]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
mm: numa: preserve PTE write permissions across a NUMA hinting fault
authorMel Gorman <mgorman@suse.de>
Wed, 25 Mar 2015 22:55:40 +0000 (15:55 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 25 Mar 2015 23:20:31 +0000 (16:20 -0700)
Protecting a PTE to trap a NUMA hinting fault clears the writable bit
and further faults are needed after trapping a NUMA hinting fault to set
the writable bit again.  This patch preserves the writable bit when
trapping NUMA hinting faults.  The impact is obvious from the number of
minor faults trapped during the basis balancing benchmark and the system
CPU usage;

  autonumabench
                                             4.0.0-rc4             4.0.0-rc4
                                              baseline              preserve
  Time System-NUMA01                  107.13 (  0.00%)      103.13 (  3.73%)
  Time System-NUMA01_THEADLOCAL       131.87 (  0.00%)       83.30 ( 36.83%)
  Time System-NUMA02                    8.95 (  0.00%)       10.72 (-19.78%)
  Time System-NUMA02_SMT                4.57 (  0.00%)        3.99 ( 12.69%)
  Time Elapsed-NUMA01                 515.78 (  0.00%)      517.26 ( -0.29%)
  Time Elapsed-NUMA01_THEADLOCAL      384.10 (  0.00%)      384.31 ( -0.05%)
  Time Elapsed-NUMA02                  48.86 (  0.00%)       48.78 (  0.16%)
  Time Elapsed-NUMA02_SMT              47.98 (  0.00%)       48.12 ( -0.29%)

               4.0.0-rc4   4.0.0-rc4
                baseline    preserve
  User          44383.95    43971.89
  System          252.61      201.24
  Elapsed         998.68     1000.94

  Minor Faults   2597249     1981230
  Major Faults       365         364

There is a similar drop in system CPU usage using Dave Chinner's xfsrepair
workload

                                      4.0.0-rc4             4.0.0-rc4
                                       baseline              preserve
  Amean    real-xfsrepair      454.14 (  0.00%)      442.36 (  2.60%)
  Amean    syst-xfsrepair      277.20 (  0.00%)      204.68 ( 26.16%)

The patch looks hacky but the alternatives looked worse.  The tidest was
to rewalk the page tables after a hinting fault but it was more complex
than this approach and the performance was worse.  It's not generally
safe to just mark the page writable during the fault if it's a write
fault as it may have been read-only for COW so that approach was
discarded.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Reported-by: Dave Chinner <david@fromorbit.com>
Tested-by: Dave Chinner <david@fromorbit.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/huge_memory.c
mm/memory.c
mm/mprotect.c

index 2f12e9fcf1a236665ae2cd3be9c79bd0b6f96e45..0a42d1521aa43a0d44e564a2bb0554ec43180bcb 100644 (file)
@@ -1260,6 +1260,7 @@ int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
        int target_nid, last_cpupid = -1;
        bool page_locked;
        bool migrated = false;
+       bool was_writable;
        int flags = 0;
 
        /* A PROT_NONE fault should not end up here */
@@ -1354,7 +1355,10 @@ int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
        goto out;
 clear_pmdnuma:
        BUG_ON(!PageLocked(page));
+       was_writable = pmd_write(pmd);
        pmd = pmd_modify(pmd, vma->vm_page_prot);
+       if (was_writable)
+               pmd = pmd_mkwrite(pmd);
        set_pmd_at(mm, haddr, pmdp, pmd);
        update_mmu_cache_pmd(vma, addr, pmdp);
        unlock_page(page);
@@ -1478,6 +1482,7 @@ int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 
        if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
                pmd_t entry;
+               bool preserve_write = prot_numa && pmd_write(*pmd);
                ret = 1;
 
                /*
@@ -1493,9 +1498,11 @@ int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
                if (!prot_numa || !pmd_protnone(*pmd)) {
                        entry = pmdp_get_and_clear_notify(mm, addr, pmd);
                        entry = pmd_modify(entry, newprot);
+                       if (preserve_write)
+                               entry = pmd_mkwrite(entry);
                        ret = HPAGE_PMD_NR;
                        set_pmd_at(mm, addr, pmd, entry);
-                       BUG_ON(pmd_write(entry));
+                       BUG_ON(!preserve_write && pmd_write(entry));
                }
                spin_unlock(ptl);
        }
index 20beb6647dba22714c52c1f6f228b30a9987a5a7..d20e12da3a3cb29bf7fc6bffa2de9fbe89f22e4f 100644 (file)
@@ -3035,6 +3035,7 @@ static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
        int last_cpupid;
        int target_nid;
        bool migrated = false;
+       bool was_writable = pte_write(pte);
        int flags = 0;
 
        /* A PROT_NONE fault should not end up here */
@@ -3059,6 +3060,8 @@ static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
        /* Make it present again */
        pte = pte_modify(pte, vma->vm_page_prot);
        pte = pte_mkyoung(pte);
+       if (was_writable)
+               pte = pte_mkwrite(pte);
        set_pte_at(mm, addr, ptep, pte);
        update_mmu_cache(vma, addr, ptep);
 
@@ -3075,11 +3078,6 @@ static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
         * to it but pte_write gets cleared during protection updates and
         * pte_dirty has unpredictable behaviour between PTE scan updates,
         * background writeback, dirty balancing and application behaviour.
-        *
-        * TODO: Note that the ideal here would be to avoid a situation where a
-        * NUMA fault is taken immediately followed by a write fault in
-        * some cases which would have lower overhead overall but would be
-        * invasive as the fault paths would need to be unified.
         */
        if (!(vma->vm_flags & VM_WRITE))
                flags |= TNF_NO_GROUP;
index 44727811bf4cf62e3579261ee9699a37fab78b3d..88584838e7046bec724d68c0cafcd94eec65a040 100644 (file)
@@ -75,6 +75,7 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
                oldpte = *pte;
                if (pte_present(oldpte)) {
                        pte_t ptent;
+                       bool preserve_write = prot_numa && pte_write(oldpte);
 
                        /*
                         * Avoid trapping faults against the zero or KSM
@@ -94,6 +95,8 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 
                        ptent = ptep_modify_prot_start(mm, addr, pte);
                        ptent = pte_modify(ptent, newprot);
+                       if (preserve_write)
+                               ptent = pte_mkwrite(ptent);
 
                        /* Avoid taking write faults for known dirty pages */
                        if (dirty_accountable && pte_dirty(ptent) &&