From: Mike Rapoport <r...@linux.ibm.com>
Hi,
This is yet another attempt to enable SPARSEMEM on ARC.
I've boot tested it on nSIM with haps_hs_defconfig with highmem and
sparsemem enabled.
With sparsemem the kernel text becomes a bit smaller, but bss and data are
slightly increased:
$ size discontig/vmlinux sparse/vmlinux
text data bss dec hex filename
4429390 785456 244580 5459426 534de2 discontig/vmlinux
4415099 786224 244844 5446167 531a17 sparse/vmlinux
I've also added a dummy global functions to wrap pfn_valid(), page_to_pfn()
and pfn_to_page(). Judging by objdump, sparsemem is a bit more efficient:
DISCONTIGMEM SPARSEMEM
<pfn_to_page>:
seths r2,0x3ffff,r0 lsr r2,r0,0xe
mpy r2,r2,1896 mpy r0,r0,0x24
add r3,r2,0x8050066c add3 r2,0x80529d1c,r2
add_s r2,r2,0x80500668 ld_s r2,[r2,0]
ld_s r3,[r3,0] bmskn r2,r2,0x3
sub_s r0,r0,r3 j_s.d [blink]
ld_s r2,[r2,0] add_s r0,r0,r2
mpy r0,r0,0x24 nop_s
j_s.d [blink]
add_s r0,r0,r2
<page_to_pfn>:
ld_s r2,[r0,0] ld_s r2,[r0,0]
lsr_s r2,r2,0x1f lsr_s r2,r2,0x1b
mpy r2,r2,1896 add3 r2,0x80529d1c,r2
add r3,r2,0x80500668 ld_s r2,[r2,0]
add_s r2,r2,0x8050066c bmskn r2,r2,0x3
ld_s r3,[r3,0] sub_s r0,r0,r2
sub_s r0,r0,r3 asr_s r0,r0,0x2
ld_s r2,[r2,0] mpy r0,r0,0x38e38e39
asr_s r0,r0,0x2 j_s [blink]
mpy r0,r0,0x38e38e39
j_s.d [blink]
add_s r0,r0,r2
nop_s
<pfn_valid>:
cmp_s r0,0x3ffff lsr_s r0,r0,0xe
mov_s r2,0 brhs.nt r0,0x20,24
mov.ls r2,0x768 add3 r0,0x80529d1c,r0
add_s r2,r2,0x80500814 breq_s r0,0,12
ld.as r3,[r2,-106] ld_s r0,[r0,0]
ld.as r2,[r2,-104] j_s.d [blink]
add_s r2,r2,r3 xbfu r0,r0,0x1
j_s.d [blink] j_s.d [blink]
seths r0,r2,r0 mov_s r0,0
nop_s
Still, SPARSEMEM has an issue with potentially wasted memory allocated for
the memory map. The memory maps are allocated for each present section,
which means that if part of the section is not populated we'll have a bunch
of unused 'struct page' objects. The smaller the section size, the smaller
is memory overhead, but the section size cannot be much smaller than the
physical address because
MAX_PHYSMEM_BITS - SECTION_SIZE_BITS
has to fit into page flags and the room there is limited.
There is yet another possibility to support separate banks. It is possible
to use FLATMEM and free the memmap allocated for the hole, like, for
instance, ARM does [1]. This will require ARC's override for pfn_valid()
that takes into account the actual memory configuration rather than relies
on the memmap.
[1] https://elixir.bootlin.com/linux/latest/source/arch/arm/mm/init.c#L305
Mike Rapoport (1):
arc: add sparsemem support
arch/arc/Kconfig | 10 ++++++++++
arch/arc/include/asm/sparsemem.h | 13 +++++++++++++
arch/arc/mm/init.c | 6 +++++-
3 files changed, 28 insertions(+), 1 deletion(-)
create mode 100644 arch/arc/include/asm/sparsemem.h
--
2.26.2
