]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/commitdiff
powerpc/mm: Lockless get_user_pages_fast() for 64-bit (v3)
authorNick Piggin <npiggin@suse.de>
Wed, 30 Jul 2008 05:23:13 +0000 (15:23 +1000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Wed, 30 Jul 2008 05:26:54 +0000 (15:26 +1000)
Implement lockless get_user_pages_fast for 64-bit powerpc.

Page table existence is guaranteed with RCU, and speculative page references
are used to take a reference to the pages without having a prior existence
guarantee on them.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Dave Kleikamp <shaggy@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/Kconfig
arch/powerpc/mm/Makefile
arch/powerpc/mm/gup.c [new file with mode: 0644]
include/asm-powerpc/pgtable-ppc64.h
include/linux/pagemap.h

index 587da5e0990f9504979f9357b2783305e6a23d02..63c9cafda9c4e34bce6d096db4bcf83b68b7cd51 100644 (file)
@@ -42,6 +42,9 @@ config GENERIC_HARDIRQS
        bool
        default y
 
+config HAVE_GET_USER_PAGES_FAST
+       def_bool PPC64
+
 config HAVE_SETUP_PER_CPU_AREA
        def_bool PPC64
 
index 1c00e0196f6c25bc99452ecdd6f04e47ee412b66..e7392b45a5ef6ae82ad68727188e257990d407ea 100644 (file)
@@ -12,7 +12,8 @@ obj-y                         := fault.o mem.o \
                                   mmu_context_$(CONFIG_WORD_SIZE).o
 hash-$(CONFIG_PPC_NATIVE)      := hash_native_64.o
 obj-$(CONFIG_PPC64)            += hash_utils_64.o \
-                                  slb_low.o slb.o stab.o mmap.o $(hash-y)
+                                  slb_low.o slb.o stab.o \
+                                  gup.o mmap.o $(hash-y)
 obj-$(CONFIG_PPC_STD_MMU_32)   += ppc_mmu_32.o
 obj-$(CONFIG_PPC_STD_MMU)      += hash_low_$(CONFIG_WORD_SIZE).o \
                                   tlb_$(CONFIG_WORD_SIZE).o
diff --git a/arch/powerpc/mm/gup.c b/arch/powerpc/mm/gup.c
new file mode 100644 (file)
index 0000000..9fdf4d6
--- /dev/null
@@ -0,0 +1,280 @@
+/*
+ * Lockless get_user_pages_fast for powerpc
+ *
+ * Copyright (C) 2008 Nick Piggin
+ * Copyright (C) 2008 Novell Inc.
+ */
+#undef DEBUG
+
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/hugetlb.h>
+#include <linux/vmstat.h>
+#include <linux/pagemap.h>
+#include <linux/rwsem.h>
+#include <asm/pgtable.h>
+
+/*
+ * The performance critical leaf functions are made noinline otherwise gcc
+ * inlines everything into a single function which results in too much
+ * register pressure.
+ */
+static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
+               unsigned long end, int write, struct page **pages, int *nr)
+{
+       unsigned long mask, result;
+       pte_t *ptep;
+
+       result = _PAGE_PRESENT|_PAGE_USER;
+       if (write)
+               result |= _PAGE_RW;
+       mask = result | _PAGE_SPECIAL;
+
+       ptep = pte_offset_kernel(&pmd, addr);
+       do {
+               pte_t pte = *ptep;
+               struct page *page;
+
+               if ((pte_val(pte) & mask) != result)
+                       return 0;
+               VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
+               page = pte_page(pte);
+               if (!page_cache_get_speculative(page))
+                       return 0;
+               if (unlikely(pte != *ptep)) {
+                       put_page(page);
+                       return 0;
+               }
+               pages[*nr] = page;
+               (*nr)++;
+
+       } while (ptep++, addr += PAGE_SIZE, addr != end);
+
+       return 1;
+}
+
+#ifdef CONFIG_HUGETLB_PAGE
+static noinline int gup_huge_pte(pte_t *ptep, struct hstate *hstate,
+                                unsigned long *addr, unsigned long end,
+                                int write, struct page **pages, int *nr)
+{
+       unsigned long mask;
+       unsigned long pte_end;
+       struct page *head, *page;
+       pte_t pte;
+       int refs;
+
+       pte_end = (*addr + huge_page_size(hstate)) & huge_page_mask(hstate);
+       if (pte_end < end)
+               end = pte_end;
+
+       pte = *ptep;
+       mask = _PAGE_PRESENT|_PAGE_USER;
+       if (write)
+               mask |= _PAGE_RW;
+       if ((pte_val(pte) & mask) != mask)
+               return 0;
+       /* hugepages are never "special" */
+       VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
+
+       refs = 0;
+       head = pte_page(pte);
+       page = head + ((*addr & ~huge_page_mask(hstate)) >> PAGE_SHIFT);
+       do {
+               VM_BUG_ON(compound_head(page) != head);
+               pages[*nr] = page;
+               (*nr)++;
+               page++;
+               refs++;
+       } while (*addr += PAGE_SIZE, *addr != end);
+
+       if (!page_cache_add_speculative(head, refs)) {
+               *nr -= refs;
+               return 0;
+       }
+       if (unlikely(pte != *ptep)) {
+               /* Could be optimized better */
+               while (*nr) {
+                       put_page(page);
+                       (*nr)--;
+               }
+       }
+
+       return 1;
+}
+#endif /* CONFIG_HUGETLB_PAGE */
+
+static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
+               int write, struct page **pages, int *nr)
+{
+       unsigned long next;
+       pmd_t *pmdp;
+
+       pmdp = pmd_offset(&pud, addr);
+       do {
+               pmd_t pmd = *pmdp;
+
+               next = pmd_addr_end(addr, end);
+               if (pmd_none(pmd))
+                       return 0;
+               if (!gup_pte_range(pmd, addr, next, write, pages, nr))
+                       return 0;
+       } while (pmdp++, addr = next, addr != end);
+
+       return 1;
+}
+
+static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
+               int write, struct page **pages, int *nr)
+{
+       unsigned long next;
+       pud_t *pudp;
+
+       pudp = pud_offset(&pgd, addr);
+       do {
+               pud_t pud = *pudp;
+
+               next = pud_addr_end(addr, end);
+               if (pud_none(pud))
+                       return 0;
+               if (!gup_pmd_range(pud, addr, next, write, pages, nr))
+                       return 0;
+       } while (pudp++, addr = next, addr != end);
+
+       return 1;
+}
+
+int get_user_pages_fast(unsigned long start, int nr_pages, int write,
+                       struct page **pages)
+{
+       struct mm_struct *mm = current->mm;
+       unsigned long addr, len, end;
+       unsigned long next;
+       pgd_t *pgdp;
+       int psize, nr = 0;
+       unsigned int shift;
+
+       pr_debug("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
+
+       start &= PAGE_MASK;
+       addr = start;
+       len = (unsigned long) nr_pages << PAGE_SHIFT;
+       end = start + len;
+
+       if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
+                                       start, len)))
+               goto slow_irqon;
+
+       pr_debug("  aligned: %lx .. %lx\n", start, end);
+
+#ifdef CONFIG_HUGETLB_PAGE
+       /* We bail out on slice boundary crossing when hugetlb is
+        * enabled in order to not have to deal with two different
+        * page table formats
+        */
+       if (addr < SLICE_LOW_TOP) {
+               if (end > SLICE_LOW_TOP)
+                       goto slow_irqon;
+
+               if (unlikely(GET_LOW_SLICE_INDEX(addr) !=
+                            GET_LOW_SLICE_INDEX(end - 1)))
+                       goto slow_irqon;
+       } else {
+               if (unlikely(GET_HIGH_SLICE_INDEX(addr) !=
+                            GET_HIGH_SLICE_INDEX(end - 1)))
+                       goto slow_irqon;
+       }
+#endif /* CONFIG_HUGETLB_PAGE */
+
+       /*
+        * XXX: batch / limit 'nr', to avoid large irq off latency
+        * needs some instrumenting to determine the common sizes used by
+        * important workloads (eg. DB2), and whether limiting the batch size
+        * will decrease performance.
+        *
+        * It seems like we're in the clear for the moment. Direct-IO is
+        * the main guy that batches up lots of get_user_pages, and even
+        * they are limited to 64-at-a-time which is not so many.
+        */
+       /*
+        * This doesn't prevent pagetable teardown, but does prevent
+        * the pagetables from being freed on powerpc.
+        *
+        * So long as we atomically load page table pointers versus teardown,
+        * we can follow the address down to the the page and take a ref on it.
+        */
+       local_irq_disable();
+
+       psize = get_slice_psize(mm, addr);
+       shift = mmu_psize_defs[psize].shift;
+
+#ifdef CONFIG_HUGETLB_PAGE
+       if (unlikely(mmu_huge_psizes[psize])) {
+               pte_t *ptep;
+               unsigned long a = addr;
+               unsigned long sz = ((1UL) << shift);
+               struct hstate *hstate = size_to_hstate(sz);
+
+               BUG_ON(!hstate);
+               /*
+                * XXX: could be optimized to avoid hstate
+                * lookup entirely (just use shift)
+                */
+
+               do {
+                       VM_BUG_ON(shift != mmu_psize_defs[get_slice_psize(mm, a)].shift);
+                       ptep = huge_pte_offset(mm, a);
+                       pr_debug(" %016lx: huge ptep %p\n", a, ptep);
+                       if (!ptep || !gup_huge_pte(ptep, hstate, &a, end, write, pages,
+                                                  &nr))
+                               goto slow;
+               } while (a != end);
+       } else
+#endif /* CONFIG_HUGETLB_PAGE */
+       {
+               pgdp = pgd_offset(mm, addr);
+               do {
+                       pgd_t pgd = *pgdp;
+
+                       VM_BUG_ON(shift != mmu_psize_defs[get_slice_psize(mm, addr)].shift);
+                       pr_debug("  %016lx: normal pgd %p\n", addr, (void *)pgd);
+                       next = pgd_addr_end(addr, end);
+                       if (pgd_none(pgd))
+                               goto slow;
+                       if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
+                               goto slow;
+               } while (pgdp++, addr = next, addr != end);
+       }
+       local_irq_enable();
+
+       VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
+       return nr;
+
+       {
+               int ret;
+
+slow:
+               local_irq_enable();
+slow_irqon:
+               pr_debug("  slow path ! nr = %d\n", nr);
+
+               /* Try to get the remaining pages with get_user_pages */
+               start += nr << PAGE_SHIFT;
+               pages += nr;
+
+               down_read(&mm->mmap_sem);
+               ret = get_user_pages(current, mm, start,
+                       (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
+               up_read(&mm->mmap_sem);
+
+               /* Have to be a bit careful with return values */
+               if (nr > 0) {
+                       if (ret < 0)
+                               ret = nr;
+                       else
+                               ret += nr;
+               }
+
+               return ret;
+       }
+}
index 5fc78c0be302fa60e972fbad50f24bb69114351b..74c6f380b8055e5f7803201e51755866d01ed904 100644 (file)
@@ -461,6 +461,8 @@ void pgtable_cache_init(void);
        return pt;
 }
 
+pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long address);
+
 #endif /* __ASSEMBLY__ */
 
 #endif /* _ASM_POWERPC_PGTABLE_PPC64_H_ */
index a39b38ccdc976e920b077fcd484cabb8cc402ba8..69ed3cb1197ac942b8d982e395fb48d5ecdbb29b 100644 (file)
@@ -143,6 +143,29 @@ static inline int page_cache_get_speculative(struct page *page)
        return 1;
 }
 
+/*
+ * Same as above, but add instead of inc (could just be merged)
+ */
+static inline int page_cache_add_speculative(struct page *page, int count)
+{
+       VM_BUG_ON(in_interrupt());
+
+#if !defined(CONFIG_SMP) && defined(CONFIG_CLASSIC_RCU)
+# ifdef CONFIG_PREEMPT
+       VM_BUG_ON(!in_atomic());
+# endif
+       VM_BUG_ON(page_count(page) == 0);
+       atomic_add(count, &page->_count);
+
+#else
+       if (unlikely(!atomic_add_unless(&page->_count, count, 0)))
+               return 0;
+#endif
+       VM_BUG_ON(PageCompound(page) && page != compound_head(page));
+
+       return 1;
+}
+
 static inline int page_freeze_refs(struct page *page, int count)
 {
        return likely(atomic_cmpxchg(&page->_count, count, 0) == count);