Print this page
7831 want vmem manual pages
7832 big theory statements need a place in the manual
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/os/vmem.c
+++ new/usr/src/uts/common/os/vmem.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
23 23 * Use is subject to license terms.
24 24 */
25 25
26 26 /*
27 27 * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
28 28 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
29 29 */
30 30
31 31 /*
32 32 * Big Theory Statement for the virtual memory allocator.
33 33 *
|
↓ open down ↓ |
33 lines elided |
↑ open up ↑ |
34 34 * For a more complete description of the main ideas, see:
35 35 *
36 36 * Jeff Bonwick and Jonathan Adams,
37 37 *
38 38 * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
39 39 * Arbitrary Resources.
40 40 *
41 41 * Proceedings of the 2001 Usenix Conference.
42 42 * Available as http://www.usenix.org/event/usenix01/bonwick.html
43 43 *
44 + * Section 1, below, is also the primary contents of vmem(9). If for some
45 + * reason you are updating this comment, you will also wish to update the
46 + * manual.
44 47 *
45 48 * 1. General Concepts
46 49 * -------------------
47 50 *
48 51 * 1.1 Overview
49 52 * ------------
50 53 * We divide the kernel address space into a number of logically distinct
51 54 * pieces, or *arenas*: text, data, heap, stack, and so on. Within these
52 55 * arenas we often subdivide further; for example, we use heap addresses
53 56 * not only for the kernel heap (kmem_alloc() space), but also for DVMA,
54 57 * bp_mapin(), /dev/kmem, and even some device mappings like the TOD chip.
55 58 * The kernel address space, therefore, is most accurately described as
56 59 * a tree of arenas in which each node of the tree *imports* some subset
57 60 * of its parent. The virtual memory allocator manages these arenas and
58 61 * supports their natural hierarchical structure.
59 62 *
60 63 * 1.2 Arenas
61 64 * ----------
62 65 * An arena is nothing more than a set of integers. These integers most
63 66 * commonly represent virtual addresses, but in fact they can represent
64 67 * anything at all. For example, we could use an arena containing the
65 68 * integers minpid through maxpid to allocate process IDs. vmem_create()
66 69 * and vmem_destroy() create and destroy vmem arenas. In order to
67 70 * differentiate between arenas used for adresses and arenas used for
68 71 * identifiers, the VMC_IDENTIFIER flag is passed to vmem_create(). This
69 72 * prevents identifier exhaustion from being diagnosed as general memory
70 73 * failure.
71 74 *
72 75 * 1.3 Spans
73 76 * ---------
74 77 * We represent the integers in an arena as a collection of *spans*, or
75 78 * contiguous ranges of integers. For example, the kernel heap consists
76 79 * of just one span: [kernelheap, ekernelheap). Spans can be added to an
77 80 * arena in two ways: explicitly, by vmem_add(), or implicitly, by
78 81 * importing, as described in Section 1.5 below.
79 82 *
80 83 * 1.4 Segments
81 84 * ------------
82 85 * Spans are subdivided into *segments*, each of which is either allocated
83 86 * or free. A segment, like a span, is a contiguous range of integers.
84 87 * Each allocated segment [addr, addr + size) represents exactly one
85 88 * vmem_alloc(size) that returned addr. Free segments represent the space
86 89 * between allocated segments. If two free segments are adjacent, we
87 90 * coalesce them into one larger segment; that is, if segments [a, b) and
88 91 * [b, c) are both free, we merge them into a single segment [a, c).
89 92 * The segments within a span are linked together in increasing-address order
90 93 * so we can easily determine whether coalescing is possible.
91 94 *
92 95 * Segments never cross span boundaries. When all segments within
93 96 * an imported span become free, we return the span to its source.
94 97 *
95 98 * 1.5 Imported Memory
96 99 * -------------------
97 100 * As mentioned in the overview, some arenas are logical subsets of
98 101 * other arenas. For example, kmem_va_arena (a virtual address cache
99 102 * that satisfies most kmem_slab_create() requests) is just a subset
100 103 * of heap_arena (the kernel heap) that provides caching for the most
101 104 * common slab sizes. When kmem_va_arena runs out of virtual memory,
102 105 * it *imports* more from the heap; we say that heap_arena is the
103 106 * *vmem source* for kmem_va_arena. vmem_create() allows you to
104 107 * specify any existing vmem arena as the source for your new arena.
105 108 * Topologically, since every arena is a child of at most one source,
106 109 * the set of all arenas forms a collection of trees.
107 110 *
108 111 * 1.6 Constrained Allocations
109 112 * ---------------------------
110 113 * Some vmem clients are quite picky about the kind of address they want.
111 114 * For example, the DVMA code may need an address that is at a particular
112 115 * phase with respect to some alignment (to get good cache coloring), or
113 116 * that lies within certain limits (the addressable range of a device),
114 117 * or that doesn't cross some boundary (a DMA counter restriction) --
115 118 * or all of the above. vmem_xalloc() allows the client to specify any
116 119 * or all of these constraints.
117 120 *
118 121 * 1.7 The Vmem Quantum
119 122 * --------------------
120 123 * Every arena has a notion of 'quantum', specified at vmem_create() time,
121 124 * that defines the arena's minimum unit of currency. Most commonly the
122 125 * quantum is either 1 or PAGESIZE, but any power of 2 is legal.
123 126 * All vmem allocations are guaranteed to be quantum-aligned.
124 127 *
125 128 * 1.8 Quantum Caching
126 129 * -------------------
127 130 * A vmem arena may be so hot (frequently used) that the scalability of vmem
128 131 * allocation is a significant concern. We address this by allowing the most
129 132 * common allocation sizes to be serviced by the kernel memory allocator,
130 133 * which provides low-latency per-cpu caching. The qcache_max argument to
131 134 * vmem_create() specifies the largest allocation size to cache.
132 135 *
133 136 * 1.9 Relationship to Kernel Memory Allocator
134 137 * -------------------------------------------
135 138 * Every kmem cache has a vmem arena as its slab supplier. The kernel memory
136 139 * allocator uses vmem_alloc() and vmem_free() to create and destroy slabs.
137 140 *
138 141 *
139 142 * 2. Implementation
140 143 * -----------------
141 144 *
142 145 * 2.1 Segment lists and markers
143 146 * -----------------------------
144 147 * The segment structure (vmem_seg_t) contains two doubly-linked lists.
145 148 *
146 149 * The arena list (vs_anext/vs_aprev) links all segments in the arena.
147 150 * In addition to the allocated and free segments, the arena contains
148 151 * special marker segments at span boundaries. Span markers simplify
149 152 * coalescing and importing logic by making it easy to tell both when
150 153 * we're at a span boundary (so we don't coalesce across it), and when
151 154 * a span is completely free (its neighbors will both be span markers).
152 155 *
153 156 * Imported spans will have vs_import set.
154 157 *
155 158 * The next-of-kin list (vs_knext/vs_kprev) links segments of the same type:
156 159 * (1) for allocated segments, vs_knext is the hash chain linkage;
157 160 * (2) for free segments, vs_knext is the freelist linkage;
158 161 * (3) for span marker segments, vs_knext is the next span marker.
159 162 *
160 163 * 2.2 Allocation hashing
161 164 * ----------------------
162 165 * We maintain a hash table of all allocated segments, hashed by address.
163 166 * This allows vmem_free() to discover the target segment in constant time.
164 167 * vmem_update() periodically resizes hash tables to keep hash chains short.
165 168 *
166 169 * 2.3 Freelist management
167 170 * -----------------------
168 171 * We maintain power-of-2 freelists for free segments, i.e. free segments
169 172 * of size >= 2^n reside in vmp->vm_freelist[n]. To ensure constant-time
170 173 * allocation, vmem_xalloc() looks not in the first freelist that *might*
171 174 * satisfy the allocation, but in the first freelist that *definitely*
172 175 * satisfies the allocation (unless VM_BESTFIT is specified, or all larger
173 176 * freelists are empty). For example, a 1000-byte allocation will be
174 177 * satisfied not from the 512..1023-byte freelist, whose members *might*
175 178 * contains a 1000-byte segment, but from a 1024-byte or larger freelist,
176 179 * the first member of which will *definitely* satisfy the allocation.
177 180 * This ensures that vmem_xalloc() works in constant time.
178 181 *
179 182 * We maintain a bit map to determine quickly which freelists are non-empty.
180 183 * vmp->vm_freemap & (1 << n) is non-zero iff vmp->vm_freelist[n] is non-empty.
181 184 *
182 185 * The different freelists are linked together into one large freelist,
183 186 * with the freelist heads serving as markers. Freelist markers simplify
184 187 * the maintenance of vm_freemap by making it easy to tell when we're taking
185 188 * the last member of a freelist (both of its neighbors will be markers).
186 189 *
187 190 * 2.4 Vmem Locking
188 191 * ----------------
189 192 * For simplicity, all arena state is protected by a per-arena lock.
190 193 * For very hot arenas, use quantum caching for scalability.
191 194 *
192 195 * 2.5 Vmem Population
193 196 * -------------------
194 197 * Any internal vmem routine that might need to allocate new segment
195 198 * structures must prepare in advance by calling vmem_populate(), which
196 199 * will preallocate enough vmem_seg_t's to get is through the entire
197 200 * operation without dropping the arena lock.
198 201 *
199 202 * 2.6 Auditing
200 203 * ------------
201 204 * If KMF_AUDIT is set in kmem_flags, we audit vmem allocations as well.
202 205 * Since virtual addresses cannot be scribbled on, there is no equivalent
203 206 * in vmem to redzone checking, deadbeef, or other kmem debugging features.
204 207 * Moreover, we do not audit frees because segment coalescing destroys the
205 208 * association between an address and its segment structure. Auditing is
206 209 * thus intended primarily to keep track of who's consuming the arena.
207 210 * Debugging support could certainly be extended in the future if it proves
208 211 * necessary, but we do so much live checking via the allocation hash table
209 212 * that even non-DEBUG systems get quite a bit of sanity checking already.
210 213 */
211 214
212 215 #include <sys/vmem_impl.h>
213 216 #include <sys/kmem.h>
214 217 #include <sys/kstat.h>
215 218 #include <sys/param.h>
216 219 #include <sys/systm.h>
217 220 #include <sys/atomic.h>
218 221 #include <sys/bitmap.h>
219 222 #include <sys/sysmacros.h>
220 223 #include <sys/cmn_err.h>
221 224 #include <sys/debug.h>
222 225 #include <sys/panic.h>
223 226
224 227 #define VMEM_INITIAL 10 /* early vmem arenas */
225 228 #define VMEM_SEG_INITIAL 200 /* early segments */
226 229
227 230 /*
228 231 * Adding a new span to an arena requires two segment structures: one to
229 232 * represent the span, and one to represent the free segment it contains.
230 233 */
231 234 #define VMEM_SEGS_PER_SPAN_CREATE 2
232 235
233 236 /*
234 237 * Allocating a piece of an existing segment requires 0-2 segment structures
235 238 * depending on how much of the segment we're allocating.
236 239 *
237 240 * To allocate the entire segment, no new segment structures are needed; we
238 241 * simply move the existing segment structure from the freelist to the
239 242 * allocation hash table.
240 243 *
241 244 * To allocate a piece from the left or right end of the segment, we must
242 245 * split the segment into two pieces (allocated part and remainder), so we
243 246 * need one new segment structure to represent the remainder.
244 247 *
245 248 * To allocate from the middle of a segment, we need two new segment strucures
246 249 * to represent the remainders on either side of the allocated part.
247 250 */
248 251 #define VMEM_SEGS_PER_EXACT_ALLOC 0
249 252 #define VMEM_SEGS_PER_LEFT_ALLOC 1
250 253 #define VMEM_SEGS_PER_RIGHT_ALLOC 1
251 254 #define VMEM_SEGS_PER_MIDDLE_ALLOC 2
252 255
253 256 /*
254 257 * vmem_populate() preallocates segment structures for vmem to do its work.
255 258 * It must preallocate enough for the worst case, which is when we must import
256 259 * a new span and then allocate from the middle of it.
257 260 */
258 261 #define VMEM_SEGS_PER_ALLOC_MAX \
259 262 (VMEM_SEGS_PER_SPAN_CREATE + VMEM_SEGS_PER_MIDDLE_ALLOC)
260 263
261 264 /*
262 265 * The segment structures themselves are allocated from vmem_seg_arena, so
263 266 * we have a recursion problem when vmem_seg_arena needs to populate itself.
264 267 * We address this by working out the maximum number of segment structures
265 268 * this act will require, and multiplying by the maximum number of threads
266 269 * that we'll allow to do it simultaneously.
267 270 *
268 271 * The worst-case segment consumption to populate vmem_seg_arena is as
269 272 * follows (depicted as a stack trace to indicate why events are occurring):
270 273 *
271 274 * (In order to lower the fragmentation in the heap_arena, we specify a
272 275 * minimum import size for the vmem_metadata_arena which is the same size
273 276 * as the kmem_va quantum cache allocations. This causes the worst-case
274 277 * allocation from the vmem_metadata_arena to be 3 segments.)
275 278 *
276 279 * vmem_alloc(vmem_seg_arena) -> 2 segs (span create + exact alloc)
277 280 * segkmem_alloc(vmem_metadata_arena)
278 281 * vmem_alloc(vmem_metadata_arena) -> 3 segs (span create + left alloc)
279 282 * vmem_alloc(heap_arena) -> 1 seg (left alloc)
280 283 * page_create()
281 284 * hat_memload()
282 285 * kmem_cache_alloc()
283 286 * kmem_slab_create()
284 287 * vmem_alloc(hat_memload_arena) -> 2 segs (span create + exact alloc)
285 288 * segkmem_alloc(heap_arena)
286 289 * vmem_alloc(heap_arena) -> 1 seg (left alloc)
287 290 * page_create()
288 291 * hat_memload() -> (hat layer won't recurse further)
289 292 *
290 293 * The worst-case consumption for each arena is 3 segment structures.
291 294 * Of course, a 3-seg reserve could easily be blown by multiple threads.
292 295 * Therefore, we serialize all allocations from vmem_seg_arena (which is OK
293 296 * because they're rare). We cannot allow a non-blocking allocation to get
294 297 * tied up behind a blocking allocation, however, so we use separate locks
295 298 * for VM_SLEEP and VM_NOSLEEP allocations. Similarly, VM_PUSHPAGE allocations
296 299 * must not block behind ordinary VM_SLEEPs. In addition, if the system is
297 300 * panicking then we must keep enough resources for panic_thread to do its
298 301 * work. Thus we have at most four threads trying to allocate from
299 302 * vmem_seg_arena, and each thread consumes at most three segment structures,
300 303 * so we must maintain a 12-seg reserve.
301 304 */
302 305 #define VMEM_POPULATE_RESERVE 12
303 306
304 307 /*
305 308 * vmem_populate() ensures that each arena has VMEM_MINFREE seg structures
306 309 * so that it can satisfy the worst-case allocation *and* participate in
307 310 * worst-case allocation from vmem_seg_arena.
308 311 */
309 312 #define VMEM_MINFREE (VMEM_POPULATE_RESERVE + VMEM_SEGS_PER_ALLOC_MAX)
310 313
311 314 static vmem_t vmem0[VMEM_INITIAL];
312 315 static vmem_t *vmem_populator[VMEM_INITIAL];
313 316 static uint32_t vmem_id;
314 317 static uint32_t vmem_populators;
315 318 static vmem_seg_t vmem_seg0[VMEM_SEG_INITIAL];
316 319 static vmem_seg_t *vmem_segfree;
317 320 static kmutex_t vmem_list_lock;
318 321 static kmutex_t vmem_segfree_lock;
319 322 static kmutex_t vmem_sleep_lock;
320 323 static kmutex_t vmem_nosleep_lock;
321 324 static kmutex_t vmem_pushpage_lock;
322 325 static kmutex_t vmem_panic_lock;
323 326 static vmem_t *vmem_list;
324 327 static vmem_t *vmem_metadata_arena;
325 328 static vmem_t *vmem_seg_arena;
326 329 static vmem_t *vmem_hash_arena;
327 330 static vmem_t *vmem_vmem_arena;
328 331 static long vmem_update_interval = 15; /* vmem_update() every 15 seconds */
329 332 uint32_t vmem_mtbf; /* mean time between failures [default: off] */
330 333 size_t vmem_seg_size = sizeof (vmem_seg_t);
331 334
332 335 static vmem_kstat_t vmem_kstat_template = {
333 336 { "mem_inuse", KSTAT_DATA_UINT64 },
334 337 { "mem_import", KSTAT_DATA_UINT64 },
335 338 { "mem_total", KSTAT_DATA_UINT64 },
336 339 { "vmem_source", KSTAT_DATA_UINT32 },
337 340 { "alloc", KSTAT_DATA_UINT64 },
338 341 { "free", KSTAT_DATA_UINT64 },
339 342 { "wait", KSTAT_DATA_UINT64 },
340 343 { "fail", KSTAT_DATA_UINT64 },
341 344 { "lookup", KSTAT_DATA_UINT64 },
342 345 { "search", KSTAT_DATA_UINT64 },
343 346 { "populate_wait", KSTAT_DATA_UINT64 },
344 347 { "populate_fail", KSTAT_DATA_UINT64 },
345 348 { "contains", KSTAT_DATA_UINT64 },
346 349 { "contains_search", KSTAT_DATA_UINT64 },
347 350 };
348 351
349 352 /*
350 353 * Insert/delete from arena list (type 'a') or next-of-kin list (type 'k').
351 354 */
352 355 #define VMEM_INSERT(vprev, vsp, type) \
353 356 { \
354 357 vmem_seg_t *vnext = (vprev)->vs_##type##next; \
355 358 (vsp)->vs_##type##next = (vnext); \
356 359 (vsp)->vs_##type##prev = (vprev); \
357 360 (vprev)->vs_##type##next = (vsp); \
358 361 (vnext)->vs_##type##prev = (vsp); \
359 362 }
360 363
361 364 #define VMEM_DELETE(vsp, type) \
362 365 { \
363 366 vmem_seg_t *vprev = (vsp)->vs_##type##prev; \
364 367 vmem_seg_t *vnext = (vsp)->vs_##type##next; \
365 368 (vprev)->vs_##type##next = (vnext); \
366 369 (vnext)->vs_##type##prev = (vprev); \
367 370 }
368 371
369 372 /*
370 373 * Get a vmem_seg_t from the global segfree list.
371 374 */
372 375 static vmem_seg_t *
373 376 vmem_getseg_global(void)
374 377 {
375 378 vmem_seg_t *vsp;
376 379
377 380 mutex_enter(&vmem_segfree_lock);
378 381 if ((vsp = vmem_segfree) != NULL)
379 382 vmem_segfree = vsp->vs_knext;
380 383 mutex_exit(&vmem_segfree_lock);
381 384
382 385 return (vsp);
383 386 }
384 387
385 388 /*
386 389 * Put a vmem_seg_t on the global segfree list.
387 390 */
388 391 static void
389 392 vmem_putseg_global(vmem_seg_t *vsp)
390 393 {
391 394 mutex_enter(&vmem_segfree_lock);
392 395 vsp->vs_knext = vmem_segfree;
393 396 vmem_segfree = vsp;
394 397 mutex_exit(&vmem_segfree_lock);
395 398 }
396 399
397 400 /*
398 401 * Get a vmem_seg_t from vmp's segfree list.
399 402 */
400 403 static vmem_seg_t *
401 404 vmem_getseg(vmem_t *vmp)
402 405 {
403 406 vmem_seg_t *vsp;
404 407
405 408 ASSERT(vmp->vm_nsegfree > 0);
406 409
407 410 vsp = vmp->vm_segfree;
408 411 vmp->vm_segfree = vsp->vs_knext;
409 412 vmp->vm_nsegfree--;
410 413
411 414 return (vsp);
412 415 }
413 416
414 417 /*
415 418 * Put a vmem_seg_t on vmp's segfree list.
416 419 */
417 420 static void
418 421 vmem_putseg(vmem_t *vmp, vmem_seg_t *vsp)
419 422 {
420 423 vsp->vs_knext = vmp->vm_segfree;
421 424 vmp->vm_segfree = vsp;
422 425 vmp->vm_nsegfree++;
423 426 }
424 427
425 428 /*
426 429 * Add vsp to the appropriate freelist.
427 430 */
428 431 static void
429 432 vmem_freelist_insert(vmem_t *vmp, vmem_seg_t *vsp)
430 433 {
431 434 vmem_seg_t *vprev;
432 435
433 436 ASSERT(*VMEM_HASH(vmp, vsp->vs_start) != vsp);
434 437
435 438 vprev = (vmem_seg_t *)&vmp->vm_freelist[highbit(VS_SIZE(vsp)) - 1];
436 439 vsp->vs_type = VMEM_FREE;
437 440 vmp->vm_freemap |= VS_SIZE(vprev);
438 441 VMEM_INSERT(vprev, vsp, k);
439 442
440 443 cv_broadcast(&vmp->vm_cv);
441 444 }
442 445
443 446 /*
444 447 * Take vsp from the freelist.
445 448 */
446 449 static void
447 450 vmem_freelist_delete(vmem_t *vmp, vmem_seg_t *vsp)
448 451 {
449 452 ASSERT(*VMEM_HASH(vmp, vsp->vs_start) != vsp);
450 453 ASSERT(vsp->vs_type == VMEM_FREE);
451 454
452 455 if (vsp->vs_knext->vs_start == 0 && vsp->vs_kprev->vs_start == 0) {
453 456 /*
454 457 * The segments on both sides of 'vsp' are freelist heads,
455 458 * so taking vsp leaves the freelist at vsp->vs_kprev empty.
456 459 */
457 460 ASSERT(vmp->vm_freemap & VS_SIZE(vsp->vs_kprev));
458 461 vmp->vm_freemap ^= VS_SIZE(vsp->vs_kprev);
459 462 }
460 463 VMEM_DELETE(vsp, k);
461 464 }
462 465
463 466 /*
464 467 * Add vsp to the allocated-segment hash table and update kstats.
465 468 */
466 469 static void
467 470 vmem_hash_insert(vmem_t *vmp, vmem_seg_t *vsp)
468 471 {
469 472 vmem_seg_t **bucket;
470 473
471 474 vsp->vs_type = VMEM_ALLOC;
472 475 bucket = VMEM_HASH(vmp, vsp->vs_start);
473 476 vsp->vs_knext = *bucket;
474 477 *bucket = vsp;
475 478
476 479 if (vmem_seg_size == sizeof (vmem_seg_t)) {
477 480 vsp->vs_depth = (uint8_t)getpcstack(vsp->vs_stack,
478 481 VMEM_STACK_DEPTH);
479 482 vsp->vs_thread = curthread;
480 483 vsp->vs_timestamp = gethrtime();
481 484 } else {
482 485 vsp->vs_depth = 0;
483 486 }
484 487
485 488 vmp->vm_kstat.vk_alloc.value.ui64++;
486 489 vmp->vm_kstat.vk_mem_inuse.value.ui64 += VS_SIZE(vsp);
487 490 }
488 491
489 492 /*
490 493 * Remove vsp from the allocated-segment hash table and update kstats.
491 494 */
492 495 static vmem_seg_t *
493 496 vmem_hash_delete(vmem_t *vmp, uintptr_t addr, size_t size)
494 497 {
495 498 vmem_seg_t *vsp, **prev_vspp;
496 499
497 500 prev_vspp = VMEM_HASH(vmp, addr);
498 501 while ((vsp = *prev_vspp) != NULL) {
499 502 if (vsp->vs_start == addr) {
500 503 *prev_vspp = vsp->vs_knext;
501 504 break;
502 505 }
503 506 vmp->vm_kstat.vk_lookup.value.ui64++;
504 507 prev_vspp = &vsp->vs_knext;
505 508 }
506 509
507 510 if (vsp == NULL)
508 511 panic("vmem_hash_delete(%p, %lx, %lu): bad free",
509 512 (void *)vmp, addr, size);
510 513 if (VS_SIZE(vsp) != size)
511 514 panic("vmem_hash_delete(%p, %lx, %lu): wrong size (expect %lu)",
512 515 (void *)vmp, addr, size, VS_SIZE(vsp));
513 516
514 517 vmp->vm_kstat.vk_free.value.ui64++;
515 518 vmp->vm_kstat.vk_mem_inuse.value.ui64 -= size;
516 519
517 520 return (vsp);
518 521 }
519 522
520 523 /*
521 524 * Create a segment spanning the range [start, end) and add it to the arena.
522 525 */
523 526 static vmem_seg_t *
524 527 vmem_seg_create(vmem_t *vmp, vmem_seg_t *vprev, uintptr_t start, uintptr_t end)
525 528 {
526 529 vmem_seg_t *newseg = vmem_getseg(vmp);
527 530
528 531 newseg->vs_start = start;
529 532 newseg->vs_end = end;
530 533 newseg->vs_type = 0;
531 534 newseg->vs_import = 0;
532 535
533 536 VMEM_INSERT(vprev, newseg, a);
534 537
535 538 return (newseg);
536 539 }
537 540
538 541 /*
539 542 * Remove segment vsp from the arena.
540 543 */
541 544 static void
542 545 vmem_seg_destroy(vmem_t *vmp, vmem_seg_t *vsp)
543 546 {
544 547 ASSERT(vsp->vs_type != VMEM_ROTOR);
545 548 VMEM_DELETE(vsp, a);
546 549
547 550 vmem_putseg(vmp, vsp);
548 551 }
549 552
550 553 /*
551 554 * Add the span [vaddr, vaddr + size) to vmp and update kstats.
552 555 */
553 556 static vmem_seg_t *
554 557 vmem_span_create(vmem_t *vmp, void *vaddr, size_t size, uint8_t import)
555 558 {
556 559 vmem_seg_t *newseg, *span;
557 560 uintptr_t start = (uintptr_t)vaddr;
558 561 uintptr_t end = start + size;
559 562
560 563 ASSERT(MUTEX_HELD(&vmp->vm_lock));
561 564
562 565 if ((start | end) & (vmp->vm_quantum - 1))
563 566 panic("vmem_span_create(%p, %p, %lu): misaligned",
564 567 (void *)vmp, vaddr, size);
565 568
566 569 span = vmem_seg_create(vmp, vmp->vm_seg0.vs_aprev, start, end);
567 570 span->vs_type = VMEM_SPAN;
568 571 span->vs_import = import;
569 572 VMEM_INSERT(vmp->vm_seg0.vs_kprev, span, k);
570 573
571 574 newseg = vmem_seg_create(vmp, span, start, end);
572 575 vmem_freelist_insert(vmp, newseg);
573 576
574 577 if (import)
575 578 vmp->vm_kstat.vk_mem_import.value.ui64 += size;
576 579 vmp->vm_kstat.vk_mem_total.value.ui64 += size;
577 580
578 581 return (newseg);
579 582 }
580 583
581 584 /*
582 585 * Remove span vsp from vmp and update kstats.
583 586 */
584 587 static void
585 588 vmem_span_destroy(vmem_t *vmp, vmem_seg_t *vsp)
586 589 {
587 590 vmem_seg_t *span = vsp->vs_aprev;
588 591 size_t size = VS_SIZE(vsp);
589 592
590 593 ASSERT(MUTEX_HELD(&vmp->vm_lock));
591 594 ASSERT(span->vs_type == VMEM_SPAN);
592 595
593 596 if (span->vs_import)
594 597 vmp->vm_kstat.vk_mem_import.value.ui64 -= size;
595 598 vmp->vm_kstat.vk_mem_total.value.ui64 -= size;
596 599
597 600 VMEM_DELETE(span, k);
598 601
599 602 vmem_seg_destroy(vmp, vsp);
600 603 vmem_seg_destroy(vmp, span);
601 604 }
602 605
603 606 /*
604 607 * Allocate the subrange [addr, addr + size) from segment vsp.
605 608 * If there are leftovers on either side, place them on the freelist.
606 609 * Returns a pointer to the segment representing [addr, addr + size).
607 610 */
608 611 static vmem_seg_t *
609 612 vmem_seg_alloc(vmem_t *vmp, vmem_seg_t *vsp, uintptr_t addr, size_t size)
610 613 {
611 614 uintptr_t vs_start = vsp->vs_start;
612 615 uintptr_t vs_end = vsp->vs_end;
613 616 size_t vs_size = vs_end - vs_start;
614 617 size_t realsize = P2ROUNDUP(size, vmp->vm_quantum);
615 618 uintptr_t addr_end = addr + realsize;
616 619
617 620 ASSERT(P2PHASE(vs_start, vmp->vm_quantum) == 0);
618 621 ASSERT(P2PHASE(addr, vmp->vm_quantum) == 0);
619 622 ASSERT(vsp->vs_type == VMEM_FREE);
620 623 ASSERT(addr >= vs_start && addr_end - 1 <= vs_end - 1);
621 624 ASSERT(addr - 1 <= addr_end - 1);
622 625
623 626 /*
624 627 * If we're allocating from the start of the segment, and the
625 628 * remainder will be on the same freelist, we can save quite
626 629 * a bit of work.
627 630 */
628 631 if (P2SAMEHIGHBIT(vs_size, vs_size - realsize) && addr == vs_start) {
629 632 ASSERT(highbit(vs_size) == highbit(vs_size - realsize));
630 633 vsp->vs_start = addr_end;
631 634 vsp = vmem_seg_create(vmp, vsp->vs_aprev, addr, addr + size);
632 635 vmem_hash_insert(vmp, vsp);
633 636 return (vsp);
634 637 }
635 638
636 639 vmem_freelist_delete(vmp, vsp);
637 640
638 641 if (vs_end != addr_end)
639 642 vmem_freelist_insert(vmp,
640 643 vmem_seg_create(vmp, vsp, addr_end, vs_end));
641 644
642 645 if (vs_start != addr)
643 646 vmem_freelist_insert(vmp,
644 647 vmem_seg_create(vmp, vsp->vs_aprev, vs_start, addr));
645 648
646 649 vsp->vs_start = addr;
647 650 vsp->vs_end = addr + size;
648 651
649 652 vmem_hash_insert(vmp, vsp);
650 653 return (vsp);
651 654 }
652 655
653 656 /*
654 657 * Returns 1 if we are populating, 0 otherwise.
655 658 * Call it if we want to prevent recursion from HAT.
656 659 */
657 660 int
658 661 vmem_is_populator()
659 662 {
660 663 return (mutex_owner(&vmem_sleep_lock) == curthread ||
661 664 mutex_owner(&vmem_nosleep_lock) == curthread ||
662 665 mutex_owner(&vmem_pushpage_lock) == curthread ||
663 666 mutex_owner(&vmem_panic_lock) == curthread);
664 667 }
665 668
666 669 /*
667 670 * Populate vmp's segfree list with VMEM_MINFREE vmem_seg_t structures.
668 671 */
669 672 static int
670 673 vmem_populate(vmem_t *vmp, int vmflag)
671 674 {
672 675 char *p;
673 676 vmem_seg_t *vsp;
674 677 ssize_t nseg;
675 678 size_t size;
676 679 kmutex_t *lp;
677 680 int i;
678 681
679 682 while (vmp->vm_nsegfree < VMEM_MINFREE &&
680 683 (vsp = vmem_getseg_global()) != NULL)
681 684 vmem_putseg(vmp, vsp);
682 685
683 686 if (vmp->vm_nsegfree >= VMEM_MINFREE)
684 687 return (1);
685 688
686 689 /*
687 690 * If we're already populating, tap the reserve.
688 691 */
689 692 if (vmem_is_populator()) {
690 693 ASSERT(vmp->vm_cflags & VMC_POPULATOR);
691 694 return (1);
692 695 }
693 696
694 697 mutex_exit(&vmp->vm_lock);
695 698
696 699 if (panic_thread == curthread)
697 700 lp = &vmem_panic_lock;
698 701 else if (vmflag & VM_NOSLEEP)
699 702 lp = &vmem_nosleep_lock;
700 703 else if (vmflag & VM_PUSHPAGE)
701 704 lp = &vmem_pushpage_lock;
702 705 else
703 706 lp = &vmem_sleep_lock;
704 707
705 708 mutex_enter(lp);
706 709
707 710 nseg = VMEM_MINFREE + vmem_populators * VMEM_POPULATE_RESERVE;
708 711 size = P2ROUNDUP(nseg * vmem_seg_size, vmem_seg_arena->vm_quantum);
709 712 nseg = size / vmem_seg_size;
710 713
711 714 /*
712 715 * The following vmem_alloc() may need to populate vmem_seg_arena
713 716 * and all the things it imports from. When doing so, it will tap
714 717 * each arena's reserve to prevent recursion (see the block comment
715 718 * above the definition of VMEM_POPULATE_RESERVE).
716 719 */
717 720 p = vmem_alloc(vmem_seg_arena, size, vmflag & VM_KMFLAGS);
718 721 if (p == NULL) {
719 722 mutex_exit(lp);
720 723 mutex_enter(&vmp->vm_lock);
721 724 vmp->vm_kstat.vk_populate_fail.value.ui64++;
722 725 return (0);
723 726 }
724 727
725 728 /*
726 729 * Restock the arenas that may have been depleted during population.
727 730 */
728 731 for (i = 0; i < vmem_populators; i++) {
729 732 mutex_enter(&vmem_populator[i]->vm_lock);
730 733 while (vmem_populator[i]->vm_nsegfree < VMEM_POPULATE_RESERVE)
731 734 vmem_putseg(vmem_populator[i],
732 735 (vmem_seg_t *)(p + --nseg * vmem_seg_size));
733 736 mutex_exit(&vmem_populator[i]->vm_lock);
734 737 }
735 738
736 739 mutex_exit(lp);
737 740 mutex_enter(&vmp->vm_lock);
738 741
739 742 /*
740 743 * Now take our own segments.
741 744 */
742 745 ASSERT(nseg >= VMEM_MINFREE);
743 746 while (vmp->vm_nsegfree < VMEM_MINFREE)
744 747 vmem_putseg(vmp, (vmem_seg_t *)(p + --nseg * vmem_seg_size));
745 748
746 749 /*
747 750 * Give the remainder to charity.
748 751 */
749 752 while (nseg > 0)
750 753 vmem_putseg_global((vmem_seg_t *)(p + --nseg * vmem_seg_size));
751 754
752 755 return (1);
753 756 }
754 757
755 758 /*
756 759 * Advance a walker from its previous position to 'afterme'.
757 760 * Note: may drop and reacquire vmp->vm_lock.
758 761 */
759 762 static void
760 763 vmem_advance(vmem_t *vmp, vmem_seg_t *walker, vmem_seg_t *afterme)
761 764 {
762 765 vmem_seg_t *vprev = walker->vs_aprev;
763 766 vmem_seg_t *vnext = walker->vs_anext;
764 767 vmem_seg_t *vsp = NULL;
765 768
766 769 VMEM_DELETE(walker, a);
767 770
768 771 if (afterme != NULL)
769 772 VMEM_INSERT(afterme, walker, a);
770 773
771 774 /*
772 775 * The walker segment's presence may have prevented its neighbors
773 776 * from coalescing. If so, coalesce them now.
774 777 */
775 778 if (vprev->vs_type == VMEM_FREE) {
776 779 if (vnext->vs_type == VMEM_FREE) {
777 780 ASSERT(vprev->vs_end == vnext->vs_start);
778 781 vmem_freelist_delete(vmp, vnext);
779 782 vmem_freelist_delete(vmp, vprev);
780 783 vprev->vs_end = vnext->vs_end;
781 784 vmem_freelist_insert(vmp, vprev);
782 785 vmem_seg_destroy(vmp, vnext);
783 786 }
784 787 vsp = vprev;
785 788 } else if (vnext->vs_type == VMEM_FREE) {
786 789 vsp = vnext;
787 790 }
788 791
789 792 /*
790 793 * vsp could represent a complete imported span,
791 794 * in which case we must return it to the source.
792 795 */
793 796 if (vsp != NULL && vsp->vs_aprev->vs_import &&
794 797 vmp->vm_source_free != NULL &&
795 798 vsp->vs_aprev->vs_type == VMEM_SPAN &&
796 799 vsp->vs_anext->vs_type == VMEM_SPAN) {
797 800 void *vaddr = (void *)vsp->vs_start;
798 801 size_t size = VS_SIZE(vsp);
799 802 ASSERT(size == VS_SIZE(vsp->vs_aprev));
800 803 vmem_freelist_delete(vmp, vsp);
801 804 vmem_span_destroy(vmp, vsp);
802 805 mutex_exit(&vmp->vm_lock);
803 806 vmp->vm_source_free(vmp->vm_source, vaddr, size);
804 807 mutex_enter(&vmp->vm_lock);
805 808 }
806 809 }
807 810
808 811 /*
809 812 * VM_NEXTFIT allocations deliberately cycle through all virtual addresses
810 813 * in an arena, so that we avoid reusing addresses for as long as possible.
811 814 * This helps to catch used-after-freed bugs. It's also the perfect policy
812 815 * for allocating things like process IDs, where we want to cycle through
813 816 * all values in order.
814 817 */
815 818 static void *
816 819 vmem_nextfit_alloc(vmem_t *vmp, size_t size, int vmflag)
817 820 {
818 821 vmem_seg_t *vsp, *rotor;
819 822 uintptr_t addr;
820 823 size_t realsize = P2ROUNDUP(size, vmp->vm_quantum);
821 824 size_t vs_size;
822 825
823 826 mutex_enter(&vmp->vm_lock);
824 827
825 828 if (vmp->vm_nsegfree < VMEM_MINFREE && !vmem_populate(vmp, vmflag)) {
826 829 mutex_exit(&vmp->vm_lock);
827 830 return (NULL);
828 831 }
829 832
830 833 /*
831 834 * The common case is that the segment right after the rotor is free,
832 835 * and large enough that extracting 'size' bytes won't change which
833 836 * freelist it's on. In this case we can avoid a *lot* of work.
834 837 * Instead of the normal vmem_seg_alloc(), we just advance the start
835 838 * address of the victim segment. Instead of moving the rotor, we
836 839 * create the new segment structure *behind the rotor*, which has
837 840 * the same effect. And finally, we know we don't have to coalesce
838 841 * the rotor's neighbors because the new segment lies between them.
839 842 */
840 843 rotor = &vmp->vm_rotor;
841 844 vsp = rotor->vs_anext;
842 845 if (vsp->vs_type == VMEM_FREE && (vs_size = VS_SIZE(vsp)) > realsize &&
843 846 P2SAMEHIGHBIT(vs_size, vs_size - realsize)) {
844 847 ASSERT(highbit(vs_size) == highbit(vs_size - realsize));
845 848 addr = vsp->vs_start;
846 849 vsp->vs_start = addr + realsize;
847 850 vmem_hash_insert(vmp,
848 851 vmem_seg_create(vmp, rotor->vs_aprev, addr, addr + size));
849 852 mutex_exit(&vmp->vm_lock);
850 853 return ((void *)addr);
851 854 }
852 855
853 856 /*
854 857 * Starting at the rotor, look for a segment large enough to
855 858 * satisfy the allocation.
856 859 */
857 860 for (;;) {
858 861 vmp->vm_kstat.vk_search.value.ui64++;
859 862 if (vsp->vs_type == VMEM_FREE && VS_SIZE(vsp) >= size)
860 863 break;
861 864 vsp = vsp->vs_anext;
862 865 if (vsp == rotor) {
863 866 /*
864 867 * We've come full circle. One possibility is that the
865 868 * there's actually enough space, but the rotor itself
866 869 * is preventing the allocation from succeeding because
867 870 * it's sitting between two free segments. Therefore,
868 871 * we advance the rotor and see if that liberates a
869 872 * suitable segment.
870 873 */
871 874 vmem_advance(vmp, rotor, rotor->vs_anext);
872 875 vsp = rotor->vs_aprev;
873 876 if (vsp->vs_type == VMEM_FREE && VS_SIZE(vsp) >= size)
874 877 break;
875 878 /*
876 879 * If there's a lower arena we can import from, or it's
877 880 * a VM_NOSLEEP allocation, let vmem_xalloc() handle it.
878 881 * Otherwise, wait until another thread frees something.
879 882 */
880 883 if (vmp->vm_source_alloc != NULL ||
881 884 (vmflag & VM_NOSLEEP)) {
882 885 mutex_exit(&vmp->vm_lock);
883 886 return (vmem_xalloc(vmp, size, vmp->vm_quantum,
884 887 0, 0, NULL, NULL, vmflag & VM_KMFLAGS));
885 888 }
886 889 vmp->vm_kstat.vk_wait.value.ui64++;
887 890 cv_wait(&vmp->vm_cv, &vmp->vm_lock);
888 891 vsp = rotor->vs_anext;
889 892 }
890 893 }
891 894
892 895 /*
893 896 * We found a segment. Extract enough space to satisfy the allocation.
894 897 */
895 898 addr = vsp->vs_start;
896 899 vsp = vmem_seg_alloc(vmp, vsp, addr, size);
897 900 ASSERT(vsp->vs_type == VMEM_ALLOC &&
898 901 vsp->vs_start == addr && vsp->vs_end == addr + size);
899 902
900 903 /*
901 904 * Advance the rotor to right after the newly-allocated segment.
902 905 * That's where the next VM_NEXTFIT allocation will begin searching.
903 906 */
904 907 vmem_advance(vmp, rotor, vsp);
905 908 mutex_exit(&vmp->vm_lock);
906 909 return ((void *)addr);
907 910 }
908 911
909 912 /*
910 913 * Checks if vmp is guaranteed to have a size-byte buffer somewhere on its
911 914 * freelist. If size is not a power-of-2, it can return a false-negative.
912 915 *
913 916 * Used to decide if a newly imported span is superfluous after re-acquiring
914 917 * the arena lock.
915 918 */
916 919 static int
917 920 vmem_canalloc(vmem_t *vmp, size_t size)
918 921 {
919 922 int hb;
920 923 int flist = 0;
921 924 ASSERT(MUTEX_HELD(&vmp->vm_lock));
922 925
923 926 if (ISP2(size))
924 927 flist = lowbit(P2ALIGN(vmp->vm_freemap, size));
925 928 else if ((hb = highbit(size)) < VMEM_FREELISTS)
926 929 flist = lowbit(P2ALIGN(vmp->vm_freemap, 1UL << hb));
927 930
928 931 return (flist);
929 932 }
930 933
931 934 /*
932 935 * Allocate size bytes at offset phase from an align boundary such that the
933 936 * resulting segment [addr, addr + size) is a subset of [minaddr, maxaddr)
934 937 * that does not straddle a nocross-aligned boundary.
935 938 */
936 939 void *
937 940 vmem_xalloc(vmem_t *vmp, size_t size, size_t align_arg, size_t phase,
938 941 size_t nocross, void *minaddr, void *maxaddr, int vmflag)
939 942 {
940 943 vmem_seg_t *vsp;
941 944 vmem_seg_t *vbest = NULL;
942 945 uintptr_t addr, taddr, start, end;
943 946 uintptr_t align = (align_arg != 0) ? align_arg : vmp->vm_quantum;
944 947 void *vaddr, *xvaddr = NULL;
945 948 size_t xsize;
946 949 int hb, flist, resv;
947 950 uint32_t mtbf;
948 951
949 952 if ((align | phase | nocross) & (vmp->vm_quantum - 1))
950 953 panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
951 954 "parameters not vm_quantum aligned",
952 955 (void *)vmp, size, align_arg, phase, nocross,
953 956 minaddr, maxaddr, vmflag);
954 957
955 958 if (nocross != 0 &&
956 959 (align > nocross || P2ROUNDUP(phase + size, align) > nocross))
957 960 panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
958 961 "overconstrained allocation",
959 962 (void *)vmp, size, align_arg, phase, nocross,
960 963 minaddr, maxaddr, vmflag);
961 964
962 965 if (phase >= align || !ISP2(align) || !ISP2(nocross))
963 966 panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
964 967 "parameters inconsistent or invalid",
965 968 (void *)vmp, size, align_arg, phase, nocross,
966 969 minaddr, maxaddr, vmflag);
967 970
968 971 if ((mtbf = vmem_mtbf | vmp->vm_mtbf) != 0 && gethrtime() % mtbf == 0 &&
969 972 (vmflag & (VM_NOSLEEP | VM_PANIC)) == VM_NOSLEEP)
970 973 return (NULL);
971 974
972 975 mutex_enter(&vmp->vm_lock);
973 976 for (;;) {
974 977 if (vmp->vm_nsegfree < VMEM_MINFREE &&
975 978 !vmem_populate(vmp, vmflag))
976 979 break;
977 980 do_alloc:
978 981 /*
979 982 * highbit() returns the highest bit + 1, which is exactly
980 983 * what we want: we want to search the first freelist whose
981 984 * members are *definitely* large enough to satisfy our
982 985 * allocation. However, there are certain cases in which we
983 986 * want to look at the next-smallest freelist (which *might*
984 987 * be able to satisfy the allocation):
985 988 *
986 989 * (1) The size is exactly a power of 2, in which case
987 990 * the smaller freelist is always big enough;
988 991 *
989 992 * (2) All other freelists are empty;
990 993 *
991 994 * (3) We're in the highest possible freelist, which is
992 995 * always empty (e.g. the 4GB freelist on 32-bit systems);
993 996 *
994 997 * (4) We're doing a best-fit or first-fit allocation.
995 998 */
996 999 if (ISP2(size)) {
997 1000 flist = lowbit(P2ALIGN(vmp->vm_freemap, size));
998 1001 } else {
999 1002 hb = highbit(size);
1000 1003 if ((vmp->vm_freemap >> hb) == 0 ||
1001 1004 hb == VMEM_FREELISTS ||
1002 1005 (vmflag & (VM_BESTFIT | VM_FIRSTFIT)))
1003 1006 hb--;
1004 1007 flist = lowbit(P2ALIGN(vmp->vm_freemap, 1UL << hb));
1005 1008 }
1006 1009
1007 1010 for (vbest = NULL, vsp = (flist == 0) ? NULL :
1008 1011 vmp->vm_freelist[flist - 1].vs_knext;
1009 1012 vsp != NULL; vsp = vsp->vs_knext) {
1010 1013 vmp->vm_kstat.vk_search.value.ui64++;
1011 1014 if (vsp->vs_start == 0) {
1012 1015 /*
1013 1016 * We're moving up to a larger freelist,
1014 1017 * so if we've already found a candidate,
1015 1018 * the fit can't possibly get any better.
1016 1019 */
1017 1020 if (vbest != NULL)
1018 1021 break;
1019 1022 /*
1020 1023 * Find the next non-empty freelist.
1021 1024 */
1022 1025 flist = lowbit(P2ALIGN(vmp->vm_freemap,
1023 1026 VS_SIZE(vsp)));
1024 1027 if (flist-- == 0)
1025 1028 break;
1026 1029 vsp = (vmem_seg_t *)&vmp->vm_freelist[flist];
1027 1030 ASSERT(vsp->vs_knext->vs_type == VMEM_FREE);
1028 1031 continue;
1029 1032 }
1030 1033 if (vsp->vs_end - 1 < (uintptr_t)minaddr)
1031 1034 continue;
1032 1035 if (vsp->vs_start > (uintptr_t)maxaddr - 1)
1033 1036 continue;
1034 1037 start = MAX(vsp->vs_start, (uintptr_t)minaddr);
1035 1038 end = MIN(vsp->vs_end - 1, (uintptr_t)maxaddr - 1) + 1;
1036 1039 taddr = P2PHASEUP(start, align, phase);
1037 1040 if (P2BOUNDARY(taddr, size, nocross))
1038 1041 taddr +=
1039 1042 P2ROUNDUP(P2NPHASE(taddr, nocross), align);
1040 1043 if ((taddr - start) + size > end - start ||
1041 1044 (vbest != NULL && VS_SIZE(vsp) >= VS_SIZE(vbest)))
1042 1045 continue;
1043 1046 vbest = vsp;
1044 1047 addr = taddr;
1045 1048 if (!(vmflag & VM_BESTFIT) || VS_SIZE(vbest) == size)
1046 1049 break;
1047 1050 }
1048 1051 if (vbest != NULL)
1049 1052 break;
1050 1053 ASSERT(xvaddr == NULL);
1051 1054 if (size == 0)
1052 1055 panic("vmem_xalloc(): size == 0");
1053 1056 if (vmp->vm_source_alloc != NULL && nocross == 0 &&
1054 1057 minaddr == NULL && maxaddr == NULL) {
1055 1058 size_t aneeded, asize;
1056 1059 size_t aquantum = MAX(vmp->vm_quantum,
1057 1060 vmp->vm_source->vm_quantum);
1058 1061 size_t aphase = phase;
1059 1062 if ((align > aquantum) &&
1060 1063 !(vmp->vm_cflags & VMC_XALIGN)) {
1061 1064 aphase = (P2PHASE(phase, aquantum) != 0) ?
1062 1065 align - vmp->vm_quantum : align - aquantum;
1063 1066 ASSERT(aphase >= phase);
1064 1067 }
1065 1068 aneeded = MAX(size + aphase, vmp->vm_min_import);
1066 1069 asize = P2ROUNDUP(aneeded, aquantum);
1067 1070
1068 1071 if (asize < size) {
1069 1072 /*
1070 1073 * The rounding induced overflow; return NULL
1071 1074 * if we are permitted to fail the allocation
1072 1075 * (and explicitly panic if we aren't).
1073 1076 */
1074 1077 if ((vmflag & VM_NOSLEEP) &&
1075 1078 !(vmflag & VM_PANIC)) {
1076 1079 mutex_exit(&vmp->vm_lock);
1077 1080 return (NULL);
1078 1081 }
1079 1082
1080 1083 panic("vmem_xalloc(): size overflow");
1081 1084 }
1082 1085
1083 1086 /*
1084 1087 * Determine how many segment structures we'll consume.
1085 1088 * The calculation must be precise because if we're
1086 1089 * here on behalf of vmem_populate(), we are taking
1087 1090 * segments from a very limited reserve.
1088 1091 */
1089 1092 if (size == asize && !(vmp->vm_cflags & VMC_XALLOC))
1090 1093 resv = VMEM_SEGS_PER_SPAN_CREATE +
1091 1094 VMEM_SEGS_PER_EXACT_ALLOC;
1092 1095 else if (phase == 0 &&
1093 1096 align <= vmp->vm_source->vm_quantum)
1094 1097 resv = VMEM_SEGS_PER_SPAN_CREATE +
1095 1098 VMEM_SEGS_PER_LEFT_ALLOC;
1096 1099 else
1097 1100 resv = VMEM_SEGS_PER_ALLOC_MAX;
1098 1101
1099 1102 ASSERT(vmp->vm_nsegfree >= resv);
1100 1103 vmp->vm_nsegfree -= resv; /* reserve our segs */
1101 1104 mutex_exit(&vmp->vm_lock);
1102 1105 if (vmp->vm_cflags & VMC_XALLOC) {
1103 1106 size_t oasize = asize;
1104 1107 vaddr = ((vmem_ximport_t *)
1105 1108 vmp->vm_source_alloc)(vmp->vm_source,
1106 1109 &asize, align, vmflag & VM_KMFLAGS);
1107 1110 ASSERT(asize >= oasize);
1108 1111 ASSERT(P2PHASE(asize,
1109 1112 vmp->vm_source->vm_quantum) == 0);
1110 1113 ASSERT(!(vmp->vm_cflags & VMC_XALIGN) ||
1111 1114 IS_P2ALIGNED(vaddr, align));
1112 1115 } else {
1113 1116 vaddr = vmp->vm_source_alloc(vmp->vm_source,
1114 1117 asize, vmflag & VM_KMFLAGS);
1115 1118 }
1116 1119 mutex_enter(&vmp->vm_lock);
1117 1120 vmp->vm_nsegfree += resv; /* claim reservation */
1118 1121 aneeded = size + align - vmp->vm_quantum;
1119 1122 aneeded = P2ROUNDUP(aneeded, vmp->vm_quantum);
1120 1123 if (vaddr != NULL) {
1121 1124 /*
1122 1125 * Since we dropped the vmem lock while
1123 1126 * calling the import function, other
1124 1127 * threads could have imported space
1125 1128 * and made our import unnecessary. In
1126 1129 * order to save space, we return
1127 1130 * excess imports immediately.
1128 1131 */
1129 1132 if (asize > aneeded &&
1130 1133 vmp->vm_source_free != NULL &&
1131 1134 vmem_canalloc(vmp, aneeded)) {
1132 1135 ASSERT(resv >=
1133 1136 VMEM_SEGS_PER_MIDDLE_ALLOC);
1134 1137 xvaddr = vaddr;
1135 1138 xsize = asize;
1136 1139 goto do_alloc;
1137 1140 }
1138 1141 vbest = vmem_span_create(vmp, vaddr, asize, 1);
1139 1142 addr = P2PHASEUP(vbest->vs_start, align, phase);
1140 1143 break;
1141 1144 } else if (vmem_canalloc(vmp, aneeded)) {
1142 1145 /*
1143 1146 * Our import failed, but another thread
1144 1147 * added sufficient free memory to the arena
1145 1148 * to satisfy our request. Go back and
1146 1149 * grab it.
1147 1150 */
1148 1151 ASSERT(resv >= VMEM_SEGS_PER_MIDDLE_ALLOC);
1149 1152 goto do_alloc;
1150 1153 }
1151 1154 }
1152 1155
1153 1156 /*
1154 1157 * If the requestor chooses to fail the allocation attempt
1155 1158 * rather than reap wait and retry - get out of the loop.
1156 1159 */
1157 1160 if (vmflag & VM_ABORT)
1158 1161 break;
1159 1162 mutex_exit(&vmp->vm_lock);
1160 1163 if (vmp->vm_cflags & VMC_IDENTIFIER)
1161 1164 kmem_reap_idspace();
1162 1165 else
1163 1166 kmem_reap();
1164 1167 mutex_enter(&vmp->vm_lock);
1165 1168 if (vmflag & VM_NOSLEEP)
1166 1169 break;
1167 1170 vmp->vm_kstat.vk_wait.value.ui64++;
1168 1171 cv_wait(&vmp->vm_cv, &vmp->vm_lock);
1169 1172 }
1170 1173 if (vbest != NULL) {
1171 1174 ASSERT(vbest->vs_type == VMEM_FREE);
1172 1175 ASSERT(vbest->vs_knext != vbest);
1173 1176 /* re-position to end of buffer */
1174 1177 if (vmflag & VM_ENDALLOC) {
1175 1178 addr += ((vbest->vs_end - (addr + size)) / align) *
1176 1179 align;
1177 1180 }
1178 1181 (void) vmem_seg_alloc(vmp, vbest, addr, size);
1179 1182 mutex_exit(&vmp->vm_lock);
1180 1183 if (xvaddr)
1181 1184 vmp->vm_source_free(vmp->vm_source, xvaddr, xsize);
1182 1185 ASSERT(P2PHASE(addr, align) == phase);
1183 1186 ASSERT(!P2BOUNDARY(addr, size, nocross));
1184 1187 ASSERT(addr >= (uintptr_t)minaddr);
1185 1188 ASSERT(addr + size - 1 <= (uintptr_t)maxaddr - 1);
1186 1189 return ((void *)addr);
1187 1190 }
1188 1191 vmp->vm_kstat.vk_fail.value.ui64++;
1189 1192 mutex_exit(&vmp->vm_lock);
1190 1193 if (vmflag & VM_PANIC)
1191 1194 panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
1192 1195 "cannot satisfy mandatory allocation",
1193 1196 (void *)vmp, size, align_arg, phase, nocross,
1194 1197 minaddr, maxaddr, vmflag);
1195 1198 ASSERT(xvaddr == NULL);
1196 1199 return (NULL);
1197 1200 }
1198 1201
1199 1202 /*
1200 1203 * Free the segment [vaddr, vaddr + size), where vaddr was a constrained
1201 1204 * allocation. vmem_xalloc() and vmem_xfree() must always be paired because
1202 1205 * both routines bypass the quantum caches.
1203 1206 */
1204 1207 void
1205 1208 vmem_xfree(vmem_t *vmp, void *vaddr, size_t size)
1206 1209 {
1207 1210 vmem_seg_t *vsp, *vnext, *vprev;
1208 1211
1209 1212 mutex_enter(&vmp->vm_lock);
1210 1213
1211 1214 vsp = vmem_hash_delete(vmp, (uintptr_t)vaddr, size);
1212 1215 vsp->vs_end = P2ROUNDUP(vsp->vs_end, vmp->vm_quantum);
1213 1216
1214 1217 /*
1215 1218 * Attempt to coalesce with the next segment.
1216 1219 */
1217 1220 vnext = vsp->vs_anext;
1218 1221 if (vnext->vs_type == VMEM_FREE) {
1219 1222 ASSERT(vsp->vs_end == vnext->vs_start);
1220 1223 vmem_freelist_delete(vmp, vnext);
1221 1224 vsp->vs_end = vnext->vs_end;
1222 1225 vmem_seg_destroy(vmp, vnext);
1223 1226 }
1224 1227
1225 1228 /*
1226 1229 * Attempt to coalesce with the previous segment.
1227 1230 */
1228 1231 vprev = vsp->vs_aprev;
1229 1232 if (vprev->vs_type == VMEM_FREE) {
1230 1233 ASSERT(vprev->vs_end == vsp->vs_start);
1231 1234 vmem_freelist_delete(vmp, vprev);
1232 1235 vprev->vs_end = vsp->vs_end;
1233 1236 vmem_seg_destroy(vmp, vsp);
1234 1237 vsp = vprev;
1235 1238 }
1236 1239
1237 1240 /*
1238 1241 * If the entire span is free, return it to the source.
1239 1242 */
1240 1243 if (vsp->vs_aprev->vs_import && vmp->vm_source_free != NULL &&
1241 1244 vsp->vs_aprev->vs_type == VMEM_SPAN &&
1242 1245 vsp->vs_anext->vs_type == VMEM_SPAN) {
1243 1246 vaddr = (void *)vsp->vs_start;
1244 1247 size = VS_SIZE(vsp);
1245 1248 ASSERT(size == VS_SIZE(vsp->vs_aprev));
1246 1249 vmem_span_destroy(vmp, vsp);
1247 1250 mutex_exit(&vmp->vm_lock);
1248 1251 vmp->vm_source_free(vmp->vm_source, vaddr, size);
1249 1252 } else {
1250 1253 vmem_freelist_insert(vmp, vsp);
1251 1254 mutex_exit(&vmp->vm_lock);
1252 1255 }
1253 1256 }
1254 1257
1255 1258 /*
1256 1259 * Allocate size bytes from arena vmp. Returns the allocated address
1257 1260 * on success, NULL on failure. vmflag specifies VM_SLEEP or VM_NOSLEEP,
1258 1261 * and may also specify best-fit, first-fit, or next-fit allocation policy
1259 1262 * instead of the default instant-fit policy. VM_SLEEP allocations are
1260 1263 * guaranteed to succeed.
1261 1264 */
1262 1265 void *
1263 1266 vmem_alloc(vmem_t *vmp, size_t size, int vmflag)
1264 1267 {
1265 1268 vmem_seg_t *vsp;
1266 1269 uintptr_t addr;
1267 1270 int hb;
1268 1271 int flist = 0;
1269 1272 uint32_t mtbf;
1270 1273
1271 1274 if (size - 1 < vmp->vm_qcache_max)
1272 1275 return (kmem_cache_alloc(vmp->vm_qcache[(size - 1) >>
1273 1276 vmp->vm_qshift], vmflag & VM_KMFLAGS));
1274 1277
1275 1278 if ((mtbf = vmem_mtbf | vmp->vm_mtbf) != 0 && gethrtime() % mtbf == 0 &&
1276 1279 (vmflag & (VM_NOSLEEP | VM_PANIC)) == VM_NOSLEEP)
1277 1280 return (NULL);
1278 1281
1279 1282 if (vmflag & VM_NEXTFIT)
1280 1283 return (vmem_nextfit_alloc(vmp, size, vmflag));
1281 1284
1282 1285 if (vmflag & (VM_BESTFIT | VM_FIRSTFIT))
1283 1286 return (vmem_xalloc(vmp, size, vmp->vm_quantum, 0, 0,
1284 1287 NULL, NULL, vmflag));
1285 1288
1286 1289 /*
1287 1290 * Unconstrained instant-fit allocation from the segment list.
1288 1291 */
1289 1292 mutex_enter(&vmp->vm_lock);
1290 1293
1291 1294 if (vmp->vm_nsegfree >= VMEM_MINFREE || vmem_populate(vmp, vmflag)) {
1292 1295 if (ISP2(size))
1293 1296 flist = lowbit(P2ALIGN(vmp->vm_freemap, size));
1294 1297 else if ((hb = highbit(size)) < VMEM_FREELISTS)
1295 1298 flist = lowbit(P2ALIGN(vmp->vm_freemap, 1UL << hb));
1296 1299 }
1297 1300
1298 1301 if (flist-- == 0) {
1299 1302 mutex_exit(&vmp->vm_lock);
1300 1303 return (vmem_xalloc(vmp, size, vmp->vm_quantum,
1301 1304 0, 0, NULL, NULL, vmflag));
1302 1305 }
1303 1306
1304 1307 ASSERT(size <= (1UL << flist));
1305 1308 vsp = vmp->vm_freelist[flist].vs_knext;
1306 1309 addr = vsp->vs_start;
1307 1310 if (vmflag & VM_ENDALLOC) {
1308 1311 addr += vsp->vs_end - (addr + size);
1309 1312 }
1310 1313 (void) vmem_seg_alloc(vmp, vsp, addr, size);
1311 1314 mutex_exit(&vmp->vm_lock);
1312 1315 return ((void *)addr);
1313 1316 }
1314 1317
1315 1318 /*
1316 1319 * Free the segment [vaddr, vaddr + size).
1317 1320 */
1318 1321 void
1319 1322 vmem_free(vmem_t *vmp, void *vaddr, size_t size)
1320 1323 {
1321 1324 if (size - 1 < vmp->vm_qcache_max)
1322 1325 kmem_cache_free(vmp->vm_qcache[(size - 1) >> vmp->vm_qshift],
1323 1326 vaddr);
1324 1327 else
1325 1328 vmem_xfree(vmp, vaddr, size);
1326 1329 }
1327 1330
1328 1331 /*
1329 1332 * Determine whether arena vmp contains the segment [vaddr, vaddr + size).
1330 1333 */
1331 1334 int
1332 1335 vmem_contains(vmem_t *vmp, void *vaddr, size_t size)
1333 1336 {
1334 1337 uintptr_t start = (uintptr_t)vaddr;
1335 1338 uintptr_t end = start + size;
1336 1339 vmem_seg_t *vsp;
1337 1340 vmem_seg_t *seg0 = &vmp->vm_seg0;
1338 1341
1339 1342 mutex_enter(&vmp->vm_lock);
1340 1343 vmp->vm_kstat.vk_contains.value.ui64++;
1341 1344 for (vsp = seg0->vs_knext; vsp != seg0; vsp = vsp->vs_knext) {
1342 1345 vmp->vm_kstat.vk_contains_search.value.ui64++;
1343 1346 ASSERT(vsp->vs_type == VMEM_SPAN);
1344 1347 if (start >= vsp->vs_start && end - 1 <= vsp->vs_end - 1)
1345 1348 break;
1346 1349 }
1347 1350 mutex_exit(&vmp->vm_lock);
1348 1351 return (vsp != seg0);
1349 1352 }
1350 1353
1351 1354 /*
1352 1355 * Add the span [vaddr, vaddr + size) to arena vmp.
1353 1356 */
1354 1357 void *
1355 1358 vmem_add(vmem_t *vmp, void *vaddr, size_t size, int vmflag)
1356 1359 {
1357 1360 if (vaddr == NULL || size == 0)
1358 1361 panic("vmem_add(%p, %p, %lu): bad arguments",
1359 1362 (void *)vmp, vaddr, size);
1360 1363
1361 1364 ASSERT(!vmem_contains(vmp, vaddr, size));
1362 1365
1363 1366 mutex_enter(&vmp->vm_lock);
1364 1367 if (vmem_populate(vmp, vmflag))
1365 1368 (void) vmem_span_create(vmp, vaddr, size, 0);
1366 1369 else
1367 1370 vaddr = NULL;
1368 1371 mutex_exit(&vmp->vm_lock);
1369 1372 return (vaddr);
1370 1373 }
1371 1374
1372 1375 /*
1373 1376 * Walk the vmp arena, applying func to each segment matching typemask.
1374 1377 * If VMEM_REENTRANT is specified, the arena lock is dropped across each
1375 1378 * call to func(); otherwise, it is held for the duration of vmem_walk()
1376 1379 * to ensure a consistent snapshot. Note that VMEM_REENTRANT callbacks
1377 1380 * are *not* necessarily consistent, so they may only be used when a hint
1378 1381 * is adequate.
1379 1382 */
1380 1383 void
1381 1384 vmem_walk(vmem_t *vmp, int typemask,
1382 1385 void (*func)(void *, void *, size_t), void *arg)
1383 1386 {
1384 1387 vmem_seg_t *vsp;
1385 1388 vmem_seg_t *seg0 = &vmp->vm_seg0;
1386 1389 vmem_seg_t walker;
1387 1390
1388 1391 if (typemask & VMEM_WALKER)
1389 1392 return;
1390 1393
1391 1394 bzero(&walker, sizeof (walker));
1392 1395 walker.vs_type = VMEM_WALKER;
1393 1396
1394 1397 mutex_enter(&vmp->vm_lock);
1395 1398 VMEM_INSERT(seg0, &walker, a);
1396 1399 for (vsp = seg0->vs_anext; vsp != seg0; vsp = vsp->vs_anext) {
1397 1400 if (vsp->vs_type & typemask) {
1398 1401 void *start = (void *)vsp->vs_start;
1399 1402 size_t size = VS_SIZE(vsp);
1400 1403 if (typemask & VMEM_REENTRANT) {
1401 1404 vmem_advance(vmp, &walker, vsp);
1402 1405 mutex_exit(&vmp->vm_lock);
1403 1406 func(arg, start, size);
1404 1407 mutex_enter(&vmp->vm_lock);
1405 1408 vsp = &walker;
1406 1409 } else {
1407 1410 func(arg, start, size);
1408 1411 }
1409 1412 }
1410 1413 }
1411 1414 vmem_advance(vmp, &walker, NULL);
1412 1415 mutex_exit(&vmp->vm_lock);
1413 1416 }
1414 1417
1415 1418 /*
1416 1419 * Return the total amount of memory whose type matches typemask. Thus:
1417 1420 *
1418 1421 * typemask VMEM_ALLOC yields total memory allocated (in use).
1419 1422 * typemask VMEM_FREE yields total memory free (available).
1420 1423 * typemask (VMEM_ALLOC | VMEM_FREE) yields total arena size.
1421 1424 */
1422 1425 size_t
1423 1426 vmem_size(vmem_t *vmp, int typemask)
1424 1427 {
1425 1428 uint64_t size = 0;
1426 1429
1427 1430 if (typemask & VMEM_ALLOC)
1428 1431 size += vmp->vm_kstat.vk_mem_inuse.value.ui64;
1429 1432 if (typemask & VMEM_FREE)
1430 1433 size += vmp->vm_kstat.vk_mem_total.value.ui64 -
1431 1434 vmp->vm_kstat.vk_mem_inuse.value.ui64;
1432 1435 return ((size_t)size);
1433 1436 }
1434 1437
1435 1438 /*
1436 1439 * Create an arena called name whose initial span is [base, base + size).
1437 1440 * The arena's natural unit of currency is quantum, so vmem_alloc()
1438 1441 * guarantees quantum-aligned results. The arena may import new spans
1439 1442 * by invoking afunc() on source, and may return those spans by invoking
1440 1443 * ffunc() on source. To make small allocations fast and scalable,
1441 1444 * the arena offers high-performance caching for each integer multiple
1442 1445 * of quantum up to qcache_max.
1443 1446 */
1444 1447 static vmem_t *
1445 1448 vmem_create_common(const char *name, void *base, size_t size, size_t quantum,
1446 1449 void *(*afunc)(vmem_t *, size_t, int),
1447 1450 void (*ffunc)(vmem_t *, void *, size_t),
1448 1451 vmem_t *source, size_t qcache_max, int vmflag)
1449 1452 {
1450 1453 int i;
1451 1454 size_t nqcache;
1452 1455 vmem_t *vmp, *cur, **vmpp;
1453 1456 vmem_seg_t *vsp;
1454 1457 vmem_freelist_t *vfp;
1455 1458 uint32_t id = atomic_inc_32_nv(&vmem_id);
1456 1459
1457 1460 if (vmem_vmem_arena != NULL) {
1458 1461 vmp = vmem_alloc(vmem_vmem_arena, sizeof (vmem_t),
1459 1462 vmflag & VM_KMFLAGS);
1460 1463 } else {
1461 1464 ASSERT(id <= VMEM_INITIAL);
1462 1465 vmp = &vmem0[id - 1];
1463 1466 }
1464 1467
1465 1468 /* An identifier arena must inherit from another identifier arena */
1466 1469 ASSERT(source == NULL || ((source->vm_cflags & VMC_IDENTIFIER) ==
1467 1470 (vmflag & VMC_IDENTIFIER)));
1468 1471
1469 1472 if (vmp == NULL)
1470 1473 return (NULL);
1471 1474 bzero(vmp, sizeof (vmem_t));
1472 1475
1473 1476 (void) snprintf(vmp->vm_name, VMEM_NAMELEN, "%s", name);
1474 1477 mutex_init(&vmp->vm_lock, NULL, MUTEX_DEFAULT, NULL);
1475 1478 cv_init(&vmp->vm_cv, NULL, CV_DEFAULT, NULL);
1476 1479 vmp->vm_cflags = vmflag;
1477 1480 vmflag &= VM_KMFLAGS;
1478 1481
1479 1482 vmp->vm_quantum = quantum;
1480 1483 vmp->vm_qshift = highbit(quantum) - 1;
1481 1484 nqcache = MIN(qcache_max >> vmp->vm_qshift, VMEM_NQCACHE_MAX);
1482 1485
1483 1486 for (i = 0; i <= VMEM_FREELISTS; i++) {
1484 1487 vfp = &vmp->vm_freelist[i];
1485 1488 vfp->vs_end = 1UL << i;
1486 1489 vfp->vs_knext = (vmem_seg_t *)(vfp + 1);
1487 1490 vfp->vs_kprev = (vmem_seg_t *)(vfp - 1);
1488 1491 }
1489 1492
1490 1493 vmp->vm_freelist[0].vs_kprev = NULL;
1491 1494 vmp->vm_freelist[VMEM_FREELISTS].vs_knext = NULL;
1492 1495 vmp->vm_freelist[VMEM_FREELISTS].vs_end = 0;
1493 1496 vmp->vm_hash_table = vmp->vm_hash0;
1494 1497 vmp->vm_hash_mask = VMEM_HASH_INITIAL - 1;
1495 1498 vmp->vm_hash_shift = highbit(vmp->vm_hash_mask);
1496 1499
1497 1500 vsp = &vmp->vm_seg0;
1498 1501 vsp->vs_anext = vsp;
1499 1502 vsp->vs_aprev = vsp;
1500 1503 vsp->vs_knext = vsp;
1501 1504 vsp->vs_kprev = vsp;
1502 1505 vsp->vs_type = VMEM_SPAN;
1503 1506
1504 1507 vsp = &vmp->vm_rotor;
1505 1508 vsp->vs_type = VMEM_ROTOR;
1506 1509 VMEM_INSERT(&vmp->vm_seg0, vsp, a);
1507 1510
1508 1511 bcopy(&vmem_kstat_template, &vmp->vm_kstat, sizeof (vmem_kstat_t));
1509 1512
1510 1513 vmp->vm_id = id;
1511 1514 if (source != NULL)
1512 1515 vmp->vm_kstat.vk_source_id.value.ui32 = source->vm_id;
1513 1516 vmp->vm_source = source;
1514 1517 vmp->vm_source_alloc = afunc;
1515 1518 vmp->vm_source_free = ffunc;
1516 1519
1517 1520 /*
1518 1521 * Some arenas (like vmem_metadata and kmem_metadata) cannot
1519 1522 * use quantum caching to lower fragmentation. Instead, we
1520 1523 * increase their imports, giving a similar effect.
1521 1524 */
1522 1525 if (vmp->vm_cflags & VMC_NO_QCACHE) {
1523 1526 vmp->vm_min_import =
1524 1527 VMEM_QCACHE_SLABSIZE(nqcache << vmp->vm_qshift);
1525 1528 nqcache = 0;
1526 1529 }
1527 1530
1528 1531 if (nqcache != 0) {
1529 1532 ASSERT(!(vmflag & VM_NOSLEEP));
1530 1533 vmp->vm_qcache_max = nqcache << vmp->vm_qshift;
1531 1534 for (i = 0; i < nqcache; i++) {
1532 1535 char buf[VMEM_NAMELEN + 21];
1533 1536 (void) sprintf(buf, "%s_%lu", vmp->vm_name,
1534 1537 (i + 1) * quantum);
1535 1538 vmp->vm_qcache[i] = kmem_cache_create(buf,
1536 1539 (i + 1) * quantum, quantum, NULL, NULL, NULL,
1537 1540 NULL, vmp, KMC_QCACHE | KMC_NOTOUCH);
1538 1541 }
1539 1542 }
1540 1543
1541 1544 if ((vmp->vm_ksp = kstat_create("vmem", vmp->vm_id, vmp->vm_name,
1542 1545 "vmem", KSTAT_TYPE_NAMED, sizeof (vmem_kstat_t) /
1543 1546 sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL)) != NULL) {
1544 1547 vmp->vm_ksp->ks_data = &vmp->vm_kstat;
1545 1548 kstat_install(vmp->vm_ksp);
1546 1549 }
1547 1550
1548 1551 mutex_enter(&vmem_list_lock);
1549 1552 vmpp = &vmem_list;
1550 1553 while ((cur = *vmpp) != NULL)
1551 1554 vmpp = &cur->vm_next;
1552 1555 *vmpp = vmp;
1553 1556 mutex_exit(&vmem_list_lock);
1554 1557
1555 1558 if (vmp->vm_cflags & VMC_POPULATOR) {
1556 1559 ASSERT(vmem_populators < VMEM_INITIAL);
1557 1560 vmem_populator[atomic_inc_32_nv(&vmem_populators) - 1] = vmp;
1558 1561 mutex_enter(&vmp->vm_lock);
1559 1562 (void) vmem_populate(vmp, vmflag | VM_PANIC);
1560 1563 mutex_exit(&vmp->vm_lock);
1561 1564 }
1562 1565
1563 1566 if ((base || size) && vmem_add(vmp, base, size, vmflag) == NULL) {
1564 1567 vmem_destroy(vmp);
1565 1568 return (NULL);
1566 1569 }
1567 1570
1568 1571 return (vmp);
1569 1572 }
1570 1573
1571 1574 vmem_t *
1572 1575 vmem_xcreate(const char *name, void *base, size_t size, size_t quantum,
1573 1576 vmem_ximport_t *afunc, vmem_free_t *ffunc, vmem_t *source,
1574 1577 size_t qcache_max, int vmflag)
1575 1578 {
1576 1579 ASSERT(!(vmflag & (VMC_POPULATOR | VMC_XALLOC)));
1577 1580 vmflag &= ~(VMC_POPULATOR | VMC_XALLOC);
1578 1581
1579 1582 return (vmem_create_common(name, base, size, quantum,
1580 1583 (vmem_alloc_t *)afunc, ffunc, source, qcache_max,
1581 1584 vmflag | VMC_XALLOC));
1582 1585 }
1583 1586
1584 1587 vmem_t *
1585 1588 vmem_create(const char *name, void *base, size_t size, size_t quantum,
1586 1589 vmem_alloc_t *afunc, vmem_free_t *ffunc, vmem_t *source,
1587 1590 size_t qcache_max, int vmflag)
1588 1591 {
1589 1592 ASSERT(!(vmflag & (VMC_XALLOC | VMC_XALIGN)));
1590 1593 vmflag &= ~(VMC_XALLOC | VMC_XALIGN);
1591 1594
1592 1595 return (vmem_create_common(name, base, size, quantum,
1593 1596 afunc, ffunc, source, qcache_max, vmflag));
1594 1597 }
1595 1598
1596 1599 /*
1597 1600 * Destroy arena vmp.
1598 1601 */
1599 1602 void
1600 1603 vmem_destroy(vmem_t *vmp)
1601 1604 {
1602 1605 vmem_t *cur, **vmpp;
1603 1606 vmem_seg_t *seg0 = &vmp->vm_seg0;
1604 1607 vmem_seg_t *vsp, *anext;
1605 1608 size_t leaked;
1606 1609 int i;
1607 1610
1608 1611 mutex_enter(&vmem_list_lock);
1609 1612 vmpp = &vmem_list;
1610 1613 while ((cur = *vmpp) != vmp)
1611 1614 vmpp = &cur->vm_next;
1612 1615 *vmpp = vmp->vm_next;
1613 1616 mutex_exit(&vmem_list_lock);
1614 1617
1615 1618 for (i = 0; i < VMEM_NQCACHE_MAX; i++)
1616 1619 if (vmp->vm_qcache[i])
1617 1620 kmem_cache_destroy(vmp->vm_qcache[i]);
1618 1621
1619 1622 leaked = vmem_size(vmp, VMEM_ALLOC);
1620 1623 if (leaked != 0)
1621 1624 cmn_err(CE_WARN, "vmem_destroy('%s'): leaked %lu %s",
1622 1625 vmp->vm_name, leaked, (vmp->vm_cflags & VMC_IDENTIFIER) ?
1623 1626 "identifiers" : "bytes");
1624 1627
1625 1628 if (vmp->vm_hash_table != vmp->vm_hash0)
1626 1629 vmem_free(vmem_hash_arena, vmp->vm_hash_table,
1627 1630 (vmp->vm_hash_mask + 1) * sizeof (void *));
1628 1631
1629 1632 /*
1630 1633 * Give back the segment structures for anything that's left in the
1631 1634 * arena, e.g. the primary spans and their free segments.
1632 1635 */
1633 1636 VMEM_DELETE(&vmp->vm_rotor, a);
1634 1637 for (vsp = seg0->vs_anext; vsp != seg0; vsp = anext) {
1635 1638 anext = vsp->vs_anext;
1636 1639 vmem_putseg_global(vsp);
1637 1640 }
1638 1641
1639 1642 while (vmp->vm_nsegfree > 0)
1640 1643 vmem_putseg_global(vmem_getseg(vmp));
1641 1644
1642 1645 kstat_delete(vmp->vm_ksp);
1643 1646
1644 1647 mutex_destroy(&vmp->vm_lock);
1645 1648 cv_destroy(&vmp->vm_cv);
1646 1649 vmem_free(vmem_vmem_arena, vmp, sizeof (vmem_t));
1647 1650 }
1648 1651
1649 1652 /*
1650 1653 * Only shrink vmem hashtable if it is 1<<vmem_rescale_minshift times (8x)
1651 1654 * larger than necessary.
1652 1655 */
1653 1656 int vmem_rescale_minshift = 3;
1654 1657
1655 1658 /*
1656 1659 * Resize vmp's hash table to keep the average lookup depth near 1.0.
1657 1660 */
1658 1661 static void
1659 1662 vmem_hash_rescale(vmem_t *vmp)
1660 1663 {
1661 1664 vmem_seg_t **old_table, **new_table, *vsp;
1662 1665 size_t old_size, new_size, h, nseg;
1663 1666
1664 1667 nseg = (size_t)(vmp->vm_kstat.vk_alloc.value.ui64 -
1665 1668 vmp->vm_kstat.vk_free.value.ui64);
1666 1669
1667 1670 new_size = MAX(VMEM_HASH_INITIAL, 1 << (highbit(3 * nseg + 4) - 2));
1668 1671 old_size = vmp->vm_hash_mask + 1;
1669 1672
1670 1673 if ((old_size >> vmem_rescale_minshift) <= new_size &&
1671 1674 new_size <= (old_size << 1))
1672 1675 return;
1673 1676
1674 1677 new_table = vmem_alloc(vmem_hash_arena, new_size * sizeof (void *),
1675 1678 VM_NOSLEEP);
1676 1679 if (new_table == NULL)
1677 1680 return;
1678 1681 bzero(new_table, new_size * sizeof (void *));
1679 1682
1680 1683 mutex_enter(&vmp->vm_lock);
1681 1684
1682 1685 old_size = vmp->vm_hash_mask + 1;
1683 1686 old_table = vmp->vm_hash_table;
1684 1687
1685 1688 vmp->vm_hash_mask = new_size - 1;
1686 1689 vmp->vm_hash_table = new_table;
1687 1690 vmp->vm_hash_shift = highbit(vmp->vm_hash_mask);
1688 1691
1689 1692 for (h = 0; h < old_size; h++) {
1690 1693 vsp = old_table[h];
1691 1694 while (vsp != NULL) {
1692 1695 uintptr_t addr = vsp->vs_start;
1693 1696 vmem_seg_t *next_vsp = vsp->vs_knext;
1694 1697 vmem_seg_t **hash_bucket = VMEM_HASH(vmp, addr);
1695 1698 vsp->vs_knext = *hash_bucket;
1696 1699 *hash_bucket = vsp;
1697 1700 vsp = next_vsp;
1698 1701 }
1699 1702 }
1700 1703
1701 1704 mutex_exit(&vmp->vm_lock);
1702 1705
1703 1706 if (old_table != vmp->vm_hash0)
1704 1707 vmem_free(vmem_hash_arena, old_table,
1705 1708 old_size * sizeof (void *));
1706 1709 }
1707 1710
1708 1711 /*
1709 1712 * Perform periodic maintenance on all vmem arenas.
1710 1713 */
1711 1714 void
1712 1715 vmem_update(void *dummy)
1713 1716 {
1714 1717 vmem_t *vmp;
1715 1718
1716 1719 mutex_enter(&vmem_list_lock);
1717 1720 for (vmp = vmem_list; vmp != NULL; vmp = vmp->vm_next) {
1718 1721 /*
1719 1722 * If threads are waiting for resources, wake them up
1720 1723 * periodically so they can issue another kmem_reap()
1721 1724 * to reclaim resources cached by the slab allocator.
1722 1725 */
1723 1726 cv_broadcast(&vmp->vm_cv);
1724 1727
1725 1728 /*
1726 1729 * Rescale the hash table to keep the hash chains short.
1727 1730 */
1728 1731 vmem_hash_rescale(vmp);
1729 1732 }
1730 1733 mutex_exit(&vmem_list_lock);
1731 1734
1732 1735 (void) timeout(vmem_update, dummy, vmem_update_interval * hz);
1733 1736 }
1734 1737
1735 1738 void
1736 1739 vmem_qcache_reap(vmem_t *vmp)
1737 1740 {
1738 1741 int i;
1739 1742
1740 1743 /*
1741 1744 * Reap any quantum caches that may be part of this vmem.
1742 1745 */
1743 1746 for (i = 0; i < VMEM_NQCACHE_MAX; i++)
1744 1747 if (vmp->vm_qcache[i])
1745 1748 kmem_cache_reap_now(vmp->vm_qcache[i]);
1746 1749 }
1747 1750
1748 1751 /*
1749 1752 * Prepare vmem for use.
1750 1753 */
1751 1754 vmem_t *
1752 1755 vmem_init(const char *heap_name,
1753 1756 void *heap_start, size_t heap_size, size_t heap_quantum,
1754 1757 void *(*heap_alloc)(vmem_t *, size_t, int),
1755 1758 void (*heap_free)(vmem_t *, void *, size_t))
1756 1759 {
1757 1760 uint32_t id;
1758 1761 int nseg = VMEM_SEG_INITIAL;
1759 1762 vmem_t *heap;
1760 1763
1761 1764 while (--nseg >= 0)
1762 1765 vmem_putseg_global(&vmem_seg0[nseg]);
1763 1766
1764 1767 heap = vmem_create(heap_name,
1765 1768 heap_start, heap_size, heap_quantum,
1766 1769 NULL, NULL, NULL, 0,
1767 1770 VM_SLEEP | VMC_POPULATOR);
1768 1771
1769 1772 vmem_metadata_arena = vmem_create("vmem_metadata",
1770 1773 NULL, 0, heap_quantum,
1771 1774 vmem_alloc, vmem_free, heap, 8 * heap_quantum,
1772 1775 VM_SLEEP | VMC_POPULATOR | VMC_NO_QCACHE);
1773 1776
1774 1777 vmem_seg_arena = vmem_create("vmem_seg",
1775 1778 NULL, 0, heap_quantum,
1776 1779 heap_alloc, heap_free, vmem_metadata_arena, 0,
1777 1780 VM_SLEEP | VMC_POPULATOR);
1778 1781
1779 1782 vmem_hash_arena = vmem_create("vmem_hash",
1780 1783 NULL, 0, 8,
1781 1784 heap_alloc, heap_free, vmem_metadata_arena, 0,
1782 1785 VM_SLEEP);
1783 1786
1784 1787 vmem_vmem_arena = vmem_create("vmem_vmem",
1785 1788 vmem0, sizeof (vmem0), 1,
1786 1789 heap_alloc, heap_free, vmem_metadata_arena, 0,
1787 1790 VM_SLEEP);
1788 1791
1789 1792 for (id = 0; id < vmem_id; id++)
1790 1793 (void) vmem_xalloc(vmem_vmem_arena, sizeof (vmem_t),
1791 1794 1, 0, 0, &vmem0[id], &vmem0[id + 1],
1792 1795 VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
1793 1796
1794 1797 return (heap);
1795 1798 }
|
↓ open down ↓ |
1742 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX