mm/xtensa: use general page fault accounting
authorPeter Xu <peterx@redhat.com>
Wed, 12 Aug 2020 01:38:53 +0000 (18:38 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 12 Aug 2020 17:58:04 +0000 (10:58 -0700)
Use the general page fault accounting by passing regs into
handle_mm_fault().  It naturally solve the issue of multiple page fault
accounting when page fault retry happened.

Remove the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf events because it's
now also done in handle_mm_fault().

Move the PERF_COUNT_SW_PAGE_FAULTS event higher before taking mmap_sem for
the fault, then it'll match with the rest of the archs.

Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Max Filippov <jcmvbkbc@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Link: http://lkml.kernel.org/r/20200707225021.200906-24-peterx@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
arch/xtensa/mm/fault.c

index e72c8c1..7666408 100644 (file)
@@ -72,6 +72,9 @@ void do_page_fault(struct pt_regs *regs)
 
        if (user_mode(regs))
                flags |= FAULT_FLAG_USER;
+
+       perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
+
 retry:
        mmap_read_lock(mm);
        vma = find_vma(mm, address);
@@ -107,7 +110,7 @@ good_area:
         * make sure we exit gracefully rather than endlessly redo
         * the fault.
         */
-       fault = handle_mm_fault(vma, address, flags, NULL);
+       fault = handle_mm_fault(vma, address, flags, regs);
 
        if (fault_signal_pending(fault, regs))
                return;
@@ -122,10 +125,6 @@ good_area:
                BUG();
        }
        if (flags & FAULT_FLAG_ALLOW_RETRY) {
-               if (fault & VM_FAULT_MAJOR)
-                       current->maj_flt++;
-               else
-                       current->min_flt++;
                if (fault & VM_FAULT_RETRY) {
                        flags |= FAULT_FLAG_TRIED;
 
@@ -139,12 +138,6 @@ good_area:
        }
 
        mmap_read_unlock(mm);
-       perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
-       if (flags & VM_FAULT_MAJOR)
-               perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
-       else
-               perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
-
        return;
 
        /* Something tried to access memory that isn't in our memory map..