Sparsemem Virtual Memmap V5

Following this email is the current state of the Sparsemem Virtual Memory Map support (description below). This version is essentially unchanged from V4. The ia64 crowd has decided they want a 4MB mapping size for the map which will come later, so this set includes only base page support for them. This patch set has been tested pretty heavily over the last few weeks.

The current aim is to bring a common virtually mapped mem_map to all architectures. This should facilitate the removal of the bespoke implementations from the architectures. This also brings performance improvements for most architecture making sparsmem vmemmap the more desirable memory model. The ultimate aim of this work is to expand sparsemem support to encompass all the features of the other memory models. This could allow us to drop support for and remove the other models in the longer term.

Below are some comparitive kernbench numbers for various architectures, comparing default memory model against SPARSEMEM VMEMMAP. All but ia64 show marginal improvement; we expect the ia64 figures to be sorted out when the larger mapping support returns.

x86-64 non-NUMA

             Base    VMEMAP    % change (-ve good)
User        85.07     84.84    -0.26
System      34.32     33.84    -1.39
Total      119.38    118.68    -0.59

ia64

             Base    VMEMAP    % change (-ve good)
User      1016.41   1016.93    0.05
System      50.83     51.02    0.36
Total     1067.25   1067.95    0.07

x86-64 NUMA
Base VMEMAP % change (-ve good)
User 30.77 431.73 0.22
System 45.39 43.98 -3.11
Total 476.17 475.71 -0.10

ppc64

             Base   VMEMAP    % change (-ve good)
User       488.77   488.35    -0.09
System      56.92    56.37    -0.97
Total      545.69   544.72    -0.18

Below are some AIM bencharks on IA64 and x86-64 (thank Bob). The seems pretty much flat as you would expect.

ia64 results 2 cpu non-numa 4Gb SCSI disk

Benchmark Version Machine Run Date
AIM Multiuser Benchmark - Suite VII "1.1" extreme Jun 1 07:17:24 2007

Tasks Jobs/Min JTI Real CPU Jobs/sec/task
1 98.9 100 58.9 1.3 1.6482
101 5547.1 95 106.0 79.4 0.9154
201 6377.7 95 183.4 158.3 0.5288
301 6932.2 95 252.7 237.3 0.3838
401 7075.8 93 329.8 316.7 0.2941
501 7235.6 94 403.0 396.2 0.2407
600 7387.5 94 472.7 475.0 0.2052

Benchmark Version Machine Run Date
AIM Multiuser Benchmark - Suite VII "1.1" vmemmap Jun 1 09:59:04 2007

Tasks Jobs/Min JTI Real CPU Jobs/sec/task
1 99.1 100 58.8 1.2 1.6509
101 5480.9 95 107.2 79.2 0.9044
201 6490.3 95 180.2 157.8 0.5382
301 6886.6 94 254.4 236.8 0.3813
401 7078.2 94 329.7 316.0 0.2942
501 7250.3 95 402.2 395.4 0.2412
600 7399.1 94 471.9 473.9 0.2055

open power 710 2 cpu, 4 Gb, SCSI and configured physically

Benchmark Version Machine Run Date
AIM Multiuser Benchmark - Suite VII "1.1" extreme May 29 15:42:53 2007

Tasks Jobs/Min JTI Real CPU Jobs/sec/task
1 25.7 100 226.3 4.3 0.4286
101 1096.0 97 536.4 199.8 0.1809
201 1236.4 96 946.1 389.1 0.1025
301 1280.5 96 1368.0 582.3 0.0709
401 1270.2 95 1837.4 771.0 0.0528
501 1251.4 96 2330.1 955.9 0.0416
601 1252.6 96 2792.4 1139.2 0.0347
701 1245.2 96 3276.5 1334.6 0.0296
918 1229.5 96 4345.4 1728.7 0.0223

Benchmark Version Machine Run Date
AIM Multiuser Benchmark - Suite VII "1.1" vmemmap May 30 07:28:26 2007

Tasks Jobs/Min JTI Real CPU Jobs/sec/task
1 25.6 100 226.9 4.3 0.4275
101 1049.3 97 560.2 198.1 0.1731
201 1199.1 97 975.6 390.7 0.0994
301 1261.7 96 1388.5 591.5 0.0699
401 1256.1 96 1858.1 771.9 0.0522
501 1220.1 96 2389.7 955.3 0.0406
601 1224.6 96 2856.3 1133.4 0.0340
701 1252.0 96 3258.7 1314.1 0.0298
915 1232.8 96 4319.7 1704.0 0.0225

amd64 2 2-core, 4Gb and SATA

Benchmark Version Machine Run Date
AIM Multiuser Benchmark - Suite VII "1.1" extreme Jun 2 03:59:48 2007

Tasks Jobs/Min JTI Real CPU Jobs/sec/task
1 13.0 100 446.4 2.1 0.2173
101 533.4 97 1102.0 110.2 0.0880
201 578.3 97 2022.8 220.8 0.0480
301 583.8 97 3000.6 332.3 0.0323
401 580.5 97 4020.1 442.2 0.0241
501 574.8 98 5072.8 558.8 0.0191
600 566.5 98 6163.8 671.0 0.0157

Benchmark Version Machine Run Date
AIM Multiuser Benchmark - Suite VII "1.1" vmemmap Jun 3 04:19:31 2007

Tasks Jobs/Min JTI Real CPU Jobs/sec/task
1 13.0 100 447.8 2.0 0.2166
101 536.5 97 1095.6 109.7 0.0885
201 567.7 97 2060.5 219.3 0.0471
301 582.1 96 3009.4 330.2 0.0322
401 578.2 96 4036.4 442.4 0.0240
501 585.1 98 4983.2 555.1 0.0195
600 565.5 98 6175.2 660.6 0.0157

This stack is against v2.6.22-rc6-mm1. It has been compile, boot and tested on x86_64, ia64 and PPC64.

Andrew, please consider for -mm.

Note that I am away from my keyboard all of next week, but I figured it better to get this out for testing.

-apw

===
SPARSEMEM is a pretty nice framework that unifies quite a bit of code over all the arches. It would be great if it could be the default so that we can get rid of various forms of DISCONTIG and other variations on memory maps. So far what has hindered this are the additional lookups that SPARSEMEM introduces for virt_to_page and page_address. This goes so far that the code to do this has to be kept in a separate function and cannot be used inline.

This patch introduces a virtual memmap mode for SPARSEMEM, in which the memmap is mapped into a virtually contigious area, only the active sections are physically backed. This allows virt_to_page page_address and cohorts become simple shift/add operations. No page flag fields, no table lookups, nothing involving memory is required.

The two key operations pfn_to_page and page_to_page become:

   #define __pfn_to_page(pfn)      (vmemmap + (pfn))
   #define __page_to_pfn(page)     ((page) - vmemmap)

By having a virtual mapping for the memmap we allow simple access without wasting physical memory. As kernel memory is typically already mapped 1:1 this introduces no additional overhead. The virtual mapping must be big enough to allow a struct page to be allocated and mapped for all valid physical pages. This vill make a virtual memmap difficult to use on 32 bit platforms that support 36 address bits.

However, if there is enough virtual space available and the arch already maps its 1-1 kernel space using TLBs (f.e. true of IA64 and x86_64) then this technique makes SPARSEMEM lookups even more efficient than CONFIG_FLATMEM. FLATMEM needs to read the contents of the mem_map variable to get the start of the memmap and then add the offset to the required entry. vmemmap is a constant to which we can simply add the offset.

This patch has the potential to allow us to make SPARSMEM the default (and even the only) option for most systems. It should be optimal on UP, SMP and NUMA on most platforms. Then we may even be able to remove the other memory models: FLATMEM, DISCONTIG etc.

V4->V5
- IA64 16MB support shelved
- rebase to current -mm

V3->V4
- SPARC64 support -- from Dave Miller
- PPC64 support -- from Andy Whitcroft
- sparsemem precense/valid split
- rename Kconfig options into SPARSEMEM configuration name space
- redundant vmemmap alignment removed
- split out PMD support to x86_64
- x86_64 Kconfig dependancies
- ia64 Kconfig dependancies
- sparc64 dependancies, cleanup defines
- cleanup function names _pop_ -> _populate_
- markup __meminit
- cleanup style
- whitespace cleanups

V2->V3
- Add IA64 16M vmemmap size support (reduces TLB pressure)
- Add function to test for eventual node/node vmemmap overlaps
- Upper / Lower boundary fix.

V1->V2
- Support for PAGE_SIZE vmemmap which allows the general use of of virtual memmap on any MMU capable platform (enabled IA64 support).
- Fix various issues as suggested by Dave Hansen.
- Add comments and error handling.