powerpc/mm: don't do tlbie for updatepp request with NO HPTE fault
authorAneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Thu, 4 Dec 2014 05:30:14 +0000 (11:00 +0530)
committerMichael Ellerman <mpe@ellerman.id.au>
Fri, 5 Dec 2014 05:26:15 +0000 (16:26 +1100)
commitaefa5688c070727b8729de1aef85cad7b9933fc7
treed6036007b0eca0b325aa3f050defb00f89341209
parentabb90ee7bca5af977b90a0c6be44f631fdacc932
powerpc/mm: don't do tlbie for updatepp request with NO HPTE fault

upatepp can get called for a nohpte fault when we find from the linux
page table that the translation was hashed before. In that case
we are sure that there is no existing translation, hence we could
avoid doing tlbie.

We could possibly race with a parallel fault filling the TLB. But
that should be ok because updatepp is only ever relaxing permissions.
We also look at linux pte permission bits when filling hash pte
permission bits. We also hold the linux pte busy bits while
inserting/updating a hashpte entry, hence a paralle update of
linux pte is not possible. On the other hand mprotect involves
ptep_modify_prot_start which cause a hpte invalidate and not updatepp.

Performance number:
We use randbox_access_bench written by Anton.

Kernel with THP disabled and smaller hash page table size.

    86.60%  random_access_b  [kernel.kallsyms]                [k] .native_hpte_updatepp
     2.10%  random_access_b  random_access_bench              [.] doit
     1.99%  random_access_b  [kernel.kallsyms]                [k] .do_raw_spin_lock
     1.85%  random_access_b  [kernel.kallsyms]                [k] .native_hpte_insert
     1.26%  random_access_b  [kernel.kallsyms]                [k] .native_flush_hash_range
     1.18%  random_access_b  [kernel.kallsyms]                [k] .__delay
     0.69%  random_access_b  [kernel.kallsyms]                [k] .native_hpte_remove
     0.37%  random_access_b  [kernel.kallsyms]                [k] .clear_user_page
     0.34%  random_access_b  [kernel.kallsyms]                [k] .__hash_page_64K
     0.32%  random_access_b  [kernel.kallsyms]                [k] fast_exception_return
     0.30%  random_access_b  [kernel.kallsyms]                [k] .hash_page_mm

With Fix:

    27.54%  random_access_b  random_access_bench              [.] doit
    22.90%  random_access_b  [kernel.kallsyms]                [k] .native_hpte_insert
     5.76%  random_access_b  [kernel.kallsyms]                [k] .native_hpte_remove
     5.20%  random_access_b  [kernel.kallsyms]                [k] fast_exception_return
     5.12%  random_access_b  [kernel.kallsyms]                [k] .__hash_page_64K
     4.80%  random_access_b  [kernel.kallsyms]                [k] .hash_page_mm
     3.31%  random_access_b  [kernel.kallsyms]                [k] data_access_common
     1.84%  random_access_b  [kernel.kallsyms]                [k] .trace_hardirqs_on_caller

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
16 files changed:
arch/powerpc/include/asm/machdep.h
arch/powerpc/include/asm/mmu-hash64.h
arch/powerpc/include/asm/tlbflush.h
arch/powerpc/kernel/exceptions-64s.S
arch/powerpc/mm/hash_low_64.S
arch/powerpc/mm/hash_native_64.c
arch/powerpc/mm/hash_utils_64.c
arch/powerpc/mm/hugepage-hash64.c
arch/powerpc/mm/hugetlbpage-hash64.c
arch/powerpc/mm/pgtable_64.c
arch/powerpc/platforms/cell/beat_htab.c
arch/powerpc/platforms/cell/spu_base.c
arch/powerpc/platforms/cell/spufs/fault.c
arch/powerpc/platforms/ps3/htab.c
arch/powerpc/platforms/pseries/lpar.c
drivers/misc/cxl/fault.c