From db151ca0c343531256b8839f938a4ecbd8b4fd7e Mon Sep 17 00:00:00 2001 From: Doug Moore Date: Sat, 29 Oct 2022 00:50:44 -0500 Subject: [PATCH] iommu_gas: start space search from 1st free space Maintain a pointer to an element in the domain map that is left of any sufficiently large free gap in the tree and start the search for free space there, rather than at the root of the tree. On find_space, move that pointer to the leftmost leaf in the subtree of nodes with free_down greater than or equal to the minimum allocation size before starting the search for space from that pointer. On removal of a node with address less than that pointer, update that pointer to point to the predecessor or successor of the removed node. In experiments with netperf streaming, this reduces by about 40% the number of map entries examined in first-fit allocation. Reviewed by: alc, kib MFC after: 3 weeks Differential Revision: https://reviews.freebsd.org/D36624 (cherry picked from commit 8b221ca631c2d28584a4b6207dc107d8ff5a1b2f) --- sys/dev/iommu/iommu.h | 1 + sys/dev/iommu/iommu_gas.c | 69 ++++++++++++++++++++++++++++----------- 2 files changed, 51 insertions(+), 19 deletions(-) diff --git a/sys/dev/iommu/iommu.h b/sys/dev/iommu/iommu.h index 2531ef09d9cd..cee59a0b55c0 100644 --- a/sys/dev/iommu/iommu.h +++ b/sys/dev/iommu/iommu.h @@ -109,6 +109,7 @@ struct iommu_domain { struct iommu_map_entries_tailq unload_entries; /* (d) Entries to unload */ struct iommu_gas_entries_tree rb_root; /* (d) */ + struct iommu_map_entry *start_gap; /* (d) */ iommu_gaddr_t end; /* (c) Highest address + 1 in the guest AS */ struct iommu_map_entry *first_place, *last_place; /* (d) */ diff --git a/sys/dev/iommu/iommu_gas.c b/sys/dev/iommu/iommu_gas.c index 8c8d0e49c378..ffb5370b507e 100644 --- a/sys/dev/iommu/iommu_gas.c +++ b/sys/dev/iommu/iommu_gas.c @@ -209,7 +209,18 @@ iommu_gas_check_free(struct iommu_domain *domain) static void iommu_gas_rb_remove(struct iommu_domain *domain, struct iommu_map_entry *entry) { + struct iommu_map_entry *nbr; + /* Removing entry may open a new free gap before domain->start_gap. */ + if (entry->end <= domain->start_gap->end) { + if (RB_RIGHT(entry, rb_entry) != NULL) + nbr = iommu_gas_entries_tree_RB_NEXT(entry); + else if (RB_LEFT(entry, rb_entry) != NULL) + nbr = RB_LEFT(entry, rb_entry); + else + nbr = RB_PARENT(entry, rb_entry); + domain->start_gap = nbr; + } RB_REMOVE(iommu_gas_entries_tree, &domain->rb_root, entry); } @@ -253,6 +264,7 @@ iommu_gas_init_domain(struct iommu_domain *domain) begin->flags = IOMMU_MAP_ENTRY_PLACE | IOMMU_MAP_ENTRY_UNMAPPED; RB_INSERT_PREV(iommu_gas_entries_tree, &domain->rb_root, end, begin); + domain->start_gap = end; domain->first_place = begin; domain->last_place = end; domain->flags |= IOMMU_DOMAIN_GAS_INITED; @@ -296,7 +308,6 @@ iommu_gas_fini_domain(struct iommu_domain *domain) } struct iommu_gas_match_args { - struct iommu_domain *domain; iommu_gaddr_t size; int offset; const struct bus_dma_tag_common *common; @@ -395,16 +406,44 @@ iommu_gas_next(struct iommu_map_entry *curr, iommu_gaddr_t min_free) return (curr); } +/* + * Address-ordered first-fit search of 'domain' for free space satisfying the + * conditions of 'a'. The space allocated is at least one page big, and is + * bounded by guard pages to left and right. The allocated space for 'domain' + * is described by an rb-tree of map entries at domain->rb_root, and + * domain->start_gap points a map entry less than or adjacent to the first + * free-space of size at least 3 pages. + */ static int -iommu_gas_find_space(struct iommu_gas_match_args *a) +iommu_gas_find_space(struct iommu_domain *domain, + struct iommu_gas_match_args *a) { - struct iommu_domain *domain; struct iommu_map_entry *curr, *first; iommu_gaddr_t addr, min_free; - IOMMU_DOMAIN_ASSERT_LOCKED(a->domain); + IOMMU_DOMAIN_ASSERT_LOCKED(domain); KASSERT(a->entry->flags == 0, - ("dirty entry %p %p", a->domain, a->entry)); + ("dirty entry %p %p", domain, a->entry)); + + /* + * start_gap may point to an entry adjacent to gaps too small for any + * new allocation. In that case, advance start_gap to the first free + * space big enough for a minimum allocation plus two guard pages. + */ + min_free = 3 * IOMMU_PAGE_SIZE; + first = domain->start_gap; + while (first != NULL && first->free_down < min_free) + first = RB_PARENT(first, rb_entry); + for (curr = first; curr != NULL; + curr = iommu_gas_next(curr, min_free)) { + if ((first = RB_LEFT(curr, rb_entry)) != NULL && + first->last + min_free <= curr->start) + break; + if ((first = RB_RIGHT(curr, rb_entry)) != NULL && + curr->end + min_free <= first->first) + break; + } + domain->start_gap = curr; /* * If the subtree doesn't have free space for the requested allocation @@ -413,20 +452,13 @@ iommu_gas_find_space(struct iommu_gas_match_args *a) min_free = 2 * IOMMU_PAGE_SIZE + roundup2(a->size + a->offset, IOMMU_PAGE_SIZE); - /* - * Find the first entry in the lower region that could abut a big-enough - * range. - */ - domain = a->domain; - curr = RB_ROOT(&domain->rb_root); - first = NULL; - while (curr != NULL && curr->free_down >= min_free) { - first = curr; - curr = RB_LEFT(curr, rb_entry); - } + /* Climb to find a node in the subtree of big-enough ranges. */ + first = curr; + while (first != NULL && first->free_down < min_free) + first = RB_PARENT(first, rb_entry); /* - * Walk the big-enough ranges until one satisfies alignment + * Walk the big-enough ranges tree until one satisfies alignment * requirements, or violates lowaddr address requirement. */ addr = a->common->lowaddr + 1; @@ -746,7 +778,6 @@ iommu_gas_map(struct iommu_domain *domain, KASSERT((flags & ~(IOMMU_MF_CANWAIT | IOMMU_MF_CANSPLIT)) == 0, ("invalid flags 0x%x", flags)); - a.domain = domain; a.size = size; a.offset = offset; a.common = common; @@ -757,7 +788,7 @@ iommu_gas_map(struct iommu_domain *domain, return (ENOMEM); a.entry = entry; IOMMU_DOMAIN_LOCK(domain); - error = iommu_gas_find_space(&a); + error = iommu_gas_find_space(domain, &a); if (error == ENOMEM) { IOMMU_DOMAIN_UNLOCK(domain); iommu_gas_free_entry(entry);