Index: sys/vm/vm_radix.c =================================================================== --- sys/vm/vm_radix.c (revision 252465) +++ sys/vm/vm_radix.c (working copy) @@ -103,8 +103,7 @@ struct vm_radix_node { static uma_zone_t vm_radix_node_zone; /* - * Allocate a radix node. Pre-allocation should ensure that the request - * will always be satisfied. + * Allocate a radix node. */ static __inline struct vm_radix_node * vm_radix_node_get(vm_pindex_t owner, uint16_t count, uint16_t clevel) @@ -112,21 +111,8 @@ vm_radix_node_get(vm_pindex_t owner, uint16_t coun struct vm_radix_node *rnode; rnode = uma_zalloc(vm_radix_node_zone, M_NOWAIT); - - /* - * The required number of nodes should already be pre-allocated - * by vm_radix_prealloc(). However, UMA can hold a few nodes - * in per-CPU buckets, which will not be accessible by the - * current CPU. Thus, the allocation could return NULL when - * the pre-allocated pool is close to exhaustion. Anyway, - * in practice this should never occur because a new node - * is not always required for insert. Thus, the pre-allocated - * pool should have some extra pages that prevent this from - * becoming a problem. - */ if (rnode == NULL) - panic("%s: uma_zalloc() returned NULL for a new node", - __func__); + return (NULL); rnode->rn_owner = owner; rnode->rn_count = count; rnode->rn_clev = clevel; @@ -309,30 +295,29 @@ vm_radix_node_zone_init(void *mem, int size __unus } /* - * Pre-allocate intermediate nodes from the UMA slab zone. + * Reserve the KVA necessary to satisfy the node allocation. + * This is particular useful in architectures not supporting direct + * mapping as they will need otherwise to carve into the kernel maps, + * resulting into deadlocks for several consumers. */ static void -vm_radix_prealloc(void *arg __unused) +vm_radix_reserve_kva(void *arg __unused) { - int nodes; /* * Calculate the number of reserved nodes, discounting the pages that * are needed to store them. */ - nodes = ((vm_paddr_t)cnt.v_page_count * PAGE_SIZE) / (PAGE_SIZE + - sizeof(struct vm_radix_node)); - if (!uma_zone_reserve_kva(vm_radix_node_zone, nodes)) + if (!uma_zone_reserve_kva(vm_radix_node_zone, + ((vm_paddr_t)cnt.v_page_count * PAGE_SIZE) / (PAGE_SIZE + + sizeof(struct vm_radix_node)))) panic("%s: unable to create new zone", __func__); - uma_prealloc(vm_radix_node_zone, nodes); } -SYSINIT(vm_radix_prealloc, SI_SUB_KMEM, SI_ORDER_SECOND, vm_radix_prealloc, - NULL); +SYSINIT(vm_radix_reserve_kva, SI_SUB_KMEM, SI_ORDER_SECOND, + vm_radix_reserve_kva, NULL); /* * Initialize the UMA slab zone. - * Until vm_radix_prealloc() is called, the zone will be served by the - * UMA boot-time pre-allocated pool of pages. */ void vm_radix_init(void) @@ -345,15 +330,14 @@ vm_radix_init(void) #else NULL, #endif - vm_radix_node_zone_init, NULL, VM_RADIX_PAD, UMA_ZONE_VM | - UMA_ZONE_NOFREE); + vm_radix_node_zone_init, NULL, VM_RADIX_PAD, UMA_ZONE_VM); } /* * Inserts the key-value pair into the trie. * Panics if the key already exists. */ -void +int vm_radix_insert(struct vm_radix *rtree, vm_page_t page) { vm_pindex_t index, newind; @@ -372,7 +356,7 @@ vm_radix_insert(struct vm_radix *rtree, vm_page_t rnode = vm_radix_getroot(rtree); if (rnode == NULL) { rtree->rt_root = (uintptr_t)page | VM_RADIX_ISLEAF; - return; + return (0); } parentp = (void **)&rtree->rt_root; for (;;) { @@ -384,17 +368,19 @@ vm_radix_insert(struct vm_radix *rtree, vm_page_t clev = vm_radix_keydiff(m->pindex, index); tmp = vm_radix_node_get(vm_radix_trimkey(index, clev + 1), 2, clev); + if (tmp == NULL) + return (ENOMEM); *parentp = tmp; vm_radix_addpage(tmp, index, clev, page); vm_radix_addpage(tmp, m->pindex, clev, m); - return; + return (0); } else if (vm_radix_keybarr(rnode, index)) break; slot = vm_radix_slot(index, rnode->rn_clev); if (rnode->rn_child[slot] == NULL) { rnode->rn_count++; vm_radix_addpage(rnode, index, rnode->rn_clev, page); - return; + return (0); } parentp = &rnode->rn_child[slot]; rnode = rnode->rn_child[slot]; @@ -409,10 +395,13 @@ vm_radix_insert(struct vm_radix *rtree, vm_page_t clev = vm_radix_keydiff(newind, index); tmp = vm_radix_node_get(vm_radix_trimkey(index, clev + 1), 2, clev); + if (tmp == NULL) + return (ENOMEM); *parentp = tmp; vm_radix_addpage(tmp, index, clev, page); slot = vm_radix_slot(newind, clev); tmp->rn_child[slot] = rnode; + return (0); } /* Index: sys/vm/vm_radix.h =================================================================== --- sys/vm/vm_radix.h (revision 252465) +++ sys/vm/vm_radix.h (working copy) @@ -36,7 +36,7 @@ #ifdef _KERNEL void vm_radix_init(void); -void vm_radix_insert(struct vm_radix *rtree, vm_page_t page); +int vm_radix_insert(struct vm_radix *rtree, vm_page_t page); vm_page_t vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index); vm_page_t vm_radix_lookup_ge(struct vm_radix *rtree, vm_pindex_t index); vm_page_t vm_radix_lookup_le(struct vm_radix *rtree, vm_pindex_t index); Index: sys/vm/sg_pager.c =================================================================== --- sys/vm/sg_pager.c (revision 252465) +++ sys/vm/sg_pager.c (working copy) @@ -42,6 +42,7 @@ __FBSDID("$FreeBSD$"); #include #include #include +#include #include #include #include @@ -190,7 +191,19 @@ sg_pager_getpages(vm_object_t object, vm_page_t *m vm_page_free(m[i]); vm_page_unlock(m[i]); } - vm_page_insert(page, object, offset); + for (;;) { + if (!vm_page_insert(page, object, offset)) + break; + VM_OBJECT_WUNLOCK(object); + VM_WAIT; + + /* + * The same conditions that applied earlier when inserting + * the pager page still apply so initial condition checks + * should not be necessary now. + */ + VM_OBJECT_WLOCK(object); + } m[reqpage] = page; page->valid = VM_PAGE_BITS_ALL; Index: sys/vm/vm_object.c =================================================================== --- sys/vm/vm_object.c (revision 252465) +++ sys/vm/vm_object.c (working copy) @@ -1351,6 +1351,16 @@ retry: VM_OBJECT_WLOCK(new_object); goto retry; } + + /* vm_page_rename() will handle dirty and cache. */ + if (vm_page_rename(m, new_object, idx)) { + VM_OBJECT_WUNLOCK(new_object); + VM_OBJECT_WUNLOCK(orig_object); + VM_WAIT; + VM_OBJECT_WLOCK(orig_object); + VM_OBJECT_WLOCK(new_object); + goto retry; + } #if VM_NRESERVLEVEL > 0 /* * If some of the reservation's allocated pages remain with @@ -1366,10 +1376,6 @@ retry: */ vm_reserv_rename(m, new_object, orig_object, offidxstart); #endif - vm_page_lock(m); - vm_page_rename(m, new_object, idx); - vm_page_unlock(m); - /* page automatically made dirty by rename and cache handled */ if (orig_object->type == OBJT_SWAP) vm_page_busy(m); } @@ -1527,21 +1533,14 @@ vm_object_backing_scan(vm_object_t object, int op) ("vm_object_backing_scan: object mismatch") ); - /* - * Destroy any associated swap - */ - if (backing_object->type == OBJT_SWAP) { - swap_pager_freespace( - backing_object, - p->pindex, - 1 - ); - } - if ( p->pindex < backing_offset_index || new_pindex >= object->size ) { + if (backing_object->type == OBJT_SWAP) + swap_pager_freespace(backing_object, + p->pindex, 1); + /* * Page is out of the parent object's range, we * can simply destroy it. @@ -1563,6 +1562,10 @@ vm_object_backing_scan(vm_object_t object, int op) (op & OBSC_COLLAPSE_NOWAIT) != 0 && (pp != NULL && pp->valid == 0) ) { + if (backing_object->type == OBJT_SWAP) + swap_pager_freespace(backing_object, + p->pindex, 1); + /* * The page in the parent is not (yet) valid. * We don't know anything about the state of @@ -1581,6 +1584,10 @@ vm_object_backing_scan(vm_object_t object, int op) pp != NULL || vm_pager_has_page(object, new_pindex, NULL, NULL) ) { + if (backing_object->type == OBJT_SWAP) + swap_pager_freespace(backing_object, + p->pindex, 1); + /* * page already exists in parent OR swap exists * for this location in the parent. Destroy @@ -1600,6 +1607,31 @@ vm_object_backing_scan(vm_object_t object, int op) continue; } + /* + * Page does not exist in parent, rename the + * page from the backing object to the main object. + * + * If the page was mapped to a process, it can remain + * mapped through the rename. + * vm_page_rename() will handle dirty and cache. + */ + if (vm_page_rename(p, object, new_pindex)) { + if (op & OBSC_COLLAPSE_NOWAIT) { + p = next; + continue; + } + VM_OBJECT_WLOCK(backing_object); + VM_OBJECT_WUNLOCK(object); + VM_WAIT; + VM_OBJECT_WLOCK(object); + VM_OBJECT_WLOCK(backing_object); + p = TAILQ_FIRST(&backing_object->memq); + continue; + } + if (backing_object->type == OBJT_SWAP) + swap_pager_freespace(backing_object, p->pindex, + 1); + #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. @@ -1607,18 +1639,6 @@ vm_object_backing_scan(vm_object_t object, int op) vm_reserv_rename(p, object, backing_object, backing_offset_index); #endif - - /* - * Page does not exist in parent, rename the - * page from the backing object to the main object. - * - * If the page was mapped to a process, it can remain - * mapped through the rename. - */ - vm_page_lock(p); - vm_page_rename(p, object, new_pindex); - vm_page_unlock(p); - /* page automatically made dirty by rename */ } p = next; } Index: sys/vm/device_pager.c =================================================================== --- sys/vm/device_pager.c (revision 252465) +++ sys/vm/device_pager.c (working copy) @@ -51,6 +51,7 @@ __FBSDID("$FreeBSD$"); #include #include #include +#include #include #include #include @@ -352,7 +353,19 @@ old_dev_pager_fault(vm_object_t object, vm_ooffset vm_page_free(*mres); vm_page_unlock(*mres); *mres = page; - vm_page_insert(page, object, pidx); + for (;;) { + if (!vm_page_insert(page, object, pidx)) + break; + VM_OBJECT_WUNLOCK(object); + VM_WAIT; + + /* + * The same conditions that applied earlier when + * inserting the pager page still apply so initial + * condition checks should not be necessary now. + */ + VM_OBJECT_WLOCK(object); + } } page->valid = VM_PAGE_BITS_ALL; return (VM_PAGER_OK); Index: sys/vm/vm_fault.c =================================================================== --- sys/vm/vm_fault.c (revision 252465) +++ sys/vm/vm_fault.c (working copy) @@ -741,9 +741,11 @@ vnode_locked: * process'es object. The page is * automatically made dirty. */ - vm_page_lock(fs.m); - vm_page_rename(fs.m, fs.first_object, fs.first_pindex); - vm_page_unlock(fs.m); + if (vm_page_rename(fs.m, fs.first_object, + fs.first_pindex)) { + unlock_and_deallocate(&fs); + goto RetryFault; + } vm_page_busy(fs.m); fs.first_m = fs.m; fs.m = NULL; Index: sys/vm/vm_page.c =================================================================== --- sys/vm/vm_page.c (revision 252465) +++ sys/vm/vm_page.c (working copy) @@ -159,11 +159,15 @@ SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFL static uma_zone_t fakepg_zone; static struct vnode *vm_page_alloc_init(vm_page_t m); +static void vm_page_alloc_turn_free(vm_page_t m); +static void vm_page_cache_turn_free(vm_page_t m); static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); static void vm_page_enqueue(int queue, vm_page_t m); static void vm_page_init_fakepg(void *dummy); -static void vm_page_insert_after(vm_page_t m, vm_object_t object, +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); SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); @@ -804,14 +808,14 @@ vm_page_dirty_KBI(vm_page_t m) * * The object must be locked. */ -void +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); - vm_page_insert_after(m, object, pindex, mpred); + return (vm_page_insert_after(m, object, pindex, mpred)); } /* @@ -824,10 +828,12 @@ vm_page_insert(vm_page_t m, vm_object_t object, vm * * The object must be locked. */ -static void +static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred) { + vm_pindex_t sidx; + vm_object_t sobj; vm_page_t msucc; VM_OBJECT_ASSERT_WLOCKED(object); @@ -844,22 +850,58 @@ vm_page_insert_after(vm_page_t m, vm_object_t obje msucc = TAILQ_FIRST(&object->memq); if (msucc != NULL) KASSERT(msucc->pindex > pindex, - ("vm_page_insert_after: msucc doesn't succeed pindex")); + ("vm_page_insert_after: msucc %lx doesn't succeed pindex %lx")); /* * Record the object/offset pair in this page */ + sobj = m->object; + sidx = m->pindex; m->object = object; m->pindex = pindex; /* * Now link into the object's ordered list of backed pages. */ + if (vm_radix_insert(&object->rtree, m)) { + m->object = sobj; + m->pindex = sidx; + return (1); + } + vm_page_insert_radixdone(m, object, mpred); + 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)); + if (mpred != NULL) { + KASSERT(mpred->object == object || + (mpred->flags & PG_SLAB) != 0, + ("vm_page_insert_after: object doesn't contain mpred")); + KASSERT(mpred->pindex < m->pindex, + ("vm_page_insert_after: mpred doesn't precede pindex")); + } + if (mpred != NULL) TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); else TAILQ_INSERT_HEAD(&object->memq, m, listq); - vm_radix_insert(&object->rtree, m); /* * Show that the object has one more resident page. @@ -1013,18 +1055,97 @@ vm_page_prev(vm_page_t m) * or vm_page_dirty() will panic. Dirty pages are not allowed * on the cache. * - * The objects must be locked. The page must be locked if it is managed. + * The objects must be locked. */ -void +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); + + 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_lock(m); vm_page_remove(m); - vm_page_insert(m, new_object, new_pindex); + vm_page_unlock(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); vm_page_dirty(m); + return (0); } /* + * Turn a cached page into a free page, by changing its attributes. + * Keep the statistics up-to-date. + * + * The free page queue must be locked. + */ +static void +vm_page_cache_turn_free(vm_page_t m) +{ + + mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); + + m->object = NULL; + m->valid = 0; + /* Clear PG_CACHED and set PG_FREE. */ + m->flags ^= PG_CACHED | PG_FREE; + KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, + ("vm_page_cache_free: page %p has inconsistent flags", m)); + cnt.v_cache_count--; + cnt.v_free_count++; +} + +/* + * Turn a partially allocated page into a free page, by changing its + * attributes. + * Keep the statistics up-to-date. + * + * The free page queue must be locked. + */ +static void +vm_page_alloc_turn_free(vm_page_t m) +{ + + mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); + + m->object = NULL; + m->valid = 0; + m->oflags = 0; + if (m->wire_count != 0) { + m->wire_count = 0; + atomic_subtract_int(&cnt.v_wire_count, 1); + } + m->flags &= ~PG_ZERO; + m->flags |= PG_FREE; + cnt.v_free_count++; +} + +/* * Convert all of the given object's cached pages that have a * pindex within the given range into free pages. If the value * zero is given for "end", then the range's upper bound is @@ -1047,14 +1168,7 @@ vm_page_cache_free(vm_object_t object, vm_pindex_t if (end != 0 && m->pindex >= end) break; vm_radix_remove(&object->cache, m->pindex); - m->object = NULL; - m->valid = 0; - /* Clear PG_CACHED and set PG_FREE. */ - m->flags ^= PG_CACHED | PG_FREE; - KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, - ("vm_page_cache_free: page %p has inconsistent flags", m)); - cnt.v_cache_count--; - cnt.v_free_count++; + vm_page_cache_turn_free(m); } empty = vm_radix_is_empty(&object->cache); mtx_unlock(&vm_page_queue_free_mtx); @@ -1134,7 +1248,8 @@ vm_page_cache_transfer(vm_object_t orig_object, vm /* Update the page's object and offset. */ m->object = new_object; m->pindex -= offidxstart; - vm_radix_insert(&new_object->cache, m); + if (vm_radix_insert(&new_object->cache, m)) + vm_page_cache_turn_free(m); } mtx_unlock(&vm_page_queue_free_mtx); } @@ -1201,7 +1316,7 @@ vm_page_alloc(vm_object_t object, vm_pindex_t pind struct vnode *vp = NULL; vm_object_t m_object; vm_page_t m, mpred; - int flags, req_class; + int flags, freelock_owned, req_class; mpred = 0; /* XXX: pacify gcc */ KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0), @@ -1222,7 +1337,15 @@ vm_page_alloc(vm_object_t object, vm_pindex_t pind KASSERT(mpred == NULL || mpred->pindex != pindex, ("vm_page_alloc: pindex already allocated")); } - mtx_lock(&vm_page_queue_free_mtx); + + /* + * The page allocation request can came from consumers which already + * hold the free page queue mutex, like vm_page_insert() in + * vm_page_cache(). + */ + freelock_owned = mtx_owned(&vm_page_queue_free_mtx); + if (!freelock_owned) + mtx_lock(&vm_page_queue_free_mtx); if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || @@ -1235,7 +1358,8 @@ vm_page_alloc(vm_object_t object, vm_pindex_t pind if (object != NULL && (m = vm_page_cache_lookup(object, pindex)) != NULL) { if ((req & VM_ALLOC_IFNOTCACHED) != 0) { - mtx_unlock(&vm_page_queue_free_mtx); + if (!freelock_owned) + mtx_unlock(&vm_page_queue_free_mtx); return (NULL); } if (vm_phys_unfree_page(m)) @@ -1248,7 +1372,8 @@ vm_page_alloc(vm_object_t object, vm_pindex_t pind panic("vm_page_alloc: cache page %p is missing" " from the free queue", m); } else if ((req & VM_ALLOC_IFCACHED) != 0) { - mtx_unlock(&vm_page_queue_free_mtx); + if (!freelock_owned) + mtx_unlock(&vm_page_queue_free_mtx); return (NULL); #if VM_NRESERVLEVEL > 0 } else if (object == NULL || (object->flags & (OBJ_COLORED | @@ -1271,7 +1396,8 @@ vm_page_alloc(vm_object_t object, vm_pindex_t pind /* * Not allocatable, give up. */ - mtx_unlock(&vm_page_queue_free_mtx); + if (!freelock_owned) + mtx_unlock(&vm_page_queue_free_mtx); atomic_add_int(&vm_pageout_deficit, max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); pagedaemon_wakeup(); @@ -1326,7 +1452,8 @@ vm_page_alloc(vm_object_t object, vm_pindex_t pind if (req & VM_ALLOC_NODUMP) flags |= PG_NODUMP; m->flags = flags; - mtx_unlock(&vm_page_queue_free_mtx); + if (!freelock_owned) + mtx_unlock(&vm_page_queue_free_mtx); m->aflags = 0; m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; @@ -1343,11 +1470,23 @@ vm_page_alloc(vm_object_t object, vm_pindex_t pind m->act_count = 0; if (object != NULL) { + if (vm_page_insert_after(m, object, pindex, mpred)) { + /* See the comment below about hold count. */ + if (vp != NULL) + vdrop(vp); + if (!freelock_owned) + mtx_lock(&vm_page_queue_free_mtx); + vm_page_alloc_turn_free(m); + if (!freelock_owned) + mtx_unlock(&vm_page_queue_free_mtx); + pagedaemon_wakeup(); + 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); - vm_page_insert_after(m, object, pindex, mpred); } else m->pindex = pindex; @@ -1413,7 +1552,7 @@ vm_page_alloc_contig(vm_object_t object, vm_pindex vm_paddr_t boundary, vm_memattr_t memattr) { struct vnode *drop; - vm_page_t deferred_vdrop_list, m, m_ret; + vm_page_t deferred_vdrop_list, m, m_tmp, m_ret; u_int flags, oflags; int req_class; @@ -1507,12 +1646,30 @@ retry: m->wire_count = 1; /* Unmanaged pages don't use "act_count". */ m->oflags = oflags; + if (object != NULL) { + if (vm_page_insert(m, object, pindex)) { + for (m_tmp = m_ret; m_tmp < m; m_tmp++) { + vm_page_lock(m_tmp); + vm_page_remove(m_tmp); + vm_page_unlock(m_tmp); + } + while (deferred_vdrop_list != NULL) { + vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev); + deferred_vdrop_list = + deferred_vdrop_list->pageq.tqe_next; + } + mtx_lock(&vm_page_queue_free_mtx); + for (m = m_ret; m < &m_ret[npages]; m++) + vm_page_alloc_turn_free(m); + mtx_unlock(&vm_page_queue_free_mtx); + if (vm_paging_needed()) + pagedaemon_wakeup(); + return (NULL); + } + } else + m->pindex = pindex; if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); - if (object != NULL) - vm_page_insert(m, object, pindex); - else - m->pindex = pindex; pindex++; } while (deferred_vdrop_list != NULL) { @@ -2183,7 +2340,6 @@ vm_page_cache(vm_page_t m) } KASSERT((m->flags & PG_CACHED) == 0, ("vm_page_cache: page %p is already cached", m)); - PCPU_INC(cnt.v_tcached); /* * Remove the page from the paging queues. @@ -2210,10 +2366,17 @@ vm_page_cache(vm_page_t m) */ m->flags &= ~PG_ZERO; mtx_lock(&vm_page_queue_free_mtx); + cache_was_empty = vm_radix_is_empty(&object->cache); + if (vm_radix_insert(&object->cache, m)) { + mtx_unlock(&vm_page_queue_free_mtx); + vm_page_free(m); + if (object->resident_page_count == 0) + vdrop(object->handle); + return; + } m->flags |= PG_CACHED; cnt.v_cache_count++; - cache_was_empty = vm_radix_is_empty(&object->cache); - vm_radix_insert(&object->cache, m); + PCPU_INC(cnt.v_tcached); #if VM_NRESERVLEVEL > 0 if (!vm_reserv_free_page(m)) { #else @@ -2807,7 +2970,13 @@ vm_page_cowfault(vm_page_t m) vm_page_lock(mnew); vm_page_free(mnew); vm_page_unlock(mnew); - vm_page_insert(m, object, pindex); + for (;;) { + if (!vm_page_insert(m, object, pindex)) + break; + VM_OBJECT_WUNLOCK(object); + VM_WAIT; + VM_OBJECT_WLOCK(object); + } } else { /* clear COW & copy page */ if (!so_zerocp_fullpage) pmap_copy_page(m, mnew); Index: sys/vm/vm_page.h =================================================================== --- sys/vm/vm_page.h (revision 252465) +++ sys/vm/vm_page.h (working copy) @@ -388,7 +388,7 @@ void vm_page_dequeue_locked(vm_page_t m); vm_page_t vm_page_find_least(vm_object_t, vm_pindex_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_insert (vm_page_t, vm_object_t, vm_pindex_t); +int vm_page_insert (vm_page_t, vm_object_t, vm_pindex_t); boolean_t vm_page_is_cached(vm_object_t object, vm_pindex_t pindex); vm_page_t vm_page_lookup (vm_object_t, vm_pindex_t); vm_page_t vm_page_next(vm_page_t m); @@ -398,7 +398,7 @@ void vm_page_putfake(vm_page_t m); void vm_page_readahead_finish(vm_page_t m); void vm_page_reference(vm_page_t m); void vm_page_remove (vm_page_t); -void vm_page_rename (vm_page_t, vm_object_t, vm_pindex_t); +int vm_page_rename (vm_page_t, vm_object_t, vm_pindex_t); void vm_page_requeue(vm_page_t m); void vm_page_requeue_locked(vm_page_t m); void vm_page_set_valid_range(vm_page_t m, int base, int size); Index: sys/i386/i386/pmap.c =================================================================== --- sys/i386/i386/pmap.c (revision 252465) +++ sys/i386/i386/pmap.c (working copy) @@ -304,7 +304,7 @@ static boolean_t pmap_enter_pde(pmap_t pmap, vm_of static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte); static void pmap_flush_page(vm_page_t m); -static void pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); +static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); static boolean_t pmap_is_modified_pvh(struct md_page *pvh); static boolean_t pmap_is_referenced_pvh(struct md_page *pvh); @@ -1604,12 +1604,12 @@ pmap_add_delayed_free_list(vm_page_t m, vm_page_t * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. */ -static __inline void +static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); - vm_radix_insert(&pmap->pm_root, mpte); + return (vm_radix_insert(&pmap->pm_root, mpte)); } /* @@ -3364,7 +3364,13 @@ setpte: ("pmap_promote_pde: page table page is out of range")); KASSERT(mpte->pindex == va >> PDRSHIFT, ("pmap_promote_pde: page table page's pindex is wrong")); - pmap_insert_pt_page(pmap, mpte); + if (pmap_insert_pt_page(pmap, mpte)) { + pmap_pde_p_failures++; + CTR2(KTR_PMAP, + "pmap_promote_pde: failure for va %#x in pmap %p", va, + pmap); + return; + } /* * Promote the pv entries. Index: sys/amd64/amd64/pmap.c =================================================================== --- sys/amd64/amd64/pmap.c (revision 252465) +++ sys/amd64/amd64/pmap.c (working copy) @@ -283,7 +283,7 @@ static boolean_t pmap_enter_pde(pmap_t pmap, vm_of static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); -static void pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); +static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); static boolean_t pmap_is_modified_pvh(struct md_page *pvh); static boolean_t pmap_is_referenced_pvh(struct md_page *pvh); static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode); @@ -1525,12 +1525,12 @@ pmap_add_delayed_free_list(vm_page_t m, vm_page_t * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. */ -static __inline void +static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); - vm_radix_insert(&pmap->pm_root, mpte); + return (vm_radix_insert(&pmap->pm_root, mpte)); } /* @@ -3392,7 +3392,13 @@ setpte: ("pmap_promote_pde: page table page is out of range")); KASSERT(mpte->pindex == pmap_pde_pindex(va), ("pmap_promote_pde: page table page's pindex is wrong")); - pmap_insert_pt_page(pmap, mpte); + if (pmap_insert_pt_page(pmap, mpte)) { + atomic_add_long(&pmap_pde_p_failures, 1); + CTR2(KTR_PMAP, + "pmap_promote_pde: failure for va %#lx in pmap %p", va, + pmap); + return; + } /* * Promote the pv entries. Index: sys/kern/subr_uio.c =================================================================== --- sys/kern/subr_uio.c (revision 252465) +++ sys/kern/subr_uio.c (working copy) @@ -56,6 +56,7 @@ __FBSDID("$FreeBSD$"); #include #include #include +#include #include #ifdef SOCKET_SEND_COW #include @@ -122,7 +123,12 @@ retry: if (uobject->backing_object != NULL) pmap_remove(map->pmap, uaddr, uaddr + PAGE_SIZE); } - vm_page_insert(kern_pg, uobject, upindex); + if (vm_page_insert(kern_pg, uobject, upindex)) { + VM_OBJECT_WUNLOCK(uobject); + VM_WAIT; + VM_OBJECT_WLOCK(uobject); + goto retry; + } vm_page_dirty(kern_pg); VM_OBJECT_WUNLOCK(uobject); vm_map_lookup_done(map, entry); Index: sys/dev/drm2/i915/i915_gem.c =================================================================== --- sys/dev/drm2/i915/i915_gem.c (revision 252465) +++ sys/dev/drm2/i915/i915_gem.c (working copy) @@ -64,6 +64,9 @@ __FBSDID("$FreeBSD$"); #include #include +#include +#include + static void i915_gem_object_flush_cpu_write_domain( struct drm_i915_gem_object *obj); static uint32_t i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, @@ -1440,8 +1443,14 @@ unlocked_vmobj: vm_page_sleep(m, "915pbs"); goto retry; } + if (vm_page_insert(m, vm_obj, OFF_TO_IDX(offset))) { + DRM_UNLOCK(dev); + VM_OBJECT_WUNLOCK(vm_obj); + VM_WAIT; + VM_OBJECT_WLOCK(vm_obj); + goto retry; + } m->valid = VM_PAGE_BITS_ALL; - vm_page_insert(m, vm_obj, OFF_TO_IDX(offset)); have_page: *mres = m; vm_page_busy(m); Index: sys/dev/drm2/ttm/ttm_bo_vm.c =================================================================== --- sys/dev/drm2/ttm/ttm_bo_vm.c (revision 252465) +++ sys/dev/drm2/ttm/ttm_bo_vm.c (working copy) @@ -47,6 +47,7 @@ __FBSDID("$FreeBSD$"); #include #include +#include #define TTM_BO_VM_NUM_PREFAULT 16 @@ -218,11 +219,16 @@ reserve: ttm_bo_unreserve(bo); goto retry; } + if (vm_page_insert(m, vm_obj, OFF_TO_IDX(offset))) { + VM_OBJECT_WUNLOCK(vm_obj); + VM_WAIT; + VM_OBJECT_WLOCK(vm_obj); + ttm_mem_io_unlock(man); + ttm_bo_unreserve(bo); + goto retry; + } m->valid = VM_PAGE_BITS_ALL; *mres = m; - vm_page_lock(m); - vm_page_insert(m, vm_obj, OFF_TO_IDX(offset)); - vm_page_unlock(m); vm_page_busy(m); if (oldm != NULL) {