From owner-svn-src-all@freebsd.org Sun Feb 26 19:54:04 2017 Return-Path: Delivered-To: svn-src-all@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 41693CEEF71; Sun, 26 Feb 2017 19:54:04 +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 03A68995; Sun, 26 Feb 2017 19:54:03 +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 v1QJs39b077065; Sun, 26 Feb 2017 19:54:03 GMT (envelope-from alc@FreeBSD.org) Received: (from alc@localhost) by repo.freebsd.org (8.15.2/8.15.2/Submit) id v1QJs2J4077061; Sun, 26 Feb 2017 19:54:02 GMT (envelope-from alc@FreeBSD.org) Message-Id: <201702261954.v1QJs2J4077061@repo.freebsd.org> X-Authentication-Warning: repo.freebsd.org: alc set sender to alc@FreeBSD.org using -f From: Alan Cox Date: Sun, 26 Feb 2017 19:54:02 +0000 (UTC) To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: svn commit: r314310 - in head/sys: amd64/amd64 amd64/include i386/i386 i386/include X-SVN-Group: head MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit X-BeenThere: svn-src-all@freebsd.org X-Mailman-Version: 2.1.23 Precedence: list List-Id: "SVN commit messages for the entire src tree \(except for " user" and " projects" \)" List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Sun, 26 Feb 2017 19:54:04 -0000 Author: alc Date: Sun Feb 26 19:54:02 2017 New Revision: 314310 URL: https://svnweb.freebsd.org/changeset/base/314310 Log: Refine the fix from r312954. Specifically, add a new PDE-only flag, PG_PROMOTED, that indicates whether lingering 4KB page mappings might need to be flushed on a PDE change that restricts or destroys a 2MB page mapping. This flag allows the pmap to avoid range invalidations that are both unnecessary and costly. Reviewed by: kib, markj MFC after: 6 weeks Differential Revision: https://reviews.freebsd.org/D9665 Modified: head/sys/amd64/amd64/pmap.c head/sys/amd64/include/pmap.h head/sys/i386/i386/pmap.c head/sys/i386/include/pmap.h Modified: head/sys/amd64/amd64/pmap.c ============================================================================== --- head/sys/amd64/amd64/pmap.c Sun Feb 26 19:51:44 2017 (r314309) +++ head/sys/amd64/amd64/pmap.c Sun Feb 26 19:54:02 2017 (r314310) @@ -613,6 +613,8 @@ static vm_page_t pmap_enter_quick_locked vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); +static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, + pd_entry_t pde); static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode); static void pmap_pde_attr(pd_entry_t *pde, int cache_bits, int mask); static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, @@ -1838,6 +1840,27 @@ pmap_update_pde(pmap_t pmap, vm_offset_t } #endif /* !SMP */ +static void +pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde) +{ + + /* + * When the PDE has PG_PROMOTED set, the 2MB page mapping was created + * by a promotion that did not invalidate the 512 4KB page mappings + * that might exist in the TLB. Consequently, at this point, the TLB + * may hold both 4KB and 2MB page mappings for the address range [va, + * va + NBPDR). Therefore, the entire range must be invalidated here. + * In contrast, when PG_PROMOTED is clear, the TLB will not hold any + * 4KB page mappings for the address range [va, va + NBPDR), and so a + * single INVLPG suffices to invalidate the 2MB page mapping from the + * TLB. + */ + if ((pde & PG_PROMOTED) != 0) + pmap_invalidate_range(pmap, va, va + NBPDR - 1); + else + pmap_invalidate_page(pmap, va); +} + #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024) void @@ -3472,7 +3495,8 @@ pmap_demote_pde_locked(pmap_t pmap, pd_e SLIST_INIT(&free); sva = trunc_2mpage(va); pmap_remove_pde(pmap, pde, sva, &free, lockp); - pmap_invalidate_range(pmap, sva, sva + NBPDR - 1); + if ((oldpde & PG_G) == 0) + pmap_invalidate_pde_page(pmap, sva, oldpde); pmap_free_zero_pages(&free); CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#lx" " in pmap %p", va, pmap); @@ -3612,25 +3636,8 @@ pmap_remove_pde(pmap_t pmap, pd_entry_t oldpde = pte_load_clear(pdq); if (oldpde & PG_W) pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; - - /* - * When workaround_erratum383 is false, a promotion to a 2M - * page mapping does not invalidate the 512 4K page mappings - * from the TLB. Consequently, at this point, the TLB may - * hold both 4K and 2M page mappings. Therefore, the entire - * range of addresses must be invalidated here. In contrast, - * when workaround_erratum383 is true, a promotion does - * invalidate the 512 4K page mappings, and so a single INVLPG - * suffices to invalidate the 2M page mapping. - */ - if ((oldpde & PG_G) != 0) { - if (workaround_erratum383) - pmap_invalidate_page(kernel_pmap, sva); - else - pmap_invalidate_range(kernel_pmap, sva, - sva + NBPDR - 1); - } - + if ((oldpde & PG_G) != 0) + pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); pmap_resident_count_dec(pmap, NBPDR / PAGE_SIZE); if (oldpde & PG_MANAGED) { CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, oldpde & PG_PS_FRAME); @@ -4010,16 +4017,16 @@ retry: if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; if (newpde != oldpde) { - if (!atomic_cmpset_long(pde, oldpde, newpde)) + /* + * As an optimization to future operations on this PDE, clear + * PG_PROMOTED. The impending invalidation will remove any + * lingering 4KB page mappings from the TLB. + */ + if (!atomic_cmpset_long(pde, oldpde, newpde & ~PG_PROMOTED)) goto retry; - if (oldpde & PG_G) { - /* See pmap_remove_pde() for explanation. */ - if (workaround_erratum383) - pmap_invalidate_page(kernel_pmap, sva); - else - pmap_invalidate_range(kernel_pmap, sva, - sva + NBPDR - 1); - } else + if ((oldpde & PG_G) != 0) + pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); + else anychanged = TRUE; } return (anychanged); @@ -4272,7 +4279,7 @@ setpte: if (workaround_erratum383) pmap_update_pde(pmap, va, pde, PG_PS | newpde); else - pde_store(pde, PG_PS | newpde); + pde_store(pde, PG_PROMOTED | PG_PS | newpde); atomic_add_long(&pmap_pde_promotions, 1); CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#lx" @@ -4585,7 +4592,8 @@ pmap_enter_pde(pmap_t pmap, vm_offset_t pmap_resident_count_inc(pmap, NBPDR / PAGE_SIZE); /* - * Map the superpage. + * Map the superpage. (This is not a promoted mapping; there will not + * be any lingering 4KB page mappings in the TLB.) */ pde_store(pde, newpde); Modified: head/sys/amd64/include/pmap.h ============================================================================== --- head/sys/amd64/include/pmap.h Sun Feb 26 19:51:44 2017 (r314309) +++ head/sys/amd64/include/pmap.h Sun Feb 26 19:54:02 2017 (r314310) @@ -109,6 +109,7 @@ #define PG_MANAGED X86_PG_AVAIL2 #define EPT_PG_EMUL_V X86_PG_AVAIL(52) #define EPT_PG_EMUL_RW X86_PG_AVAIL(53) +#define PG_PROMOTED X86_PG_AVAIL(54) /* PDE only */ #define PG_FRAME (0x000ffffffffff000ul) #define PG_PS_FRAME (0x000fffffffe00000ul) Modified: head/sys/i386/i386/pmap.c ============================================================================== --- head/sys/i386/i386/pmap.c Sun Feb 26 19:51:44 2017 (r314309) +++ head/sys/i386/i386/pmap.c Sun Feb 26 19:54:02 2017 (r314310) @@ -301,6 +301,8 @@ static vm_page_t pmap_enter_quick_locked vm_page_t m, vm_prot_t prot, vm_page_t mpte); static void pmap_flush_page(vm_page_t m); static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); +static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, + pd_entry_t pde); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); static boolean_t pmap_is_modified_pvh(struct md_page *pvh); static boolean_t pmap_is_referenced_pvh(struct md_page *pvh); @@ -1259,6 +1261,27 @@ pmap_update_pde(pmap_t pmap, vm_offset_t } #endif /* !SMP */ +static void +pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde) +{ + + /* + * When the PDE has PG_PROMOTED set, the 2- or 4MB page mapping was + * created by a promotion that did not invalidate the 512 or 1024 4KB + * page mappings that might exist in the TLB. Consequently, at this + * point, the TLB may hold both 4KB and 2- or 4MB page mappings for + * the address range [va, va + NBPDR). Therefore, the entire range + * must be invalidated here. In contrast, when PG_PROMOTED is clear, + * the TLB will not hold any 4KB page mappings for the address range + * [va, va + NBPDR), and so a single INVLPG suffices to invalidate the + * 2- or 4MB page mapping from the TLB. + */ + if ((pde & PG_PROMOTED) != 0) + pmap_invalidate_range(pmap, va, va + NBPDR - 1); + else + pmap_invalidate_page(pmap, va); +} + #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024) void @@ -2649,7 +2672,8 @@ pmap_demote_pde(pmap_t pmap, pd_entry_t SLIST_INIT(&free); sva = trunc_4mpage(va); pmap_remove_pde(pmap, pde, sva, &free); - pmap_invalidate_range(pmap, sva, sva + NBPDR - 1); + if ((oldpde & PG_G) == 0) + pmap_invalidate_pde_page(pmap, sva, oldpde); pmap_free_zero_pages(&free); CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#x" " in pmap %p", va, pmap); @@ -2819,23 +2843,9 @@ pmap_remove_pde(pmap_t pmap, pd_entry_t /* * Machines that don't support invlpg, also don't support * PG_G. - * - * When workaround_erratum383 is false, a promotion to a 2M/4M - * page mapping does not invalidate the 512/1024 4K page mappings - * from the TLB. Consequently, at this point, the TLB may - * hold both 4K and 2M/4M page mappings. Therefore, the entire - * range of addresses must be invalidated here. In contrast, - * when workaround_erratum383 is true, a promotion does - * invalidate the 512/1024 4K page mappings, and so a single INVLPG - * suffices to invalidate the 2M/4M page mapping. - */ - if ((oldpde & PG_G) != 0) { - if (workaround_erratum383) - pmap_invalidate_page(kernel_pmap, sva); - else - pmap_invalidate_range(kernel_pmap, sva, - sva + NBPDR - 1); - } + */ + if ((oldpde & PG_G) != 0) + pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; if (oldpde & PG_MANAGED) { @@ -3143,16 +3153,16 @@ retry: newpde |= pg_nx; #endif if (newpde != oldpde) { - if (!pde_cmpset(pde, oldpde, newpde)) + /* + * As an optimization to future operations on this PDE, clear + * PG_PROMOTED. The impending invalidation will remove any + * lingering 4KB page mappings from the TLB. + */ + if (!pde_cmpset(pde, oldpde, newpde & ~PG_PROMOTED)) goto retry; - if (oldpde & PG_G) { - /* See pmap_remove_pde() for explanation. */ - if (workaround_erratum383) - pmap_invalidate_page(kernel_pmap, sva); - else - pmap_invalidate_range(kernel_pmap, sva, - sva + NBPDR - 1); - } else + if ((oldpde & PG_G) != 0) + pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); + else anychanged = TRUE; } return (anychanged); @@ -3437,9 +3447,9 @@ setpte: if (workaround_erratum383) pmap_update_pde(pmap, va, pde, PG_PS | newpde); else if (pmap == kernel_pmap) - pmap_kenter_pde(va, PG_PS | newpde); + pmap_kenter_pde(va, PG_PROMOTED | PG_PS | newpde); else - pde_store(pde, PG_PS | newpde); + pde_store(pde, PG_PROMOTED | PG_PS | newpde); pmap_pde_promotions++; CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#x" @@ -3722,7 +3732,8 @@ pmap_enter_pde(pmap_t pmap, vm_offset_t pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; /* - * Map the superpage. + * Map the superpage. (This is not a promoted mapping; there will not + * be any lingering 4KB page mappings in the TLB.) */ pde_store(pde, newpde); Modified: head/sys/i386/include/pmap.h ============================================================================== --- head/sys/i386/include/pmap.h Sun Feb 26 19:51:44 2017 (r314309) +++ head/sys/i386/include/pmap.h Sun Feb 26 19:54:02 2017 (r314310) @@ -71,6 +71,7 @@ /* Our various interpretations of the above */ #define PG_W PG_AVAIL1 /* "Wired" pseudoflag */ #define PG_MANAGED PG_AVAIL2 +#define PG_PROMOTED PG_AVAIL3 /* PDE only */ #if defined(PAE) || defined(PAE_TABLES) #define PG_FRAME (0x000ffffffffff000ull) #define PG_PS_FRAME (0x000fffffffe00000ull)