diff --git a/sys/arm/nvidia/drm2/tegra_bo.c b/sys/arm/nvidia/drm2/tegra_bo.c index ae0c48d6b677..afec9f50ff50 100644 --- a/sys/arm/nvidia/drm2/tegra_bo.c +++ b/sys/arm/nvidia/drm2/tegra_bo.c @@ -1,360 +1,363 @@ /*- * Copyright (c) 2015 Michal Meloun * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static void tegra_bo_destruct(struct tegra_bo *bo) { vm_page_t m; size_t size; int i; if (bo->cdev_pager == NULL) return; size = round_page(bo->gem_obj.size); if (bo->vbase != 0) pmap_qremove(bo->vbase, bo->npages); VM_OBJECT_WLOCK(bo->cdev_pager); for (i = 0; i < bo->npages; i++) { m = bo->m[i]; vm_page_busy_acquire(m, 0); cdev_pager_free_page(bo->cdev_pager, m); m->flags &= ~PG_FICTITIOUS; vm_page_unwire_noq(m); vm_page_free(m); } VM_OBJECT_WUNLOCK(bo->cdev_pager); vm_object_deallocate(bo->cdev_pager); if (bo->vbase != 0) vmem_free(kmem_arena, bo->vbase, size); } static void tegra_bo_free_object(struct drm_gem_object *gem_obj) { struct tegra_bo *bo; bo = container_of(gem_obj, struct tegra_bo, gem_obj); drm_gem_free_mmap_offset(gem_obj); drm_gem_object_release(gem_obj); tegra_bo_destruct(bo); free(bo->m, DRM_MEM_DRIVER); free(bo, DRM_MEM_DRIVER); } static int tegra_bo_alloc_contig(size_t npages, u_long alignment, vm_memattr_t memattr, vm_page_t **ret_page) { vm_page_t m; - int tries, i; + int err, i, tries; vm_paddr_t low, high, boundary; low = 0; high = -1UL; boundary = 0; tries = 0; retry: m = vm_page_alloc_noobj_contig(VM_ALLOC_WIRED | VM_ALLOC_ZERO, npages, low, high, alignment, boundary, memattr); if (m == NULL) { if (tries < 3) { - if (!vm_page_reclaim_contig(0, npages, low, high, - alignment, boundary)) + err = vm_page_reclaim_contig(0, npages, low, high, + alignment, boundary); + if (err == ENOMEM) vm_wait(NULL); + else if (err != 0) + return (ENOMEM); tries++; goto retry; } return (ENOMEM); } for (i = 0; i < npages; i++, m++) { m->valid = VM_PAGE_BITS_ALL; (*ret_page)[i] = m; } return (0); } /* Initialize pager and insert all object pages to it*/ static int tegra_bo_init_pager(struct tegra_bo *bo) { vm_page_t m; size_t size; int i; size = round_page(bo->gem_obj.size); bo->pbase = VM_PAGE_TO_PHYS(bo->m[0]); if (vmem_alloc(kmem_arena, size, M_WAITOK | M_BESTFIT, &bo->vbase)) return (ENOMEM); VM_OBJECT_WLOCK(bo->cdev_pager); for (i = 0; i < bo->npages; i++) { m = bo->m[i]; /* * XXX This is a temporary hack. * We need pager suitable for paging (mmap) managed * real (non-fictitious) pages. * - managed pages are needed for clean module unload. * - aliasing fictitious page to real one is bad, * pmap cannot handle this situation without issues * It expects that * paddr = PHYS_TO_VM_PAGE(VM_PAGE_TO_PHYS(paddr)) * for every single page passed to pmap. */ m->oflags &= ~VPO_UNMANAGED; m->flags |= PG_FICTITIOUS; if (vm_page_insert(m, bo->cdev_pager, i) != 0) return (EINVAL); } VM_OBJECT_WUNLOCK(bo->cdev_pager); pmap_qenter(bo->vbase, bo->m, bo->npages); return (0); } /* Allocate memory for frame buffer */ static int tegra_bo_alloc(struct drm_device *drm, struct tegra_bo *bo) { size_t size; int rv; size = bo->gem_obj.size; bo->npages = atop(size); bo->m = malloc(sizeof(vm_page_t *) * bo->npages, DRM_MEM_DRIVER, M_WAITOK | M_ZERO); rv = tegra_bo_alloc_contig(bo->npages, PAGE_SIZE, VM_MEMATTR_WRITE_COMBINING, &(bo->m)); if (rv != 0) { DRM_WARNING("Cannot allocate memory for gem object.\n"); return (rv); } rv = tegra_bo_init_pager(bo); if (rv != 0) { DRM_WARNING("Cannot initialize gem object pager.\n"); return (rv); } return (0); } int tegra_bo_create(struct drm_device *drm, size_t size, struct tegra_bo **res_bo) { struct tegra_bo *bo; int rv; if (size <= 0) return (-EINVAL); bo = malloc(sizeof(*bo), DRM_MEM_DRIVER, M_WAITOK | M_ZERO); size = round_page(size); rv = drm_gem_object_init(drm, &bo->gem_obj, size); if (rv != 0) { free(bo, DRM_MEM_DRIVER); return (rv); } rv = drm_gem_create_mmap_offset(&bo->gem_obj); if (rv != 0) { drm_gem_object_release(&bo->gem_obj); free(bo, DRM_MEM_DRIVER); return (rv); } bo->cdev_pager = cdev_pager_allocate(&bo->gem_obj, OBJT_MGTDEVICE, drm->driver->gem_pager_ops, size, 0, 0, NULL); rv = tegra_bo_alloc(drm, bo); if (rv != 0) { tegra_bo_free_object(&bo->gem_obj); return (rv); } *res_bo = bo; return (0); } static int tegra_bo_create_with_handle(struct drm_file *file, struct drm_device *drm, size_t size, uint32_t *handle, struct tegra_bo **res_bo) { int rv; struct tegra_bo *bo; rv = tegra_bo_create(drm, size, &bo); if (rv != 0) return (rv); rv = drm_gem_handle_create(file, &bo->gem_obj, handle); if (rv != 0) { tegra_bo_free_object(&bo->gem_obj); drm_gem_object_release(&bo->gem_obj); return (rv); } drm_gem_object_unreference_unlocked(&bo->gem_obj); *res_bo = bo; return (0); } static int tegra_bo_dumb_create(struct drm_file *file, struct drm_device *drm_dev, struct drm_mode_create_dumb *args) { struct tegra_drm *drm; struct tegra_bo *bo; int rv; drm = container_of(drm_dev, struct tegra_drm, drm_dev); args->pitch= (args->width * args->bpp + 7) / 8; args->pitch = roundup(args->pitch, drm->pitch_align); args->size = args->pitch * args->height; rv = tegra_bo_create_with_handle(file, drm_dev, args->size, &args->handle, &bo); return (rv); } static int tegra_bo_dumb_map_offset(struct drm_file *file_priv, struct drm_device *drm_dev, uint32_t handle, uint64_t *offset) { struct drm_gem_object *gem_obj; int rv; DRM_LOCK(drm_dev); gem_obj = drm_gem_object_lookup(drm_dev, file_priv, handle); if (gem_obj == NULL) { device_printf(drm_dev->dev, "Object not found\n"); DRM_UNLOCK(drm_dev); return (-EINVAL); } rv = drm_gem_create_mmap_offset(gem_obj); if (rv != 0) goto fail; *offset = DRM_GEM_MAPPING_OFF(gem_obj->map_list.key) | DRM_GEM_MAPPING_KEY; drm_gem_object_unreference(gem_obj); DRM_UNLOCK(drm_dev); return (0); fail: drm_gem_object_unreference(gem_obj); DRM_UNLOCK(drm_dev); return (rv); } static int tegra_bo_dumb_destroy(struct drm_file *file_priv, struct drm_device *drm_dev, unsigned int handle) { int rv; rv = drm_gem_handle_delete(file_priv, handle); return (rv); } /* * mmap support */ static int tegra_gem_pager_fault(vm_object_t vm_obj, vm_ooffset_t offset, int prot, vm_page_t *mres) { #ifdef DRM_PAGER_DEBUG DRM_DEBUG("object %p offset %jd prot %d mres %p\n", vm_obj, (intmax_t)offset, prot, mres); #endif return (VM_PAGER_FAIL); } static int tegra_gem_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { if (color != NULL) *color = 0; return (0); } static void tegra_gem_pager_dtor(void *handle) { } static struct cdev_pager_ops tegra_gem_pager_ops = { .cdev_pg_fault = tegra_gem_pager_fault, .cdev_pg_ctor = tegra_gem_pager_ctor, .cdev_pg_dtor = tegra_gem_pager_dtor }; /* Fill up relevant fields in drm_driver ops */ void tegra_bo_driver_register(struct drm_driver *drm_drv) { drm_drv->gem_free_object = tegra_bo_free_object; drm_drv->gem_pager_ops = &tegra_gem_pager_ops; drm_drv->dumb_create = tegra_bo_dumb_create; drm_drv->dumb_map_offset = tegra_bo_dumb_map_offset; drm_drv->dumb_destroy = tegra_bo_dumb_destroy; } diff --git a/sys/compat/linuxkpi/common/src/linux_page.c b/sys/compat/linuxkpi/common/src/linux_page.c index bb5a1411eeec..b910ada90a4d 100644 --- a/sys/compat/linuxkpi/common/src/linux_page.c +++ b/sys/compat/linuxkpi/common/src/linux_page.c @@ -1,556 +1,558 @@ /*- * Copyright (c) 2010 Isilon Systems, Inc. * Copyright (c) 2016 Matthew Macy (mmacy@mattmacy.io) * Copyright (c) 2017 Mellanox Technologies, Ltd. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef __i386__ DEFINE_IDR(mtrr_idr); static MALLOC_DEFINE(M_LKMTRR, "idr", "Linux MTRR compat"); extern int pat_works; #endif void si_meminfo(struct sysinfo *si) { si->totalram = physmem; si->freeram = vm_free_count(); si->totalhigh = 0; si->freehigh = 0; si->mem_unit = PAGE_SIZE; } void * linux_page_address(struct page *page) { if (page->object != kernel_object) { return (PMAP_HAS_DMAP ? ((void *)(uintptr_t)PHYS_TO_DMAP(page_to_phys(page))) : NULL); } return ((void *)(uintptr_t)(VM_MIN_KERNEL_ADDRESS + IDX_TO_OFF(page->pindex))); } struct page * linux_alloc_pages(gfp_t flags, unsigned int order) { struct page *page; if (PMAP_HAS_DMAP) { unsigned long npages = 1UL << order; int req = VM_ALLOC_WIRED; if ((flags & M_ZERO) != 0) req |= VM_ALLOC_ZERO; if (order == 0 && (flags & GFP_DMA32) == 0) { page = vm_page_alloc_noobj(req); if (page == NULL) return (NULL); } else { vm_paddr_t pmax = (flags & GFP_DMA32) ? BUS_SPACE_MAXADDR_32BIT : BUS_SPACE_MAXADDR; if ((flags & __GFP_NORETRY) != 0) req |= VM_ALLOC_NORECLAIM; retry: page = vm_page_alloc_noobj_contig(req, npages, 0, pmax, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); if (page == NULL) { if ((flags & (M_WAITOK | __GFP_NORETRY)) == M_WAITOK) { - if (!vm_page_reclaim_contig(req, - npages, 0, pmax, PAGE_SIZE, 0)) { + int err = vm_page_reclaim_contig(req, + npages, 0, pmax, PAGE_SIZE, 0); + if (err == ENOMEM) vm_wait(NULL); - } + else if (err != 0) + return (NULL); flags &= ~M_WAITOK; goto retry; } return (NULL); } } } else { vm_offset_t vaddr; vaddr = linux_alloc_kmem(flags, order); if (vaddr == 0) return (NULL); page = virt_to_page((void *)vaddr); KASSERT(vaddr == (vm_offset_t)page_address(page), ("Page address mismatch")); } return (page); } static void _linux_free_kmem(vm_offset_t addr, unsigned int order) { size_t size = ((size_t)PAGE_SIZE) << order; kmem_free((void *)addr, size); } void linux_free_pages(struct page *page, unsigned int order) { if (PMAP_HAS_DMAP) { unsigned long npages = 1UL << order; unsigned long x; for (x = 0; x != npages; x++) { vm_page_t pgo = page + x; if (vm_page_unwire_noq(pgo)) vm_page_free(pgo); } } else { vm_offset_t vaddr; vaddr = (vm_offset_t)page_address(page); _linux_free_kmem(vaddr, order); } } vm_offset_t linux_alloc_kmem(gfp_t flags, unsigned int order) { size_t size = ((size_t)PAGE_SIZE) << order; void *addr; addr = kmem_alloc_contig(size, flags & GFP_NATIVE_MASK, 0, ((flags & GFP_DMA32) == 0) ? -1UL : BUS_SPACE_MAXADDR_32BIT, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); return ((vm_offset_t)addr); } void linux_free_kmem(vm_offset_t addr, unsigned int order) { KASSERT((addr & ~PAGE_MASK) == 0, ("%s: addr %p is not page aligned", __func__, (void *)addr)); if (addr >= VM_MIN_KERNEL_ADDRESS && addr < VM_MAX_KERNEL_ADDRESS) { _linux_free_kmem(addr, order); } else { vm_page_t page; page = PHYS_TO_VM_PAGE(DMAP_TO_PHYS(addr)); linux_free_pages(page, order); } } static int linux_get_user_pages_internal(vm_map_t map, unsigned long start, int nr_pages, int write, struct page **pages) { vm_prot_t prot; size_t len; int count; prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ; len = ptoa((vm_offset_t)nr_pages); count = vm_fault_quick_hold_pages(map, start, len, prot, pages, nr_pages); return (count == -1 ? -EFAULT : nr_pages); } int __get_user_pages_fast(unsigned long start, int nr_pages, int write, struct page **pages) { vm_map_t map; vm_page_t *mp; vm_offset_t va; vm_offset_t end; vm_prot_t prot; int count; if (nr_pages == 0 || in_interrupt()) return (0); MPASS(pages != NULL); map = &curthread->td_proc->p_vmspace->vm_map; end = start + ptoa((vm_offset_t)nr_pages); if (!vm_map_range_valid(map, start, end)) return (-EINVAL); prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ; for (count = 0, mp = pages, va = start; va < end; mp++, va += PAGE_SIZE, count++) { *mp = pmap_extract_and_hold(map->pmap, va, prot); if (*mp == NULL) break; if ((prot & VM_PROT_WRITE) != 0 && (*mp)->dirty != VM_PAGE_BITS_ALL) { /* * Explicitly dirty the physical page. Otherwise, the * caller's changes may go unnoticed because they are * performed through an unmanaged mapping or by a DMA * operation. * * The object lock is not held here. * See vm_page_clear_dirty_mask(). */ vm_page_dirty(*mp); } } return (count); } long get_user_pages_remote(struct task_struct *task, struct mm_struct *mm, unsigned long start, unsigned long nr_pages, unsigned int gup_flags, struct page **pages, struct vm_area_struct **vmas) { vm_map_t map; map = &task->task_thread->td_proc->p_vmspace->vm_map; return (linux_get_user_pages_internal(map, start, nr_pages, !!(gup_flags & FOLL_WRITE), pages)); } long get_user_pages(unsigned long start, unsigned long nr_pages, unsigned int gup_flags, struct page **pages, struct vm_area_struct **vmas) { vm_map_t map; map = &curthread->td_proc->p_vmspace->vm_map; return (linux_get_user_pages_internal(map, start, nr_pages, !!(gup_flags & FOLL_WRITE), pages)); } int is_vmalloc_addr(const void *addr) { return (vtoslab((vm_offset_t)addr & ~UMA_SLAB_MASK) != NULL); } vm_fault_t lkpi_vmf_insert_pfn_prot_locked(struct vm_area_struct *vma, unsigned long addr, unsigned long pfn, pgprot_t prot) { vm_object_t vm_obj = vma->vm_obj; vm_object_t tmp_obj; vm_page_t page; vm_pindex_t pindex; VM_OBJECT_ASSERT_WLOCKED(vm_obj); pindex = OFF_TO_IDX(addr - vma->vm_start); if (vma->vm_pfn_count == 0) vma->vm_pfn_first = pindex; MPASS(pindex <= OFF_TO_IDX(vma->vm_end)); retry: page = vm_page_grab(vm_obj, pindex, VM_ALLOC_NOCREAT); if (page == NULL) { page = PHYS_TO_VM_PAGE(IDX_TO_OFF(pfn)); if (!vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) goto retry; if (page->object != NULL) { tmp_obj = page->object; vm_page_xunbusy(page); VM_OBJECT_WUNLOCK(vm_obj); VM_OBJECT_WLOCK(tmp_obj); if (page->object == tmp_obj && vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) { KASSERT(page->object == tmp_obj, ("page has changed identity")); KASSERT((page->oflags & VPO_UNMANAGED) == 0, ("page does not belong to shmem")); vm_pager_page_unswapped(page); if (pmap_page_is_mapped(page)) { vm_page_xunbusy(page); VM_OBJECT_WUNLOCK(tmp_obj); printf("%s: page rename failed: page " "is mapped\n", __func__); VM_OBJECT_WLOCK(vm_obj); return (VM_FAULT_NOPAGE); } vm_page_remove(page); } VM_OBJECT_WUNLOCK(tmp_obj); VM_OBJECT_WLOCK(vm_obj); goto retry; } if (vm_page_insert(page, vm_obj, pindex)) { vm_page_xunbusy(page); return (VM_FAULT_OOM); } vm_page_valid(page); } pmap_page_set_memattr(page, pgprot2cachemode(prot)); vma->vm_pfn_count++; return (VM_FAULT_NOPAGE); } int lkpi_remap_pfn_range(struct vm_area_struct *vma, unsigned long start_addr, unsigned long start_pfn, unsigned long size, pgprot_t prot) { vm_object_t vm_obj; unsigned long addr, pfn; int err = 0; vm_obj = vma->vm_obj; VM_OBJECT_WLOCK(vm_obj); for (addr = start_addr, pfn = start_pfn; addr < start_addr + size; addr += PAGE_SIZE) { vm_fault_t ret; retry: ret = lkpi_vmf_insert_pfn_prot_locked(vma, addr, pfn, prot); if ((ret & VM_FAULT_OOM) != 0) { VM_OBJECT_WUNLOCK(vm_obj); vm_wait(NULL); VM_OBJECT_WLOCK(vm_obj); goto retry; } if ((ret & VM_FAULT_ERROR) != 0) { err = -EFAULT; break; } pfn++; } VM_OBJECT_WUNLOCK(vm_obj); if (unlikely(err)) { zap_vma_ptes(vma, start_addr, (pfn - start_pfn) << PAGE_SHIFT); return (err); } return (0); } int lkpi_io_mapping_map_user(struct io_mapping *iomap, struct vm_area_struct *vma, unsigned long addr, unsigned long pfn, unsigned long size) { pgprot_t prot; int ret; prot = cachemode2protval(iomap->attr); ret = lkpi_remap_pfn_range(vma, addr, pfn, size, prot); return (ret); } /* * Although FreeBSD version of unmap_mapping_range has semantics and types of * parameters compatible with Linux version, the values passed in are different * @obj should match to vm_private_data field of vm_area_struct returned by * mmap file operation handler, see linux_file_mmap_single() sources * @holelen should match to size of area to be munmapped. */ void lkpi_unmap_mapping_range(void *obj, loff_t const holebegin __unused, loff_t const holelen, int even_cows __unused) { vm_object_t devobj; vm_page_t page; int i, page_count; devobj = cdev_pager_lookup(obj); if (devobj != NULL) { page_count = OFF_TO_IDX(holelen); VM_OBJECT_WLOCK(devobj); retry: for (i = 0; i < page_count; i++) { page = vm_page_lookup(devobj, i); if (page == NULL) continue; if (!vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) goto retry; cdev_pager_free_page(devobj, page); } VM_OBJECT_WUNLOCK(devobj); vm_object_deallocate(devobj); } } int lkpi_arch_phys_wc_add(unsigned long base, unsigned long size) { #ifdef __i386__ struct mem_range_desc *mrdesc; int error, id, act; /* If PAT is available, do nothing */ if (pat_works) return (0); mrdesc = malloc(sizeof(*mrdesc), M_LKMTRR, M_WAITOK); mrdesc->mr_base = base; mrdesc->mr_len = size; mrdesc->mr_flags = MDF_WRITECOMBINE; strlcpy(mrdesc->mr_owner, "drm", sizeof(mrdesc->mr_owner)); act = MEMRANGE_SET_UPDATE; error = mem_range_attr_set(mrdesc, &act); if (error == 0) { error = idr_get_new(&mtrr_idr, mrdesc, &id); MPASS(idr_find(&mtrr_idr, id) == mrdesc); if (error != 0) { act = MEMRANGE_SET_REMOVE; mem_range_attr_set(mrdesc, &act); } } if (error != 0) { free(mrdesc, M_LKMTRR); pr_warn( "Failed to add WC MTRR for [%p-%p]: %d; " "performance may suffer\n", (void *)base, (void *)(base + size - 1), error); } else pr_warn("Successfully added WC MTRR for [%p-%p]\n", (void *)base, (void *)(base + size - 1)); return (error != 0 ? -error : id + __MTRR_ID_BASE); #else return (0); #endif } void lkpi_arch_phys_wc_del(int reg) { #ifdef __i386__ struct mem_range_desc *mrdesc; int act; /* Check if arch_phys_wc_add() failed. */ if (reg < __MTRR_ID_BASE) return; mrdesc = idr_find(&mtrr_idr, reg - __MTRR_ID_BASE); MPASS(mrdesc != NULL); idr_remove(&mtrr_idr, reg - __MTRR_ID_BASE); act = MEMRANGE_SET_REMOVE; mem_range_attr_set(mrdesc, &act); free(mrdesc, M_LKMTRR); #endif } /* * This is a highly simplified version of the Linux page_frag_cache. * We only support up-to 1 single page as fragment size and we will * always return a full page. This may be wasteful on small objects * but the only known consumer (mt76) is either asking for a half-page * or a full page. If this was to become a problem we can implement * a more elaborate version. */ void * linuxkpi_page_frag_alloc(struct page_frag_cache *pfc, size_t fragsz, gfp_t gfp) { vm_page_t pages; if (fragsz == 0) return (NULL); KASSERT(fragsz <= PAGE_SIZE, ("%s: fragsz %zu > PAGE_SIZE not yet " "supported", __func__, fragsz)); pages = alloc_pages(gfp, flsl(howmany(fragsz, PAGE_SIZE) - 1)); if (pages == NULL) return (NULL); pfc->va = linux_page_address(pages); /* Passed in as "count" to __page_frag_cache_drain(). Unused by us. */ pfc->pagecnt_bias = 0; return (pfc->va); } void linuxkpi_page_frag_free(void *addr) { vm_page_t page; page = virt_to_page(addr); linux_free_pages(page, 0); } void linuxkpi__page_frag_cache_drain(struct page *page, size_t count __unused) { linux_free_pages(page, 0); } diff --git a/sys/dev/drm2/ttm/ttm_bo.c b/sys/dev/drm2/ttm/ttm_bo.c index f60cb4f37535..b57d961338f4 100644 --- a/sys/dev/drm2/ttm/ttm_bo.c +++ b/sys/dev/drm2/ttm/ttm_bo.c @@ -1,1892 +1,1892 @@ /************************************************************************** * * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the * "Software"), to deal in the Software without restriction, including * without limitation the rights to use, copy, modify, merge, publish, * distribute, sub license, and/or sell copies of the Software, and to * permit persons to whom the Software is furnished to do so, subject to * the following conditions: * * The above copyright notice and this permission notice (including the * next paragraph) shall be included in all copies or substantial portions * of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE * USE OR OTHER DEALINGS IN THE SOFTWARE. * **************************************************************************/ /* * Authors: Thomas Hellstrom */ #include #include #include #include #include #include #define TTM_ASSERT_LOCKED(param) #define TTM_DEBUG(fmt, arg...) #define TTM_BO_HASH_ORDER 13 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo); static int ttm_bo_swapout(struct ttm_mem_shrink *shrink); static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob); MALLOC_DEFINE(M_TTM_BO, "ttm_bo", "TTM Buffer Objects"); static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type) { int i; for (i = 0; i <= TTM_PL_PRIV5; i++) if (flags & (1 << i)) { *mem_type = i; return 0; } return -EINVAL; } static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) { struct ttm_mem_type_manager *man = &bdev->man[mem_type]; printf(" has_type: %d\n", man->has_type); printf(" use_type: %d\n", man->use_type); printf(" flags: 0x%08X\n", man->flags); printf(" gpu_offset: 0x%08lX\n", man->gpu_offset); printf(" size: %ju\n", (uintmax_t)man->size); printf(" available_caching: 0x%08X\n", man->available_caching); printf(" default_caching: 0x%08X\n", man->default_caching); if (mem_type != TTM_PL_SYSTEM) (*man->func->debug)(man, TTM_PFX); } static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, struct ttm_placement *placement) { int i, ret, mem_type; printf("No space for %p (%lu pages, %luK, %luM)\n", bo, bo->mem.num_pages, bo->mem.size >> 10, bo->mem.size >> 20); for (i = 0; i < placement->num_placement; i++) { ret = ttm_mem_type_from_flags(placement->placement[i], &mem_type); if (ret) return; printf(" placement[%d]=0x%08X (%d)\n", i, placement->placement[i], mem_type); ttm_mem_type_debug(bo->bdev, mem_type); } } #if 0 static ssize_t ttm_bo_global_show(struct ttm_bo_global *glob, char *buffer) { return snprintf(buffer, PAGE_SIZE, "%lu\n", (unsigned long) atomic_read(&glob->bo_count)); } #endif static inline uint32_t ttm_bo_type_flags(unsigned type) { return 1 << (type); } static void ttm_bo_release_list(struct ttm_buffer_object *bo) { struct ttm_bo_device *bdev = bo->bdev; size_t acc_size = bo->acc_size; MPASS(atomic_read(&bo->list_kref) == 0); MPASS(atomic_read(&bo->kref) == 0); MPASS(atomic_read(&bo->cpu_writers) == 0); MPASS(bo->sync_obj == NULL); MPASS(bo->mem.mm_node == NULL); MPASS(list_empty(&bo->lru)); MPASS(list_empty(&bo->ddestroy)); if (bo->ttm) ttm_tt_destroy(bo->ttm); atomic_dec(&bo->glob->bo_count); if (bo->destroy) bo->destroy(bo); else { free(bo, M_TTM_BO); } ttm_mem_global_free(bdev->glob->mem_glob, acc_size); } static int ttm_bo_wait_unreserved_locked(struct ttm_buffer_object *bo, bool interruptible) { const char *wmsg; int flags, ret; ret = 0; if (interruptible) { flags = PCATCH; wmsg = "ttbowi"; } else { flags = 0; wmsg = "ttbowu"; } while (ttm_bo_is_reserved(bo)) { ret = -msleep(bo, &bo->glob->lru_lock, flags, wmsg, 0); if (ret == -EINTR || ret == -ERESTART) ret = -ERESTARTSYS; if (ret != 0) break; } return (ret); } void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_mem_type_manager *man; MPASS(ttm_bo_is_reserved(bo)); if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { MPASS(list_empty(&bo->lru)); man = &bdev->man[bo->mem.mem_type]; list_add_tail(&bo->lru, &man->lru); refcount_acquire(&bo->list_kref); if (bo->ttm != NULL) { list_add_tail(&bo->swap, &bo->glob->swap_lru); refcount_acquire(&bo->list_kref); } } } int ttm_bo_del_from_lru(struct ttm_buffer_object *bo) { int put_count = 0; if (!list_empty(&bo->swap)) { list_del_init(&bo->swap); ++put_count; } if (!list_empty(&bo->lru)) { list_del_init(&bo->lru); ++put_count; } /* * TODO: Add a driver hook to delete from * driver-specific LRU's here. */ return put_count; } int ttm_bo_reserve_nolru(struct ttm_buffer_object *bo, bool interruptible, bool no_wait, bool use_sequence, uint32_t sequence) { int ret; while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { /** * Deadlock avoidance for multi-bo reserving. */ if (use_sequence && bo->seq_valid) { /** * We've already reserved this one. */ if (unlikely(sequence == bo->val_seq)) return -EDEADLK; /** * Already reserved by a thread that will not back * off for us. We need to back off. */ if (unlikely(sequence - bo->val_seq < (1U << 31))) return -EAGAIN; } if (no_wait) return -EBUSY; ret = ttm_bo_wait_unreserved_locked(bo, interruptible); if (unlikely(ret)) return ret; } if (use_sequence) { bool wake_up = false; /** * Wake up waiters that may need to recheck for deadlock, * if we decreased the sequence number. */ if (unlikely((bo->val_seq - sequence < (1U << 31)) || !bo->seq_valid)) wake_up = true; /* * In the worst case with memory ordering these values can be * seen in the wrong order. However since we call wake_up_all * in that case, this will hopefully not pose a problem, * and the worst case would only cause someone to accidentally * hit -EAGAIN in ttm_bo_reserve when they see old value of * val_seq. However this would only happen if seq_valid was * written before val_seq was, and just means some slightly * increased cpu usage */ bo->val_seq = sequence; bo->seq_valid = true; if (wake_up) wakeup(bo); } else { bo->seq_valid = false; } return 0; } void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count, bool never_free) { u_int old; old = atomic_fetchadd_int(&bo->list_kref, -count); if (old <= count) { if (never_free) panic("ttm_bo_ref_buf"); ttm_bo_release_list(bo); } } int ttm_bo_reserve(struct ttm_buffer_object *bo, bool interruptible, bool no_wait, bool use_sequence, uint32_t sequence) { struct ttm_bo_global *glob = bo->glob; int put_count = 0; int ret; mtx_lock(&bo->glob->lru_lock); ret = ttm_bo_reserve_nolru(bo, interruptible, no_wait, use_sequence, sequence); if (likely(ret == 0)) { put_count = ttm_bo_del_from_lru(bo); mtx_unlock(&glob->lru_lock); ttm_bo_list_ref_sub(bo, put_count, true); } else mtx_unlock(&bo->glob->lru_lock); return ret; } int ttm_bo_reserve_slowpath_nolru(struct ttm_buffer_object *bo, bool interruptible, uint32_t sequence) { bool wake_up = false; int ret; while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { if (bo->seq_valid && sequence == bo->val_seq) { DRM_ERROR( "%s: bo->seq_valid && sequence == bo->val_seq", __func__); } ret = ttm_bo_wait_unreserved_locked(bo, interruptible); if (unlikely(ret)) return ret; } if ((bo->val_seq - sequence < (1U << 31)) || !bo->seq_valid) wake_up = true; /** * Wake up waiters that may need to recheck for deadlock, * if we decreased the sequence number. */ bo->val_seq = sequence; bo->seq_valid = true; if (wake_up) wakeup(bo); return 0; } int ttm_bo_reserve_slowpath(struct ttm_buffer_object *bo, bool interruptible, uint32_t sequence) { struct ttm_bo_global *glob = bo->glob; int put_count, ret; mtx_lock(&glob->lru_lock); ret = ttm_bo_reserve_slowpath_nolru(bo, interruptible, sequence); if (likely(!ret)) { put_count = ttm_bo_del_from_lru(bo); mtx_unlock(&glob->lru_lock); ttm_bo_list_ref_sub(bo, put_count, true); } else mtx_unlock(&glob->lru_lock); return ret; } void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo) { ttm_bo_add_to_lru(bo); atomic_set(&bo->reserved, 0); wakeup(bo); } void ttm_bo_unreserve(struct ttm_buffer_object *bo) { struct ttm_bo_global *glob = bo->glob; mtx_lock(&glob->lru_lock); ttm_bo_unreserve_locked(bo); mtx_unlock(&glob->lru_lock); } /* * Call bo->mutex locked. */ static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_bo_global *glob = bo->glob; int ret = 0; uint32_t page_flags = 0; TTM_ASSERT_LOCKED(&bo->mutex); bo->ttm = NULL; if (bdev->need_dma32) page_flags |= TTM_PAGE_FLAG_DMA32; switch (bo->type) { case ttm_bo_type_device: if (zero_alloc) page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; case ttm_bo_type_kernel: bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, page_flags, glob->dummy_read_page); if (unlikely(bo->ttm == NULL)) ret = -ENOMEM; break; case ttm_bo_type_sg: bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, page_flags | TTM_PAGE_FLAG_SG, glob->dummy_read_page); if (unlikely(bo->ttm == NULL)) { ret = -ENOMEM; break; } bo->ttm->sg = bo->sg; break; default: printf("[TTM] Illegal buffer object type\n"); ret = -EINVAL; break; } return ret; } static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem, bool evict, bool interruptible, bool no_wait_gpu) { struct ttm_bo_device *bdev = bo->bdev; bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; int ret = 0; if (old_is_pci || new_is_pci || ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { ret = ttm_mem_io_lock(old_man, true); if (unlikely(ret != 0)) goto out_err; ttm_bo_unmap_virtual_locked(bo); ttm_mem_io_unlock(old_man); } /* * Create and bind a ttm if required. */ if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { if (bo->ttm == NULL) { bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); ret = ttm_bo_add_ttm(bo, zero); if (ret) goto out_err; } ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); if (ret) goto out_err; if (mem->mem_type != TTM_PL_SYSTEM) { ret = ttm_tt_bind(bo->ttm, mem); if (ret) goto out_err; } if (bo->mem.mem_type == TTM_PL_SYSTEM) { if (bdev->driver->move_notify) bdev->driver->move_notify(bo, mem); bo->mem = *mem; mem->mm_node = NULL; goto moved; } } if (bdev->driver->move_notify) bdev->driver->move_notify(bo, mem); if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem); else if (bdev->driver->move) ret = bdev->driver->move(bo, evict, interruptible, no_wait_gpu, mem); else ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem); if (ret) { if (bdev->driver->move_notify) { struct ttm_mem_reg tmp_mem = *mem; *mem = bo->mem; bo->mem = tmp_mem; bdev->driver->move_notify(bo, mem); bo->mem = *mem; *mem = tmp_mem; } goto out_err; } moved: if (bo->evicted) { ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); if (ret) printf("[TTM] Can not flush read caches\n"); bo->evicted = false; } if (bo->mem.mm_node) { bo->offset = (bo->mem.start << PAGE_SHIFT) + bdev->man[bo->mem.mem_type].gpu_offset; bo->cur_placement = bo->mem.placement; } else bo->offset = 0; return 0; out_err: new_man = &bdev->man[bo->mem.mem_type]; if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) { ttm_tt_unbind(bo->ttm); ttm_tt_destroy(bo->ttm); bo->ttm = NULL; } return ret; } /** * Call bo::reserved. * Will release GPU memory type usage on destruction. * This is the place to put in driver specific hooks to release * driver private resources. * Will release the bo::reserved lock. */ static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) { if (bo->bdev->driver->move_notify) bo->bdev->driver->move_notify(bo, NULL); if (bo->ttm) { ttm_tt_unbind(bo->ttm); ttm_tt_destroy(bo->ttm); bo->ttm = NULL; } ttm_bo_mem_put(bo, &bo->mem); atomic_set(&bo->reserved, 0); wakeup(&bo); /* * Since the final reference to this bo may not be dropped by * the current task we have to put a memory barrier here to make * sure the changes done in this function are always visible. * * This function only needs protection against the final kref_put. */ mb(); } static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_bo_global *glob = bo->glob; struct ttm_bo_driver *driver = bdev->driver; void *sync_obj = NULL; int put_count; int ret; mtx_lock(&glob->lru_lock); ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); mtx_lock(&bdev->fence_lock); (void) ttm_bo_wait(bo, false, false, true); if (!ret && !bo->sync_obj) { mtx_unlock(&bdev->fence_lock); put_count = ttm_bo_del_from_lru(bo); mtx_unlock(&glob->lru_lock); ttm_bo_cleanup_memtype_use(bo); ttm_bo_list_ref_sub(bo, put_count, true); return; } if (bo->sync_obj) sync_obj = driver->sync_obj_ref(bo->sync_obj); mtx_unlock(&bdev->fence_lock); if (!ret) { atomic_set(&bo->reserved, 0); wakeup(bo); } refcount_acquire(&bo->list_kref); list_add_tail(&bo->ddestroy, &bdev->ddestroy); mtx_unlock(&glob->lru_lock); if (sync_obj) { driver->sync_obj_flush(sync_obj); driver->sync_obj_unref(&sync_obj); } taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, ((hz / 100) < 1) ? 1 : hz / 100); } /** * function ttm_bo_cleanup_refs_and_unlock * If bo idle, remove from delayed- and lru lists, and unref. * If not idle, do nothing. * * Must be called with lru_lock and reservation held, this function * will drop both before returning. * * @interruptible Any sleeps should occur interruptibly. * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. */ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo, bool interruptible, bool no_wait_gpu) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_bo_driver *driver = bdev->driver; struct ttm_bo_global *glob = bo->glob; int put_count; int ret; mtx_lock(&bdev->fence_lock); ret = ttm_bo_wait(bo, false, false, true); if (ret && !no_wait_gpu) { void *sync_obj; /* * Take a reference to the fence and unreserve, * at this point the buffer should be dead, so * no new sync objects can be attached. */ sync_obj = driver->sync_obj_ref(bo->sync_obj); mtx_unlock(&bdev->fence_lock); atomic_set(&bo->reserved, 0); wakeup(bo); mtx_unlock(&glob->lru_lock); ret = driver->sync_obj_wait(sync_obj, false, interruptible); driver->sync_obj_unref(&sync_obj); if (ret) return ret; /* * remove sync_obj with ttm_bo_wait, the wait should be * finished, and no new wait object should have been added. */ mtx_lock(&bdev->fence_lock); ret = ttm_bo_wait(bo, false, false, true); mtx_unlock(&bdev->fence_lock); if (ret) return ret; mtx_lock(&glob->lru_lock); ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); /* * We raced, and lost, someone else holds the reservation now, * and is probably busy in ttm_bo_cleanup_memtype_use. * * Even if it's not the case, because we finished waiting any * delayed destruction would succeed, so just return success * here. */ if (ret) { mtx_unlock(&glob->lru_lock); return 0; } } else mtx_unlock(&bdev->fence_lock); if (ret || unlikely(list_empty(&bo->ddestroy))) { atomic_set(&bo->reserved, 0); wakeup(bo); mtx_unlock(&glob->lru_lock); return ret; } put_count = ttm_bo_del_from_lru(bo); list_del_init(&bo->ddestroy); ++put_count; mtx_unlock(&glob->lru_lock); ttm_bo_cleanup_memtype_use(bo); ttm_bo_list_ref_sub(bo, put_count, true); return 0; } /** * Traverse the delayed list, and call ttm_bo_cleanup_refs on all * encountered buffers. */ static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) { struct ttm_bo_global *glob = bdev->glob; struct ttm_buffer_object *entry = NULL; int ret = 0; mtx_lock(&glob->lru_lock); if (list_empty(&bdev->ddestroy)) goto out_unlock; entry = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object, ddestroy); refcount_acquire(&entry->list_kref); for (;;) { struct ttm_buffer_object *nentry = NULL; if (entry->ddestroy.next != &bdev->ddestroy) { nentry = list_first_entry(&entry->ddestroy, struct ttm_buffer_object, ddestroy); refcount_acquire(&nentry->list_kref); } ret = ttm_bo_reserve_nolru(entry, false, true, false, 0); if (remove_all && ret) { ret = ttm_bo_reserve_nolru(entry, false, false, false, 0); } if (!ret) ret = ttm_bo_cleanup_refs_and_unlock(entry, false, !remove_all); else mtx_unlock(&glob->lru_lock); if (refcount_release(&entry->list_kref)) ttm_bo_release_list(entry); entry = nentry; if (ret || !entry) goto out; mtx_lock(&glob->lru_lock); if (list_empty(&entry->ddestroy)) break; } out_unlock: mtx_unlock(&glob->lru_lock); out: if (entry && refcount_release(&entry->list_kref)) ttm_bo_release_list(entry); return ret; } static void ttm_bo_delayed_workqueue(void *arg, int pending __unused) { struct ttm_bo_device *bdev = arg; if (ttm_bo_delayed_delete(bdev, false)) { taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, ((hz / 100) < 1) ? 1 : hz / 100); } } static void ttm_bo_release(struct ttm_buffer_object *bo) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; rw_wlock(&bdev->vm_lock); if (likely(bo->vm_node != NULL)) { RB_REMOVE(ttm_bo_device_buffer_objects, &bdev->addr_space_rb, bo); drm_mm_put_block(bo->vm_node); bo->vm_node = NULL; } rw_wunlock(&bdev->vm_lock); ttm_mem_io_lock(man, false); ttm_mem_io_free_vm(bo); ttm_mem_io_unlock(man); ttm_bo_cleanup_refs_or_queue(bo); if (refcount_release(&bo->list_kref)) ttm_bo_release_list(bo); } void ttm_bo_unref(struct ttm_buffer_object **p_bo) { struct ttm_buffer_object *bo = *p_bo; *p_bo = NULL; if (refcount_release(&bo->kref)) ttm_bo_release(bo); } int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) { int pending; if (taskqueue_cancel_timeout(taskqueue_thread, &bdev->wq, &pending)) taskqueue_drain_timeout(taskqueue_thread, &bdev->wq); return (pending); } void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) { if (resched) { taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, ((hz / 100) < 1) ? 1 : hz / 100); } } static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible, bool no_wait_gpu) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_mem_reg evict_mem; struct ttm_placement placement; int ret = 0; mtx_lock(&bdev->fence_lock); ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); mtx_unlock(&bdev->fence_lock); if (unlikely(ret != 0)) { if (ret != -ERESTARTSYS) { printf("[TTM] Failed to expire sync object before buffer eviction\n"); } goto out; } MPASS(ttm_bo_is_reserved(bo)); evict_mem = bo->mem; evict_mem.mm_node = NULL; evict_mem.bus.io_reserved_vm = false; evict_mem.bus.io_reserved_count = 0; placement.fpfn = 0; placement.lpfn = 0; placement.num_placement = 0; placement.num_busy_placement = 0; bdev->driver->evict_flags(bo, &placement); ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible, no_wait_gpu); if (ret) { if (ret != -ERESTARTSYS) { printf("[TTM] Failed to find memory space for buffer 0x%p eviction\n", bo); ttm_bo_mem_space_debug(bo, &placement); } goto out; } ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible, no_wait_gpu); if (ret) { if (ret != -ERESTARTSYS) printf("[TTM] Buffer eviction failed\n"); ttm_bo_mem_put(bo, &evict_mem); goto out; } bo->evicted = true; out: return ret; } static int ttm_mem_evict_first(struct ttm_bo_device *bdev, uint32_t mem_type, bool interruptible, bool no_wait_gpu) { struct ttm_bo_global *glob = bdev->glob; struct ttm_mem_type_manager *man = &bdev->man[mem_type]; struct ttm_buffer_object *bo; int ret = -EBUSY, put_count; mtx_lock(&glob->lru_lock); list_for_each_entry(bo, &man->lru, lru) { ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); if (!ret) break; } if (ret) { mtx_unlock(&glob->lru_lock); return ret; } refcount_acquire(&bo->list_kref); if (!list_empty(&bo->ddestroy)) { ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible, no_wait_gpu); if (refcount_release(&bo->list_kref)) ttm_bo_release_list(bo); return ret; } put_count = ttm_bo_del_from_lru(bo); mtx_unlock(&glob->lru_lock); MPASS(ret == 0); ttm_bo_list_ref_sub(bo, put_count, true); ret = ttm_bo_evict(bo, interruptible, no_wait_gpu); ttm_bo_unreserve(bo); if (refcount_release(&bo->list_kref)) ttm_bo_release_list(bo); return ret; } void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) { struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; if (mem->mm_node) (*man->func->put_node)(man, mem); } /** * Repeatedly evict memory from the LRU for @mem_type until we create enough * space, or we've evicted everything and there isn't enough space. */ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, uint32_t mem_type, struct ttm_placement *placement, struct ttm_mem_reg *mem, bool interruptible, bool no_wait_gpu) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_mem_type_manager *man = &bdev->man[mem_type]; int ret; do { ret = (*man->func->get_node)(man, bo, placement, mem); if (unlikely(ret != 0)) return ret; if (mem->mm_node) break; ret = ttm_mem_evict_first(bdev, mem_type, interruptible, no_wait_gpu); if (unlikely(ret != 0)) return ret; } while (1); if (mem->mm_node == NULL) return -ENOMEM; mem->mem_type = mem_type; return 0; } static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, uint32_t cur_placement, uint32_t proposed_placement) { uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; /** * Keep current caching if possible. */ if ((cur_placement & caching) != 0) result |= (cur_placement & caching); else if ((man->default_caching & caching) != 0) result |= man->default_caching; else if ((TTM_PL_FLAG_CACHED & caching) != 0) result |= TTM_PL_FLAG_CACHED; else if ((TTM_PL_FLAG_WC & caching) != 0) result |= TTM_PL_FLAG_WC; else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) result |= TTM_PL_FLAG_UNCACHED; return result; } static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, uint32_t mem_type, uint32_t proposed_placement, uint32_t *masked_placement) { uint32_t cur_flags = ttm_bo_type_flags(mem_type); if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0) return false; if ((proposed_placement & man->available_caching) == 0) return false; cur_flags |= (proposed_placement & man->available_caching); *masked_placement = cur_flags; return true; } /** * Creates space for memory region @mem according to its type. * * This function first searches for free space in compatible memory types in * the priority order defined by the driver. If free space isn't found, then * ttm_bo_mem_force_space is attempted in priority order to evict and find * space. */ int ttm_bo_mem_space(struct ttm_buffer_object *bo, struct ttm_placement *placement, struct ttm_mem_reg *mem, bool interruptible, bool no_wait_gpu) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_mem_type_manager *man; uint32_t mem_type = TTM_PL_SYSTEM; uint32_t cur_flags = 0; bool type_found = false; bool type_ok = false; bool has_erestartsys = false; int i, ret; mem->mm_node = NULL; for (i = 0; i < placement->num_placement; ++i) { ret = ttm_mem_type_from_flags(placement->placement[i], &mem_type); if (ret) return ret; man = &bdev->man[mem_type]; type_ok = ttm_bo_mt_compatible(man, mem_type, placement->placement[i], &cur_flags); if (!type_ok) continue; cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags); /* * Use the access and other non-mapping-related flag bits from * the memory placement flags to the current flags */ ttm_flag_masked(&cur_flags, placement->placement[i], ~TTM_PL_MASK_MEMTYPE); if (mem_type == TTM_PL_SYSTEM) break; if (man->has_type && man->use_type) { type_found = true; ret = (*man->func->get_node)(man, bo, placement, mem); if (unlikely(ret)) return ret; } if (mem->mm_node) break; } if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { mem->mem_type = mem_type; mem->placement = cur_flags; return 0; } if (!type_found) return -EINVAL; for (i = 0; i < placement->num_busy_placement; ++i) { ret = ttm_mem_type_from_flags(placement->busy_placement[i], &mem_type); if (ret) return ret; man = &bdev->man[mem_type]; if (!man->has_type) continue; if (!ttm_bo_mt_compatible(man, mem_type, placement->busy_placement[i], &cur_flags)) continue; cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags); /* * Use the access and other non-mapping-related flag bits from * the memory placement flags to the current flags */ ttm_flag_masked(&cur_flags, placement->busy_placement[i], ~TTM_PL_MASK_MEMTYPE); if (mem_type == TTM_PL_SYSTEM) { mem->mem_type = mem_type; mem->placement = cur_flags; mem->mm_node = NULL; return 0; } ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem, interruptible, no_wait_gpu); if (ret == 0 && mem->mm_node) { mem->placement = cur_flags; return 0; } if (ret == -ERESTARTSYS) has_erestartsys = true; } ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; return ret; } static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, struct ttm_placement *placement, bool interruptible, bool no_wait_gpu) { int ret = 0; struct ttm_mem_reg mem; struct ttm_bo_device *bdev = bo->bdev; MPASS(ttm_bo_is_reserved(bo)); /* * FIXME: It's possible to pipeline buffer moves. * Have the driver move function wait for idle when necessary, * instead of doing it here. */ mtx_lock(&bdev->fence_lock); ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); mtx_unlock(&bdev->fence_lock); if (ret) return ret; mem.num_pages = bo->num_pages; mem.size = mem.num_pages << PAGE_SHIFT; mem.page_alignment = bo->mem.page_alignment; mem.bus.io_reserved_vm = false; mem.bus.io_reserved_count = 0; /* * Determine where to move the buffer. */ ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_gpu); if (ret) goto out_unlock; ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_gpu); out_unlock: if (ret && mem.mm_node) ttm_bo_mem_put(bo, &mem); return ret; } static int ttm_bo_mem_compat(struct ttm_placement *placement, struct ttm_mem_reg *mem) { int i; if (mem->mm_node && placement->lpfn != 0 && (mem->start < placement->fpfn || mem->start + mem->num_pages > placement->lpfn)) return -1; for (i = 0; i < placement->num_placement; i++) { if ((placement->placement[i] & mem->placement & TTM_PL_MASK_CACHING) && (placement->placement[i] & mem->placement & TTM_PL_MASK_MEM)) return i; } return -1; } int ttm_bo_validate(struct ttm_buffer_object *bo, struct ttm_placement *placement, bool interruptible, bool no_wait_gpu) { int ret; MPASS(ttm_bo_is_reserved(bo)); /* Check that range is valid */ if (placement->lpfn || placement->fpfn) if (placement->fpfn > placement->lpfn || (placement->lpfn - placement->fpfn) < bo->num_pages) return -EINVAL; /* * Check whether we need to move buffer. */ ret = ttm_bo_mem_compat(placement, &bo->mem); if (ret < 0) { ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_gpu); if (ret) return ret; } else { /* * Use the access and other non-mapping-related flag bits from * the compatible memory placement flags to the active flags */ ttm_flag_masked(&bo->mem.placement, placement->placement[ret], ~TTM_PL_MASK_MEMTYPE); } /* * We might need to add a TTM. */ if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { ret = ttm_bo_add_ttm(bo, true); if (ret) return ret; } return 0; } int ttm_bo_check_placement(struct ttm_buffer_object *bo, struct ttm_placement *placement) { MPASS(!((placement->fpfn || placement->lpfn) && (bo->mem.num_pages > (placement->lpfn - placement->fpfn)))); return 0; } int ttm_bo_init(struct ttm_bo_device *bdev, struct ttm_buffer_object *bo, unsigned long size, enum ttm_bo_type type, struct ttm_placement *placement, uint32_t page_alignment, bool interruptible, struct vm_object *persistent_swap_storage, size_t acc_size, struct sg_table *sg, void (*destroy) (struct ttm_buffer_object *)) { int ret = 0; unsigned long num_pages; struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false); if (ret) { printf("[TTM] Out of kernel memory\n"); if (destroy) (*destroy)(bo); else free(bo, M_TTM_BO); return -ENOMEM; } num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; if (num_pages == 0) { printf("[TTM] Illegal buffer object size\n"); if (destroy) (*destroy)(bo); else free(bo, M_TTM_BO); ttm_mem_global_free(mem_glob, acc_size); return -EINVAL; } bo->destroy = destroy; refcount_init(&bo->kref, 1); refcount_init(&bo->list_kref, 1); atomic_set(&bo->cpu_writers, 0); atomic_set(&bo->reserved, 1); INIT_LIST_HEAD(&bo->lru); INIT_LIST_HEAD(&bo->ddestroy); INIT_LIST_HEAD(&bo->swap); INIT_LIST_HEAD(&bo->io_reserve_lru); bo->bdev = bdev; bo->glob = bdev->glob; bo->type = type; bo->num_pages = num_pages; bo->mem.size = num_pages << PAGE_SHIFT; bo->mem.mem_type = TTM_PL_SYSTEM; bo->mem.num_pages = bo->num_pages; bo->mem.mm_node = NULL; bo->mem.page_alignment = page_alignment; bo->mem.bus.io_reserved_vm = false; bo->mem.bus.io_reserved_count = 0; bo->priv_flags = 0; bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); bo->seq_valid = false; bo->persistent_swap_storage = persistent_swap_storage; bo->acc_size = acc_size; bo->sg = sg; atomic_inc(&bo->glob->bo_count); ret = ttm_bo_check_placement(bo, placement); if (unlikely(ret != 0)) goto out_err; /* * For ttm_bo_type_device buffers, allocate * address space from the device. */ if (bo->type == ttm_bo_type_device || bo->type == ttm_bo_type_sg) { ret = ttm_bo_setup_vm(bo); if (ret) goto out_err; } ret = ttm_bo_validate(bo, placement, interruptible, false); if (ret) goto out_err; ttm_bo_unreserve(bo); return 0; out_err: ttm_bo_unreserve(bo); ttm_bo_unref(&bo); return ret; } size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, unsigned long bo_size, unsigned struct_size) { unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; size_t size = 0; size += ttm_round_pot(struct_size); size += PAGE_ALIGN(npages * sizeof(void *)); size += ttm_round_pot(sizeof(struct ttm_tt)); return size; } size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, unsigned long bo_size, unsigned struct_size) { unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; size_t size = 0; size += ttm_round_pot(struct_size); size += PAGE_ALIGN(npages * sizeof(void *)); size += PAGE_ALIGN(npages * sizeof(dma_addr_t)); size += ttm_round_pot(sizeof(struct ttm_dma_tt)); return size; } int ttm_bo_create(struct ttm_bo_device *bdev, unsigned long size, enum ttm_bo_type type, struct ttm_placement *placement, uint32_t page_alignment, bool interruptible, struct vm_object *persistent_swap_storage, struct ttm_buffer_object **p_bo) { struct ttm_buffer_object *bo; size_t acc_size; int ret; bo = malloc(sizeof(*bo), M_TTM_BO, M_WAITOK | M_ZERO); acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, interruptible, persistent_swap_storage, acc_size, NULL, NULL); if (likely(ret == 0)) *p_bo = bo; return ret; } static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, unsigned mem_type, bool allow_errors) { struct ttm_mem_type_manager *man = &bdev->man[mem_type]; struct ttm_bo_global *glob = bdev->glob; int ret; /* * Can't use standard list traversal since we're unlocking. */ mtx_lock(&glob->lru_lock); while (!list_empty(&man->lru)) { mtx_unlock(&glob->lru_lock); ret = ttm_mem_evict_first(bdev, mem_type, false, false); if (ret) { if (allow_errors) { return ret; } else { printf("[TTM] Cleanup eviction failed\n"); } } mtx_lock(&glob->lru_lock); } mtx_unlock(&glob->lru_lock); return 0; } int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) { struct ttm_mem_type_manager *man; int ret = -EINVAL; if (mem_type >= TTM_NUM_MEM_TYPES) { printf("[TTM] Illegal memory type %d\n", mem_type); return ret; } man = &bdev->man[mem_type]; if (!man->has_type) { printf("[TTM] Trying to take down uninitialized memory manager type %u\n", mem_type); return ret; } man->use_type = false; man->has_type = false; ret = 0; if (mem_type > 0) { ttm_bo_force_list_clean(bdev, mem_type, false); ret = (*man->func->takedown)(man); } return ret; } int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) { struct ttm_mem_type_manager *man = &bdev->man[mem_type]; if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { printf("[TTM] Illegal memory manager memory type %u\n", mem_type); return -EINVAL; } if (!man->has_type) { printf("[TTM] Memory type %u has not been initialized\n", mem_type); return 0; } return ttm_bo_force_list_clean(bdev, mem_type, true); } int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, unsigned long p_size) { int ret = -EINVAL; struct ttm_mem_type_manager *man; MPASS(type < TTM_NUM_MEM_TYPES); man = &bdev->man[type]; MPASS(!man->has_type); man->io_reserve_fastpath = true; man->use_io_reserve_lru = false; sx_init(&man->io_reserve_mutex, "ttmman"); INIT_LIST_HEAD(&man->io_reserve_lru); ret = bdev->driver->init_mem_type(bdev, type, man); if (ret) return ret; man->bdev = bdev; ret = 0; if (type != TTM_PL_SYSTEM) { ret = (*man->func->init)(man, p_size); if (ret) return ret; } man->has_type = true; man->use_type = true; man->size = p_size; INIT_LIST_HEAD(&man->lru); return 0; } static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob) { ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink); vm_page_free(glob->dummy_read_page); } void ttm_bo_global_release(struct drm_global_reference *ref) { struct ttm_bo_global *glob = ref->object; if (refcount_release(&glob->kobj_ref)) ttm_bo_global_kobj_release(glob); } int ttm_bo_global_init(struct drm_global_reference *ref) { struct ttm_bo_global_ref *bo_ref = container_of(ref, struct ttm_bo_global_ref, ref); struct ttm_bo_global *glob = ref->object; int ret; int tries; sx_init(&glob->device_list_mutex, "ttmdlm"); mtx_init(&glob->lru_lock, "ttmlru", NULL, MTX_DEF); glob->mem_glob = bo_ref->mem_glob; tries = 0; retry: glob->dummy_read_page = vm_page_alloc_noobj_contig(0, 1, 0, VM_MAX_ADDRESS, PAGE_SIZE, 0, VM_MEMATTR_UNCACHEABLE); if (unlikely(glob->dummy_read_page == NULL)) { - if (tries < 1 && vm_page_reclaim_contig(0, 1, 0, - VM_MAX_ADDRESS, PAGE_SIZE, 0)) { + if (tries < 1 && (vm_page_reclaim_contig(0, 1, 0, + VM_MAX_ADDRESS, PAGE_SIZE, 0) == 0)) { tries++; goto retry; } ret = -ENOMEM; goto out_no_drp; } INIT_LIST_HEAD(&glob->swap_lru); INIT_LIST_HEAD(&glob->device_list); ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); if (unlikely(ret != 0)) { printf("[TTM] Could not register buffer object swapout\n"); goto out_no_shrink; } atomic_set(&glob->bo_count, 0); refcount_init(&glob->kobj_ref, 1); return (0); out_no_shrink: vm_page_free(glob->dummy_read_page); out_no_drp: free(glob, M_DRM_GLOBAL); return ret; } int ttm_bo_device_release(struct ttm_bo_device *bdev) { int ret = 0; unsigned i = TTM_NUM_MEM_TYPES; struct ttm_mem_type_manager *man; struct ttm_bo_global *glob = bdev->glob; while (i--) { man = &bdev->man[i]; if (man->has_type) { man->use_type = false; if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { ret = -EBUSY; printf("[TTM] DRM memory manager type %d is not clean\n", i); } man->has_type = false; } } sx_xlock(&glob->device_list_mutex); list_del(&bdev->device_list); sx_xunlock(&glob->device_list_mutex); if (taskqueue_cancel_timeout(taskqueue_thread, &bdev->wq, NULL)) taskqueue_drain_timeout(taskqueue_thread, &bdev->wq); while (ttm_bo_delayed_delete(bdev, true)) ; mtx_lock(&glob->lru_lock); if (list_empty(&bdev->ddestroy)) TTM_DEBUG("Delayed destroy list was clean\n"); if (list_empty(&bdev->man[0].lru)) TTM_DEBUG("Swap list was clean\n"); mtx_unlock(&glob->lru_lock); MPASS(drm_mm_clean(&bdev->addr_space_mm)); rw_wlock(&bdev->vm_lock); drm_mm_takedown(&bdev->addr_space_mm); rw_wunlock(&bdev->vm_lock); return ret; } int ttm_bo_device_init(struct ttm_bo_device *bdev, struct ttm_bo_global *glob, struct ttm_bo_driver *driver, uint64_t file_page_offset, bool need_dma32) { int ret = -EINVAL; rw_init(&bdev->vm_lock, "ttmvml"); bdev->driver = driver; memset(bdev->man, 0, sizeof(bdev->man)); /* * Initialize the system memory buffer type. * Other types need to be driver / IOCTL initialized. */ ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); if (unlikely(ret != 0)) goto out_no_sys; RB_INIT(&bdev->addr_space_rb); ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000); if (unlikely(ret != 0)) goto out_no_addr_mm; TIMEOUT_TASK_INIT(taskqueue_thread, &bdev->wq, 0, ttm_bo_delayed_workqueue, bdev); INIT_LIST_HEAD(&bdev->ddestroy); bdev->dev_mapping = NULL; bdev->glob = glob; bdev->need_dma32 = need_dma32; bdev->val_seq = 0; mtx_init(&bdev->fence_lock, "ttmfence", NULL, MTX_DEF); sx_xlock(&glob->device_list_mutex); list_add_tail(&bdev->device_list, &glob->device_list); sx_xunlock(&glob->device_list_mutex); return 0; out_no_addr_mm: ttm_bo_clean_mm(bdev, 0); out_no_sys: return ret; } /* * buffer object vm functions. */ bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) { struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { if (mem->mem_type == TTM_PL_SYSTEM) return false; if (man->flags & TTM_MEMTYPE_FLAG_CMA) return false; if (mem->placement & TTM_PL_FLAG_CACHED) return false; } return true; } void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) { ttm_bo_release_mmap(bo); ttm_mem_io_free_vm(bo); } void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) { struct ttm_bo_device *bdev = bo->bdev; struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; ttm_mem_io_lock(man, false); ttm_bo_unmap_virtual_locked(bo); ttm_mem_io_unlock(man); } static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo) { struct ttm_bo_device *bdev = bo->bdev; /* The caller acquired bdev->vm_lock. */ RB_INSERT(ttm_bo_device_buffer_objects, &bdev->addr_space_rb, bo); } /** * ttm_bo_setup_vm: * * @bo: the buffer to allocate address space for * * Allocate address space in the drm device so that applications * can mmap the buffer and access the contents. This only * applies to ttm_bo_type_device objects as others are not * placed in the drm device address space. */ static int ttm_bo_setup_vm(struct ttm_buffer_object *bo) { struct ttm_bo_device *bdev = bo->bdev; int ret; retry_pre_get: ret = drm_mm_pre_get(&bdev->addr_space_mm); if (unlikely(ret != 0)) return ret; rw_wlock(&bdev->vm_lock); bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm, bo->mem.num_pages, 0, 0); if (unlikely(bo->vm_node == NULL)) { ret = -ENOMEM; goto out_unlock; } bo->vm_node = drm_mm_get_block_atomic(bo->vm_node, bo->mem.num_pages, 0); if (unlikely(bo->vm_node == NULL)) { rw_wunlock(&bdev->vm_lock); goto retry_pre_get; } ttm_bo_vm_insert_rb(bo); rw_wunlock(&bdev->vm_lock); bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT; return 0; out_unlock: rw_wunlock(&bdev->vm_lock); return ret; } int ttm_bo_wait(struct ttm_buffer_object *bo, bool lazy, bool interruptible, bool no_wait) { struct ttm_bo_driver *driver = bo->bdev->driver; struct ttm_bo_device *bdev = bo->bdev; void *sync_obj; int ret = 0; if (likely(bo->sync_obj == NULL)) return 0; while (bo->sync_obj) { if (driver->sync_obj_signaled(bo->sync_obj)) { void *tmp_obj = bo->sync_obj; bo->sync_obj = NULL; clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); mtx_unlock(&bdev->fence_lock); driver->sync_obj_unref(&tmp_obj); mtx_lock(&bdev->fence_lock); continue; } if (no_wait) return -EBUSY; sync_obj = driver->sync_obj_ref(bo->sync_obj); mtx_unlock(&bdev->fence_lock); ret = driver->sync_obj_wait(sync_obj, lazy, interruptible); if (unlikely(ret != 0)) { driver->sync_obj_unref(&sync_obj); mtx_lock(&bdev->fence_lock); return ret; } mtx_lock(&bdev->fence_lock); if (likely(bo->sync_obj == sync_obj)) { void *tmp_obj = bo->sync_obj; bo->sync_obj = NULL; clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); mtx_unlock(&bdev->fence_lock); driver->sync_obj_unref(&sync_obj); driver->sync_obj_unref(&tmp_obj); mtx_lock(&bdev->fence_lock); } else { mtx_unlock(&bdev->fence_lock); driver->sync_obj_unref(&sync_obj); mtx_lock(&bdev->fence_lock); } } return 0; } int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) { struct ttm_bo_device *bdev = bo->bdev; int ret = 0; /* * Using ttm_bo_reserve makes sure the lru lists are updated. */ ret = ttm_bo_reserve(bo, true, no_wait, false, 0); if (unlikely(ret != 0)) return ret; mtx_lock(&bdev->fence_lock); ret = ttm_bo_wait(bo, false, true, no_wait); mtx_unlock(&bdev->fence_lock); if (likely(ret == 0)) atomic_inc(&bo->cpu_writers); ttm_bo_unreserve(bo); return ret; } void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) { atomic_dec(&bo->cpu_writers); } /** * A buffer object shrink method that tries to swap out the first * buffer object on the bo_global::swap_lru list. */ static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) { struct ttm_bo_global *glob = container_of(shrink, struct ttm_bo_global, shrink); struct ttm_buffer_object *bo; int ret = -EBUSY; int put_count; uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); mtx_lock(&glob->lru_lock); list_for_each_entry(bo, &glob->swap_lru, swap) { ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); if (!ret) break; } if (ret) { mtx_unlock(&glob->lru_lock); return ret; } refcount_acquire(&bo->list_kref); if (!list_empty(&bo->ddestroy)) { ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); if (refcount_release(&bo->list_kref)) ttm_bo_release_list(bo); return ret; } put_count = ttm_bo_del_from_lru(bo); mtx_unlock(&glob->lru_lock); ttm_bo_list_ref_sub(bo, put_count, true); /** * Wait for GPU, then move to system cached. */ mtx_lock(&bo->bdev->fence_lock); ret = ttm_bo_wait(bo, false, false, false); mtx_unlock(&bo->bdev->fence_lock); if (unlikely(ret != 0)) goto out; if ((bo->mem.placement & swap_placement) != swap_placement) { struct ttm_mem_reg evict_mem; evict_mem = bo->mem; evict_mem.mm_node = NULL; evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; evict_mem.mem_type = TTM_PL_SYSTEM; ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, false, false); if (unlikely(ret != 0)) goto out; } ttm_bo_unmap_virtual(bo); /** * Swap out. Buffer will be swapped in again as soon as * anyone tries to access a ttm page. */ if (bo->bdev->driver->swap_notify) bo->bdev->driver->swap_notify(bo); ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); out: /** * * Unreserve without putting on LRU to avoid swapping out an * already swapped buffer. */ atomic_set(&bo->reserved, 0); wakeup(bo); if (refcount_release(&bo->list_kref)) ttm_bo_release_list(bo); return ret; } void ttm_bo_swapout_all(struct ttm_bo_device *bdev) { while (ttm_bo_swapout(&bdev->glob->shrink) == 0) ; } diff --git a/sys/dev/drm2/ttm/ttm_page_alloc.c b/sys/dev/drm2/ttm/ttm_page_alloc.c index 67c484218cc2..7518ecb4dfd1 100644 --- a/sys/dev/drm2/ttm/ttm_page_alloc.c +++ b/sys/dev/drm2/ttm/ttm_page_alloc.c @@ -1,919 +1,922 @@ /* * Copyright (c) Red Hat Inc. * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sub license, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the * next paragraph) shall be included in all copies or substantial portions * of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER * DEALINGS IN THE SOFTWARE. * * Authors: Dave Airlie * Jerome Glisse * Pauli Nieminen */ /* * Copyright (c) 2013 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. */ /* simple list based uncached page pool * - Pool collects resently freed pages for reuse * - Use page->lru to keep a free list * - doesn't track currently in use pages */ #include #include #include #include #include #include #define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(vm_page_t)) #define SMALL_ALLOCATION 16 #define FREE_ALL_PAGES (~0U) /* times are in msecs */ #define PAGE_FREE_INTERVAL 1000 /** * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages. * * @lock: Protects the shared pool from concurrnet access. Must be used with * irqsave/irqrestore variants because pool allocator maybe called from * delayed work. * @fill_lock: Prevent concurrent calls to fill. * @list: Pool of free uc/wc pages for fast reuse. * @gfp_flags: Flags to pass for alloc_page. * @npages: Number of pages in pool. */ struct ttm_page_pool { struct mtx lock; bool fill_lock; bool dma32; struct pglist list; int ttm_page_alloc_flags; unsigned npages; char *name; unsigned long nfrees; unsigned long nrefills; }; /** * Limits for the pool. They are handled without locks because only place where * they may change is in sysfs store. They won't have immediate effect anyway * so forcing serialization to access them is pointless. */ struct ttm_pool_opts { unsigned alloc_size; unsigned max_size; unsigned small; }; #define NUM_POOLS 4 /** * struct ttm_pool_manager - Holds memory pools for fst allocation * * Manager is read only object for pool code so it doesn't need locking. * * @free_interval: minimum number of jiffies between freeing pages from pool. * @page_alloc_inited: reference counting for pool allocation. * @work: Work that is used to shrink the pool. Work is only run when there is * some pages to free. * @small_allocation: Limit in number of pages what is small allocation. * * @pools: All pool objects in use. **/ struct ttm_pool_manager { unsigned int kobj_ref; eventhandler_tag lowmem_handler; struct ttm_pool_opts options; union { struct ttm_page_pool u_pools[NUM_POOLS]; struct _utag { struct ttm_page_pool u_wc_pool; struct ttm_page_pool u_uc_pool; struct ttm_page_pool u_wc_pool_dma32; struct ttm_page_pool u_uc_pool_dma32; } _ut; } _u; }; #define pools _u.u_pools #define wc_pool _u._ut.u_wc_pool #define uc_pool _u._ut.u_uc_pool #define wc_pool_dma32 _u._ut.u_wc_pool_dma32 #define uc_pool_dma32 _u._ut.u_uc_pool_dma32 MALLOC_DEFINE(M_TTM_POOLMGR, "ttm_poolmgr", "TTM Pool Manager"); static void ttm_vm_page_free(vm_page_t m) { KASSERT(m->object == NULL, ("ttm page %p is owned", m)); KASSERT(vm_page_wired(m), ("ttm lost wire %p", m)); KASSERT((m->flags & PG_FICTITIOUS) != 0, ("ttm lost fictitious %p", m)); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("ttm got unmanaged %p", m)); m->flags &= ~PG_FICTITIOUS; m->oflags |= VPO_UNMANAGED; vm_page_unwire_noq(m); vm_page_free(m); } static vm_memattr_t ttm_caching_state_to_vm(enum ttm_caching_state cstate) { switch (cstate) { case tt_uncached: return (VM_MEMATTR_UNCACHEABLE); case tt_wc: return (VM_MEMATTR_WRITE_COMBINING); case tt_cached: return (VM_MEMATTR_WRITE_BACK); } panic("caching state %d\n", cstate); } static vm_page_t ttm_vm_page_alloc_dma32(int req, vm_memattr_t memattr) { vm_page_t p; - int tries; + int err, tries; for (tries = 0; ; tries++) { p = vm_page_alloc_noobj_contig(req, 1, 0, 0xffffffff, PAGE_SIZE, 0, memattr); if (p != NULL || tries > 2) return (p); - if (!vm_page_reclaim_contig(req, 1, 0, 0xffffffff, - PAGE_SIZE, 0)) + err = vm_page_reclaim_contig(req, 1, 0, 0xffffffff, + PAGE_SIZE, 0); + if (err == ENOMEM) vm_wait(NULL); + else if (err != 0) + return (NULL); } } static vm_page_t ttm_vm_page_alloc_any(int req, vm_memattr_t memattr) { vm_page_t p; p = vm_page_alloc_noobj(req | VM_ALLOC_WAITOK); pmap_page_set_memattr(p, memattr); return (p); } static vm_page_t ttm_vm_page_alloc(int flags, enum ttm_caching_state cstate) { vm_page_t p; vm_memattr_t memattr; int req; memattr = ttm_caching_state_to_vm(cstate); req = VM_ALLOC_WIRED; if ((flags & TTM_PAGE_FLAG_ZERO_ALLOC) != 0) req |= VM_ALLOC_ZERO; if ((flags & TTM_PAGE_FLAG_DMA32) != 0) p = ttm_vm_page_alloc_dma32(req, memattr); else p = ttm_vm_page_alloc_any(req, memattr); if (p != NULL) { p->oflags &= ~VPO_UNMANAGED; p->flags |= PG_FICTITIOUS; } return (p); } static void ttm_pool_kobj_release(struct ttm_pool_manager *m) { free(m, M_TTM_POOLMGR); } #if 0 /* XXXKIB sysctl */ static ssize_t ttm_pool_store(struct ttm_pool_manager *m, struct attribute *attr, const char *buffer, size_t size) { int chars; unsigned val; chars = sscanf(buffer, "%u", &val); if (chars == 0) return size; /* Convert kb to number of pages */ val = val / (PAGE_SIZE >> 10); if (attr == &ttm_page_pool_max) m->options.max_size = val; else if (attr == &ttm_page_pool_small) m->options.small = val; else if (attr == &ttm_page_pool_alloc_size) { if (val > NUM_PAGES_TO_ALLOC*8) { pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n", NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7), NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10)); return size; } else if (val > NUM_PAGES_TO_ALLOC) { pr_warn("Setting allocation size to larger than %lu is not recommended\n", NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10)); } m->options.alloc_size = val; } return size; } static ssize_t ttm_pool_show(struct ttm_pool_manager *m, struct attribute *attr, char *buffer) { unsigned val = 0; if (attr == &ttm_page_pool_max) val = m->options.max_size; else if (attr == &ttm_page_pool_small) val = m->options.small; else if (attr == &ttm_page_pool_alloc_size) val = m->options.alloc_size; val = val * (PAGE_SIZE >> 10); return snprintf(buffer, PAGE_SIZE, "%u\n", val); } #endif static struct ttm_pool_manager *_manager; static int set_pages_array_wb(vm_page_t *pages, int addrinarray) { #ifdef TTM_HAS_AGP int i; for (i = 0; i < addrinarray; i++) pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_BACK); #endif return 0; } static int set_pages_array_wc(vm_page_t *pages, int addrinarray) { #ifdef TTM_HAS_AGP int i; for (i = 0; i < addrinarray; i++) pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_COMBINING); #endif return 0; } static int set_pages_array_uc(vm_page_t *pages, int addrinarray) { #ifdef TTM_HAS_AGP int i; for (i = 0; i < addrinarray; i++) pmap_page_set_memattr(pages[i], VM_MEMATTR_UNCACHEABLE); #endif return 0; } /** * Select the right pool or requested caching state and ttm flags. */ static struct ttm_page_pool *ttm_get_pool(int flags, enum ttm_caching_state cstate) { int pool_index; if (cstate == tt_cached) return NULL; if (cstate == tt_wc) pool_index = 0x0; else pool_index = 0x1; if (flags & TTM_PAGE_FLAG_DMA32) pool_index |= 0x2; return &_manager->pools[pool_index]; } /* set memory back to wb and free the pages. */ static void ttm_pages_put(vm_page_t *pages, unsigned npages) { unsigned i; /* Our VM handles vm memattr automatically on the page free. */ if (set_pages_array_wb(pages, npages)) printf("[TTM] Failed to set %d pages to wb!\n", npages); for (i = 0; i < npages; ++i) ttm_vm_page_free(pages[i]); } static void ttm_pool_update_free_locked(struct ttm_page_pool *pool, unsigned freed_pages) { pool->npages -= freed_pages; pool->nfrees += freed_pages; } /** * Free pages from pool. * * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC * number of pages in one go. * * @pool: to free the pages from * @free_all: If set to true will free all pages in pool **/ static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free) { vm_page_t p, p1; vm_page_t *pages_to_free; unsigned freed_pages = 0, npages_to_free = nr_free; unsigned i; if (NUM_PAGES_TO_ALLOC < nr_free) npages_to_free = NUM_PAGES_TO_ALLOC; pages_to_free = malloc(npages_to_free * sizeof(vm_page_t), M_TEMP, M_WAITOK | M_ZERO); restart: mtx_lock(&pool->lock); TAILQ_FOREACH_REVERSE_SAFE(p, &pool->list, pglist, plinks.q, p1) { if (freed_pages >= npages_to_free) break; pages_to_free[freed_pages++] = p; /* We can only remove NUM_PAGES_TO_ALLOC at a time. */ if (freed_pages >= NUM_PAGES_TO_ALLOC) { /* remove range of pages from the pool */ for (i = 0; i < freed_pages; i++) TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q); ttm_pool_update_free_locked(pool, freed_pages); /** * Because changing page caching is costly * we unlock the pool to prevent stalling. */ mtx_unlock(&pool->lock); ttm_pages_put(pages_to_free, freed_pages); if (likely(nr_free != FREE_ALL_PAGES)) nr_free -= freed_pages; if (NUM_PAGES_TO_ALLOC >= nr_free) npages_to_free = nr_free; else npages_to_free = NUM_PAGES_TO_ALLOC; freed_pages = 0; /* free all so restart the processing */ if (nr_free) goto restart; /* Not allowed to fall through or break because * following context is inside spinlock while we are * outside here. */ goto out; } } /* remove range of pages from the pool */ if (freed_pages) { for (i = 0; i < freed_pages; i++) TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q); ttm_pool_update_free_locked(pool, freed_pages); nr_free -= freed_pages; } mtx_unlock(&pool->lock); if (freed_pages) ttm_pages_put(pages_to_free, freed_pages); out: free(pages_to_free, M_TEMP); return nr_free; } /* Get good estimation how many pages are free in pools */ static int ttm_pool_get_num_unused_pages(void) { unsigned i; int total = 0; for (i = 0; i < NUM_POOLS; ++i) total += _manager->pools[i].npages; return total; } /** * Callback for mm to request pool to reduce number of page held. */ static int ttm_pool_mm_shrink(void *arg) { static unsigned int start_pool = 0; unsigned i; unsigned pool_offset = atomic_fetchadd_int(&start_pool, 1); struct ttm_page_pool *pool; int shrink_pages = 100; /* XXXKIB */ pool_offset = pool_offset % NUM_POOLS; /* select start pool in round robin fashion */ for (i = 0; i < NUM_POOLS; ++i) { unsigned nr_free = shrink_pages; if (shrink_pages == 0) break; pool = &_manager->pools[(i + pool_offset)%NUM_POOLS]; shrink_pages = ttm_page_pool_free(pool, nr_free); } /* return estimated number of unused pages in pool */ return ttm_pool_get_num_unused_pages(); } static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager) { manager->lowmem_handler = EVENTHANDLER_REGISTER(vm_lowmem, ttm_pool_mm_shrink, manager, EVENTHANDLER_PRI_ANY); } static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager) { EVENTHANDLER_DEREGISTER(vm_lowmem, manager->lowmem_handler); } static int ttm_set_pages_caching(vm_page_t *pages, enum ttm_caching_state cstate, unsigned cpages) { int r = 0; /* Set page caching */ switch (cstate) { case tt_uncached: r = set_pages_array_uc(pages, cpages); if (r) printf("[TTM] Failed to set %d pages to uc!\n", cpages); break; case tt_wc: r = set_pages_array_wc(pages, cpages); if (r) printf("[TTM] Failed to set %d pages to wc!\n", cpages); break; default: break; } return r; } /** * Free pages the pages that failed to change the caching state. If there is * any pages that have changed their caching state already put them to the * pool. */ static void ttm_handle_caching_state_failure(struct pglist *pages, int ttm_flags, enum ttm_caching_state cstate, vm_page_t *failed_pages, unsigned cpages) { unsigned i; /* Failed pages have to be freed */ for (i = 0; i < cpages; ++i) { TAILQ_REMOVE(pages, failed_pages[i], plinks.q); ttm_vm_page_free(failed_pages[i]); } } /** * Allocate new pages with correct caching. * * This function is reentrant if caller updates count depending on number of * pages returned in pages array. */ static int ttm_alloc_new_pages(struct pglist *pages, int ttm_alloc_flags, int ttm_flags, enum ttm_caching_state cstate, unsigned count) { vm_page_t *caching_array; vm_page_t p; int r = 0; unsigned i, cpages; unsigned max_cpages = min(count, (unsigned)(PAGE_SIZE/sizeof(vm_page_t))); /* allocate array for page caching change */ caching_array = malloc(max_cpages * sizeof(vm_page_t), M_TEMP, M_WAITOK | M_ZERO); for (i = 0, cpages = 0; i < count; ++i) { p = ttm_vm_page_alloc(ttm_alloc_flags, cstate); if (!p) { printf("[TTM] Unable to get page %u\n", i); /* store already allocated pages in the pool after * setting the caching state */ if (cpages) { r = ttm_set_pages_caching(caching_array, cstate, cpages); if (r) ttm_handle_caching_state_failure(pages, ttm_flags, cstate, caching_array, cpages); } r = -ENOMEM; goto out; } #ifdef CONFIG_HIGHMEM /* KIB: nop */ /* gfp flags of highmem page should never be dma32 so we * we should be fine in such case */ if (!PageHighMem(p)) #endif { caching_array[cpages++] = p; if (cpages == max_cpages) { r = ttm_set_pages_caching(caching_array, cstate, cpages); if (r) { ttm_handle_caching_state_failure(pages, ttm_flags, cstate, caching_array, cpages); goto out; } cpages = 0; } } TAILQ_INSERT_HEAD(pages, p, plinks.q); } if (cpages) { r = ttm_set_pages_caching(caching_array, cstate, cpages); if (r) ttm_handle_caching_state_failure(pages, ttm_flags, cstate, caching_array, cpages); } out: free(caching_array, M_TEMP); return r; } /** * Fill the given pool if there aren't enough pages and the requested number of * pages is small. */ static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool, int ttm_flags, enum ttm_caching_state cstate, unsigned count) { vm_page_t p; int r; unsigned cpages = 0; /** * Only allow one pool fill operation at a time. * If pool doesn't have enough pages for the allocation new pages are * allocated from outside of pool. */ if (pool->fill_lock) return; pool->fill_lock = true; /* If allocation request is small and there are not enough * pages in a pool we fill the pool up first. */ if (count < _manager->options.small && count > pool->npages) { struct pglist new_pages; unsigned alloc_size = _manager->options.alloc_size; /** * Can't change page caching if in irqsave context. We have to * drop the pool->lock. */ mtx_unlock(&pool->lock); TAILQ_INIT(&new_pages); r = ttm_alloc_new_pages(&new_pages, pool->ttm_page_alloc_flags, ttm_flags, cstate, alloc_size); mtx_lock(&pool->lock); if (!r) { TAILQ_CONCAT(&pool->list, &new_pages, plinks.q); ++pool->nrefills; pool->npages += alloc_size; } else { printf("[TTM] Failed to fill pool (%p)\n", pool); /* If we have any pages left put them to the pool. */ TAILQ_FOREACH(p, &pool->list, plinks.q) { ++cpages; } TAILQ_CONCAT(&pool->list, &new_pages, plinks.q); pool->npages += cpages; } } pool->fill_lock = false; } /** * Cut 'count' number of pages from the pool and put them on the return list. * * @return count of pages still required to fulfill the request. */ static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool, struct pglist *pages, int ttm_flags, enum ttm_caching_state cstate, unsigned count) { vm_page_t p; unsigned i; mtx_lock(&pool->lock); ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count); if (count >= pool->npages) { /* take all pages from the pool */ TAILQ_CONCAT(pages, &pool->list, plinks.q); count -= pool->npages; pool->npages = 0; goto out; } for (i = 0; i < count; i++) { p = TAILQ_FIRST(&pool->list); TAILQ_REMOVE(&pool->list, p, plinks.q); TAILQ_INSERT_TAIL(pages, p, plinks.q); } pool->npages -= count; count = 0; out: mtx_unlock(&pool->lock); return count; } /* Put all pages in pages list to correct pool to wait for reuse */ static void ttm_put_pages(vm_page_t *pages, unsigned npages, int flags, enum ttm_caching_state cstate) { struct ttm_page_pool *pool = ttm_get_pool(flags, cstate); unsigned i; if (pool == NULL) { /* No pool for this memory type so free the pages */ for (i = 0; i < npages; i++) { if (pages[i]) { ttm_vm_page_free(pages[i]); pages[i] = NULL; } } return; } mtx_lock(&pool->lock); for (i = 0; i < npages; i++) { if (pages[i]) { TAILQ_INSERT_TAIL(&pool->list, pages[i], plinks.q); pages[i] = NULL; pool->npages++; } } /* Check that we don't go over the pool limit */ npages = 0; if (pool->npages > _manager->options.max_size) { npages = pool->npages - _manager->options.max_size; /* free at least NUM_PAGES_TO_ALLOC number of pages * to reduce calls to set_memory_wb */ if (npages < NUM_PAGES_TO_ALLOC) npages = NUM_PAGES_TO_ALLOC; } mtx_unlock(&pool->lock); if (npages) ttm_page_pool_free(pool, npages); } /* * On success pages list will hold count number of correctly * cached pages. */ static int ttm_get_pages(vm_page_t *pages, unsigned npages, int flags, enum ttm_caching_state cstate) { struct ttm_page_pool *pool = ttm_get_pool(flags, cstate); struct pglist plist; vm_page_t p = NULL; int gfp_flags; unsigned count; int r; /* No pool for cached pages */ if (pool == NULL) { for (r = 0; r < npages; ++r) { p = ttm_vm_page_alloc(flags, cstate); if (!p) { printf("[TTM] Unable to allocate page\n"); return -ENOMEM; } pages[r] = p; } return 0; } /* combine zero flag to pool flags */ gfp_flags = flags | pool->ttm_page_alloc_flags; /* First we take pages from the pool */ TAILQ_INIT(&plist); npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages); count = 0; TAILQ_FOREACH(p, &plist, plinks.q) { pages[count++] = p; } /* clear the pages coming from the pool if requested */ if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) { TAILQ_FOREACH(p, &plist, plinks.q) { pmap_zero_page(p); } } /* If pool didn't have enough pages allocate new one. */ if (npages > 0) { /* ttm_alloc_new_pages doesn't reference pool so we can run * multiple requests in parallel. **/ TAILQ_INIT(&plist); r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate, npages); TAILQ_FOREACH(p, &plist, plinks.q) { pages[count++] = p; } if (r) { /* If there is any pages in the list put them back to * the pool. */ printf("[TTM] Failed to allocate extra pages for large request\n"); ttm_put_pages(pages, count, flags, cstate); return r; } } return 0; } static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags, char *name) { mtx_init(&pool->lock, "ttmpool", NULL, MTX_DEF); pool->fill_lock = false; TAILQ_INIT(&pool->list); pool->npages = pool->nfrees = 0; pool->ttm_page_alloc_flags = flags; pool->name = name; } int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages) { if (_manager != NULL) printf("[TTM] manager != NULL\n"); printf("[TTM] Initializing pool allocator\n"); _manager = malloc(sizeof(*_manager), M_TTM_POOLMGR, M_WAITOK | M_ZERO); ttm_page_pool_init_locked(&_manager->wc_pool, 0, "wc"); ttm_page_pool_init_locked(&_manager->uc_pool, 0, "uc"); ttm_page_pool_init_locked(&_manager->wc_pool_dma32, TTM_PAGE_FLAG_DMA32, "wc dma"); ttm_page_pool_init_locked(&_manager->uc_pool_dma32, TTM_PAGE_FLAG_DMA32, "uc dma"); _manager->options.max_size = max_pages; _manager->options.small = SMALL_ALLOCATION; _manager->options.alloc_size = NUM_PAGES_TO_ALLOC; refcount_init(&_manager->kobj_ref, 1); ttm_pool_mm_shrink_init(_manager); return 0; } void ttm_page_alloc_fini(void) { int i; printf("[TTM] Finalizing pool allocator\n"); ttm_pool_mm_shrink_fini(_manager); for (i = 0; i < NUM_POOLS; ++i) ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES); if (refcount_release(&_manager->kobj_ref)) ttm_pool_kobj_release(_manager); _manager = NULL; } int ttm_pool_populate(struct ttm_tt *ttm) { struct ttm_mem_global *mem_glob = ttm->glob->mem_glob; unsigned i; int ret; if (ttm->state != tt_unpopulated) return 0; for (i = 0; i < ttm->num_pages; ++i) { ret = ttm_get_pages(&ttm->pages[i], 1, ttm->page_flags, ttm->caching_state); if (ret != 0) { ttm_pool_unpopulate(ttm); return -ENOMEM; } ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i], false, false); if (unlikely(ret != 0)) { ttm_pool_unpopulate(ttm); return -ENOMEM; } } if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) { ret = ttm_tt_swapin(ttm); if (unlikely(ret != 0)) { ttm_pool_unpopulate(ttm); return ret; } } ttm->state = tt_unbound; return 0; } void ttm_pool_unpopulate(struct ttm_tt *ttm) { unsigned i; for (i = 0; i < ttm->num_pages; ++i) { if (ttm->pages[i]) { ttm_mem_global_free_page(ttm->glob->mem_glob, ttm->pages[i]); ttm_put_pages(&ttm->pages[i], 1, ttm->page_flags, ttm->caching_state); } } ttm->state = tt_unpopulated; } #if 0 /* XXXKIB sysctl */ int ttm_page_alloc_debugfs(struct seq_file *m, void *data) { struct ttm_page_pool *p; unsigned i; char *h[] = {"pool", "refills", "pages freed", "size"}; if (!_manager) { seq_printf(m, "No pool allocator running.\n"); return 0; } seq_printf(m, "%6s %12s %13s %8s\n", h[0], h[1], h[2], h[3]); for (i = 0; i < NUM_POOLS; ++i) { p = &_manager->pools[i]; seq_printf(m, "%6s %12ld %13ld %8d\n", p->name, p->nrefills, p->nfrees, p->npages); } return 0; } #endif diff --git a/sys/kern/uipc_ktls.c b/sys/kern/uipc_ktls.c index 60ec05d3d94a..64150086658a 100644 --- a/sys/kern/uipc_ktls.c +++ b/sys/kern/uipc_ktls.c @@ -1,3361 +1,3362 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2014-2019 Netflix Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include #include "opt_inet.h" #include "opt_inet6.h" #include "opt_kern_tls.h" #include "opt_ratelimit.h" #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) #include #endif #include #include #include #ifdef RSS #include #include #endif #include #include #include #include #include #ifdef TCP_OFFLOAD #include #endif #include #include #include #include #include #include struct ktls_wq { struct mtx mtx; STAILQ_HEAD(, mbuf) m_head; STAILQ_HEAD(, socket) so_head; bool running; int lastallocfail; } __aligned(CACHE_LINE_SIZE); struct ktls_reclaim_thread { uint64_t wakeups; uint64_t reclaims; struct thread *td; int running; }; struct ktls_domain_info { int count; int cpu[MAXCPU]; struct ktls_reclaim_thread reclaim_td; }; struct ktls_domain_info ktls_domains[MAXMEMDOM]; static struct ktls_wq *ktls_wq; static struct proc *ktls_proc; static uma_zone_t ktls_session_zone; static uma_zone_t ktls_buffer_zone; static uint16_t ktls_cpuid_lookup[MAXCPU]; static int ktls_init_state; static struct sx ktls_init_lock; SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init"); SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Kernel TLS offload"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Kernel TLS offload stats"); #ifdef RSS static int ktls_bind_threads = 1; #else static int ktls_bind_threads; #endif SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, &ktls_bind_threads, 0, "Bind crypto threads to cores (1) or cores and domains (2) at boot"); static u_int ktls_maxlen = 16384; SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN, &ktls_maxlen, 0, "Maximum TLS record size"); static int ktls_number_threads; SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, &ktls_number_threads, 0, "Number of TLS threads in thread-pool"); unsigned int ktls_ifnet_max_rexmit_pct = 2; SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN, &ktls_ifnet_max_rexmit_pct, 2, "Max percent bytes retransmitted before ifnet TLS is disabled"); static bool ktls_offload_enable; SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN, &ktls_offload_enable, 0, "Enable support for kernel TLS offload"); static bool ktls_cbc_enable = true; SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN, &ktls_cbc_enable, 1, "Enable support of AES-CBC crypto for kernel TLS"); static bool ktls_sw_buffer_cache = true; SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN, &ktls_sw_buffer_cache, 1, "Enable caching of output buffers for SW encryption"); static int ktls_max_reclaim = 1024; SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN, &ktls_max_reclaim, 128, "Max number of 16k buffers to reclaim in thread context"); static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active); SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, &ktls_tasks_active, "Number of active tasks"); static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD, &ktls_cnt_tx_pending, "Number of TLS 1.0 records waiting for earlier TLS records"); static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD, &ktls_cnt_tx_queued, "Number of TLS records in queue to tasks for SW encryption"); static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD, &ktls_cnt_rx_queued, "Number of TLS sockets in queue to tasks for SW decryption"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_total); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, CTLFLAG_RD, &ktls_offload_total, "Total successful TLS setups (parameters set)"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, CTLFLAG_RD, &ktls_offload_enable_calls, "Total number of TLS enable calls made"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_active); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, &ktls_offload_active, "Total Active TLS sessions"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD, &ktls_offload_corrupted_records, "Total corrupted TLS records received"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, &ktls_offload_failed_crypto, "Total TLS crypto failures"); static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD, &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD, &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet"); static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD, &ktls_destroy_task, "Number of times ktls session was destroyed via taskqueue"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Software TLS session stats"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Hardware (ifnet) TLS session stats"); #ifdef TCP_OFFLOAD SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "TOE TLS session stats"); #endif static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc); SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, "Active number of software TLS sessions using AES-CBC"); static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm); SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, "Active number of software TLS sessions using AES-GCM"); static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20); SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD, &ktls_sw_chacha20, "Active number of software TLS sessions using Chacha20-Poly1305"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, &ktls_ifnet_cbc, "Active number of ifnet TLS sessions using AES-CBC"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, &ktls_ifnet_gcm, "Active number of ifnet TLS sessions using AES-GCM"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD, &ktls_ifnet_chacha20, "Active number of ifnet TLS sessions using Chacha20-Poly1305"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, &ktls_ifnet_reset_dropped, "TLS sessions dropped after failing to update ifnet send tag"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, &ktls_ifnet_reset_failed, "TLS sessions that failed to allocate a new ifnet send tag"); static int ktls_ifnet_permitted; SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, &ktls_ifnet_permitted, 1, "Whether to permit hardware (ifnet) TLS sessions"); #ifdef TCP_OFFLOAD static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc); SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD, &ktls_toe_cbc, "Active number of TOE TLS sessions using AES-CBC"); static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm); SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD, &ktls_toe_gcm, "Active number of TOE TLS sessions using AES-GCM"); static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20); SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD, &ktls_toe_chacha20, "Active number of TOE TLS sessions using Chacha20-Poly1305"); #endif static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); static void ktls_reset_receive_tag(void *context, int pending); static void ktls_reset_send_tag(void *context, int pending); static void ktls_work_thread(void *ctx); static void ktls_reclaim_thread(void *ctx); static u_int ktls_get_cpu(struct socket *so) { struct inpcb *inp; #ifdef NUMA struct ktls_domain_info *di; #endif u_int cpuid; inp = sotoinpcb(so); #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid != NETISR_CPUID_NONE) return (cpuid); #endif /* * Just use the flowid to shard connections in a repeatable * fashion. Note that TLS 1.0 sessions rely on the * serialization provided by having the same connection use * the same queue. */ #ifdef NUMA if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) { di = &ktls_domains[inp->inp_numa_domain]; cpuid = di->cpu[inp->inp_flowid % di->count]; } else #endif cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; return (cpuid); } static int ktls_buffer_import(void *arg, void **store, int count, int domain, int flags) { vm_page_t m; int i, req; KASSERT((ktls_maxlen & PAGE_MASK) == 0, ("%s: ktls max length %d is not page size-aligned", __func__, ktls_maxlen)); req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags); for (i = 0; i < count; i++) { m = vm_page_alloc_noobj_contig_domain(domain, req, atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); if (m == NULL) break; store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); } return (i); } static void ktls_buffer_release(void *arg __unused, void **store, int count) { vm_page_t m; int i, j; for (i = 0; i < count; i++) { m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i])); for (j = 0; j < atop(ktls_maxlen); j++) { (void)vm_page_unwire_noq(m + j); vm_page_free(m + j); } } } static void ktls_free_mext_contig(struct mbuf *m) { M_ASSERTEXTPG(m); uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0])); } static int ktls_init(void) { struct thread *td; struct pcpu *pc; int count, domain, error, i; ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, M_WAITOK | M_ZERO); ktls_session_zone = uma_zcreate("ktls_session", sizeof(struct ktls_session), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); if (ktls_sw_buffer_cache) { ktls_buffer_zone = uma_zcache_create("ktls_buffers", roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL, ktls_buffer_import, ktls_buffer_release, NULL, UMA_ZONE_FIRSTTOUCH); } /* * Initialize the workqueues to run the TLS work. We create a * work queue for each CPU. */ CPU_FOREACH(i) { STAILQ_INIT(&ktls_wq[i].m_head); STAILQ_INIT(&ktls_wq[i].so_head); mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); if (ktls_bind_threads > 1) { pc = pcpu_find(i); domain = pc->pc_domain; count = ktls_domains[domain].count; ktls_domains[domain].cpu[count] = i; ktls_domains[domain].count++; } ktls_cpuid_lookup[ktls_number_threads] = i; ktls_number_threads++; } /* * If we somehow have an empty domain, fall back to choosing * among all KTLS threads. */ if (ktls_bind_threads > 1) { for (i = 0; i < vm_ndomains; i++) { if (ktls_domains[i].count == 0) { ktls_bind_threads = 1; break; } } } /* Start kthreads for each workqueue. */ CPU_FOREACH(i) { error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); if (error) { printf("Can't add KTLS thread %d error %d\n", i, error); return (error); } } /* * Start an allocation thread per-domain to perform blocking allocations * of 16k physically contiguous TLS crypto destination buffers. */ if (ktls_sw_buffer_cache) { for (domain = 0; domain < vm_ndomains; domain++) { if (VM_DOMAIN_EMPTY(domain)) continue; if (CPU_EMPTY(&cpuset_domain[domain])) continue; error = kproc_kthread_add(ktls_reclaim_thread, &ktls_domains[domain], &ktls_proc, &ktls_domains[domain].reclaim_td.td, 0, 0, "KTLS", "reclaim_%d", domain); if (error) { printf("Can't add KTLS reclaim thread %d error %d\n", domain, error); return (error); } } } if (bootverbose) printf("KTLS: Initialized %d threads\n", ktls_number_threads); return (0); } static int ktls_start_kthreads(void) { int error, state; start: state = atomic_load_acq_int(&ktls_init_state); if (__predict_true(state > 0)) return (0); if (state < 0) return (ENXIO); sx_xlock(&ktls_init_lock); if (ktls_init_state != 0) { sx_xunlock(&ktls_init_lock); goto start; } error = ktls_init(); if (error == 0) state = 1; else state = -1; atomic_store_rel_int(&ktls_init_state, state); sx_xunlock(&ktls_init_lock); return (error); } static int ktls_create_session(struct socket *so, struct tls_enable *en, struct ktls_session **tlsp, int direction) { struct ktls_session *tls; int error; /* Only TLS 1.0 - 1.3 are supported. */ if (en->tls_vmajor != TLS_MAJOR_VER_ONE) return (EINVAL); if (en->tls_vminor < TLS_MINOR_VER_ZERO || en->tls_vminor > TLS_MINOR_VER_THREE) return (EINVAL); if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) return (EINVAL); if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) return (EINVAL); if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) return (EINVAL); /* All supported algorithms require a cipher key. */ if (en->cipher_key_len == 0) return (EINVAL); /* No flags are currently supported. */ if (en->flags != 0) return (EINVAL); /* Common checks for supported algorithms. */ switch (en->cipher_algorithm) { case CRYPTO_AES_NIST_GCM_16: /* * auth_algorithm isn't used, but permit GMAC values * for compatibility. */ switch (en->auth_algorithm) { case 0: #ifdef COMPAT_FREEBSD12 /* XXX: Really 13.0-current COMPAT. */ case CRYPTO_AES_128_NIST_GMAC: case CRYPTO_AES_192_NIST_GMAC: case CRYPTO_AES_256_NIST_GMAC: #endif break; default: return (EINVAL); } if (en->auth_key_len != 0) return (EINVAL); switch (en->tls_vminor) { case TLS_MINOR_VER_TWO: if (en->iv_len != TLS_AEAD_GCM_LEN) return (EINVAL); break; case TLS_MINOR_VER_THREE: if (en->iv_len != TLS_1_3_GCM_IV_LEN) return (EINVAL); break; default: return (EINVAL); } break; case CRYPTO_AES_CBC: switch (en->auth_algorithm) { case CRYPTO_SHA1_HMAC: break; case CRYPTO_SHA2_256_HMAC: case CRYPTO_SHA2_384_HMAC: if (en->tls_vminor != TLS_MINOR_VER_TWO) return (EINVAL); break; default: return (EINVAL); } if (en->auth_key_len == 0) return (EINVAL); /* * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2 * use explicit IVs. */ switch (en->tls_vminor) { case TLS_MINOR_VER_ZERO: if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) return (EINVAL); break; case TLS_MINOR_VER_ONE: case TLS_MINOR_VER_TWO: /* Ignore any supplied IV. */ en->iv_len = 0; break; default: return (EINVAL); } break; case CRYPTO_CHACHA20_POLY1305: if (en->auth_algorithm != 0 || en->auth_key_len != 0) return (EINVAL); if (en->tls_vminor != TLS_MINOR_VER_TWO && en->tls_vminor != TLS_MINOR_VER_THREE) return (EINVAL); if (en->iv_len != TLS_CHACHA20_IV_LEN) return (EINVAL); break; default: return (EINVAL); } error = ktls_start_kthreads(); if (error != 0) return (error); tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); counter_u64_add(ktls_offload_active, 1); refcount_init(&tls->refcount, 1); if (direction == KTLS_RX) { TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls); } else { TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); tls->inp = so->so_pcb; in_pcbref(tls->inp); tls->tx = true; } tls->wq_index = ktls_get_cpu(so); tls->params.cipher_algorithm = en->cipher_algorithm; tls->params.auth_algorithm = en->auth_algorithm; tls->params.tls_vmajor = en->tls_vmajor; tls->params.tls_vminor = en->tls_vminor; tls->params.flags = en->flags; tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); /* Set the header and trailer lengths. */ tls->params.tls_hlen = sizeof(struct tls_record_layer); switch (en->cipher_algorithm) { case CRYPTO_AES_NIST_GCM_16: /* * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte * nonce. TLS 1.3 uses a 12 byte implicit IV. */ if (en->tls_vminor < TLS_MINOR_VER_THREE) tls->params.tls_hlen += sizeof(uint64_t); tls->params.tls_tlen = AES_GMAC_HASH_LEN; tls->params.tls_bs = 1; break; case CRYPTO_AES_CBC: switch (en->auth_algorithm) { case CRYPTO_SHA1_HMAC: if (en->tls_vminor == TLS_MINOR_VER_ZERO) { /* Implicit IV, no nonce. */ tls->sequential_records = true; tls->next_seqno = be64dec(en->rec_seq); STAILQ_INIT(&tls->pending_records); } else { tls->params.tls_hlen += AES_BLOCK_LEN; } tls->params.tls_tlen = AES_BLOCK_LEN + SHA1_HASH_LEN; break; case CRYPTO_SHA2_256_HMAC: tls->params.tls_hlen += AES_BLOCK_LEN; tls->params.tls_tlen = AES_BLOCK_LEN + SHA2_256_HASH_LEN; break; case CRYPTO_SHA2_384_HMAC: tls->params.tls_hlen += AES_BLOCK_LEN; tls->params.tls_tlen = AES_BLOCK_LEN + SHA2_384_HASH_LEN; break; default: panic("invalid hmac"); } tls->params.tls_bs = AES_BLOCK_LEN; break; case CRYPTO_CHACHA20_POLY1305: /* * Chacha20 uses a 12 byte implicit IV. */ tls->params.tls_tlen = POLY1305_HASH_LEN; tls->params.tls_bs = 1; break; default: panic("invalid cipher"); } /* * TLS 1.3 includes optional padding which we do not support, * and also puts the "real" record type at the end of the * encrypted data. */ if (en->tls_vminor == TLS_MINOR_VER_THREE) tls->params.tls_tlen += sizeof(uint8_t); KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, ("TLS header length too long: %d", tls->params.tls_hlen)); KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, ("TLS trailer length too long: %d", tls->params.tls_tlen)); if (en->auth_key_len != 0) { tls->params.auth_key_len = en->auth_key_len; tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, M_WAITOK); error = copyin(en->auth_key, tls->params.auth_key, en->auth_key_len); if (error) goto out; } tls->params.cipher_key_len = en->cipher_key_len; tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); error = copyin(en->cipher_key, tls->params.cipher_key, en->cipher_key_len); if (error) goto out; /* * This holds the implicit portion of the nonce for AEAD * ciphers and the initial implicit IV for TLS 1.0. The * explicit portions of the IV are generated in ktls_frame(). */ if (en->iv_len != 0) { tls->params.iv_len = en->iv_len; error = copyin(en->iv, tls->params.iv, en->iv_len); if (error) goto out; /* * For TLS 1.2 with GCM, generate an 8-byte nonce as a * counter to generate unique explicit IVs. * * Store this counter in the last 8 bytes of the IV * array so that it is 8-byte aligned. */ if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && en->tls_vminor == TLS_MINOR_VER_TWO) arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0); } *tlsp = tls; return (0); out: ktls_free(tls); return (error); } static struct ktls_session * ktls_clone_session(struct ktls_session *tls, int direction) { struct ktls_session *tls_new; tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); counter_u64_add(ktls_offload_active, 1); refcount_init(&tls_new->refcount, 1); if (direction == KTLS_RX) { TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag, tls_new); } else { TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new); tls_new->inp = tls->inp; tls_new->tx = true; in_pcbref(tls_new->inp); } /* Copy fields from existing session. */ tls_new->params = tls->params; tls_new->wq_index = tls->wq_index; /* Deep copy keys. */ if (tls_new->params.auth_key != NULL) { tls_new->params.auth_key = malloc(tls->params.auth_key_len, M_KTLS, M_WAITOK); memcpy(tls_new->params.auth_key, tls->params.auth_key, tls->params.auth_key_len); } tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, M_WAITOK); memcpy(tls_new->params.cipher_key, tls->params.cipher_key, tls->params.cipher_key_len); return (tls_new); } #ifdef TCP_OFFLOAD static int ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction) { struct inpcb *inp; struct tcpcb *tp; int error; inp = so->so_pcb; INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } if (inp->inp_socket == NULL) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); if (!(tp->t_flags & TF_TOE)) { INP_WUNLOCK(inp); return (EOPNOTSUPP); } error = tcp_offload_alloc_tls_session(tp, tls, direction); INP_WUNLOCK(inp); if (error == 0) { tls->mode = TCP_TLS_MODE_TOE; switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_toe_cbc, 1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_toe_gcm, 1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_toe_chacha20, 1); break; } } return (error); } #endif /* * Common code used when first enabling ifnet TLS on a connection or * when allocating a new ifnet TLS session due to a routing change. * This function allocates a new TLS send tag on whatever interface * the connection is currently routed over. */ static int ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, struct m_snd_tag **mstp) { union if_snd_tag_alloc_params params; struct ifnet *ifp; struct nhop_object *nh; struct tcpcb *tp; int error; INP_RLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } if (inp->inp_socket == NULL) { INP_RUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); /* * Check administrative controls on ifnet TLS to determine if * ifnet TLS should be denied. * * - Always permit 'force' requests. * - ktls_ifnet_permitted == 0: always deny. */ if (!force && ktls_ifnet_permitted == 0) { INP_RUNLOCK(inp); return (ENXIO); } /* * XXX: Use the cached route in the inpcb to find the * interface. This should perhaps instead use * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only * enabled after a connection has completed key negotiation in * userland, the cached route will be present in practice. */ nh = inp->inp_route.ro_nh; if (nh == NULL) { INP_RUNLOCK(inp); return (ENXIO); } ifp = nh->nh_ifp; if_ref(ifp); /* * Allocate a TLS + ratelimit tag if the connection has an * existing pacing rate. */ if (tp->t_pacing_rate != -1 && (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) { params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT; params.tls_rate_limit.inp = inp; params.tls_rate_limit.tls = tls; params.tls_rate_limit.max_rate = tp->t_pacing_rate; } else { params.hdr.type = IF_SND_TAG_TYPE_TLS; params.tls.inp = inp; params.tls.tls = tls; } params.hdr.flowid = inp->inp_flowid; params.hdr.flowtype = inp->inp_flowtype; params.hdr.numa_domain = inp->inp_numa_domain; INP_RUNLOCK(inp); if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) { error = EOPNOTSUPP; goto out; } if (inp->inp_vflag & INP_IPV6) { if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) { error = EOPNOTSUPP; goto out; } } else { if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) { error = EOPNOTSUPP; goto out; } } error = m_snd_tag_alloc(ifp, ¶ms, mstp); out: if_rele(ifp); return (error); } /* * Allocate an initial TLS receive tag for doing HW decryption of TLS * data. * * This function allocates a new TLS receive tag on whatever interface * the connection is currently routed over. If the connection ends up * using a different interface for receive this will get fixed up via * ktls_input_ifp_mismatch as future packets arrive. */ static int ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls, struct m_snd_tag **mstp) { union if_snd_tag_alloc_params params; struct ifnet *ifp; struct nhop_object *nh; int error; if (!ktls_ocf_recrypt_supported(tls)) return (ENXIO); INP_RLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } if (inp->inp_socket == NULL) { INP_RUNLOCK(inp); return (ECONNRESET); } /* * Check administrative controls on ifnet TLS to determine if * ifnet TLS should be denied. */ if (ktls_ifnet_permitted == 0) { INP_RUNLOCK(inp); return (ENXIO); } /* * XXX: As with ktls_alloc_snd_tag, use the cached route in * the inpcb to find the interface. */ nh = inp->inp_route.ro_nh; if (nh == NULL) { INP_RUNLOCK(inp); return (ENXIO); } ifp = nh->nh_ifp; if_ref(ifp); tls->rx_ifp = ifp; params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; params.hdr.flowid = inp->inp_flowid; params.hdr.flowtype = inp->inp_flowtype; params.hdr.numa_domain = inp->inp_numa_domain; params.tls_rx.inp = inp; params.tls_rx.tls = tls; params.tls_rx.vlan_id = 0; INP_RUNLOCK(inp); if (inp->inp_vflag & INP_IPV6) { if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) { error = EOPNOTSUPP; goto out; } } else { if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) { error = EOPNOTSUPP; goto out; } } error = m_snd_tag_alloc(ifp, ¶ms, mstp); /* * If this connection is over a vlan, vlan_snd_tag_alloc * rewrites vlan_id with the saved interface. Save the VLAN * ID for use in ktls_reset_receive_tag which allocates new * receive tags directly from the leaf interface bypassing * if_vlan. */ if (error == 0) tls->rx_vlan_id = params.tls_rx.vlan_id; out: return (error); } static int ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction, bool force) { struct m_snd_tag *mst; int error; switch (direction) { case KTLS_TX: error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); if (__predict_false(error != 0)) goto done; break; case KTLS_RX: KASSERT(!force, ("%s: forced receive tag", __func__)); error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst); if (__predict_false(error != 0)) goto done; break; default: __assert_unreachable(); } tls->mode = TCP_TLS_MODE_IFNET; tls->snd_tag = mst; switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_ifnet_cbc, 1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_ifnet_gcm, 1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_ifnet_chacha20, 1); break; default: break; } done: return (error); } static void ktls_use_sw(struct ktls_session *tls) { tls->mode = TCP_TLS_MODE_SW; switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_sw_cbc, 1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_sw_gcm, 1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_sw_chacha20, 1); break; } } static int ktls_try_sw(struct ktls_session *tls, int direction) { int error; error = ktls_ocf_try(tls, direction); if (error) return (error); ktls_use_sw(tls); return (0); } /* * KTLS RX stores data in the socket buffer as a list of TLS records, * where each record is stored as a control message containg the TLS * header followed by data mbufs containing the decrypted data. This * is different from KTLS TX which always uses an mb_ext_pgs mbuf for * both encrypted and decrypted data. TLS records decrypted by a NIC * should be queued to the socket buffer as records, but encrypted * data which needs to be decrypted by software arrives as a stream of * regular mbufs which need to be converted. In addition, there may * already be pending encrypted data in the socket buffer when KTLS RX * is enabled. * * To manage not-yet-decrypted data for KTLS RX, the following scheme * is used: * * - A single chain of NOTREADY mbufs is hung off of sb_mtls. * * - ktls_check_rx checks this chain of mbufs reading the TLS header * from the first mbuf. Once all of the data for that TLS record is * queued, the socket is queued to a worker thread. * * - The worker thread calls ktls_decrypt to decrypt TLS records in * the TLS chain. Each TLS record is detached from the TLS chain, * decrypted, and inserted into the regular socket buffer chain as * record starting with a control message holding the TLS header and * a chain of mbufs holding the encrypted data. */ static void sb_mark_notready(struct sockbuf *sb) { struct mbuf *m; m = sb->sb_mb; sb->sb_mtls = m; sb->sb_mb = NULL; sb->sb_mbtail = NULL; sb->sb_lastrecord = NULL; for (; m != NULL; m = m->m_next) { KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL", __func__)); KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail", __func__)); KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len", __func__)); m->m_flags |= M_NOTREADY; sb->sb_acc -= m->m_len; sb->sb_tlscc += m->m_len; sb->sb_mtlstail = m; } KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc, ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc, sb->sb_ccc)); } /* * Return information about the pending TLS data in a socket * buffer. On return, 'seqno' is set to the sequence number * of the next TLS record to be received, 'resid' is set to * the amount of bytes still needed for the last pending * record. The function returns 'false' if the last pending * record contains a partial TLS header. In that case, 'resid' * is the number of bytes needed to complete the TLS header. */ bool ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp) { struct tls_record_layer hdr; struct mbuf *m; uint64_t seqno; size_t resid; u_int offset, record_len; SOCKBUF_LOCK_ASSERT(sb); MPASS(sb->sb_flags & SB_TLS_RX); seqno = sb->sb_tls_seqno; resid = sb->sb_tlscc; m = sb->sb_mtls; offset = 0; if (resid == 0) { *seqnop = seqno; *residp = 0; return (true); } for (;;) { seqno++; if (resid < sizeof(hdr)) { *seqnop = seqno; *residp = sizeof(hdr) - resid; return (false); } m_copydata(m, offset, sizeof(hdr), (void *)&hdr); record_len = sizeof(hdr) + ntohs(hdr.tls_length); if (resid <= record_len) { *seqnop = seqno; *residp = record_len - resid; return (true); } resid -= record_len; while (record_len != 0) { if (m->m_len - offset > record_len) { offset += record_len; break; } record_len -= (m->m_len - offset); offset = 0; m = m->m_next; } } } int ktls_enable_rx(struct socket *so, struct tls_enable *en) { struct ktls_session *tls; int error; if (!ktls_offload_enable) return (ENOTSUP); counter_u64_add(ktls_offload_enable_calls, 1); /* * This should always be true since only the TCP socket option * invokes this function. */ if (so->so_proto->pr_protocol != IPPROTO_TCP) return (EINVAL); /* * XXX: Don't overwrite existing sessions. We should permit * this to support rekeying in the future. */ if (so->so_rcv.sb_tls_info != NULL) return (EALREADY); if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) return (ENOTSUP); error = ktls_create_session(so, en, &tls, KTLS_RX); if (error) return (error); error = ktls_ocf_try(tls, KTLS_RX); if (error) { ktls_free(tls); return (error); } /* * Serialize with soreceive_generic() and make sure that we're not * operating on a listening socket. */ error = SOCK_IO_RECV_LOCK(so, SBL_WAIT); if (error) { ktls_free(tls); return (error); } /* Mark the socket as using TLS offload. */ SOCK_RECVBUF_LOCK(so); if (__predict_false(so->so_rcv.sb_tls_info != NULL)) error = EALREADY; else if ((so->so_rcv.sb_flags & SB_SPLICED) != 0) error = EINVAL; if (error != 0) { SOCK_RECVBUF_UNLOCK(so); SOCK_IO_RECV_UNLOCK(so); ktls_free(tls); return (EALREADY); } so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq); so->so_rcv.sb_tls_info = tls; so->so_rcv.sb_flags |= SB_TLS_RX; /* Mark existing data as not ready until it can be decrypted. */ sb_mark_notready(&so->so_rcv); ktls_check_rx(&so->so_rcv); SOCK_RECVBUF_UNLOCK(so); SOCK_IO_RECV_UNLOCK(so); /* Prefer TOE -> ifnet TLS -> software TLS. */ #ifdef TCP_OFFLOAD error = ktls_try_toe(so, tls, KTLS_RX); if (error) #endif error = ktls_try_ifnet(so, tls, KTLS_RX, false); if (error) ktls_use_sw(tls); counter_u64_add(ktls_offload_total, 1); return (0); } int ktls_enable_tx(struct socket *so, struct tls_enable *en) { struct ktls_session *tls; struct inpcb *inp; struct tcpcb *tp; int error; if (!ktls_offload_enable) return (ENOTSUP); counter_u64_add(ktls_offload_enable_calls, 1); /* * This should always be true since only the TCP socket option * invokes this function. */ if (so->so_proto->pr_protocol != IPPROTO_TCP) return (EINVAL); /* * XXX: Don't overwrite existing sessions. We should permit * this to support rekeying in the future. */ if (so->so_snd.sb_tls_info != NULL) return (EALREADY); if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) return (ENOTSUP); /* TLS requires ext pgs */ if (mb_use_ext_pgs == 0) return (ENXIO); error = ktls_create_session(so, en, &tls, KTLS_TX); if (error) return (error); /* Prefer TOE -> ifnet TLS -> software TLS. */ #ifdef TCP_OFFLOAD error = ktls_try_toe(so, tls, KTLS_TX); if (error) #endif error = ktls_try_ifnet(so, tls, KTLS_TX, false); if (error) error = ktls_try_sw(tls, KTLS_TX); if (error) { ktls_free(tls); return (error); } /* * Serialize with sosend_generic() and make sure that we're not * operating on a listening socket. */ error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); if (error) { ktls_free(tls); return (error); } /* * Write lock the INP when setting sb_tls_info so that * routines in tcp_ratelimit.c can read sb_tls_info while * holding the INP lock. */ inp = so->so_pcb; INP_WLOCK(inp); SOCK_SENDBUF_LOCK(so); if (__predict_false(so->so_snd.sb_tls_info != NULL)) error = EALREADY; else if ((so->so_snd.sb_flags & SB_SPLICED) != 0) error = EINVAL; if (error != 0) { SOCK_SENDBUF_UNLOCK(so); INP_WUNLOCK(inp); SOCK_IO_SEND_UNLOCK(so); ktls_free(tls); return (error); } so->so_snd.sb_tls_seqno = be64dec(en->rec_seq); so->so_snd.sb_tls_info = tls; if (tls->mode != TCP_TLS_MODE_SW) { tp = intotcpcb(inp); MPASS(tp->t_nic_ktls_xmit == 0); tp->t_nic_ktls_xmit = 1; if (tp->t_fb->tfb_hwtls_change != NULL) (*tp->t_fb->tfb_hwtls_change)(tp, 1); } SOCK_SENDBUF_UNLOCK(so); INP_WUNLOCK(inp); SOCK_IO_SEND_UNLOCK(so); counter_u64_add(ktls_offload_total, 1); return (0); } int ktls_get_rx_mode(struct socket *so, int *modep) { struct ktls_session *tls; struct inpcb *inp __diagused; if (SOLISTENING(so)) return (EINVAL); inp = so->so_pcb; INP_WLOCK_ASSERT(inp); SOCK_RECVBUF_LOCK(so); tls = so->so_rcv.sb_tls_info; if (tls == NULL) *modep = TCP_TLS_MODE_NONE; else *modep = tls->mode; SOCK_RECVBUF_UNLOCK(so); return (0); } /* * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number. * * This function gets information about the next TCP- and TLS- * sequence number to be processed by the TLS receive worker * thread. The information is extracted from the given "inpcb" * structure. The values are stored in host endian format at the two * given output pointer locations. The TCP sequence number points to * the beginning of the TLS header. * * This function returns zero on success, else a non-zero error code * is returned. */ int ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq) { struct socket *so; struct tcpcb *tp; INP_RLOCK(inp); so = inp->inp_socket; if (__predict_false(so == NULL)) { INP_RUNLOCK(inp); return (EINVAL); } if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); MPASS(tp != NULL); SOCKBUF_LOCK(&so->so_rcv); *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc; *tlsseq = so->so_rcv.sb_tls_seqno; SOCKBUF_UNLOCK(&so->so_rcv); INP_RUNLOCK(inp); return (0); } int ktls_get_tx_mode(struct socket *so, int *modep) { struct ktls_session *tls; struct inpcb *inp __diagused; if (SOLISTENING(so)) return (EINVAL); inp = so->so_pcb; INP_WLOCK_ASSERT(inp); SOCK_SENDBUF_LOCK(so); tls = so->so_snd.sb_tls_info; if (tls == NULL) *modep = TCP_TLS_MODE_NONE; else *modep = tls->mode; SOCK_SENDBUF_UNLOCK(so); return (0); } /* * Switch between SW and ifnet TLS sessions as requested. */ int ktls_set_tx_mode(struct socket *so, int mode) { struct ktls_session *tls, *tls_new; struct inpcb *inp; struct tcpcb *tp; int error; if (SOLISTENING(so)) return (EINVAL); switch (mode) { case TCP_TLS_MODE_SW: case TCP_TLS_MODE_IFNET: break; default: return (EINVAL); } inp = so->so_pcb; INP_WLOCK_ASSERT(inp); tp = intotcpcb(inp); if (mode == TCP_TLS_MODE_IFNET) { /* Don't allow enabling ifnet ktls multiple times */ if (tp->t_nic_ktls_xmit) return (EALREADY); /* * Don't enable ifnet ktls if we disabled it due to an * excessive retransmission rate */ if (tp->t_nic_ktls_xmit_dis) return (ENXIO); } SOCKBUF_LOCK(&so->so_snd); tls = so->so_snd.sb_tls_info; if (tls == NULL) { SOCKBUF_UNLOCK(&so->so_snd); return (0); } if (tls->mode == mode) { SOCKBUF_UNLOCK(&so->so_snd); return (0); } tls = ktls_hold(tls); SOCKBUF_UNLOCK(&so->so_snd); INP_WUNLOCK(inp); tls_new = ktls_clone_session(tls, KTLS_TX); if (mode == TCP_TLS_MODE_IFNET) error = ktls_try_ifnet(so, tls_new, KTLS_TX, true); else error = ktls_try_sw(tls_new, KTLS_TX); if (error) { counter_u64_add(ktls_switch_failed, 1); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (error); } error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); if (error) { counter_u64_add(ktls_switch_failed, 1); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (error); } /* * If we raced with another session change, keep the existing * session. */ if (tls != so->so_snd.sb_tls_info) { counter_u64_add(ktls_switch_failed, 1); SOCK_IO_SEND_UNLOCK(so); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (EBUSY); } INP_WLOCK(inp); SOCKBUF_LOCK(&so->so_snd); so->so_snd.sb_tls_info = tls_new; if (tls_new->mode != TCP_TLS_MODE_SW) { MPASS(tp->t_nic_ktls_xmit == 0); tp->t_nic_ktls_xmit = 1; if (tp->t_fb->tfb_hwtls_change != NULL) (*tp->t_fb->tfb_hwtls_change)(tp, 1); } SOCKBUF_UNLOCK(&so->so_snd); SOCK_IO_SEND_UNLOCK(so); /* * Drop two references on 'tls'. The first is for the * ktls_hold() above. The second drops the reference from the * socket buffer. */ KASSERT(tls->refcount >= 2, ("too few references on old session")); ktls_free(tls); ktls_free(tls); if (mode == TCP_TLS_MODE_IFNET) counter_u64_add(ktls_switch_to_ifnet, 1); else counter_u64_add(ktls_switch_to_sw, 1); return (0); } /* * Try to allocate a new TLS receive tag. This task is scheduled when * sbappend_ktls_rx detects an input path change. If a new tag is * allocated, replace the tag in the TLS session. If a new tag cannot * be allocated, let the session fall back to software decryption. */ static void ktls_reset_receive_tag(void *context, int pending) { union if_snd_tag_alloc_params params; struct ktls_session *tls; struct m_snd_tag *mst; struct inpcb *inp; struct ifnet *ifp; struct socket *so; int error; MPASS(pending == 1); tls = context; so = tls->so; inp = so->so_pcb; ifp = NULL; INP_RLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); goto out; } SOCKBUF_LOCK(&so->so_rcv); mst = tls->snd_tag; tls->snd_tag = NULL; if (mst != NULL) m_snd_tag_rele(mst); ifp = tls->rx_ifp; if_ref(ifp); SOCKBUF_UNLOCK(&so->so_rcv); params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; params.hdr.flowid = inp->inp_flowid; params.hdr.flowtype = inp->inp_flowtype; params.hdr.numa_domain = inp->inp_numa_domain; params.tls_rx.inp = inp; params.tls_rx.tls = tls; params.tls_rx.vlan_id = tls->rx_vlan_id; INP_RUNLOCK(inp); if (inp->inp_vflag & INP_IPV6) { if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0) goto out; } else { if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0) goto out; } error = m_snd_tag_alloc(ifp, ¶ms, &mst); if (error == 0) { SOCKBUF_LOCK(&so->so_rcv); tls->snd_tag = mst; SOCKBUF_UNLOCK(&so->so_rcv); counter_u64_add(ktls_ifnet_reset, 1); } else { /* * Just fall back to software decryption if a tag * cannot be allocated leaving the connection intact. * If a future input path change switches to another * interface this connection will resume ifnet TLS. */ counter_u64_add(ktls_ifnet_reset_failed, 1); } out: mtx_pool_lock(mtxpool_sleep, tls); tls->reset_pending = false; mtx_pool_unlock(mtxpool_sleep, tls); if (ifp != NULL) if_rele(ifp); CURVNET_SET(so->so_vnet); sorele(so); CURVNET_RESTORE(); ktls_free(tls); } /* * Try to allocate a new TLS send tag. This task is scheduled when * ip_output detects a route change while trying to transmit a packet * holding a TLS record. If a new tag is allocated, replace the tag * in the TLS session. Subsequent packets on the connection will use * the new tag. If a new tag cannot be allocated, drop the * connection. */ static void ktls_reset_send_tag(void *context, int pending) { struct epoch_tracker et; struct ktls_session *tls; struct m_snd_tag *old, *new; struct inpcb *inp; struct tcpcb *tp; int error; MPASS(pending == 1); tls = context; inp = tls->inp; /* * Free the old tag first before allocating a new one. * ip[6]_output_send() will treat a NULL send tag the same as * an ifp mismatch and drop packets until a new tag is * allocated. * * Write-lock the INP when changing tls->snd_tag since * ip[6]_output_send() holds a read-lock when reading the * pointer. */ INP_WLOCK(inp); old = tls->snd_tag; tls->snd_tag = NULL; INP_WUNLOCK(inp); if (old != NULL) m_snd_tag_rele(old); error = ktls_alloc_snd_tag(inp, tls, true, &new); if (error == 0) { INP_WLOCK(inp); tls->snd_tag = new; mtx_pool_lock(mtxpool_sleep, tls); tls->reset_pending = false; mtx_pool_unlock(mtxpool_sleep, tls); INP_WUNLOCK(inp); counter_u64_add(ktls_ifnet_reset, 1); /* * XXX: Should we kick tcp_output explicitly now that * the send tag is fixed or just rely on timers? */ } else { NET_EPOCH_ENTER(et); INP_WLOCK(inp); if (!(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); CURVNET_SET(inp->inp_vnet); tp = tcp_drop(tp, ECONNABORTED); CURVNET_RESTORE(); if (tp != NULL) { counter_u64_add(ktls_ifnet_reset_dropped, 1); INP_WUNLOCK(inp); } } else INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); counter_u64_add(ktls_ifnet_reset_failed, 1); /* * Leave reset_pending true to avoid future tasks while * the socket goes away. */ } ktls_free(tls); } void ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp) { struct ktls_session *tls; struct socket *so; SOCKBUF_LOCK_ASSERT(sb); KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", __func__, sb)); so = __containerof(sb, struct socket, so_rcv); tls = sb->sb_tls_info; if_rele(tls->rx_ifp); if_ref(ifp); tls->rx_ifp = ifp; /* * See if we should schedule a task to update the receive tag for * this session. */ mtx_pool_lock(mtxpool_sleep, tls); if (!tls->reset_pending) { (void) ktls_hold(tls); soref(so); tls->so = so; tls->reset_pending = true; taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); } mtx_pool_unlock(mtxpool_sleep, tls); } int ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) { if (inp == NULL) return (ENOBUFS); INP_LOCK_ASSERT(inp); /* * See if we should schedule a task to update the send tag for * this session. */ mtx_pool_lock(mtxpool_sleep, tls); if (!tls->reset_pending) { (void) ktls_hold(tls); tls->reset_pending = true; taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); } mtx_pool_unlock(mtxpool_sleep, tls); return (ENOBUFS); } #ifdef RATELIMIT int ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate) { union if_snd_tag_modify_params params = { .rate_limit.max_rate = max_pacing_rate, .rate_limit.flags = M_NOWAIT, }; struct m_snd_tag *mst; /* Can't get to the inp, but it should be locked. */ /* INP_LOCK_ASSERT(inp); */ MPASS(tls->mode == TCP_TLS_MODE_IFNET); if (tls->snd_tag == NULL) { /* * Resetting send tag, ignore this change. The * pending reset may or may not see this updated rate * in the tcpcb. If it doesn't, we will just lose * this rate change. */ return (0); } mst = tls->snd_tag; MPASS(mst != NULL); MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT); return (mst->sw->snd_tag_modify(mst, ¶ms)); } #endif static void ktls_destroy_help(void *context, int pending __unused) { ktls_destroy(context); } void ktls_destroy(struct ktls_session *tls) { struct inpcb *inp; struct tcpcb *tp; bool wlocked; MPASS(tls->refcount == 0); inp = tls->inp; if (tls->tx) { wlocked = INP_WLOCKED(inp); if (!wlocked && !INP_TRY_WLOCK(inp)) { /* * rwlocks read locks are anonymous, and there * is no way to know if our current thread * holds an rlock on the inp. As a rough * estimate, check to see if the thread holds * *any* rlocks at all. If it does not, then we * know that we don't hold the inp rlock, and * can safely take the wlock */ if (curthread->td_rw_rlocks == 0) { INP_WLOCK(inp); } else { /* * We might hold the rlock, so let's * do the destroy in a taskqueue * context to avoid a potential * deadlock. This should be very * rare. */ counter_u64_add(ktls_destroy_task, 1); TASK_INIT(&tls->destroy_task, 0, ktls_destroy_help, tls); (void)taskqueue_enqueue(taskqueue_thread, &tls->destroy_task); return; } } } if (tls->sequential_records) { struct mbuf *m, *n; int page_count; STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) { page_count = m->m_epg_enc_cnt; while (page_count > 0) { KASSERT(page_count >= m->m_epg_nrdy, ("%s: too few pages", __func__)); page_count -= m->m_epg_nrdy; m = m_free(m); } } } counter_u64_add(ktls_offload_active, -1); switch (tls->mode) { case TCP_TLS_MODE_SW: switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_sw_cbc, -1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_sw_gcm, -1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_sw_chacha20, -1); break; } break; case TCP_TLS_MODE_IFNET: switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_ifnet_cbc, -1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_ifnet_gcm, -1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_ifnet_chacha20, -1); break; } if (tls->snd_tag != NULL) m_snd_tag_rele(tls->snd_tag); if (tls->rx_ifp != NULL) if_rele(tls->rx_ifp); if (tls->tx) { INP_WLOCK_ASSERT(inp); tp = intotcpcb(inp); MPASS(tp->t_nic_ktls_xmit == 1); tp->t_nic_ktls_xmit = 0; } break; #ifdef TCP_OFFLOAD case TCP_TLS_MODE_TOE: switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_toe_cbc, -1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_toe_gcm, -1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_toe_chacha20, -1); break; } break; #endif } if (tls->ocf_session != NULL) ktls_ocf_free(tls); if (tls->params.auth_key != NULL) { zfree(tls->params.auth_key, M_KTLS); tls->params.auth_key = NULL; tls->params.auth_key_len = 0; } if (tls->params.cipher_key != NULL) { zfree(tls->params.cipher_key, M_KTLS); tls->params.cipher_key = NULL; tls->params.cipher_key_len = 0; } if (tls->tx) { INP_WLOCK_ASSERT(inp); if (!in_pcbrele_wlocked(inp) && !wlocked) INP_WUNLOCK(inp); } explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); uma_zfree(ktls_session_zone, tls); } void ktls_seq(struct sockbuf *sb, struct mbuf *m) { for (; m != NULL; m = m->m_next) { KASSERT((m->m_flags & M_EXTPG) != 0, ("ktls_seq: mapped mbuf %p", m)); m->m_epg_seqno = sb->sb_tls_seqno; sb->sb_tls_seqno++; } } /* * Add TLS framing (headers and trailers) to a chain of mbufs. Each * mbuf in the chain must be an unmapped mbuf. The payload of the * mbuf must be populated with the payload of each TLS record. * * The record_type argument specifies the TLS record type used when * populating the TLS header. * * The enq_count argument on return is set to the number of pages of * payload data for this entire chain that need to be encrypted via SW * encryption. The returned value should be passed to ktls_enqueue * when scheduling encryption of this chain of mbufs. To handle the * special case of empty fragments for TLS 1.0 sessions, an empty * fragment counts as one page. */ void ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, uint8_t record_type) { struct tls_record_layer *tlshdr; struct mbuf *m; uint64_t *noncep; uint16_t tls_len; int maxlen __diagused; maxlen = tls->params.max_frame_len; *enq_cnt = 0; for (m = top; m != NULL; m = m->m_next) { /* * All mbufs in the chain should be TLS records whose * payload does not exceed the maximum frame length. * * Empty TLS 1.0 records are permitted when using CBC. */ KASSERT(m->m_len <= maxlen && m->m_len >= 0 && (m->m_len > 0 || ktls_permit_empty_frames(tls)), ("ktls_frame: m %p len %d", m, m->m_len)); /* * TLS frames require unmapped mbufs to store session * info. */ KASSERT((m->m_flags & M_EXTPG) != 0, ("ktls_frame: mapped mbuf %p (top = %p)", m, top)); tls_len = m->m_len; /* Save a reference to the session. */ m->m_epg_tls = ktls_hold(tls); m->m_epg_hdrlen = tls->params.tls_hlen; m->m_epg_trllen = tls->params.tls_tlen; if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { int bs, delta; /* * AES-CBC pads messages to a multiple of the * block size. Note that the padding is * applied after the digest and the encryption * is done on the "plaintext || mac || padding". * At least one byte of padding is always * present. * * Compute the final trailer length assuming * at most one block of padding. * tls->params.tls_tlen is the maximum * possible trailer length (padding + digest). * delta holds the number of excess padding * bytes if the maximum were used. Those * extra bytes are removed. */ bs = tls->params.tls_bs; delta = (tls_len + tls->params.tls_tlen) & (bs - 1); m->m_epg_trllen -= delta; } m->m_len += m->m_epg_hdrlen + m->m_epg_trllen; /* Populate the TLS header. */ tlshdr = (void *)m->m_epg_hdr; tlshdr->tls_vmajor = tls->params.tls_vmajor; /* * TLS 1.3 masquarades as TLS 1.2 with a record type * of TLS_RLTYPE_APP. */ if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { tlshdr->tls_vminor = TLS_MINOR_VER_TWO; tlshdr->tls_type = TLS_RLTYPE_APP; /* save the real record type for later */ m->m_epg_record_type = record_type; m->m_epg_trail[0] = record_type; } else { tlshdr->tls_vminor = tls->params.tls_vminor; tlshdr->tls_type = record_type; } tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); /* * Store nonces / explicit IVs after the end of the * TLS header. * * For GCM with TLS 1.2, an 8 byte nonce is copied * from the end of the IV. The nonce is then * incremented for use by the next record. * * For CBC, a random nonce is inserted for TLS 1.1+. */ if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && tls->params.tls_vminor == TLS_MINOR_VER_TWO) { noncep = (uint64_t *)(tls->params.iv + 8); be64enc(tlshdr + 1, *noncep); (*noncep)++; } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && tls->params.tls_vminor >= TLS_MINOR_VER_ONE) arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); /* * When using SW encryption, mark the mbuf not ready. * It will be marked ready via sbready() after the * record has been encrypted. * * When using ifnet TLS, unencrypted TLS records are * sent down the stack to the NIC. */ if (tls->mode == TCP_TLS_MODE_SW) { m->m_flags |= M_NOTREADY; if (__predict_false(tls_len == 0)) { /* TLS 1.0 empty fragment. */ m->m_epg_nrdy = 1; } else m->m_epg_nrdy = m->m_epg_npgs; *enq_cnt += m->m_epg_nrdy; } } } bool ktls_permit_empty_frames(struct ktls_session *tls) { return (tls->params.cipher_algorithm == CRYPTO_AES_CBC && tls->params.tls_vminor == TLS_MINOR_VER_ZERO); } void ktls_check_rx(struct sockbuf *sb) { struct tls_record_layer hdr; struct ktls_wq *wq; struct socket *so; bool running; SOCKBUF_LOCK_ASSERT(sb); KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", __func__, sb)); so = __containerof(sb, struct socket, so_rcv); if (sb->sb_flags & SB_TLS_RX_RUNNING) return; /* Is there enough queued for a TLS header? */ if (sb->sb_tlscc < sizeof(hdr)) { if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0) so->so_error = EMSGSIZE; return; } m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr); /* Is the entire record queued? */ if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) { if ((sb->sb_state & SBS_CANTRCVMORE) != 0) so->so_error = EMSGSIZE; return; } sb->sb_flags |= SB_TLS_RX_RUNNING; soref(so); wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index]; mtx_lock(&wq->mtx); STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); counter_u64_add(ktls_cnt_rx_queued, 1); } static struct mbuf * ktls_detach_record(struct sockbuf *sb, int len) { struct mbuf *m, *n, *top; int remain; SOCKBUF_LOCK_ASSERT(sb); MPASS(len <= sb->sb_tlscc); /* * If TLS chain is the exact size of the record, * just grab the whole record. */ top = sb->sb_mtls; if (sb->sb_tlscc == len) { sb->sb_mtls = NULL; sb->sb_mtlstail = NULL; goto out; } /* * While it would be nice to use m_split() here, we need * to know exactly what m_split() allocates to update the * accounting, so do it inline instead. */ remain = len; for (m = top; remain > m->m_len; m = m->m_next) remain -= m->m_len; /* Easy case: don't have to split 'm'. */ if (remain == m->m_len) { sb->sb_mtls = m->m_next; if (sb->sb_mtls == NULL) sb->sb_mtlstail = NULL; m->m_next = NULL; goto out; } /* * Need to allocate an mbuf to hold the remainder of 'm'. Try * with M_NOWAIT first. */ n = m_get(M_NOWAIT, MT_DATA); if (n == NULL) { /* * Use M_WAITOK with socket buffer unlocked. If * 'sb_mtls' changes while the lock is dropped, return * NULL to force the caller to retry. */ SOCKBUF_UNLOCK(sb); n = m_get(M_WAITOK, MT_DATA); SOCKBUF_LOCK(sb); if (sb->sb_mtls != top) { m_free(n); return (NULL); } } n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED)); /* Store remainder in 'n'. */ n->m_len = m->m_len - remain; if (m->m_flags & M_EXT) { n->m_data = m->m_data + remain; mb_dupcl(n, m); } else { bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len); } /* Trim 'm' and update accounting. */ m->m_len -= n->m_len; sb->sb_tlscc -= n->m_len; sb->sb_ccc -= n->m_len; /* Account for 'n'. */ sballoc_ktls_rx(sb, n); /* Insert 'n' into the TLS chain. */ sb->sb_mtls = n; n->m_next = m->m_next; if (sb->sb_mtlstail == m) sb->sb_mtlstail = n; /* Detach the record from the TLS chain. */ m->m_next = NULL; out: MPASS(m_length(top, NULL) == len); for (m = top; m != NULL; m = m->m_next) sbfree_ktls_rx(sb, m); sb->sb_tlsdcc = len; sb->sb_ccc += len; SBCHECK(sb); return (top); } /* * Determine the length of the trailing zero padding and find the real * record type in the byte before the padding. * * Walking the mbuf chain backwards is clumsy, so another option would * be to scan forwards remembering the last non-zero byte before the * trailer. However, it would be expensive to scan the entire record. * Instead, find the last non-zero byte of each mbuf in the chain * keeping track of the relative offset of that nonzero byte. * * trail_len is the size of the MAC/tag on input and is set to the * size of the full trailer including padding and the record type on * return. */ static int tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len, int *trailer_len, uint8_t *record_typep) { char *cp; u_int digest_start, last_offset, m_len, offset; uint8_t record_type; digest_start = tls_len - *trailer_len; last_offset = 0; offset = 0; for (; m != NULL && offset < digest_start; offset += m->m_len, m = m->m_next) { /* Don't look for padding in the tag. */ m_len = min(digest_start - offset, m->m_len); cp = mtod(m, char *); /* Find last non-zero byte in this mbuf. */ while (m_len > 0 && cp[m_len - 1] == 0) m_len--; if (m_len > 0) { record_type = cp[m_len - 1]; last_offset = offset + m_len; } } if (last_offset < tls->params.tls_hlen) return (EBADMSG); *record_typep = record_type; *trailer_len = tls_len - last_offset + 1; return (0); } /* * Check if a mbuf chain is fully decrypted at the given offset and * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data * is encrypted. */ ktls_mbuf_crypto_st_t ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len) { int m_flags_ored = 0; int m_flags_anded = -1; for (; mb != NULL; mb = mb->m_next) { if (offset < mb->m_len) break; offset -= mb->m_len; } offset += len; for (; mb != NULL; mb = mb->m_next) { m_flags_ored |= mb->m_flags; m_flags_anded &= mb->m_flags; if (offset <= mb->m_len) break; offset -= mb->m_len; } MPASS(mb != NULL || offset == 0); if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED) return (KTLS_MBUF_CRYPTO_ST_MIXED); else return ((m_flags_ored & M_DECRYPTED) ? KTLS_MBUF_CRYPTO_ST_DECRYPTED : KTLS_MBUF_CRYPTO_ST_ENCRYPTED); } /* * ktls_resync_ifnet - get HW TLS RX back on track after packet loss */ static int ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num) { union if_snd_tag_modify_params params; struct m_snd_tag *mst; struct inpcb *inp; struct tcpcb *tp; mst = so->so_rcv.sb_tls_info->snd_tag; if (__predict_false(mst == NULL)) return (EINVAL); inp = sotoinpcb(so); if (__predict_false(inp == NULL)) return (EINVAL); INP_RLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); MPASS(tp != NULL); /* Get the TCP sequence number of the next valid TLS header. */ SOCKBUF_LOCK(&so->so_rcv); params.tls_rx.tls_hdr_tcp_sn = tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len; params.tls_rx.tls_rec_length = tls_len; params.tls_rx.tls_seq_number = tls_rcd_num; SOCKBUF_UNLOCK(&so->so_rcv); INP_RUNLOCK(inp); MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX); return (mst->sw->snd_tag_modify(mst, ¶ms)); } static void ktls_drop(struct socket *so, int error) { struct epoch_tracker et; struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp; NET_EPOCH_ENTER(et); INP_WLOCK(inp); if (!(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); CURVNET_SET(inp->inp_vnet); tp = tcp_drop(tp, error); CURVNET_RESTORE(); if (tp != NULL) INP_WUNLOCK(inp); } else { so->so_error = error; SOCK_RECVBUF_LOCK(so); sorwakeup_locked(so); INP_WUNLOCK(inp); } NET_EPOCH_EXIT(et); } static void ktls_decrypt(struct socket *so) { char tls_header[MBUF_PEXT_HDR_LEN]; struct ktls_session *tls; struct sockbuf *sb; struct tls_record_layer *hdr; struct tls_get_record tgr; struct mbuf *control, *data, *m; ktls_mbuf_crypto_st_t state; uint64_t seqno; int error, remain, tls_len, trail_len; bool tls13; uint8_t vminor, record_type; hdr = (struct tls_record_layer *)tls_header; sb = &so->so_rcv; SOCKBUF_LOCK(sb); KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING, ("%s: socket %p not running", __func__, so)); tls = sb->sb_tls_info; MPASS(tls != NULL); tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE); if (tls13) vminor = TLS_MINOR_VER_TWO; else vminor = tls->params.tls_vminor; for (;;) { /* Is there enough queued for a TLS header? */ if (sb->sb_tlscc < tls->params.tls_hlen) break; m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header); tls_len = sizeof(*hdr) + ntohs(hdr->tls_length); if (hdr->tls_vmajor != tls->params.tls_vmajor || hdr->tls_vminor != vminor) error = EINVAL; else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP) error = EINVAL; else if (tls_len < tls->params.tls_hlen || tls_len > tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 + tls->params.tls_tlen) error = EMSGSIZE; else error = 0; if (__predict_false(error != 0)) { /* * We have a corrupted record and are likely * out of sync. The connection isn't * recoverable at this point, so abort it. */ SOCKBUF_UNLOCK(sb); counter_u64_add(ktls_offload_corrupted_records, 1); ktls_drop(so, error); goto deref; } /* Is the entire record queued? */ if (sb->sb_tlscc < tls_len) break; /* * Split out the portion of the mbuf chain containing * this TLS record. */ data = ktls_detach_record(sb, tls_len); if (data == NULL) continue; MPASS(sb->sb_tlsdcc == tls_len); seqno = sb->sb_tls_seqno; sb->sb_tls_seqno++; SBCHECK(sb); SOCKBUF_UNLOCK(sb); /* get crypto state for this TLS record */ state = ktls_mbuf_crypto_state(data, 0, tls_len); switch (state) { case KTLS_MBUF_CRYPTO_ST_MIXED: error = ktls_ocf_recrypt(tls, hdr, data, seqno); if (error) break; /* FALLTHROUGH */ case KTLS_MBUF_CRYPTO_ST_ENCRYPTED: error = ktls_ocf_decrypt(tls, hdr, data, seqno, &trail_len); if (__predict_true(error == 0)) { if (tls13) { error = tls13_find_record_type(tls, data, tls_len, &trail_len, &record_type); } else { record_type = hdr->tls_type; } } break; case KTLS_MBUF_CRYPTO_ST_DECRYPTED: /* * NIC TLS is only supported for AEAD * ciphersuites which used a fixed sized * trailer. */ if (tls13) { trail_len = tls->params.tls_tlen - 1; error = tls13_find_record_type(tls, data, tls_len, &trail_len, &record_type); } else { trail_len = tls->params.tls_tlen; error = 0; record_type = hdr->tls_type; } break; default: error = EINVAL; break; } if (error) { counter_u64_add(ktls_offload_failed_crypto, 1); SOCKBUF_LOCK(sb); if (sb->sb_tlsdcc == 0) { /* * sbcut/drop/flush discarded these * mbufs. */ m_freem(data); break; } /* * Drop this TLS record's data, but keep * decrypting subsequent records. */ sb->sb_ccc -= tls_len; sb->sb_tlsdcc = 0; if (error != EMSGSIZE) error = EBADMSG; CURVNET_SET(so->so_vnet); so->so_error = error; sorwakeup_locked(so); CURVNET_RESTORE(); m_freem(data); SOCKBUF_LOCK(sb); continue; } /* Allocate the control mbuf. */ memset(&tgr, 0, sizeof(tgr)); tgr.tls_type = record_type; tgr.tls_vmajor = hdr->tls_vmajor; tgr.tls_vminor = hdr->tls_vminor; tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen - trail_len); control = sbcreatecontrol(&tgr, sizeof(tgr), TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK); SOCKBUF_LOCK(sb); if (sb->sb_tlsdcc == 0) { /* sbcut/drop/flush discarded these mbufs. */ MPASS(sb->sb_tlscc == 0); m_freem(data); m_freem(control); break; } /* * Clear the 'dcc' accounting in preparation for * adding the decrypted record. */ sb->sb_ccc -= tls_len; sb->sb_tlsdcc = 0; SBCHECK(sb); /* If there is no payload, drop all of the data. */ if (tgr.tls_length == htobe16(0)) { m_freem(data); data = NULL; } else { /* Trim header. */ remain = tls->params.tls_hlen; while (remain > 0) { if (data->m_len > remain) { data->m_data += remain; data->m_len -= remain; break; } remain -= data->m_len; data = m_free(data); } /* Trim trailer and clear M_NOTREADY. */ remain = be16toh(tgr.tls_length); m = data; for (m = data; remain > m->m_len; m = m->m_next) { m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); remain -= m->m_len; } m->m_len = remain; m_freem(m->m_next); m->m_next = NULL; m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); /* Set EOR on the final mbuf. */ m->m_flags |= M_EOR; } sbappendcontrol_locked(sb, data, control, 0); if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) { sb->sb_flags |= SB_TLS_RX_RESYNC; SOCKBUF_UNLOCK(sb); ktls_resync_ifnet(so, tls_len, seqno); SOCKBUF_LOCK(sb); } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) { sb->sb_flags &= ~SB_TLS_RX_RESYNC; SOCKBUF_UNLOCK(sb); ktls_resync_ifnet(so, 0, seqno); SOCKBUF_LOCK(sb); } } sb->sb_flags &= ~SB_TLS_RX_RUNNING; if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0) so->so_error = EMSGSIZE; sorwakeup_locked(so); deref: SOCKBUF_UNLOCK_ASSERT(sb); CURVNET_SET(so->so_vnet); sorele(so); CURVNET_RESTORE(); } void ktls_enqueue_to_free(struct mbuf *m) { struct ktls_wq *wq; bool running; /* Mark it for freeing. */ m->m_epg_flags |= EPG_FLAG_2FREE; wq = &ktls_wq[m->m_epg_tls->wq_index]; mtx_lock(&wq->mtx); STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); } static void * ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m) { void *buf; int domain, running; if (m->m_epg_npgs <= 2) return (NULL); if (ktls_buffer_zone == NULL) return (NULL); if ((u_int)(ticks - wq->lastallocfail) < hz) { /* * Rate-limit allocation attempts after a failure. * ktls_buffer_import() will acquire a per-domain mutex to check * the free page queues and may fail consistently if memory is * fragmented. */ return (NULL); } buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM); if (buf == NULL) { domain = PCPU_GET(domain); wq->lastallocfail = ticks; /* * Note that this check is "racy", but the races are * harmless, and are either a spurious wakeup if * multiple threads fail allocations before the alloc * thread wakes, or waiting an extra second in case we * see an old value of running == true. */ if (!VM_DOMAIN_EMPTY(domain)) { running = atomic_load_int(&ktls_domains[domain].reclaim_td.running); if (!running) wakeup(&ktls_domains[domain].reclaim_td); } } return (buf); } static int ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m, struct ktls_session *tls, struct ktls_ocf_encrypt_state *state) { vm_page_t pg; int error, i, len, off; KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY), ("%p not unready & nomap mbuf\n", m)); KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen, ("page count %d larger than maximum frame length %d", m->m_epg_npgs, ktls_maxlen)); /* Anonymous mbufs are encrypted in place. */ if ((m->m_epg_flags & EPG_FLAG_ANON) != 0) return (ktls_ocf_encrypt(state, tls, m, NULL, 0)); /* * For file-backed mbufs (from sendfile), anonymous wired * pages are allocated and used as the encryption destination. */ if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) { len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len - m->m_epg_1st_off; state->dst_iov[0].iov_base = (char *)state->cbuf + m->m_epg_1st_off; state->dst_iov[0].iov_len = len; state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf); i = 1; } else { off = m->m_epg_1st_off; for (i = 0; i < m->m_epg_npgs; i++, off = 0) { pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP | VM_ALLOC_WIRED | VM_ALLOC_WAITOK); len = m_epg_pagelen(m, i, off); state->parray[i] = VM_PAGE_TO_PHYS(pg); state->dst_iov[i].iov_base = (char *)PHYS_TO_DMAP(state->parray[i]) + off; state->dst_iov[i].iov_len = len; } } KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small")); state->dst_iov[i].iov_base = m->m_epg_trail; state->dst_iov[i].iov_len = m->m_epg_trllen; error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1); if (__predict_false(error != 0)) { /* Free the anonymous pages. */ if (state->cbuf != NULL) uma_zfree(ktls_buffer_zone, state->cbuf); else { for (i = 0; i < m->m_epg_npgs; i++) { pg = PHYS_TO_VM_PAGE(state->parray[i]); (void)vm_page_unwire_noq(pg); vm_page_free(pg); } } } return (error); } /* Number of TLS records in a batch passed to ktls_enqueue(). */ static u_int ktls_batched_records(struct mbuf *m) { int page_count, records; records = 0; page_count = m->m_epg_enc_cnt; while (page_count > 0) { records++; page_count -= m->m_epg_nrdy; m = m->m_next; } KASSERT(page_count == 0, ("%s: mismatched page count", __func__)); return (records); } void ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) { struct ktls_session *tls; struct ktls_wq *wq; int queued; bool running; KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY)), ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf")); m->m_epg_enc_cnt = page_count; /* * Save a pointer to the socket. The caller is responsible * for taking an additional reference via soref(). */ m->m_epg_so = so; queued = 1; tls = m->m_epg_tls; wq = &ktls_wq[tls->wq_index]; mtx_lock(&wq->mtx); if (__predict_false(tls->sequential_records)) { /* * For TLS 1.0, records must be encrypted * sequentially. For a given connection, all records * queued to the associated work queue are processed * sequentially. However, sendfile(2) might complete * I/O requests spanning multiple TLS records out of * order. Here we ensure TLS records are enqueued to * the work queue in FIFO order. * * tls->next_seqno holds the sequence number of the * next TLS record that should be enqueued to the work * queue. If this next record is not tls->next_seqno, * it must be a future record, so insert it, sorted by * TLS sequence number, into tls->pending_records and * return. * * If this TLS record matches tls->next_seqno, place * it in the work queue and then check * tls->pending_records to see if any * previously-queued records are now ready for * encryption. */ if (m->m_epg_seqno != tls->next_seqno) { struct mbuf *n, *p; p = NULL; STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) { if (n->m_epg_seqno > m->m_epg_seqno) break; p = n; } if (n == NULL) STAILQ_INSERT_TAIL(&tls->pending_records, m, m_epg_stailq); else if (p == NULL) STAILQ_INSERT_HEAD(&tls->pending_records, m, m_epg_stailq); else STAILQ_INSERT_AFTER(&tls->pending_records, p, m, m_epg_stailq); mtx_unlock(&wq->mtx); counter_u64_add(ktls_cnt_tx_pending, 1); return; } tls->next_seqno += ktls_batched_records(m); STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); while (!STAILQ_EMPTY(&tls->pending_records)) { struct mbuf *n; n = STAILQ_FIRST(&tls->pending_records); if (n->m_epg_seqno != tls->next_seqno) break; queued++; STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq); tls->next_seqno += ktls_batched_records(n); STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq); } counter_u64_add(ktls_cnt_tx_pending, -(queued - 1)); } else STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); counter_u64_add(ktls_cnt_tx_queued, queued); } /* * Once a file-backed mbuf (from sendfile) has been encrypted, free * the pages from the file and replace them with the anonymous pages * allocated in ktls_encrypt_record(). */ static void ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state) { int i; MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0); /* Free the old pages. */ m->m_ext.ext_free(m); /* Replace them with the new pages. */ if (state->cbuf != NULL) { for (i = 0; i < m->m_epg_npgs; i++) m->m_epg_pa[i] = state->parray[0] + ptoa(i); /* Contig pages should go back to the cache. */ m->m_ext.ext_free = ktls_free_mext_contig; } else { for (i = 0; i < m->m_epg_npgs; i++) m->m_epg_pa[i] = state->parray[i]; /* Use the basic free routine. */ m->m_ext.ext_free = mb_free_mext_pgs; } /* Pages are now writable. */ m->m_epg_flags |= EPG_FLAG_ANON; } static __noinline void ktls_encrypt(struct ktls_wq *wq, struct mbuf *top) { struct ktls_ocf_encrypt_state state; struct ktls_session *tls; struct socket *so; struct mbuf *m; int error, npages, total_pages; so = top->m_epg_so; tls = top->m_epg_tls; KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); #ifdef INVARIANTS top->m_epg_so = NULL; #endif total_pages = top->m_epg_enc_cnt; npages = 0; /* * Encrypt the TLS records in the chain of mbufs starting with * 'top'. 'total_pages' gives us a total count of pages and is * used to know when we have finished encrypting the TLS * records originally queued with 'top'. * * NB: These mbufs are queued in the socket buffer and * 'm_next' is traversing the mbufs in the socket buffer. The * socket buffer lock is not held while traversing this chain. * Since the mbufs are all marked M_NOTREADY their 'm_next' * pointers should be stable. However, the 'm_next' of the * last mbuf encrypted is not necessarily NULL. It can point * to other mbufs appended while 'top' was on the TLS work * queue. * * Each mbuf holds an entire TLS record. */ error = 0; for (m = top; npages != total_pages; m = m->m_next) { KASSERT(m->m_epg_tls == tls, ("different TLS sessions in a single mbuf chain: %p vs %p", tls, m->m_epg_tls)); KASSERT(npages + m->m_epg_npgs <= total_pages, ("page count mismatch: top %p, total_pages %d, m %p", top, total_pages, m)); error = ktls_encrypt_record(wq, m, tls, &state); if (error) { counter_u64_add(ktls_offload_failed_crypto, 1); break; } if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) ktls_finish_nonanon(m, &state); npages += m->m_epg_nrdy; /* * Drop a reference to the session now that it is no * longer needed. Existing code depends on encrypted * records having no associated session vs * yet-to-be-encrypted records having an associated * session. */ m->m_epg_tls = NULL; ktls_free(tls); } CURVNET_SET(so->so_vnet); if (error == 0) { (void)so->so_proto->pr_ready(so, top, npages); } else { ktls_drop(so, EIO); mb_free_notready(top, total_pages); } sorele(so); CURVNET_RESTORE(); } void ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error) { struct ktls_session *tls; struct socket *so; struct mbuf *m; int npages; m = state->m; if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) ktls_finish_nonanon(m, state); so = state->so; free(state, M_KTLS); /* * Drop a reference to the session now that it is no longer * needed. Existing code depends on encrypted records having * no associated session vs yet-to-be-encrypted records having * an associated session. */ tls = m->m_epg_tls; m->m_epg_tls = NULL; ktls_free(tls); if (error != 0) counter_u64_add(ktls_offload_failed_crypto, 1); CURVNET_SET(so->so_vnet); npages = m->m_epg_nrdy; if (error == 0) { (void)so->so_proto->pr_ready(so, m, npages); } else { ktls_drop(so, EIO); mb_free_notready(m, npages); } sorele(so); CURVNET_RESTORE(); } /* * Similar to ktls_encrypt, but used with asynchronous OCF backends * (coprocessors) where encryption does not use host CPU resources and * it can be beneficial to queue more requests than CPUs. */ static __noinline void ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top) { struct ktls_ocf_encrypt_state *state; struct ktls_session *tls; struct socket *so; struct mbuf *m, *n; int error, mpages, npages, total_pages; so = top->m_epg_so; tls = top->m_epg_tls; KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); #ifdef INVARIANTS top->m_epg_so = NULL; #endif total_pages = top->m_epg_enc_cnt; npages = 0; error = 0; for (m = top; npages != total_pages; m = n) { KASSERT(m->m_epg_tls == tls, ("different TLS sessions in a single mbuf chain: %p vs %p", tls, m->m_epg_tls)); KASSERT(npages + m->m_epg_npgs <= total_pages, ("page count mismatch: top %p, total_pages %d, m %p", top, total_pages, m)); state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO); soref(so); state->so = so; state->m = m; mpages = m->m_epg_nrdy; n = m->m_next; error = ktls_encrypt_record(wq, m, tls, state); if (error) { counter_u64_add(ktls_offload_failed_crypto, 1); free(state, M_KTLS); CURVNET_SET(so->so_vnet); sorele(so); CURVNET_RESTORE(); break; } npages += mpages; } CURVNET_SET(so->so_vnet); if (error != 0) { ktls_drop(so, EIO); mb_free_notready(m, total_pages - npages); } sorele(so); CURVNET_RESTORE(); } static int ktls_bind_domain(int domain) { int error; error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]); if (error != 0) return (error); curthread->td_domain.dr_policy = DOMAINSET_PREF(domain); return (0); } static void ktls_reclaim_thread(void *ctx) { struct ktls_domain_info *ktls_domain = ctx; struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td; struct sysctl_oid *oid; char name[80]; int error, domain; domain = ktls_domain - ktls_domains; if (bootverbose) printf("Starting KTLS reclaim thread for domain %d\n", domain); error = ktls_bind_domain(domain); if (error) printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n", domain, error); snprintf(name, sizeof(name), "domain%d", domain); oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims", CTLFLAG_RD, &sc->reclaims, 0, "buffers reclaimed"); SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups", CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups"); SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running", CTLFLAG_RD, &sc->running, 0, "thread running"); for (;;) { atomic_store_int(&sc->running, 0); tsleep(sc, PZERO | PNOLOCK, "-", 0); atomic_store_int(&sc->running, 1); sc->wakeups++; /* * Below we attempt to reclaim ktls_max_reclaim * buffers using vm_page_reclaim_contig_domain_ext(). * We do this here, as this function can take several * seconds to scan all of memory and it does not * matter if this thread pauses for a while. If we * block a ktls worker thread, we risk developing * backlogs of buffers to be encrypted, leading to * surges of traffic and potential NIC output drops. */ - if (!vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL, - atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, ktls_max_reclaim)) { + if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL, + atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, + ktls_max_reclaim) != 0) { vm_wait_domain(domain); } else { sc->reclaims += ktls_max_reclaim; } } } static void ktls_work_thread(void *ctx) { struct ktls_wq *wq = ctx; struct mbuf *m, *n; struct socket *so, *son; STAILQ_HEAD(, mbuf) local_m_head; STAILQ_HEAD(, socket) local_so_head; int cpu; cpu = wq - ktls_wq; if (bootverbose) printf("Starting KTLS worker thread for CPU %d\n", cpu); /* * Bind to a core. If ktls_bind_threads is > 1, then * we bind to the NUMA domain instead. */ if (ktls_bind_threads) { int error; if (ktls_bind_threads > 1) { struct pcpu *pc = pcpu_find(cpu); error = ktls_bind_domain(pc->pc_domain); } else { cpuset_t mask; CPU_SETOF(cpu, &mask); error = cpuset_setthread(curthread->td_tid, &mask); } if (error) printf("Unable to bind KTLS worker thread for CPU %d: error %d\n", cpu, error); } #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) fpu_kern_thread(0); #endif for (;;) { mtx_lock(&wq->mtx); while (STAILQ_EMPTY(&wq->m_head) && STAILQ_EMPTY(&wq->so_head)) { wq->running = false; mtx_sleep(wq, &wq->mtx, 0, "-", 0); wq->running = true; } STAILQ_INIT(&local_m_head); STAILQ_CONCAT(&local_m_head, &wq->m_head); STAILQ_INIT(&local_so_head); STAILQ_CONCAT(&local_so_head, &wq->so_head); mtx_unlock(&wq->mtx); STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) { if (m->m_epg_flags & EPG_FLAG_2FREE) { ktls_free(m->m_epg_tls); m_free_raw(m); } else { if (m->m_epg_tls->sync_dispatch) ktls_encrypt(wq, m); else ktls_encrypt_async(wq, m); counter_u64_add(ktls_cnt_tx_queued, -1); } } STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) { ktls_decrypt(so); counter_u64_add(ktls_cnt_rx_queued, -1); } } } static void ktls_disable_ifnet_help(void *context, int pending __unused) { struct ktls_session *tls; struct inpcb *inp; struct tcpcb *tp; struct socket *so; int err; tls = context; inp = tls->inp; if (inp == NULL) return; INP_WLOCK(inp); so = inp->inp_socket; MPASS(so != NULL); if (inp->inp_flags & INP_DROPPED) { goto out; } if (so->so_snd.sb_tls_info != NULL) err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW); else err = ENXIO; if (err == 0) { counter_u64_add(ktls_ifnet_disable_ok, 1); /* ktls_set_tx_mode() drops inp wlock, so recheck flags */ if ((inp->inp_flags & INP_DROPPED) == 0 && (tp = intotcpcb(inp)) != NULL && tp->t_fb->tfb_hwtls_change != NULL) (*tp->t_fb->tfb_hwtls_change)(tp, 0); } else { counter_u64_add(ktls_ifnet_disable_fail, 1); } out: CURVNET_SET(so->so_vnet); sorele(so); CURVNET_RESTORE(); INP_WUNLOCK(inp); ktls_free(tls); } /* * Called when re-transmits are becoming a substantial portion of the * sends on this connection. When this happens, we transition the * connection to software TLS. This is needed because most inline TLS * NICs keep crypto state only for in-order transmits. This means * that to handle a TCP rexmit (which is out-of-order), the NIC must * re-DMA the entire TLS record up to and including the current * segment. This means that when re-transmitting the last ~1448 byte * segment of a 16KB TLS record, we could wind up re-DMA'ing an order * of magnitude more data than we are sending. This can cause the * PCIe link to saturate well before the network, which can cause * output drops, and a general loss of capacity. */ void ktls_disable_ifnet(void *arg) { struct tcpcb *tp; struct inpcb *inp; struct socket *so; struct ktls_session *tls; tp = arg; inp = tptoinpcb(tp); INP_WLOCK_ASSERT(inp); so = inp->inp_socket; SOCK_LOCK(so); tls = so->so_snd.sb_tls_info; if (tp->t_nic_ktls_xmit_dis == 1) { SOCK_UNLOCK(so); return; } /* * note that t_nic_ktls_xmit_dis is never cleared; disabling * ifnet can only be done once per connection, so we never want * to do it again */ (void)ktls_hold(tls); soref(so); tp->t_nic_ktls_xmit_dis = 1; SOCK_UNLOCK(so); TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls); (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task); } diff --git a/sys/kern/uipc_shm.c b/sys/kern/uipc_shm.c index c8d5521c5ab1..a07dd097f7b8 100644 --- a/sys/kern/uipc_shm.c +++ b/sys/kern/uipc_shm.c @@ -1,2250 +1,2251 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson * Copyright 2020 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by BAE Systems, the University of * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent * Computing (TC) research program. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * Support for shared swap-backed anonymous memory objects via * shm_open(2), shm_rename(2), and shm_unlink(2). * While most of the implementation is here, vm_mmap.c contains * mapping logic changes. * * posixshmcontrol(1) allows users to inspect the state of the memory * objects. Per-uid swap resource limit controls total amount of * memory that user can consume for anonymous objects, including * shared. */ #include #include "opt_capsicum.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct shm_mapping { char *sm_path; Fnv32_t sm_fnv; struct shmfd *sm_shmfd; LIST_ENTRY(shm_mapping) sm_link; }; static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor"); static LIST_HEAD(, shm_mapping) *shm_dictionary; static struct sx shm_dict_lock; static struct mtx shm_timestamp_lock; static u_long shm_hash; static struct unrhdr64 shm_ino_unr; static dev_t shm_dev_ino; #define SHM_HASH(fnv) (&shm_dictionary[(fnv) & shm_hash]) static void shm_init(void *arg); static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd); static struct shmfd *shm_lookup(char *path, Fnv32_t fnv); static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred); static void shm_doremove(struct shm_mapping *map); static int shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie); static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie); static int shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out); static int shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags); static fo_rdwr_t shm_read; static fo_rdwr_t shm_write; static fo_truncate_t shm_truncate; static fo_ioctl_t shm_ioctl; static fo_stat_t shm_stat; static fo_close_t shm_close; static fo_chmod_t shm_chmod; static fo_chown_t shm_chown; static fo_seek_t shm_seek; static fo_fill_kinfo_t shm_fill_kinfo; static fo_mmap_t shm_mmap; static fo_get_seals_t shm_get_seals; static fo_add_seals_t shm_add_seals; static fo_fallocate_t shm_fallocate; static fo_fspacectl_t shm_fspacectl; /* File descriptor operations. */ const struct fileops shm_ops = { .fo_read = shm_read, .fo_write = shm_write, .fo_truncate = shm_truncate, .fo_ioctl = shm_ioctl, .fo_poll = invfo_poll, .fo_kqfilter = invfo_kqfilter, .fo_stat = shm_stat, .fo_close = shm_close, .fo_chmod = shm_chmod, .fo_chown = shm_chown, .fo_sendfile = vn_sendfile, .fo_seek = shm_seek, .fo_fill_kinfo = shm_fill_kinfo, .fo_mmap = shm_mmap, .fo_get_seals = shm_get_seals, .fo_add_seals = shm_add_seals, .fo_fallocate = shm_fallocate, .fo_fspacectl = shm_fspacectl, .fo_cmp = file_kcmp_generic, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE, }; FEATURE(posix_shm, "POSIX shared memory"); static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, ""); static int largepage_reclaim_tries = 1; SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries, CTLFLAG_RWTUN, &largepage_reclaim_tries, 0, "Number of contig reclaims before giving up for default alloc policy"); #define shm_rangelock_unlock(shmfd, cookie) \ rangelock_unlock(&(shmfd)->shm_rl, (cookie), &(shmfd)->shm_mtx) #define shm_rangelock_rlock(shmfd, start, end) \ rangelock_rlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx) #define shm_rangelock_tryrlock(shmfd, start, end) \ rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx) #define shm_rangelock_wlock(shmfd, start, end) \ rangelock_wlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx) static int uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio) { vm_page_t m; vm_pindex_t idx; size_t tlen; int error, offset, rv; idx = OFF_TO_IDX(uio->uio_offset); offset = uio->uio_offset & PAGE_MASK; tlen = MIN(PAGE_SIZE - offset, len); rv = vm_page_grab_valid_unlocked(&m, obj, idx, VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT); if (rv == VM_PAGER_OK) goto found; /* * Read I/O without either a corresponding resident page or swap * page: use zero_region. This is intended to avoid instantiating * pages on read from a sparse region. */ VM_OBJECT_WLOCK(obj); m = vm_page_lookup(obj, idx); if (uio->uio_rw == UIO_READ && m == NULL && !vm_pager_has_page(obj, idx, NULL, NULL)) { VM_OBJECT_WUNLOCK(obj); return (uiomove(__DECONST(void *, zero_region), tlen, uio)); } /* * Although the tmpfs vnode lock is held here, it is * nonetheless safe to sleep waiting for a free page. The * pageout daemon does not need to acquire the tmpfs vnode * lock to page out tobj's pages because tobj is a OBJT_SWAP * type object. */ rv = vm_page_grab_valid(&m, obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); if (rv != VM_PAGER_OK) { VM_OBJECT_WUNLOCK(obj); if (bootverbose) { printf("uiomove_object: vm_obj %p idx %jd " "pager error %d\n", obj, idx, rv); } return (rv == VM_PAGER_AGAIN ? ENOSPC : EIO); } VM_OBJECT_WUNLOCK(obj); found: error = uiomove_fromphys(&m, offset, tlen, uio); if (uio->uio_rw == UIO_WRITE && error == 0) vm_page_set_dirty(m); vm_page_activate(m); vm_page_sunbusy(m); return (error); } int uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio) { ssize_t resid; size_t len; int error; error = 0; while ((resid = uio->uio_resid) > 0) { if (obj_size <= uio->uio_offset) break; len = MIN(obj_size - uio->uio_offset, resid); if (len == 0) break; error = uiomove_object_page(obj, len, uio); if (error != 0 || resid == uio->uio_resid) break; } return (error); } static u_long count_largepages[MAXPAGESIZES]; static int shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx, int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last) { vm_page_t m __diagused; int psind; psind = object->un_pager.phys.data_val; if (psind == 0 || pidx >= object->size) return (VM_PAGER_FAIL); *first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE); /* * We only busy the first page in the superpage run. It is * useless to busy whole run since we only remove full * superpage, and it takes too long to busy e.g. 512 * 512 == * 262144 pages constituing 1G amd64 superage. */ m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT); MPASS(m != NULL); *last = *first + atop(pagesizes[psind]) - 1; return (VM_PAGER_OK); } static boolean_t shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { int psind; psind = object->un_pager.phys.data_val; if (psind == 0 || pindex >= object->size) return (FALSE); if (before != NULL) { *before = pindex - rounddown2(pindex, pagesizes[psind] / PAGE_SIZE); } if (after != NULL) { *after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) - pindex; } return (TRUE); } static void shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred) { } static void shm_largepage_phys_dtor(vm_object_t object) { int psind; psind = object->un_pager.phys.data_val; if (psind != 0) { atomic_subtract_long(&count_largepages[psind], object->size / (pagesizes[psind] / PAGE_SIZE)); vm_wire_sub(object->size); } else { KASSERT(object->size == 0, ("largepage phys obj %p not initialized bit size %#jx > 0", object, (uintmax_t)object->size)); } } static const struct phys_pager_ops shm_largepage_phys_ops = { .phys_pg_populate = shm_largepage_phys_populate, .phys_pg_haspage = shm_largepage_phys_haspage, .phys_pg_ctor = shm_largepage_phys_ctor, .phys_pg_dtor = shm_largepage_phys_dtor, }; bool shm_largepage(struct shmfd *shmfd) { return (shmfd->shm_object->type == OBJT_PHYS); } static void shm_pager_freespace(vm_object_t obj, vm_pindex_t start, vm_size_t size) { struct shmfd *shm; vm_size_t c; swap_pager_freespace(obj, start, size, &c); if (c == 0) return; shm = obj->un_pager.swp.swp_priv; if (shm == NULL) return; KASSERT(shm->shm_pages >= c, ("shm %p pages %jd free %jd", shm, (uintmax_t)shm->shm_pages, (uintmax_t)c)); shm->shm_pages -= c; } static void shm_page_inserted(vm_object_t obj, vm_page_t m) { struct shmfd *shm; shm = obj->un_pager.swp.swp_priv; if (shm == NULL) return; if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) shm->shm_pages += 1; } static void shm_page_removed(vm_object_t obj, vm_page_t m) { struct shmfd *shm; shm = obj->un_pager.swp.swp_priv; if (shm == NULL) return; if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) { KASSERT(shm->shm_pages >= 1, ("shm %p pages %jd free 1", shm, (uintmax_t)shm->shm_pages)); shm->shm_pages -= 1; } } static struct pagerops shm_swap_pager_ops = { .pgo_kvme_type = KVME_TYPE_SWAP, .pgo_freespace = shm_pager_freespace, .pgo_page_inserted = shm_page_inserted, .pgo_page_removed = shm_page_removed, }; static int shmfd_pager_type = -1; static int shm_seek(struct file *fp, off_t offset, int whence, struct thread *td) { struct shmfd *shmfd; off_t foffset; int error; shmfd = fp->f_data; foffset = foffset_lock(fp, 0); error = 0; switch (whence) { case L_INCR: if (foffset < 0 || (offset > 0 && foffset > OFF_MAX - offset)) { error = EOVERFLOW; break; } offset += foffset; break; case L_XTND: if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) { error = EOVERFLOW; break; } offset += shmfd->shm_size; break; case L_SET: break; default: error = EINVAL; } if (error == 0) { if (offset < 0 || offset > shmfd->shm_size) error = EINVAL; else td->td_uretoff.tdu_off = offset; } foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); return (error); } static int shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif foffset_lock_uio(fp, uio, flags); rl_cookie = shm_rangelock_rlock(shmfd, uio->uio_offset, uio->uio_offset + uio->uio_resid); error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); shm_rangelock_unlock(shmfd, rl_cookie); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; off_t newsize; KASSERT((flags & FOF_OFFSET) == 0 || uio->uio_offset >= 0, ("%s: negative offset", __func__)); shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0) return (EINVAL); foffset_lock_uio(fp, uio, flags); if (uio->uio_resid > OFF_MAX - uio->uio_offset) { /* * Overflow is only an error if we're supposed to expand on * write. Otherwise, we'll just truncate the write to the * size of the file, which can only grow up to OFF_MAX. */ if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) { foffset_unlock_uio(fp, uio, flags); return (EFBIG); } newsize = atomic_load_64(&shmfd->shm_size); } else { newsize = uio->uio_offset + uio->uio_resid; } if ((flags & FOF_OFFSET) == 0) rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); else rl_cookie = shm_rangelock_wlock(shmfd, uio->uio_offset, MAX(newsize, uio->uio_offset)); if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) { error = EPERM; } else { error = 0; if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 && newsize > shmfd->shm_size) { error = shm_dotruncate_cookie(shmfd, newsize, rl_cookie); } if (error == 0) error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); } shm_rangelock_unlock(shmfd, rl_cookie); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_truncate(struct file *fp, off_t length, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif return (shm_dotruncate(shmfd, length)); } int shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; struct shm_largepage_conf *conf; void *rl_cookie; shmfd = fp->f_data; switch (com) { case FIONBIO: case FIOASYNC: /* * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work, * just like it would on an unlinked regular file */ return (0); case FIOSSHMLPGCNF: if (!shm_largepage(shmfd)) return (ENOTTY); conf = data; if (shmfd->shm_lp_psind != 0 && conf->psind != shmfd->shm_lp_psind) return (EINVAL); if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES || pagesizes[conf->psind] == 0) return (EINVAL); if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT && conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT && conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD) return (EINVAL); rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); shmfd->shm_lp_psind = conf->psind; shmfd->shm_lp_alloc_policy = conf->alloc_policy; shmfd->shm_object->un_pager.phys.data_val = conf->psind; shm_rangelock_unlock(shmfd, rl_cookie); return (0); case FIOGSHMLPGCNF: if (!shm_largepage(shmfd)) return (ENOTTY); conf = data; rl_cookie = shm_rangelock_rlock(shmfd, 0, OFF_MAX); conf->psind = shmfd->shm_lp_psind; conf->alloc_policy = shmfd->shm_lp_alloc_policy; shm_rangelock_unlock(shmfd, rl_cookie); return (0); default: return (ENOTTY); } } static int shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif /* * Attempt to return sanish values for fstat() on a memory file * descriptor. */ bzero(sb, sizeof(*sb)); sb->st_blksize = PAGE_SIZE; sb->st_size = shmfd->shm_size; mtx_lock(&shm_timestamp_lock); sb->st_atim = shmfd->shm_atime; sb->st_ctim = shmfd->shm_ctime; sb->st_mtim = shmfd->shm_mtime; sb->st_birthtim = shmfd->shm_birthtime; sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */ sb->st_uid = shmfd->shm_uid; sb->st_gid = shmfd->shm_gid; mtx_unlock(&shm_timestamp_lock); sb->st_dev = shm_dev_ino; sb->st_ino = shmfd->shm_ino; sb->st_nlink = shmfd->shm_object->ref_count; if (shm_largepage(shmfd)) { sb->st_blocks = shmfd->shm_object->size / (pagesizes[shmfd->shm_lp_psind] >> PAGE_SHIFT); } else { sb->st_blocks = shmfd->shm_pages; } return (0); } static int shm_close(struct file *fp, struct thread *td) { struct shmfd *shmfd; shmfd = fp->f_data; fp->f_data = NULL; shm_drop(shmfd); return (0); } static int shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) { int error; char *path; const char *pr_path; size_t pr_pathlen; path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); pr_path = td->td_ucred->cr_prison->pr_path; /* Construct a full pathname for jailed callers. */ pr_pathlen = strcmp(pr_path, "/") == 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); error = copyinstr(userpath_in, path + pr_pathlen, MAXPATHLEN - pr_pathlen, NULL); if (error != 0) goto out; #ifdef KTRACE if (KTRPOINT(curthread, KTR_NAMEI)) ktrnamei(path); #endif /* Require paths to start with a '/' character. */ if (path[pr_pathlen] != '/') { error = EINVAL; goto out; } *path_out = path; out: if (error != 0) free(path, M_SHMFD); return (error); } static int shm_partial_page_invalidate(vm_object_t object, vm_pindex_t idx, int base, int end) { vm_page_t m; int rv; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(base >= 0, ("%s: base %d", __func__, base)); KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base, end)); retry: m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT); if (m != NULL) { MPASS(vm_page_all_valid(m)); } else if (vm_pager_has_page(object, idx, NULL, NULL)) { m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); if (m == NULL) goto retry; vm_object_pip_add(object, 1); VM_OBJECT_WUNLOCK(object); rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); VM_OBJECT_WLOCK(object); vm_object_pip_wakeup(object); if (rv == VM_PAGER_OK) { /* * Since the page was not resident, and therefore not * recently accessed, immediately enqueue it for * asynchronous laundering. The current operation is * not regarded as an access. */ vm_page_launder(m); } else { vm_page_free(m); VM_OBJECT_WUNLOCK(object); return (EIO); } } if (m != NULL) { pmap_zero_page_area(m, base, end - base); KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m)); vm_page_set_dirty(m); vm_page_xunbusy(m); } return (0); } static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie) { vm_object_t object; vm_pindex_t nobjsize; vm_ooffset_t delta; int base, error; KASSERT(length >= 0, ("shm_dotruncate: length < 0")); object = shmfd->shm_object; VM_OBJECT_ASSERT_WLOCKED(object); rangelock_cookie_assert(rl_cookie, RA_WLOCKED); if (length == shmfd->shm_size) return (0); nobjsize = OFF_TO_IDX(length + PAGE_MASK); /* Are we shrinking? If so, trim the end. */ if (length < shmfd->shm_size) { if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) return (EPERM); /* * Disallow any requests to shrink the size if this * object is mapped into the kernel. */ if (shmfd->shm_kmappings > 0) return (EBUSY); /* * Zero the truncated part of the last page. */ base = length & PAGE_MASK; if (base != 0) { error = shm_partial_page_invalidate(object, OFF_TO_IDX(length), base, PAGE_SIZE); if (error) return (error); } delta = IDX_TO_OFF(object->size - nobjsize); if (nobjsize < object->size) vm_object_page_remove(object, nobjsize, object->size, 0); /* Free the swap accounted for shm */ swap_release_by_cred(delta, object->cred); object->charge -= delta; } else { if ((shmfd->shm_seals & F_SEAL_GROW) != 0) return (EPERM); /* Try to reserve additional swap space. */ delta = IDX_TO_OFF(nobjsize - object->size); if (!swap_reserve_by_cred(delta, object->cred)) return (ENOMEM); object->charge += delta; } shmfd->shm_size = length; mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_ctime); shmfd->shm_mtime = shmfd->shm_ctime; mtx_unlock(&shm_timestamp_lock); object->size = nobjsize; return (0); } static int shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie) { vm_object_t object; vm_page_t m; vm_pindex_t newobjsz; vm_pindex_t oldobjsz __unused; int aflags, error, i, psind, try; KASSERT(length >= 0, ("shm_dotruncate: length < 0")); object = shmfd->shm_object; VM_OBJECT_ASSERT_WLOCKED(object); rangelock_cookie_assert(rl_cookie, RA_WLOCKED); oldobjsz = object->size; newobjsz = OFF_TO_IDX(length); if (length == shmfd->shm_size) return (0); psind = shmfd->shm_lp_psind; if (psind == 0 && length != 0) return (EINVAL); if ((length & (pagesizes[psind] - 1)) != 0) return (EINVAL); if (length < shmfd->shm_size) { if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) return (EPERM); if (shmfd->shm_kmappings > 0) return (EBUSY); return (ENOTSUP); /* Pages are unmanaged. */ #if 0 vm_object_page_remove(object, newobjsz, oldobjsz, 0); object->size = newobjsz; shmfd->shm_size = length; return (0); #endif } if ((shmfd->shm_seals & F_SEAL_GROW) != 0) return (EPERM); aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO; if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT) aflags |= VM_ALLOC_WAITFAIL; try = 0; /* * Extend shmfd and object, keeping all already fully * allocated large pages intact even on error, because dropped * object lock might allowed mapping of them. */ while (object->size < newobjsz) { m = vm_page_alloc_contig(object, object->size, aflags, pagesizes[psind] / PAGE_SIZE, 0, ~0, pagesizes[psind], 0, VM_MEMATTR_DEFAULT); if (m == NULL) { VM_OBJECT_WUNLOCK(object); if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT || (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_DEFAULT && try >= largepage_reclaim_tries)) { VM_OBJECT_WLOCK(object); return (ENOMEM); } error = vm_page_reclaim_contig(aflags, pagesizes[psind] / PAGE_SIZE, 0, ~0, - pagesizes[psind], 0) ? 0 : - vm_wait_intr(object); + pagesizes[psind], 0); + if (error == ENOMEM) + error = vm_wait_intr(object); if (error != 0) { VM_OBJECT_WLOCK(object); return (error); } try++; VM_OBJECT_WLOCK(object); continue; } try = 0; for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) { if ((m[i].flags & PG_ZERO) == 0) pmap_zero_page(&m[i]); vm_page_valid(&m[i]); vm_page_xunbusy(&m[i]); } object->size += OFF_TO_IDX(pagesizes[psind]); shmfd->shm_size += pagesizes[psind]; atomic_add_long(&count_largepages[psind], 1); vm_wire_add(atop(pagesizes[psind])); } return (0); } static int shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie) { int error; VM_OBJECT_WLOCK(shmfd->shm_object); error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd, length, rl_cookie) : shm_dotruncate_locked(shmfd, length, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); return (error); } int shm_dotruncate(struct shmfd *shmfd, off_t length) { void *rl_cookie; int error; rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); error = shm_dotruncate_cookie(shmfd, length, rl_cookie); shm_rangelock_unlock(shmfd, rl_cookie); return (error); } /* * shmfd object management including creation and reference counting * routines. */ struct shmfd * shm_alloc(struct ucred *ucred, mode_t mode, bool largepage) { struct shmfd *shmfd; vm_object_t obj; if (largepage) { obj = phys_pager_allocate(NULL, &shm_largepage_phys_ops, NULL, 0, VM_PROT_DEFAULT, 0, ucred); } else { obj = vm_pager_allocate(shmfd_pager_type, NULL, 0, VM_PROT_DEFAULT, 0, ucred); } if (obj == NULL) { /* * swap reservation limits can cause object allocation * to fail. */ return (NULL); } shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); shmfd->shm_uid = ucred->cr_uid; shmfd->shm_gid = ucred->cr_gid; shmfd->shm_mode = mode; if (largepage) { obj->un_pager.phys.phys_priv = shmfd; shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT; } else { obj->un_pager.swp.swp_priv = shmfd; } VM_OBJECT_WLOCK(obj); vm_object_set_flag(obj, OBJ_POSIXSHM); VM_OBJECT_WUNLOCK(obj); shmfd->shm_object = obj; vfs_timestamp(&shmfd->shm_birthtime); shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = shmfd->shm_birthtime; shmfd->shm_ino = alloc_unr64(&shm_ino_unr); refcount_init(&shmfd->shm_refs, 1); mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); rangelock_init(&shmfd->shm_rl); #ifdef MAC mac_posixshm_init(shmfd); mac_posixshm_create(ucred, shmfd); #endif return (shmfd); } struct shmfd * shm_hold(struct shmfd *shmfd) { refcount_acquire(&shmfd->shm_refs); return (shmfd); } void shm_drop(struct shmfd *shmfd) { vm_object_t obj; if (refcount_release(&shmfd->shm_refs)) { #ifdef MAC mac_posixshm_destroy(shmfd); #endif rangelock_destroy(&shmfd->shm_rl); mtx_destroy(&shmfd->shm_mtx); obj = shmfd->shm_object; VM_OBJECT_WLOCK(obj); if (shm_largepage(shmfd)) obj->un_pager.phys.phys_priv = NULL; else obj->un_pager.swp.swp_priv = NULL; VM_OBJECT_WUNLOCK(obj); vm_object_deallocate(obj); free(shmfd, M_SHMFD); } } /* * Determine if the credentials have sufficient permissions for a * specified combination of FREAD and FWRITE. */ int shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) { accmode_t accmode; int error; accmode = 0; if (flags & FREAD) accmode |= VREAD; if (flags & FWRITE) accmode |= VWRITE; mtx_lock(&shm_timestamp_lock); error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, accmode, ucred); mtx_unlock(&shm_timestamp_lock); return (error); } static void shm_init(void *arg) { char name[32]; int i; mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); sx_init(&shm_dict_lock, "shm dictionary"); shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); new_unrhdr64(&shm_ino_unr, 1); shm_dev_ino = devfs_alloc_cdp_inode(); KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops, OBJT_SWAP); MPASS(shmfd_pager_type != -1); for (i = 1; i < MAXPAGESIZES; i++) { if (pagesizes[i] == 0) break; #define M (1024 * 1024) #define G (1024 * M) if (pagesizes[i] >= G) snprintf(name, sizeof(name), "%luG", pagesizes[i] / G); else if (pagesizes[i] >= M) snprintf(name, sizeof(name), "%luM", pagesizes[i] / M); else snprintf(name, sizeof(name), "%lu", pagesizes[i]); #undef G #undef M SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages), OID_AUTO, name, CTLFLAG_RD, &count_largepages[i], "number of non-transient largepages allocated"); } } SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); /* * Remove all shared memory objects that belong to a prison. */ void shm_remove_prison(struct prison *pr) { struct shm_mapping *shmm, *tshmm; u_long i; sx_xlock(&shm_dict_lock); for (i = 0; i < shm_hash + 1; i++) { LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) { if (shmm->sm_shmfd->shm_object->cred && shmm->sm_shmfd->shm_object->cred->cr_prison == pr) shm_doremove(shmm); } } sx_xunlock(&shm_dict_lock); } /* * Dictionary management. We maintain an in-kernel dictionary to map * paths to shmfd objects. We use the FNV hash on the path to store * the mappings in a hash table. */ static struct shmfd * shm_lookup(char *path, Fnv32_t fnv) { struct shm_mapping *map; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) return (map->sm_shmfd); } return (NULL); } static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) { struct shm_mapping *map; map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); map->sm_path = path; map->sm_fnv = fnv; map->sm_shmfd = shm_hold(shmfd); shmfd->shm_path = path; LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); } static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) { struct shm_mapping *map; int error; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) { #ifdef MAC error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); if (error) return (error); #endif error = shm_access(map->sm_shmfd, ucred, FREAD | FWRITE); if (error) return (error); shm_doremove(map); return (0); } } return (ENOENT); } static void shm_doremove(struct shm_mapping *map) { map->sm_shmfd->shm_path = NULL; LIST_REMOVE(map, sm_link); shm_drop(map->sm_shmfd); free(map->sm_path, M_SHMFD); free(map, M_SHMFD); } int kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode, int shmflags, struct filecaps *fcaps, const char *name __unused) { struct pwddesc *pdp; struct shmfd *shmfd; struct file *fp; char *path; void *rl_cookie; Fnv32_t fnv; mode_t cmode; int error, fd, initial_seals; bool largepage; if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE | SHM_LARGEPAGE)) != 0) return (EINVAL); initial_seals = F_SEAL_SEAL; if ((shmflags & SHM_ALLOW_SEALING) != 0) initial_seals &= ~F_SEAL_SEAL; AUDIT_ARG_FFLAGS(flags); AUDIT_ARG_MODE(mode); if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) return (EINVAL); if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0) return (EINVAL); largepage = (shmflags & SHM_LARGEPAGE) != 0; if (largepage && !PMAP_HAS_LARGEPAGES) return (ENOTTY); /* * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. * If the decision is made later to allow additional seals, care must be * taken below to ensure that the seals are properly set if the shmfd * already existed -- this currently assumes that only F_SEAL_SEAL can * be set and doesn't take further precautions to ensure the validity of * the seals being added with respect to current mappings. */ if ((initial_seals & ~F_SEAL_SEAL) != 0) return (EINVAL); if (userpath != SHM_ANON) { error = shm_copyin_path(td, userpath, &path); if (error != 0) return (error); #ifdef CAPABILITY_MODE /* * shm_open(2) is only allowed for anonymous objects. */ if (CAP_TRACING(td)) ktrcapfail(CAPFAIL_NAMEI, path); if (IN_CAPABILITY_MODE(td)) { error = ECAPMODE; goto outnofp; } #endif AUDIT_ARG_UPATH1_CANON(path); } else { path = NULL; } pdp = td->td_proc->p_pd; cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS; /* * shm_open(2) created shm should always have O_CLOEXEC set, as mandated * by POSIX. We allow it to be unset here so that an in-kernel * interface may be written as a thin layer around shm, optionally not * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally * in sys_shm_open() to keep this implementation compliant. */ error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); if (error != 0) goto outnofp; /* A SHM_ANON path pointer creates an anonymous object. */ if (userpath == SHM_ANON) { /* A read-only anonymous object is pointless. */ if ((flags & O_ACCMODE) == O_RDONLY) { error = EINVAL; goto out; } shmfd = shm_alloc(td->td_ucred, cmode, largepage); if (shmfd == NULL) { error = ENOMEM; goto out; } shmfd->shm_seals = initial_seals; shmfd->shm_flags = shmflags; } else { fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); shmfd = shm_lookup(path, fnv); if (shmfd == NULL) { /* Object does not yet exist, create it if requested. */ if (flags & O_CREAT) { #ifdef MAC error = mac_posixshm_check_create(td->td_ucred, path); if (error == 0) { #endif shmfd = shm_alloc(td->td_ucred, cmode, largepage); if (shmfd == NULL) { error = ENOMEM; } else { shmfd->shm_seals = initial_seals; shmfd->shm_flags = shmflags; shm_insert(path, fnv, shmfd); path = NULL; } #ifdef MAC } #endif } else { error = ENOENT; } } else { /* * Object already exists, obtain a new reference if * requested and permitted. */ rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); /* * kern_shm_open() likely shouldn't ever error out on * trying to set a seal that already exists, unlike * F_ADD_SEALS. This would break terribly as * shm_open(2) actually sets F_SEAL_SEAL to maintain * historical behavior where the underlying file could * not be sealed. */ initial_seals &= ~shmfd->shm_seals; /* * initial_seals can't set additional seals if we've * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, * then we've already removed that one from * initial_seals. This is currently redundant as we * only allow setting F_SEAL_SEAL at creation time, but * it's cheap to check and decreases the effort required * to allow additional seals. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && initial_seals != 0) error = EPERM; else if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) error = EEXIST; else if (shmflags != 0 && shmflags != shmfd->shm_flags) error = EINVAL; else { #ifdef MAC error = mac_posixshm_check_open(td->td_ucred, shmfd, FFLAGS(flags & O_ACCMODE)); if (error == 0) #endif error = shm_access(shmfd, td->td_ucred, FFLAGS(flags & O_ACCMODE)); } /* * Truncate the file back to zero length if * O_TRUNC was specified and the object was * opened with read/write. */ if (error == 0 && (flags & (O_ACCMODE | O_TRUNC)) == (O_RDWR | O_TRUNC)) { VM_OBJECT_WLOCK(shmfd->shm_object); #ifdef MAC error = mac_posixshm_check_truncate( td->td_ucred, fp->f_cred, shmfd); if (error == 0) #endif error = shm_dotruncate_locked(shmfd, 0, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); } if (error == 0) { /* * Currently we only allow F_SEAL_SEAL to be * set initially. As noted above, this would * need to be reworked should that change. */ shmfd->shm_seals |= initial_seals; shm_hold(shmfd); } shm_rangelock_unlock(shmfd, rl_cookie); } sx_xunlock(&shm_dict_lock); if (error != 0) goto out; } finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); td->td_retval[0] = fd; fdrop(fp, td); free(path, M_SHMFD); return (0); out: fdclose(td, fp, fd); fdrop(fp, td); outnofp: free(path, M_SHMFD); return (error); } /* System calls. */ #ifdef COMPAT_FREEBSD12 int freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) { return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode, NULL)); } #endif int sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) { char *path; Fnv32_t fnv; int error; error = shm_copyin_path(td, uap->path, &path); if (error != 0) return (error); AUDIT_ARG_UPATH1_CANON(path); fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); error = shm_remove(path, fnv, td->td_ucred); sx_xunlock(&shm_dict_lock); free(path, M_SHMFD); return (error); } int sys_shm_rename(struct thread *td, struct shm_rename_args *uap) { char *path_from = NULL, *path_to = NULL; Fnv32_t fnv_from, fnv_to; struct shmfd *fd_from; struct shmfd *fd_to; int error; int flags; flags = uap->flags; AUDIT_ARG_FFLAGS(flags); /* * Make sure the user passed only valid flags. * If you add a new flag, please add a new term here. */ if ((flags & ~( SHM_RENAME_NOREPLACE | SHM_RENAME_EXCHANGE )) != 0) { error = EINVAL; goto out; } /* * EXCHANGE and NOREPLACE don't quite make sense together. Let's * force the user to choose one or the other. */ if ((flags & SHM_RENAME_NOREPLACE) != 0 && (flags & SHM_RENAME_EXCHANGE) != 0) { error = EINVAL; goto out; } /* Renaming to or from anonymous makes no sense */ if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) { error = EINVAL; goto out; } error = shm_copyin_path(td, uap->path_from, &path_from); if (error != 0) goto out; error = shm_copyin_path(td, uap->path_to, &path_to); if (error != 0) goto out; AUDIT_ARG_UPATH1_CANON(path_from); AUDIT_ARG_UPATH2_CANON(path_to); /* Rename with from/to equal is a no-op */ if (strcmp(path_from, path_to) == 0) goto out; fnv_from = fnv_32_str(path_from, FNV1_32_INIT); fnv_to = fnv_32_str(path_to, FNV1_32_INIT); sx_xlock(&shm_dict_lock); fd_from = shm_lookup(path_from, fnv_from); if (fd_from == NULL) { error = ENOENT; goto out_locked; } fd_to = shm_lookup(path_to, fnv_to); if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { error = EEXIST; goto out_locked; } /* * Unconditionally prevents shm_remove from invalidating the 'from' * shm's state. */ shm_hold(fd_from); error = shm_remove(path_from, fnv_from, td->td_ucred); /* * One of my assumptions failed if ENOENT (e.g. locking didn't * protect us) */ KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", path_from)); if (error != 0) { shm_drop(fd_from); goto out_locked; } /* * If we are exchanging, we need to ensure the shm_remove below * doesn't invalidate the dest shm's state. */ if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) shm_hold(fd_to); /* * NOTE: if path_to is not already in the hash, c'est la vie; * it simply means we have nothing already at path_to to unlink. * That is the ENOENT case. * * If we somehow don't have access to unlink this guy, but * did for the shm at path_from, then relink the shm to path_from * and abort with EACCES. * * All other errors: that is weird; let's relink and abort the * operation. */ error = shm_remove(path_to, fnv_to, td->td_ucred); if (error != 0 && error != ENOENT) { shm_insert(path_from, fnv_from, fd_from); shm_drop(fd_from); /* Don't free path_from now, since the hash references it */ path_from = NULL; goto out_locked; } error = 0; shm_insert(path_to, fnv_to, fd_from); /* Don't free path_to now, since the hash references it */ path_to = NULL; /* We kept a ref when we removed, and incremented again in insert */ shm_drop(fd_from); KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_from->shm_refs)); if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { shm_insert(path_from, fnv_from, fd_to); path_from = NULL; shm_drop(fd_to); KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_to->shm_refs)); } out_locked: sx_xunlock(&shm_dict_lock); out: free(path_from, M_SHMFD); free(path_to, M_SHMFD); return (error); } static int shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr, vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags, vm_ooffset_t foff, struct thread *td) { struct vmspace *vms; vm_map_entry_t next_entry, prev_entry; vm_offset_t align, mask, maxaddr; int docow, error, rv, try; bool curmap; if (shmfd->shm_lp_psind == 0) return (EINVAL); /* MAP_PRIVATE is disabled */ if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL | MAP_NOCORE | MAP_32BIT | MAP_ALIGNMENT_MASK)) != 0) return (EINVAL); vms = td->td_proc->p_vmspace; curmap = map == &vms->vm_map; if (curmap) { error = kern_mmap_racct_check(td, map, size); if (error != 0) return (error); } docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT; docow |= MAP_INHERIT_SHARE; if ((flags & MAP_NOCORE) != 0) docow |= MAP_DISABLE_COREDUMP; mask = pagesizes[shmfd->shm_lp_psind] - 1; if ((foff & mask) != 0) return (EINVAL); maxaddr = vm_map_max(map); if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR) maxaddr = MAP_32BIT_MAX_ADDR; if (size == 0 || (size & mask) != 0 || (*addr != 0 && ((*addr & mask) != 0 || *addr + size < *addr || *addr + size > maxaddr))) return (EINVAL); align = flags & MAP_ALIGNMENT_MASK; if (align == 0) { align = pagesizes[shmfd->shm_lp_psind]; } else if (align == MAP_ALIGNED_SUPER) { if (shmfd->shm_lp_psind != 1) return (EINVAL); align = pagesizes[1]; } else { align >>= MAP_ALIGNMENT_SHIFT; align = 1ULL << align; /* Also handles overflow. */ if (align < pagesizes[shmfd->shm_lp_psind]) return (EINVAL); } vm_map_lock(map); if ((flags & MAP_FIXED) == 0) { try = 1; if (curmap && (*addr == 0 || (*addr >= round_page((vm_offset_t)vms->vm_taddr) && *addr < round_page((vm_offset_t)vms->vm_daddr + lim_max(td, RLIMIT_DATA))))) { *addr = roundup2((vm_offset_t)vms->vm_daddr + lim_max(td, RLIMIT_DATA), pagesizes[shmfd->shm_lp_psind]); } again: rv = vm_map_find_aligned(map, addr, size, maxaddr, align); if (rv != KERN_SUCCESS) { if (try == 1) { try = 2; *addr = vm_map_min(map); if ((*addr & mask) != 0) *addr = (*addr + mask) & mask; goto again; } goto fail1; } } else if ((flags & MAP_EXCL) == 0) { rv = vm_map_delete(map, *addr, *addr + size); if (rv != KERN_SUCCESS) goto fail1; } else { error = ENOSPC; if (vm_map_lookup_entry(map, *addr, &prev_entry)) goto fail; next_entry = vm_map_entry_succ(prev_entry); if (next_entry->start < *addr + size) goto fail; } rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size, prot, max_prot, docow); fail1: error = vm_mmap_to_errno(rv); fail: vm_map_unlock(map); return (error); } static int shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, vm_prot_t prot, vm_prot_t max_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { struct shmfd *shmfd; vm_prot_t maxprot; int error; bool writecnt; void *rl_cookie; shmfd = fp->f_data; maxprot = VM_PROT_NONE; rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize); /* FREAD should always be set. */ if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; /* * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared * mapping with a write seal applied. Private mappings are always * writeable. */ if ((flags & MAP_SHARED) == 0) { if ((max_maxprot & VM_PROT_WRITE) != 0) maxprot |= VM_PROT_WRITE; writecnt = false; } else { if ((fp->f_flag & FWRITE) != 0 && (shmfd->shm_seals & F_SEAL_WRITE) == 0) maxprot |= VM_PROT_WRITE; /* * Any mappings from a writable descriptor may be upgraded to * VM_PROT_WRITE with mprotect(2), unless a write-seal was * applied between the open and subsequent mmap(2). We want to * reject application of a write seal as long as any such * mapping exists so that the seal cannot be trivially bypassed. */ writecnt = (maxprot & VM_PROT_WRITE) != 0; if (!writecnt && (prot & VM_PROT_WRITE) != 0) { error = EACCES; goto out; } } maxprot &= max_maxprot; /* See comment in vn_mmap(). */ if ( #ifdef _LP64 objsize > OFF_MAX || #endif foff > OFF_MAX - objsize) { error = EINVAL; goto out; } #ifdef MAC error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); if (error != 0) goto out; #endif mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_atime); mtx_unlock(&shm_timestamp_lock); vm_object_reference(shmfd->shm_object); if (shm_largepage(shmfd)) { writecnt = false; error = shm_mmap_large(shmfd, map, addr, objsize, prot, maxprot, flags, foff, td); } else { if (writecnt) { vm_pager_update_writecount(shmfd->shm_object, 0, objsize); } error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, shmfd->shm_object, foff, writecnt, td); } if (error != 0) { if (writecnt) vm_pager_release_writecount(shmfd->shm_object, 0, objsize); vm_object_deallocate(shmfd->shm_object); } out: shm_rangelock_unlock(shmfd, rl_cookie); return (error); } static int shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); /* * SUSv4 says that x bits of permission need not be affected. * Be consistent with our shm_open there. */ #ifdef MAC error = mac_posixshm_check_setmode(active_cred, shmfd, mode); if (error != 0) goto out; #endif error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, VADMIN, active_cred); if (error != 0) goto out; shmfd->shm_mode = mode & ACCESSPERMS; out: mtx_unlock(&shm_timestamp_lock); return (error); } static int shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); #ifdef MAC error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); if (error != 0) goto out; #endif if (uid == (uid_t)-1) uid = shmfd->shm_uid; if (gid == (gid_t)-1) gid = shmfd->shm_gid; if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) goto out; shmfd->shm_uid = uid; shmfd->shm_gid = gid; out: mtx_unlock(&shm_timestamp_lock); return (error); } /* * Helper routines to allow the backing object of a shared memory file * descriptor to be mapped in the kernel. */ int shm_map(struct file *fp, size_t size, off_t offset, void **memp) { struct shmfd *shmfd; vm_offset_t kva, ofs; vm_object_t obj; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; obj = shmfd->shm_object; VM_OBJECT_WLOCK(obj); /* * XXXRW: This validation is probably insufficient, and subject to * sign errors. It should be fixed. */ if (offset >= shmfd->shm_size || offset + size > round_page(shmfd->shm_size)) { VM_OBJECT_WUNLOCK(obj); return (EINVAL); } shmfd->shm_kmappings++; vm_object_reference_locked(obj); VM_OBJECT_WUNLOCK(obj); /* Map the object into the kernel_map and wire it. */ kva = vm_map_min(kernel_map); ofs = offset & PAGE_MASK; offset = trunc_page(offset); size = round_page(size + ofs); rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, VM_PROT_READ | VM_PROT_WRITE, 0); if (rv == KERN_SUCCESS) { rv = vm_map_wire(kernel_map, kva, kva + size, VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); if (rv == KERN_SUCCESS) { *memp = (void *)(kva + ofs); return (0); } vm_map_remove(kernel_map, kva, kva + size); } else vm_object_deallocate(obj); /* On failure, drop our mapping reference. */ VM_OBJECT_WLOCK(obj); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (vm_mmap_to_errno(rv)); } /* * We require the caller to unmap the entire entry. This allows us to * safely decrement shm_kmappings when a mapping is removed. */ int shm_unmap(struct file *fp, void *mem, size_t size) { struct shmfd *shmfd; vm_map_entry_t entry; vm_offset_t kva, ofs; vm_object_t obj; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; vm_map_t map; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; kva = (vm_offset_t)mem; ofs = kva & PAGE_MASK; kva = trunc_page(kva); size = round_page(size + ofs); map = kernel_map; rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, &obj, &pindex, &prot, &wired); if (rv != KERN_SUCCESS) return (EINVAL); if (entry->start != kva || entry->end != kva + size) { vm_map_lookup_done(map, entry); return (EINVAL); } vm_map_lookup_done(map, entry); if (obj != shmfd->shm_object) return (EINVAL); vm_map_remove(map, kva, kva + size); VM_OBJECT_WLOCK(obj); KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (0); } static int shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) { const char *path, *pr_path; size_t pr_pathlen; bool visible; sx_assert(&shm_dict_lock, SA_LOCKED); kif->kf_type = KF_TYPE_SHM; kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; if (shmfd->shm_path != NULL) { path = shmfd->shm_path; pr_path = curthread->td_ucred->cr_prison->pr_path; if (strcmp(pr_path, "/") != 0) { /* Return the jail-rooted pathname. */ pr_pathlen = strlen(pr_path); visible = strncmp(path, pr_path, pr_pathlen) == 0 && path[pr_pathlen] == '/'; if (list && !visible) return (EPERM); if (visible) path += pr_pathlen; } strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); } return (0); } static int shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp __unused) { int res; sx_slock(&shm_dict_lock); res = shm_fill_kinfo_locked(fp->f_data, kif, false); sx_sunlock(&shm_dict_lock); return (res); } static int shm_add_seals(struct file *fp, int seals) { struct shmfd *shmfd; void *rl_cookie; vm_ooffset_t writemappings; int error, nseals; error = 0; shmfd = fp->f_data; rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); /* Even already-set seals should result in EPERM. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { error = EPERM; goto out; } nseals = seals & ~shmfd->shm_seals; if ((nseals & F_SEAL_WRITE) != 0) { if (shm_largepage(shmfd)) { error = ENOTSUP; goto out; } /* * The rangelock above prevents writable mappings from being * added after we've started applying seals. The RLOCK here * is to avoid torn reads on ILP32 arches as unmapping/reducing * writemappings will be done without a rangelock. */ VM_OBJECT_RLOCK(shmfd->shm_object); writemappings = shmfd->shm_object->un_pager.swp.writemappings; VM_OBJECT_RUNLOCK(shmfd->shm_object); /* kmappings are also writable */ if (writemappings > 0) { error = EBUSY; goto out; } } shmfd->shm_seals |= nseals; out: shm_rangelock_unlock(shmfd, rl_cookie); return (error); } static int shm_get_seals(struct file *fp, int *seals) { struct shmfd *shmfd; shmfd = fp->f_data; *seals = shmfd->shm_seals; return (0); } static int shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags) { vm_object_t object; vm_pindex_t pistart, pi, piend; vm_ooffset_t off, len; int startofs, endofs, end; int error; off = *offset; len = *length; KASSERT(off + len <= (vm_ooffset_t)OFF_MAX, ("off + len overflows")); if (off + len > shmfd->shm_size) len = shmfd->shm_size - off; object = shmfd->shm_object; startofs = off & PAGE_MASK; endofs = (off + len) & PAGE_MASK; pistart = OFF_TO_IDX(off); piend = OFF_TO_IDX(off + len); pi = OFF_TO_IDX(off + PAGE_MASK); error = 0; /* Handle the case when offset is on or beyond shm size. */ if ((off_t)len <= 0) { *length = 0; return (0); } VM_OBJECT_WLOCK(object); if (startofs != 0) { end = pistart != piend ? PAGE_SIZE : endofs; error = shm_partial_page_invalidate(object, pistart, startofs, end); if (error) goto out; off += end - startofs; len -= end - startofs; } if (pi < piend) { vm_object_page_remove(object, pi, piend, 0); off += IDX_TO_OFF(piend - pi); len -= IDX_TO_OFF(piend - pi); } if (endofs != 0 && pistart != piend) { error = shm_partial_page_invalidate(object, piend, 0, endofs); if (error) goto out; off += endofs; len -= endofs; } out: VM_OBJECT_WUNLOCK(shmfd->shm_object); *offset = off; *length = len; return (error); } static int shm_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags, struct ucred *active_cred, struct thread *td) { void *rl_cookie; struct shmfd *shmfd; off_t off, len; int error; KASSERT(cmd == SPACECTL_DEALLOC, ("shm_fspacectl: Invalid cmd")); KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0, ("shm_fspacectl: non-zero flags")); KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset, ("shm_fspacectl: offset/length overflow or underflow")); error = EINVAL; shmfd = fp->f_data; off = *offset; len = *length; rl_cookie = shm_rangelock_wlock(shmfd, off, off + len); switch (cmd) { case SPACECTL_DEALLOC: if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) { error = EPERM; break; } error = shm_deallocate(shmfd, &off, &len, flags); *offset = off; *length = len; break; default: __assert_unreachable(); } shm_rangelock_unlock(shmfd, rl_cookie); return (error); } static int shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) { void *rl_cookie; struct shmfd *shmfd; size_t size; int error; /* This assumes that the caller already checked for overflow. */ error = 0; shmfd = fp->f_data; size = offset + len; /* * Just grab the rangelock for the range that we may be attempting to * grow, rather than blocking read/write for regions we won't be * touching while this (potential) resize is in progress. Other * attempts to resize the shmfd will have to take a write lock from 0 to * OFF_MAX, so this being potentially beyond the current usable range of * the shmfd is not necessarily a concern. If other mechanisms are * added to grow a shmfd, this may need to be re-evaluated. */ rl_cookie = shm_rangelock_wlock(shmfd, offset, size); if (size > shmfd->shm_size) error = shm_dotruncate_cookie(shmfd, size, rl_cookie); shm_rangelock_unlock(shmfd, rl_cookie); /* Translate to posix_fallocate(2) return value as needed. */ if (error == ENOMEM) error = ENOSPC; return (error); } static int sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) { struct shm_mapping *shmm; struct sbuf sb; struct kinfo_file kif; u_long i; int error, error2; sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); sbuf_clear_flags(&sb, SBUF_INCLUDENUL); error = 0; sx_slock(&shm_dict_lock); for (i = 0; i < shm_hash + 1; i++) { LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { error = shm_fill_kinfo_locked(shmm->sm_shmfd, &kif, true); if (error == EPERM) { error = 0; continue; } if (error != 0) break; pack_kinfo(&kif); error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 0 : ENOMEM; if (error != 0) break; } } sx_sunlock(&shm_dict_lock); error2 = sbuf_finish(&sb); sbuf_delete(&sb); return (error != 0 ? error : error2); } SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, NULL, 0, sysctl_posix_shm_list, "", "POSIX SHM list"); int kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode, struct filecaps *caps) { return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL)); } /* * This version of the shm_open() interface leaves CLOEXEC behavior up to the * caller, and libc will enforce it for the traditional shm_open() call. This * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This * interface also includes a 'name' argument that is currently unused, but could * potentially be exported later via some interface for debugging purposes. * From the kernel's perspective, it is optional. Individual consumers like * memfd_create() may require it in order to be compatible with other systems * implementing the same function. */ int sys_shm_open2(struct thread *td, struct shm_open2_args *uap) { return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, uap->shmflags, NULL, uap->name)); } int shm_get_path(struct vm_object *obj, char *path, size_t sz) { struct shmfd *shmfd; int error; error = 0; shmfd = NULL; sx_slock(&shm_dict_lock); VM_OBJECT_RLOCK(obj); if ((obj->flags & OBJ_POSIXSHM) == 0) { error = EINVAL; } else { if (obj->type == shmfd_pager_type) shmfd = obj->un_pager.swp.swp_priv; else if (obj->type == OBJT_PHYS) shmfd = obj->un_pager.phys.phys_priv; if (shmfd == NULL) { error = ENXIO; } else { strlcpy(path, shmfd->shm_path == NULL ? "anon" : shmfd->shm_path, sz); } } if (error != 0) path[0] = '\0'; VM_OBJECT_RUNLOCK(obj); sx_sunlock(&shm_dict_lock); return (error); } diff --git a/sys/vm/vm_domainset.c b/sys/vm/vm_domainset.c index b666324db494..ff0e476a68b9 100644 --- a/sys/vm/vm_domainset.c +++ b/sys/vm/vm_domainset.c @@ -1,356 +1,378 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2017, Jeffrey Roberson * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * */ #include #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef NUMA /* * Iterators are written such that the first nowait pass has as short a * codepath as possible to eliminate bloat from the allocator. It is * assumed that most allocations are successful. */ static int vm_domainset_default_stride = 64; /* * Determine which policy is to be used for this allocation. */ static void vm_domainset_iter_init(struct vm_domainset_iter *di, struct domainset *ds, int *iter, struct vm_object *obj, vm_pindex_t pindex) { di->di_domain = ds; di->di_iter = iter; di->di_policy = ds->ds_policy; + DOMAINSET_COPY(&ds->ds_mask, &di->di_valid_mask); if (di->di_policy == DOMAINSET_POLICY_INTERLEAVE) { #if VM_NRESERVLEVEL > 0 if (vm_object_reserv(obj)) { /* * Color the pindex so we end up on the correct * reservation boundary. */ pindex += obj->pg_color; pindex >>= VM_LEVEL_0_ORDER; } else #endif pindex /= vm_domainset_default_stride; /* * Offset pindex so the first page of each object does * not end up in domain 0. */ if (obj != NULL) pindex += (((uintptr_t)obj) / sizeof(*obj)); di->di_offset = pindex; } /* Skip domains below min on the first pass. */ di->di_minskip = true; } static void vm_domainset_iter_rr(struct vm_domainset_iter *di, int *domain) { *domain = di->di_domain->ds_order[ ++(*di->di_iter) % di->di_domain->ds_cnt]; } static void vm_domainset_iter_prefer(struct vm_domainset_iter *di, int *domain) { int d; do { d = di->di_domain->ds_order[ ++(*di->di_iter) % di->di_domain->ds_cnt]; } while (d == di->di_domain->ds_prefer); *domain = d; } static void vm_domainset_iter_interleave(struct vm_domainset_iter *di, int *domain) { int d; d = di->di_offset % di->di_domain->ds_cnt; *di->di_iter = d; *domain = di->di_domain->ds_order[d]; } static void vm_domainset_iter_next(struct vm_domainset_iter *di, int *domain) { KASSERT(di->di_n > 0, ("%s: Invalid n %d", __func__, di->di_n)); switch (di->di_policy) { case DOMAINSET_POLICY_FIRSTTOUCH: /* * To prevent impossible allocations we convert an invalid * first-touch to round-robin. */ /* FALLTHROUGH */ case DOMAINSET_POLICY_INTERLEAVE: /* FALLTHROUGH */ case DOMAINSET_POLICY_ROUNDROBIN: vm_domainset_iter_rr(di, domain); break; case DOMAINSET_POLICY_PREFER: vm_domainset_iter_prefer(di, domain); break; default: panic("%s: Unknown policy %d", __func__, di->di_policy); } KASSERT(*domain < vm_ndomains, ("%s: Invalid domain %d", __func__, *domain)); } static void vm_domainset_iter_first(struct vm_domainset_iter *di, int *domain) { switch (di->di_policy) { case DOMAINSET_POLICY_FIRSTTOUCH: *domain = PCPU_GET(domain); - if (DOMAINSET_ISSET(*domain, &di->di_domain->ds_mask)) { + if (DOMAINSET_ISSET(*domain, &di->di_valid_mask)) { /* * Add an extra iteration because we will visit the * current domain a second time in the rr iterator. */ di->di_n = di->di_domain->ds_cnt + 1; break; } /* * To prevent impossible allocations we convert an invalid * first-touch to round-robin. */ /* FALLTHROUGH */ case DOMAINSET_POLICY_ROUNDROBIN: di->di_n = di->di_domain->ds_cnt; vm_domainset_iter_rr(di, domain); break; case DOMAINSET_POLICY_PREFER: *domain = di->di_domain->ds_prefer; di->di_n = di->di_domain->ds_cnt; break; case DOMAINSET_POLICY_INTERLEAVE: vm_domainset_iter_interleave(di, domain); di->di_n = di->di_domain->ds_cnt; break; default: panic("%s: Unknown policy %d", __func__, di->di_policy); } KASSERT(di->di_n > 0, ("%s: Invalid n %d", __func__, di->di_n)); KASSERT(*domain < vm_ndomains, ("%s: Invalid domain %d", __func__, *domain)); } void vm_domainset_iter_page_init(struct vm_domainset_iter *di, struct vm_object *obj, vm_pindex_t pindex, int *domain, int *req) { struct domainset_ref *dr; /* * Object policy takes precedence over thread policy. The policies * are immutable and unsynchronized. Updates can race but pointer * loads are assumed to be atomic. */ if (obj != NULL && obj->domain.dr_policy != NULL) dr = &obj->domain; else dr = &curthread->td_domain; vm_domainset_iter_init(di, dr->dr_policy, &dr->dr_iter, obj, pindex); di->di_flags = *req; *req = (di->di_flags & ~(VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) | VM_ALLOC_NOWAIT; vm_domainset_iter_first(di, domain); if (vm_page_count_min_domain(*domain)) vm_domainset_iter_page(di, obj, domain); } int vm_domainset_iter_page(struct vm_domainset_iter *di, struct vm_object *obj, int *domain) { + if (__predict_false(DOMAINSET_EMPTY(&di->di_valid_mask))) + return (ENOMEM); /* If there are more domains to visit we run the iterator. */ while (--di->di_n != 0) { vm_domainset_iter_next(di, domain); - if (!di->di_minskip || !vm_page_count_min_domain(*domain)) + if (DOMAINSET_ISSET(*domain, &di->di_valid_mask) && + (!di->di_minskip || !vm_page_count_min_domain(*domain))) return (0); } /* If we skipped domains below min restart the search. */ if (di->di_minskip) { di->di_minskip = false; vm_domainset_iter_first(di, domain); return (0); } /* If we visited all domains and this was a NOWAIT we return error. */ if ((di->di_flags & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) == 0) return (ENOMEM); /* Wait for one of the domains to accumulate some free pages. */ if (obj != NULL) VM_OBJECT_WUNLOCK(obj); - vm_wait_doms(&di->di_domain->ds_mask, 0); + vm_wait_doms(&di->di_valid_mask, 0); if (obj != NULL) VM_OBJECT_WLOCK(obj); if ((di->di_flags & VM_ALLOC_WAITFAIL) != 0) return (ENOMEM); /* Restart the search. */ vm_domainset_iter_first(di, domain); return (0); } static void _vm_domainset_iter_policy_init(struct vm_domainset_iter *di, int *domain, int *flags) { di->di_flags = *flags; *flags = (di->di_flags & ~M_WAITOK) | M_NOWAIT; vm_domainset_iter_first(di, domain); if (vm_page_count_min_domain(*domain)) vm_domainset_iter_policy(di, domain); } void vm_domainset_iter_policy_init(struct vm_domainset_iter *di, struct domainset *ds, int *domain, int *flags) { vm_domainset_iter_init(di, ds, &curthread->td_domain.dr_iter, NULL, 0); _vm_domainset_iter_policy_init(di, domain, flags); } void vm_domainset_iter_policy_ref_init(struct vm_domainset_iter *di, struct domainset_ref *dr, int *domain, int *flags) { vm_domainset_iter_init(di, dr->dr_policy, &dr->dr_iter, NULL, 0); _vm_domainset_iter_policy_init(di, domain, flags); } int vm_domainset_iter_policy(struct vm_domainset_iter *di, int *domain) { + if (DOMAINSET_EMPTY(&di->di_valid_mask)) + return (ENOMEM); /* If there are more domains to visit we run the iterator. */ while (--di->di_n != 0) { vm_domainset_iter_next(di, domain); - if (!di->di_minskip || !vm_page_count_min_domain(*domain)) + if (DOMAINSET_ISSET(*domain, &di->di_valid_mask) && + (!di->di_minskip || !vm_page_count_min_domain(*domain))) return (0); } /* If we skipped domains below min restart the search. */ if (di->di_minskip) { di->di_minskip = false; vm_domainset_iter_first(di, domain); return (0); } /* If we visited all domains and this was a NOWAIT we return error. */ if ((di->di_flags & M_WAITOK) == 0) return (ENOMEM); /* Wait for one of the domains to accumulate some free pages. */ - vm_wait_doms(&di->di_domain->ds_mask, 0); + vm_wait_doms(&di->di_valid_mask, 0); /* Restart the search. */ vm_domainset_iter_first(di, domain); return (0); } +void +vm_domainset_iter_ignore(struct vm_domainset_iter *di, int domain) +{ + KASSERT(DOMAINSET_ISSET(domain, &di->di_valid_mask), + ("%s: domain %d not present in di_valid_mask for di %p", + __func__, domain, di)); + DOMAINSET_CLR(domain, &di->di_valid_mask); +} + #else /* !NUMA */ int vm_domainset_iter_page(struct vm_domainset_iter *di, struct vm_object *obj, int *domain) { return (EJUSTRETURN); } void vm_domainset_iter_page_init(struct vm_domainset_iter *di, struct vm_object *obj, vm_pindex_t pindex, int *domain, int *flags) { *domain = 0; } int vm_domainset_iter_policy(struct vm_domainset_iter *di, int *domain) { return (EJUSTRETURN); } void vm_domainset_iter_policy_init(struct vm_domainset_iter *di, struct domainset *ds, int *domain, int *flags) { *domain = 0; } void vm_domainset_iter_policy_ref_init(struct vm_domainset_iter *di, struct domainset_ref *dr, int *domain, int *flags) { *domain = 0; } +void +vm_domainset_iter_ignore(struct vm_domainset_iter *di __unused, + int domain __unused) +{ +} + #endif /* NUMA */ diff --git a/sys/vm/vm_domainset.h b/sys/vm/vm_domainset.h index 4ce619069dd0..d2cfe362ae78 100644 --- a/sys/vm/vm_domainset.h +++ b/sys/vm/vm_domainset.h @@ -1,53 +1,55 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2017, Jeffrey Roberson * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #ifndef __VM_DOMAINSET_H__ #define __VM_DOMAINSET_H__ struct vm_domainset_iter { struct domainset *di_domain; unsigned int *di_iter; + domainset_t di_valid_mask; vm_pindex_t di_offset; int di_flags; uint16_t di_policy; domainid_t di_n; bool di_minskip; }; int vm_domainset_iter_page(struct vm_domainset_iter *, struct vm_object *, int *); void vm_domainset_iter_page_init(struct vm_domainset_iter *, struct vm_object *, vm_pindex_t, int *, int *); int vm_domainset_iter_policy(struct vm_domainset_iter *, int *); void vm_domainset_iter_policy_init(struct vm_domainset_iter *, struct domainset *, int *, int *); void vm_domainset_iter_policy_ref_init(struct vm_domainset_iter *, struct domainset_ref *, int *, int *); +void vm_domainset_iter_ignore(struct vm_domainset_iter *, int); int vm_wait_doms(const domainset_t *, int mflags); #endif /* __VM_DOMAINSET_H__ */ diff --git a/sys/vm/vm_kern.c b/sys/vm/vm_kern.c index 387d73d2e46e..f11dfd745720 100644 --- a/sys/vm/vm_kern.c +++ b/sys/vm/vm_kern.c @@ -1,1023 +1,1043 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_kern.c 8.3 (Berkeley) 1/12/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Kernel memory management. */ #include #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct vm_map kernel_map_store; struct vm_map exec_map_store; struct vm_map pipe_map_store; const void *zero_region; CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0); /* NB: Used by kernel debuggers. */ const u_long vm_maxuser_address = VM_MAXUSER_ADDRESS; u_int exec_map_entry_size; u_int exec_map_entries; SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD, SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS, "Min kernel address"); SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD, #if defined(__arm__) &vm_max_kernel_address, 0, #else SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS, #endif "Max kernel address"); #if VM_NRESERVLEVEL > 0 #define KVA_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT) #else /* On non-superpage architectures we want large import sizes. */ #define KVA_QUANTUM_SHIFT (8 + PAGE_SHIFT) #endif #define KVA_QUANTUM (1ul << KVA_QUANTUM_SHIFT) #define KVA_NUMA_IMPORT_QUANTUM (KVA_QUANTUM * 128) extern void uma_startup2(void); /* * kva_alloc: * * Allocate a virtual address range with no underlying object and * no initial mapping to physical memory. Any mapping from this * range to physical memory must be explicitly created prior to * its use, typically with pmap_qenter(). Any attempt to create * a mapping on demand through vm_fault() will result in a panic. */ vm_offset_t kva_alloc(vm_size_t size) { vm_offset_t addr; TSENTER(); size = round_page(size); if (vmem_xalloc(kernel_arena, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, M_BESTFIT | M_NOWAIT, &addr)) return (0); TSEXIT(); return (addr); } /* * kva_alloc_aligned: * * Allocate a virtual address range as in kva_alloc where the base * address is aligned to align. */ vm_offset_t kva_alloc_aligned(vm_size_t size, vm_size_t align) { vm_offset_t addr; TSENTER(); size = round_page(size); if (vmem_xalloc(kernel_arena, size, align, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, M_BESTFIT | M_NOWAIT, &addr)) return (0); TSEXIT(); return (addr); } /* * kva_free: * * Release a region of kernel virtual memory allocated * with kva_alloc, and return the physical pages * associated with that region. * * This routine may not block on kernel maps. */ void kva_free(vm_offset_t addr, vm_size_t size) { size = round_page(size); vmem_xfree(kernel_arena, addr, size); } /* * Update sanitizer shadow state to reflect a new allocation. Force inlining to * help make KMSAN origin tracking more precise. */ static __always_inline void kmem_alloc_san(vm_offset_t addr, vm_size_t size, vm_size_t asize, int flags) { if ((flags & M_ZERO) == 0) { kmsan_mark((void *)addr, asize, KMSAN_STATE_UNINIT); kmsan_orig((void *)addr, asize, KMSAN_TYPE_KMEM, KMSAN_RET_ADDR); } else { kmsan_mark((void *)addr, asize, KMSAN_STATE_INITED); } kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE); } static vm_page_t kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain, int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { vm_page_t m; int tries; bool wait, reclaim; VM_OBJECT_ASSERT_WLOCKED(object); wait = (pflags & VM_ALLOC_WAITOK) != 0; reclaim = (pflags & VM_ALLOC_NORECLAIM) == 0; pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); pflags |= VM_ALLOC_NOWAIT; for (tries = wait ? 3 : 1;; tries--) { m = vm_page_alloc_contig_domain(object, pindex, domain, pflags, npages, low, high, alignment, boundary, memattr); if (m != NULL || tries == 0 || !reclaim) break; VM_OBJECT_WUNLOCK(object); - if (!vm_page_reclaim_contig_domain(domain, pflags, npages, - low, high, alignment, boundary) && wait) + if (vm_page_reclaim_contig_domain(domain, pflags, npages, + low, high, alignment, boundary) == ENOMEM && wait) vm_wait_domain(domain); VM_OBJECT_WLOCK(object); } return (m); } /* * Allocates a region from the kernel address map and physical pages * within the specified address range to the kernel object. Creates a * wired mapping from this region to these pages, and returns the * region's starting virtual address. The allocated pages are not * necessarily physically contiguous. If M_ZERO is specified through the * given flags, then the pages are zeroed before they are mapped. */ static void * kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr) { vmem_t *vmem; vm_object_t object; vm_offset_t addr, i, offset; vm_page_t m; vm_size_t asize; int pflags; vm_prot_t prot; object = kernel_object; asize = round_page(size); vmem = vm_dom[domain].vmd_kernel_arena; if (vmem_alloc(vmem, asize, M_BESTFIT | flags, &addr)) return (0); offset = addr - VM_MIN_KERNEL_ADDRESS; pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; VM_OBJECT_WLOCK(object); for (i = 0; i < asize; i += PAGE_SIZE) { m = kmem_alloc_contig_pages(object, atop(offset + i), domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr); if (m == NULL) { VM_OBJECT_WUNLOCK(object); kmem_unback(object, addr, i); vmem_free(vmem, addr, asize); return (0); } KASSERT(vm_page_domain(m) == domain, ("kmem_alloc_attr_domain: Domain mismatch %d != %d", vm_page_domain(m), domain)); if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); vm_page_valid(m); pmap_enter(kernel_pmap, addr + i, m, prot, prot | PMAP_ENTER_WIRED, 0); } VM_OBJECT_WUNLOCK(object); kmem_alloc_san(addr, size, asize, flags); return ((void *)addr); } void * kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr) { return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low, high, memattr)); } void * kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr) { struct vm_domainset_iter di; + vm_page_t bounds[2]; void *addr; int domain; + int start_segind; + + start_segind = -1; vm_domainset_iter_policy_init(&di, ds, &domain, &flags); do { addr = kmem_alloc_attr_domain(domain, size, flags, low, high, memattr); if (addr != NULL) break; + if (start_segind == -1) + start_segind = vm_phys_lookup_segind(low); + if (vm_phys_find_range(bounds, start_segind, domain, + atop(round_page(size)), low, high) == -1) { + vm_domainset_iter_ignore(&di, domain); + } } while (vm_domainset_iter_policy(&di, &domain) == 0); return (addr); } /* * Allocates a region from the kernel address map and physically * contiguous pages within the specified address range to the kernel * object. Creates a wired mapping from this region to these pages, and * returns the region's starting virtual address. If M_ZERO is specified * through the given flags, then the pages are zeroed before they are * mapped. */ static void * kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { vmem_t *vmem; vm_object_t object; vm_offset_t addr, offset, tmp; vm_page_t end_m, m; vm_size_t asize; u_long npages; int pflags; object = kernel_object; asize = round_page(size); vmem = vm_dom[domain].vmd_kernel_arena; if (vmem_alloc(vmem, asize, flags | M_BESTFIT, &addr)) return (NULL); offset = addr - VM_MIN_KERNEL_ADDRESS; pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; npages = atop(asize); VM_OBJECT_WLOCK(object); m = kmem_alloc_contig_pages(object, atop(offset), domain, pflags, npages, low, high, alignment, boundary, memattr); if (m == NULL) { VM_OBJECT_WUNLOCK(object); vmem_free(vmem, addr, asize); return (NULL); } KASSERT(vm_page_domain(m) == domain, ("kmem_alloc_contig_domain: Domain mismatch %d != %d", vm_page_domain(m), domain)); end_m = m + npages; tmp = addr; for (; m < end_m; m++) { if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); vm_page_valid(m); pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW, VM_PROT_RW | PMAP_ENTER_WIRED, 0); tmp += PAGE_SIZE; } VM_OBJECT_WUNLOCK(object); kmem_alloc_san(addr, size, asize, flags); return ((void *)addr); } void * kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low, high, alignment, boundary, memattr)); } void * kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; + vm_page_t bounds[2]; void *addr; int domain; + int start_segind; + + start_segind = -1; vm_domainset_iter_policy_init(&di, ds, &domain, &flags); do { addr = kmem_alloc_contig_domain(domain, size, flags, low, high, alignment, boundary, memattr); if (addr != NULL) break; + if (start_segind == -1) + start_segind = vm_phys_lookup_segind(low); + if (vm_phys_find_range(bounds, start_segind, domain, + atop(round_page(size)), low, high) == -1) { + vm_domainset_iter_ignore(&di, domain); + } } while (vm_domainset_iter_policy(&di, &domain) == 0); return (addr); } /* * kmem_subinit: * * Initializes a map to manage a subrange * of the kernel virtual address space. * * Arguments are as follows: * * parent Map to take range from * min, max Returned endpoints of map * size Size of range to find * superpage_align Request that min is superpage aligned */ void kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max, vm_size_t size, bool superpage_align) { int ret; size = round_page(size); *min = vm_map_min(parent); ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ? VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_ACC_NO_CHARGE); if (ret != KERN_SUCCESS) panic("kmem_subinit: bad status return of %d", ret); *max = *min + size; vm_map_init(map, vm_map_pmap(parent), *min, *max); if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS) panic("kmem_subinit: unable to change range to submap"); } /* * kmem_malloc_domain: * * Allocate wired-down pages in the kernel's address space. */ static void * kmem_malloc_domain(int domain, vm_size_t size, int flags) { vmem_t *arena; vm_offset_t addr; vm_size_t asize; int rv; if (__predict_true((flags & M_EXEC) == 0)) arena = vm_dom[domain].vmd_kernel_arena; else arena = vm_dom[domain].vmd_kernel_rwx_arena; asize = round_page(size); if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr)) return (0); rv = kmem_back_domain(domain, kernel_object, addr, asize, flags); if (rv != KERN_SUCCESS) { vmem_free(arena, addr, asize); return (0); } kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE); return ((void *)addr); } void * kmem_malloc(vm_size_t size, int flags) { void * p; TSENTER(); p = kmem_malloc_domainset(DOMAINSET_RR(), size, flags); TSEXIT(); return (p); } void * kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags) { struct vm_domainset_iter di; void *addr; int domain; vm_domainset_iter_policy_init(&di, ds, &domain, &flags); do { addr = kmem_malloc_domain(domain, size, flags); if (addr != NULL) break; } while (vm_domainset_iter_policy(&di, &domain) == 0); return (addr); } /* * kmem_back_domain: * * Allocate physical pages from the specified domain for the specified * virtual address range. */ int kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr, vm_size_t size, int flags) { vm_offset_t offset, i; vm_page_t m, mpred; vm_prot_t prot; int pflags; KASSERT(object == kernel_object, ("kmem_back_domain: only supports kernel object.")); offset = addr - VM_MIN_KERNEL_ADDRESS; pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); if (flags & M_WAITOK) pflags |= VM_ALLOC_WAITFAIL; prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; i = 0; VM_OBJECT_WLOCK(object); retry: mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i)); for (; i < size; i += PAGE_SIZE, mpred = m) { m = vm_page_alloc_domain_after(object, atop(offset + i), domain, pflags, mpred); /* * Ran out of space, free everything up and return. Don't need * to lock page queues here as we know that the pages we got * aren't on any queues. */ if (m == NULL) { if ((flags & M_NOWAIT) == 0) goto retry; VM_OBJECT_WUNLOCK(object); kmem_unback(object, addr, i); return (KERN_NO_SPACE); } KASSERT(vm_page_domain(m) == domain, ("kmem_back_domain: Domain mismatch %d != %d", vm_page_domain(m), domain)); if (flags & M_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("kmem_malloc: page %p is managed", m)); vm_page_valid(m); pmap_enter(kernel_pmap, addr + i, m, prot, prot | PMAP_ENTER_WIRED, 0); if (__predict_false((prot & VM_PROT_EXECUTE) != 0)) m->oflags |= VPO_KMEM_EXEC; } VM_OBJECT_WUNLOCK(object); kmem_alloc_san(addr, size, size, flags); return (KERN_SUCCESS); } /* * kmem_back: * * Allocate physical pages for the specified virtual address range. */ int kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags) { vm_offset_t end, next, start; int domain, rv; KASSERT(object == kernel_object, ("kmem_back: only supports kernel object.")); for (start = addr, end = addr + size; addr < end; addr = next) { /* * We must ensure that pages backing a given large virtual page * all come from the same physical domain. */ if (vm_ndomains > 1) { domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains; while (VM_DOMAIN_EMPTY(domain)) domain++; next = roundup2(addr + 1, KVA_QUANTUM); if (next > end || next < start) next = end; } else { domain = 0; next = end; } rv = kmem_back_domain(domain, object, addr, next - addr, flags); if (rv != KERN_SUCCESS) { kmem_unback(object, start, addr - start); break; } } return (rv); } /* * kmem_unback: * * Unmap and free the physical pages underlying the specified virtual * address range. * * A physical page must exist within the specified object at each index * that is being unmapped. */ static struct vmem * _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) { struct vmem *arena; vm_page_t m, next; vm_offset_t end, offset; int domain; KASSERT(object == kernel_object, ("kmem_unback: only supports kernel object.")); if (size == 0) return (NULL); pmap_remove(kernel_pmap, addr, addr + size); offset = addr - VM_MIN_KERNEL_ADDRESS; end = offset + size; VM_OBJECT_WLOCK(object); m = vm_page_lookup(object, atop(offset)); domain = vm_page_domain(m); if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0)) arena = vm_dom[domain].vmd_kernel_arena; else arena = vm_dom[domain].vmd_kernel_rwx_arena; for (; offset < end; offset += PAGE_SIZE, m = next) { next = vm_page_next(m); vm_page_xbusy_claim(m); vm_page_unwire_noq(m); vm_page_free(m); } VM_OBJECT_WUNLOCK(object); return (arena); } void kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) { (void)_kmem_unback(object, addr, size); } /* * kmem_free: * * Free memory allocated with kmem_malloc. The size must match the * original allocation. */ void kmem_free(void *addr, vm_size_t size) { struct vmem *arena; size = round_page(size); kasan_mark(addr, size, size, 0); arena = _kmem_unback(kernel_object, (uintptr_t)addr, size); if (arena != NULL) vmem_free(arena, (uintptr_t)addr, size); } /* * kmap_alloc_wait: * * Allocates pageable memory from a sub-map of the kernel. If the submap * has no room, the caller sleeps waiting for more memory in the submap. * * This routine may block. */ vm_offset_t kmap_alloc_wait(vm_map_t map, vm_size_t size) { vm_offset_t addr; size = round_page(size); if (!swap_reserve(size)) return (0); for (;;) { /* * To make this work for more than one map, use the map's lock * to lock out sleepers/wakers. */ vm_map_lock(map); addr = vm_map_findspace(map, vm_map_min(map), size); if (addr + size <= vm_map_max(map)) break; /* no space now; see if we can ever get space */ if (vm_map_max(map) - vm_map_min(map) < size) { vm_map_unlock(map); swap_release(size); return (0); } map->needs_wakeup = TRUE; vm_map_unlock_and_wait(map, 0); } vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW, MAP_ACC_CHARGED); vm_map_unlock(map); return (addr); } /* * kmap_free_wakeup: * * Returns memory to a submap of the kernel, and wakes up any processes * waiting for memory in that map. */ void kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size) { vm_map_lock(map); (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size)); if (map->needs_wakeup) { map->needs_wakeup = FALSE; vm_map_wakeup(map); } vm_map_unlock(map); } void kmem_init_zero_region(void) { vm_offset_t addr, i; vm_page_t m; /* * Map a single physical page of zeros to a larger virtual range. * This requires less looping in places that want large amounts of * zeros, while not using much more physical resources. */ addr = kva_alloc(ZERO_REGION_SIZE); m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO); for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE) pmap_qenter(addr + i, &m, 1); pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ); zero_region = (const void *)addr; } /* * Import KVA from the kernel map into the kernel arena. */ static int kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp) { vm_offset_t addr; int result; TSENTER(); KASSERT((size % KVA_QUANTUM) == 0, ("kva_import: Size %jd is not a multiple of %d", (intmax_t)size, (int)KVA_QUANTUM)); addr = vm_map_min(kernel_map); result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0, VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); if (result != KERN_SUCCESS) { TSEXIT(); return (ENOMEM); } *addrp = addr; TSEXIT(); return (0); } /* * Import KVA from a parent arena into a per-domain arena. Imports must be * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size. */ static int kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp) { KASSERT((size % KVA_QUANTUM) == 0, ("kva_import_domain: Size %jd is not a multiple of %d", (intmax_t)size, (int)KVA_QUANTUM)); return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags, addrp)); } /* * kmem_init: * * Create the kernel map; insert a mapping covering kernel text, * data, bss, and all space allocated thus far (`boostrap' data). The * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and * `start' as allocated, and the range between `start' and `end' as free. * Create the kernel vmem arena and its per-domain children. */ void kmem_init(vm_offset_t start, vm_offset_t end) { vm_size_t quantum; int domain; vm_map_init(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end); kernel_map->system_map = 1; vm_map_lock(kernel_map); /* N.B.: cannot use kgdb to debug, starting with this assignment ... */ (void)vm_map_insert(kernel_map, NULL, 0, #ifdef __amd64__ KERNBASE, #else VM_MIN_KERNEL_ADDRESS, #endif start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); /* ... and ending with the completion of the above `insert' */ #ifdef __amd64__ /* * Mark KVA used for the page array as allocated. Other platforms * that handle vm_page_array allocation can simply adjust virtual_avail * instead. */ (void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array, (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size * sizeof(struct vm_page)), VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); #endif vm_map_unlock(kernel_map); /* * Use a large import quantum on NUMA systems. This helps minimize * interleaving of superpages, reducing internal fragmentation within * the per-domain arenas. */ if (vm_ndomains > 1 && PMAP_HAS_DMAP) quantum = KVA_NUMA_IMPORT_QUANTUM; else quantum = KVA_QUANTUM; /* * Initialize the kernel_arena. This can grow on demand. */ vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0); vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum); for (domain = 0; domain < vm_ndomains; domain++) { /* * Initialize the per-domain arenas. These are used to color * the KVA space in a way that ensures that virtual large pages * are backed by memory from the same physical domain, * maximizing the potential for superpage promotion. */ vm_dom[domain].vmd_kernel_arena = vmem_create( "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); vmem_set_import(vm_dom[domain].vmd_kernel_arena, kva_import_domain, NULL, kernel_arena, quantum); /* * In architectures with superpages, maintain separate arenas * for allocations with permissions that differ from the * "standard" read/write permissions used for kernel memory, * so as not to inhibit superpage promotion. * * Use the base import quantum since this arena is rarely used. */ #if VM_NRESERVLEVEL > 0 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create( "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena, kva_import_domain, (vmem_release_t *)vmem_xfree, kernel_arena, KVA_QUANTUM); #else vm_dom[domain].vmd_kernel_rwx_arena = vm_dom[domain].vmd_kernel_arena; #endif } /* * This must be the very first call so that the virtual address * space used for early allocations is properly marked used in * the map. */ uma_startup2(); } /* * kmem_bootstrap_free: * * Free pages backing preloaded data (e.g., kernel modules) to the * system. Currently only supported on platforms that create a * vm_phys segment for preloaded data. */ void kmem_bootstrap_free(vm_offset_t start, vm_size_t size) { #if defined(__i386__) || defined(__amd64__) struct vm_domain *vmd; vm_offset_t end, va; vm_paddr_t pa; vm_page_t m; end = trunc_page(start + size); start = round_page(start); #ifdef __amd64__ /* * Preloaded files do not have execute permissions by default on amd64. * Restore the default permissions to ensure that the direct map alias * is updated. */ pmap_change_prot(start, end - start, VM_PROT_RW); #endif for (va = start; va < end; va += PAGE_SIZE) { pa = pmap_kextract(va); m = PHYS_TO_VM_PAGE(pa); vmd = vm_pagequeue_domain(m); vm_domain_free_lock(vmd); vm_phys_free_pages(m, 0); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, 1); vm_cnt.v_page_count++; } pmap_remove(kernel_pmap, start, end); (void)vmem_add(kernel_arena, start, end - start, M_WAITOK); #endif } #ifdef PMAP_WANT_ACTIVE_CPUS_NAIVE void pmap_active_cpus(pmap_t pmap, cpuset_t *res) { struct thread *td; struct proc *p; struct vmspace *vm; int c; CPU_ZERO(res); CPU_FOREACH(c) { td = cpuid_to_pcpu[c]->pc_curthread; p = td->td_proc; if (p == NULL) continue; vm = vmspace_acquire_ref(p); if (vm == NULL) continue; if (pmap == vmspace_pmap(vm)) CPU_SET(c, res); vmspace_free(vm); } } #endif /* * Allow userspace to directly trigger the VM drain routine for testing * purposes. */ static int debug_vm_lowmem(SYSCTL_HANDLER_ARGS) { int error, i; i = 0; error = sysctl_handle_int(oidp, &i, 0, req); if (error != 0) return (error); if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0) return (EINVAL); if (i != 0) EVENTHANDLER_INVOKE(vm_lowmem, i); return (0); } SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I", "set to trigger vm_lowmem event with given flags"); static int debug_uma_reclaim(SYSCTL_HANDLER_ARGS) { int error, i; i = 0; error = sysctl_handle_int(oidp, &i, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN && i != UMA_RECLAIM_DRAIN_CPU) return (EINVAL); uma_reclaim(i); return (0); } SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim, "I", "set to generate request to reclaim uma caches"); static int debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS) { int domain, error, request; request = 0; error = sysctl_handle_int(oidp, &request, 0, req); if (error != 0 || req->newptr == NULL) return (error); domain = request >> 4; request &= 0xf; if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN && request != UMA_RECLAIM_DRAIN_CPU) return (EINVAL); if (domain < 0 || domain >= vm_ndomains) return (EINVAL); uma_reclaim_domain(request, domain); return (0); } SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim_domain, "I", ""); diff --git a/sys/vm/vm_page.c b/sys/vm/vm_page.c index 5b5282dc5c9b..6daaaa79163e 100644 --- a/sys/vm/vm_page.c +++ b/sys/vm/vm_page.c @@ -1,5668 +1,5698 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 */ /*- * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Resident memory management module. */ #include #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct vm_domain vm_dom[MAXMEMDOM]; DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; struct mtx_padalign __exclusive_cache_line vm_domainset_lock; /* The following fields are protected by the domainset lock. */ domainset_t __exclusive_cache_line vm_min_domains; domainset_t __exclusive_cache_line vm_severe_domains; static int vm_min_waiters; static int vm_severe_waiters; static int vm_pageproc_waiters; static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "VM page statistics"); static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, CTLFLAG_RD, &pqstate_commit_retries, "Number of failed per-page atomic queue state updates"); static COUNTER_U64_DEFINE_EARLY(queue_ops); SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, CTLFLAG_RD, &queue_ops, "Number of batched queue operations"); static COUNTER_U64_DEFINE_EARLY(queue_nops); SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, CTLFLAG_RD, &queue_nops, "Number of batched queue operations with no effects"); /* * bogus page -- for I/O to/from partially complete buffers, * or for paging into sparsely invalid regions. */ vm_page_t bogus_page; vm_page_t vm_page_array; long vm_page_array_size; long first_page; struct bitset *vm_page_dump; long vm_page_dump_pages; static TAILQ_HEAD(, vm_page) blacklist_head; static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); static uma_zone_t fakepg_zone; static void vm_page_alloc_check(vm_page_t m); static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); static void vm_page_enqueue(vm_page_t m, uint8_t queue); static bool vm_page_free_prep(vm_page_t m); static void vm_page_free_toq(vm_page_t m); static void vm_page_init(void *dummy); static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred); static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred); static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, const uint16_t nflag); static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high); static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req); static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags); static void vm_page_zone_release(void *arg, void **store, int cnt); SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); static void vm_page_init(void *dummy) { fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED); } static int pgcache_zone_max_pcpu; SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0, "Per-CPU page cache size"); /* * The cache page zone is initialized later since we need to be able to allocate * pages before UMA is fully initialized. */ static void vm_page_init_cache_zones(void *dummy __unused) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int cache, domain, maxcache, pool; TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu); maxcache = pgcache_zone_max_pcpu * mp_ncpus; for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (pool = 0; pool < VM_NFREEPOOL; pool++) { pgcache = &vmd->vmd_pgcache[pool]; pgcache->domain = domain; pgcache->pool = pool; pgcache->zone = uma_zcache_create("vm pgcache", PAGE_SIZE, NULL, NULL, NULL, NULL, vm_page_zone_import, vm_page_zone_release, pgcache, UMA_ZONE_VM); /* * Limit each pool's zone to 0.1% of the pages in the * domain. */ cache = maxcache != 0 ? maxcache : vmd->vmd_page_count / 1000; uma_zone_set_maxcache(pgcache->zone, cache); } } } SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ #if PAGE_SIZE == 32768 #ifdef CTASSERT CTASSERT(sizeof(u_long) >= 8); #endif #endif /* * vm_set_page_size: * * Sets the page size, perhaps based upon the memory * size. Must be called before any use of page-size * dependent functions. */ void vm_set_page_size(void) { if (vm_cnt.v_page_size == 0) vm_cnt.v_page_size = PAGE_SIZE; if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) panic("vm_set_page_size: page size not a power of two"); } /* * vm_page_blacklist_next: * * Find the next entry in the provided string of blacklist * addresses. Entries are separated by space, comma, or newline. * If an invalid integer is encountered then the rest of the * string is skipped. Updates the list pointer to the next * character, or NULL if the string is exhausted or invalid. */ static vm_paddr_t vm_page_blacklist_next(char **list, char *end) { vm_paddr_t bad; char *cp, *pos; if (list == NULL || *list == NULL) return (0); if (**list =='\0') { *list = NULL; return (0); } /* * If there's no end pointer then the buffer is coming from * the kenv and we know it's null-terminated. */ if (end == NULL) end = *list + strlen(*list); /* Ensure that strtoq() won't walk off the end */ if (*end != '\0') { if (*end == '\n' || *end == ' ' || *end == ',') *end = '\0'; else { printf("Blacklist not terminated, skipping\n"); *list = NULL; return (0); } } for (pos = *list; *pos != '\0'; pos = cp) { bad = strtoq(pos, &cp, 0); if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { if (bad == 0) { if (++cp < end) continue; else break; } } else break; if (*cp == '\0' || ++cp >= end) *list = NULL; else *list = cp; return (trunc_page(bad)); } printf("Garbage in RAM blacklist, skipping\n"); *list = NULL; return (0); } bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose) { struct vm_domain *vmd; vm_page_t m; bool found; m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) return (true); /* page does not exist, no failure */ vmd = vm_pagequeue_domain(m); vm_domain_free_lock(vmd); found = vm_phys_unfree_page(m); vm_domain_free_unlock(vmd); if (found) { vm_domain_freecnt_inc(vmd, -1); TAILQ_INSERT_TAIL(&blacklist_head, m, listq); if (verbose) printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); } return (found); } /* * vm_page_blacklist_check: * * Iterate through the provided string of blacklist addresses, pulling * each entry out of the physical allocator free list and putting it * onto a list for reporting via the vm.page_blacklist sysctl. */ static void vm_page_blacklist_check(char *list, char *end) { vm_paddr_t pa; char *next; next = list; while (next != NULL) { if ((pa = vm_page_blacklist_next(&next, end)) == 0) continue; vm_page_blacklist_add(pa, bootverbose); } } /* * vm_page_blacklist_load: * * Search for a special module named "ram_blacklist". It'll be a * plain text file provided by the user via the loader directive * of the same name. */ static void vm_page_blacklist_load(char **list, char **end) { void *mod; u_char *ptr; u_int len; mod = NULL; ptr = NULL; mod = preload_search_by_type("ram_blacklist"); if (mod != NULL) { ptr = preload_fetch_addr(mod); len = preload_fetch_size(mod); } if (ptr != NULL && len > 0) { *list = ptr; *end = ptr + len - 1; } else { *list = NULL; *end = NULL; } return; } static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) { vm_page_t m; struct sbuf sbuf; int error, first; first = 1; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); TAILQ_FOREACH(m, &blacklist_head, listq) { sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", (uintmax_t)m->phys_addr); first = 0; } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * Initialize a dummy page for use in scans of the specified paging queue. * In principle, this function only needs to set the flag PG_MARKER. * Nonetheless, it write busies the page as a safety precaution. */ void vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) { bzero(marker, sizeof(*marker)); marker->flags = PG_MARKER; marker->a.flags = aflags; marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; marker->a.queue = queue; } static void vm_page_domain_init(int domain) { struct vm_domain *vmd; struct vm_pagequeue *pq; int i; vmd = VM_DOMAIN(domain); bzero(vmd, sizeof(*vmd)); *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = "vm inactive pagequeue"; *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = "vm active pagequeue"; *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = "vm laundry pagequeue"; *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = "vm unswappable pagequeue"; vmd->vmd_domain = domain; vmd->vmd_page_count = 0; vmd->vmd_free_count = 0; vmd->vmd_segs = 0; vmd->vmd_oom = false; vmd->vmd_helper_threads_enabled = true; for (i = 0; i < PQ_COUNT; i++) { pq = &vmd->vmd_pagequeues[i]; TAILQ_INIT(&pq->pq_pl); mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", MTX_DEF | MTX_DUPOK); pq->pq_pdpages = 0; vm_page_init_marker(&vmd->vmd_markers[i], i, 0); } mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); /* * inacthead is used to provide FIFO ordering for LRU-bypassing * insertions. */ vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, &vmd->vmd_inacthead, plinks.q); /* * The clock pages are used to implement active queue scanning without * requeues. Scans start at clock[0], which is advanced after the scan * ends. When the two clock hands meet, they are reset and scanning * resumes from the head of the queue. */ vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[1], plinks.q); } /* * Initialize a physical page in preparation for adding it to the free * lists. */ void vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) { m->object = NULL; m->ref_count = 0; m->busy_lock = VPB_FREED; m->flags = m->a.flags = 0; m->phys_addr = pa; m->a.queue = PQ_NONE; m->psind = 0; m->segind = segind; m->order = VM_NFREEORDER; m->pool = VM_FREEPOOL_DEFAULT; m->valid = m->dirty = 0; pmap_page_init(m); } #ifndef PMAP_HAS_PAGE_ARRAY static vm_paddr_t vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) { vm_paddr_t new_end; /* * Reserve an unmapped guard page to trap access to vm_page_array[-1]. * However, because this page is allocated from KVM, out-of-bounds * accesses using the direct map will not be trapped. */ *vaddr += PAGE_SIZE; /* * Allocate physical memory for the page structures, and map it. */ new_end = trunc_page(end - page_range * sizeof(struct vm_page)); vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); vm_page_array_size = page_range; return (new_end); } #endif /* * vm_page_startup: * * Initializes the resident memory module. Allocates physical memory for * bootstrapping UMA and some data structures that are used to manage * physical pages. Initializes these structures, and populates the free * page queues. */ vm_offset_t vm_page_startup(vm_offset_t vaddr) { struct vm_phys_seg *seg; struct vm_domain *vmd; vm_page_t m; char *list, *listend; vm_paddr_t end, high_avail, low_avail, new_end, size; vm_paddr_t page_range __unused; vm_paddr_t last_pa, pa, startp, endp; u_long pagecount; #if MINIDUMP_PAGE_TRACKING u_long vm_page_dump_size; #endif int biggestone, i, segind; #ifdef WITNESS vm_offset_t mapped; int witness_size; #endif #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) long ii; #endif vaddr = round_page(vaddr); vm_phys_early_startup(); biggestone = vm_phys_avail_largest(); end = phys_avail[biggestone+1]; /* * Initialize the page and queue locks. */ mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); for (i = 0; i < PA_LOCK_COUNT; i++) mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); for (i = 0; i < vm_ndomains; i++) vm_page_domain_init(i); new_end = end; #ifdef WITNESS witness_size = round_page(witness_startup_count()); new_end -= witness_size; mapped = pmap_map(&vaddr, new_end, new_end + witness_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, witness_size); witness_startup((void *)mapped); #endif #if MINIDUMP_PAGE_TRACKING /* * Allocate a bitmap to indicate that a random physical page * needs to be included in a minidump. * * The amd64 port needs this to indicate which direct map pages * need to be dumped, via calls to dump_add_page()/dump_drop_page(). * * However, i386 still needs this workspace internally within the * minidump code. In theory, they are not needed on i386, but are * included should the sf_buf code decide to use them. */ last_pa = 0; vm_page_dump_pages = 0; for (i = 0; dump_avail[i + 1] != 0; i += 2) { vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - dump_avail[i] / PAGE_SIZE; if (dump_avail[i + 1] > last_pa) last_pa = dump_avail[i + 1]; } vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); new_end -= vm_page_dump_size; vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)vm_page_dump, vm_page_dump_size); #else (void)last_pa; #endif #if defined(__aarch64__) || defined(__amd64__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include the UMA bootstrap pages, witness pages and vm_page_dump * in a crash dump. When pmap_map() uses the direct map, they are * not automatically included. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; #ifdef __amd64__ /* * Request that the physical pages underlying the message buffer be * included in a crash dump. Since the message buffer is accessed * through the direct map, they are not automatically included. */ pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); last_pa = pa + round_page(msgbufsize); while (pa < last_pa) { dump_add_page(pa); pa += PAGE_SIZE; } #endif /* * Determine the lowest and highest physical addresses and, in the case * of VM_PHYSSEG_SPARSE, the exact size of the available physical * memory. vm_phys_early_startup() already checked that phys_avail[] * has at least one element. */ #ifdef VM_PHYSSEG_SPARSE size = phys_avail[1] - phys_avail[0]; #endif low_avail = phys_avail[0]; high_avail = phys_avail[1]; for (i = 2; phys_avail[i + 1] != 0; i += 2) { #ifdef VM_PHYSSEG_SPARSE size += phys_avail[i + 1] - phys_avail[i]; #endif if (phys_avail[i] < low_avail) low_avail = phys_avail[i]; if (phys_avail[i + 1] > high_avail) high_avail = phys_avail[i + 1]; } for (i = 0; i < vm_phys_nsegs; i++) { #ifdef VM_PHYSSEG_SPARSE size += vm_phys_segs[i].end - vm_phys_segs[i].start; #endif if (vm_phys_segs[i].start < low_avail) low_avail = vm_phys_segs[i].start; if (vm_phys_segs[i].end > high_avail) high_avail = vm_phys_segs[i].end; } first_page = low_avail / PAGE_SIZE; #ifdef VM_PHYSSEG_DENSE size = high_avail - low_avail; #endif #ifdef PMAP_HAS_PAGE_ARRAY pmap_page_array_startup(size / PAGE_SIZE); biggestone = vm_phys_avail_largest(); end = new_end = phys_avail[biggestone + 1]; #else #ifdef VM_PHYSSEG_DENSE /* * In the VM_PHYSSEG_DENSE case, the number of pages can account for * the overhead of a page structure per page only if vm_page_array is * allocated from the last physical memory chunk. Otherwise, we must * allocate page structures representing the physical memory * underlying vm_page_array, even though they will not be used. */ if (new_end != high_avail) page_range = size / PAGE_SIZE; else #endif { page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); /* * If the partial bytes remaining are large enough for * a page (PAGE_SIZE) without a corresponding * 'struct vm_page', then new_end will contain an * extra page after subtracting the length of the VM * page array. Compensate by subtracting an extra * page from new_end. */ if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { if (new_end == high_avail) high_avail -= PAGE_SIZE; new_end -= PAGE_SIZE; } } end = new_end; new_end = vm_page_array_alloc(&vaddr, end, page_range); #endif #if VM_NRESERVLEVEL > 0 /* * Allocate physical memory for the reservation management system's * data structures, and map it. */ new_end = vm_reserv_startup(&vaddr, new_end); #endif #if defined(__aarch64__) || defined(__amd64__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include vm_page_array and vm_reserv_array in a crash dump. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; /* * Add physical memory segments corresponding to the available * physical pages. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); /* * Initialize the physical memory allocator. */ vm_phys_init(); /* * Initialize the page structures and add every available page to the * physical memory allocator's free lists. */ #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) for (ii = 0; ii < vm_page_array_size; ii++) { m = &vm_page_array[ii]; vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); m->flags = PG_FICTITIOUS; } #endif vm_cnt.v_page_count = 0; for (segind = 0; segind < vm_phys_nsegs; segind++) { seg = &vm_phys_segs[segind]; for (m = seg->first_page, pa = seg->start; pa < seg->end; m++, pa += PAGE_SIZE) vm_page_init_page(m, pa, segind); /* * Add the segment's pages that are covered by one of * phys_avail's ranges to the free lists. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) { if (seg->end <= phys_avail[i] || seg->start >= phys_avail[i + 1]) continue; startp = MAX(seg->start, phys_avail[i]); endp = MIN(seg->end, phys_avail[i + 1]); pagecount = (u_long)atop(endp - startp); if (pagecount == 0) continue; m = seg->first_page + atop(startp - seg->start); vmd = VM_DOMAIN(seg->domain); vm_domain_free_lock(vmd); vm_phys_enqueue_contig(m, pagecount); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, pagecount); vm_cnt.v_page_count += (u_int)pagecount; vmd->vmd_page_count += (u_int)pagecount; vmd->vmd_segs |= 1UL << segind; } } /* * Remove blacklisted pages from the physical memory allocator. */ TAILQ_INIT(&blacklist_head); vm_page_blacklist_load(&list, &listend); vm_page_blacklist_check(list, listend); list = kern_getenv("vm.blacklist"); vm_page_blacklist_check(list, NULL); freeenv(list); #if VM_NRESERVLEVEL > 0 /* * Initialize the reservation management system. */ vm_reserv_init(); #endif return (vaddr); } void vm_page_reference(vm_page_t m) { vm_page_aflag_set(m, PGA_REFERENCED); } /* * vm_page_trybusy * * Helper routine for grab functions to trylock busy. * * Returns true on success and false on failure. */ static bool vm_page_trybusy(vm_page_t m, int allocflags) { if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) return (vm_page_trysbusy(m)); else return (vm_page_tryxbusy(m)); } /* * vm_page_tryacquire * * Helper routine for grab functions to trylock busy and wire. * * Returns true on success and false on failure. */ static inline bool vm_page_tryacquire(vm_page_t m, int allocflags) { bool locked; locked = vm_page_trybusy(m, allocflags); if (locked && (allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); return (locked); } /* * vm_page_busy_acquire: * * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop * and drop the object lock if necessary. */ bool vm_page_busy_acquire(vm_page_t m, int allocflags) { vm_object_t obj; bool locked; /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = atomic_load_ptr(&m->object); for (;;) { if (vm_page_tryacquire(m, allocflags)) return (true); if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (false); if (obj != NULL) locked = VM_OBJECT_WOWNED(obj); else locked = false; MPASS(locked || vm_page_wired(m)); if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, locked) && locked) VM_OBJECT_WLOCK(obj); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (false); KASSERT(m->object == obj || m->object == NULL, ("vm_page_busy_acquire: page %p does not belong to %p", m, obj)); } } /* * vm_page_busy_downgrade: * * Downgrade an exclusive busy page into a single shared busy page. */ void vm_page_busy_downgrade(vm_page_t m) { u_int x; vm_page_assert_xbusied(m); x = vm_page_busy_fetch(m); for (;;) { if (atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_SHARERS_WORD(1))) break; } if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); } /* * * vm_page_busy_tryupgrade: * * Attempt to upgrade a single shared busy into an exclusive busy. */ int vm_page_busy_tryupgrade(vm_page_t m) { u_int ce, x; vm_page_assert_sbusied(m); x = vm_page_busy_fetch(m); ce = VPB_CURTHREAD_EXCLUSIVE; for (;;) { if (VPB_SHARERS(x) > 1) return (0); KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_busy_tryupgrade: invalid lock state")); if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, ce | (x & VPB_BIT_WAITERS))) continue; return (1); } } /* * vm_page_sbusied: * * Return a positive value if the page is shared busied, 0 otherwise. */ int vm_page_sbusied(vm_page_t m) { u_int x; x = vm_page_busy_fetch(m); return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); } /* * vm_page_sunbusy: * * Shared unbusy a page. */ void vm_page_sunbusy(vm_page_t m) { u_int x; vm_page_assert_sbusied(m); x = vm_page_busy_fetch(m); for (;;) { KASSERT(x != VPB_FREED, ("vm_page_sunbusy: Unlocking freed page.")); if (VPB_SHARERS(x) > 1) { if (atomic_fcmpset_int(&m->busy_lock, &x, x - VPB_ONE_SHARER)) break; continue; } KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_sunbusy: invalid lock state")); if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) continue; if ((x & VPB_BIT_WAITERS) == 0) break; wakeup(m); break; } } /* * vm_page_busy_sleep: * * Sleep if the page is busy, using the page pointer as wchan. * This is used to implement the hard-path of the busying mechanism. * * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function * will not sleep if the page is shared-busy. * * The object lock must be held on entry. * * Returns true if it slept and dropped the object lock, or false * if there was no sleep and the lock is still held. */ bool vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) { vm_object_t obj; obj = m->object; VM_OBJECT_ASSERT_LOCKED(obj); return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, true)); } /* * vm_page_busy_sleep_unlocked: * * Sleep if the page is busy, using the page pointer as wchan. * This is used to implement the hard-path of busying mechanism. * * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function * will not sleep if the page is shared-busy. * * The object lock must not be held on entry. The operation will * return if the page changes identity. */ void vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags) { VM_OBJECT_ASSERT_UNLOCKED(obj); (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); } /* * _vm_page_busy_sleep: * * Internal busy sleep function. Verifies the page identity and * lockstate against parameters. Returns true if it sleeps and * false otherwise. * * allocflags uses VM_ALLOC_* flags to specify the lock required. * * If locked is true the lock will be dropped for any true returns * and held for any false returns. */ static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked) { bool xsleep; u_int x; /* * If the object is busy we must wait for that to drain to zero * before trying the page again. */ if (obj != NULL && vm_object_busied(obj)) { if (locked) VM_OBJECT_DROP(obj); vm_object_busy_wait(obj, wmesg); return (true); } if (!vm_page_busied(m)) return (false); xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; sleepq_lock(m); x = vm_page_busy_fetch(m); do { /* * If the page changes objects or becomes unlocked we can * simply return. */ if (x == VPB_UNBUSIED || (xsleep && (x & VPB_BIT_SHARED) != 0) || m->object != obj || m->pindex != pindex) { sleepq_release(m); return (false); } if ((x & VPB_BIT_WAITERS) != 0) break; } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); if (locked) VM_OBJECT_DROP(obj); DROP_GIANT(); sleepq_add(m, NULL, wmesg, 0, 0); sleepq_wait(m, PVM); PICKUP_GIANT(); return (true); } /* * vm_page_trysbusy: * * Try to shared busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_trysbusy(vm_page_t m) { vm_object_t obj; u_int x; obj = m->object; x = vm_page_busy_fetch(m); for (;;) { if ((x & VPB_BIT_SHARED) == 0) return (0); /* * Reduce the window for transient busies that will trigger * false negatives in vm_page_ps_test(). */ if (obj != NULL && vm_object_busied(obj)) return (0); if (atomic_fcmpset_acq_int(&m->busy_lock, &x, x + VPB_ONE_SHARER)) break; } /* Refetch the object now that we're guaranteed that it is stable. */ obj = m->object; if (obj != NULL && vm_object_busied(obj)) { vm_page_sunbusy(m); return (0); } return (1); } /* * vm_page_tryxbusy: * * Try to exclusive busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_tryxbusy(vm_page_t m) { vm_object_t obj; if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, VPB_CURTHREAD_EXCLUSIVE) == 0) return (0); obj = m->object; if (obj != NULL && vm_object_busied(obj)) { vm_page_xunbusy(m); return (0); } return (1); } static void vm_page_xunbusy_hard_tail(vm_page_t m) { atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); /* Wake the waiter. */ wakeup(m); } /* * vm_page_xunbusy_hard: * * Called when unbusy has failed because there is a waiter. */ void vm_page_xunbusy_hard(vm_page_t m) { vm_page_assert_xbusied(m); vm_page_xunbusy_hard_tail(m); } void vm_page_xunbusy_hard_unchecked(vm_page_t m) { vm_page_assert_xbusied_unchecked(m); vm_page_xunbusy_hard_tail(m); } static void vm_page_busy_free(vm_page_t m) { u_int x; atomic_thread_fence_rel(); x = atomic_swap_int(&m->busy_lock, VPB_FREED); if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); } /* * vm_page_unhold_pages: * * Unhold each of the pages that is referenced by the given array. */ void vm_page_unhold_pages(vm_page_t *ma, int count) { for (; count != 0; count--) { vm_page_unwire(*ma, PQ_ACTIVE); ma++; } } vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa) { vm_page_t m; #ifdef VM_PHYSSEG_SPARSE m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) m = vm_phys_fictitious_to_vm_page(pa); return (m); #elif defined(VM_PHYSSEG_DENSE) long pi; pi = atop(pa); if (pi >= first_page && (pi - first_page) < vm_page_array_size) { m = &vm_page_array[pi - first_page]; return (m); } return (vm_phys_fictitious_to_vm_page(pa)); #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif } /* * vm_page_getfake: * * Create a fictitious page with the specified physical address and * memory attribute. The memory attribute is the only the machine- * dependent aspect of a fictitious page that must be initialized. */ vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) { vm_page_t m; m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); vm_page_initfake(m, paddr, memattr); return (m); } void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { if ((m->flags & PG_FICTITIOUS) != 0) { /* * The page's memattr might have changed since the * previous initialization. Update the pmap to the * new memattr. */ goto memattr; } m->phys_addr = paddr; m->a.queue = PQ_NONE; /* Fictitious pages don't use "segind". */ m->flags = PG_FICTITIOUS; /* Fictitious pages don't use "order" or "pool". */ m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; /* Fictitious pages are unevictable. */ m->ref_count = 1; pmap_page_init(m); memattr: pmap_page_set_memattr(m, memattr); } /* * vm_page_putfake: * * Release a fictitious page. */ void vm_page_putfake(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_putfake: bad page %p", m)); vm_page_assert_xbusied(m); vm_page_busy_free(m); uma_zfree(fakepg_zone, m); } /* * vm_page_updatefake: * * Update the given fictitious page to the specified physical address and * memory attribute. */ void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_updatefake: bad page %p", m)); m->phys_addr = paddr; pmap_page_set_memattr(m, memattr); } /* * vm_page_free: * * Free a page. */ void vm_page_free(vm_page_t m) { m->flags &= ~PG_ZERO; vm_page_free_toq(m); } /* * vm_page_free_zero: * * Free a page to the zerod-pages queue */ void vm_page_free_zero(vm_page_t m) { m->flags |= PG_ZERO; vm_page_free_toq(m); } /* * Unbusy and handle the page queueing for a page from a getpages request that * was optionally read ahead or behind. */ void vm_page_readahead_finish(vm_page_t m) { /* We shouldn't put invalid pages on queues. */ KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); /* * Since the page is not the actually needed one, whether it should * be activated or deactivated is not obvious. Empirical results * have shown that deactivating the page is usually the best choice, * unless the page is wanted by another thread. */ if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); vm_page_xunbusy_unchecked(m); } /* * Destroy the identity of an invalid page and free it if possible. * This is intended to be used when reading a page from backing store fails. */ void vm_page_free_invalid(vm_page_t m) { KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); KASSERT(m->object != NULL, ("page %p has no object", m)); VM_OBJECT_ASSERT_WLOCKED(m->object); /* * We may be attempting to free the page as part of the handling for an * I/O error, in which case the page was xbusied by a different thread. */ vm_page_xbusy_claim(m); /* * If someone has wired this page while the object lock * was not held, then the thread that unwires is responsible * for freeing the page. Otherwise just free the page now. * The wire count of this unmapped page cannot change while * we have the page xbusy and the page's object wlocked. */ if (vm_page_remove(m)) vm_page_free(m); } /* * vm_page_dirty_KBI: [ internal use only ] * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). * * This function should only be called by vm_page_dirty(). */ void vm_page_dirty_KBI(vm_page_t m) { /* Refer to this operation by its public name. */ KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); m->dirty = VM_PAGE_BITS_ALL; } /* * vm_page_insert: [ internal use only ] * * Inserts the given mem entry into the object and object list. * * The object must be locked. */ int vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) { vm_page_t mpred; VM_OBJECT_ASSERT_WLOCKED(object); mpred = vm_radix_lookup_le(&object->rtree, pindex); return (vm_page_insert_after(m, object, pindex, mpred)); } /* * vm_page_insert_after: * * Inserts the page "m" into the specified object at offset "pindex". * * The page "mpred" must immediately precede the offset "pindex" within * the specified object. * * The object must be locked. */ static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred) { vm_page_t msucc; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(m->object == NULL, ("vm_page_insert_after: page already inserted")); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_after: object doesn't contain mpred")); KASSERT(mpred->pindex < pindex, ("vm_page_insert_after: mpred doesn't precede pindex")); msucc = TAILQ_NEXT(mpred, listq); } else msucc = TAILQ_FIRST(&object->memq); if (msucc != NULL) KASSERT(msucc->pindex > pindex, ("vm_page_insert_after: msucc doesn't succeed pindex")); /* * Record the object/offset pair in this page. */ m->object = object; m->pindex = pindex; m->ref_count |= VPRC_OBJREF; /* * Now link into the object's ordered list of backed pages. */ if (vm_radix_insert(&object->rtree, m)) { m->object = NULL; m->pindex = 0; m->ref_count &= ~VPRC_OBJREF; return (1); } vm_page_insert_radixdone(m, object, mpred); vm_pager_page_inserted(object, m); return (0); } /* * vm_page_insert_radixdone: * * Complete page "m" insertion into the specified object after the * radix trie hooking. * * The page "mpred" must precede the offset "m->pindex" within the * specified object. * * The object must be locked. */ static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object != NULL && m->object == object, ("vm_page_insert_radixdone: page %p has inconsistent object", m)); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("vm_page_insert_radixdone: page %p is missing object ref", m)); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_radixdone: object doesn't contain mpred")); KASSERT(mpred->pindex < m->pindex, ("vm_page_insert_radixdone: mpred doesn't precede pindex")); } if (mpred != NULL) TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); else TAILQ_INSERT_HEAD(&object->memq, m, listq); /* * Show that the object has one more resident page. */ object->resident_page_count++; /* * Hold the vnode until the last page is released. */ if (object->resident_page_count == 1 && object->type == OBJT_VNODE) vhold(object->handle); /* * Since we are inserting a new and possibly dirty page, * update the object's generation count. */ if (pmap_page_is_write_mapped(m)) vm_object_set_writeable_dirty(object); } /* * Do the work to remove a page from its object. The caller is responsible for * updating the page's fields to reflect this removal. */ static void vm_page_object_remove(vm_page_t m) { vm_object_t object; vm_page_t mrem __diagused; vm_page_assert_xbusied(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("page %p is missing its object ref", m)); /* Deferred free of swap space. */ if ((m->a.flags & PGA_SWAP_FREE) != 0) vm_pager_page_unswapped(m); vm_pager_page_removed(object, m); m->object = NULL; mrem = vm_radix_remove(&object->rtree, m->pindex); KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); /* * Now remove from the object's list of backed pages. */ TAILQ_REMOVE(&object->memq, m, listq); /* * And show that the object has one fewer resident page. */ object->resident_page_count--; /* * The vnode may now be recycled. */ if (object->resident_page_count == 0 && object->type == OBJT_VNODE) vdrop(object->handle); } /* * vm_page_remove: * * Removes the specified page from its containing object, but does not * invalidate any backing storage. Returns true if the object's reference * was the last reference to the page, and false otherwise. * * The object must be locked and the page must be exclusively busied. * The exclusive busy will be released on return. If this is not the * final ref and the caller does not hold a wire reference it may not * continue to access the page. */ bool vm_page_remove(vm_page_t m) { bool dropped; dropped = vm_page_remove_xbusy(m); vm_page_xunbusy(m); return (dropped); } /* * vm_page_remove_xbusy * * Removes the page but leaves the xbusy held. Returns true if this * removed the final ref and false otherwise. */ bool vm_page_remove_xbusy(vm_page_t m) { vm_page_object_remove(m); return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); } /* * vm_page_lookup: * * Returns the page associated with the object/offset * pair specified; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_lookup(vm_object_t object, vm_pindex_t pindex) { VM_OBJECT_ASSERT_LOCKED(object); return (vm_radix_lookup(&object->rtree, pindex)); } /* * vm_page_lookup_unlocked: * * Returns the page associated with the object/offset pair specified; * if none is found, NULL is returned. The page may be no longer be * present in the object at the time that this function returns. Only * useful for opportunistic checks such as inmem(). */ vm_page_t vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) { return (vm_radix_lookup_unlocked(&object->rtree, pindex)); } /* * vm_page_relookup: * * Returns a page that must already have been busied by * the caller. Used for bogus page replacement. */ vm_page_t vm_page_relookup(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; m = vm_radix_lookup_unlocked(&object->rtree, pindex); KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && m->object == object && m->pindex == pindex, ("vm_page_relookup: Invalid page %p", m)); return (m); } /* * This should only be used by lockless functions for releasing transient * incorrect acquires. The page may have been freed after we acquired a * busy lock. In this case busy_lock == VPB_FREED and we have nothing * further to do. */ static void vm_page_busy_release(vm_page_t m) { u_int x; x = vm_page_busy_fetch(m); for (;;) { if (x == VPB_FREED) break; if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { if (atomic_fcmpset_int(&m->busy_lock, &x, x - VPB_ONE_SHARER)) break; continue; } KASSERT((x & VPB_BIT_SHARED) != 0 || (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, ("vm_page_busy_release: %p xbusy not owned.", m)); if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) continue; if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); break; } } /* * vm_page_find_least: * * Returns the page associated with the object with least pindex * greater than or equal to the parameter pindex, or NULL. * * The object must be locked. */ vm_page_t vm_page_find_least(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; VM_OBJECT_ASSERT_LOCKED(object); if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) m = vm_radix_lookup_ge(&object->rtree, pindex); return (m); } /* * Returns the given page's successor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_next(vm_page_t m) { vm_page_t next; VM_OBJECT_ASSERT_LOCKED(m->object); if ((next = TAILQ_NEXT(m, listq)) != NULL) { MPASS(next->object == m->object); if (next->pindex != m->pindex + 1) next = NULL; } return (next); } /* * Returns the given page's predecessor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_prev(vm_page_t m) { vm_page_t prev; VM_OBJECT_ASSERT_LOCKED(m->object); if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { MPASS(prev->object == m->object); if (prev->pindex != m->pindex - 1) prev = NULL; } return (prev); } /* * Uses the page mnew as a replacement for an existing page at index * pindex which must be already present in the object. * * Both pages must be exclusively busied on enter. The old page is * unbusied on exit. * * A return value of true means mold is now free. If this is not the * final ref and the caller does not hold a wire reference it may not * continue to access the page. */ static bool vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold) { vm_page_t mret __diagused; bool dropped; VM_OBJECT_ASSERT_WLOCKED(object); vm_page_assert_xbusied(mold); KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, ("vm_page_replace: page %p already in object", mnew)); /* * This function mostly follows vm_page_insert() and * vm_page_remove() without the radix, object count and vnode * dance. Double check such functions for more comments. */ mnew->object = object; mnew->pindex = pindex; atomic_set_int(&mnew->ref_count, VPRC_OBJREF); mret = vm_radix_replace(&object->rtree, mnew); KASSERT(mret == mold, ("invalid page replacement, mold=%p, mret=%p", mold, mret)); KASSERT((mold->oflags & VPO_UNMANAGED) == (mnew->oflags & VPO_UNMANAGED), ("vm_page_replace: mismatched VPO_UNMANAGED")); /* Keep the resident page list in sorted order. */ TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); TAILQ_REMOVE(&object->memq, mold, listq); mold->object = NULL; /* * The object's resident_page_count does not change because we have * swapped one page for another, but the generation count should * change if the page is dirty. */ if (pmap_page_is_write_mapped(mnew)) vm_object_set_writeable_dirty(object); dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; vm_page_xunbusy(mold); return (dropped); } void vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold) { vm_page_assert_xbusied(mnew); if (vm_page_replace_hold(mnew, object, pindex, mold)) vm_page_free(mold); } /* * vm_page_rename: * * Move the given memory entry from its * current object to the specified target object/offset. * * This routine dirties the page if it is valid, as callers are expected to * transfer backing storage only after moving the page. Dirtying the page * ensures that the destination object retains the most recent copy of the * page. * * The objects must be locked. */ int vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) { vm_page_t mpred; vm_pindex_t opidx; VM_OBJECT_ASSERT_WLOCKED(new_object); KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); KASSERT(mpred == NULL || mpred->pindex != new_pindex, ("vm_page_rename: pindex already renamed")); /* * Create a custom version of vm_page_insert() which does not depend * by m_prev and can cheat on the implementation aspects of the * function. */ opidx = m->pindex; m->pindex = new_pindex; if (vm_radix_insert(&new_object->rtree, m)) { m->pindex = opidx; return (1); } /* * The operation cannot fail anymore. The removal must happen before * the listq iterator is tainted. */ m->pindex = opidx; vm_page_object_remove(m); /* Return back to the new pindex to complete vm_page_insert(). */ m->pindex = new_pindex; m->object = new_object; vm_page_insert_radixdone(m, new_object, mpred); if (vm_page_any_valid(m)) vm_page_dirty(m); vm_pager_page_inserted(new_object, m); return (0); } /* * vm_page_alloc: * * Allocate and return a page that is associated with the specified * object and offset pair. By default, this page is exclusive busied. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) { return (vm_page_alloc_after(object, pindex, req, vm_radix_lookup_le(&object->rtree, pindex))); } vm_page_t vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req) { return (vm_page_alloc_domain_after(object, pindex, domain, req, vm_radix_lookup_le(&object->rtree, pindex))); } /* * Allocate a page in the specified object with the given page index. To * optimize insertion of the page into the object, the caller must also specifiy * the resident page in the object with largest index smaller than the given * page index, or NULL if no such page exists. */ vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req, vm_page_t mpred) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_domain_after(object, pindex, domain, req, mpred); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } /* * Returns true if the number of free pages exceeds the minimum * for the request class and false otherwise. */ static int _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) { u_int limit, old, new; if (req_class == VM_ALLOC_INTERRUPT) limit = 0; else if (req_class == VM_ALLOC_SYSTEM) limit = vmd->vmd_interrupt_free_min; else limit = vmd->vmd_free_reserved; /* * Attempt to reserve the pages. Fail if we're below the limit. */ limit += npages; old = atomic_load_int(&vmd->vmd_free_count); do { if (old < limit) return (0); new = old - npages; } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); /* Wake the page daemon if we've crossed the threshold. */ if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) pagedaemon_wakeup(vmd->vmd_domain); /* Only update bitsets on transitions. */ if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) vm_domain_set(vmd); return (1); } int vm_domain_allocate(struct vm_domain *vmd, int req, int npages) { int req_class; /* * The page daemon is allowed to dig deeper into the free page list. */ req_class = req & VM_ALLOC_CLASS_MASK; if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; return (_vm_domain_allocate(vmd, req_class, npages)); } vm_page_t vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, int req, vm_page_t mpred) { struct vm_domain *vmd; vm_page_t m; int flags; #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \ VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \ VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) KASSERT((req & ~VPA_FLAGS) == 0, ("invalid request %#x", req)); KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("invalid request %#x", req)); KASSERT(mpred == NULL || mpred->pindex < pindex, ("mpred %p doesn't precede pindex 0x%jx", mpred, (uintmax_t)pindex)); VM_OBJECT_ASSERT_WLOCKED(object); flags = 0; m = NULL; if (!vm_pager_can_alloc_page(object, pindex)) return (NULL); again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the page from a reservation? */ if (vm_object_reserv(object) && (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != NULL) { goto found; } #endif vmd = VM_DOMAIN(domain); if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, M_NOWAIT | M_NOVM); if (m != NULL) { flags |= PG_PCPU_CACHE; goto found; } } if (vm_domain_allocate(vmd, req, 1)) { /* * If not, allocate it from the free page queues. */ vm_domain_free_lock(vmd); m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); vm_domain_free_unlock(vmd); if (m == NULL) { vm_domain_freecnt_inc(vmd, 1); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_inactive(domain)) goto again; #endif } } if (m == NULL) { /* * Not allocatable, give up. */ if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } /* * At this point we had better have found a good page. */ found: vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ flags |= m->flags & PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; m->flags = flags; m->a.flags = 0; m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; else if ((req & VM_ALLOC_SBUSY) != 0) m->busy_lock = VPB_SHARERS_WORD(1); else m->busy_lock = VPB_UNBUSIED; if (req & VM_ALLOC_WIRED) { vm_wire_add(1); m->ref_count = 1; } m->a.act_count = 0; if (vm_page_insert_after(m, object, pindex, mpred)) { if (req & VM_ALLOC_WIRED) { vm_wire_sub(1); m->ref_count = 0; } KASSERT(m->object == NULL, ("page %p has object", m)); m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } /* Ignore device objects; the pager sets "memattr" for them. */ if (object->memattr != VM_MEMATTR_DEFAULT && (object->flags & OBJ_FICTITIOUS) == 0) pmap_page_set_memattr(m, object->memattr); return (m); } /* * vm_page_alloc_contig: * * Allocate a contiguous set of physical pages of the given size "npages" * from the free lists. All of the physical pages must be at or above * the given physical address "low" and below the given physical address * "high". The given value "alignment" determines the alignment of the * first physical page in the set. If the given value "boundary" is * non-zero, then the set of physical pages cannot cross any physical * address boundary that is a multiple of that value. Both "alignment" * and "boundary" must be a power of two. * * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, * then the memory attribute setting for the physical pages is configured * to the object's memory attribute setting. Otherwise, the memory * attribute setting for the physical pages is configured to "memattr", * overriding the object's memory attribute setting. However, if the * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the * memory attribute setting for the physical pages cannot be configured * to VM_MEMATTR_DEFAULT. * * The specified object may not contain fictitious pages. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; + vm_page_t bounds[2]; vm_page_t m; int domain; + int start_segind; + + start_segind = -1; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_contig_domain(object, pindex, domain, req, npages, low, high, alignment, boundary, memattr); if (m != NULL) break; + if (start_segind == -1) + start_segind = vm_phys_lookup_segind(low); + if (vm_phys_find_range(bounds, start_segind, domain, + npages, low, high) == -1) { + vm_domainset_iter_ignore(&di, domain); + } } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } static vm_page_t vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domain *vmd; vm_page_t m_ret; /* * Can we allocate the pages without the number of free pages falling * below the lower bound for the allocation class? */ vmd = VM_DOMAIN(domain); if (!vm_domain_allocate(vmd, req, npages)) return (NULL); /* * Try to allocate the pages from the free page queues. */ vm_domain_free_lock(vmd); m_ret = vm_phys_alloc_contig(domain, npages, low, high, alignment, boundary); vm_domain_free_unlock(vmd); if (m_ret != NULL) return (m_ret); #if VM_NRESERVLEVEL > 0 /* * Try to break a reservation to allocate the pages. */ if ((req & VM_ALLOC_NORECLAIM) == 0) { m_ret = vm_reserv_reclaim_contig(domain, npages, low, high, alignment, boundary); if (m_ret != NULL) return (m_ret); } #endif vm_domain_freecnt_inc(vmd, npages); return (NULL); } vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { vm_page_t m, m_ret, mpred; u_int busy_lock, flags, oflags; #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM) KASSERT((req & ~VPAC_FLAGS) == 0, ("invalid request %#x", req)); KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("invalid request %#x", req)); KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), ("invalid request %#x", req)); VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_FICTITIOUS) == 0, ("vm_page_alloc_contig: object %p has fictitious pages", object)); KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); mpred = vm_radix_lookup_le(&object->rtree, pindex); KASSERT(mpred == NULL || mpred->pindex != pindex, ("vm_page_alloc_contig: pindex already allocated")); for (;;) { #if VM_NRESERVLEVEL > 0 /* * Can we allocate the pages from a reservation? */ if (vm_object_reserv(object) && (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, mpred, npages, low, high, alignment, boundary)) != NULL) { break; } #endif if ((m_ret = vm_page_find_contig_domain(domain, req, npages, low, high, alignment, boundary)) != NULL) break; if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req)) return (NULL); } for (m = m_ret; m < &m_ret[npages]; m++) { vm_page_dequeue(m); vm_page_alloc_check(m); } /* * Initialize the pages. Only the PG_ZERO flag is inherited. */ flags = PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) busy_lock = VPB_CURTHREAD_EXCLUSIVE; else if ((req & VM_ALLOC_SBUSY) != 0) busy_lock = VPB_SHARERS_WORD(1); else busy_lock = VPB_UNBUSIED; if ((req & VM_ALLOC_WIRED) != 0) vm_wire_add(npages); if (object->memattr != VM_MEMATTR_DEFAULT && memattr == VM_MEMATTR_DEFAULT) memattr = object->memattr; for (m = m_ret; m < &m_ret[npages]; m++) { m->a.flags = 0; m->flags = (m->flags | PG_NODUMP) & flags; m->busy_lock = busy_lock; if ((req & VM_ALLOC_WIRED) != 0) m->ref_count = 1; m->a.act_count = 0; m->oflags = oflags; if (vm_page_insert_after(m, object, pindex, mpred)) { if ((req & VM_ALLOC_WIRED) != 0) vm_wire_sub(npages); KASSERT(m->object == NULL, ("page %p has object", m)); mpred = m; for (m = m_ret; m < &m_ret[npages]; m++) { if (m <= mpred && (req & VM_ALLOC_WIRED) != 0) m->ref_count = 0; m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); } if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } mpred = m; if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); pindex++; } return (m_ret); } /* * Allocate a physical page that is not intended to be inserted into a VM * object. If the "freelist" parameter is not equal to VM_NFREELIST, then only * pages from the specified vm_phys freelist will be returned. */ static __always_inline vm_page_t _vm_page_alloc_noobj_domain(int domain, const int freelist, int req) { struct vm_domain *vmd; vm_page_t m; int flags; #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \ VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \ VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) KASSERT((req & ~VPAN_FLAGS) == 0, ("invalid request %#x", req)); flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0; vmd = VM_DOMAIN(domain); again: if (freelist == VM_NFREELIST && vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, M_NOWAIT | M_NOVM); if (m != NULL) { flags |= PG_PCPU_CACHE; goto found; } } if (vm_domain_allocate(vmd, req, 1)) { vm_domain_free_lock(vmd); if (freelist == VM_NFREELIST) m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); else m = vm_phys_alloc_freelist_pages(domain, freelist, VM_FREEPOOL_DIRECT, 0); vm_domain_free_unlock(vmd); if (m == NULL) { vm_domain_freecnt_inc(vmd, 1); #if VM_NRESERVLEVEL > 0 if (freelist == VM_NFREELIST && vm_reserv_reclaim_inactive(domain)) goto again; #endif } } if (m == NULL) { if (vm_domain_alloc_fail(vmd, NULL, req)) goto again; return (NULL); } found: vm_page_dequeue(m); vm_page_alloc_check(m); /* * Consumers should not rely on a useful default pindex value. */ m->pindex = 0xdeadc0dedeadc0de; m->flags = (m->flags & PG_ZERO) | flags; m->a.flags = 0; m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; if ((req & VM_ALLOC_WIRED) != 0) { vm_wire_add(1); m->ref_count = 1; } if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); return (m); } vm_page_t vm_page_alloc_freelist(int freelist, int req) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_freelist_domain(domain, freelist, req); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_freelist_domain(int domain, int freelist, int req) { KASSERT(freelist >= 0 && freelist < VM_NFREELIST, ("%s: invalid freelist %d", __func__, freelist)); return (_vm_page_alloc_noobj_domain(domain, freelist, req)); } vm_page_t vm_page_alloc_noobj(int req) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_noobj_domain(domain, req); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_noobj_domain(int domain, int req) { return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req)); } vm_page_t vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, high, alignment, boundary, memattr); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { vm_page_t m, m_ret; u_int flags; #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM) KASSERT((req & ~VPANC_FLAGS) == 0, ("invalid request %#x", req)); KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), ("invalid request %#x", req)); KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("invalid request %#x", req)); KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); while ((m_ret = vm_page_find_contig_domain(domain, req, npages, low, high, alignment, boundary)) == NULL) { if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) return (NULL); } /* * Initialize the pages. Only the PG_ZERO flag is inherited. */ flags = PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; if ((req & VM_ALLOC_WIRED) != 0) vm_wire_add(npages); for (m = m_ret; m < &m_ret[npages]; m++) { vm_page_dequeue(m); vm_page_alloc_check(m); /* * Consumers should not rely on a useful default pindex value. */ m->pindex = 0xdeadc0dedeadc0de; m->a.flags = 0; m->flags = (m->flags | PG_NODUMP) & flags; m->busy_lock = VPB_UNBUSIED; if ((req & VM_ALLOC_WIRED) != 0) m->ref_count = 1; m->a.act_count = 0; m->oflags = VPO_UNMANAGED; /* * Zero the page before updating any mappings since the page is * not yet shared with any devices which might require the * non-default memory attribute. pmap_page_set_memattr() * flushes data caches before returning. */ if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); } return (m_ret); } /* * Check a page that has been freshly dequeued from a freelist. */ static void vm_page_alloc_check(vm_page_t m) { KASSERT(m->object == NULL, ("page %p has object", m)); KASSERT(m->a.queue == PQ_NONE && (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has unexpected queue %d, flags %#x", m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); KASSERT(m->ref_count == 0, ("page %p has references", m)); KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); KASSERT(m->dirty == 0, ("page %p is dirty", m)); KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has unexpected memattr %d", m, pmap_page_get_memattr(m))); KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); pmap_vm_page_alloc_check(m); } static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); /* * The page daemon should avoid creating extra memory pressure since its * main purpose is to replenish the store of free pages. */ if (vmd->vmd_severeset || curproc == pageproc || !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) return (0); domain = vmd->vmd_domain; vm_domain_free_lock(vmd); i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, (vm_page_t *)store); vm_domain_free_unlock(vmd); if (cnt != i) vm_domain_freecnt_inc(vmd, cnt - i); return (i); } static void vm_page_zone_release(void *arg, void **store, int cnt) { struct vm_domain *vmd; struct vm_pgcache *pgcache; vm_page_t m; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); vm_domain_free_lock(vmd); for (i = 0; i < cnt; i++) { m = (vm_page_t)store[i]; vm_phys_free_pages(m, 0); } vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } #define VPSC_ANY 0 /* No restrictions. */ #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ #define VPSC_NOSUPER 2 /* Skip superpages. */ /* * vm_page_scan_contig: * * Scan vm_page_array[] between the specified entries "m_start" and * "m_end" for a run of contiguous physical pages that satisfy the * specified conditions, and return the lowest page in the run. The * specified "alignment" determines the alignment of the lowest physical * page in the run. If the specified "boundary" is non-zero, then the * run of physical pages cannot span a physical address that is a * multiple of "boundary". * * "m_end" is never dereferenced, so it need not point to a vm_page * structure within vm_page_array[]. * * "npages" must be greater than zero. "m_start" and "m_end" must not * span a hole (or discontiguity) in the physical address space. Both * "alignment" and "boundary" must be a power of two. */ static vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options) { vm_object_t object; vm_paddr_t pa; vm_page_t m, m_run; #if VM_NRESERVLEVEL > 0 int level; #endif int m_inc, order, run_ext, run_len; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); m_run = NULL; run_len = 0; for (m = m_start; m < m_end && run_len < npages; m += m_inc) { KASSERT((m->flags & PG_MARKER) == 0, ("page %p is PG_MARKER", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, ("fictitious page %p has invalid ref count", m)); /* * If the current page would be the start of a run, check its * physical address against the end, alignment, and boundary * conditions. If it doesn't satisfy these conditions, either * terminate the scan or advance to the next page that * satisfies the failed condition. */ if (run_len == 0) { KASSERT(m_run == NULL, ("m_run != NULL")); if (m + npages > m_end) break; pa = VM_PAGE_TO_PHYS(m); if (!vm_addr_align_ok(pa, alignment)) { m_inc = atop(roundup2(pa, alignment) - pa); continue; } if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { m_inc = atop(roundup2(pa, boundary) - pa); continue; } } else KASSERT(m_run != NULL, ("m_run == NULL")); retry: m_inc = 1; if (vm_page_wired(m)) run_ext = 0; #if VM_NRESERVLEVEL > 0 else if ((level = vm_reserv_level(m)) >= 0 && (options & VPSC_NORESERV) != 0) { run_ext = 0; /* Advance to the end of the reservation. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); } #endif else if ((object = atomic_load_ptr(&m->object)) != NULL) { /* * The page is considered eligible for relocation if * and only if it could be laundered or reclaimed by * the page daemon. */ VM_OBJECT_RLOCK(object); if (object != m->object) { VM_OBJECT_RUNLOCK(object); goto retry; } /* Don't care: PG_NODUMP, PG_ZERO. */ if ((object->flags & OBJ_SWAP) == 0 && object->type != OBJT_VNODE) { run_ext = 0; #if VM_NRESERVLEVEL > 0 } else if ((options & VPSC_NOSUPER) != 0 && (level = vm_reserv_level_iffullpop(m)) >= 0) { run_ext = 0; /* Advance to the end of the superpage. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); #endif } else if (object->memattr == VM_MEMATTR_DEFAULT && vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { /* * The page is allocated but eligible for * relocation. Extend the current run by one * page. */ KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: PGA_NOSYNC. */ run_ext = 1; } else run_ext = 0; VM_OBJECT_RUNLOCK(object); #if VM_NRESERVLEVEL > 0 } else if (level >= 0) { /* * The page is reserved but not yet allocated. In * other words, it is still free. Extend the current * run by one page. */ run_ext = 1; #endif } else if ((order = m->order) < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it is the * first page in a power-of-two-sized run of * contiguous free pages. Add these pages to the end * of the current run, and jump ahead. */ run_ext = 1 << order; m_inc = 1 << order; } else { /* * Skip the page for one of the following reasons: (1) * It is enqueued in the physical memory allocator's * free page queues. However, it is not the first * page in a run of contiguous free pages. (This case * rarely occurs because the scan is performed in * ascending order.) (2) It is not reserved, and it is * transitioning from free to allocated. (Conversely, * the transition from allocated to free for managed * pages is blocked by the page busy lock.) (3) It is * allocated but not contained by an object and not * wired, e.g., allocated by Xen's balloon driver. */ run_ext = 0; } /* * Extend or reset the current run of pages. */ if (run_ext > 0) { if (run_len == 0) m_run = m; run_len += run_ext; } else { if (run_len > 0) { m_run = NULL; run_len = 0; } } } if (run_len >= npages) return (m_run); return (NULL); } /* * vm_page_reclaim_run: * * Try to relocate each of the allocated virtual pages within the * specified run of physical pages to a new physical address. Free the * physical pages underlying the relocated virtual pages. A virtual page * is relocatable if and only if it could be laundered or reclaimed by * the page daemon. Whenever possible, a virtual page is relocated to a * physical address above "high". * * Returns 0 if every physical page within the run was already free or * just freed by a successful relocation. Otherwise, returns a non-zero * value indicating why the last attempt to relocate a virtual page was * unsuccessful. * * "req_class" must be an allocation class. */ static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high) { struct vm_domain *vmd; struct spglist free; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_end, m_new; int error, order, req; KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, ("req_class is not an allocation class")); SLIST_INIT(&free); error = 0; m = m_run; m_end = m_run + npages; for (; error == 0 && m < m_end; m++) { KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, ("page %p is PG_FICTITIOUS or PG_MARKER", m)); /* * Racily check for wirings. Races are handled once the object * lock is held and the page is unmapped. */ if (vm_page_wired(m)) error = EBUSY; else if ((object = atomic_load_ptr(&m->object)) != NULL) { /* * The page is relocated if and only if it could be * laundered or reclaimed by the page daemon. */ VM_OBJECT_WLOCK(object); /* Don't care: PG_NODUMP, PG_ZERO. */ if (m->object != object || ((object->flags & OBJ_SWAP) == 0 && object->type != OBJT_VNODE)) error = EINVAL; else if (object->memattr != VM_MEMATTR_DEFAULT) error = EINVAL; else if (vm_page_queue(m) != PQ_NONE && vm_page_tryxbusy(m) != 0) { if (vm_page_wired(m)) { vm_page_xunbusy(m); error = EBUSY; goto unlock; } KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT(m->oflags == 0, ("page %p has unexpected oflags", m)); /* Don't care: PGA_NOSYNC. */ if (!vm_page_none_valid(m)) { /* * First, try to allocate a new page * that is above "high". Failing * that, try to allocate a new page * that is below "m_run". Allocate * the new page between the end of * "m_run" and "high" only as a last * resort. */ req = req_class; if ((m->flags & PG_NODUMP) != 0) req |= VM_ALLOC_NODUMP; if (trunc_page(high) != ~(vm_paddr_t)PAGE_MASK) { m_new = vm_page_alloc_noobj_contig( req, 1, round_page(high), ~(vm_paddr_t)0, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } else m_new = NULL; if (m_new == NULL) { pa = VM_PAGE_TO_PHYS(m_run); m_new = vm_page_alloc_noobj_contig( req, 1, 0, pa - 1, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { pa += ptoa(npages); m_new = vm_page_alloc_noobj_contig( req, 1, pa, high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { vm_page_xunbusy(m); error = ENOMEM; goto unlock; } /* * Unmap the page and check for new * wirings that may have been acquired * through a pmap lookup. */ if (object->ref_count != 0 && !vm_page_try_remove_all(m)) { vm_page_xunbusy(m); vm_page_free(m_new); error = EBUSY; goto unlock; } /* * Replace "m" with the new page. For * vm_page_replace(), "m" must be busy * and dequeued. Finally, change "m" * as if vm_page_free() was called. */ m_new->a.flags = m->a.flags & ~PGA_QUEUE_STATE_MASK; KASSERT(m_new->oflags == VPO_UNMANAGED, ("page %p is managed", m_new)); m_new->oflags = 0; pmap_copy_page(m, m_new); m_new->valid = m->valid; m_new->dirty = m->dirty; m->flags &= ~PG_ZERO; vm_page_dequeue(m); if (vm_page_replace_hold(m_new, object, m->pindex, m) && vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); /* * The new page must be deactivated * before the object is unlocked. */ vm_page_deactivate(m_new); } else { m->flags &= ~PG_ZERO; vm_page_dequeue(m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); KASSERT(m->dirty == 0, ("page %p is dirty", m)); } } else error = EBUSY; unlock: VM_OBJECT_WUNLOCK(object); } else { MPASS(vm_page_domain(m) == domain); vmd = VM_DOMAIN(domain); vm_domain_free_lock(vmd); order = m->order; if (order < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it * is the first page in a power-of-two-sized * run of contiguous free pages. Jump ahead * to the last page within that run, and * continue from there. */ m += (1 << order) - 1; } #if VM_NRESERVLEVEL > 0 else if (vm_reserv_is_page_free(m)) order = 0; #endif vm_domain_free_unlock(vmd); if (order == VM_NFREEORDER) error = EINVAL; } } if ((m = SLIST_FIRST(&free)) != NULL) { int cnt; vmd = VM_DOMAIN(domain); cnt = 0; vm_domain_free_lock(vmd); do { MPASS(vm_page_domain(m) == domain); SLIST_REMOVE_HEAD(&free, plinks.s.ss); vm_phys_free_pages(m, 0); cnt++; } while ((m = SLIST_FIRST(&free)) != NULL); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } return (error); } #define NRUNS 16 #define RUN_INDEX(count, nruns) ((count) % (nruns)) #define MIN_RECLAIM 8 /* * vm_page_reclaim_contig: * * Reclaim allocated, contiguous physical memory satisfying the specified * conditions by relocating the virtual pages using that physical memory. * Returns true if reclamation is successful and false otherwise. Since * relocation requires the allocation of physical pages, reclamation may * fail due to a shortage of free pages. When reclamation fails, callers * are expected to perform vm_wait() before retrying a failed allocation * operation, e.g., vm_page_alloc_contig(). * * The caller must always specify an allocation class through "req". * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * The optional allocation flags are ignored. * * "npages" must be greater than zero. Both "alignment" and "boundary" * must be a power of two. */ -bool +int vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, int desired_runs) { struct vm_domain *vmd; vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; u_long count, minalign, reclaimed; - int error, i, min_reclaim, nruns, options, req_class, segind; - bool ret; + int error, i, min_reclaim, nruns, options, req_class; + int segind, start_segind; + int ret; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); - ret = false; + ret = ENOMEM; /* * If the caller wants to reclaim multiple runs, try to allocate * space to store the runs. If that fails, fall back to the old * behavior of just reclaiming MIN_RECLAIM pages. */ if (desired_runs > 1) m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), M_TEMP, M_NOWAIT); else m_runs = NULL; if (m_runs == NULL) { m_runs = _m_runs; nruns = NRUNS; } else { nruns = NRUNS + desired_runs - 1; } min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); /* * The caller will attempt an allocation after some runs have been * reclaimed and added to the vm_phys buddy lists. Due to limitations * of vm_phys_alloc_contig(), round up the requested length to the next * power of two or maximum chunk size, and ensure that each run is * suitably aligned. */ minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); npages = roundup2(npages, minalign); if (alignment < ptoa(minalign)) alignment = ptoa(minalign); /* * The page daemon is allowed to dig deeper into the free page list. */ req_class = req & VM_ALLOC_CLASS_MASK; if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; + start_segind = vm_phys_lookup_segind(low); + /* * Return if the number of free pages cannot satisfy the requested * allocation. */ vmd = VM_DOMAIN(domain); count = vmd->vmd_free_count; if (count < npages + vmd->vmd_free_reserved || (count < npages + vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || (count < npages && req_class == VM_ALLOC_INTERRUPT)) goto done; /* * Scan up to three times, relaxing the restrictions ("options") on * the reclamation of reservations and superpages each time. */ for (options = VPSC_NORESERV;;) { + bool phys_range_exists = false; + /* * Find the highest runs that satisfy the given constraints * and restrictions, and record them in "m_runs". */ count = 0; - segind = vm_phys_lookup_segind(low); + segind = start_segind; while ((segind = vm_phys_find_range(bounds, segind, domain, npages, low, high)) != -1) { + phys_range_exists = true; while ((m_run = vm_page_scan_contig(npages, bounds[0], bounds[1], alignment, boundary, options))) { bounds[0] = m_run + npages; m_runs[RUN_INDEX(count, nruns)] = m_run; count++; } segind++; } + if (!phys_range_exists) { + ret = ERANGE; + goto done; + } + /* * Reclaim the highest runs in LIFO (descending) order until * the number of reclaimed pages, "reclaimed", is at least * "min_reclaim". Reset "reclaimed" each time because each * reclamation is idempotent, and runs will (likely) recur * from one scan to the next as restrictions are relaxed. */ reclaimed = 0; for (i = 0; count > 0 && i < nruns; i++) { count--; m_run = m_runs[RUN_INDEX(count, nruns)]; error = vm_page_reclaim_run(req_class, domain, npages, m_run, high); if (error == 0) { reclaimed += npages; if (reclaimed >= min_reclaim) { - ret = true; + ret = 0; goto done; } } } /* * Either relax the restrictions on the next scan or return if * the last scan had no restrictions. */ if (options == VPSC_NORESERV) options = VPSC_NOSUPER; else if (options == VPSC_NOSUPER) options = VPSC_ANY; else if (options == VPSC_ANY) { - ret = reclaimed != 0; + if (reclaimed != 0) + ret = 0; goto done; } } done: if (m_runs != _m_runs) free(m_runs, M_TEMP); return (ret); } -bool +int vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high, alignment, boundary, 1)); } -bool +int vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domainset_iter di; - int domain; - bool ret; + int domain, ret, status; + + ret = ERANGE; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { - ret = vm_page_reclaim_contig_domain(domain, req, npages, low, + status = vm_page_reclaim_contig_domain(domain, req, npages, low, high, alignment, boundary); - if (ret) - break; + if (status == 0) + return (0); + else if (status == ERANGE) + vm_domainset_iter_ignore(&di, domain); + else { + KASSERT(status == ENOMEM, ("Unrecognized error %d " + "from vm_page_reclaim_contig_domain()", status)); + ret = ENOMEM; + } } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (ret); } /* * Set the domain in the appropriate page level domainset. */ void vm_domain_set(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (!vmd->vmd_minset && vm_paging_min(vmd)) { vmd->vmd_minset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); } if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { vmd->vmd_severeset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); } mtx_unlock(&vm_domainset_lock); } /* * Clear the domain from the appropriate page level domainset. */ void vm_domain_clear(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_minset && !vm_paging_min(vmd)) { vmd->vmd_minset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); if (vm_min_waiters != 0) { vm_min_waiters = 0; wakeup(&vm_min_domains); } } if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { vmd->vmd_severeset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); if (vm_severe_waiters != 0) { vm_severe_waiters = 0; wakeup(&vm_severe_domains); } } /* * If pageout daemon needs pages, then tell it that there are * some free. */ if (vmd->vmd_pageout_pages_needed && vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { wakeup(&vmd->vmd_pageout_pages_needed); vmd->vmd_pageout_pages_needed = 0; } /* See comments in vm_wait_doms(). */ if (vm_pageproc_waiters) { vm_pageproc_waiters = 0; wakeup(&vm_pageproc_waiters); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the min threshold globally. */ void vm_wait_min(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_min()) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the severe threshold globally. */ void vm_wait_severe(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_severe()) { vm_severe_waiters++; msleep(&vm_severe_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } u_int vm_wait_count(void) { return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); } int vm_wait_doms(const domainset_t *wdoms, int mflags) { int error; error = 0; /* * We use racey wakeup synchronization to avoid expensive global * locking for the pageproc when sleeping with a non-specific vm_wait. * To handle this, we only sleep for one tick in this instance. It * is expected that most allocations for the pageproc will come from * kmem or vm_page_grab* which will use the more specific and * race-free vm_wait_domain(). */ if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); vm_pageproc_waiters++; error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP | mflags, "pageprocwait", 1); } else { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(wdoms)) { if (pageproc == NULL) panic("vm_wait in early boot"); vm_min_waiters++; error = msleep(&vm_min_domains, &vm_domainset_lock, PVM | PDROP | mflags, "vmwait", 0); } else mtx_unlock(&vm_domainset_lock); } return (error); } /* * vm_wait_domain: * * Sleep until free pages are available for allocation. * - Called in various places after failed memory allocations. */ void vm_wait_domain(int domain) { struct vm_domain *vmd; domainset_t wdom; vmd = VM_DOMAIN(domain); vm_domain_free_assert_unlocked(vmd); if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { vmd->vmd_pageout_pages_needed = 1; msleep(&vmd->vmd_pageout_pages_needed, &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); } else mtx_unlock(&vm_domainset_lock); } else { DOMAINSET_ZERO(&wdom); DOMAINSET_SET(vmd->vmd_domain, &wdom); vm_wait_doms(&wdom, 0); } } static int vm_wait_flags(vm_object_t obj, int mflags) { struct domainset *d; d = NULL; /* * Carefully fetch pointers only once: the struct domainset * itself is ummutable but the pointer might change. */ if (obj != NULL) d = obj->domain.dr_policy; if (d == NULL) d = curthread->td_domain.dr_policy; return (vm_wait_doms(&d->ds_mask, mflags)); } /* * vm_wait: * * Sleep until free pages are available for allocation in the * affinity domains of the obj. If obj is NULL, the domain set * for the calling thread is used. * Called in various places after failed memory allocations. */ void vm_wait(vm_object_t obj) { (void)vm_wait_flags(obj, 0); } int vm_wait_intr(vm_object_t obj) { return (vm_wait_flags(obj, PCATCH)); } /* * vm_domain_alloc_fail: * * Called when a page allocation function fails. Informs the * pagedaemon and performs the requested wait. Requires the * domain_free and object lock on entry. Returns with the * object lock held and free lock released. Returns an error when * retry is necessary. * */ static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) { vm_domain_free_assert_unlocked(vmd); atomic_add_int(&vmd->vmd_pageout_deficit, max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_wait_domain(vmd->vmd_domain); if (object != NULL) VM_OBJECT_WLOCK(object); if (req & VM_ALLOC_WAITOK) return (EAGAIN); } return (0); } /* * vm_waitpfault: * * Sleep until free pages are available for allocation. * - Called only in vm_fault so that processes page faulting * can be easily tracked. * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing * processes will be able to grab memory first. Do not change * this balance without careful testing first. */ void vm_waitpfault(struct domainset *dset, int timo) { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(&dset->ds_mask)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, "pfault", timo); } else mtx_unlock(&vm_domainset_lock); } static struct vm_pagequeue * _vm_page_pagequeue(vm_page_t m, uint8_t queue) { return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); } #ifdef INVARIANTS static struct vm_pagequeue * vm_page_pagequeue(vm_page_t m) { return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); } #endif static __always_inline bool vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { vm_page_astate_t tmp; tmp = *old; do { if (__predict_true(vm_page_astate_fcmpset(m, old, new))) return (true); counter_u64_add(pqstate_commit_retries, 1); } while (old->_bits == tmp._bits); return (false); } /* * Do the work of committing a queue state update that moves the page out of * its current queue. */ static bool _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { vm_page_t next; vm_pagequeue_assert_locked(pq); KASSERT(vm_page_pagequeue(m) == pq, ("%s: queue %p does not match page %p", __func__, pq, m)); KASSERT(old->queue != PQ_NONE && new.queue != old->queue, ("%s: invalid queue indices %d %d", __func__, old->queue, new.queue)); /* * Once the queue index of the page changes there is nothing * synchronizing with further updates to the page's physical * queue state. Therefore we must speculatively remove the page * from the queue now and be prepared to roll back if the queue * state update fails. If the page is not physically enqueued then * we just update its queue index. */ if ((old->flags & PGA_ENQUEUED) != 0) { new.flags &= ~PGA_ENQUEUED; next = TAILQ_NEXT(m, plinks.q); TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_pagequeue_cnt_dec(pq); if (!vm_page_pqstate_fcmpset(m, old, new)) { if (next == NULL) TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); else TAILQ_INSERT_BEFORE(next, m, plinks.q); vm_pagequeue_cnt_inc(pq); return (false); } else { return (true); } } else { return (vm_page_pqstate_fcmpset(m, old, new)); } } static bool vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { struct vm_pagequeue *pq; vm_page_astate_t as; bool ret; pq = _vm_page_pagequeue(m, old->queue); /* * The queue field and PGA_ENQUEUED flag are stable only so long as the * corresponding page queue lock is held. */ vm_pagequeue_lock(pq); as = vm_page_astate_load(m); if (__predict_false(as._bits != old->_bits)) { *old = as; ret = false; } else { ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); } vm_pagequeue_unlock(pq); return (ret); } /* * Commit a queue state update that enqueues or requeues a page. */ static bool _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { struct vm_domain *vmd; vm_pagequeue_assert_locked(pq); KASSERT(old->queue != PQ_NONE && new.queue == old->queue, ("%s: invalid queue indices %d %d", __func__, old->queue, new.queue)); new.flags |= PGA_ENQUEUED; if (!vm_page_pqstate_fcmpset(m, old, new)) return (false); if ((old->flags & PGA_ENQUEUED) != 0) TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); else vm_pagequeue_cnt_inc(pq); /* * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if * both flags are set in close succession, only PGA_REQUEUE_HEAD will be * applied, even if it was set first. */ if ((old->flags & PGA_REQUEUE_HEAD) != 0) { vmd = vm_pagequeue_domain(m); KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], ("%s: invalid page queue for page %p", __func__, m)); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); } else { TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); } return (true); } /* * Commit a queue state update that encodes a request for a deferred queue * operation. */ static bool vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { KASSERT(old->queue == new.queue || new.queue != PQ_NONE, ("%s: invalid state, queue %d flags %x", __func__, new.queue, new.flags)); if (old->_bits != new._bits && !vm_page_pqstate_fcmpset(m, old, new)) return (false); vm_page_pqbatch_submit(m, new.queue); return (true); } /* * A generic queue state update function. This handles more cases than the * specialized functions above. */ bool vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { if (old->_bits == new._bits) return (true); if (old->queue != PQ_NONE && new.queue != old->queue) { if (!vm_page_pqstate_commit_dequeue(m, old, new)) return (false); if (new.queue != PQ_NONE) vm_page_pqbatch_submit(m, new.queue); } else { if (!vm_page_pqstate_fcmpset(m, old, new)) return (false); if (new.queue != PQ_NONE && ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) vm_page_pqbatch_submit(m, new.queue); } return (true); } /* * Apply deferred queue state updates to a page. */ static inline void vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) { vm_page_astate_t new, old; CRITICAL_ASSERT(curthread); vm_pagequeue_assert_locked(pq); KASSERT(queue < PQ_COUNT, ("%s: invalid queue index %d", __func__, queue)); KASSERT(pq == _vm_page_pagequeue(m, queue), ("%s: page %p does not belong to queue %p", __func__, m, pq)); for (old = vm_page_astate_load(m);;) { if (__predict_false(old.queue != queue || (old.flags & PGA_QUEUE_OP_MASK) == 0)) { counter_u64_add(queue_nops, 1); break; } KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is unmanaged", __func__, m)); new = old; if ((old.flags & PGA_DEQUEUE) != 0) { new.flags &= ~PGA_QUEUE_OP_MASK; new.queue = PQ_NONE; if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, m, &old, new))) { counter_u64_add(queue_ops, 1); break; } } else { new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); if (__predict_true(_vm_page_pqstate_commit_requeue(pq, m, &old, new))) { counter_u64_add(queue_ops, 1); break; } } } } static void vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, uint8_t queue) { int i; for (i = 0; i < bq->bq_cnt; i++) vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); vm_batchqueue_init(bq); } /* * vm_page_pqbatch_submit: [ internal use only ] * * Enqueue a page in the specified page queue's batched work queue. * The caller must have encoded the requested operation in the page * structure's a.flags field. */ void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) { struct vm_batchqueue *bq; struct vm_pagequeue *pq; int domain, slots_remaining; KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); domain = vm_page_domain(m); critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); slots_remaining = vm_batchqueue_insert(bq, m); if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { /* keep building the bq */ critical_exit(); return; } else if (slots_remaining > 0 ) { /* Try to process the bq if we can get the lock */ pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; if (vm_pagequeue_trylock(pq)) { vm_pqbatch_process(pq, bq, queue); vm_pagequeue_unlock(pq); } critical_exit(); return; } critical_exit(); /* if we make it here, the bq is full so wait for the lock */ pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); vm_pqbatch_process(pq, bq, queue); vm_pqbatch_process_page(pq, m, queue); vm_pagequeue_unlock(pq); critical_exit(); } /* * vm_page_pqbatch_drain: [ internal use only ] * * Force all per-CPU page queue batch queues to be drained. This is * intended for use in severe memory shortages, to ensure that pages * do not remain stuck in the batch queues. */ void vm_page_pqbatch_drain(void) { struct thread *td; struct vm_domain *vmd; struct vm_pagequeue *pq; int cpu, domain, queue; td = curthread; CPU_FOREACH(cpu) { thread_lock(td); sched_bind(td, cpu); thread_unlock(td); for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (queue = 0; queue < PQ_COUNT; queue++) { pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); critical_enter(); vm_pqbatch_process(pq, DPCPU_PTR(pqbatch[domain][queue]), queue); critical_exit(); vm_pagequeue_unlock(pq); } } } thread_lock(td); sched_unbind(td); thread_unlock(td); } /* * vm_page_dequeue_deferred: [ internal use only ] * * Request removal of the given page from its current page * queue. Physical removal from the queue may be deferred * indefinitely. */ void vm_page_dequeue_deferred(vm_page_t m) { vm_page_astate_t new, old; old = vm_page_astate_load(m); do { if (old.queue == PQ_NONE) { KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p has unexpected queue state", __func__, m)); break; } new = old; new.flags |= PGA_DEQUEUE; } while (!vm_page_pqstate_commit_request(m, &old, new)); } /* * vm_page_dequeue: * * Remove the page from whichever page queue it's in, if any, before * returning. */ void vm_page_dequeue(vm_page_t m) { vm_page_astate_t new, old; old = vm_page_astate_load(m); do { if (old.queue == PQ_NONE) { KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p has unexpected queue state", __func__, m)); break; } new = old; new.flags &= ~PGA_QUEUE_OP_MASK; new.queue = PQ_NONE; } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); } /* * Schedule the given page for insertion into the specified page queue. * Physical insertion of the page may be deferred indefinitely. */ static void vm_page_enqueue(vm_page_t m, uint8_t queue) { KASSERT(m->a.queue == PQ_NONE && (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p is already enqueued", __func__, m)); KASSERT(m->ref_count > 0, ("%s: page %p does not carry any references", __func__, m)); m->a.queue = queue; if ((m->a.flags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_page_pqbatch_submit(m, queue); } /* * vm_page_free_prep: * * Prepares the given page to be put on the free list, * disassociating it from any VM object. The caller may return * the page to the free list only if this function returns true. * * The object, if it exists, must be locked, and then the page must * be xbusy. Otherwise the page must be not busied. A managed * page must be unmapped. */ static bool vm_page_free_prep(vm_page_t m) { /* * Synchronize with threads that have dropped a reference to this * page. */ atomic_thread_fence_acq(); #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { uint64_t *p; int i; p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", m, i, (uintmax_t)*p)); } #endif if ((m->oflags & VPO_UNMANAGED) == 0) { KASSERT(!pmap_page_is_mapped(m), ("vm_page_free_prep: freeing mapped page %p", m)); KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, ("vm_page_free_prep: mapping flags set in page %p", m)); } else { KASSERT(m->a.queue == PQ_NONE, ("vm_page_free_prep: unmanaged page %p is queued", m)); } VM_CNT_INC(v_tfree); if (m->object != NULL) { KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == ((m->object->flags & OBJ_UNMANAGED) != 0), ("vm_page_free_prep: managed flag mismatch for page %p", m)); vm_page_assert_xbusied(m); /* * The object reference can be released without an atomic * operation. */ KASSERT((m->flags & PG_FICTITIOUS) != 0 || m->ref_count == VPRC_OBJREF, ("vm_page_free_prep: page %p has unexpected ref_count %u", m, m->ref_count)); vm_page_object_remove(m); m->ref_count -= VPRC_OBJREF; } else vm_page_assert_unbusied(m); vm_page_busy_free(m); /* * If fictitious remove object association and * return. */ if ((m->flags & PG_FICTITIOUS) != 0) { KASSERT(m->ref_count == 1, ("fictitious page %p is referenced", m)); KASSERT(m->a.queue == PQ_NONE, ("fictitious page %p is queued", m)); return (false); } /* * Pages need not be dequeued before they are returned to the physical * memory allocator, but they must at least be marked for a deferred * dequeue. */ if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_dequeue_deferred(m); m->valid = 0; vm_page_undirty(m); if (m->ref_count != 0) panic("vm_page_free_prep: page %p has references", m); /* * Restore the default memory attribute to the page. */ if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); #if VM_NRESERVLEVEL > 0 /* * Determine whether the page belongs to a reservation. If the page was * allocated from a per-CPU cache, it cannot belong to a reservation, so * as an optimization, we avoid the check in that case. */ if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) return (false); #endif return (true); } /* * vm_page_free_toq: * * Returns the given page to the free list, disassociating it * from any VM object. * * The object must be locked. The page must be exclusively busied if it * belongs to an object. */ static void vm_page_free_toq(vm_page_t m) { struct vm_domain *vmd; uma_zone_t zone; if (!vm_page_free_prep(m)) return; vmd = vm_pagequeue_domain(m); zone = vmd->vmd_pgcache[m->pool].zone; if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { uma_zfree(zone, m); return; } vm_domain_free_lock(vmd); vm_phys_free_pages(m, 0); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, 1); } /* * vm_page_free_pages_toq: * * Returns a list of pages to the free list, disassociating it * from any VM object. In other words, this is equivalent to * calling vm_page_free_toq() for each page of a list of VM objects. */ int vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) { vm_page_t m; int count; if (SLIST_EMPTY(free)) return (0); count = 0; while ((m = SLIST_FIRST(free)) != NULL) { count++; SLIST_REMOVE_HEAD(free, plinks.s.ss); vm_page_free_toq(m); } if (update_wire_count) vm_wire_sub(count); return (count); } /* * Mark this page as wired down. For managed pages, this prevents reclamation * by the page daemon, or when the containing object, if any, is destroyed. */ void vm_page_wire(vm_page_t m) { u_int old; #ifdef INVARIANTS if (m->object != NULL && !vm_page_busied(m) && !vm_object_busied(m->object)) VM_OBJECT_ASSERT_LOCKED(m->object); #endif KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(m->ref_count) >= 1, ("vm_page_wire: fictitious page %p has zero wirings", m)); old = atomic_fetchadd_int(&m->ref_count, 1); KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, ("vm_page_wire: counter overflow for page %p", m)); if (VPRC_WIRE_COUNT(old) == 0) { if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); vm_wire_add(1); } } /* * Attempt to wire a mapped page following a pmap lookup of that page. * This may fail if a thread is concurrently tearing down mappings of the page. * The transient failure is acceptable because it translates to the * failure of the caller pmap_extract_and_hold(), which should be then * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). */ bool vm_page_wire_mapped(vm_page_t m) { u_int old; old = atomic_load_int(&m->ref_count); do { KASSERT(old > 0, ("vm_page_wire_mapped: wiring unreferenced page %p", m)); if ((old & VPRC_BLOCKED) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); if (VPRC_WIRE_COUNT(old) == 0) { if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); vm_wire_add(1); } return (true); } /* * Release a wiring reference to a managed page. If the page still belongs to * an object, update its position in the page queues to reflect the reference. * If the wiring was the last reference to the page, free the page. */ static void vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) { u_int old; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is unmanaged", __func__, m)); /* * Update LRU state before releasing the wiring reference. * Use a release store when updating the reference count to * synchronize with vm_page_free_prep(). */ old = atomic_load_int(&m->ref_count); do { u_int count; KASSERT(VPRC_WIRE_COUNT(old) > 0, ("vm_page_unwire: wire count underflow for page %p", m)); count = old & ~VPRC_BLOCKED; if (count > VPRC_OBJREF + 1) { /* * The page has at least one other wiring reference. An * earlier iteration of this loop may have called * vm_page_release_toq() and cleared PGA_DEQUEUE, so * re-set it if necessary. */ if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); } else if (count == VPRC_OBJREF + 1) { /* * This is the last wiring. Clear PGA_DEQUEUE and * update the page's queue state to reflect the * reference. If the page does not belong to an object * (i.e., the VPRC_OBJREF bit is clear), we only need to * clear leftover queue state. */ vm_page_release_toq(m, nqueue, noreuse); } else if (count == 1) { vm_page_aflag_clear(m, PGA_DEQUEUE); } } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); if (VPRC_WIRE_COUNT(old) == 1) { vm_wire_sub(1); if (old == 1) vm_page_free(m); } } /* * Release one wiring of the specified page, potentially allowing it to be * paged out. * * Only managed pages belonging to an object can be paged out. If the number * of wirings transitions to zero and the page is eligible for page out, then * the page is added to the specified paging queue. If the released wiring * represented the last reference to the page, the page is freed. */ void vm_page_unwire(vm_page_t m, uint8_t nqueue) { KASSERT(nqueue < PQ_COUNT, ("vm_page_unwire: invalid queue %u request for page %p", nqueue, m)); if ((m->oflags & VPO_UNMANAGED) != 0) { if (vm_page_unwire_noq(m) && m->ref_count == 0) vm_page_free(m); return; } vm_page_unwire_managed(m, nqueue, false); } /* * Unwire a page without (re-)inserting it into a page queue. It is up * to the caller to enqueue, requeue, or free the page as appropriate. * In most cases involving managed pages, vm_page_unwire() should be used * instead. */ bool vm_page_unwire_noq(vm_page_t m) { u_int old; old = vm_page_drop(m, 1); KASSERT(VPRC_WIRE_COUNT(old) != 0, ("%s: counter underflow for page %p", __func__, m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, ("%s: missing ref on fictitious page %p", __func__, m)); if (VPRC_WIRE_COUNT(old) > 1) return (false); if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_clear(m, PGA_DEQUEUE); vm_wire_sub(1); return (true); } /* * Ensure that the page ends up in the specified page queue. If the page is * active or being moved to the active queue, ensure that its act_count is * at least ACT_INIT but do not otherwise mess with it. */ static __always_inline void vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) { vm_page_astate_t old, new; KASSERT(m->ref_count > 0, ("%s: page %p does not carry any references", __func__, m)); KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, ("%s: invalid flags %x", __func__, nflag)); if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) return; old = vm_page_astate_load(m); do { if ((old.flags & PGA_DEQUEUE) != 0) break; new = old; new.flags &= ~PGA_QUEUE_OP_MASK; if (nqueue == PQ_ACTIVE) new.act_count = max(old.act_count, ACT_INIT); if (old.queue == nqueue) { /* * There is no need to requeue pages already in the * active queue. */ if (nqueue != PQ_ACTIVE || (old.flags & PGA_ENQUEUED) == 0) new.flags |= nflag; } else { new.flags |= nflag; new.queue = nqueue; } } while (!vm_page_pqstate_commit(m, &old, new)); } /* * Put the specified page on the active list (if appropriate). */ void vm_page_activate(vm_page_t m) { vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); } /* * Move the specified page to the tail of the inactive queue, or requeue * the page if it is already in the inactive queue. */ void vm_page_deactivate(vm_page_t m) { vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); } void vm_page_deactivate_noreuse(vm_page_t m) { vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); } /* * Put a page in the laundry, or requeue it if it is already there. */ void vm_page_launder(vm_page_t m) { vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); } /* * Put a page in the PQ_UNSWAPPABLE holding queue. */ void vm_page_unswappable(vm_page_t m) { VM_OBJECT_ASSERT_LOCKED(m->object); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p already unswappable", m)); vm_page_dequeue(m); vm_page_enqueue(m, PQ_UNSWAPPABLE); } /* * Release a page back to the page queues in preparation for unwiring. */ static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) { vm_page_astate_t old, new; uint16_t nflag; /* * Use a check of the valid bits to determine whether we should * accelerate reclamation of the page. The object lock might not be * held here, in which case the check is racy. At worst we will either * accelerate reclamation of a valid page and violate LRU, or * unnecessarily defer reclamation of an invalid page. * * If we were asked to not cache the page, place it near the head of the * inactive queue so that is reclaimed sooner. */ if (noreuse || vm_page_none_valid(m)) { nqueue = PQ_INACTIVE; nflag = PGA_REQUEUE_HEAD; } else { nflag = PGA_REQUEUE; } old = vm_page_astate_load(m); do { new = old; /* * If the page is already in the active queue and we are not * trying to accelerate reclamation, simply mark it as * referenced and avoid any queue operations. */ new.flags &= ~PGA_QUEUE_OP_MASK; if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && (old.flags & PGA_ENQUEUED) != 0) new.flags |= PGA_REFERENCED; else { new.flags |= nflag; new.queue = nqueue; } } while (!vm_page_pqstate_commit(m, &old, new)); } /* * Unwire a page and either attempt to free it or re-add it to the page queues. */ void vm_page_release(vm_page_t m, int flags) { vm_object_t object; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release: page %p is unmanaged", m)); if ((flags & VPR_TRYFREE) != 0) { for (;;) { object = atomic_load_ptr(&m->object); if (object == NULL) break; /* Depends on type-stability. */ if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) break; if (object == m->object) { vm_page_release_locked(m, flags); VM_OBJECT_WUNLOCK(object); return; } VM_OBJECT_WUNLOCK(object); } } vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); } /* See vm_page_release(). */ void vm_page_release_locked(vm_page_t m, int flags) { VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release_locked: page %p is unmanaged", m)); if (vm_page_unwire_noq(m)) { if ((flags & VPR_TRYFREE) != 0 && (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && m->dirty == 0 && vm_page_tryxbusy(m)) { /* * An unlocked lookup may have wired the page before the * busy lock was acquired, in which case the page must * not be freed. */ if (__predict_true(!vm_page_wired(m))) { vm_page_free(m); return; } vm_page_xunbusy(m); } else { vm_page_release_toq(m, PQ_INACTIVE, flags != 0); } } } static bool vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) { u_int old; KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, ("vm_page_try_blocked_op: page %p has no object", m)); KASSERT(vm_page_busied(m), ("vm_page_try_blocked_op: page %p is not busy", m)); VM_OBJECT_ASSERT_LOCKED(m->object); old = atomic_load_int(&m->ref_count); do { KASSERT(old != 0, ("vm_page_try_blocked_op: page %p has no references", m)); KASSERT((old & VPRC_BLOCKED) == 0, ("vm_page_try_blocked_op: page %p blocks wirings", m)); if (VPRC_WIRE_COUNT(old) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); (op)(m); /* * If the object is read-locked, new wirings may be created via an * object lookup. */ old = vm_page_drop(m, VPRC_BLOCKED); KASSERT(!VM_OBJECT_WOWNED(m->object) || old == (VPRC_BLOCKED | VPRC_OBJREF), ("vm_page_try_blocked_op: unexpected refcount value %u for %p", old, m)); return (true); } /* * Atomically check for wirings and remove all mappings of the page. */ bool vm_page_try_remove_all(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_all)); } /* * Atomically check for wirings and remove all writeable mappings of the page. */ bool vm_page_try_remove_write(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_write)); } /* * vm_page_advise * * Apply the specified advice to the given page. */ void vm_page_advise(vm_page_t m, int advice) { VM_OBJECT_ASSERT_WLOCKED(m->object); vm_page_assert_xbusied(m); if (advice == MADV_FREE) /* * Mark the page clean. This will allow the page to be freed * without first paging it out. MADV_FREE pages are often * quickly reused by malloc(3), so we do not do anything that * would result in a page fault on a later access. */ vm_page_undirty(m); else if (advice != MADV_DONTNEED) { if (advice == MADV_WILLNEED) vm_page_activate(m); return; } if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) vm_page_dirty(m); /* * Clear any references to the page. Otherwise, the page daemon will * immediately reactivate the page. */ vm_page_aflag_clear(m, PGA_REFERENCED); /* * Place clean pages near the head of the inactive queue rather than * the tail, thus defeating the queue's LRU operation and ensuring that * the page will be reused quickly. Dirty pages not already in the * laundry are moved there. */ if (m->dirty == 0) vm_page_deactivate_noreuse(m); else if (!vm_page_in_laundry(m)) vm_page_launder(m); } /* * vm_page_grab_release * * Helper routine for grab functions to release busy on return. */ static inline void vm_page_grab_release(vm_page_t m, int allocflags) { if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) vm_page_sunbusy(m); else vm_page_xunbusy(m); } } /* * vm_page_grab_sleep * * Sleep for busy according to VM_ALLOC_ parameters. Returns true * if the caller should retry and false otherwise. * * If the object is locked on entry the object will be unlocked with * false returns and still locked but possibly having been dropped * with true returns. */ static bool vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (false); /* * Reference the page before unlocking and sleeping so that * the page daemon is less likely to reclaim it. */ if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_reference(m); if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && locked) VM_OBJECT_WLOCK(object); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (false); return (true); } /* * Assert that the grab flags are valid. */ static inline void vm_page_grab_check(int allocflags) { KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || (allocflags & VM_ALLOC_WIRED) != 0, ("vm_page_grab*: the pages must be busied or wired")); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); } /* * Calculate the page allocation flags for grab. */ static inline int vm_page_grab_pflags(int allocflags) { int pflags; pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) pflags |= VM_ALLOC_SBUSY; return (pflags); } /* * Grab a page, waiting until we are waken up due to the page * changing state. We keep on waiting, if the page continues * to be in the object. If the page doesn't exist, first allocate it * and then conditionally zero it. * * This routine may sleep. * * The object must be locked on entry. The lock will, however, be released * and reacquired if the routine sleeps. */ vm_page_t vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; VM_OBJECT_ASSERT_WLOCKED(object); vm_page_grab_check(allocflags); retrylookup: if ((m = vm_page_lookup(object, pindex)) != NULL) { if (!vm_page_tryacquire(m, allocflags)) { if (vm_page_grab_sleep(object, m, pindex, "pgrbwt", allocflags, true)) goto retrylookup; return (NULL); } goto out; } if ((allocflags & VM_ALLOC_NOCREAT) != 0) return (NULL); m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags)); if (m == NULL) { if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) return (NULL); goto retrylookup; } if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); out: vm_page_grab_release(m, allocflags); return (m); } /* * Locklessly attempt to acquire a page given a (object, pindex) tuple * and an optional previous page to avoid the radix lookup. The resulting * page will be validated against the identity tuple and busied or wired * as requested. A NULL *mp return guarantees that the page was not in * radix at the time of the call but callers must perform higher level * synchronization or retry the operation under a lock if they require * an atomic answer. This is the only lock free validation routine, * other routines can depend on the resulting page state. * * The return value indicates whether the operation failed due to caller * flags. The return is tri-state with mp: * * (true, *mp != NULL) - The operation was successful. * (true, *mp == NULL) - The page was not found in tree. * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition. */ static bool vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t prev, vm_page_t *mp, int allocflags) { vm_page_t m; vm_page_grab_check(allocflags); MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev)); *mp = NULL; for (;;) { /* * We may see a false NULL here because the previous page * has been removed or just inserted and the list is loaded * without barriers. Switch to radix to verify. */ if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex || atomic_load_ptr(&m->object) != object) { prev = NULL; /* * This guarantees the result is instantaneously * correct. */ m = vm_radix_lookup_unlocked(&object->rtree, pindex); } if (m == NULL) return (true); if (vm_page_trybusy(m, allocflags)) { if (m->object == object && m->pindex == pindex) break; /* relookup. */ vm_page_busy_release(m); cpu_spinwait(); continue; } if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", allocflags, false)) return (false); } if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); vm_page_grab_release(m, allocflags); *mp = m; return (true); } /* * Try to locklessly grab a page and fall back to the object lock if NOCREAT * is not set. */ vm_page_t vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; vm_page_grab_check(allocflags); if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags)) return (NULL); if (m != NULL) return (m); /* * The radix lockless lookup should never return a false negative * errors. If the user specifies NOCREAT they are guaranteed there * was no page present at the instant of the call. A NOCREAT caller * must handle create races gracefully. */ if ((allocflags & VM_ALLOC_NOCREAT) != 0) return (NULL); VM_OBJECT_WLOCK(object); m = vm_page_grab(object, pindex, allocflags); VM_OBJECT_WUNLOCK(object); return (m); } /* * Grab a page and make it valid, paging in if necessary. Pages missing from * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought * in simultaneously. Additional pages will be left on a paging queue but * will neither be wired nor busy regardless of allocflags. */ int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; vm_page_t ma[VM_INITIAL_PAGEIN]; int after, i, pflags, rv; KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); KASSERT((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); VM_OBJECT_ASSERT_WLOCKED(object); pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); pflags |= VM_ALLOC_WAITFAIL; retrylookup: if ((m = vm_page_lookup(object, pindex)) != NULL) { /* * If the page is fully valid it can only become invalid * with the object lock held. If it is not valid it can * become valid with the busy lock held. Therefore, we * may unnecessarily lock the exclusive busy here if we * race with I/O completion not using the object lock. * However, we will not end up with an invalid page and a * shared lock. */ if (!vm_page_trybusy(m, vm_page_all_valid(m) ? allocflags : 0)) { (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", allocflags, true); goto retrylookup; } if (vm_page_all_valid(m)) goto out; if ((allocflags & VM_ALLOC_NOCREAT) != 0) { vm_page_busy_release(m); *mp = NULL; return (VM_PAGER_FAIL); } } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { *mp = NULL; return (VM_PAGER_FAIL); } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { if (!vm_pager_can_alloc_page(object, pindex)) { *mp = NULL; return (VM_PAGER_AGAIN); } goto retrylookup; } vm_page_assert_xbusied(m); if (vm_pager_has_page(object, pindex, NULL, &after)) { after = MIN(after, VM_INITIAL_PAGEIN); after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); after = MAX(after, 1); ma[0] = m; for (i = 1; i < after; i++) { if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { if (vm_page_any_valid(ma[i]) || !vm_page_tryxbusy(ma[i])) break; } else { ma[i] = vm_page_alloc(object, m->pindex + i, VM_ALLOC_NORMAL); if (ma[i] == NULL) break; } } after = i; vm_object_pip_add(object, after); VM_OBJECT_WUNLOCK(object); rv = vm_pager_get_pages(object, ma, after, NULL, NULL); VM_OBJECT_WLOCK(object); vm_object_pip_wakeupn(object, after); /* Pager may have replaced a page. */ m = ma[0]; if (rv != VM_PAGER_OK) { for (i = 0; i < after; i++) { if (!vm_page_wired(ma[i])) vm_page_free(ma[i]); else vm_page_xunbusy(ma[i]); } *mp = NULL; return (rv); } for (i = 1; i < after; i++) vm_page_readahead_finish(ma[i]); MPASS(vm_page_all_valid(m)); } else { vm_page_zero_invalid(m, TRUE); } out: if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) vm_page_busy_downgrade(m); else if ((allocflags & VM_ALLOC_NOBUSY) != 0) vm_page_busy_release(m); *mp = m; return (VM_PAGER_OK); } /* * Locklessly grab a valid page. If the page is not valid or not yet * allocated this will fall back to the object lock method. */ int vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; int flags; int error; KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " "mismatch")); KASSERT((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); /* * Attempt a lockless lookup and busy. We need at least an sbusy * before we can inspect the valid field and return a wired page. */ flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags)) return (VM_PAGER_FAIL); if ((m = *mp) != NULL) { if (vm_page_all_valid(m)) { if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); vm_page_grab_release(m, allocflags); return (VM_PAGER_OK); } vm_page_busy_release(m); } if ((allocflags & VM_ALLOC_NOCREAT) != 0) { *mp = NULL; return (VM_PAGER_FAIL); } VM_OBJECT_WLOCK(object); error = vm_page_grab_valid(mp, object, pindex, allocflags); VM_OBJECT_WUNLOCK(object); return (error); } /* * Return the specified range of pages from the given object. For each * page offset within the range, if a page already exists within the object * at that offset and it is busy, then wait for it to change state. If, * instead, the page doesn't exist, then allocate it. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs the pages * * The caller must always specify that the pages are to be busied and/or * wired. * * optional allocation flags: * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NOWAIT do not sleep * VM_ALLOC_SBUSY set page to sbusy state * VM_ALLOC_WIRED wire the pages * VM_ALLOC_ZERO zero and validate any invalid pages * * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it * may return a partial prefix of the requested range. */ int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count) { vm_page_t m, mpred; int pflags; int i; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); KASSERT(count > 0, ("vm_page_grab_pages: invalid page count %d", count)); vm_page_grab_check(allocflags); pflags = vm_page_grab_pflags(allocflags); i = 0; retrylookup: m = vm_radix_lookup_le(&object->rtree, pindex + i); if (m == NULL || m->pindex != pindex + i) { mpred = m; m = NULL; } else mpred = TAILQ_PREV(m, pglist, listq); for (; i < count; i++) { if (m != NULL) { if (!vm_page_tryacquire(m, allocflags)) { if (vm_page_grab_sleep(object, m, pindex + i, "grbmaw", allocflags, true)) goto retrylookup; break; } } else { if ((allocflags & VM_ALLOC_NOCREAT) != 0) break; m = vm_page_alloc_after(object, pindex + i, pflags | VM_ALLOC_COUNT(count - i), mpred); if (m == NULL) { if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) break; goto retrylookup; } } if (vm_page_none_valid(m) && (allocflags & VM_ALLOC_ZERO) != 0) { if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); vm_page_valid(m); } vm_page_grab_release(m, allocflags); ma[i] = mpred = m; m = vm_page_next(m); } return (i); } /* * Unlocked variant of vm_page_grab_pages(). This accepts the same flags * and will fall back to the locked variant to handle allocation. */ int vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count) { vm_page_t m, pred; int flags; int i; KASSERT(count > 0, ("vm_page_grab_pages_unlocked: invalid page count %d", count)); vm_page_grab_check(allocflags); /* * Modify flags for lockless acquire to hold the page until we * set it valid if necessary. */ flags = allocflags & ~VM_ALLOC_NOBUSY; pred = NULL; for (i = 0; i < count; i++, pindex++) { if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags)) return (i); if (m == NULL) break; if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); vm_page_valid(m); } /* m will still be wired or busy according to flags. */ vm_page_grab_release(m, allocflags); pred = ma[i] = m; } if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) return (i); count -= i; VM_OBJECT_WLOCK(object); i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); VM_OBJECT_WUNLOCK(object); return (i); } /* * Mapping function for valid or dirty bits in a page. * * Inputs are required to range within a page. */ vm_page_bits_t vm_page_bits(int base, int size) { int first_bit; int last_bit; KASSERT( base + size <= PAGE_SIZE, ("vm_page_bits: illegal base/size %d/%d", base, size) ); if (size == 0) /* handle degenerate case */ return (0); first_bit = base >> DEV_BSHIFT; last_bit = (base + size - 1) >> DEV_BSHIFT; return (((vm_page_bits_t)2 << last_bit) - ((vm_page_bits_t)1 << first_bit)); } void vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) { #if PAGE_SIZE == 32768 atomic_set_64((uint64_t *)bits, set); #elif PAGE_SIZE == 16384 atomic_set_32((uint32_t *)bits, set); #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) atomic_set_16((uint16_t *)bits, set); #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) atomic_set_8((uint8_t *)bits, set); #else /* PAGE_SIZE <= 8192 */ uintptr_t addr; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); atomic_set_32((uint32_t *)addr, set << shift); #endif /* PAGE_SIZE */ } static inline void vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) { #if PAGE_SIZE == 32768 atomic_clear_64((uint64_t *)bits, clear); #elif PAGE_SIZE == 16384 atomic_clear_32((uint32_t *)bits, clear); #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) atomic_clear_16((uint16_t *)bits, clear); #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) atomic_clear_8((uint8_t *)bits, clear); #else /* PAGE_SIZE <= 8192 */ uintptr_t addr; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); atomic_clear_32((uint32_t *)addr, clear << shift); #endif /* PAGE_SIZE */ } static inline vm_page_bits_t vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) { #if PAGE_SIZE == 32768 uint64_t old; old = *bits; while (atomic_fcmpset_64(bits, &old, newbits) == 0); return (old); #elif PAGE_SIZE == 16384 uint32_t old; old = *bits; while (atomic_fcmpset_32(bits, &old, newbits) == 0); return (old); #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) uint16_t old; old = *bits; while (atomic_fcmpset_16(bits, &old, newbits) == 0); return (old); #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) uint8_t old; old = *bits; while (atomic_fcmpset_8(bits, &old, newbits) == 0); return (old); #else /* PAGE_SIZE <= 4096*/ uintptr_t addr; uint32_t old, new, mask; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, swap, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); mask = VM_PAGE_BITS_ALL << shift; old = *bits; do { new = old & ~mask; new |= newbits << shift; } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); return (old >> shift); #endif /* PAGE_SIZE */ } /* * vm_page_set_valid_range: * * Sets portions of a page valid. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zeroed. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_valid_range(vm_page_t m, int base, int size) { int endoff, frag; vm_page_bits_t pagebits; vm_page_assert_busied(m); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Assert that no previously invalid block that is now being validated * is already dirty. */ KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, ("vm_page_set_valid_range: page %p is dirty", m)); /* * Set valid bits inclusive of any overlap. */ pagebits = vm_page_bits(base, size); if (vm_page_xbusied(m)) m->valid |= pagebits; else vm_page_bits_set(m, &m->valid, pagebits); } /* * Set the page dirty bits and free the invalid swap space if * present. Returns the previous dirty bits. */ vm_page_bits_t vm_page_set_dirty(vm_page_t m) { vm_page_bits_t old; VM_PAGE_OBJECT_BUSY_ASSERT(m); if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { old = m->dirty; m->dirty = VM_PAGE_BITS_ALL; } else old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) vm_pager_page_unswapped(m); return (old); } /* * Clear the given bits from the specified page's dirty field. */ static __inline void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) { vm_page_assert_busied(m); /* * If the page is xbusied and not write mapped we are the * only thread that can modify dirty bits. Otherwise, The pmap * layer can call vm_page_dirty() without holding a distinguished * lock. The combination of page busy and atomic operations * suffice to guarantee consistency of the page dirty field. */ if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) m->dirty &= ~pagebits; else vm_page_bits_clear(m, &m->dirty, pagebits); } /* * vm_page_set_validclean: * * Sets portions of a page valid and clean. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zero'd. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_validclean(vm_page_t m, int base, int size) { vm_page_bits_t oldvalid, pagebits; int endoff, frag; vm_page_assert_busied(m); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Set valid, clear dirty bits. If validating the entire * page we can safely clear the pmap modify bit. We also * use this opportunity to clear the PGA_NOSYNC flag. If a process * takes a write fault on a MAP_NOSYNC memory area the flag will * be set again. * * We set valid bits inclusive of any overlap, but we can only * clear dirty bits for DEV_BSIZE chunks that are fully within * the range. */ oldvalid = m->valid; pagebits = vm_page_bits(base, size); if (vm_page_xbusied(m)) m->valid |= pagebits; else vm_page_bits_set(m, &m->valid, pagebits); #if 0 /* NOT YET */ if ((frag = base & (DEV_BSIZE - 1)) != 0) { frag = DEV_BSIZE - frag; base += frag; size -= frag; if (size < 0) size = 0; } pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); #endif if (base == 0 && size == PAGE_SIZE) { /* * The page can only be modified within the pmap if it is * mapped, and it can only be mapped if it was previously * fully valid. */ if (oldvalid == VM_PAGE_BITS_ALL) /* * Perform the pmap_clear_modify() first. Otherwise, * a concurrent pmap operation, such as * pmap_protect(), could clear a modification in the * pmap and set the dirty field on the page before * pmap_clear_modify() had begun and after the dirty * field was cleared here. */ pmap_clear_modify(m); m->dirty = 0; vm_page_aflag_clear(m, PGA_NOSYNC); } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) m->dirty &= ~pagebits; else vm_page_clear_dirty_mask(m, pagebits); } void vm_page_clear_dirty(vm_page_t m, int base, int size) { vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); } /* * vm_page_set_invalid: * * Invalidates DEV_BSIZE'd chunks within a page. Both the * valid and dirty bits for the effected areas are cleared. */ void vm_page_set_invalid(vm_page_t m, int base, int size) { vm_page_bits_t bits; vm_object_t object; /* * The object lock is required so that pages can't be mapped * read-only while we're in the process of invalidating them. */ object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); vm_page_assert_busied(m); if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + size >= object->un_pager.vnp.vnp_size) bits = VM_PAGE_BITS_ALL; else bits = vm_page_bits(base, size); if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) pmap_remove_all(m); KASSERT((bits == 0 && vm_page_all_valid(m)) || !pmap_page_is_mapped(m), ("vm_page_set_invalid: page %p is mapped", m)); if (vm_page_xbusied(m)) { m->valid &= ~bits; m->dirty &= ~bits; } else { vm_page_bits_clear(m, &m->valid, bits); vm_page_bits_clear(m, &m->dirty, bits); } } /* * vm_page_invalid: * * Invalidates the entire page. The page must be busy, unmapped, and * the enclosing object must be locked. The object locks protects * against concurrent read-only pmap enter which is done without * busy. */ void vm_page_invalid(vm_page_t m) { vm_page_assert_busied(m); VM_OBJECT_ASSERT_WLOCKED(m->object); MPASS(!pmap_page_is_mapped(m)); if (vm_page_xbusied(m)) m->valid = 0; else vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); } /* * vm_page_zero_invalid() * * The kernel assumes that the invalid portions of a page contain * garbage, but such pages can be mapped into memory by user code. * When this occurs, we must zero out the non-valid portions of the * page so user code sees what it expects. * * Pages are most often semi-valid when the end of a file is mapped * into memory and the file's size is not page aligned. */ void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) { int b; int i; /* * Scan the valid bits looking for invalid sections that * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the * valid bit may be set ) have already been zeroed by * vm_page_set_validclean(). */ for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { if (i == (PAGE_SIZE / DEV_BSIZE) || (m->valid & ((vm_page_bits_t)1 << i))) { if (i > b) { pmap_zero_page_area(m, b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); } b = i + 1; } } /* * setvalid is TRUE when we can safely set the zero'd areas * as being valid. We can do this if there are no cache consistency * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. */ if (setvalid) vm_page_valid(m); } /* * vm_page_is_valid: * * Is (partial) page valid? Note that the case where size == 0 * will return FALSE in the degenerate case where the page is * entirely invalid, and TRUE otherwise. * * Some callers envoke this routine without the busy lock held and * handle races via higher level locks. Typical callers should * hold a busy lock to prevent invalidation. */ int vm_page_is_valid(vm_page_t m, int base, int size) { vm_page_bits_t bits; bits = vm_page_bits(base, size); return (vm_page_any_valid(m) && (m->valid & bits) == bits); } /* * Returns true if all of the specified predicates are true for the entire * (super)page and false otherwise. */ bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) { vm_object_t object; int i, npages; object = m->object; if (skip_m != NULL && skip_m->object != object) return (false); VM_OBJECT_ASSERT_LOCKED(object); npages = atop(pagesizes[m->psind]); /* * The physically contiguous pages that make up a superpage, i.e., a * page with a page size index ("psind") greater than zero, will * occupy adjacent entries in vm_page_array[]. */ for (i = 0; i < npages; i++) { /* Always test object consistency, including "skip_m". */ if (m[i].object != object) return (false); if (&m[i] == skip_m) continue; if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) return (false); if ((flags & PS_ALL_DIRTY) != 0) { /* * Calling vm_page_test_dirty() or pmap_is_modified() * might stop this case from spuriously returning * "false". However, that would require a write lock * on the object containing "m[i]". */ if (m[i].dirty != VM_PAGE_BITS_ALL) return (false); } if ((flags & PS_ALL_VALID) != 0 && m[i].valid != VM_PAGE_BITS_ALL) return (false); } return (true); } /* * Set the page's dirty bits if the page is modified. */ void vm_page_test_dirty(vm_page_t m) { vm_page_assert_busied(m); if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) vm_page_dirty(m); } void vm_page_valid(vm_page_t m) { vm_page_assert_busied(m); if (vm_page_xbusied(m)) m->valid = VM_PAGE_BITS_ALL; else vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); } void vm_page_lock_KBI(vm_page_t m, const char *file, int line) { mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); } void vm_page_unlock_KBI(vm_page_t m, const char *file, int line) { mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); } int vm_page_trylock_KBI(vm_page_t m, const char *file, int line) { return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); } #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) { vm_page_lock_assert_KBI(m, MA_OWNED, file, line); } void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) { mtx_assert_(vm_page_lockptr(m), a, file, line); } #endif #ifdef INVARIANTS void vm_page_object_busy_assert(vm_page_t m) { /* * Certain of the page's fields may only be modified by the * holder of a page or object busy. */ if (m->object != NULL && !vm_page_busied(m)) VM_OBJECT_ASSERT_BUSY(m->object); } void vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) { if ((bits & PGA_WRITEABLE) == 0) return; /* * The PGA_WRITEABLE flag can only be set if the page is * managed, is exclusively busied or the object is locked. * Currently, this flag is only set by pmap_enter(). */ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("PGA_WRITEABLE on unmanaged page")); if (!vm_page_xbusied(m)) VM_OBJECT_ASSERT_BUSY(m->object); } #endif #include "opt_ddb.h" #ifdef DDB #include #include DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) { db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); } DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) { int dom; db_printf("pq_free %d\n", vm_free_count()); for (dom = 0; dom < vm_ndomains; dom++) { db_printf( "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", dom, vm_dom[dom].vmd_page_count, vm_dom[dom].vmd_free_count, vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); } } DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) { vm_page_t m; boolean_t phys, virt; if (!have_addr) { db_printf("show pginfo addr\n"); return; } phys = strchr(modif, 'p') != NULL; virt = strchr(modif, 'v') != NULL; if (virt) m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); else if (phys) m = PHYS_TO_VM_PAGE(addr); else m = (vm_page_t)addr; db_printf( "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, m->a.queue, m->ref_count, m->a.flags, m->oflags, m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); } #endif /* DDB */ diff --git a/sys/vm/vm_page.h b/sys/vm/vm_page.h index 68f68d62d2be..e8d1497e2afc 100644 --- a/sys/vm/vm_page.h +++ b/sys/vm/vm_page.h @@ -1,1037 +1,1037 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_page.h 8.2 (Berkeley) 12/13/93 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Resident memory system definitions. */ #ifndef _VM_PAGE_ #define _VM_PAGE_ #include #include /* * Management of resident (logical) pages. * * A small structure is kept for each resident * page, indexed by page number. Each structure * is an element of several collections: * * A radix tree used to quickly * perform object/offset lookups * * A list of all pages for a given object, * so they can be quickly deactivated at * time of deallocation. * * An ordered list of pages due for pageout. * * In addition, the structure contains the object * and offset to which this page belongs (for pageout), * and sundry status bits. * * In general, operations on this structure's mutable fields are * synchronized using either one of or a combination of locks. If a * field is annotated with two of these locks then holding either is * sufficient for read access but both are required for write access. * The queue lock for a page depends on the value of its queue field and is * described in detail below. * * The following annotations are possible: * (A) the field must be accessed using atomic(9) and may require * additional synchronization. * (B) the page busy lock. * (C) the field is immutable. * (F) the per-domain lock for the free queues. * (M) Machine dependent, defined by pmap layer. * (O) the object that the page belongs to. * (Q) the page's queue lock. * * The busy lock is an embedded reader-writer lock that protects the * page's contents and identity (i.e., its tuple) as * well as certain valid/dirty modifications. To avoid bloating the * the page structure, the busy lock lacks some of the features available * the kernel's general-purpose synchronization primitives. As a result, * busy lock ordering rules are not verified, lock recursion is not * detected, and an attempt to xbusy a busy page or sbusy an xbusy page * results will trigger a panic rather than causing the thread to block. * vm_page_sleep_if_busy() can be used to sleep until the page's busy * state changes, after which the caller must re-lookup the page and * re-evaluate its state. vm_page_busy_acquire() will block until * the lock is acquired. * * The valid field is protected by the page busy lock (B) and object * lock (O). Transitions from invalid to valid are generally done * via I/O or zero filling and do not require the object lock. * These must be protected with the busy lock to prevent page-in or * creation races. Page invalidation generally happens as a result * of truncate or msync. When invalidated, pages must not be present * in pmap and must hold the object lock to prevent concurrent * speculative read-only mappings that do not require busy. I/O * routines may check for validity without a lock if they are prepared * to handle invalidation races with higher level locks (vnode) or are * unconcerned with races so long as they hold a reference to prevent * recycling. When a valid bit is set while holding a shared busy * lock (A) atomic operations are used to protect against concurrent * modification. * * In contrast, the synchronization of accesses to the page's * dirty field is a mix of machine dependent (M) and busy (B). In * the machine-independent layer, the page busy must be held to * operate on the field. However, the pmap layer is permitted to * set all bits within the field without holding that lock. If the * underlying architecture does not support atomic read-modify-write * operations on the field's type, then the machine-independent * layer uses a 32-bit atomic on the aligned 32-bit word that * contains the dirty field. In the machine-independent layer, * the implementation of read-modify-write operations on the * field is encapsulated in vm_page_clear_dirty_mask(). An * exclusive busy lock combined with pmap_remove_{write/all}() is the * only way to ensure a page can not become dirty. I/O generally * removes the page from pmap to ensure exclusive access and atomic * writes. * * The ref_count field tracks references to the page. References that * prevent the page from being reclaimable are called wirings and are * counted in the low bits of ref_count. The containing object's * reference, if one exists, is counted using the VPRC_OBJREF bit in the * ref_count field. Additionally, the VPRC_BLOCKED bit is used to * atomically check for wirings and prevent new wirings via * pmap_extract_and_hold(). When a page belongs to an object, it may be * wired only when the object is locked, or the page is busy, or by * pmap_extract_and_hold(). As a result, if the object is locked and the * page is not busy (or is exclusively busied by the current thread), and * the page is unmapped, its wire count will not increase. The ref_count * field is updated using atomic operations in most cases, except when it * is known that no other references to the page exist, such as in the page * allocator. A page may be present in the page queues, or even actively * scanned by the page daemon, without an explicitly counted referenced. * The page daemon must therefore handle the possibility of a concurrent * free of the page. * * The queue state of a page consists of the queue and act_count fields of * its atomically updated state, and the subset of atomic flags specified * by PGA_QUEUE_STATE_MASK. The queue field contains the page's page queue * index, or PQ_NONE if it does not belong to a page queue. To modify the * queue field, the page queue lock corresponding to the old value must be * held, unless that value is PQ_NONE, in which case the queue index must * be updated using an atomic RMW operation. There is one exception to * this rule: the page daemon may transition the queue field from * PQ_INACTIVE to PQ_NONE immediately prior to freeing the page during an * inactive queue scan. At that point the page is already dequeued and no * other references to that vm_page structure can exist. The PGA_ENQUEUED * flag, when set, indicates that the page structure is physically inserted * into the queue corresponding to the page's queue index, and may only be * set or cleared with the corresponding page queue lock held. * * To avoid contention on page queue locks, page queue operations (enqueue, * dequeue, requeue) are batched using fixed-size per-CPU queues. A * deferred operation is requested by setting one of the flags in * PGA_QUEUE_OP_MASK and inserting an entry into a batch queue. When a * queue is full, an attempt to insert a new entry will lock the page * queues and trigger processing of the pending entries. The * type-stability of vm_page structures is crucial to this scheme since the * processing of entries in a given batch queue may be deferred * indefinitely. In particular, a page may be freed with pending batch * queue entries. The page queue operation flags must be set using atomic * RWM operations. */ #if PAGE_SIZE == 4096 #define VM_PAGE_BITS_ALL 0xffu typedef uint8_t vm_page_bits_t; #elif PAGE_SIZE == 8192 #define VM_PAGE_BITS_ALL 0xffffu typedef uint16_t vm_page_bits_t; #elif PAGE_SIZE == 16384 #define VM_PAGE_BITS_ALL 0xffffffffu typedef uint32_t vm_page_bits_t; #elif PAGE_SIZE == 32768 #define VM_PAGE_BITS_ALL 0xfffffffffffffffflu typedef uint64_t vm_page_bits_t; #endif typedef union vm_page_astate { struct { uint16_t flags; uint8_t queue; uint8_t act_count; }; uint32_t _bits; } vm_page_astate_t; struct vm_page { union { TAILQ_ENTRY(vm_page) q; /* page queue or free list (Q) */ struct { SLIST_ENTRY(vm_page) ss; /* private slists */ } s; struct { u_long p; u_long v; } memguard; struct { void *slab; void *zone; } uma; } plinks; TAILQ_ENTRY(vm_page) listq; /* pages in same object (O) */ vm_object_t object; /* which object am I in (O) */ vm_pindex_t pindex; /* offset into object (O,P) */ vm_paddr_t phys_addr; /* physical address of page (C) */ struct md_page md; /* machine dependent stuff */ u_int ref_count; /* page references (A) */ u_int busy_lock; /* busy owners lock (A) */ union vm_page_astate a; /* state accessed atomically (A) */ uint8_t order; /* index of the buddy queue (F) */ uint8_t pool; /* vm_phys freepool index (F) */ uint8_t flags; /* page PG_* flags (P) */ uint8_t oflags; /* page VPO_* flags (O) */ int8_t psind; /* pagesizes[] index (O) */ int8_t segind; /* vm_phys segment index (C) */ /* NOTE that these must support one bit per DEV_BSIZE in a page */ /* so, on normal X86 kernels, they must be at least 8 bits wide */ vm_page_bits_t valid; /* valid DEV_BSIZE chunk map (O,B) */ vm_page_bits_t dirty; /* dirty DEV_BSIZE chunk map (M,B) */ }; /* * Special bits used in the ref_count field. * * ref_count is normally used to count wirings that prevent the page from being * reclaimed, but also supports several special types of references that do not * prevent reclamation. Accesses to the ref_count field must be atomic unless * the page is unallocated. * * VPRC_OBJREF is the reference held by the containing object. It can set or * cleared only when the corresponding object's write lock is held. * * VPRC_BLOCKED is used to atomically block wirings via pmap lookups while * attempting to tear down all mappings of a given page. The page busy lock and * object write lock must both be held in order to set or clear this bit. */ #define VPRC_BLOCKED 0x40000000u /* mappings are being removed */ #define VPRC_OBJREF 0x80000000u /* object reference, cleared with (O) */ #define VPRC_WIRE_COUNT(c) ((c) & ~(VPRC_BLOCKED | VPRC_OBJREF)) #define VPRC_WIRE_COUNT_MAX (~(VPRC_BLOCKED | VPRC_OBJREF)) /* * Page flags stored in oflags: * * Access to these page flags is synchronized by the lock on the object * containing the page (O). * * Note: VPO_UNMANAGED (used by OBJT_DEVICE, OBJT_PHYS and OBJT_SG) * indicates that the page is not under PV management but * otherwise should be treated as a normal page. Pages not * under PV management cannot be paged out via the * object/vm_page_t because there is no knowledge of their pte * mappings, and such pages are also not on any PQ queue. * */ #define VPO_KMEM_EXEC 0x01 /* kmem mapping allows execution */ #define VPO_SWAPSLEEP 0x02 /* waiting for swap to finish */ #define VPO_UNMANAGED 0x04 /* no PV management for page */ #define VPO_SWAPINPROG 0x08 /* swap I/O in progress on page */ /* * Busy page implementation details. * The algorithm is taken mostly by rwlock(9) and sx(9) locks implementation, * even if the support for owner identity is removed because of size * constraints. Checks on lock recursion are then not possible, while the * lock assertions effectiveness is someway reduced. */ #define VPB_BIT_SHARED 0x01 #define VPB_BIT_EXCLUSIVE 0x02 #define VPB_BIT_WAITERS 0x04 #define VPB_BIT_FLAGMASK \ (VPB_BIT_SHARED | VPB_BIT_EXCLUSIVE | VPB_BIT_WAITERS) #define VPB_SHARERS_SHIFT 3 #define VPB_SHARERS(x) \ (((x) & ~VPB_BIT_FLAGMASK) >> VPB_SHARERS_SHIFT) #define VPB_SHARERS_WORD(x) ((x) << VPB_SHARERS_SHIFT | VPB_BIT_SHARED) #define VPB_ONE_SHARER (1 << VPB_SHARERS_SHIFT) #define VPB_SINGLE_EXCLUSIVE VPB_BIT_EXCLUSIVE #ifdef INVARIANTS #define VPB_CURTHREAD_EXCLUSIVE \ (VPB_BIT_EXCLUSIVE | ((u_int)(uintptr_t)curthread & ~VPB_BIT_FLAGMASK)) #else #define VPB_CURTHREAD_EXCLUSIVE VPB_SINGLE_EXCLUSIVE #endif #define VPB_UNBUSIED VPB_SHARERS_WORD(0) /* Freed lock blocks both shared and exclusive. */ #define VPB_FREED (0xffffffff - VPB_BIT_SHARED) #define PQ_NONE 255 #define PQ_INACTIVE 0 #define PQ_ACTIVE 1 #define PQ_LAUNDRY 2 #define PQ_UNSWAPPABLE 3 #define PQ_COUNT 4 #ifndef VM_PAGE_HAVE_PGLIST TAILQ_HEAD(pglist, vm_page); #define VM_PAGE_HAVE_PGLIST #endif SLIST_HEAD(spglist, vm_page); #ifdef _KERNEL extern vm_page_t bogus_page; #endif /* _KERNEL */ extern struct mtx_padalign pa_lock[]; #if defined(__arm__) #define PDRSHIFT PDR_SHIFT #elif !defined(PDRSHIFT) #define PDRSHIFT 21 #endif #define pa_index(pa) ((pa) >> PDRSHIFT) #define PA_LOCKPTR(pa) ((struct mtx *)(&pa_lock[pa_index(pa) % PA_LOCK_COUNT])) #define PA_LOCKOBJPTR(pa) ((struct lock_object *)PA_LOCKPTR((pa))) #define PA_LOCK(pa) mtx_lock(PA_LOCKPTR(pa)) #define PA_TRYLOCK(pa) mtx_trylock(PA_LOCKPTR(pa)) #define PA_UNLOCK(pa) mtx_unlock(PA_LOCKPTR(pa)) #define PA_UNLOCK_COND(pa) \ do { \ if ((pa) != 0) { \ PA_UNLOCK((pa)); \ (pa) = 0; \ } \ } while (0) #define PA_LOCK_ASSERT(pa, a) mtx_assert(PA_LOCKPTR(pa), (a)) #if defined(KLD_MODULE) && !defined(KLD_TIED) #define vm_page_lock(m) vm_page_lock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_unlock(m) vm_page_unlock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_trylock(m) vm_page_trylock_KBI((m), LOCK_FILE, LOCK_LINE) #else /* !KLD_MODULE */ #define vm_page_lockptr(m) (PA_LOCKPTR(VM_PAGE_TO_PHYS((m)))) #define vm_page_lock(m) mtx_lock(vm_page_lockptr((m))) #define vm_page_unlock(m) mtx_unlock(vm_page_lockptr((m))) #define vm_page_trylock(m) mtx_trylock(vm_page_lockptr((m))) #endif #if defined(INVARIANTS) #define vm_page_assert_locked(m) \ vm_page_assert_locked_KBI((m), __FILE__, __LINE__) #define vm_page_lock_assert(m, a) \ vm_page_lock_assert_KBI((m), (a), __FILE__, __LINE__) #else #define vm_page_assert_locked(m) #define vm_page_lock_assert(m, a) #endif /* * The vm_page's aflags are updated using atomic operations. To set or clear * these flags, the functions vm_page_aflag_set() and vm_page_aflag_clear() * must be used. Neither these flags nor these functions are part of the KBI. * * PGA_REFERENCED may be cleared only if the page is locked. It is set by * both the MI and MD VM layers. However, kernel loadable modules should not * directly set this flag. They should call vm_page_reference() instead. * * PGA_WRITEABLE is set exclusively on managed pages by pmap_enter(). * When it does so, the object must be locked, or the page must be * exclusive busied. The MI VM layer must never access this flag * directly. Instead, it should call pmap_page_is_write_mapped(). * * PGA_EXECUTABLE may be set by pmap routines, and indicates that a page has * at least one executable mapping. It is not consumed by the MI VM layer. * * PGA_NOSYNC must be set and cleared with the page busy lock held. * * PGA_ENQUEUED is set and cleared when a page is inserted into or removed * from a page queue, respectively. It determines whether the plinks.q field * of the page is valid. To set or clear this flag, page's "queue" field must * be a valid queue index, and the corresponding page queue lock must be held. * * PGA_DEQUEUE is set when the page is scheduled to be dequeued from a page * queue, and cleared when the dequeue request is processed. A page may * have PGA_DEQUEUE set and PGA_ENQUEUED cleared, for instance if a dequeue * is requested after the page is scheduled to be enqueued but before it is * actually inserted into the page queue. * * PGA_REQUEUE is set when the page is scheduled to be enqueued or requeued * in its page queue. * * PGA_REQUEUE_HEAD is a special flag for enqueuing pages near the head of * the inactive queue, thus bypassing LRU. * * The PGA_DEQUEUE, PGA_REQUEUE and PGA_REQUEUE_HEAD flags must be set using an * atomic RMW operation to ensure that the "queue" field is a valid queue index, * and the corresponding page queue lock must be held when clearing any of the * flags. * * PGA_SWAP_FREE is used to defer freeing swap space to the pageout daemon * when the context that dirties the page does not have the object write lock * held. */ #define PGA_WRITEABLE 0x0001 /* page may be mapped writeable */ #define PGA_REFERENCED 0x0002 /* page has been referenced */ #define PGA_EXECUTABLE 0x0004 /* page may be mapped executable */ #define PGA_ENQUEUED 0x0008 /* page is enqueued in a page queue */ #define PGA_DEQUEUE 0x0010 /* page is due to be dequeued */ #define PGA_REQUEUE 0x0020 /* page is due to be requeued */ #define PGA_REQUEUE_HEAD 0x0040 /* page requeue should bypass LRU */ #define PGA_NOSYNC 0x0080 /* do not collect for syncer */ #define PGA_SWAP_FREE 0x0100 /* page with swap space was dirtied */ #define PGA_SWAP_SPACE 0x0200 /* page has allocated swap space */ #define PGA_QUEUE_OP_MASK (PGA_DEQUEUE | PGA_REQUEUE | PGA_REQUEUE_HEAD) #define PGA_QUEUE_STATE_MASK (PGA_ENQUEUED | PGA_QUEUE_OP_MASK) /* * Page flags. Updates to these flags are not synchronized, and thus they must * be set during page allocation or free to avoid races. * * The PG_PCPU_CACHE flag is set at allocation time if the page was * allocated from a per-CPU cache. It is cleared the next time that the * page is allocated from the physical memory allocator. */ #define PG_PCPU_CACHE 0x01 /* was allocated from per-CPU caches */ #define PG_FICTITIOUS 0x02 /* physical page doesn't exist */ #define PG_ZERO 0x04 /* page is zeroed */ #define PG_MARKER 0x08 /* special queue marker page */ #define PG_NODUMP 0x10 /* don't include this page in a dump */ /* * Misc constants. */ #define ACT_DECLINE 1 #define ACT_ADVANCE 3 #define ACT_INIT 5 #define ACT_MAX 64 #ifdef _KERNEL #include #include /* * Each pageable resident page falls into one of five lists: * * free * Available for allocation now. * * inactive * Low activity, candidates for reclamation. * This list is approximately LRU ordered. * * laundry * This is the list of pages that should be * paged out next. * * unswappable * Dirty anonymous pages that cannot be paged * out because no swap device is configured. * * active * Pages that are "active", i.e., they have been * recently referenced. * */ extern vm_page_t vm_page_array; /* First resident page in table */ extern long vm_page_array_size; /* number of vm_page_t's */ extern long first_page; /* first physical page number */ #define VM_PAGE_TO_PHYS(entry) ((entry)->phys_addr) /* * PHYS_TO_VM_PAGE() returns the vm_page_t object that represents a memory * page to which the given physical address belongs. The correct vm_page_t * object is returned for addresses that are not page-aligned. */ vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa); /* * Page allocation parameters for vm_page for the functions * vm_page_alloc(), vm_page_grab(), vm_page_alloc_contig() and * vm_page_alloc_freelist(). Some functions support only a subset * of the flags, and ignore others, see the flags legend. * * The meaning of VM_ALLOC_ZERO differs slightly between the vm_page_alloc*() * and the vm_page_grab*() functions. See these functions for details. * * Bits 0 - 1 define class. * Bits 2 - 15 dedicated for flags. * Legend: * (a) - vm_page_alloc() supports the flag. * (c) - vm_page_alloc_contig() supports the flag. * (g) - vm_page_grab() supports the flag. * (n) - vm_page_alloc_noobj() and vm_page_alloc_freelist() support the flag. * (p) - vm_page_grab_pages() supports the flag. * Bits above 15 define the count of additional pages that the caller * intends to allocate. */ #define VM_ALLOC_NORMAL 0 #define VM_ALLOC_INTERRUPT 1 #define VM_ALLOC_SYSTEM 2 #define VM_ALLOC_CLASS_MASK 3 #define VM_ALLOC_WAITOK 0x0008 /* (acn) Sleep and retry */ #define VM_ALLOC_WAITFAIL 0x0010 /* (acn) Sleep and return error */ #define VM_ALLOC_WIRED 0x0020 /* (acgnp) Allocate a wired page */ #define VM_ALLOC_ZERO 0x0040 /* (acgnp) Allocate a zeroed page */ #define VM_ALLOC_NORECLAIM 0x0080 /* (c) Do not reclaim after failure */ #define VM_ALLOC_AVAIL0 0x0100 #define VM_ALLOC_NOBUSY 0x0200 /* (acgp) Do not excl busy the page */ #define VM_ALLOC_NOCREAT 0x0400 /* (gp) Don't create a page */ #define VM_ALLOC_AVAIL1 0x0800 #define VM_ALLOC_IGN_SBUSY 0x1000 /* (gp) Ignore shared busy flag */ #define VM_ALLOC_NODUMP 0x2000 /* (ag) don't include in dump */ #define VM_ALLOC_SBUSY 0x4000 /* (acgp) Shared busy the page */ #define VM_ALLOC_NOWAIT 0x8000 /* (acgnp) Do not sleep */ #define VM_ALLOC_COUNT_MAX 0xffff #define VM_ALLOC_COUNT_SHIFT 16 #define VM_ALLOC_COUNT_MASK (VM_ALLOC_COUNT(VM_ALLOC_COUNT_MAX)) #define VM_ALLOC_COUNT(count) ({ \ KASSERT((count) <= VM_ALLOC_COUNT_MAX, \ ("%s: invalid VM_ALLOC_COUNT value", __func__)); \ (count) << VM_ALLOC_COUNT_SHIFT; \ }) #ifdef M_NOWAIT static inline int malloc2vm_flags(int malloc_flags) { int pflags; KASSERT((malloc_flags & M_USE_RESERVE) == 0 || (malloc_flags & M_NOWAIT) != 0, ("M_USE_RESERVE requires M_NOWAIT")); pflags = (malloc_flags & M_USE_RESERVE) != 0 ? VM_ALLOC_INTERRUPT : VM_ALLOC_SYSTEM; if ((malloc_flags & M_ZERO) != 0) pflags |= VM_ALLOC_ZERO; if ((malloc_flags & M_NODUMP) != 0) pflags |= VM_ALLOC_NODUMP; if ((malloc_flags & M_NOWAIT)) pflags |= VM_ALLOC_NOWAIT; if ((malloc_flags & M_WAITOK)) pflags |= VM_ALLOC_WAITOK; if ((malloc_flags & M_NORECLAIM)) pflags |= VM_ALLOC_NORECLAIM; return (pflags); } #endif /* * Predicates supported by vm_page_ps_test(): * * PS_ALL_DIRTY is true only if the entire (super)page is dirty. * However, it can be spuriously false when the (super)page has become * dirty in the pmap but that information has not been propagated to the * machine-independent layer. */ #define PS_ALL_DIRTY 0x1 #define PS_ALL_VALID 0x2 #define PS_NONE_BUSY 0x4 bool vm_page_busy_acquire(vm_page_t m, int allocflags); void vm_page_busy_downgrade(vm_page_t m); int vm_page_busy_tryupgrade(vm_page_t m); bool vm_page_busy_sleep(vm_page_t m, const char *msg, int allocflags); void vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags); void vm_page_free(vm_page_t m); void vm_page_free_zero(vm_page_t m); void vm_page_activate (vm_page_t); void vm_page_advise(vm_page_t m, int advice); vm_page_t vm_page_alloc(vm_object_t, vm_pindex_t, int); vm_page_t vm_page_alloc_domain(vm_object_t, vm_pindex_t, int, int); vm_page_t vm_page_alloc_after(vm_object_t, vm_pindex_t, int, vm_page_t); vm_page_t vm_page_alloc_domain_after(vm_object_t, vm_pindex_t, int, int, vm_page_t); vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_freelist(int, int); vm_page_t vm_page_alloc_freelist_domain(int, int, int); vm_page_t vm_page_alloc_noobj(int); vm_page_t vm_page_alloc_noobj_domain(int, int); vm_page_t vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); void vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set); bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose); vm_page_t vm_page_grab(vm_object_t, vm_pindex_t, int); vm_page_t vm_page_grab_unlocked(vm_object_t, vm_pindex_t, int); int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count); int vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count); int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags); int vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags); void vm_page_deactivate(vm_page_t); void vm_page_deactivate_noreuse(vm_page_t); void vm_page_dequeue(vm_page_t m); void vm_page_dequeue_deferred(vm_page_t m); vm_page_t vm_page_find_least(vm_object_t, vm_pindex_t); void vm_page_free_invalid(vm_page_t); vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags); void vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind); int vm_page_insert (vm_page_t, vm_object_t, vm_pindex_t); void vm_page_invalid(vm_page_t m); void vm_page_launder(vm_page_t m); vm_page_t vm_page_lookup(vm_object_t, vm_pindex_t); vm_page_t vm_page_lookup_unlocked(vm_object_t, vm_pindex_t); vm_page_t vm_page_next(vm_page_t m); void vm_page_pqbatch_drain(void); void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue); bool vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new); vm_page_t vm_page_prev(vm_page_t m); bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m); void vm_page_putfake(vm_page_t m); void vm_page_readahead_finish(vm_page_t m); -bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, +int vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); -bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, +int vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); -bool vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, +int vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, int desired_runs); void vm_page_reference(vm_page_t m); #define VPR_TRYFREE 0x01 #define VPR_NOREUSE 0x02 void vm_page_release(vm_page_t m, int flags); void vm_page_release_locked(vm_page_t m, int flags); vm_page_t vm_page_relookup(vm_object_t, vm_pindex_t); bool vm_page_remove(vm_page_t); bool vm_page_remove_xbusy(vm_page_t); int vm_page_rename(vm_page_t, vm_object_t, vm_pindex_t); void vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold); int vm_page_sbusied(vm_page_t m); vm_page_bits_t vm_page_set_dirty(vm_page_t m); void vm_page_set_valid_range(vm_page_t m, int base, int size); vm_offset_t vm_page_startup(vm_offset_t vaddr); void vm_page_sunbusy(vm_page_t m); bool vm_page_try_remove_all(vm_page_t m); bool vm_page_try_remove_write(vm_page_t m); int vm_page_trysbusy(vm_page_t m); int vm_page_tryxbusy(vm_page_t m); void vm_page_unhold_pages(vm_page_t *ma, int count); void vm_page_unswappable(vm_page_t m); void vm_page_unwire(vm_page_t m, uint8_t queue); bool vm_page_unwire_noq(vm_page_t m); void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_wire(vm_page_t); bool vm_page_wire_mapped(vm_page_t m); void vm_page_xunbusy_hard(vm_page_t m); void vm_page_xunbusy_hard_unchecked(vm_page_t m); void vm_page_set_validclean (vm_page_t, int, int); void vm_page_clear_dirty(vm_page_t, int, int); void vm_page_set_invalid(vm_page_t, int, int); void vm_page_valid(vm_page_t m); int vm_page_is_valid(vm_page_t, int, int); void vm_page_test_dirty(vm_page_t); vm_page_bits_t vm_page_bits(int base, int size); void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid); int vm_page_free_pages_toq(struct spglist *free, bool update_wire_count); void vm_page_dirty_KBI(vm_page_t m); void vm_page_lock_KBI(vm_page_t m, const char *file, int line); void vm_page_unlock_KBI(vm_page_t m, const char *file, int line); int vm_page_trylock_KBI(vm_page_t m, const char *file, int line); #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line); void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line); #endif #define vm_page_busy_fetch(m) atomic_load_int(&(m)->busy_lock) #define vm_page_assert_busied(m) \ KASSERT(vm_page_busied(m), \ ("vm_page_assert_busied: page %p not busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_sbusied(m) \ KASSERT(vm_page_sbusied(m), \ ("vm_page_assert_sbusied: page %p not shared busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_unbusied(m) \ KASSERT((vm_page_busy_fetch(m) & ~VPB_BIT_WAITERS) != \ VPB_CURTHREAD_EXCLUSIVE, \ ("vm_page_assert_xbusied: page %p busy_lock %#x owned" \ " by me @ %s:%d", \ (m), (m)->busy_lock, __FILE__, __LINE__)); \ #define vm_page_assert_xbusied_unchecked(m) do { \ KASSERT(vm_page_xbusied(m), \ ("vm_page_assert_xbusied: page %p not exclusive busy @ %s:%d", \ (m), __FILE__, __LINE__)); \ } while (0) #define vm_page_assert_xbusied(m) do { \ vm_page_assert_xbusied_unchecked(m); \ KASSERT((vm_page_busy_fetch(m) & ~VPB_BIT_WAITERS) == \ VPB_CURTHREAD_EXCLUSIVE, \ ("vm_page_assert_xbusied: page %p busy_lock %#x not owned" \ " by me @ %s:%d", \ (m), (m)->busy_lock, __FILE__, __LINE__)); \ } while (0) #define vm_page_busied(m) \ (vm_page_busy_fetch(m) != VPB_UNBUSIED) #define vm_page_xbusied(m) \ ((vm_page_busy_fetch(m) & VPB_SINGLE_EXCLUSIVE) != 0) #define vm_page_busy_freed(m) \ (vm_page_busy_fetch(m) == VPB_FREED) /* Note: page m's lock must not be owned by the caller. */ #define vm_page_xunbusy(m) do { \ if (!atomic_cmpset_rel_int(&(m)->busy_lock, \ VPB_CURTHREAD_EXCLUSIVE, VPB_UNBUSIED)) \ vm_page_xunbusy_hard(m); \ } while (0) #define vm_page_xunbusy_unchecked(m) do { \ if (!atomic_cmpset_rel_int(&(m)->busy_lock, \ VPB_CURTHREAD_EXCLUSIVE, VPB_UNBUSIED)) \ vm_page_xunbusy_hard_unchecked(m); \ } while (0) #ifdef INVARIANTS void vm_page_object_busy_assert(vm_page_t m); #define VM_PAGE_OBJECT_BUSY_ASSERT(m) vm_page_object_busy_assert(m) void vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits); #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) \ vm_page_assert_pga_writeable(m, bits) /* * Claim ownership of a page's xbusy state. In non-INVARIANTS kernels this * operation is a no-op since ownership is not tracked. In particular * this macro does not provide any synchronization with the previous owner. */ #define vm_page_xbusy_claim(m) do { \ u_int _busy_lock; \ \ vm_page_assert_xbusied_unchecked((m)); \ do { \ _busy_lock = vm_page_busy_fetch(m); \ } while (!atomic_cmpset_int(&(m)->busy_lock, _busy_lock, \ (_busy_lock & VPB_BIT_FLAGMASK) | VPB_CURTHREAD_EXCLUSIVE)); \ } while (0) #else #define VM_PAGE_OBJECT_BUSY_ASSERT(m) (void)0 #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) (void)0 #define vm_page_xbusy_claim(m) #endif #if BYTE_ORDER == BIG_ENDIAN #define VM_PAGE_AFLAG_SHIFT 16 #else #define VM_PAGE_AFLAG_SHIFT 0 #endif /* * Load a snapshot of a page's 32-bit atomic state. */ static inline vm_page_astate_t vm_page_astate_load(vm_page_t m) { vm_page_astate_t a; a._bits = atomic_load_32(&m->a._bits); return (a); } /* * Atomically compare and set a page's atomic state. */ static inline bool vm_page_astate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { KASSERT(new.queue == PQ_INACTIVE || (new.flags & PGA_REQUEUE_HEAD) == 0, ("%s: invalid head requeue request for page %p", __func__, m)); KASSERT((new.flags & PGA_ENQUEUED) == 0 || new.queue != PQ_NONE, ("%s: setting PGA_ENQUEUED with PQ_NONE in page %p", __func__, m)); KASSERT(new._bits != old->_bits, ("%s: bits are unchanged", __func__)); return (atomic_fcmpset_32(&m->a._bits, &old->_bits, new._bits) != 0); } /* * Clear the given bits in the specified page. */ static inline void vm_page_aflag_clear(vm_page_t m, uint16_t bits) { uint32_t *addr, val; /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->a; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_clear_32(addr, val); } /* * Set the given bits in the specified page. */ static inline void vm_page_aflag_set(vm_page_t m, uint16_t bits) { uint32_t *addr, val; VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits); /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->a; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_set_32(addr, val); } /* * vm_page_dirty: * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). */ static __inline void vm_page_dirty(vm_page_t m) { /* Use vm_page_dirty_KBI() under INVARIANTS to save memory. */ #if (defined(KLD_MODULE) && !defined(KLD_TIED)) || defined(INVARIANTS) vm_page_dirty_KBI(m); #else m->dirty = VM_PAGE_BITS_ALL; #endif } /* * vm_page_undirty: * * Set page to not be dirty. Note: does not clear pmap modify bits */ static __inline void vm_page_undirty(vm_page_t m) { VM_PAGE_OBJECT_BUSY_ASSERT(m); m->dirty = 0; } static inline uint8_t _vm_page_queue(vm_page_astate_t as) { if ((as.flags & PGA_DEQUEUE) != 0) return (PQ_NONE); return (as.queue); } /* * vm_page_queue: * * Return the index of the queue containing m. */ static inline uint8_t vm_page_queue(vm_page_t m) { return (_vm_page_queue(vm_page_astate_load(m))); } static inline bool vm_page_active(vm_page_t m) { return (vm_page_queue(m) == PQ_ACTIVE); } static inline bool vm_page_inactive(vm_page_t m) { return (vm_page_queue(m) == PQ_INACTIVE); } static inline bool vm_page_in_laundry(vm_page_t m) { uint8_t queue; queue = vm_page_queue(m); return (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE); } static inline void vm_page_clearref(vm_page_t m) { u_int r; r = m->ref_count; while (atomic_fcmpset_int(&m->ref_count, &r, r & (VPRC_BLOCKED | VPRC_OBJREF)) == 0) ; } /* * vm_page_drop: * * Release a reference to a page and return the old reference count. */ static inline u_int vm_page_drop(vm_page_t m, u_int val) { u_int old; /* * Synchronize with vm_page_free_prep(): ensure that all updates to the * page structure are visible before it is freed. */ atomic_thread_fence_rel(); old = atomic_fetchadd_int(&m->ref_count, -val); KASSERT(old != VPRC_BLOCKED, ("vm_page_drop: page %p has an invalid refcount value", m)); return (old); } /* * vm_page_wired: * * Perform a racy check to determine whether a reference prevents the page * from being reclaimable. If the page's object is locked, and the page is * unmapped and exclusively busied by the current thread, no new wirings * may be created. */ static inline bool vm_page_wired(vm_page_t m) { return (VPRC_WIRE_COUNT(m->ref_count) > 0); } static inline bool vm_page_all_valid(vm_page_t m) { return (m->valid == VM_PAGE_BITS_ALL); } static inline bool vm_page_any_valid(vm_page_t m) { return (m->valid != 0); } static inline bool vm_page_none_valid(vm_page_t m) { return (m->valid == 0); } static inline int vm_page_domain(vm_page_t m) { #ifdef NUMA int domn, segind; segind = m->segind; KASSERT(segind < vm_phys_nsegs, ("segind %d m %p", segind, m)); domn = vm_phys_segs[segind].domain; KASSERT(domn >= 0 && domn < vm_ndomains, ("domain %d m %p", domn, m)); return (domn); #else return (0); #endif } #endif /* _KERNEL */ #endif /* !_VM_PAGE_ */