From owner-svn-src-stable-11@freebsd.org Sat Oct 1 19:30:29 2016 Return-Path: Delivered-To: svn-src-stable-11@mailman.ysv.freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2001:1900:2254:206a::19:1]) by mailman.ysv.freebsd.org (Postfix) with ESMTP id 9740AA94231; Sat, 1 Oct 2016 19:30:29 +0000 (UTC) (envelope-from alc@FreeBSD.org) Received: from repo.freebsd.org (repo.freebsd.org [IPv6:2610:1c1:1:6068::e6a:0]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (Client did not present a certificate) by mx1.freebsd.org (Postfix) with ESMTPS id 6C2DDB87; Sat, 1 Oct 2016 19:30:29 +0000 (UTC) (envelope-from alc@FreeBSD.org) Received: from repo.freebsd.org ([127.0.1.37]) by repo.freebsd.org (8.15.2/8.15.2) with ESMTP id u91JUSBl072655; Sat, 1 Oct 2016 19:30:28 GMT (envelope-from alc@FreeBSD.org) Received: (from alc@localhost) by repo.freebsd.org (8.15.2/8.15.2/Submit) id u91JUSYF072651; Sat, 1 Oct 2016 19:30:28 GMT (envelope-from alc@FreeBSD.org) Message-Id: <201610011930.u91JUSYF072651@repo.freebsd.org> X-Authentication-Warning: repo.freebsd.org: alc set sender to alc@FreeBSD.org using -f From: Alan Cox Date: Sat, 1 Oct 2016 19:30:28 +0000 (UTC) To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-stable@freebsd.org, svn-src-stable-11@freebsd.org Subject: svn commit: r306558 - in stable/11/sys: amd64/amd64 arm/arm i386/i386 sparc64/sparc64 X-SVN-Group: stable-11 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit X-BeenThere: svn-src-stable-11@freebsd.org X-Mailman-Version: 2.1.23 Precedence: list List-Id: SVN commit messages for only the 11-stable src tree List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Sat, 01 Oct 2016 19:30:29 -0000 Author: alc Date: Sat Oct 1 19:30:28 2016 New Revision: 306558 URL: https://svnweb.freebsd.org/changeset/base/306558 Log: MFC r305213,305319,305398 As an optimization to the machine-independent layer, change the machine- dependent pmap_ts_referenced() so that it updates the page's dirty field if a modified bit is found while counting reference bits. This opportunistic update can be performed at low cost and can eliminate the need for some future calls to pmap_is_modified() by the machine- independent layer. Replace the number 4 in sparc64's pmap_ts_referenced() by PMAP_TS_REFERENCED_MAX, like we've done elsewhere, e.g., amd64. Modified: stable/11/sys/amd64/amd64/pmap.c stable/11/sys/arm/arm/pmap-v6.c stable/11/sys/i386/i386/pmap.c stable/11/sys/sparc64/sparc64/pmap.c Directory Properties: stable/11/ (props changed) Modified: stable/11/sys/amd64/amd64/pmap.c ============================================================================== --- stable/11/sys/amd64/amd64/pmap.c Sat Oct 1 17:57:32 2016 (r306557) +++ stable/11/sys/amd64/amd64/pmap.c Sat Oct 1 19:30:28 2016 (r306558) @@ -5831,6 +5831,14 @@ safe_to_clear_referenced(pmap_t pmap, pt * should be tested and standardized at some point in the future for * optimal aging of shared pages. * + * As an optimization, update the page's dirty field if a modified bit is + * found while counting reference bits. This opportunistic update can be + * performed at low cost and can eliminate the need for some future calls + * to pmap_is_modified(). However, since this function stops after + * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some + * dirty pages. Those dirty pages will only be detected by a future call + * to pmap_is_modified(). + * * A DI block is not needed within this function, because * invalidations are performed before the PV list lock is * released. @@ -5843,7 +5851,7 @@ pmap_ts_referenced(vm_page_t m) pmap_t pmap; struct rwlock *lock; pd_entry_t oldpde, *pde; - pt_entry_t *pte, PG_A; + pt_entry_t *pte, PG_A, PG_M, PG_RW; vm_offset_t va; vm_paddr_t pa; int cleared, md_gen, not_cleared, pvh_gen; @@ -5878,9 +5886,19 @@ retry: } } PG_A = pmap_accessed_bit(pmap); + PG_M = pmap_modified_bit(pmap); + PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, pv->pv_va); oldpde = *pde; + if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) { + /* + * Although "oldpde" is mapping a 2MB page, because + * this function is called at a 4KB page granularity, + * we only update the 4KB page under test. + */ + vm_page_dirty(m); + } if ((*pde & PG_A) != 0) { /* * Since this reference bit is shared by 512 4KB @@ -5974,11 +5992,15 @@ small_mappings: } } PG_A = pmap_accessed_bit(pmap); + PG_M = pmap_modified_bit(pmap); + PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_ts_referenced: found a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); + if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) + vm_page_dirty(m); if ((*pte & PG_A) != 0) { if (safe_to_clear_referenced(pmap, *pte)) { atomic_clear_long(pte, PG_A); Modified: stable/11/sys/arm/arm/pmap-v6.c ============================================================================== --- stable/11/sys/arm/arm/pmap-v6.c Sat Oct 1 17:57:32 2016 (r306557) +++ stable/11/sys/arm/arm/pmap-v6.c Sat Oct 1 19:30:28 2016 (r306558) @@ -5178,6 +5178,14 @@ pmap_is_referenced(vm_page_t m) * XXX: The exact number of bits to check and clear is a matter that * should be tested and standardized at some point in the future for * optimal aging of shared pages. + * + * As an optimization, update the page's dirty field if a modified bit is + * found while counting reference bits. This opportunistic update can be + * performed at low cost and can eliminate the need for some future calls + * to pmap_is_modified(). However, since this function stops after + * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some + * dirty pages. Those dirty pages will only be detected by a future call + * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) @@ -5186,7 +5194,7 @@ pmap_ts_referenced(vm_page_t m) pv_entry_t pv, pvf; pmap_t pmap; pt1_entry_t *pte1p, opte1; - pt2_entry_t *pte2p; + pt2_entry_t *pte2p, opte2; vm_paddr_t pa; int rtval = 0; @@ -5205,6 +5213,14 @@ pmap_ts_referenced(vm_page_t m) PMAP_LOCK(pmap); pte1p = pmap_pte1(pmap, pv->pv_va); opte1 = pte1_load(pte1p); + if (pte1_is_dirty(opte1)) { + /* + * Although "opte1" is mapping a 1MB page, because + * this function is called at a 4KB page granularity, + * we only update the 4KB page under test. + */ + vm_page_dirty(m); + } if ((opte1 & PTE1_A) != 0) { /* * Since this reference bit is shared by 256 4KB pages, @@ -5253,7 +5269,10 @@ small_mappings: ("%s: not found a link in page %p's pv list", __func__, m)); pte2p = pmap_pte2_quick(pmap, pv->pv_va); - if ((pte2_load(pte2p) & PTE2_A) != 0) { + opte2 = pte2_load(pte2p); + if (pte2_is_dirty(opte2)) + vm_page_dirty(m); + if ((opte2 & PTE2_A) != 0) { pte2_clear_bit(pte2p, PTE2_A); pmap_tlb_flush(pmap, pv->pv_va); rtval++; Modified: stable/11/sys/i386/i386/pmap.c ============================================================================== --- stable/11/sys/i386/i386/pmap.c Sat Oct 1 17:57:32 2016 (r306557) +++ stable/11/sys/i386/i386/pmap.c Sat Oct 1 19:30:28 2016 (r306558) @@ -4802,6 +4802,14 @@ retry: * XXX: The exact number of bits to check and clear is a matter that * should be tested and standardized at some point in the future for * optimal aging of shared pages. + * + * As an optimization, update the page's dirty field if a modified bit is + * found while counting reference bits. This opportunistic update can be + * performed at low cost and can eliminate the need for some future calls + * to pmap_is_modified(). However, since this function stops after + * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some + * dirty pages. Those dirty pages will only be detected by a future call + * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) @@ -4828,6 +4836,14 @@ pmap_ts_referenced(vm_page_t m) pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); + if ((*pde & (PG_M | PG_RW)) == (PG_M | PG_RW)) { + /* + * Although "*pde" is mapping a 2/4MB page, because + * this function is called at a 4KB page granularity, + * we only update the 4KB page under test. + */ + vm_page_dirty(m); + } if ((*pde & PG_A) != 0) { /* * Since this reference bit is shared by either 1024 @@ -4876,6 +4892,8 @@ small_mappings: ("pmap_ts_referenced: found a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); + if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) + vm_page_dirty(m); if ((*pte & PG_A) != 0) { atomic_clear_int((u_int *)pte, PG_A); pmap_invalidate_page(pmap, pv->pv_va); Modified: stable/11/sys/sparc64/sparc64/pmap.c ============================================================================== --- stable/11/sys/sparc64/sparc64/pmap.c Sat Oct 1 17:57:32 2016 (r306557) +++ stable/11/sys/sparc64/sparc64/pmap.c Sat Oct 1 19:30:28 2016 (r306558) @@ -2106,6 +2106,8 @@ pmap_page_is_mapped(vm_page_t m) return (rv); } +#define PMAP_TS_REFERENCED_MAX 5 + /* * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it @@ -2115,6 +2117,14 @@ pmap_page_is_mapped(vm_page_t m) * XXX: The exact number of bits to check and clear is a matter that * should be tested and standardized at some point in the future for * optimal aging of shared pages. + * + * As an optimization, update the page's dirty field if a modified bit is + * found while counting reference bits. This opportunistic update can be + * performed at low cost and can eliminate the need for some future calls + * to pmap_is_modified(). However, since this function stops after + * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some + * dirty pages. Those dirty pages will only be detected by a future call + * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) @@ -2138,7 +2148,10 @@ pmap_ts_referenced(vm_page_t m) if ((tp->tte_data & TD_PV) == 0) continue; data = atomic_clear_long(&tp->tte_data, TD_REF); - if ((data & TD_REF) != 0 && ++count > 4) + if ((data & TD_W) != 0) + vm_page_dirty(m); + if ((data & TD_REF) != 0 && ++count >= + PMAP_TS_REFERENCED_MAX) break; } while ((tp = tpn) != NULL && tp != tpf); }