]> bbs.cooldavid.org Git - net-next-2.6.git/commit
powerpc/mm: Bump SECTION_SIZE_BITS from 16MB to 256MB
authorAnton Blanchard <anton@samba.org>
Thu, 25 Feb 2010 20:18:46 +0000 (20:18 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Wed, 7 Apr 2010 08:00:49 +0000 (18:00 +1000)
commit7545ba6f82924d4523f8f8a2baf2e517a750265d
tree7e2dc55ae231cb6a2a5122df6e18a4a1e452b1fe
parente9bbc8cde0e3c33b42ddbe1b02108cb5c97275eb
powerpc/mm: Bump SECTION_SIZE_BITS from 16MB to 256MB

The current setting for SECTION_SIZE_BITS is quite small compared to
everyone else:

arch/powerpc/include/asm/sparsemem.h:#define SECTION_SIZE_BITS  24

arch/sparc/include/asm/sparsemem.h:#define SECTION_SIZE_BITS    30
arch/ia64/include/asm/sparsemem.h:#define SECTION_SIZE_BITS     (30)
arch/s390/include/asm/sparsemem.h:#define SECTION_SIZE_BITS     28
arch/x86/include/asm/sparsemem.h:# define SECTION_SIZE_BITS     27

And it has proven to be an issue during boot on very large machines.
If hotplug memory is enabled, drivers/base/memory.c does this:

       for (i = 0; i < NR_MEM_SECTIONS; i++) {
                if (!present_section_nr(i))
                        continue;
                err = add_memory_block(0, __nr_to_section(i), MEM_ONLINE,
                                        0, BOOT);
                if (!ret)
                        ret = err;
        }

Which creates a sysfs directory for every 16MB of memory. As a result
I'm seeing up to 30 minutes spent here during boot:

c000000000248ee0 .__sysfs_add_one+0x28/0x128
c0000000002492a8 .sysfs_add_one+0x38/0x188
c000000000249c88 .create_dir+0x70/0x138
c000000000249d98 .sysfs_create_dir+0x48/0x78
c00000000032bad8 .kobject_add_internal+0x140/0x308
c00000000032beb4 .kobject_init_and_add+0x4c/0x68
c00000000046c2c0 .sysdev_register+0xa0/0x220
c00000000047b1dc .add_memory_block+0x124/0x1e8
c0000000008d1f28 .memory_dev_init+0xf4/0x168
c0000000008d1b64 .driver_init+0x50/0x64
c000000000890378 .do_basic_setup+0x40/0xd4

I assume there are some O(n^2) issues in sysfs as we add all the memory
nodes. Bumping SECTION_SIZE_BITS to 256 MB drops the time to about 10
seconds and results in a much smaller /sys.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/include/asm/sparsemem.h