Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S262925AbVD2UAu (ORCPT ); Fri, 29 Apr 2005 16:00:50 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S262921AbVD2UAt (ORCPT ); Fri, 29 Apr 2005 16:00:49 -0400 Received: from omx2-ext.sgi.com ([192.48.171.19]:18149 "EHLO omx2.sgi.com") by vger.kernel.org with ESMTP id S262922AbVD2T7J (ORCPT ); Fri, 29 Apr 2005 15:59:09 -0400 Date: Fri, 29 Apr 2005 12:59:06 -0700 (PDT) From: Christoph Lameter To: linux-kernel@vger.kernel.org Cc: linux-mm@kvack.org, linux-ia64@vger.kernel.org, Christoph Lameter Message-Id: <20050429195907.15694.75458.sendpatchset@schroedinger.engr.sgi.com> In-Reply-To: <20050429195901.15694.28520.sendpatchset@schroedinger.engr.sgi.com> References: <20050429195901.15694.28520.sendpatchset@schroedinger.engr.sgi.com> Subject: [PATCH 1/3] Page Fault Scalability V20: Avoid spurious page faults Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 14306 Lines: 354 Updating a page table entry (pte) can be difficult since the MMU may modify the pte concurrently. The current approach taken is to first exchange the pte contents with zero. Clearing the pte by writing zero to it also resets the present bit, which will stop the MMU from modifying the pte and allows the processing of the bits that were set. Then the pte is set to its new value. While the present bit is not set, accesses to the page mapped by the pte will results in page faults, which may install a new pte over the non present entry. In order to avoid that scenario the page_table_lock is held. An access will still result in a page fault but the fault handler will also try to acquire the page_table_lock. The page_table_lock is released after the pte has been setup by the first process. The second process will now acquire the page_table_lock and find that there is already a pte setup for this page and return without having done anything. This means that a useless page fault has been generated. However, most architectures have the capability to atomically exchange the value of the pte. For those the clearing of pte before setting them to a new value is not necessary. The use of atomic exchanges avoids useless page faults. The following patch introduces two new atomic operations ptep_xchg and ptep_cmpxchg that may be provided by an architecture. The fallback in include/asm-generic/pgtable.h is to simulate both operations through the existing ptep_get_and_clear function. So there is essentially no change if atomic operations on ptes have not been defined. Architectures that do not support atomic operations on ptes may continue to use the clearing of a pte. Atomic operations are enabled for i386, ia64 and x86_64 if a suitable CPU is configured in SMP mode. Generic atomic definitions for ptep_xchg and ptep_cmpxchg have been provided based on the existing xchg() and cmpxchg() functions that already work atomically on many platforms. The provided generic atomic functions may be overridden as usual by defining the appropriate__HAVE_ARCH_xxx constant and providing a different implementation. This patch is a piece of my attempt to reduce the use of the page_table_lock in the page fault handler through atomic operations. This is only possible if it can be ensured that a pte is never cleared if the pte is in use even when the page_table_lock is not held. Clearing a pte before setting it to another value could result in a situation in which a fault generated by another cpu could install a pte which is then immediately overwritten by the first CPU setting the pte to a valid value again. This patch is necessary for the other patches removing the use of the page_table_lock to work properly. Some numbers: AIM7 Benchmark on an 8 processor system: w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 471.79 100 471.7899 12.34 2.30 Wed Mar 30 21:29:54 2005 100 18068.92 89 180.6892 32.21 158.37 Wed Mar 30 21:30:27 2005 200 21427.39 84 107.1369 54.32 315.84 Wed Mar 30 21:31:22 2005 300 21500.87 82 71.6696 81.21 473.74 Wed Mar 30 21:32:43 2005 400 24886.42 83 62.2160 93.55 633.73 Wed Mar 30 21:34:23 2005 500 25658.89 81 51.3178 113.41 789.44 Wed Mar 30 21:36:17 2005 600 25693.47 81 42.8225 135.91 949.00 Wed Mar 30 21:38:33 2005 700 26098.32 80 37.2833 156.10 1108.17 Wed Mar 30 21:41:10 2005 800 26334.25 80 32.9178 176.80 1266.73 Wed Mar 30 21:44:07 2005 900 26913.85 80 29.9043 194.62 1422.11 Wed Mar 30 21:47:22 2005 1000 26749.89 80 26.7499 217.57 1583.95 Wed Mar 30 21:51:01 2005 w/patch: Tasks jobs/min jti jobs/min/task real cpu 1 470.30 100 470.3030 12.38 2.33 Wed Mar 30 21:57:27 2005 100 18465.05 89 184.6505 31.52 158.62 Wed Mar 30 21:57:58 2005 200 22399.26 86 111.9963 51.97 315.95 Wed Mar 30 21:58:51 2005 300 24274.61 84 80.9154 71.93 475.04 Wed Mar 30 22:00:03 2005 400 25120.86 82 62.8021 92.67 634.10 Wed Mar 30 22:01:36 2005 500 25742.87 81 51.4857 113.04 791.13 Wed Mar 30 22:03:30 2005 600 26322.73 82 43.8712 132.66 948.31 Wed Mar 30 22:05:43 2005 700 25718.40 80 36.7406 158.41 1112.30 Wed Mar 30 22:08:22 2005 800 26361.08 80 32.9514 176.62 1269.94 Wed Mar 30 22:11:19 2005 900 26975.67 81 29.9730 194.17 1424.56 Wed Mar 30 22:14:33 2005 1000 26765.51 80 26.7655 217.44 1585.27 Wed Mar 30 22:18:12 2005 There are some minor performance improvements and some minimal losses for other numbers of tasks. The improvement may be due to the avoidance of one store and the avoidance of useless page faults. Signed-off-by: Christoph Lameter Index: linux-2.6.11/mm/rmap.c =================================================================== --- linux-2.6.11.orig/mm/rmap.c 2005-04-29 08:25:55.000000000 -0700 +++ linux-2.6.11/mm/rmap.c 2005-04-29 08:26:12.000000000 -0700 @@ -574,11 +574,6 @@ static int try_to_unmap_one(struct page /* Nuke the page table entry. */ flush_cache_page(vma, address, page_to_pfn(page)); - pteval = ptep_clear_flush(vma, address, pte); - - /* Move the dirty bit to the physical page now the pte is gone. */ - if (pte_dirty(pteval)) - set_page_dirty(page); if (PageAnon(page)) { swp_entry_t entry = { .val = page->private }; @@ -593,10 +588,15 @@ static int try_to_unmap_one(struct page list_add(&mm->mmlist, &init_mm.mmlist); spin_unlock(&mmlist_lock); } - set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); + pteval = ptep_xchg_flush(vma, address, pte, swp_entry_to_pte(entry)); BUG_ON(pte_file(*pte)); dec_mm_counter(mm, anon_rss); - } + } else + pteval = ptep_clear_flush(vma, address, pte); + + /* Move the dirty bit to the physical page now the pte is gone. */ + if (pte_dirty(pteval)) + set_page_dirty(page); inc_mm_counter(mm, rss); page_remove_rmap(page); @@ -689,15 +689,15 @@ static void try_to_unmap_cluster(unsigne if (ptep_clear_flush_young(vma, address, pte)) continue; - /* Nuke the page table entry. */ flush_cache_page(vma, address, pfn); - pteval = ptep_clear_flush(vma, address, pte); /* If nonlinear, store the file page offset in the pte. */ if (page->index != linear_page_index(vma, address)) - set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); + pteval = ptep_xchg_flush(vma, address, pte, pgoff_to_pte(page->index)); + else + pteval = ptep_clear_flush(vma, address, pte); - /* Move the dirty bit to the physical page now the pte is gone. */ + /* Move the dirty bit to the physical page now that the pte is gone. */ if (pte_dirty(pteval)) set_page_dirty(page); Index: linux-2.6.11/mm/mprotect.c =================================================================== --- linux-2.6.11.orig/mm/mprotect.c 2005-04-29 08:25:55.000000000 -0700 +++ linux-2.6.11/mm/mprotect.c 2005-04-29 08:26:12.000000000 -0700 @@ -32,17 +32,19 @@ static void change_pte_range(struct mm_s pte = pte_offset_map(pmd, addr); do { - if (pte_present(*pte)) { - pte_t ptent; + pte_t ptent; +redo: + ptent = *pte; + if (!pte_present(ptent)) + continue; - /* Avoid an SMP race with hardware updated dirty/clean - * bits by wiping the pte and then setting the new pte - * into place. - */ - ptent = pte_modify(ptep_get_and_clear(mm, addr, pte), newprot); - set_pte_at(mm, addr, pte, ptent); - lazy_mmu_prot_update(ptent); - } + /* Deal with a potential SMP race with hardware/arch + * interrupt updating dirty/clean bits through the use + * of ptep_cmpxchg. + */ + if (!ptep_cmpxchg(mm, addr, pte, ptent, pte_modify(ptent, newprot))) + goto redo; + lazy_mmu_prot_update(ptent); } while (pte++, addr += PAGE_SIZE, addr != end); pte_unmap(pte - 1); } Index: linux-2.6.11/include/asm-generic/pgtable.h =================================================================== --- linux-2.6.11.orig/include/asm-generic/pgtable.h 2005-04-29 08:25:54.000000000 -0700 +++ linux-2.6.11/include/asm-generic/pgtable.h 2005-04-29 08:26:12.000000000 -0700 @@ -111,6 +111,92 @@ do { \ }) #endif +#ifdef CONFIG_ATOMIC_TABLE_OPS + +/* + * The architecture does support atomic table operations. + * We may be able to provide atomic ptep_xchg and ptep_cmpxchg using + * cmpxchg and xchg. + */ +#ifndef __HAVE_ARCH_PTEP_XCHG +#define ptep_xchg(__mm, __address, __ptep, __pteval) \ + __pte(xchg(&pte_val(*(__ptep)), pte_val(__pteval))) +#endif + +#ifndef __HAVE_ARCH_PTEP_CMPXCHG +#define ptep_cmpxchg(__mm, __address, __ptep,__oldval,__newval) \ + (cmpxchg(&pte_val(*(__ptep)), \ + pte_val(__oldval), \ + pte_val(__newval) \ + ) == pte_val(__oldval) \ + ) +#endif + +#ifndef __HAVE_ARCH_PTEP_XCHG_FLUSH +#define ptep_xchg_flush(__vma, __address, __ptep, __pteval) \ +({ \ + pte_t __pte = ptep_xchg(__vma, __address, __ptep, __pteval); \ + flush_tlb_page(__vma, __address); \ + __pte; \ +}) +#endif + +#else + +/* + * No support for atomic operations on the page table. + * Exchanging of pte values is done by first swapping zeros into + * a pte and then putting new content into the pte entry. + * However, these functions will generate an empty pte for a + * short time frame. This means that the page_table_lock must be held + * to avoid a page fault that would install a new entry. + */ +#ifndef __HAVE_ARCH_PTEP_XCHG +#define ptep_xchg(__mm, __address, __ptep, __pteval) \ +({ \ + pte_t __pte = ptep_get_and_clear(__mm, __address, __ptep); \ + set_pte_at(__mm, __address, __ptep, __pteval); \ + __pte; \ +}) +#endif + +#ifndef __HAVE_ARCH_PTEP_XCHG_FLUSH +#ifndef __HAVE_ARCH_PTEP_XCHG +#define ptep_xchg_flush(__vma, __address, __ptep, __pteval) \ +({ \ + pte_t __pte = ptep_clear_flush(__vma, __address, __ptep); \ + set_pte_at((__vma)->mm, __address, __ptep, __pteval); \ + __pte; \ +}) +#else +#define ptep_xchg_flush(__vma, __address, __ptep, __pteval) \ +({ \ + pte_t __pte = ptep_xchg((__vma)->mm, __address, __ptep, __pteval);\ + flush_tlb_page(__vma, __address); \ + __pte; \ +}) +#endif +#endif + +/* + * The fallback function for ptep_cmpxchg avoids any real use of cmpxchg + * since cmpxchg may not be available on certain architectures. Instead + * the clearing of a pte is used as a form of locking mechanism. + * This approach will only work if the page_table_lock is held to insure + * that the pte is not populated by a page fault generated on another + * CPU. + */ +#ifndef __HAVE_ARCH_PTEP_CMPXCHG +#define ptep_cmpxchg(__mm, __address, __ptep, __old, __new) \ +({ \ + pte_t prev = ptep_get_and_clear(__mm, __address, __ptep); \ + int r = pte_val(prev) == pte_val(__old); \ + set_pte_at(__mm, __address, __ptep, r ? (__new) : prev); \ + r; \ +}) +#endif +#endif + #ifndef __HAVE_ARCH_PTEP_SET_WRPROTECT static inline void ptep_set_wrprotect(struct mm_struct *mm, unsigned long address, pte_t *ptep) { Index: linux-2.6.11/arch/ia64/Kconfig =================================================================== --- linux-2.6.11.orig/arch/ia64/Kconfig 2005-04-29 08:26:06.000000000 -0700 +++ linux-2.6.11/arch/ia64/Kconfig 2005-04-29 09:23:07.000000000 -0700 @@ -273,6 +273,11 @@ config PREEMPT Say Y here if you are building a kernel for a desktop, embedded or real-time system. Say N if you are unsure. +config ATOMIC_TABLE_OPS + bool + depends on SMP + default y + config HAVE_DEC_LOCK bool depends on (SMP || PREEMPT) Index: linux-2.6.11/arch/i386/Kconfig =================================================================== --- linux-2.6.11.orig/arch/i386/Kconfig 2005-04-29 08:25:50.000000000 -0700 +++ linux-2.6.11/arch/i386/Kconfig 2005-04-29 08:26:12.000000000 -0700 @@ -886,6 +886,11 @@ config HAVE_DEC_LOCK depends on (SMP || PREEMPT) && X86_CMPXCHG default y +config ATOMIC_TABLE_OPS + bool + depends on SMP && X86_CMPXCHG && !X86_PAE + default y + # turning this on wastes a bunch of space. # Summit needs it only when NUMA is on config BOOT_IOREMAP Index: linux-2.6.11/arch/x86_64/Kconfig =================================================================== --- linux-2.6.11.orig/arch/x86_64/Kconfig 2005-04-29 08:25:51.000000000 -0700 +++ linux-2.6.11/arch/x86_64/Kconfig 2005-04-29 08:26:12.000000000 -0700 @@ -223,6 +223,11 @@ config PREEMPT Say Y here if you are feeling brave and building a kernel for a desktop, embedded or real-time system. Say N if you are unsure. +config ATOMIC_TABLE_OPS + bool + depends on SMP + default y + config PREEMPT_BKL bool "Preempt The Big Kernel Lock" depends on PREEMPT Index: linux-2.6.11/mm/memory.c =================================================================== --- linux-2.6.11.orig/mm/memory.c 2005-04-29 08:25:55.000000000 -0700 +++ linux-2.6.11/mm/memory.c 2005-04-29 08:26:12.000000000 -0700 @@ -551,15 +551,19 @@ static void zap_pte_range(struct mmu_gat page->index > details->last_index)) continue; } - ptent = ptep_get_and_clear(tlb->mm, addr, pte); - tlb_remove_tlb_entry(tlb, pte, addr); - if (unlikely(!page)) + if (unlikely(!page)) { + ptent = ptep_get_and_clear(tlb->mm, addr, pte); + tlb_remove_tlb_entry(tlb, pte, addr); continue; + } if (unlikely(details) && details->nonlinear_vma && linear_page_index(details->nonlinear_vma, addr) != page->index) - set_pte_at(tlb->mm, addr, pte, - pgoff_to_pte(page->index)); + ptent = ptep_xchg(tlb->mm, addr, pte, + pgoff_to_pte(page->index)); + else + ptent = ptep_get_and_clear(tlb->mm, addr, pte); + tlb_remove_tlb_entry(tlb, pte, addr); if (pte_dirty(ptent)) set_page_dirty(page); if (PageAnon(page)) - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/