_pages() variants. Signed-off-by:
-
Andrea Arcangeli authored
Add a new compound_lock() needed to serialize put_page against __split_huge_page_refcount(). Signed-off-by:
-
Andrea Arcangeli authored
Define MADV_HUGEPAGE. Signed-off-by:
Arnd Bergmann <arnd@arndb.de> Acked-by:
-
Andrea Arcangeli authored
Documentation/vm/transhuge.txt Signed-off-by:
-
Andrea Arcangeli authored
page_count shows the count of the head page, but the actual check is done on the tail page, so show what is really being checked. Signed-off-by:
-
Hugh Dickins authored
When a swapcache page is replaced by a ksm page, it's best to free that swap immediately. Reported-by:
-
Minchan Kim authored
I think determine_dirtyable_memory() is a rather costly function since it need many atomic reads for gathering zone/global page state. But when we use vm_dirty_bytes && dirty_background_bytes, we don't need that costly calculation. This patch eliminates such unnecessary overhead. NOTE : newly added if condition might add overhead in normal path. But it should be _really_ small because anyway we need the access both vm_dirty_bytes and dirty_background_bytes so it is likely to hit the cache. [akpm@linux-foundation.org: fix used-uninitialised warning] Signed-off-by:Minchan Kim <minchan.kim@gmail.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by:
-
Volodymyr G. Lukiianyk authored
When numa_zonelist_order parameter is set to "node" or "zone" on the command line it's still showing as "default" in sysctl. That's because early_param parsing function changes only user_zonelist_order variable. Fix this by copying user-provided string to numa_zonelist_order if it was successfully parsed. Signed-off-by:
Volodymyr G Lukiianyk <volodymyrgl@gmail.com> Acked-by:
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by:
-
Mel Gorman authored
When kswapd is woken up for a high-order allocation, it takes account of the highest usable zone by the caller (the classzone idx). During allocation, this index is used to select the lowmem_reserve[] that should be applied to the watermark calculation in zone_watermark_ok(). When balancing a node, kswapd considers the highest unbalanced zone to be the classzone index. This will always be at least be the callers classzone_idx and can be higher. However, sleeping_prematurely() always considers the lowest zone (e.g. ZONE_DMA) to be the classzone index. This means that sleeping_prematurely() can consider a zone to be balanced that is unusable by the allocation request that originally woke kswapd. This patch changes sleeping_prematurely() to use a classzone_idx matching the value it used in balance_pgdat(). Signed-off-by:
Eric B Munson <emunson@mgebm.net> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Simon Kirby <sim@hostway.ca> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Shaohua Li <shaohua.li@intel.com> Cc: Dave Hansen <dave@linux.vnet.ibm.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by:
-
Mel Gorman authored
After DEF_PRIORITY, balance_pgdat() considers all_unreclaimable zones to be balanced but sleeping_prematurely does not. This can force kswapd to stay awake longer than it should. This patch fixes it. Signed-off-by:
-
Mel Gorman authored
When kswapd wakes up, it reads its order and classzone from pgdat and calls balance_pgdat. While its awake, it potentially reclaimes at a high order and a low classzone index. This might have been a once-off that was not required by subsequent callers. However, because the pgdat values were not reset, they remain artifically high while balance_pgdat() is running and potentially kswapd enters a second unnecessary reclaim cycle. Reset the pgdat order and classzone index after reading. Signed-off-by:
-
Mel Gorman authored
Before kswapd goes to sleep, it uses sleeping_prematurely() to check if there was a race pushing a zone below its watermark. If the race happened, it stays awake. However, balance_pgdat() can decide to reclaim at order-0 if it decides that high-order reclaim is not working as expected. This information is not passed back to sleeping_prematurely(). The impact is that kswapd remains awake reclaiming pages long after it should have gone to sleep. This patch passes the adjusted order to sleeping_prematurely and uses the same logic as balance_pgdat to decide if it's ok to go to sleep. Signed-off-by:
-
Mel Gorman authored
When reclaiming for high-orders, kswapd is responsible for balancing a node but it should not reclaim excessively. It avoids excessive reclaim by considering if any zone in a node is balanced then the node is balanced. In the cases where there are imbalanced zone sizes (e.g. ZONE_DMA with both ZONE_DMA32 and ZONE_NORMAL), kswapd can go to sleep prematurely as just one small zone was balanced. This alters the sleep logic of kswapd slightly. It counts the number of pages that make up the balanced zones. If the total number of balanced pages is more than a quarter of the zone, kswapd will go back to sleep. This should keep a node balanced without reclaiming an excessive number of pages. Signed-off-by:
-
Mel Gorman authored
Simon Kirby reported the following problem We're seeing cases on a number of servers where cache never fully grows to use all available memory. Sometimes we see servers with 4 GB of memory that never seem to have less than 1.5 GB free, even with a constantly-active VM. In some cases, these servers also swap out while this happens, even though they are constantly reading the working set into memory. We have been seeing this happening for a long time; I don't think it's anything recent, and it still happens on 2.6.36. After some debugging work by Simon, Dave Hansen and others, the prevaling theory became that kswapd is reclaiming order-3 pages requested by SLUB too aggressive about it. There are two apparent problems here. On the target machine, there is a small Normal zone in comparison to DMA32. As kswapd tries to balance all zones, it would continually try reclaiming for Normal even though DMA32 was balanced enough for callers. The second problem is that sleeping_prematurely() does not use the same logic as balance_pgdat() when deciding whether to sleep or not. This keeps kswapd artifically awake. A number of tests were run and the figures from previous postings will look very different for a few reasons. One, the old figures were forcing my network card to use GFP_ATOMIC in attempt to replicate Simon's problem. Second, I previous specified slub_min_order=3 again in an attempt to reproduce Simon's problem. In this posting, I'm depending on Simon to say whether his problem is fixed or not and these figures are to show the impact to the ordinary cases. Finally, the "vmscan" figures are taken from /proc/vmstat instead of the tracepoints. There is less information but recording is less disruptive. The first test of relevance was postmark with a process running in the background reading a large amount of anonymous memory in blocks. The objective was to vaguely simulate what was happening on Simon's machine and it's memory intensive enough to have kswapd awake. POSTMARK traceonly kanyzone Transactions per second: 156.00 ( 0.00%) 153.00 (-1.96%) Data megabytes read per second: 21.51 ( 0.00%) 21.52 ( 0.05%) Data megabytes written per second: 29.28 ( 0.00%) 29.11 (-0.58%) Files created alone per second: 250.00 ( 0.00%) 416.00 (39.90%) Files create/transact per second: 79.00 ( 0.00%) 76.00 (-3.95%) Files deleted alone per second: 520.00 ( 0.00%) 420.00 (-23.81%) Files delete/transact per second: 79.00 ( 0.00%) 76.00 (-3.95%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 16.58 17.4 Total Elapsed Time (seconds) 218.48 222.47 VMstat Reclaim Statistics: vmscan Direct reclaims 0 4 Direct reclaim pages scanned 0 203 Direct reclaim pages reclaimed 0 184 Kswapd pages scanned 326631 322018 Kswapd pages reclaimed 312632 309784 Kswapd low wmark quickly 1 4 Kswapd high wmark quickly 122 475 Kswapd skip congestion_wait 1 0 Pages activated 700040 705317 Pages deactivated 212113 203922 Pages written 9875 6363 Total pages scanned 326631 322221 Total pages reclaimed 312632 309968 %age total pages scanned/reclaimed 95.71% 96.20% %age total pages scanned/written 3.02% 1.97% proc vmstat: Faults Major Faults 300 254 Minor Faults 645183 660284 Page ins 493588 486704 Page outs 4960088 4986704 Swap ins 1230 661 Swap outs 9869 6355 Performance is mildly affected because kswapd is no longer doing as much work and the background memory consumer process is getting in the way. Note that kswapd scanned and reclaimed fewer pages as it's less aggressive and overall fewer pages were scanned and reclaimed. Swap in/out is particularly reduced again reflecting kswapd throwing out fewer pages. The slight performance impact is unfortunate here but it looks like a direct result of kswapd being less aggressive. As the bug report is about too many pages being freed by kswapd, it may have to be accepted for now. The second test is a streaming IO benchmark that was previously used by Johannes to show regressions in page reclaim. MICRO traceonly kanyzone User/Sys Time Running Test (seconds) 29.29 28.87 Total Elapsed Time (seconds) 492.18 488.79 VMstat Reclaim Statistics: vmscan Direct reclaims 2128 1460 Direct reclaim pages scanned 2284822 1496067 Direct reclaim pages reclaimed 148919 110937 Kswapd pages scanned 15450014 16202876 Kswapd pages reclaimed 8503697 8537897 Kswapd low wmark quickly 3100 3397 Kswapd high wmark quickly 1860 7243 Kswapd skip congestion_wait 708 801 Pages activated 9635 9573 Pages deactivated 1432 1271 Pages written 223 1130 Total pages scanned 17734836 17698943 Total pages reclaimed 8652616 8648834 %age total pages scanned/reclaimed 48.79% 48.87% %age total pages scanned/written 0.00% 0.01% proc vmstat: Faults Major Faults 165 221 Minor Faults 9655785 9656506 Page ins 3880 7228 Page outs 37692940 37480076 Swap ins 0 69 Swap outs 19 15 Again fewer pages are scanned and reclaimed as expected and this time the test completed faster. Note that kswapd is hitting its watermarks faster (low and high wmark quickly) which I expect is due to kswapd reclaiming fewer pages. I also ran fs-mark, iozone and sysbench but there is nothing interesting to report in the figures. Performance is not significantly changed and the reclaim statistics look reasonable. Tgis patch: When the allocator enters its slow path, kswapd is woken up to balance the node. It continues working until all zones within the node are balanced. For order-0 allocations, this makes perfect sense but for higher orders it can have unintended side-effects. If the zone sizes are imbalanced, kswapd may reclaim heavily within a smaller zone discarding an excessive number of pages. The user-visible behaviour is that kswapd is awake and reclaiming even though plenty of pages are free from a suitable zone. This patch alters the "balance" logic for high-order reclaim allowing kswapd to stop if any suitable zone becomes balanced to reduce the number of pages it reclaims from other zones. kswapd still tries to ensure that order-0 watermarks for all zones are met before sleeping. Signed-off-by:
-
