- 07 Dec, 2006 40 commits
-
-
Andy Whitcroft authored
NUMA node ids are passed as either int or unsigned int almost exclusivly page_to_nid and zone_to_nid both return unsigned long. This is a throw back to when page_to_nid was a #define and was thus exposing the real type of the page flags field. In addition to fixing up the definitions of page_to_nid and zone_to_nid I audited the users of these functions identifying the following incorrect uses: 1) mm/page_alloc.c show_node() -- printk dumping the node id, 2) include/asm-ia64/pgalloc.h pgtable_quicklist_free() -- comparison against numa_node_id() which returns an int from cpu_to_node(), and 3) mm/mpolicy.c check_pte_range -- used as an index in node_isset which uses bit_set which in generic code takes an int. Signed-off-by:
Andy Whitcroft <apw@shadowen.org> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: "Luck, Tony" <tony.luck@intel.com> Signed-off-by:
Andrew Morton <akpm@osdl.org> Signed-off-by:
Linus Torvalds <torvalds@osdl.org>
-
Christoph Lameter authored
drain_node_pages() currently drains the complete pageset of all pages. If there are a large number of pages in the queues then we may hold off interrupts for too long. Duplicate the method used in free_hot_cold_page. Only drain pcp->batch pages at one time. Signed-off-by:
Christoph Lameter <clameter@sgi.com> Signed-off-by:
-
Christoph Lameter authored
Remove all uses of kmem_cache_t (the most were left in slab.h). The typedef for kmem_cache_t is then only necessary for other kernel subsystems. Add a comment to that effect. Signed-off-by:
-
Christoph Lameter authored
The names_cachep is used for getname() and putname(). So lets put it into fs.h near those two definitions. Signed-off-by:
-
Christoph Lameter authored
fs_cachep is only used in kernel/exit.c and in kernel/fork.c. It is used to store fs_struct items so it should be placed in linux/fs_struct.h Signed-off-by:
-
Christoph Lameter authored
filp_cachep is only used in fs/file_table.c and in fs/dcache.c where it is defined. Move it to related definitions in linux/file.h. Signed-off-by:
-
Christoph Lameter authored
Proper place is in file.h since files_cachep uses are rated to file I/O. Signed-off-by:
-
Christoph Lameter authored
vm_area_cachep is used to store vm_area_structs. So move to mm.h. Signed-off-by:
-
Christoph Lameter authored
Move sighand_cachep definitioni to linux/signal.h The sighand cache is only used in fs/exec.c and kernel/fork.c. It is defined in kernel/fork.c but only used in fs/exec.c. The sighand_cachep is related to signal processing. So add the definition to signal.h. Signed-off-by:
-
Christoph Lameter authored
Remove bio_cachep from slab.h - it no longer exists. Signed-off-by:
-
Adrian Bunk authored
This patch makes the needlessly global "global_faults" static. Signed-off-by:
Adrian Bunk <bunk@stusta.de> Signed-off-by:
-
Christian Krafft authored
When booting a NUMA system with nodes that have no memory (eg by limiting memory), bootmem_alloc_core tried to find pages in an uninitialized bootmem_map. This caused a null pointer access. This fix adds a check, so that NULL is returned. That will enable the caller (bootmem_alloc_nopanic) to alloc memory on other without a panic. Signed-off-by:
Christian Krafft <krafft@de.ibm.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Martin Bligh <mbligh@google.com> Signed-off-by:
-
Alan Stern authored
The patch (as824b) makes percpu_free() ignore NULL arguments, as one would expect for a deallocation routine. (Note that free_percpu is #defined as percpu_free in include/linux/percpu.h.) A few callers are updated to remove now-unneeded tests for NULL. A few other callers already seem to assume that passing a NULL pointer to percpu_free() is okay! The patch also removes an unnecessary NULL check in percpu_depopulate(). Signed-off-by:
Alan Stern <stern@rowland.harvard.edu> Signed-off-by:
-
Christoph Hellwig authored
Node-aware allocation of skbs for the receive path. Details: - __alloc_skb gets a new node argument and cals the node-aware slab functions with it. - netdev_alloc_skb passed the node number it gets from dev_to_node to it, everyone else passes -1 (any node) Signed-off-by:Christoph Hellwig <hch@lst.de> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: "David S. Miller" <davem@davemloft.net> Signed-off-by:
-
Christoph Hellwig authored
For node-aware skb allocations we need information about the node in struct net_device or struct device. Davem suggested to put it into struct device which this patch does. In particular: - struct device gets a new int numa_node member if CONFIG_NUMA is set - there are two new helpers, dev_to_node and set_dev_node to transparently deal with the non-numa case - for pci devices the node-info is set to the value we get from pcibus_to_node. Note that for some architectures pcibus_to_node doesn't work yet at the time we call it currently. This is harmless and will just mean skb allocations aren't node-local on this architectures until the implementation of pcibus_to_node on these architectures have been updated (There are patches for x86 and x86_64 floating around) [akpm@osdl.org: cleanup] Signed-off-by:
-
Christoph Hellwig authored
We have variants of kmalloc and kmem_cache_alloc that leave leak tracking to the caller. This is used for subsystem-specific allocators like skb_alloc. To make skb_alloc node-aware we need similar routines for the node-aware slab allocator, which this patch adds. Note that the code is rather ugly, but it mirrors the non-node-aware code 1:1: [akpm@osdl.org: add module export] Signed-off-by:
-
Suleiman Souhlal authored
It would be possible for /proc/swaps to not always print out the header: swapon /dev/hdc2 swapon /dev/hde2 swapoff /dev/hdc2 At this point /proc/swaps would not have a header. Signed-off-by:
Suleiman Souhlal <suleiman@google.com> Cc: Hugh Dickins <hugh@veritas.com> Signed-off-by:
-
Kirill Korotaev authored
OOM can panic due to the processes stuck in __alloc_pages() doing infinite rebalance loop while no memory can be reclaimed. OOM killer tries to kill some processes, but unfortunetaly, rebalance label was moved by someone below the TIF_MEMDIE check, so buddy allocator doesn't see that process is OOM-killed and it can simply fail the allocation :/ Observed in reality on RHEL4(2.6.9)+OpenVZ kernel when a user doing some memory allocation tricks triggered OOM panic. Signed-off-by:
Denis Lunev <den@sw.ru> Signed-off-by:
Kirill Korotaev <dev@openvz.org> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by:
-
Rik Bobbaers authored
mm is defined as vma->vm_mm, so use that. Acked-by:
Hugh Dickins <hugh@veritas.com> Signed-off-by:
-
Guillem Jover authored
Currently a user process cannot rise its own oom_adj value (i.e. unprotecting itself from the OOM killer). As this value is stored in the task structure it gets inherited and the unprivileged childs will be unable to rise it. The EPERM will be handled by the generic proc fs layer, as only processes with the proper caps or the owner of the process will be able to write to the file. So we allow only the processes with CAP_SYS_RESOURCE to lower the value, otherwise it will get an EACCES which seems more appropriate than EPERM. Signed-off-by:
Guillem Jover <guillem.jover@nokia.com> Acked-by:
Andrea Arcangeli <andrea@novell.com> Signed-off-by:
-
Jeremy Fitzhardinge authored
kunmap_atomic() will call kpte_clear_flush with vaddr/ptep arguments which don't correspond if the vaddr is just a normal lowmem address (ie, not in the KMAP area). This patch makes sure that the pte is only cleared if kmap area was actually used for the mapping. Signed-off-by:
Jeremy Fitzhardinge <jeremy@xensource.com> Signed-off-by:
Rusty Russell <rusty@rustcorp.com.au> Cc: Zachary Amsden <zach@vmware.com> Signed-off-by:
-
Peter Zijlstra authored
Make kmap_atomic/kunmap_atomic denote a pagefault disabled scope. All non trivial implementations already do this anyway. Signed-off-by:
Peter Zijlstra <a.p.zijlstra@chello.nl> Acked-by:
Nick Piggin <npiggin@suse.de> Signed-off-by:
-
Peter Zijlstra authored
Introduce pagefault_{disable,enable}() and use these where previously we did manual preempt increments/decrements to make the pagefault handler do the atomic thing. Currently they still rely on the increased preempt count, but do not rely on the disabled preemption, this might go away in the future. (NOTE: the extra barrier() in pagefault_disable might fix some holes on machines which have too many registers for their own good) [heiko.carstens@de.ibm.com: s390 fix] Signed-off-by:Heiko Carstens <heiko.carstens@de.ibm.com> Signed-off-by:
-
Peter Zijlstra authored
In light of the recent pagefault and filemap_copy_from_user work I've gone through all the arch pagefault handlers to make sure the inc_preempt_count() 'feature' works as expected. Several sections of code (including the new filemap_copy_from_user) rely on the fact that faults do not take locks under increased preempt count. arch/x86_64 - good arch/powerpc - good arch/cris - fixed arch/i386 - good arch/parisc - fixed arch/sh - good arch/sparc - good arch/s390 - good arch/m68k - fixed arch/ppc - good arch/alpha - fixed arch/mips - good arch/sparc64 - good arch/ia64 - good arch/arm - fixed arch/um - good arch/avr32 - good arch/h8300 - NA arch/m32r - good arch/v850 - good arch/frv - fixed arch/m68knommu - NA arch/arm26 - fixed arch/sh64 - fixed arch/xtensa - good Signed-off-by:
-
Paul Menage authored
When using numa=fake on non-NUMA hardware there is no benefit to having the alien caches, and they consume much memory. Add a kernel boot option to disable them. Christoph sayeth "This is good to have even on large NUMA. The problem is that the alien caches grow by the square of the size of the system in terms of nodes." Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by:
-
Ravikiran G Thirumalai authored
Here's an attempt towards doing away with lock_cpu_hotplug in the slab subsystem. This approach also fixes a bug which shows up when cpus are being offlined/onlined and slab caches are being tuned simultaneously. http://marc.theaimsgroup.com/?l=linux-kernel&m=116098888100481&w=2 The patch has been stress tested overnight on a 2 socket 4 core AMD box with repeated cpu online and offline, while dbench and kernbench process are running, and slab caches being tuned at the same time. There were no lockdep warnings either. (This test on 2,6.18 as 2.6.19-rc crashes at __drain_pages http://marc.theaimsgroup.com/?l=linux-kernel&m=116172164217678&w=2 ) The approach here is to hold cache_chain_mutex from CPU_UP_PREPARE until CPU_ONLINE (similar in approach as worqueue_mutex) . Slab code sensitive to cpu_online_map (kmem_cache_create, kmem_cache_destroy, slabinfo_write, __cache_shrink) is already serialized with cache_chain_mutex. (This patch lengthens cache_chain_mutex hold time at kmem_cache_destroy to cover this). This patch also takes the cache_chain_sem at kmem_cache_shrink to protect sanity of cpu_online_map at __cache_shrink, as viewed by slab. (kmem_cache_shrink->__cache_shrink->drain_cpu_caches). But, really, kmem_cache_shrink is used at just one place in the acpi subsystem! Do we really need to keep kmem_cache_shrink at all? Another note. Looks like a cpu hotplug event can send CPU_UP_CANCELED to a registered subsystem even if the subsystem did not receive CPU_UP_PREPARE. This could be due to a subsystem registered for notification earlier than the current subsystem crapping out with NOTIFY_BAD. Badness can occur with in the CPU_UP_CANCELED code path at slab if this happens (The same would apply for workqueue.c as well). To overcome this, we might have to use either a) a per subsystem flag and avoid handling of CPU_UP_CANCELED, or b) Use a special notifier events like LOCK_ACQUIRE/RELEASE as Gautham was using in his experiments, or c) Do not send CPU_UP_CANCELED to a subsystem which did not receive CPU_UP_PREPARE. I would prefer c). Signed-off-by:
Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by:
Shai Fultheim <shai@scalex86.org> Signed-off-by:
-
Kevin Hilman authored
When CONFIG_SLAB_DEBUG is used in combination with ARCH_SLAB_MINALIGN, some debug flags should be disabled which depend on BYTES_PER_WORD alignment. The disabling of these debug flags is not properly handled when BYTES_PER_WORD < ARCH_SLAB_MEMALIGN < cache_line_size() This patch fixes that and also adds an alignment check to cache_alloc_debugcheck_after() when ARCH_SLAB_MINALIGN is used. Signed-off-by:
Kevin Hilman <khilman@mvista.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by:
-
Chen, Kenneth W authored
Imprecise RSS accounting is an irritating ill effect with pt sharing. After consulted with several VM experts, I have tried various methods to solve that problem: (1) iterate through all mm_structs that share the PT and increment count; (2) keep RSS count in page table structure and then sum them up at reporting time. None of the above methods yield any satisfactory implementation. Since process RSS accounting is pure information only, I propose we don't count them at all for hugetlb page. rlimit has such field, though there is absolutely no enforcement on limiting that resource. One other method is to account all RSS at hugetlb mmap time regardless they are faulted or not. I opt for the simplicity of no accounting at all. Hugetlb page are special, they are reserved up front in global reservation pool and is not reclaimable. From physical memory resource point of view, it is already consumed regardless whether there are users using them. If the concern is that RSS can be used to control resource allocation, we already can specify hugetlb fs size limit and sysadmin can enforce that at mount time. Combined with the two points mentioned above, I fail to see if there is anything got affected because of this patch. Signed-off-by:
Ken Chen <kenneth.w.chen@intel.com> Acked-by:
-
Chen, Kenneth W authored
Following up with the work on shared page table done by Dave McCracken. This set of patch target shared page table for hugetlb memory only. The shared page table is particular useful in the situation of large number of independent processes sharing large shared memory segments. In the normal page case, the amount of memory saved from process' page table is quite significant. For hugetlb, the saving on page table memory is not the primary objective (as hugetlb itself already cuts down page table overhead significantly), instead, the purpose of using shared page table on hugetlb is to allow faster TLB refill and smaller cache pollution upon TLB miss. With PT sharing, pte entries are shared among hundreds of processes, the cache consumption used by all the page table is smaller and in return, application gets much higher cache hit ratio. One other effect is that cache hit ratio with hardware page walker hitting on pte in cache will be higher and this helps to reduce tlb miss latency. These two effects contribute to higher application performance. Signed-off-by:
-
Andrew Morton authored
Despaghettify balance_pdgat() a bit. Signed-off-by:
-
Nick Piggin authored
Add an arch_alloc_page to match arch_free_page. Signed-off-by:
-
Ashwin Chaugule authored
The new swap token patches replace the current token traversal algo. The old algo had a crude timeout parameter that was used to handover the token from one task to another. This algo, transfers the token to the tasks that are in need of the token. The urgency for the token is based on the number of times a task is required to swap-in pages. Accordingly, the priority of a task is incremented if it has been badly affected due to swap-outs. To ensure that the token doesnt bounce around rapidly, the token holders are given a priority boost. The priority of tasks is also decremented, if their rate of swap-in's keeps reducing. This way, the condition to check whether to pre-empt the swap token, is a matter of comparing two task's priority fields. [akpm@osdl.org: cleanups] Signed-off-by:
Ashwin Chaugule <ashwin.chaugule@celunite.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by:
-
Ashwin Chaugule authored
Make sure the contention for the token happens _before_ any read-in and kicks the swap-token algo only when the VM is under pressure. Signed-off-by:
-
Nick Piggin authored
Some drivers are returning OOM when it is not in response to a memory shortage. Signed-off-by:
-
Nick Piggin authored
Don't cause all threads in all other thread groups to gain TIF_MEMDIE otherwise we'll get a thundering herd eating our memory reserve. This may not be the optimal scheme, but it fits our policy of allowing just one TIF_MEMDIE in the system at once. Signed-off-by:
-
Nick Piggin authored
Clean up the OOM killer messages to be more consistent. Signed-off-by:
-
Nick Piggin authored
Abort the kill if any of our threads have OOM_DISABLE set. Having this test here also prevents any OOM_DISABLE child of the "selected" process from being killed. Signed-off-by:
-
Paul Jackson authored
Rearrange the struct members in the 'struct zonelist_cache' structure, so as to put the readonly (once initialized) z_to_n[] array first, where it will come right after the zones[] array in struct zonelist. This pretty much eliminates the chance that the two frequently written elements of 'struct zonelist_cache', the fullzones bitmap and last_full_zap times, will end up on the same cache line as the performance sensitive, frequently read, never (after init) written zones[] array. Keeping frequently written data off frequently read cache lines is good for performance. Thanks to Rohit Seth for the suggestion. Signed-off-by:
Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Paul Menage <menage@google.com> Signed-off-by:
-
Paul Jackson authored
Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: -
Christoph Lameter authored
The zone table is mostly not needed. If we have a node in the page flags then we can get to the zone via NODE_DATA() which is much more likely to be already in the cpu cache. In case of SMP and UP NODE_DATA() is a constant pointer which allows us to access an exact replica of zonetable in the node_zones field. In all of the above cases there will be no need at all for the zone table. The only remaining case is if in a NUMA system the node numbers do not fit into the page flags. In that case we make sparse generate a table that maps sections to nodes and use that table to to figure out the node number. This table is sized to fit in a single cache line for the known 32 bit NUMA platform which makes it very likely that the information can be obtained without a cache miss. For sparsemem the zone table seems to be have been fairly large based on the maximum possible number of sections and the number of zones per node. There is some memory saving by removing zone_table. The main benefit is to reduce the cache foootprint of the VM from the frequent lookups of zones. Plus it simplifies the page allocator. [akpm@osdl.org: build fix] Signed-off-by:
-
