Only call cpu_icache_sync_range when inserting an executable page. If the

page is non-executable the contents of the i-cache are unimportant so this
call is just adding unneeded overhead when inserting pages.

While doing research using gem5 with an O3 pipeline and 1k/32k/1M iTLB/L1
iCache/L2 Bjoern Zeeb (bz@) observed a fairly high rate of calls into
arm64_icache_sync_range() from pmap_enter() along with a high number of
instruction fetches and iTLB/iCache hits.

Limiting the calls to arm64_icache_sync_range() to only executable pages,
we observe the iTLB and iCache Hit going down by about 43%. These numbers
are quite misleading when looked at alone as at the same time instructions
retired were reduced by 19.2% and instruction fetches were reduced by 38.8%.
Overall this reduced the runtime of the test program by 22.4%.

On Juno hardware, in steady-state, running the same test, using the cycle
count to determine runtime, we do see a reduction of up to 28.9% in runtime.

While these numbers certainly depend on the program executed, we expect an
overall performance improvement.

Reported by:	bz
Obtained from:	ABT Systems Ltd
MFC after:	1 week
Sponsored by:	The FreeBSD Foundation
This commit is contained in:
Andrew Turner 2016-09-07 16:22:05 +00:00
parent d204cea9f8
commit 3b34364450
Notes: svn2git 2020-12-20 02:59:44 +00:00
svn path=/head/; revision=305545

View File

@ -2939,8 +2939,9 @@ pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
pmap_invalidate_page(pmap, va);
if (pmap != pmap_kernel()) {
if (pmap == &curproc->p_vmspace->vm_pmap)
cpu_icache_sync_range(va, PAGE_SIZE);
if (pmap == &curproc->p_vmspace->vm_pmap &&
(prot & VM_PROT_EXECUTE) != 0)
cpu_icache_sync_range(va, PAGE_SIZE);
if ((mpte == NULL || mpte->wire_count == NL3PG) &&
pmap_superpages_enabled() &&