- 11 Sep, 2009 1 commit
-
-
Lee Schermerhorn authored
workload running on recent distro [SLES11] and mainline kernels on an 8-socket, 32-core, 256GB x86_64 platform. On the SLES11 kernel [2.6.27.19+] with Barcelona processors, as we increased the load [10s of thousands of tasks], the throughput would vary between two "plateaus"--one at ~65K jobs per minute and one at ~130K jpm. The simple patch below causes the results to smooth out at the ~130k plateau. But wait, there's more: We do not see this behavior on smaller platforms--e.g., 4 socket/8 core. This could be the result of the larger number of cpus on the larger platform--a scalability issue--or it could be the result of the larger number of interconnect "hops" between some nodes in this platform and how the tasks for a given load end up distributed over the nodes' cpus and memories--a stochastic NUMA effect. The variability in the results are less pronounced [on the same platform] with Shanghai processors and with mainline kernels. With 31-rc6 on Shanghai processors and 288 file systems on 288 fibre attached storage volumes, the curves [jpm vs load] are both quite flat with the patched kernel consistently producing ~3.9% better throughput [~80K jpm vs ~77K jpm] than the unpatched kernel. Profiling indicated that the "slow" runs were incurring high[er] contention on an anon_vma lock in vma_adjust(), apparently called from the sbrk() system call. The patch: A comment in mm/mmap.c:vma_adjust() suggests that we don't really need the anon_vma lock when we're only adjusting the end of a vma, as is the case for brk(). The comment questions whether it's worth while to optimize for this case. Apparently, on the newer, larger x86_64 platforms, with interesting NUMA topologies, it is worth while--especially considering that the patch [if correct!] is quite simple. We can detect this condition--no overlap with next vma--by noting a NULL "importer". The anon_vma pointer will also be NULL in this case, so simply avoid loading vma->anon_vma to avoid the lock. However, we apparently DO need to take the anon_vma lock when we're inserting a vma ['insert' non-NULL] even when we have no overlap [NULL "importer"], so we need to check for 'insert', as well. I have tested with and without the 'file || ' test in the patch. This does not seem to matter for stability nor performance. I left this check/filter in, so we only optimize away the anon_vma lock acquisition when adjusting the end of a non- importing, non-inserting, anon vma. Signed-off-by:
Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Eric Whitney <eric.whitney@hp.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org>
-
- 10 Sep, 2009 2 commits
-
-
David Miller authored
where D-cache aliasing is an issue. vmalloc_user() ensures that the kernel side mapping is SHMLBA aligned, and on platforms with D-cache aliasing matters the presence of VM_SHARED will similarly SHMLBA align the user side mapping. Thus the kernel and the user will be writing to the same D-cache aliases and we'll avoid inconsistencies and corruption. The only trick with this change is that vfree() cannot be invoked from interrupt context, and thus it's not allowed from RCU callbacks. We deal with this by using schedule_work(). Since the ring buffer is now completely linear even on the kernel side, several simplifications are probably now possible in the code where we add entries to the ring. With help from Peter Zijlstra. Signed-off-by:
David S. Miller <davem@davemloft.net> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Jens Axboe <jens.axboe@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by:
-
David Miller authored
is necessary for this to work on platforms with cpu D-caches which can have aliases. Otherwise kernel side writes won't be seen properly in userspace and vice versa. If the kernel side mapping and the user side one have the same alignment, modulo SHMLBA, this can work as long as VM_SHARED is shared of VMA and for all current users this is true. VM_SHARED will force SHMLBA alignment of the user side mmap on platforms with D-cache aliasing matters. The bulk of this patch is just making it so that a specific alignment can be passed down into __get_vm_area_node(). All existing callers pass in '1' which preserves existing behavior. vmalloc_user() gives SHMLBA for the alignment. As a side effect this should get the video media drivers and other vmalloc_user() users into more working shape on such systems. Signed-off-by:
-
- 09 Sep, 2009 10 commits
-
-
Hugh Dickins authored
CONFIG_TMPFS is off: that's a little anomalous, and I'd intended to make more sense of it by removing CONFIG_TMPFS altogether, always enabling its code when CONFIG_SHMEM; but so many defconfigs have CONFIG_SHMEM on CONFIG_TMPFS off that we'd better leave that as is. But there is no point in asking for CONFIG_TMPFS if CONFIG_SHMEM is off: make TMPFS depend on SHMEM, which also prevents TMPFS_POSIX_ACL shmem_acl.o being pointlessly built into the kernel when SHMEM is off. And a selfish change, to prevent the world from being rebuilt when I switch between CONFIG_SHMEM on and off: the only CONFIG_SHMEM in the header files is mm.h shmem_lock() - give that a shmem.c stub instead. Signed-off-by:
Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by:
Matt Mackall <mpm@selenic.com> Signed-off-by:
-
Huang Shijie authored
the bit VM_LOCKED which is set by calc_vm_flag_bits(). So there is no need to reset it again, just remove it. Signed-off-by:
Huang Shijie <shijie8@gmail.com> Acked-by:
-
Hugh Dickins authored
them into FOLL flags for follow_page(). Though oddly named, the FOLL flags are more widely used, so pass them to __get_user_pages() now. Sorry, VM flags, VM_FAULT flags and FAULT_FLAGs are still distinct. (The patch to __get_user_pages() looks peculiar, with both gup_flags and foll_flags: the gup_flags remain constant; but as before there's an exceptional case, out of scope of the patch, in which foll_flags per page have FOLL_WRITE masked off.) Signed-off-by:
-
Hugh Dickins authored
advantage of the ZERO_PAGE: which we stopped do_anonymous_page() from using in 2.6.24. And there were a couple of regression reports on LKML. Following suggestions from Linus, reinstate do_anonymous_page() use of the ZERO_PAGE; but this time avoid dirtying its struct page cacheline with (map)count updates - let vm_normal_page() regard it as abnormal. Use it only on arches which __HAVE_ARCH_PTE_SPECIAL (x86, s390, sh32, most powerpc): that's not essential, but minimizes additional branches (keeping them in the unlikely pte_special case); and incidentally excludes mips (some models of which needed eight colours of ZERO_PAGE to avoid costly exceptions). Don't be fanatical about avoiding ZERO_PAGE updates: get_user_pages() callers won't want to make exceptions for it, so increment its count there. Changes to mlock and migration? happily seems not needed. In most places it's quicker to check pfn than struct page address: prepare a __read_mostly zero_pfn for that. Does get_dump_page() still need its ZERO_PAGE check? probably not, but keep it anyway. Signed-off-by:
Rik van Riel <riel@redhat.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Minchan Kim <minchan.kim@gmail.com> Signed-off-by:
-
Hugh Dickins authored
If it's not going to mark the pte writable, then it won't help to mark it dirty here, and clogs up memory with pages which will need swap instead of being thrown away. Especially wrong if no overcommit is chosen, and this vma is not yet VM_ACCOUNTed - we could exceed the limit and OOM despite no overcommit. Signed-off-by:
-
Hugh Dickins authored
either: pass the foll_flags down to it, instead of just the write bit. Remove that obscure huge_zeropage_ok() test. The decision is easy, though unlike the non-huge case - here vm_ops->fault is always set. But we know that a fault would serve up zeroes, unless there's already a hugetlbfs pagecache page to back the range. (Alternatively, since hugetlb pages aren't swapped out under pressure, you could save more dump space by arguing that a page not yet faulted into this process cannot be relevant to the dump; but that would be more surprising.) Signed-off-by:
-
Hugh Dickins authored
confusion and bugs: why does it test VM_SHARED? for the very good but unsatisfying reason that VMware crashed without. As we look to maybe reinstating anonymous use of the ZERO_PAGE, we need to sort this out. Easily done: it's silly for __get_user_pages() and follow_page() to be guessing whether it's safe to assume that they're being used for a coredump (which can take a shortcut snapshot where other uses must handle a fault) - just tell them with GUP_FLAGS_DUMP and FOLL_DUMP. get_dump_page() doesn't even want a ZERO_PAGE: an error suits fine. Signed-off-by:
Mel Gorman <mel@csn.ul.ie> Reviewed-by:
Minchan Kim <minchan.kim@gmail.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Nick Piggin <npiggin@suse.de> Signed-off-by:
-
Hugh Dickins authored
interface for the CONFIG_ELF_CORE dumpers to use, instead of calling get_user_pages() directly. They're not interested in errors: they just want to use holes as much as possible, to save space and make sure that the data is aligned where the headers said it would be. Oh, and don't use that horrid DUMP_SEEK(off) macro! Signed-off-by:
-
Hugh Dickins authored
flags added solely to prevent __get_user_pages() from doing some of what it usually does, in the munlock case: we can now remove them. Signed-off-by:
-
Hugh Dickins authored
by SIGKILL, the subsequent munlock() takes unnecessarily long because its use of __get_user_pages() insists on faulting in all the pages which mlock() never reached. It's worse than slowness if mlock() is terminated by Out Of Memory kill: the munlock_vma_pages_all() in exit_mmap() insists on faulting in all the pages which mlock() could not find memory for; so innocent bystanders are killed too, and perhaps the system hangs. __get_user_pages() does a lot that's silly for munlock(): so remove the munlock option from __mlock_vma_pages_range(), and use a simple loop of follow_page()s in munlock_vma_pages_range() instead; ignoring absent pages, and not marking present pages as accessed or dirty. (Change munlock() to only go so far as mlock() reached? That does not work out, given the convention that mlock() claims complete success even when it has to give up early - in part so that an underlying file can be extended later, and those pages locked which earlier would give SIGBUS.) Signed-off-by:
-
- 03 Sep, 2009 1 commit
-
-
Minchan Kim authored
nodes [0,1] in system. barrios:~$ cat /proc/zoneinfo | egrep 'numa|zone' Node 0, zone DMA numa_hit 33226 numa_miss 1739 numa_foreign 27978 .. .. Node 1, zone DMA numa_hit 307 numa_miss 46900 numa_foreign 0 1) In node 0, NUMA_MISS means it wanted to allocate page in node 1 but ended up with page in node 0 2) In node 0, NUMA_FOREIGN means it wanted to allocate page in node 0 but ended up with page from Node 1. But now, numastat explains it oppositely about (MISS, FOREIGN). Let's fix up with viewpoint of zone. Signed-off-by:
Christoph Lameter <cl@linux-foundation.org> Signed-off-by:
-
- 04 Sep, 2009 1 commit
-
-
Vincent Li authored
movement, either. Already, in commit 5343daceec, KOSAKI did it about shrink_inactive_list. This patch removes unnecessary overhead of page accounting and locking in shrink_active_list as follow-up work of commit 5343daceec. Signed-off-by:
Vincent Li <macli@brc.ubc.ca> Reviewed-by:
KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reviewed-by:
Wu Fengguang <fengguang.wu@intel.com> Acked-by:
-
- 02 Sep, 2009 3 commits
-
-
Andrew Morton authored
#45: FILE: mm/page_alloc.c:551: + * Remove pages from lists in a round-robin fashion. A batch_free total: 0 errors, 1 warnings, 46 lines checked ./patches/page-allocator-maintain-rolling-count-of-pages-to-free-from-the-pcp.patch has style problems, please review. If any of these errors are false positives report them to the maintainer, see CHECKPATCH in MAINTAINERS. Please run checkpatch prior to sending patches Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by: -
Mel Gorman authored
encountered. In the event one of the lists has more pages than another, there may be numerous checks for list_empty() which is undesirable. This patch maintains a count of pages to free which is incremented when empty lists are encountered. The intention is that more pages will then be freed from fuller lists than the empty ones reducing the number of empty list checks in the free path. Signed-off-by:
-
Mel Gorman authored
by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by:
Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by:
-
- 27 Aug, 2009 3 commits
-
-
KOSAKI Motohiro authored
and new line handling. this patch fixes it. Signed-off-by:
-
KOSAKI Motohiro authored
thas shread the same mm. it mean vfork parent will be killed. This is definitely incorrect. another process have another oom_adj. we shouldn't ignore their oom_adj (it might have OOM_DISABLE). following caller hit the minefield. =============================== switch (constraint) { case CONSTRAINT_MEMORY_POLICY: oom_kill_process(current, gfp_mask, order, 0, NULL, "No available memory (MPOL_BIND)"); break; Note: force_sig(SIGKILL) send SIGKILL to all thread in the process. We don't need to care multi thread in here. Signed-off-by: -
KOSAKI Motohiro authored
as per-process too. it makes consistency more and makes speed up select_bad_process(). Signed-off-by:
-
- 12 Sep, 2009 1 commit
-
-
KOSAKI Motohiro authored
treating code in android driver also should be changed. Signed-off-by:
-
- 27 Aug, 2009 2 commits
-
-
KOSAKI Motohiro authored
__out_of_memory() select_bad_process() for each task badness() calculate badness of one task oom_kill_process() search child oom_kill_task() kill target task and mm shared tasks with it example, process-A have two thread, thread-A and thread-B and it have very fat memory and each thread have following oom_adj and oom_score. thread-A: oom_adj = OOM_DISABLE, oom_score = 0 thread-B: oom_adj = 0, oom_score = very-high Then, select_bad_process() select thread-B, but oom_kill_task() refuse kill the task because thread-A have OOM_DISABLE. Thus __out_of_memory() call select_bad_process() again. but select_bad_process() select the same task. It mean kernel fall in livelock. The fact is, select_bad_process() must select killable task. otherwise OOM logic go into livelock. And root cause is, oom_adj shouldn't be per-thread value. it should be per-process value because OOM-killer kill a process, not thread. Thus This patch moves oomkilladj (now more appropriately named oom_adj) from struct task_struct to struct signal_struct. it naturally prevent select_bad_process() choose wrong task. Signed-off-by: -
Eric B Munson authored
directory and a reference to the example in hugetlbpage.txt. Signed-off-by:
Eric B Munson <ebmunson@us.ibm.com> Acked-by:
David Rientjes <rientjes@google.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Adam Litke <agl@us.ibm.com> Cc: David Gibson <david@gibson.dropbear.id.au> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Signed-off-by:
-
- 10 Sep, 2009 2 commits
-
-
Eric B Munson authored
will look like anonymous memory to userspace. This is accomplished by using a file on the internal vfsmount. MAP_HUGETLB is a modifier of MAP_ANONYMOUS and so must be specified with it. The region will behave the same as a MAP_ANONYMOUS region using small pages. Signed-off-by:
-
Eric B Munson authored
anonymous mapping be backed with huge pages. This mapping will borrow functionality from the huge page shm code to create a file on the kernel internal mount and use it to approximate an anonymous mapping. The MAP_HUGETLB flag is a modifier to MAP_ANONYMOUS and will not work without both flags being preset. A new flag is necessary because there is no other way to hook into huge pages without creating a file on a hugetlbfs mount which wouldn't be MAP_ANONYMOUS. To userspace, this mapping will behave just like an anonymous mapping because the file is not accessible outside of the kernel. This patchset is meant to simplify the programming model. Presently there is a large chunk of boiler platecode, contained in libhugetlbfs, required to create private, hugepage backed mappings. This patch set would allow use of hugepages without linking to libhugetlbfs or having hugetblfs mounted. Unification of the VM code would provide these same benefits, but it has been resisted each time that it has been suggested for several reasons: it would break PAGE_SIZE assumptions across the kernel, it makes page-table abstractions really expensive, and it does not provide any benefit on architectures that do not support huge pages, incurring fast path penalties without providing any benefit on these architectures. This patch: There are two means of creating mappings backed by huge pages: 1. mmap() a file created on hugetlbfs 2. Use shm which creates a file on an internal mount which essentially maps it MAP_SHARED The internal mount is only used for shared mappings but there is very little that stops it being used for private mappings. This patch extends hugetlbfs_file_setup() to deal with the creation of files that will be mapped MAP_PRIVATE on the internal hugetlbfs mount. This extended API is used in a subsequent patch to implement the MAP_HUGETLB mmap() flag. Signed-off-by:
-
- 25 Aug, 2009 1 commit
-
-
Vincent Li authored
page_queue_congested(). Remove the page_queue_congested() comment in vmscan pageout() too. Signed-off-by:
-
- 22 Aug, 2009 2 commits
-
-
Wu Fengguang authored
which case shrink_list() _still_ calls isolate_pages() with the much larger SWAP_CLUSTER_MAX. It effectively scales up the inactive list scan rate by up to 32 times. For example, with 16k inactive pages and DEF_PRIORITY=12, (16k >> 12)=4. So when shrink_zone() expects to scan 4 pages in the active/inactive list, the active list will be scanned 4 pages, while the inactive list will be (over) scanned SWAP_CLUSTER_MAX=32 pages in effect. And that could break the balance between the two lists. It can further impact the scan of anon active list, due to the anon active/inactive ratio rebalance logic in balance_pgdat()/shrink_zone(): inactive anon list over scanned => inactive_anon_is_low() == TRUE => shrink_active_list() => active anon list over scanned So the end result may be - anon inactive => over scanned - anon active => over scanned (maybe not as much) - file inactive => over scanned - file active => under scanned (relatively) The accesses to nr_saved_scan are not lock protected and so not 100% accurate, however we can tolerate small errors and the resulted small imbalanced scan rates between zones. Cc: Rik van Riel <riel@redhat.com> Reviewed-by:Balbir Singh <balbir@linux.vnet.ibm.com> Reviewed-by:
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by:
-
Alexey Dobriyan authored
Acked-by:
-
- 21 Aug, 2009 1 commit
-
-
Vincent Li authored
but it actually returns zone's LRU list number pages. Signed-off-by:
-
- 20 Aug, 2009 5 commits
-
-
Jan Beulich authored
may want a sub-page sized amount of memory. Overall this seems more consistent with the other hash table allocations, and allows making two supposedly mm-only variables really mm-only (nr_{kernel,all}_pages). Signed-off-by:Jan Beulich <jbeulich@novell.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by:
-
Jan Beulich authored
addressing) memory anyway, using the ..._low() variant only makes sense when the physical address range of the allocated memory must fulfill further constraints, espacially since on 64-bits (or more generally in all cases where the pools the two variants allocate from are than the full available range. Probably the use in alloc_tce_table() could also be eliminated (based on code inspection of pci-calgary_64.c), but that seems too risky given I know nothing about that hardware and have no way to test it. Signed-off-by:
-
Jan Beulich authored
pages found in a system, as that generally includes (perhaps a huge amount of) non-RAM pages. The amount of what actually is usable as storage should instead be used as a basis here. Some of the calculations (i.e. those not intending to use high memory) should likely even use (totalram_pages - totalhigh_pages). Signed-off-by:
Rusty Russell <rusty@rustcorp.com.au> Acked-by:
Ingo Molnar <mingo@elte.hu> Cc: Dave Airlie <airlied@linux.ie> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: "David S. Miller" <davem@davemloft.net> Cc: Patrick McHardy <kaber@trash.net> Signed-off-by:
-
Jan Beulich authored
pages found in a system, as that generally includes (perhaps a huge amount of) non-RAM pages. The amount of what actually is usable as storage should instead be used as a basis here. In line with that, the memory hotplug code should update num_physpages in a way that it retains its original (post-boot) meaning; in particular, decreasing the value should at best be done with great care - this patch doesn't try to ever decrease this value at all as it doesn't really seem meaningful to do so. Signed-off-by:
-
Mel Gorman authored
that depended on high-order atomic allocations were failing. The solution was to preserve a property in the buddy allocator which tended to keep the minimum number of free pages in the zone at the lower physical addresses and contiguous. To preserve this property, MIGRATE_RESERVE was introduced and a number of pageblocks at the start of a zone would be marked "reserve", the number of which depended on min_free_kbytes. Anti-fragmentation works by avoiding the mixing of page migratetypes within the same pageblock. One way of helping this is to increase min_free_kbytes because it becomes less like that it will be necessary to place pages of of MIGRATE_RESERVE is unbounded, the free memory is kept there in large contiguous blocks instead of helping anti-fragmentation as much as it should. With the page-allocator tracepoint patches applied, it was found during anti-fragmentation tests that the number of fragmentation-related events were far higher than expected even with min_free_kbytes at higher values. This patch limits the number of MIGRATE_RESERVE blocks that exist per zone to two. For example, with a sufficient min_free_kbytes, 4MB of memory will be kept aside on an x86-64 and remain more or less free and contiguous for the systems uptime. This should be sufficient for devices depending on high-order atomic allocations while helping fragmentation control when min_free_kbytes is tuned appropriately. As side-effect of this patch is that the reserve variable is converted to int as unsigned long was the wrong type to use when ensuring that only the required number of reserve blocks are created. With the patches applied, fragmentation-related events as measured by the page allocator tracepoints were significantly reduced when running some fragmentation stress-tests on systems with min_free_kbytes tuned to a value appropriate for hugepage allocations at runtime. On x86, the events recorded were reduced by 99.8%, on x86-64 by 99.72% and on ppc64 by 99.83%. Signed-off-by:
-
- 13 Aug, 2009 2 commits
-
-
Johannes Weiner authored
cache page freeable. Signed-off-by:
Johannes Weiner <hannes@cmpxchg.org> Cc: Christoph Lameter <cl@linux-foundation.org> Reviewed-by:
-
Johannes Weiner authored
negations from the two callsites using it for arithmetic. Signed-off-by:
-
- 24 Aug, 2009 1 commit
-
-
Johannes Weiner authored
arithmetic, which was always a bit weird. Now that page_lru_base_type() exists for LRU arithmetic, make page_is_file_cache() a real predicate function and adjust the boolean-using callsites to drop those pesky double negations. Signed-off-by:
-
- 27 Aug, 2009 1 commit
-
-
Hugh Dickins authored
killer after an hour or two. Bisection led to mm-return-boolean-from-page_is_file_cache.patch, but really it's the prior mm-introduce-page_lru_base_type.patch that's at fault. It converted page_lru() to use page_lru_base_type(), but forgot to convert del_page_from_lru() - which then decremented the wrong stats once page_is_file_cache() was changed to a boolean. Fix that, move page_lru_base_type() before del_page_from_lru(), and mark it "inline" like the other mm_inline.h functions. Signed-off-by:
-
- 24 Aug, 2009 1 commit
-
-
Johannes Weiner authored
another helper with a more appropriate name and convert the non-boolean users of page_is_file_cache() accordingly. This new helper gives the LRU base type a page is supposed to live on, inactive anon or inactive file. Signed-off-by:
-
