Remove DISCONTIGMEM support from x86-32
RFC PATCH 1/1 Remove DISCONTIGMEM support from x86-32
Hello,
This patchset removes DISCONTIGMEM support from x86-32 leaving it with FLATMEM for UMA and SPARSEMEM for NUMA. At this point, DISCONTIGMEM doesn't really have much benefit over SPARSEMEM. The only remaining benefit is that it can have finer granularity with memsections regardless of the number of bits available in page->flags.
There are two obstacles to removing DISCONTIGMEM support in x86-32. One is that ARCH_SUPPORTS_MEMORY_FAILURE is dependent on !SPARSEMEM, so if a NUMA kernel wants to use memory failure, it has to use DISCONTIGMEM. At least SLE11SP1 kernel seems to have chosen that route.
This is solved by reducing the number of page->flags bits used for memsection selection by one, which I think is the right thing to do regardless of this change. Please read the patch description on the second patch for details.
The other obstable is that due to the limited number of page->flags bits available and the relatively large physical address area when PAE is enabled, the granularity of sparse memsections is somewhat coarse. It was originally 512MiB but increased to 1GiB with the second patch.
For unaligned memsections, this can result in some amount of wasted struct page allocations, which I think shouldn't be a problem but not completely sure and thus the RFC status.
This patchset contains the following four patches.
0001-x86-Clean-up-memory-model-related-configs-in-arch-x8.patch 0002-x86-32-Increment-SECTION_SIZE_BITS-to-30-when-X86_PA.patch 0003-x86-32-Remove-restrictions-on-ARCH_SUPPORTS_MEMORY_F.patch 0004-x86-32-NUMA-Remove-support-for-DISCONTIGMEM.patch
0001 is a cleanup which I'll probably queue regardless of this series unless there are objections. 0002-0003 solve the first obstacle described above and 0004 removes DISCONTIGMEM.
The patchset is also available in the following git branch.
git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git x86_32-remove-DISCONTIGMEM
diffstat follows.
arch/x86/Kconfig | 49 +++++++++------------------------ arch/x86/include/asm/mmzone_32.h | 48 -------------------------------- arch/x86/include/asm/pgtable_32.h | 5 --- arch/x86/include/asm/sparsemem.h | 2 - arch/x86/mm/numa_32.c | 56 ++++---------------------------------- 5 files changed, 22 insertions(+), 138 deletions(-)
Thanks.
Signature · 1 line
tejun
PATCH 1/4 x86: Clean up memory model related configs in arch/x86/Kconfig
* Remove bogus dependency on ARCH_SELECT_MEMORY_MODEL from ARCH_FLATMEM_ENABLE. ENABLE configs don't interfere with SELECT_MEMORY_MODEL. They just need to indicate whether the specific memory model is supported.
* Relocate HAVE_ARCH_ALLOC_REMAP, ARCH_PROC_KCORE_TEXT and ARCH_SPARSEMEM_DEFAULT so that memory model related configs are together in consistent order.
--- arch/x86/Kconfig | 26 +++++++++++++------------- 1 files changed, 13 insertions(+), 13 deletions(-)
arch/x86/Kconfig +13 -13
@@ -1223,6 +1223,10 @@ config HAVE_ARCH_BOOTMEM def_bool y depends on X86_32 && NUMA +config HAVE_ARCH_ALLOC_REMAP + def_bool y + depends on X86_32 && NUMA + config ARCH_HAVE_MEMORY_PRESENT def_bool y depends on X86_32 && DISCONTIGMEM
@@ -1231,13 +1235,9 @@ config NEED_NODE_MEMMAP_SIZE def_bool y depends on X86_32 && (DISCONTIGMEM || SPARSEMEM) -config HAVE_ARCH_ALLOC_REMAP - def_bool y - depends on X86_32 && NUMA - config ARCH_FLATMEM_ENABLE def_bool y - depends on X86_32 && ARCH_SELECT_MEMORY_MODEL && !NUMA + depends on X86_32 && !NUMA config ARCH_DISCONTIGMEM_ENABLE def_bool y
@@ -1247,20 +1247,16 @@ config ARCH_DISCONTIGMEM_DEFAULT def_bool y depends on NUMA && X86_32 -config ARCH_PROC_KCORE_TEXT - def_bool y - depends on X86_64 && PROC_KCORE - -config ARCH_SPARSEMEM_DEFAULT - def_bool y - depends on X86_64 - config ARCH_SPARSEMEM_ENABLE def_bool y depends on X86_64 || NUMA || (EXPERIMENTAL && X86_32) || X86_32_NON_STANDARD select SPARSEMEM_STATIC if X86_32 select SPARSEMEM_VMEMMAP_ENABLE if X86_64 +config ARCH_SPARSEMEM_DEFAULT + def_bool y + depends on X86_64 + config ARCH_SELECT_MEMORY_MODEL def_bool y depends on ARCH_SPARSEMEM_ENABLE
@@ -1269,6 +1265,10 @@ config ARCH_MEMORY_PROBE def_bool X86_64 depends on MEMORY_HOTPLUG +config ARCH_PROC_KCORE_TEXT + def_bool y + depends on X86_64 && PROC_KCORE + config ILLEGAL_POINTER_VALUE hex default 0 if X86_32
Signature · 1 line
1.7.1
Re: [PATCH 1/4] x86: Clean up memory model related configs in arch/x86/Kconfig
Quoted 4 lines · Tejun Heo
* Remove bogus dependency on
ARCH_SELECT_MEMORY_MODELfromARCH_FLATMEM_ENABLE. ENABLE configs don't interfere withSELECT_MEMORY_MODEL. They just need to indicate whether the specific memory model is supported.
PATCH 2/4 x86-32: Increment SECTION_SIZE_BITS to 30 when X86_PAE
In reply to Tejun HeoWhen X86_PAE is enabled, SECTION_SIZE_BITS is 29 and MAX_PHYSADDR_BITS 36 making SECTIONS_WIDTH 7, which is different from !X86_PAE case where the width is 6. On x86, when all the related options are enabled, NR_PAGEFLAGS reaches 26 and thus can accomodate only 6 more bits, not 7.
This direcrepancy leads to confusing Kconfig restrictions like disabling of ARCH_SUPPORTS_MEMORY_FAILURE depending on combinations of options.
This patch bumps SECTION_SIZE_BITS to 30 when X86_PAE. This change increases sparse memory model granuality from 512MiB to 1GiB which in turn can increase the amount of wasted memory in memmap when the memory sections aren't aligned. The amount of wastage can be a few MiB per unaligned section.
--- arch/x86/include/asm/sparsemem.h | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/include/asm/sparsemem.h +1 -1
@@ -16,7 +16,7 @@#ifdef CONFIG_X86_32 # ifdef CONFIG_X86_PAE -# define SECTION_SIZE_BITS 29 +# define SECTION_SIZE_BITS 30# define MAX_PHYSADDR_BITS 36# define MAX_PHYSMEM_BITS 36# else
Signature · 1 line
1.7.1
PATCH 3/4 x86-32: Remove restrictions on ARCH_SUPPORTS_MEMORY_FAILURE
In reply to Tejun Heod949f36f18 (x86: Fix hwpoison code related build failure on 32-bit NUMAQ) disabled ARCH_SUPPORTS_MEMORY_FAILURE on some 32bit configurations because the extra page flag usage triggered build failure when combined with memory section/node bits.
The root cause of build failures depending on config options is SECTIONS_WIDTH difference depending on X86_PAE. As the previous patch removes the problem, we can remove the restrictions on MEMORY_FAILURE.
This makes the configuration more consistent. For example, TRANSPARENT_HUGEPAGE also consumes a page flag but didn't trigger build failure thanks to the removal of PG_buddy and Kconfig ends up enforcing rather arbitrary preference for TRANSPARENT_HUGEPAGE over MEMORY_FAILURE.
--- arch/x86/Kconfig | 11 +---------- 1 files changed, 1 insertions(+), 10 deletions(-)
arch/x86/Kconfig +1 -10
@@ -446,16 +446,6 @@ config X86_NUMAQ of Flat Logical. You will need a new lynxer.elf file to flash your firmware with - send email to <[email protected]>. -config X86_SUPPORTS_MEMORY_FAILURE - def_bool y - # MCE code calls memory_failure(): - depends on X86_MCE - # On 32-bit this adds too big of NODES_SHIFT and we run out of page flags: - depends on !X86_NUMAQ - # On 32-bit SPARSEMEM adds too big of SECTIONS_WIDTH: - depends on X86_64 || !SPARSEMEM - select ARCH_SUPPORTS_MEMORY_FAILURE - config X86_VISWS bool "SGI 320/540 (Visual Workstation)" depends on X86_32 && PCI && X86_MPPARSE && PCI_GODIRECT
@@ -845,6 +835,7 @@ config X86_REROUTE_FOR_BROKEN_BOOT_IRQS config X86_MCE bool "Machine Check / overheating reporting" + select ARCH_SUPPORTS_MEMORY_FAILURE ---help--- Machine Check support allows the processor to notify the kernel if it detects a problem (e.g. overheating, data corruption).
Signature · 1 line
1.7.1
PATCH 4/4 x86-32, NUMA: Remove support for DISCONTIGMEM
In reply to Tejun HeoThere's no meaningful performance difference between DISCONTIGMEM and SPARSEMEM, and sparse memory model has been stable for quite some time now.
At this point, the only difference between the two is that sparsemem might waste some amount of memory for memmap if the memory sections aren't aligned to 1 << SECTION_SIZE_BITS which is 1GiB is PAE is enabled and 64MiB otherwise.
Drop DISCONTIGMEM support and use SPARSEMEM on NUMA configurations. Note that due to the shortage of vmalloc space, SPARSEMEM_VMMEMAP can't be enabled and thus there's performance difference between FLATMEM and SPARSEMEM, so x86-32 can't drop FLATMEM and just use sparse like x86-64.
The following is from Christoph Lameter's page alloc benchmark module.
DISCONTIGMEM
Single thread testing
=====================
1. Repeatedly allocate then free test
1000 times alloc_page(,0) -> 1017 cycles __free_pages(,0)-> 359 cycles
1000 times alloc_page(,1) -> 1041 cycles __free_pages(,1)-> 424 cycles
1000 times alloc_page(,2) -> 1215 cycles __free_pages(,2)-> 565 cycles
1000 times alloc_page(,3) -> 1256 cycles __free_pages(,3)-> 613 cycles
1000 times alloc_page(,4) -> 1554 cycles __free_pages(,4)-> 697 cycles
1000 times alloc_page(,5) -> 1990 cycles __free_pages(,5)-> 1114 cycles
1000 times alloc_page(,6) -> 3072 cycles __free_pages(,6)-> 1788 cycles
1000 times alloc_page(,7) -> 5215 cycles __free_pages(,7)-> 5550 cycles
2. alloc/free test
1000 times alloc( ,0)/free -> 662 cycles
1000 times alloc( ,1)/free -> 853 cycles
1000 times alloc( ,2)/free -> 926 cycles
1000 times alloc( ,3)/free -> 1013 cycles
1000 times alloc( ,4)/free -> 1249 cycles
1000 times alloc( ,5)/free -> 1614 cycles
1000 times alloc( ,6)/free -> 2454 cycles
1000 times alloc( ,7)/free -> 4000 cycles
Concurrent allocs
=================
Page alloc N*alloc N*free(0): 0=5143/5868 1=5245/6051 2=5199/5982 3=5220/6045 4=5191/5876 5=5270/6066 6=966/5902 7=5243/6035 Average=4684/5978 Page alloc N*alloc N*free(1): 0=19537/24032 1=19569/24080 2=19555/24039 3=19580/24073 4=19590/24075 5=19595/24080 6=19585/24073 7=19586/24079 Average=19574/24066 Page alloc N*alloc N*free(2): 0=19466/24490 1=19433/24439 2=19470/24502 3=19472/24493 4=19450/24496 5=19441/24483 6=19446/24488 7=19428/24478 Average=19451/24484 Page alloc N*alloc N*free(3): 0=19072/25386 1=19036/25381 2=19029/25377 3=19034/25376 4=18799/25382 5=18905/25347 6=18880/25266 7=18891/25379 Average=18956/25362 Page alloc N*alloc N*free(4): 0=19252/26454 1=19183/26451 2=19199/26454 3=19205/26450 4=19180/26458 5=19170/26437 6=19131/26373 7=19179/26447 Average=19188/26440 ----Fastpath---
Page N*(alloc free)(0): 0=639 1=639 2=639 3=638 4=641 5=642 6=655 7=638 Average=641
Page N*(alloc free)(1): 0=36697 1=36717 2=36746 3=36747 4=36748 5=36753 6=36751 7=36751 Average=36739
Page N*(alloc free)(2): 0=35753 1=35756 2=35780 3=35757 4=35782 5=35784 6=35778 7=35779 Average=35771
Page N*(alloc free)(3): 0=37499 1=37527 2=37544 3=37547 4=37545 5=37550 6=37547 7=37543 Average=37538
Page N*(alloc free)(4): 0=36374 1=36381 2=36378 3=36395 4=36391 5=36390 6=36390 7=36391 Average=36386
Remote free test
================
N*remote free(0): 0=5/4157 1=4907/0 2=4838/0 3=4846/0 4=4891/0 5=4883/0 6=4880/0 7=4904/0 Average=4269/519
N*remote free(1): 0=4/5832 1=15285/0 2=15243/0 3=15270/0 4=15299/0 5=15306/0 6=15304/0 7=15294/0 Average=13376/729
N*remote free(2): 0=4/9441 1=15489/0 2=15492/0 3=15483/0 4=15440/0 5=15476/0 6=15468/0 7=15471/0 Average=13540/1180
N*remote free(3): 0=5/13038 1=15725/0 2=15727/0 3=15711/0 4=15713/0 5=15710/0 6=15704/0 7=15710/0 Average=13751/1629
N*remote free(4): 0=6/21888 1=16137/0 2=16142/0 3=16157/0 4=16156/0 5=16160/0 6=16156/0 7=16158/0 Average=14134/2736
SPARSEMEM
Single thread testing
=====================
1. Repeatedly allocate then free test
1000 times alloc_page(,0) -> 1033 cycles __free_pages(,0)-> 358 cycles
1000 times alloc_page(,1) -> 1061 cycles __free_pages(,1)-> 477 cycles
1000 times alloc_page(,2) -> 1191 cycles __free_pages(,2)-> 578 cycles
1000 times alloc_page(,3) -> 1219 cycles __free_pages(,3)-> 604 cycles
1000 times alloc_page(,4) -> 1440 cycles __free_pages(,4)-> 681 cycles
1000 times alloc_page(,5) -> 1820 cycles __free_pages(,5)-> 1179 cycles
1000 times alloc_page(,6) -> 2757 cycles __free_pages(,6)-> 2062 cycles
1000 times alloc_page(,7) -> 4601 cycles __free_pages(,7)-> 5539 cycles
2. alloc/free test
1000 times alloc( ,0)/free -> 649 cycles
1000 times alloc( ,1)/free -> 850 cycles
1000 times alloc( ,2)/free -> 886 cycles
1000 times alloc( ,3)/free -> 984 cycles
1000 times alloc( ,4)/free -> 1161 cycles
1000 times alloc( ,5)/free -> 1458 cycles
1000 times alloc( ,6)/free -> 2145 cycles
1000 times alloc( ,7)/free -> 3293 cycles
Concurrent allocs
=================
Page alloc N*alloc N*free(0): 0=5054/6282 1=5032/6212 2=4994/6246 3=5034/6295 4=4987/6057 5=928/6211 6=5103/6311 7=5067/6267 Average=4525/6235 Page alloc N*alloc N*free(1): 0=19551/24798 1=19599/24851 2=19609/24871 3=19599/24821 4=19604/24869 5=19615/24866 6=19615/24841 7=19616/24872 Average=19601/24848 Page alloc N*alloc N*free(2): 0=19332/24555 1=19342/24568 2=19363/24588 3=19363/24594 4=19362/24593 5=19365/24595 6=19360/24590 7=19360/24586 Average=19356/24584 Page alloc N*alloc N*free(3): 0=19145/25868 1=19099/25865 2=19094/25858 3=19089/25865 4=18941/25873 5=18985/25840 6=18962/25838 7=18932/25836 Average=19031/25855 Page alloc N*alloc N*free(4): 0=19202/27147 1=19151/27142 2=19155/27143 3=19144/27141 4=19104/27145 5=19051/27082 6=19068/27113 7=19080/27139 Average=19119/27132 ----Fastpath---
Page N*(alloc free)(0): 0=683 1=702 2=667 3=666 4=668 5=671 6=683 7=685 Average=678
Page N*(alloc free)(1): 0=37829 1=37843 2=37856 3=37843 4=37856 5=37857 6=37853 7=37854 Average=37849
Page N*(alloc free)(2): 0=36966 1=36952 2=36970 3=36977 4=36988 5=36988 6=36990 7=36987 Average=36977
Page N*(alloc free)(3): 0=40647 1=40626 2=40662 3=40653 4=40664 5=40665 6=40667 7=40660 Average=40656
Page N*(alloc free)(4): 0=38194 1=38153 2=38145 3=38152 4=38178 5=38170 6=38174 7=38182 Average=38169
Remote free test
================
N*remote free(0): 0=4/4335 1=4791/0 2=4895/0 3=4886/0 4=4901/0 5=4856/0 6=4902/0 7=4909/0 Average=4268/542
N*remote free(1): 0=4/5972 1=15298/0 2=15275/0 3=15284/0 4=15314/0 5=15309/0 6=15313/0 7=15299/0 Average=13387/746
N*remote free(2): 0=4/9684 1=15441/0 2=15433/0 3=15418/0 4=15423/0 5=15388/0 6=15432/0 7=15430/0 Average=13496/1210
N*remote free(3): 0=5/13320 1=15716/0 2=15720/0 3=15716/0 4=15683/0 5=15645/0 6=15682/0 7=15688/0 Average=13732/1665
N*remote free(4): 0=6/22219 1=15946/0 2=15957/0 3=15949/0 4=15928/0 5=15924/0 6=15948/0 7=15942/0 Average=13950/2777
--- arch/x86/Kconfig | 18 ++---------- arch/x86/include/asm/mmzone_32.h | 48 ------------------------------- arch/x86/include/asm/pgtable_32.h | 5 +-- arch/x86/mm/numa_32.c | 56 ++++--------------------------------- 4 files changed, 10 insertions(+), 117 deletions(-)
arch/x86/Kconfig +3 -15
@@ -1218,26 +1218,14 @@ config HAVE_ARCH_ALLOC_REMAP def_bool y depends on X86_32 && NUMA -config ARCH_HAVE_MEMORY_PRESENT - def_bool y - depends on X86_32 && DISCONTIGMEM - config NEED_NODE_MEMMAP_SIZE def_bool y - depends on X86_32 && (DISCONTIGMEM || SPARSEMEM) + depends on X86_32 && SPARSEMEM config ARCH_FLATMEM_ENABLE def_bool y depends on X86_32 && !NUMA -config ARCH_DISCONTIGMEM_ENABLE - def_bool y - depends on NUMA && X86_32 - -config ARCH_DISCONTIGMEM_DEFAULT - def_bool y - depends on NUMA && X86_32 - config ARCH_SPARSEMEM_ENABLE def_bool y depends on X86_64 || NUMA || (EXPERIMENTAL && X86_32) || X86_32_NON_STANDARD
@@ -1246,11 +1234,11 @@ config ARCH_SPARSEMEM_ENABLE config ARCH_SPARSEMEM_DEFAULT def_bool y - depends on X86_64 + depends on X86_64 || (X86_32 && NUMA) config ARCH_SELECT_MEMORY_MODEL def_bool y - depends on ARCH_SPARSEMEM_ENABLE + depends on ARCH_FLATMEM_ENABLE config ARCH_MEMORY_PROBE def_bool X86_64
arch/x86/include/asm/mmzone_32.h +0 -48
54 lines not on this page · raw message
arch/x86/include/asm/pgtable_32.h +1 -4
@@ -74,10 +74,7 @@ do { \#endif /* !__ASSEMBLY__ */ -/* -* kern_addr_valid() is (1) for FLATMEM and (0) for -* SPARSEMEM and DISCONTIGMEM -*/ +/* kern_addr_valid() is (1) for FLATMEM and (0) for SPARSEMEM */ #ifdef CONFIG_FLATMEM #define kern_addr_valid (addr) (1)#else
arch/x86/mm/numa_32.c +6 -50
68 lines not on this page · raw message
Signature · 1 line
1.7.1