mm, hugetlbfs: pass fault address to no page handler
authorHuang Ying <ying.huang@intel.com>
Fri, 8 Jun 2018 00:08:08 +0000 (17:08 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 8 Jun 2018 00:34:36 +0000 (17:34 -0700)
This is to take better advantage of general huge page clearing
optimization (commit c79b57e462b5: "mm: hugetlb: clear target sub-page
last when clearing huge page") for hugetlbfs.

In the general optimization patch, the sub-page to access will be
cleared last to avoid the cache lines of to access sub-page to be
evicted when clearing other sub-pages.  This works better if we have the
address of the sub-page to access, that is, the fault address inside the
huge page.  So the hugetlbfs no page fault handler is changed to pass
that information.  This will benefit workloads which don't access the
begin of the hugetlbfs huge page after the page fault under heavy cache
contention for shared last level cache.

The patch is a generic optimization which should benefit quite some
workloads, not for a specific use case.  To demonstrate the performance
benefit of the patch, we tested it with vm-scalability run on hugetlbfs.

With this patch, the throughput increases ~28.1% in vm-scalability
anon-w-seq test case with 88 processes on a 2 socket Xeon E5 2699 v4
system (44 cores, 88 threads).  The test case creates 88 processes, each
process mmaps a big anonymous memory area with MAP_HUGETLB and writes to
it from the end to the begin.  For each process, other processes could
be seen as other workload which generates heavy cache pressure.  At the
same time, the cache miss rate reduced from ~36.3% to ~25.6%, the IPC
(instruction per cycle) increased from 0.3 to 0.37, and the time spent
in user space is reduced ~19.3%.

Link: http://lkml.kernel.org/r/20180517083539.9242-1-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andi Kleen <andi.kleen@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Matthew Wilcox <mawilcox@microsoft.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shaohua Li <shli@fb.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Punit Agrawal <punit.agrawal@arm.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/hugetlb.c

index 6c6decc634696c18863651ec88f2f9672c4f3bb2..696befffe6f7237b0d58f36839092e9f2a96aeec 100644 (file)
@@ -3686,6 +3686,7 @@ static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
        struct page *page;
        pte_t new_pte;
        spinlock_t *ptl;
+       unsigned long haddr = address & huge_page_mask(h);
 
        /*
         * Currently, we are forced to kill the process in the event the
@@ -3716,7 +3717,7 @@ retry:
                        u32 hash;
                        struct vm_fault vmf = {
                                .vma = vma,
-                               .address = address,
+                               .address = haddr,
                                .flags = flags,
                                /*
                                 * Hard to debug if it ends up being
@@ -3733,14 +3734,14 @@ retry:
                         * fault to make calling code simpler.
                         */
                        hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
-                                                       idx, address);
+                                                       idx, haddr);
                        mutex_unlock(&hugetlb_fault_mutex_table[hash]);
                        ret = handle_userfault(&vmf, VM_UFFD_MISSING);
                        mutex_lock(&hugetlb_fault_mutex_table[hash]);
                        goto out;
                }
 
-               page = alloc_huge_page(vma, address, 0);
+               page = alloc_huge_page(vma, haddr, 0);
                if (IS_ERR(page)) {
                        ret = PTR_ERR(page);
                        if (ret == -ENOMEM)
@@ -3789,12 +3790,12 @@ retry:
         * the spinlock.
         */
        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
-               if (vma_needs_reservation(h, vma, address) < 0) {
+               if (vma_needs_reservation(h, vma, haddr) < 0) {
                        ret = VM_FAULT_OOM;
                        goto backout_unlocked;
                }
                /* Just decrements count, does not deallocate */
-               vma_end_reservation(h, vma, address);
+               vma_end_reservation(h, vma, haddr);
        }
 
        ptl = huge_pte_lock(h, mm, ptep);
@@ -3808,17 +3809,17 @@ retry:
 
        if (anon_rmap) {
                ClearPagePrivate(page);
-               hugepage_add_new_anon_rmap(page, vma, address);
+               hugepage_add_new_anon_rmap(page, vma, haddr);
        } else
                page_dup_rmap(page, true);
        new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
                                && (vma->vm_flags & VM_SHARED)));
-       set_huge_pte_at(mm, address, ptep, new_pte);
+       set_huge_pte_at(mm, haddr, ptep, new_pte);
 
        hugetlb_count_add(pages_per_huge_page(h), mm);
        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
                /* Optimization, do the COW without a second fault */
-               ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
+               ret = hugetlb_cow(mm, vma, haddr, ptep, page, ptl);
        }
 
        spin_unlock(ptl);
@@ -3830,7 +3831,7 @@ backout:
        spin_unlock(ptl);
 backout_unlocked:
        unlock_page(page);
-       restore_reserve_on_error(h, vma, address, page);
+       restore_reserve_on_error(h, vma, haddr, page);
        put_page(page);
        goto out;
 }
@@ -3883,10 +3884,9 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        struct hstate *h = hstate_vma(vma);
        struct address_space *mapping;
        int need_wait_lock = 0;
+       unsigned long haddr = address & huge_page_mask(h);
 
-       address &= huge_page_mask(h);
-
-       ptep = huge_pte_offset(mm, address, huge_page_size(h));
+       ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
        if (ptep) {
                entry = huge_ptep_get(ptep);
                if (unlikely(is_hugetlb_entry_migration(entry))) {
@@ -3896,20 +3896,20 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
                        return VM_FAULT_HWPOISON_LARGE |
                                VM_FAULT_SET_HINDEX(hstate_index(h));
        } else {
-               ptep = huge_pte_alloc(mm, address, huge_page_size(h));
+               ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
                if (!ptep)
                        return VM_FAULT_OOM;
        }
 
        mapping = vma->vm_file->f_mapping;
-       idx = vma_hugecache_offset(h, vma, address);
+       idx = vma_hugecache_offset(h, vma, haddr);
 
        /*
         * Serialize hugepage allocation and instantiation, so that we don't
         * get spurious allocation failures if two CPUs race to instantiate
         * the same page in the page cache.
         */
-       hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
+       hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr);
        mutex_lock(&hugetlb_fault_mutex_table[hash]);
 
        entry = huge_ptep_get(ptep);
@@ -3939,16 +3939,16 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
         * consumed.
         */
        if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
-               if (vma_needs_reservation(h, vma, address) < 0) {
+               if (vma_needs_reservation(h, vma, haddr) < 0) {
                        ret = VM_FAULT_OOM;
                        goto out_mutex;
                }
                /* Just decrements count, does not deallocate */
-               vma_end_reservation(h, vma, address);
+               vma_end_reservation(h, vma, haddr);
 
                if (!(vma->vm_flags & VM_MAYSHARE))
                        pagecache_page = hugetlbfs_pagecache_page(h,
-                                                               vma, address);
+                                                               vma, haddr);
        }
 
        ptl = huge_pte_lock(h, mm, ptep);
@@ -3973,16 +3973,16 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 
        if (flags & FAULT_FLAG_WRITE) {
                if (!huge_pte_write(entry)) {
-                       ret = hugetlb_cow(mm, vma, address, ptep,
+                       ret = hugetlb_cow(mm, vma, haddr, ptep,
                                          pagecache_page, ptl);
                        goto out_put_page;
                }
                entry = huge_pte_mkdirty(entry);
        }
        entry = pte_mkyoung(entry);
-       if (huge_ptep_set_access_flags(vma, address, ptep, entry,
+       if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
                                                flags & FAULT_FLAG_WRITE))
-               update_mmu_cache(vma, address, ptep);
+               update_mmu_cache(vma, haddr, ptep);
 out_put_page:
        if (page != pagecache_page)
                unlock_page(page);