]> git.proxmox.com Git - mirror_ubuntu-eoan-kernel.git/commitdiff
Revert "mm, thp: consolidate THP gfp handling into alloc_hugepage_direct_gfpmask"
authorDavid Rientjes <rientjes@google.com>
Fri, 7 Dec 2018 22:50:16 +0000 (14:50 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 8 Dec 2018 18:26:20 +0000 (10:26 -0800)
This reverts commit 89c83fb539f95491be80cdd5158e6f0ce329e317.

This should have been done as part of 2f0799a0ffc0 ("mm, thp: restore
node-local hugepage allocations").  The movement of the thp allocation
policy from alloc_pages_vma() to alloc_hugepage_direct_gfpmask() was
intended to only set __GFP_THISNODE for mempolicies that are not
MPOL_BIND whereas the revert could set this regardless of mempolicy.

While the check for MPOL_BIND between alloc_hugepage_direct_gfpmask()
and alloc_pages_vma() was racy, that has since been removed since the
revert.  What is left is the possibility to use __GFP_THISNODE in
policy_node() when it is unexpected because the special handling for
hugepages in alloc_pages_vma()  was removed as part of the consolidation.

Secondly, prior to 89c83fb539f9, alloc_pages_vma() implemented a somewhat
different policy for hugepage allocations, which were allocated through
alloc_hugepage_vma().  For hugepage allocations, if the allocating
process's node is in the set of allowed nodes, allocate with
__GFP_THISNODE for that node (for MPOL_PREFERRED, use that node with
__GFP_THISNODE instead).  This was changed for shmem_alloc_hugepage() to
allow fallback to other nodes in 89c83fb539f9 as it did for new_page() in
mm/mempolicy.c which is functionally different behavior and removes the
requirement to only allocate hugepages locally.

So this commit does a full revert of 89c83fb539f9 instead of the partial
revert that was done in 2f0799a0ffc0.  The result is the same thp
allocation policy for 4.20 that was in 4.19.

Fixes: 89c83fb539f9 ("mm, thp: consolidate THP gfp handling into alloc_hugepage_direct_gfpmask")
Fixes: 2f0799a0ffc0 ("mm, thp: restore node-local hugepage allocations")
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/gfp.h
mm/huge_memory.c
mm/mempolicy.c
mm/shmem.c

index 76f8db0b0e715c016cc00cb95aa9a269f12c075d..0705164f928c949720d9a0927f1564119146c3fa 100644 (file)
@@ -510,18 +510,22 @@ alloc_pages(gfp_t gfp_mask, unsigned int order)
 }
 extern struct page *alloc_pages_vma(gfp_t gfp_mask, int order,
                        struct vm_area_struct *vma, unsigned long addr,
-                       int node);
+                       int node, bool hugepage);
+#define alloc_hugepage_vma(gfp_mask, vma, addr, order) \
+       alloc_pages_vma(gfp_mask, order, vma, addr, numa_node_id(), true)
 #else
 #define alloc_pages(gfp_mask, order) \
                alloc_pages_node(numa_node_id(), gfp_mask, order)
-#define alloc_pages_vma(gfp_mask, order, vma, addr, node)\
+#define alloc_pages_vma(gfp_mask, order, vma, addr, node, false)\
+       alloc_pages(gfp_mask, order)
+#define alloc_hugepage_vma(gfp_mask, vma, addr, order) \
        alloc_pages(gfp_mask, order)
 #endif
 #define alloc_page(gfp_mask) alloc_pages(gfp_mask, 0)
 #define alloc_page_vma(gfp_mask, vma, addr)                    \
-       alloc_pages_vma(gfp_mask, 0, vma, addr, numa_node_id())
+       alloc_pages_vma(gfp_mask, 0, vma, addr, numa_node_id(), false)
 #define alloc_page_vma_node(gfp_mask, vma, addr, node)         \
-       alloc_pages_vma(gfp_mask, 0, vma, addr, node)
+       alloc_pages_vma(gfp_mask, 0, vma, addr, node, false)
 
 extern unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order);
 extern unsigned long get_zeroed_page(gfp_t gfp_mask);
index f2d19e4fe854f3941bbc24dd1e1cc76d0622b7eb..5da55b38b1b7fd2878a20b41c180b839096fd872 100644 (file)
@@ -629,30 +629,30 @@ release:
  *         available
  * never: never stall for any thp allocation
  */
-static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma, unsigned long addr)
+static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 {
        const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
-       const gfp_t gfp_mask = GFP_TRANSHUGE_LIGHT | __GFP_THISNODE;
 
        /* Always do synchronous compaction */
        if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
-               return GFP_TRANSHUGE | __GFP_THISNODE |
-                      (vma_madvised ? 0 : __GFP_NORETRY);
+               return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 
        /* Kick kcompactd and fail quickly */
        if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
-               return gfp_mask | __GFP_KSWAPD_RECLAIM;
+               return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 
        /* Synchronous compaction if madvised, otherwise kick kcompactd */
        if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
-               return gfp_mask | (vma_madvised ? __GFP_DIRECT_RECLAIM :
-                                                 __GFP_KSWAPD_RECLAIM);
+               return GFP_TRANSHUGE_LIGHT |
+                       (vma_madvised ? __GFP_DIRECT_RECLAIM :
+                                       __GFP_KSWAPD_RECLAIM);
 
        /* Only do synchronous compaction if madvised */
        if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
-               return gfp_mask | (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
+               return GFP_TRANSHUGE_LIGHT |
+                      (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 
-       return gfp_mask;
+       return GFP_TRANSHUGE_LIGHT;
 }
 
 /* Caller must hold page table lock. */
@@ -724,8 +724,8 @@ vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
                        pte_free(vma->vm_mm, pgtable);
                return ret;
        }
-       gfp = alloc_hugepage_direct_gfpmask(vma, haddr);
-       page = alloc_pages_vma(gfp, HPAGE_PMD_ORDER, vma, haddr, numa_node_id());
+       gfp = alloc_hugepage_direct_gfpmask(vma);
+       page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
        if (unlikely(!page)) {
                count_vm_event(THP_FAULT_FALLBACK);
                return VM_FAULT_FALLBACK;
@@ -1295,9 +1295,8 @@ vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
 alloc:
        if (transparent_hugepage_enabled(vma) &&
            !transparent_hugepage_debug_cow()) {
-               huge_gfp = alloc_hugepage_direct_gfpmask(vma, haddr);
-               new_page = alloc_pages_vma(huge_gfp, HPAGE_PMD_ORDER, vma,
-                               haddr, numa_node_id());
+               huge_gfp = alloc_hugepage_direct_gfpmask(vma);
+               new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
        } else
                new_page = NULL;
 
index 69e278b469efd6850113bf96a69954c0c651cfb3..d4496d9d34f533dcd66accb7d92a69a03feae65c 100644 (file)
@@ -1116,8 +1116,8 @@ static struct page *new_page(struct page *page, unsigned long start)
        } else if (PageTransHuge(page)) {
                struct page *thp;
 
-               thp = alloc_pages_vma(GFP_TRANSHUGE, HPAGE_PMD_ORDER, vma,
-                               address, numa_node_id());
+               thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
+                                        HPAGE_PMD_ORDER);
                if (!thp)
                        return NULL;
                prep_transhuge_page(thp);
@@ -2011,6 +2011,7 @@ static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
  *     @vma:  Pointer to VMA or NULL if not available.
  *     @addr: Virtual Address of the allocation. Must be inside the VMA.
  *     @node: Which node to prefer for allocation (modulo policy).
+ *     @hugepage: for hugepages try only the preferred node if possible
  *
  *     This function allocates a page from the kernel page pool and applies
  *     a NUMA policy associated with the VMA or the current process.
@@ -2021,7 +2022,7 @@ static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
  */
 struct page *
 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
-               unsigned long addr, int node)
+               unsigned long addr, int node, bool hugepage)
 {
        struct mempolicy *pol;
        struct page *page;
@@ -2039,6 +2040,31 @@ alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
                goto out;
        }
 
+       if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
+               int hpage_node = node;
+
+               /*
+                * For hugepage allocation and non-interleave policy which
+                * allows the current node (or other explicitly preferred
+                * node) we only try to allocate from the current/preferred
+                * node and don't fall back to other nodes, as the cost of
+                * remote accesses would likely offset THP benefits.
+                *
+                * If the policy is interleave, or does not allow the current
+                * node in its nodemask, we allocate the standard way.
+                */
+               if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
+                       hpage_node = pol->v.preferred_node;
+
+               nmask = policy_nodemask(gfp, pol);
+               if (!nmask || node_isset(hpage_node, *nmask)) {
+                       mpol_cond_put(pol);
+                       page = __alloc_pages_node(hpage_node,
+                                               gfp | __GFP_THISNODE, order);
+                       goto out;
+               }
+       }
+
        nmask = policy_nodemask(gfp, pol);
        preferred_nid = policy_node(gfp, pol, node);
        page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
index cddc72ac44d868492e95a4badde4ea748d4aafeb..921f80488bb3fdd03cc7ce64a5a5a6d0f015b794 100644 (file)
@@ -1439,7 +1439,7 @@ static struct page *shmem_alloc_hugepage(gfp_t gfp,
 
        shmem_pseudo_vma_init(&pvma, info, hindex);
        page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
-                       HPAGE_PMD_ORDER, &pvma, 0, numa_node_id());
+                       HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
        shmem_pseudo_vma_destroy(&pvma);
        if (page)
                prep_transhuge_page(page);