4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * Big Theory Statement for the virtual memory allocator.
29 * For a more complete description of the main ideas, see:
31 * Jeff Bonwick and Jonathan Adams,
33 * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
34 * Arbitrary Resources.
36 * Proceedings of the 2001 Usenix Conference.
37 * Available as http://www.usenix.org/event/usenix01/bonwick.html
45 * We divide the kernel address space into a number of logically distinct
46 * pieces, or *arenas*: text, data, heap, stack, and so on. Within these
47 * arenas we often subdivide further; for example, we use heap addresses
48 * not only for the kernel heap (kmem_alloc() space), but also for DVMA,
49 * bp_mapin(), /dev/kmem, and even some device mappings like the TOD chip.
50 * The kernel address space, therefore, is most accurately described as
51 * a tree of arenas in which each node of the tree *imports* some subset
52 * of its parent. The virtual memory allocator manages these arenas and
53 * supports their natural hierarchical structure.
57 * An arena is nothing more than a set of integers. These integers most
58 * commonly represent virtual addresses, but in fact they can represent
59 * anything at all. For example, we could use an arena containing the
60 * integers minpid through maxpid to allocate process IDs. vmem_create()
61 * and vmem_destroy() create and destroy vmem arenas. In order to
62 * differentiate between arenas used for adresses and arenas used for
63 * identifiers, the VMC_IDENTIFIER flag is passed to vmem_create(). This
64 * prevents identifier exhaustion from being diagnosed as general memory
69 * We represent the integers in an arena as a collection of *spans*, or
70 * contiguous ranges of integers. For example, the kernel heap consists
71 * of just one span: [kernelheap, ekernelheap). Spans can be added to an
72 * arena in two ways: explicitly, by vmem_add(), or implicitly, by
73 * importing, as described in Section 1.5 below.
77 * Spans are subdivided into *segments*, each of which is either allocated
78 * or free. A segment, like a span, is a contiguous range of integers.
79 * Each allocated segment [addr, addr + size) represents exactly one
80 * vmem_alloc(size) that returned addr. Free segments represent the space
81 * between allocated segments. If two free segments are adjacent, we
82 * coalesce them into one larger segment; that is, if segments [a, b) and
83 * [b, c) are both free, we merge them into a single segment [a, c).
84 * The segments within a span are linked together in increasing-address order
85 * so we can easily determine whether coalescing is possible.
87 * Segments never cross span boundaries. When all segments within
88 * an imported span become free, we return the span to its source.
92 * As mentioned in the overview, some arenas are logical subsets of
93 * other arenas. For example, kmem_va_arena (a virtual address cache
94 * that satisfies most kmem_slab_create() requests) is just a subset
95 * of heap_arena (the kernel heap) that provides caching for the most
96 * common slab sizes. When kmem_va_arena runs out of virtual memory,
97 * it *imports* more from the heap; we say that heap_arena is the
98 * *vmem source* for kmem_va_arena. vmem_create() allows you to
99 * specify any existing vmem arena as the source for your new arena.
100 * Topologically, since every arena is a child of at most one source,
101 * the set of all arenas forms a collection of trees.
103 * 1.6 Constrained Allocations
104 * ---------------------------
105 * Some vmem clients are quite picky about the kind of address they want.
106 * For example, the DVMA code may need an address that is at a particular
107 * phase with respect to some alignment (to get good cache coloring), or
108 * that lies within certain limits (the addressable range of a device),
109 * or that doesn't cross some boundary (a DMA counter restriction) --
110 * or all of the above. vmem_xalloc() allows the client to specify any
111 * or all of these constraints.
113 * 1.7 The Vmem Quantum
114 * --------------------
115 * Every arena has a notion of 'quantum', specified at vmem_create() time,
116 * that defines the arena's minimum unit of currency. Most commonly the
117 * quantum is either 1 or PAGESIZE, but any power of 2 is legal.
118 * All vmem allocations are guaranteed to be quantum-aligned.
120 * 1.8 Quantum Caching
121 * -------------------
122 * A vmem arena may be so hot (frequently used) that the scalability of vmem
123 * allocation is a significant concern. We address this by allowing the most
124 * common allocation sizes to be serviced by the kernel memory allocator,
125 * which provides low-latency per-cpu caching. The qcache_max argument to
126 * vmem_create() specifies the largest allocation size to cache.
128 * 1.9 Relationship to Kernel Memory Allocator
129 * -------------------------------------------
130 * Every kmem cache has a vmem arena as its slab supplier. The kernel memory
131 * allocator uses vmem_alloc() and vmem_free() to create and destroy slabs.
137 * 2.1 Segment lists and markers
138 * -----------------------------
139 * The segment structure (vmem_seg_t) contains two doubly-linked lists.
141 * The arena list (vs_anext/vs_aprev) links all segments in the arena.
142 * In addition to the allocated and free segments, the arena contains
143 * special marker segments at span boundaries. Span markers simplify
144 * coalescing and importing logic by making it easy to tell both when
145 * we're at a span boundary (so we don't coalesce across it), and when
146 * a span is completely free (its neighbors will both be span markers).
148 * Imported spans will have vs_import set.
150 * The next-of-kin list (vs_knext/vs_kprev) links segments of the same type:
151 * (1) for allocated segments, vs_knext is the hash chain linkage;
152 * (2) for free segments, vs_knext is the freelist linkage;
153 * (3) for span marker segments, vs_knext is the next span marker.
155 * 2.2 Allocation hashing
156 * ----------------------
157 * We maintain a hash table of all allocated segments, hashed by address.
158 * This allows vmem_free() to discover the target segment in constant time.
159 * vmem_update() periodically resizes hash tables to keep hash chains short.
161 * 2.3 Freelist management
162 * -----------------------
163 * We maintain power-of-2 freelists for free segments, i.e. free segments
164 * of size >= 2^n reside in vmp->vm_freelist[n]. To ensure constant-time
165 * allocation, vmem_xalloc() looks not in the first freelist that *might*
166 * satisfy the allocation, but in the first freelist that *definitely*
167 * satisfies the allocation (unless VM_BESTFIT is specified, or all larger
168 * freelists are empty). For example, a 1000-byte allocation will be
169 * satisfied not from the 512..1023-byte freelist, whose members *might*
170 * contains a 1000-byte segment, but from a 1024-byte or larger freelist,
171 * the first member of which will *definitely* satisfy the allocation.
172 * This ensures that vmem_xalloc() works in constant time.
174 * We maintain a bit map to determine quickly which freelists are non-empty.
175 * vmp->vm_freemap & (1 << n) is non-zero iff vmp->vm_freelist[n] is non-empty.
177 * The different freelists are linked together into one large freelist,
178 * with the freelist heads serving as markers. Freelist markers simplify
179 * the maintenance of vm_freemap by making it easy to tell when we're taking
180 * the last member of a freelist (both of its neighbors will be markers).
184 * For simplicity, all arena state is protected by a per-arena lock.
185 * For very hot arenas, use quantum caching for scalability.
187 * 2.5 Vmem Population
188 * -------------------
189 * Any internal vmem routine that might need to allocate new segment
190 * structures must prepare in advance by calling vmem_populate(), which
191 * will preallocate enough vmem_seg_t's to get is through the entire
192 * operation without dropping the arena lock.
196 * If KMF_AUDIT is set in kmem_flags, we audit vmem allocations as well.
197 * Since virtual addresses cannot be scribbled on, there is no equivalent
198 * in vmem to redzone checking, deadbeef, or other kmem debugging features.
199 * Moreover, we do not audit frees because segment coalescing destroys the
200 * association between an address and its segment structure. Auditing is
201 * thus intended primarily to keep track of who's consuming the arena.
202 * Debugging support could certainly be extended in the future if it proves
203 * necessary, but we do so much live checking via the allocation hash table
204 * that even non-DEBUG systems get quite a bit of sanity checking already.
207 #include <sys/vmem_impl.h>
208 #include <sys/kmem.h>
209 #include <sys/kstat.h>
210 #include <sys/param.h>
211 #include <sys/systm.h>
212 #include <sys/atomic.h>
213 #include <sys/bitmap.h>
214 #include <sys/sysmacros.h>
215 #include <sys/cmn_err.h>
216 #include <sys/debug.h>
217 #include <sys/panic.h>
219 #define VMEM_INITIAL 10 /* early vmem arenas */
220 #define VMEM_SEG_INITIAL 200 /* early segments */
223 * Adding a new span to an arena requires two segment structures: one to
224 * represent the span, and one to represent the free segment it contains.
226 #define VMEM_SEGS_PER_SPAN_CREATE 2
229 * Allocating a piece of an existing segment requires 0-2 segment structures
230 * depending on how much of the segment we're allocating.
232 * To allocate the entire segment, no new segment structures are needed; we
233 * simply move the existing segment structure from the freelist to the
234 * allocation hash table.
236 * To allocate a piece from the left or right end of the segment, we must
237 * split the segment into two pieces (allocated part and remainder), so we
238 * need one new segment structure to represent the remainder.
240 * To allocate from the middle of a segment, we need two new segment strucures
241 * to represent the remainders on either side of the allocated part.
243 #define VMEM_SEGS_PER_EXACT_ALLOC 0
244 #define VMEM_SEGS_PER_LEFT_ALLOC 1
245 #define VMEM_SEGS_PER_RIGHT_ALLOC 1
246 #define VMEM_SEGS_PER_MIDDLE_ALLOC 2
249 * vmem_populate() preallocates segment structures for vmem to do its work.
250 * It must preallocate enough for the worst case, which is when we must import
251 * a new span and then allocate from the middle of it.
253 #define VMEM_SEGS_PER_ALLOC_MAX \
254 (VMEM_SEGS_PER_SPAN_CREATE + VMEM_SEGS_PER_MIDDLE_ALLOC)
257 * The segment structures themselves are allocated from vmem_seg_arena, so
258 * we have a recursion problem when vmem_seg_arena needs to populate itself.
259 * We address this by working out the maximum number of segment structures
260 * this act will require, and multiplying by the maximum number of threads
261 * that we'll allow to do it simultaneously.
263 * The worst-case segment consumption to populate vmem_seg_arena is as
264 * follows (depicted as a stack trace to indicate why events are occurring):
266 * (In order to lower the fragmentation in the heap_arena, we specify a
267 * minimum import size for the vmem_metadata_arena which is the same size
268 * as the kmem_va quantum cache allocations. This causes the worst-case
269 * allocation from the vmem_metadata_arena to be 3 segments.)
271 * vmem_alloc(vmem_seg_arena) -> 2 segs (span create + exact alloc)
272 * segkmem_alloc(vmem_metadata_arena)
273 * vmem_alloc(vmem_metadata_arena) -> 3 segs (span create + left alloc)
274 * vmem_alloc(heap_arena) -> 1 seg (left alloc)
279 * vmem_alloc(hat_memload_arena) -> 2 segs (span create + exact alloc)
280 * segkmem_alloc(heap_arena)
281 * vmem_alloc(heap_arena) -> 1 seg (left alloc)
283 * hat_memload() -> (hat layer won't recurse further)
285 * The worst-case consumption for each arena is 3 segment structures.
286 * Of course, a 3-seg reserve could easily be blown by multiple threads.
287 * Therefore, we serialize all allocations from vmem_seg_arena (which is OK
288 * because they're rare). We cannot allow a non-blocking allocation to get
289 * tied up behind a blocking allocation, however, so we use separate locks
290 * for VM_SLEEP and VM_NOSLEEP allocations. Similarly, VM_PUSHPAGE allocations
291 * must not block behind ordinary VM_SLEEPs. In addition, if the system is
292 * panicking then we must keep enough resources for panic_thread to do its
293 * work. Thus we have at most four threads trying to allocate from
294 * vmem_seg_arena, and each thread consumes at most three segment structures,
295 * so we must maintain a 12-seg reserve.
297 #define VMEM_POPULATE_RESERVE 12
300 * vmem_populate() ensures that each arena has VMEM_MINFREE seg structures
301 * so that it can satisfy the worst-case allocation *and* participate in
302 * worst-case allocation from vmem_seg_arena.
304 #define VMEM_MINFREE (VMEM_POPULATE_RESERVE + VMEM_SEGS_PER_ALLOC_MAX)
306 static vmem_t vmem0
[VMEM_INITIAL
];
307 static vmem_t
*vmem_populator
[VMEM_INITIAL
];
308 static uint32_t vmem_id
;
309 static uint32_t vmem_populators
;
310 static vmem_seg_t vmem_seg0
[VMEM_SEG_INITIAL
];
311 static vmem_seg_t
*vmem_segfree
;
312 static kmutex_t vmem_list_lock
;
313 static kmutex_t vmem_segfree_lock
;
314 static kmutex_t vmem_sleep_lock
;
315 static kmutex_t vmem_nosleep_lock
;
316 static kmutex_t vmem_pushpage_lock
;
317 static kmutex_t vmem_panic_lock
;
318 static vmem_t
*vmem_list
;
319 static vmem_t
*vmem_metadata_arena
;
320 static vmem_t
*vmem_seg_arena
;
321 static vmem_t
*vmem_hash_arena
;
322 static vmem_t
*vmem_vmem_arena
;
323 static long vmem_update_interval
= 15; /* vmem_update() every 15 seconds */
324 uint32_t vmem_mtbf
; /* mean time between failures [default: off] */
325 size_t vmem_seg_size
= sizeof (vmem_seg_t
);
327 static vmem_kstat_t vmem_kstat_template
= {
328 { "mem_inuse", KSTAT_DATA_UINT64
},
329 { "mem_import", KSTAT_DATA_UINT64
},
330 { "mem_total", KSTAT_DATA_UINT64
},
331 { "vmem_source", KSTAT_DATA_UINT32
},
332 { "alloc", KSTAT_DATA_UINT64
},
333 { "free", KSTAT_DATA_UINT64
},
334 { "wait", KSTAT_DATA_UINT64
},
335 { "fail", KSTAT_DATA_UINT64
},
336 { "lookup", KSTAT_DATA_UINT64
},
337 { "search", KSTAT_DATA_UINT64
},
338 { "populate_wait", KSTAT_DATA_UINT64
},
339 { "populate_fail", KSTAT_DATA_UINT64
},
340 { "contains", KSTAT_DATA_UINT64
},
341 { "contains_search", KSTAT_DATA_UINT64
},
345 * Insert/delete from arena list (type 'a') or next-of-kin list (type 'k').
347 #define VMEM_INSERT(vprev, vsp, type) \
349 vmem_seg_t *vnext = (vprev)->vs_##type##next; \
350 (vsp)->vs_##type##next = (vnext); \
351 (vsp)->vs_##type##prev = (vprev); \
352 (vprev)->vs_##type##next = (vsp); \
353 (vnext)->vs_##type##prev = (vsp); \
356 #define VMEM_DELETE(vsp, type) \
358 vmem_seg_t *vprev = (vsp)->vs_##type##prev; \
359 vmem_seg_t *vnext = (vsp)->vs_##type##next; \
360 (vprev)->vs_##type##next = (vnext); \
361 (vnext)->vs_##type##prev = (vprev); \
365 * Get a vmem_seg_t from the global segfree list.
368 vmem_getseg_global(void)
372 mutex_enter(&vmem_segfree_lock
);
373 if ((vsp
= vmem_segfree
) != NULL
)
374 vmem_segfree
= vsp
->vs_knext
;
375 mutex_exit(&vmem_segfree_lock
);
381 * Put a vmem_seg_t on the global segfree list.
384 vmem_putseg_global(vmem_seg_t
*vsp
)
386 mutex_enter(&vmem_segfree_lock
);
387 vsp
->vs_knext
= vmem_segfree
;
389 mutex_exit(&vmem_segfree_lock
);
393 * Get a vmem_seg_t from vmp's segfree list.
396 vmem_getseg(vmem_t
*vmp
)
400 ASSERT(vmp
->vm_nsegfree
> 0);
402 vsp
= vmp
->vm_segfree
;
403 vmp
->vm_segfree
= vsp
->vs_knext
;
410 * Put a vmem_seg_t on vmp's segfree list.
413 vmem_putseg(vmem_t
*vmp
, vmem_seg_t
*vsp
)
415 vsp
->vs_knext
= vmp
->vm_segfree
;
416 vmp
->vm_segfree
= vsp
;
421 * Add vsp to the appropriate freelist.
424 vmem_freelist_insert(vmem_t
*vmp
, vmem_seg_t
*vsp
)
428 ASSERT(*VMEM_HASH(vmp
, vsp
->vs_start
) != vsp
);
430 vprev
= (vmem_seg_t
*)&vmp
->vm_freelist
[highbit(VS_SIZE(vsp
)) - 1];
431 vsp
->vs_type
= VMEM_FREE
;
432 vmp
->vm_freemap
|= VS_SIZE(vprev
);
433 VMEM_INSERT(vprev
, vsp
, k
);
435 cv_broadcast(&vmp
->vm_cv
);
439 * Take vsp from the freelist.
442 vmem_freelist_delete(vmem_t
*vmp
, vmem_seg_t
*vsp
)
444 ASSERT(*VMEM_HASH(vmp
, vsp
->vs_start
) != vsp
);
445 ASSERT(vsp
->vs_type
== VMEM_FREE
);
447 if (vsp
->vs_knext
->vs_start
== 0 && vsp
->vs_kprev
->vs_start
== 0) {
449 * The segments on both sides of 'vsp' are freelist heads,
450 * so taking vsp leaves the freelist at vsp->vs_kprev empty.
452 ASSERT(vmp
->vm_freemap
& VS_SIZE(vsp
->vs_kprev
));
453 vmp
->vm_freemap
^= VS_SIZE(vsp
->vs_kprev
);
459 * Add vsp to the allocated-segment hash table and update kstats.
462 vmem_hash_insert(vmem_t
*vmp
, vmem_seg_t
*vsp
)
466 vsp
->vs_type
= VMEM_ALLOC
;
467 bucket
= VMEM_HASH(vmp
, vsp
->vs_start
);
468 vsp
->vs_knext
= *bucket
;
471 if (vmem_seg_size
== sizeof (vmem_seg_t
)) {
472 vsp
->vs_depth
= (uint8_t)getpcstack(vsp
->vs_stack
,
474 vsp
->vs_thread
= curthread
;
475 vsp
->vs_timestamp
= gethrtime();
480 vmp
->vm_kstat
.vk_alloc
.value
.ui64
++;
481 vmp
->vm_kstat
.vk_mem_inuse
.value
.ui64
+= VS_SIZE(vsp
);
485 * Remove vsp from the allocated-segment hash table and update kstats.
488 vmem_hash_delete(vmem_t
*vmp
, uintptr_t addr
, size_t size
)
490 vmem_seg_t
*vsp
, **prev_vspp
;
492 prev_vspp
= VMEM_HASH(vmp
, addr
);
493 while ((vsp
= *prev_vspp
) != NULL
) {
494 if (vsp
->vs_start
== addr
) {
495 *prev_vspp
= vsp
->vs_knext
;
498 vmp
->vm_kstat
.vk_lookup
.value
.ui64
++;
499 prev_vspp
= &vsp
->vs_knext
;
503 panic("vmem_hash_delete(%p, %lx, %lu): bad free",
504 (void *)vmp
, addr
, size
);
505 if (VS_SIZE(vsp
) != size
)
506 panic("vmem_hash_delete(%p, %lx, %lu): wrong size (expect %lu)",
507 (void *)vmp
, addr
, size
, VS_SIZE(vsp
));
509 vmp
->vm_kstat
.vk_free
.value
.ui64
++;
510 vmp
->vm_kstat
.vk_mem_inuse
.value
.ui64
-= size
;
516 * Create a segment spanning the range [start, end) and add it to the arena.
519 vmem_seg_create(vmem_t
*vmp
, vmem_seg_t
*vprev
, uintptr_t start
, uintptr_t end
)
521 vmem_seg_t
*newseg
= vmem_getseg(vmp
);
523 newseg
->vs_start
= start
;
524 newseg
->vs_end
= end
;
526 newseg
->vs_import
= 0;
528 VMEM_INSERT(vprev
, newseg
, a
);
534 * Remove segment vsp from the arena.
537 vmem_seg_destroy(vmem_t
*vmp
, vmem_seg_t
*vsp
)
539 ASSERT(vsp
->vs_type
!= VMEM_ROTOR
);
542 vmem_putseg(vmp
, vsp
);
546 * Add the span [vaddr, vaddr + size) to vmp and update kstats.
549 vmem_span_create(vmem_t
*vmp
, void *vaddr
, size_t size
, uint8_t import
)
551 vmem_seg_t
*newseg
, *span
;
552 uintptr_t start
= (uintptr_t)vaddr
;
553 uintptr_t end
= start
+ size
;
555 ASSERT(MUTEX_HELD(&vmp
->vm_lock
));
557 if ((start
| end
) & (vmp
->vm_quantum
- 1))
558 panic("vmem_span_create(%p, %p, %lu): misaligned",
559 (void *)vmp
, vaddr
, size
);
561 span
= vmem_seg_create(vmp
, vmp
->vm_seg0
.vs_aprev
, start
, end
);
562 span
->vs_type
= VMEM_SPAN
;
563 span
->vs_import
= import
;
564 VMEM_INSERT(vmp
->vm_seg0
.vs_kprev
, span
, k
);
566 newseg
= vmem_seg_create(vmp
, span
, start
, end
);
567 vmem_freelist_insert(vmp
, newseg
);
570 vmp
->vm_kstat
.vk_mem_import
.value
.ui64
+= size
;
571 vmp
->vm_kstat
.vk_mem_total
.value
.ui64
+= size
;
577 * Remove span vsp from vmp and update kstats.
580 vmem_span_destroy(vmem_t
*vmp
, vmem_seg_t
*vsp
)
582 vmem_seg_t
*span
= vsp
->vs_aprev
;
583 size_t size
= VS_SIZE(vsp
);
585 ASSERT(MUTEX_HELD(&vmp
->vm_lock
));
586 ASSERT(span
->vs_type
== VMEM_SPAN
);
589 vmp
->vm_kstat
.vk_mem_import
.value
.ui64
-= size
;
590 vmp
->vm_kstat
.vk_mem_total
.value
.ui64
-= size
;
592 VMEM_DELETE(span
, k
);
594 vmem_seg_destroy(vmp
, vsp
);
595 vmem_seg_destroy(vmp
, span
);
599 * Allocate the subrange [addr, addr + size) from segment vsp.
600 * If there are leftovers on either side, place them on the freelist.
601 * Returns a pointer to the segment representing [addr, addr + size).
604 vmem_seg_alloc(vmem_t
*vmp
, vmem_seg_t
*vsp
, uintptr_t addr
, size_t size
)
606 uintptr_t vs_start
= vsp
->vs_start
;
607 uintptr_t vs_end
= vsp
->vs_end
;
608 size_t vs_size
= vs_end
- vs_start
;
609 size_t realsize
= P2ROUNDUP(size
, vmp
->vm_quantum
);
610 uintptr_t addr_end
= addr
+ realsize
;
612 ASSERT(P2PHASE(vs_start
, vmp
->vm_quantum
) == 0);
613 ASSERT(P2PHASE(addr
, vmp
->vm_quantum
) == 0);
614 ASSERT(vsp
->vs_type
== VMEM_FREE
);
615 ASSERT(addr
>= vs_start
&& addr_end
- 1 <= vs_end
- 1);
616 ASSERT(addr
- 1 <= addr_end
- 1);
619 * If we're allocating from the start of the segment, and the
620 * remainder will be on the same freelist, we can save quite
623 if (P2SAMEHIGHBIT(vs_size
, vs_size
- realsize
) && addr
== vs_start
) {
624 ASSERT(highbit(vs_size
) == highbit(vs_size
- realsize
));
625 vsp
->vs_start
= addr_end
;
626 vsp
= vmem_seg_create(vmp
, vsp
->vs_aprev
, addr
, addr
+ size
);
627 vmem_hash_insert(vmp
, vsp
);
631 vmem_freelist_delete(vmp
, vsp
);
633 if (vs_end
!= addr_end
)
634 vmem_freelist_insert(vmp
,
635 vmem_seg_create(vmp
, vsp
, addr_end
, vs_end
));
637 if (vs_start
!= addr
)
638 vmem_freelist_insert(vmp
,
639 vmem_seg_create(vmp
, vsp
->vs_aprev
, vs_start
, addr
));
641 vsp
->vs_start
= addr
;
642 vsp
->vs_end
= addr
+ size
;
644 vmem_hash_insert(vmp
, vsp
);
649 * Returns 1 if we are populating, 0 otherwise.
650 * Call it if we want to prevent recursion from HAT.
655 return (mutex_owner(&vmem_sleep_lock
) == curthread
||
656 mutex_owner(&vmem_nosleep_lock
) == curthread
||
657 mutex_owner(&vmem_pushpage_lock
) == curthread
||
658 mutex_owner(&vmem_panic_lock
) == curthread
);
662 * Populate vmp's segfree list with VMEM_MINFREE vmem_seg_t structures.
665 vmem_populate(vmem_t
*vmp
, int vmflag
)
674 while (vmp
->vm_nsegfree
< VMEM_MINFREE
&&
675 (vsp
= vmem_getseg_global()) != NULL
)
676 vmem_putseg(vmp
, vsp
);
678 if (vmp
->vm_nsegfree
>= VMEM_MINFREE
)
682 * If we're already populating, tap the reserve.
684 if (vmem_is_populator()) {
685 ASSERT(vmp
->vm_cflags
& VMC_POPULATOR
);
689 mutex_exit(&vmp
->vm_lock
);
691 if (panic_thread
== curthread
)
692 lp
= &vmem_panic_lock
;
693 else if (vmflag
& VM_NOSLEEP
)
694 lp
= &vmem_nosleep_lock
;
695 else if (vmflag
& VM_PUSHPAGE
)
696 lp
= &vmem_pushpage_lock
;
698 lp
= &vmem_sleep_lock
;
702 nseg
= VMEM_MINFREE
+ vmem_populators
* VMEM_POPULATE_RESERVE
;
703 size
= P2ROUNDUP(nseg
* vmem_seg_size
, vmem_seg_arena
->vm_quantum
);
704 nseg
= size
/ vmem_seg_size
;
707 * The following vmem_alloc() may need to populate vmem_seg_arena
708 * and all the things it imports from. When doing so, it will tap
709 * each arena's reserve to prevent recursion (see the block comment
710 * above the definition of VMEM_POPULATE_RESERVE).
712 p
= vmem_alloc(vmem_seg_arena
, size
, vmflag
& VM_KMFLAGS
);
715 mutex_enter(&vmp
->vm_lock
);
716 vmp
->vm_kstat
.vk_populate_fail
.value
.ui64
++;
721 * Restock the arenas that may have been depleted during population.
723 for (i
= 0; i
< vmem_populators
; i
++) {
724 mutex_enter(&vmem_populator
[i
]->vm_lock
);
725 while (vmem_populator
[i
]->vm_nsegfree
< VMEM_POPULATE_RESERVE
)
726 vmem_putseg(vmem_populator
[i
],
727 (vmem_seg_t
*)(p
+ --nseg
* vmem_seg_size
));
728 mutex_exit(&vmem_populator
[i
]->vm_lock
);
732 mutex_enter(&vmp
->vm_lock
);
735 * Now take our own segments.
737 ASSERT(nseg
>= VMEM_MINFREE
);
738 while (vmp
->vm_nsegfree
< VMEM_MINFREE
)
739 vmem_putseg(vmp
, (vmem_seg_t
*)(p
+ --nseg
* vmem_seg_size
));
742 * Give the remainder to charity.
745 vmem_putseg_global((vmem_seg_t
*)(p
+ --nseg
* vmem_seg_size
));
751 * Advance a walker from its previous position to 'afterme'.
752 * Note: may drop and reacquire vmp->vm_lock.
755 vmem_advance(vmem_t
*vmp
, vmem_seg_t
*walker
, vmem_seg_t
*afterme
)
757 vmem_seg_t
*vprev
= walker
->vs_aprev
;
758 vmem_seg_t
*vnext
= walker
->vs_anext
;
759 vmem_seg_t
*vsp
= NULL
;
761 VMEM_DELETE(walker
, a
);
764 VMEM_INSERT(afterme
, walker
, a
);
767 * The walker segment's presence may have prevented its neighbors
768 * from coalescing. If so, coalesce them now.
770 if (vprev
->vs_type
== VMEM_FREE
) {
771 if (vnext
->vs_type
== VMEM_FREE
) {
772 ASSERT(vprev
->vs_end
== vnext
->vs_start
);
773 vmem_freelist_delete(vmp
, vnext
);
774 vmem_freelist_delete(vmp
, vprev
);
775 vprev
->vs_end
= vnext
->vs_end
;
776 vmem_freelist_insert(vmp
, vprev
);
777 vmem_seg_destroy(vmp
, vnext
);
780 } else if (vnext
->vs_type
== VMEM_FREE
) {
785 * vsp could represent a complete imported span,
786 * in which case we must return it to the source.
788 if (vsp
!= NULL
&& vsp
->vs_aprev
->vs_import
&&
789 vmp
->vm_source_free
!= NULL
&&
790 vsp
->vs_aprev
->vs_type
== VMEM_SPAN
&&
791 vsp
->vs_anext
->vs_type
== VMEM_SPAN
) {
792 void *vaddr
= (void *)vsp
->vs_start
;
793 size_t size
= VS_SIZE(vsp
);
794 ASSERT(size
== VS_SIZE(vsp
->vs_aprev
));
795 vmem_freelist_delete(vmp
, vsp
);
796 vmem_span_destroy(vmp
, vsp
);
797 mutex_exit(&vmp
->vm_lock
);
798 vmp
->vm_source_free(vmp
->vm_source
, vaddr
, size
);
799 mutex_enter(&vmp
->vm_lock
);
804 * VM_NEXTFIT allocations deliberately cycle through all virtual addresses
805 * in an arena, so that we avoid reusing addresses for as long as possible.
806 * This helps to catch used-after-freed bugs. It's also the perfect policy
807 * for allocating things like process IDs, where we want to cycle through
808 * all values in order.
811 vmem_nextfit_alloc(vmem_t
*vmp
, size_t size
, int vmflag
)
813 vmem_seg_t
*vsp
, *rotor
;
815 size_t realsize
= P2ROUNDUP(size
, vmp
->vm_quantum
);
818 mutex_enter(&vmp
->vm_lock
);
820 if (vmp
->vm_nsegfree
< VMEM_MINFREE
&& !vmem_populate(vmp
, vmflag
)) {
821 mutex_exit(&vmp
->vm_lock
);
826 * The common case is that the segment right after the rotor is free,
827 * and large enough that extracting 'size' bytes won't change which
828 * freelist it's on. In this case we can avoid a *lot* of work.
829 * Instead of the normal vmem_seg_alloc(), we just advance the start
830 * address of the victim segment. Instead of moving the rotor, we
831 * create the new segment structure *behind the rotor*, which has
832 * the same effect. And finally, we know we don't have to coalesce
833 * the rotor's neighbors because the new segment lies between them.
835 rotor
= &vmp
->vm_rotor
;
836 vsp
= rotor
->vs_anext
;
837 if (vsp
->vs_type
== VMEM_FREE
&& (vs_size
= VS_SIZE(vsp
)) > realsize
&&
838 P2SAMEHIGHBIT(vs_size
, vs_size
- realsize
)) {
839 ASSERT(highbit(vs_size
) == highbit(vs_size
- realsize
));
840 addr
= vsp
->vs_start
;
841 vsp
->vs_start
= addr
+ realsize
;
842 vmem_hash_insert(vmp
,
843 vmem_seg_create(vmp
, rotor
->vs_aprev
, addr
, addr
+ size
));
844 mutex_exit(&vmp
->vm_lock
);
845 return ((void *)addr
);
849 * Starting at the rotor, look for a segment large enough to
850 * satisfy the allocation.
853 vmp
->vm_kstat
.vk_search
.value
.ui64
++;
854 if (vsp
->vs_type
== VMEM_FREE
&& VS_SIZE(vsp
) >= size
)
859 * We've come full circle. One possibility is that the
860 * there's actually enough space, but the rotor itself
861 * is preventing the allocation from succeeding because
862 * it's sitting between two free segments. Therefore,
863 * we advance the rotor and see if that liberates a
866 vmem_advance(vmp
, rotor
, rotor
->vs_anext
);
867 vsp
= rotor
->vs_aprev
;
868 if (vsp
->vs_type
== VMEM_FREE
&& VS_SIZE(vsp
) >= size
)
871 * If there's a lower arena we can import from, or it's
872 * a VM_NOSLEEP allocation, let vmem_xalloc() handle it.
873 * Otherwise, wait until another thread frees something.
875 if (vmp
->vm_source_alloc
!= NULL
||
876 (vmflag
& VM_NOSLEEP
)) {
877 mutex_exit(&vmp
->vm_lock
);
878 return (vmem_xalloc(vmp
, size
, vmp
->vm_quantum
,
879 0, 0, NULL
, NULL
, vmflag
& VM_KMFLAGS
));
881 vmp
->vm_kstat
.vk_wait
.value
.ui64
++;
882 cv_wait(&vmp
->vm_cv
, &vmp
->vm_lock
);
883 vsp
= rotor
->vs_anext
;
888 * We found a segment. Extract enough space to satisfy the allocation.
890 addr
= vsp
->vs_start
;
891 vsp
= vmem_seg_alloc(vmp
, vsp
, addr
, size
);
892 ASSERT(vsp
->vs_type
== VMEM_ALLOC
&&
893 vsp
->vs_start
== addr
&& vsp
->vs_end
== addr
+ size
);
896 * Advance the rotor to right after the newly-allocated segment.
897 * That's where the next VM_NEXTFIT allocation will begin searching.
899 vmem_advance(vmp
, rotor
, vsp
);
900 mutex_exit(&vmp
->vm_lock
);
901 return ((void *)addr
);
905 * Checks if vmp is guaranteed to have a size-byte buffer somewhere on its
906 * freelist. If size is not a power-of-2, it can return a false-negative.
908 * Used to decide if a newly imported span is superfluous after re-acquiring
912 vmem_canalloc(vmem_t
*vmp
, size_t size
)
916 ASSERT(MUTEX_HELD(&vmp
->vm_lock
));
918 if ((size
& (size
- 1)) == 0)
919 flist
= lowbit(P2ALIGN(vmp
->vm_freemap
, size
));
920 else if ((hb
= highbit(size
)) < VMEM_FREELISTS
)
921 flist
= lowbit(P2ALIGN(vmp
->vm_freemap
, 1UL << hb
));
927 * Allocate size bytes at offset phase from an align boundary such that the
928 * resulting segment [addr, addr + size) is a subset of [minaddr, maxaddr)
929 * that does not straddle a nocross-aligned boundary.
932 vmem_xalloc(vmem_t
*vmp
, size_t size
, size_t align_arg
, size_t phase
,
933 size_t nocross
, void *minaddr
, void *maxaddr
, int vmflag
)
936 vmem_seg_t
*vbest
= NULL
;
937 uintptr_t addr
, taddr
, start
, end
;
938 uintptr_t align
= (align_arg
!= 0) ? align_arg
: vmp
->vm_quantum
;
939 void *vaddr
, *xvaddr
= NULL
;
944 if ((align
| phase
| nocross
) & (vmp
->vm_quantum
- 1))
945 panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
946 "parameters not vm_quantum aligned",
947 (void *)vmp
, size
, align_arg
, phase
, nocross
,
948 minaddr
, maxaddr
, vmflag
);
951 (align
> nocross
|| P2ROUNDUP(phase
+ size
, align
) > nocross
))
952 panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
953 "overconstrained allocation",
954 (void *)vmp
, size
, align_arg
, phase
, nocross
,
955 minaddr
, maxaddr
, vmflag
);
957 if (phase
>= align
|| (align
& (align
- 1)) != 0 ||
958 (nocross
& (nocross
- 1)) != 0)
959 panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
960 "parameters inconsistent or invalid",
961 (void *)vmp
, size
, align_arg
, phase
, nocross
,
962 minaddr
, maxaddr
, vmflag
);
964 if ((mtbf
= vmem_mtbf
| vmp
->vm_mtbf
) != 0 && gethrtime() % mtbf
== 0 &&
965 (vmflag
& (VM_NOSLEEP
| VM_PANIC
)) == VM_NOSLEEP
)
968 mutex_enter(&vmp
->vm_lock
);
970 if (vmp
->vm_nsegfree
< VMEM_MINFREE
&&
971 !vmem_populate(vmp
, vmflag
))
975 * highbit() returns the highest bit + 1, which is exactly
976 * what we want: we want to search the first freelist whose
977 * members are *definitely* large enough to satisfy our
978 * allocation. However, there are certain cases in which we
979 * want to look at the next-smallest freelist (which *might*
980 * be able to satisfy the allocation):
982 * (1) The size is exactly a power of 2, in which case
983 * the smaller freelist is always big enough;
985 * (2) All other freelists are empty;
987 * (3) We're in the highest possible freelist, which is
988 * always empty (e.g. the 4GB freelist on 32-bit systems);
990 * (4) We're doing a best-fit or first-fit allocation.
992 if ((size
& (size
- 1)) == 0) {
993 flist
= lowbit(P2ALIGN(vmp
->vm_freemap
, size
));
996 if ((vmp
->vm_freemap
>> hb
) == 0 ||
997 hb
== VMEM_FREELISTS
||
998 (vmflag
& (VM_BESTFIT
| VM_FIRSTFIT
)))
1000 flist
= lowbit(P2ALIGN(vmp
->vm_freemap
, 1UL << hb
));
1003 for (vbest
= NULL
, vsp
= (flist
== 0) ? NULL
:
1004 vmp
->vm_freelist
[flist
- 1].vs_knext
;
1005 vsp
!= NULL
; vsp
= vsp
->vs_knext
) {
1006 vmp
->vm_kstat
.vk_search
.value
.ui64
++;
1007 if (vsp
->vs_start
== 0) {
1009 * We're moving up to a larger freelist,
1010 * so if we've already found a candidate,
1011 * the fit can't possibly get any better.
1016 * Find the next non-empty freelist.
1018 flist
= lowbit(P2ALIGN(vmp
->vm_freemap
,
1022 vsp
= (vmem_seg_t
*)&vmp
->vm_freelist
[flist
];
1023 ASSERT(vsp
->vs_knext
->vs_type
== VMEM_FREE
);
1026 if (vsp
->vs_end
- 1 < (uintptr_t)minaddr
)
1028 if (vsp
->vs_start
> (uintptr_t)maxaddr
- 1)
1030 start
= MAX(vsp
->vs_start
, (uintptr_t)minaddr
);
1031 end
= MIN(vsp
->vs_end
- 1, (uintptr_t)maxaddr
- 1) + 1;
1032 taddr
= P2PHASEUP(start
, align
, phase
);
1033 if (P2BOUNDARY(taddr
, size
, nocross
))
1035 P2ROUNDUP(P2NPHASE(taddr
, nocross
), align
);
1036 if ((taddr
- start
) + size
> end
- start
||
1037 (vbest
!= NULL
&& VS_SIZE(vsp
) >= VS_SIZE(vbest
)))
1041 if (!(vmflag
& VM_BESTFIT
) || VS_SIZE(vbest
) == size
)
1046 ASSERT(xvaddr
== NULL
);
1048 panic("vmem_xalloc(): size == 0");
1049 if (vmp
->vm_source_alloc
!= NULL
&& nocross
== 0 &&
1050 minaddr
== NULL
&& maxaddr
== NULL
) {
1051 size_t aneeded
, asize
;
1052 size_t aquantum
= MAX(vmp
->vm_quantum
,
1053 vmp
->vm_source
->vm_quantum
);
1054 size_t aphase
= phase
;
1055 if ((align
> aquantum
) &&
1056 !(vmp
->vm_cflags
& VMC_XALIGN
)) {
1057 aphase
= (P2PHASE(phase
, aquantum
) != 0) ?
1058 align
- vmp
->vm_quantum
: align
- aquantum
;
1059 ASSERT(aphase
>= phase
);
1061 aneeded
= MAX(size
+ aphase
, vmp
->vm_min_import
);
1062 asize
= P2ROUNDUP(aneeded
, aquantum
);
1065 * Determine how many segment structures we'll consume.
1066 * The calculation must be precise because if we're
1067 * here on behalf of vmem_populate(), we are taking
1068 * segments from a very limited reserve.
1070 if (size
== asize
&& !(vmp
->vm_cflags
& VMC_XALLOC
))
1071 resv
= VMEM_SEGS_PER_SPAN_CREATE
+
1072 VMEM_SEGS_PER_EXACT_ALLOC
;
1073 else if (phase
== 0 &&
1074 align
<= vmp
->vm_source
->vm_quantum
)
1075 resv
= VMEM_SEGS_PER_SPAN_CREATE
+
1076 VMEM_SEGS_PER_LEFT_ALLOC
;
1078 resv
= VMEM_SEGS_PER_ALLOC_MAX
;
1080 ASSERT(vmp
->vm_nsegfree
>= resv
);
1081 vmp
->vm_nsegfree
-= resv
; /* reserve our segs */
1082 mutex_exit(&vmp
->vm_lock
);
1083 if (vmp
->vm_cflags
& VMC_XALLOC
) {
1084 size_t oasize
= asize
;
1085 vaddr
= ((vmem_ximport_t
*)
1086 vmp
->vm_source_alloc
)(vmp
->vm_source
,
1087 &asize
, align
, vmflag
& VM_KMFLAGS
);
1088 ASSERT(asize
>= oasize
);
1089 ASSERT(P2PHASE(asize
,
1090 vmp
->vm_source
->vm_quantum
) == 0);
1091 ASSERT(!(vmp
->vm_cflags
& VMC_XALIGN
) ||
1092 IS_P2ALIGNED(vaddr
, align
));
1094 vaddr
= vmp
->vm_source_alloc(vmp
->vm_source
,
1095 asize
, vmflag
& VM_KMFLAGS
);
1097 mutex_enter(&vmp
->vm_lock
);
1098 vmp
->vm_nsegfree
+= resv
; /* claim reservation */
1099 aneeded
= size
+ align
- vmp
->vm_quantum
;
1100 aneeded
= P2ROUNDUP(aneeded
, vmp
->vm_quantum
);
1101 if (vaddr
!= NULL
) {
1103 * Since we dropped the vmem lock while
1104 * calling the import function, other
1105 * threads could have imported space
1106 * and made our import unnecessary. In
1107 * order to save space, we return
1108 * excess imports immediately.
1110 if (asize
> aneeded
&&
1111 vmp
->vm_source_free
!= NULL
&&
1112 vmem_canalloc(vmp
, aneeded
)) {
1114 VMEM_SEGS_PER_MIDDLE_ALLOC
);
1119 vbest
= vmem_span_create(vmp
, vaddr
, asize
, 1);
1120 addr
= P2PHASEUP(vbest
->vs_start
, align
, phase
);
1122 } else if (vmem_canalloc(vmp
, aneeded
)) {
1124 * Our import failed, but another thread
1125 * added sufficient free memory to the arena
1126 * to satisfy our request. Go back and
1129 ASSERT(resv
>= VMEM_SEGS_PER_MIDDLE_ALLOC
);
1135 * If the requestor chooses to fail the allocation attempt
1136 * rather than reap wait and retry - get out of the loop.
1138 if (vmflag
& VM_ABORT
)
1140 mutex_exit(&vmp
->vm_lock
);
1141 if (vmp
->vm_cflags
& VMC_IDENTIFIER
)
1142 kmem_reap_idspace();
1145 mutex_enter(&vmp
->vm_lock
);
1146 if (vmflag
& VM_NOSLEEP
)
1148 vmp
->vm_kstat
.vk_wait
.value
.ui64
++;
1149 cv_wait(&vmp
->vm_cv
, &vmp
->vm_lock
);
1151 if (vbest
!= NULL
) {
1152 ASSERT(vbest
->vs_type
== VMEM_FREE
);
1153 ASSERT(vbest
->vs_knext
!= vbest
);
1154 (void) vmem_seg_alloc(vmp
, vbest
, addr
, size
);
1155 mutex_exit(&vmp
->vm_lock
);
1157 vmp
->vm_source_free(vmp
->vm_source
, xvaddr
, xsize
);
1158 ASSERT(P2PHASE(addr
, align
) == phase
);
1159 ASSERT(!P2BOUNDARY(addr
, size
, nocross
));
1160 ASSERT(addr
>= (uintptr_t)minaddr
);
1161 ASSERT(addr
+ size
- 1 <= (uintptr_t)maxaddr
- 1);
1162 return ((void *)addr
);
1164 vmp
->vm_kstat
.vk_fail
.value
.ui64
++;
1165 mutex_exit(&vmp
->vm_lock
);
1166 if (vmflag
& VM_PANIC
)
1167 panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
1168 "cannot satisfy mandatory allocation",
1169 (void *)vmp
, size
, align_arg
, phase
, nocross
,
1170 minaddr
, maxaddr
, vmflag
);
1171 ASSERT(xvaddr
== NULL
);
1176 * Free the segment [vaddr, vaddr + size), where vaddr was a constrained
1177 * allocation. vmem_xalloc() and vmem_xfree() must always be paired because
1178 * both routines bypass the quantum caches.
1181 vmem_xfree(vmem_t
*vmp
, void *vaddr
, size_t size
)
1183 vmem_seg_t
*vsp
, *vnext
, *vprev
;
1185 mutex_enter(&vmp
->vm_lock
);
1187 vsp
= vmem_hash_delete(vmp
, (uintptr_t)vaddr
, size
);
1188 vsp
->vs_end
= P2ROUNDUP(vsp
->vs_end
, vmp
->vm_quantum
);
1191 * Attempt to coalesce with the next segment.
1193 vnext
= vsp
->vs_anext
;
1194 if (vnext
->vs_type
== VMEM_FREE
) {
1195 ASSERT(vsp
->vs_end
== vnext
->vs_start
);
1196 vmem_freelist_delete(vmp
, vnext
);
1197 vsp
->vs_end
= vnext
->vs_end
;
1198 vmem_seg_destroy(vmp
, vnext
);
1202 * Attempt to coalesce with the previous segment.
1204 vprev
= vsp
->vs_aprev
;
1205 if (vprev
->vs_type
== VMEM_FREE
) {
1206 ASSERT(vprev
->vs_end
== vsp
->vs_start
);
1207 vmem_freelist_delete(vmp
, vprev
);
1208 vprev
->vs_end
= vsp
->vs_end
;
1209 vmem_seg_destroy(vmp
, vsp
);
1214 * If the entire span is free, return it to the source.
1216 if (vsp
->vs_aprev
->vs_import
&& vmp
->vm_source_free
!= NULL
&&
1217 vsp
->vs_aprev
->vs_type
== VMEM_SPAN
&&
1218 vsp
->vs_anext
->vs_type
== VMEM_SPAN
) {
1219 vaddr
= (void *)vsp
->vs_start
;
1220 size
= VS_SIZE(vsp
);
1221 ASSERT(size
== VS_SIZE(vsp
->vs_aprev
));
1222 vmem_span_destroy(vmp
, vsp
);
1223 mutex_exit(&vmp
->vm_lock
);
1224 vmp
->vm_source_free(vmp
->vm_source
, vaddr
, size
);
1226 vmem_freelist_insert(vmp
, vsp
);
1227 mutex_exit(&vmp
->vm_lock
);
1232 * Allocate size bytes from arena vmp. Returns the allocated address
1233 * on success, NULL on failure. vmflag specifies VM_SLEEP or VM_NOSLEEP,
1234 * and may also specify best-fit, first-fit, or next-fit allocation policy
1235 * instead of the default instant-fit policy. VM_SLEEP allocations are
1236 * guaranteed to succeed.
1239 vmem_alloc(vmem_t
*vmp
, size_t size
, int vmflag
)
1247 if (size
- 1 < vmp
->vm_qcache_max
)
1248 return (kmem_cache_alloc(vmp
->vm_qcache
[(size
- 1) >>
1249 vmp
->vm_qshift
], vmflag
& VM_KMFLAGS
));
1251 if ((mtbf
= vmem_mtbf
| vmp
->vm_mtbf
) != 0 && gethrtime() % mtbf
== 0 &&
1252 (vmflag
& (VM_NOSLEEP
| VM_PANIC
)) == VM_NOSLEEP
)
1255 if (vmflag
& VM_NEXTFIT
)
1256 return (vmem_nextfit_alloc(vmp
, size
, vmflag
));
1258 if (vmflag
& (VM_BESTFIT
| VM_FIRSTFIT
))
1259 return (vmem_xalloc(vmp
, size
, vmp
->vm_quantum
, 0, 0,
1260 NULL
, NULL
, vmflag
));
1263 * Unconstrained instant-fit allocation from the segment list.
1265 mutex_enter(&vmp
->vm_lock
);
1267 if (vmp
->vm_nsegfree
>= VMEM_MINFREE
|| vmem_populate(vmp
, vmflag
)) {
1268 if ((size
& (size
- 1)) == 0)
1269 flist
= lowbit(P2ALIGN(vmp
->vm_freemap
, size
));
1270 else if ((hb
= highbit(size
)) < VMEM_FREELISTS
)
1271 flist
= lowbit(P2ALIGN(vmp
->vm_freemap
, 1UL << hb
));
1275 mutex_exit(&vmp
->vm_lock
);
1276 return (vmem_xalloc(vmp
, size
, vmp
->vm_quantum
,
1277 0, 0, NULL
, NULL
, vmflag
));
1280 ASSERT(size
<= (1UL << flist
));
1281 vsp
= vmp
->vm_freelist
[flist
].vs_knext
;
1282 addr
= vsp
->vs_start
;
1283 (void) vmem_seg_alloc(vmp
, vsp
, addr
, size
);
1284 mutex_exit(&vmp
->vm_lock
);
1285 return ((void *)addr
);
1289 * Free the segment [vaddr, vaddr + size).
1292 vmem_free(vmem_t
*vmp
, void *vaddr
, size_t size
)
1294 if (size
- 1 < vmp
->vm_qcache_max
)
1295 kmem_cache_free(vmp
->vm_qcache
[(size
- 1) >> vmp
->vm_qshift
],
1298 vmem_xfree(vmp
, vaddr
, size
);
1302 * Determine whether arena vmp contains the segment [vaddr, vaddr + size).
1305 vmem_contains(vmem_t
*vmp
, void *vaddr
, size_t size
)
1307 uintptr_t start
= (uintptr_t)vaddr
;
1308 uintptr_t end
= start
+ size
;
1310 vmem_seg_t
*seg0
= &vmp
->vm_seg0
;
1312 mutex_enter(&vmp
->vm_lock
);
1313 vmp
->vm_kstat
.vk_contains
.value
.ui64
++;
1314 for (vsp
= seg0
->vs_knext
; vsp
!= seg0
; vsp
= vsp
->vs_knext
) {
1315 vmp
->vm_kstat
.vk_contains_search
.value
.ui64
++;
1316 ASSERT(vsp
->vs_type
== VMEM_SPAN
);
1317 if (start
>= vsp
->vs_start
&& end
- 1 <= vsp
->vs_end
- 1)
1320 mutex_exit(&vmp
->vm_lock
);
1321 return (vsp
!= seg0
);
1325 * Add the span [vaddr, vaddr + size) to arena vmp.
1328 vmem_add(vmem_t
*vmp
, void *vaddr
, size_t size
, int vmflag
)
1330 if (vaddr
== NULL
|| size
== 0)
1331 panic("vmem_add(%p, %p, %lu): bad arguments",
1332 (void *)vmp
, vaddr
, size
);
1334 ASSERT(!vmem_contains(vmp
, vaddr
, size
));
1336 mutex_enter(&vmp
->vm_lock
);
1337 if (vmem_populate(vmp
, vmflag
))
1338 (void) vmem_span_create(vmp
, vaddr
, size
, 0);
1341 mutex_exit(&vmp
->vm_lock
);
1346 * Walk the vmp arena, applying func to each segment matching typemask.
1347 * If VMEM_REENTRANT is specified, the arena lock is dropped across each
1348 * call to func(); otherwise, it is held for the duration of vmem_walk()
1349 * to ensure a consistent snapshot. Note that VMEM_REENTRANT callbacks
1350 * are *not* necessarily consistent, so they may only be used when a hint
1354 vmem_walk(vmem_t
*vmp
, int typemask
,
1355 void (*func
)(void *, void *, size_t), void *arg
)
1358 vmem_seg_t
*seg0
= &vmp
->vm_seg0
;
1361 if (typemask
& VMEM_WALKER
)
1364 bzero(&walker
, sizeof (walker
));
1365 walker
.vs_type
= VMEM_WALKER
;
1367 mutex_enter(&vmp
->vm_lock
);
1368 VMEM_INSERT(seg0
, &walker
, a
);
1369 for (vsp
= seg0
->vs_anext
; vsp
!= seg0
; vsp
= vsp
->vs_anext
) {
1370 if (vsp
->vs_type
& typemask
) {
1371 void *start
= (void *)vsp
->vs_start
;
1372 size_t size
= VS_SIZE(vsp
);
1373 if (typemask
& VMEM_REENTRANT
) {
1374 vmem_advance(vmp
, &walker
, vsp
);
1375 mutex_exit(&vmp
->vm_lock
);
1376 func(arg
, start
, size
);
1377 mutex_enter(&vmp
->vm_lock
);
1380 func(arg
, start
, size
);
1384 vmem_advance(vmp
, &walker
, NULL
);
1385 mutex_exit(&vmp
->vm_lock
);
1389 * Return the total amount of memory whose type matches typemask. Thus:
1391 * typemask VMEM_ALLOC yields total memory allocated (in use).
1392 * typemask VMEM_FREE yields total memory free (available).
1393 * typemask (VMEM_ALLOC | VMEM_FREE) yields total arena size.
1396 vmem_size(vmem_t
*vmp
, int typemask
)
1400 if (typemask
& VMEM_ALLOC
)
1401 size
+= vmp
->vm_kstat
.vk_mem_inuse
.value
.ui64
;
1402 if (typemask
& VMEM_FREE
)
1403 size
+= vmp
->vm_kstat
.vk_mem_total
.value
.ui64
-
1404 vmp
->vm_kstat
.vk_mem_inuse
.value
.ui64
;
1405 return ((size_t)size
);
1409 * Create an arena called name whose initial span is [base, base + size).
1410 * The arena's natural unit of currency is quantum, so vmem_alloc()
1411 * guarantees quantum-aligned results. The arena may import new spans
1412 * by invoking afunc() on source, and may return those spans by invoking
1413 * ffunc() on source. To make small allocations fast and scalable,
1414 * the arena offers high-performance caching for each integer multiple
1415 * of quantum up to qcache_max.
1418 vmem_create_common(const char *name
, void *base
, size_t size
, size_t quantum
,
1419 void *(*afunc
)(vmem_t
*, size_t, int),
1420 void (*ffunc
)(vmem_t
*, void *, size_t),
1421 vmem_t
*source
, size_t qcache_max
, int vmflag
)
1425 vmem_t
*vmp
, *cur
, **vmpp
;
1427 vmem_freelist_t
*vfp
;
1428 uint32_t id
= atomic_add_32_nv(&vmem_id
, 1);
1430 if (vmem_vmem_arena
!= NULL
) {
1431 vmp
= vmem_alloc(vmem_vmem_arena
, sizeof (vmem_t
),
1432 vmflag
& VM_KMFLAGS
);
1434 ASSERT(id
<= VMEM_INITIAL
);
1435 vmp
= &vmem0
[id
- 1];
1438 /* An identifier arena must inherit from another identifier arena */
1439 ASSERT(source
== NULL
|| ((source
->vm_cflags
& VMC_IDENTIFIER
) ==
1440 (vmflag
& VMC_IDENTIFIER
)));
1444 bzero(vmp
, sizeof (vmem_t
));
1446 (void) snprintf(vmp
->vm_name
, VMEM_NAMELEN
, "%s", name
);
1447 mutex_init(&vmp
->vm_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1448 cv_init(&vmp
->vm_cv
, NULL
, CV_DEFAULT
, NULL
);
1449 vmp
->vm_cflags
= vmflag
;
1450 vmflag
&= VM_KMFLAGS
;
1452 vmp
->vm_quantum
= quantum
;
1453 vmp
->vm_qshift
= highbit(quantum
) - 1;
1454 nqcache
= MIN(qcache_max
>> vmp
->vm_qshift
, VMEM_NQCACHE_MAX
);
1456 for (i
= 0; i
<= VMEM_FREELISTS
; i
++) {
1457 vfp
= &vmp
->vm_freelist
[i
];
1458 vfp
->vs_end
= 1UL << i
;
1459 vfp
->vs_knext
= (vmem_seg_t
*)(vfp
+ 1);
1460 vfp
->vs_kprev
= (vmem_seg_t
*)(vfp
- 1);
1463 vmp
->vm_freelist
[0].vs_kprev
= NULL
;
1464 vmp
->vm_freelist
[VMEM_FREELISTS
].vs_knext
= NULL
;
1465 vmp
->vm_freelist
[VMEM_FREELISTS
].vs_end
= 0;
1466 vmp
->vm_hash_table
= vmp
->vm_hash0
;
1467 vmp
->vm_hash_mask
= VMEM_HASH_INITIAL
- 1;
1468 vmp
->vm_hash_shift
= highbit(vmp
->vm_hash_mask
);
1470 vsp
= &vmp
->vm_seg0
;
1471 vsp
->vs_anext
= vsp
;
1472 vsp
->vs_aprev
= vsp
;
1473 vsp
->vs_knext
= vsp
;
1474 vsp
->vs_kprev
= vsp
;
1475 vsp
->vs_type
= VMEM_SPAN
;
1477 vsp
= &vmp
->vm_rotor
;
1478 vsp
->vs_type
= VMEM_ROTOR
;
1479 VMEM_INSERT(&vmp
->vm_seg0
, vsp
, a
);
1481 bcopy(&vmem_kstat_template
, &vmp
->vm_kstat
, sizeof (vmem_kstat_t
));
1485 vmp
->vm_kstat
.vk_source_id
.value
.ui32
= source
->vm_id
;
1486 vmp
->vm_source
= source
;
1487 vmp
->vm_source_alloc
= afunc
;
1488 vmp
->vm_source_free
= ffunc
;
1491 * Some arenas (like vmem_metadata and kmem_metadata) cannot
1492 * use quantum caching to lower fragmentation. Instead, we
1493 * increase their imports, giving a similar effect.
1495 if (vmp
->vm_cflags
& VMC_NO_QCACHE
) {
1496 vmp
->vm_min_import
=
1497 VMEM_QCACHE_SLABSIZE(nqcache
<< vmp
->vm_qshift
);
1502 ASSERT(!(vmflag
& VM_NOSLEEP
));
1503 vmp
->vm_qcache_max
= nqcache
<< vmp
->vm_qshift
;
1504 for (i
= 0; i
< nqcache
; i
++) {
1505 char buf
[VMEM_NAMELEN
+ 21];
1506 (void) sprintf(buf
, "%s_%lu", vmp
->vm_name
,
1508 vmp
->vm_qcache
[i
] = kmem_cache_create(buf
,
1509 (i
+ 1) * quantum
, quantum
, NULL
, NULL
, NULL
,
1510 NULL
, vmp
, KMC_QCACHE
| KMC_NOTOUCH
);
1514 if ((vmp
->vm_ksp
= kstat_create("vmem", vmp
->vm_id
, vmp
->vm_name
,
1515 "vmem", KSTAT_TYPE_NAMED
, sizeof (vmem_kstat_t
) /
1516 sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
)) != NULL
) {
1517 vmp
->vm_ksp
->ks_data
= &vmp
->vm_kstat
;
1518 kstat_install(vmp
->vm_ksp
);
1521 mutex_enter(&vmem_list_lock
);
1523 while ((cur
= *vmpp
) != NULL
)
1524 vmpp
= &cur
->vm_next
;
1526 mutex_exit(&vmem_list_lock
);
1528 if (vmp
->vm_cflags
& VMC_POPULATOR
) {
1529 ASSERT(vmem_populators
< VMEM_INITIAL
);
1530 vmem_populator
[atomic_add_32_nv(&vmem_populators
, 1) - 1] = vmp
;
1531 mutex_enter(&vmp
->vm_lock
);
1532 (void) vmem_populate(vmp
, vmflag
| VM_PANIC
);
1533 mutex_exit(&vmp
->vm_lock
);
1536 if ((base
|| size
) && vmem_add(vmp
, base
, size
, vmflag
) == NULL
) {
1545 vmem_xcreate(const char *name
, void *base
, size_t size
, size_t quantum
,
1546 vmem_ximport_t
*afunc
, vmem_free_t
*ffunc
, vmem_t
*source
,
1547 size_t qcache_max
, int vmflag
)
1549 ASSERT(!(vmflag
& (VMC_POPULATOR
| VMC_XALLOC
)));
1550 vmflag
&= ~(VMC_POPULATOR
| VMC_XALLOC
);
1552 return (vmem_create_common(name
, base
, size
, quantum
,
1553 (vmem_alloc_t
*)afunc
, ffunc
, source
, qcache_max
,
1554 vmflag
| VMC_XALLOC
));
1558 vmem_create(const char *name
, void *base
, size_t size
, size_t quantum
,
1559 vmem_alloc_t
*afunc
, vmem_free_t
*ffunc
, vmem_t
*source
,
1560 size_t qcache_max
, int vmflag
)
1562 ASSERT(!(vmflag
& (VMC_XALLOC
| VMC_XALIGN
)));
1563 vmflag
&= ~(VMC_XALLOC
| VMC_XALIGN
);
1565 return (vmem_create_common(name
, base
, size
, quantum
,
1566 afunc
, ffunc
, source
, qcache_max
, vmflag
));
1570 * Destroy arena vmp.
1573 vmem_destroy(vmem_t
*vmp
)
1575 vmem_t
*cur
, **vmpp
;
1576 vmem_seg_t
*seg0
= &vmp
->vm_seg0
;
1581 mutex_enter(&vmem_list_lock
);
1583 while ((cur
= *vmpp
) != vmp
)
1584 vmpp
= &cur
->vm_next
;
1585 *vmpp
= vmp
->vm_next
;
1586 mutex_exit(&vmem_list_lock
);
1588 for (i
= 0; i
< VMEM_NQCACHE_MAX
; i
++)
1589 if (vmp
->vm_qcache
[i
])
1590 kmem_cache_destroy(vmp
->vm_qcache
[i
]);
1592 leaked
= vmem_size(vmp
, VMEM_ALLOC
);
1594 cmn_err(CE_WARN
, "vmem_destroy('%s'): leaked %lu %s",
1595 vmp
->vm_name
, leaked
, (vmp
->vm_cflags
& VMC_IDENTIFIER
) ?
1596 "identifiers" : "bytes");
1598 if (vmp
->vm_hash_table
!= vmp
->vm_hash0
)
1599 vmem_free(vmem_hash_arena
, vmp
->vm_hash_table
,
1600 (vmp
->vm_hash_mask
+ 1) * sizeof (void *));
1603 * Give back the segment structures for anything that's left in the
1604 * arena, e.g. the primary spans and their free segments.
1606 VMEM_DELETE(&vmp
->vm_rotor
, a
);
1607 for (vsp
= seg0
->vs_anext
; vsp
!= seg0
; vsp
= vsp
->vs_anext
)
1608 vmem_putseg_global(vsp
);
1610 while (vmp
->vm_nsegfree
> 0)
1611 vmem_putseg_global(vmem_getseg(vmp
));
1613 kstat_delete(vmp
->vm_ksp
);
1615 mutex_destroy(&vmp
->vm_lock
);
1616 cv_destroy(&vmp
->vm_cv
);
1617 vmem_free(vmem_vmem_arena
, vmp
, sizeof (vmem_t
));
1621 * Resize vmp's hash table to keep the average lookup depth near 1.0.
1624 vmem_hash_rescale(vmem_t
*vmp
)
1626 vmem_seg_t
**old_table
, **new_table
, *vsp
;
1627 size_t old_size
, new_size
, h
, nseg
;
1629 nseg
= (size_t)(vmp
->vm_kstat
.vk_alloc
.value
.ui64
-
1630 vmp
->vm_kstat
.vk_free
.value
.ui64
);
1632 new_size
= MAX(VMEM_HASH_INITIAL
, 1 << (highbit(3 * nseg
+ 4) - 2));
1633 old_size
= vmp
->vm_hash_mask
+ 1;
1635 if ((old_size
>> 1) <= new_size
&& new_size
<= (old_size
<< 1))
1638 new_table
= vmem_alloc(vmem_hash_arena
, new_size
* sizeof (void *),
1640 if (new_table
== NULL
)
1642 bzero(new_table
, new_size
* sizeof (void *));
1644 mutex_enter(&vmp
->vm_lock
);
1646 old_size
= vmp
->vm_hash_mask
+ 1;
1647 old_table
= vmp
->vm_hash_table
;
1649 vmp
->vm_hash_mask
= new_size
- 1;
1650 vmp
->vm_hash_table
= new_table
;
1651 vmp
->vm_hash_shift
= highbit(vmp
->vm_hash_mask
);
1653 for (h
= 0; h
< old_size
; h
++) {
1655 while (vsp
!= NULL
) {
1656 uintptr_t addr
= vsp
->vs_start
;
1657 vmem_seg_t
*next_vsp
= vsp
->vs_knext
;
1658 vmem_seg_t
**hash_bucket
= VMEM_HASH(vmp
, addr
);
1659 vsp
->vs_knext
= *hash_bucket
;
1665 mutex_exit(&vmp
->vm_lock
);
1667 if (old_table
!= vmp
->vm_hash0
)
1668 vmem_free(vmem_hash_arena
, old_table
,
1669 old_size
* sizeof (void *));
1673 * Perform periodic maintenance on all vmem arenas.
1676 vmem_update(void *dummy
)
1680 mutex_enter(&vmem_list_lock
);
1681 for (vmp
= vmem_list
; vmp
!= NULL
; vmp
= vmp
->vm_next
) {
1683 * If threads are waiting for resources, wake them up
1684 * periodically so they can issue another kmem_reap()
1685 * to reclaim resources cached by the slab allocator.
1687 cv_broadcast(&vmp
->vm_cv
);
1690 * Rescale the hash table to keep the hash chains short.
1692 vmem_hash_rescale(vmp
);
1694 mutex_exit(&vmem_list_lock
);
1696 (void) timeout(vmem_update
, dummy
, vmem_update_interval
* hz
);
1700 * Prepare vmem for use.
1703 vmem_init(const char *heap_name
,
1704 void *heap_start
, size_t heap_size
, size_t heap_quantum
,
1705 void *(*heap_alloc
)(vmem_t
*, size_t, int),
1706 void (*heap_free
)(vmem_t
*, void *, size_t))
1709 int nseg
= VMEM_SEG_INITIAL
;
1713 vmem_putseg_global(&vmem_seg0
[nseg
]);
1715 heap
= vmem_create(heap_name
,
1716 heap_start
, heap_size
, heap_quantum
,
1717 NULL
, NULL
, NULL
, 0,
1718 VM_SLEEP
| VMC_POPULATOR
);
1720 vmem_metadata_arena
= vmem_create("vmem_metadata",
1721 NULL
, 0, heap_quantum
,
1722 vmem_alloc
, vmem_free
, heap
, 8 * heap_quantum
,
1723 VM_SLEEP
| VMC_POPULATOR
| VMC_NO_QCACHE
);
1725 vmem_seg_arena
= vmem_create("vmem_seg",
1726 NULL
, 0, heap_quantum
,
1727 heap_alloc
, heap_free
, vmem_metadata_arena
, 0,
1728 VM_SLEEP
| VMC_POPULATOR
);
1730 vmem_hash_arena
= vmem_create("vmem_hash",
1732 heap_alloc
, heap_free
, vmem_metadata_arena
, 0,
1735 vmem_vmem_arena
= vmem_create("vmem_vmem",
1736 vmem0
, sizeof (vmem0
), 1,
1737 heap_alloc
, heap_free
, vmem_metadata_arena
, 0,
1740 for (id
= 0; id
< vmem_id
; id
++)
1741 (void) vmem_xalloc(vmem_vmem_arena
, sizeof (vmem_t
),
1742 1, 0, 0, &vmem0
[id
], &vmem0
[id
+ 1],
1743 VM_NOSLEEP
| VM_BESTFIT
| VM_PANIC
);