1 /*
2 * Copyright (c) 2016-2021 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28
29 /* BEGIN CSTYLED */
30 /*
31 * A region represents a collection of one or more similarly-sized memory
32 * segments, each of which is a contiguous range of integers. A segment
33 * is either allocated or free, and is treated as disjoint from all other
34 * segments. That is, the contiguity applies only at the segment level,
35 * and a region with multiple segments is not contiguous at the region level.
36 * A segment always belongs to the segment freelist, or the allocated-address
37 * hash chain, as described below.
38 *
39 * The optional SKMEM_REGION_CR_NOREDIRECT flag indicates that the region
40 * stays intact even after a defunct. Otherwise, the segments belonging
41 * to the region will be freed at defunct time, and the span covered by
42 * the region will be redirected to zero-filled anonymous memory.
43 *
44 * Memory for a region is always created as pageable and purgeable. It is
45 * the client's responsibility to prepare (wire) it, and optionally insert
46 * it to the IOMMU, at segment construction time. When the segment is
47 * freed, the client is responsible for removing it from IOMMU (if needed),
48 * and complete (unwire) it.
49 *
50 * When the region is created with SKMEM_REGION_CR_PERSISTENT, the memory
51 * is immediately wired upon allocation (segment removed from freelist).
52 * It gets unwired when memory is discarded (segment inserted to freelist).
53 *
54 * The chronological life cycle of a segment is as such:
55 *
56 * SKSEG_STATE_DETACHED
57 * SKSEG_STATE_{MAPPED,MAPPED_WIRED}
58 * [segment allocated, useable by client]
59 * ...
60 * [client frees segment]
61 * SKSEG_STATE_{MAPPED,MAPPED_WIRED}
62 * [reclaim]
63 * SKSEG_STATE_DETACHED
64 *
65 * The region can also be marked as user-mappable (SKMEM_REGION_CR_MMAPOK);
66 * this allows it to be further marked with SKMEM_REGION_CR_UREADONLY to
67 * prevent modifications by the user task. Only user-mappable regions will
68 * be considered for inclusion during skmem_arena_mmap().
69 *
70 * Every skmem allocator has a region as its slab supplier. Each slab is
71 * exactly a segment. The allocator uses skmem_region_{alloc,free}() to
72 * create and destroy slabs.
73 *
74 * A region may be mirrored by another region; the latter acts as the master
75 * controller for both regions. Mirrored (slave) regions cannot be used
76 * directly by the skmem allocator. Region mirroring technique is used for
77 * managing shadow objects {umd,kmd} and {usd,ksd}, where an object in one
78 * region has the same size and lifetime as its shadow counterpart.
79 *
80 * CREATION/DESTRUCTION:
81 *
82 * At creation time, all segments are allocated and are immediately inserted
83 * into the freelist. Allocating a purgeable segment has very little cost,
84 * as it is not backed by physical memory until it is accessed. Immediate
85 * insertion into the freelist causes the mapping to be further torn down.
86 *
87 * At destruction time, the freelist is emptied, and each segment is then
88 * destroyed. The system will assert if it detects there are outstanding
89 * segments not yet returned to the region (not freed by the client.)
90 *
91 * ALLOCATION:
92 *
93 * Allocating involves searching the freelist for a segment; if found, the
94 * segment is removed from the freelist and is inserted into the allocated-
95 * address hash chain. The address of the memory object represented by
96 * the segment is used as hash key. The use of allocated-address hash chain
97 * is needed since we return the address of the memory object, and not the
98 * segment's itself, to the client.
99 *
100 * DEALLOCATION:
101 *
102 * Freeing a memory object causes the chain to be searched for a matching
103 * segment. The system will assert if a segment cannot be found, since
104 * that indicates that the memory object address is invalid. Once found,
105 * the segment is removed from the allocated-address hash chain, and is
106 * inserted to the freelist.
107 *
108 * Segment allocation and deallocation can be expensive. Because of this,
109 * we expect that most clients will utilize the skmem_cache slab allocator
110 * as the frontend instead.
111 */
112 /* END CSTYLED */
113
114 #include <skywalk/os_skywalk_private.h>
115 #define _FN_KPRINTF /* don't redefine kprintf() */
116 #include <pexpert/pexpert.h> /* for PE_parse_boot_argn */
117
118 static void skmem_region_destroy(struct skmem_region *skr);
119 static void skmem_region_depopulate(struct skmem_region *);
120 static int sksegment_cmp(const struct sksegment *, const struct sksegment *);
121 static struct sksegment *sksegment_create(struct skmem_region *, uint32_t);
122 static void sksegment_destroy(struct skmem_region *, struct sksegment *);
123 static void sksegment_freelist_insert(struct skmem_region *,
124 struct sksegment *, boolean_t);
125 static struct sksegment *sksegment_freelist_remove(struct skmem_region *,
126 struct sksegment *, uint32_t, boolean_t);
127 static struct sksegment *sksegment_freelist_grow(struct skmem_region *);
128 static struct sksegment *sksegment_alloc_with_idx(struct skmem_region *,
129 uint32_t);
130 static void *skmem_region_alloc_common(struct skmem_region *,
131 struct sksegment *);
132 static void *skmem_region_mirror_alloc(struct skmem_region *,
133 struct sksegment *, struct sksegment **);
134 static void skmem_region_applyall(void (*)(struct skmem_region *));
135 static void skmem_region_update(struct skmem_region *);
136 static void skmem_region_update_func(thread_call_param_t, thread_call_param_t);
137 static inline void skmem_region_retain_locked(struct skmem_region *);
138 static inline boolean_t skmem_region_release_locked(struct skmem_region *);
139 static int skmem_region_mib_get_sysctl SYSCTL_HANDLER_ARGS;
140
141 RB_PROTOTYPE_PREV(segtfreehead, sksegment, sg_node, sksegment_cmp);
142 RB_GENERATE_PREV(segtfreehead, sksegment, sg_node, sksegment_cmp);
143
144 SYSCTL_PROC(_kern_skywalk_stats, OID_AUTO, region,
145 CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_LOCKED,
146 0, 0, skmem_region_mib_get_sysctl, "S,sk_stats_region",
147 "Skywalk region statistics");
148
149 static LCK_ATTR_DECLARE(skmem_region_lock_attr, 0, 0);
150 static LCK_GRP_DECLARE(skmem_region_lock_grp, "skmem_region");
151 static LCK_MTX_DECLARE_ATTR(skmem_region_lock, &skmem_region_lock_grp,
152 &skmem_region_lock_attr);
153
154 /* protected by skmem_region_lock */
155 static TAILQ_HEAD(, skmem_region) skmem_region_head;
156
157 static thread_call_t skmem_region_update_tc;
158
159 #define SKMEM_REGION_UPDATE_INTERVAL 13 /* 13 seconds */
160 static uint32_t skmem_region_update_interval = SKMEM_REGION_UPDATE_INTERVAL;
161
162 #define SKMEM_WDT_MAXTIME 30 /* # of secs before watchdog */
163 #define SKMEM_WDT_PURGE 3 /* retry purge threshold */
164
165 #if (DEVELOPMENT || DEBUG)
166 /* Mean Time Between Failures (ms) */
167 static volatile uint64_t skmem_region_mtbf;
168
169 static int skmem_region_mtbf_sysctl(struct sysctl_oid *, void *, int,
170 struct sysctl_req *);
171
172 SYSCTL_PROC(_kern_skywalk_mem, OID_AUTO, region_mtbf,
173 CTLTYPE_QUAD | CTLFLAG_RW | CTLFLAG_LOCKED, NULL, 0,
174 skmem_region_mtbf_sysctl, "Q", "Region MTBF (ms)");
175
176 SYSCTL_UINT(_kern_skywalk_mem, OID_AUTO, region_update_interval,
177 CTLFLAG_RW | CTLFLAG_LOCKED, &skmem_region_update_interval,
178 SKMEM_REGION_UPDATE_INTERVAL, "Region update interval (sec)");
179 #endif /* (DEVELOPMENT || DEBUG) */
180
181 #define SKMEM_REGION_LOCK() \
182 lck_mtx_lock(&skmem_region_lock)
183 #define SKMEM_REGION_LOCK_ASSERT_HELD() \
184 LCK_MTX_ASSERT(&skmem_region_lock, LCK_MTX_ASSERT_OWNED)
185 #define SKMEM_REGION_LOCK_ASSERT_NOTHELD() \
186 LCK_MTX_ASSERT(&skmem_region_lock, LCK_MTX_ASSERT_NOTOWNED)
187 #define SKMEM_REGION_UNLOCK() \
188 lck_mtx_unlock(&skmem_region_lock)
189
190 /*
191 * Hash table bounds. Start with the initial value, and rescale up to
192 * the specified limit. Ideally we don't need a limit, but in practice
193 * this helps guard against runaways. These values should be revisited
194 * in future and be adjusted as needed.
195 */
196 #define SKMEM_REGION_HASH_INITIAL 32 /* initial hash table size */
197 #define SKMEM_REGION_HASH_LIMIT 4096 /* hash table size limit */
198
199 #define SKMEM_REGION_HASH_INDEX(_a, _s, _m) \
200 (((_a) + ((_a) >> (_s)) + ((_a) >> ((_s) << 1))) & (_m))
201 #define SKMEM_REGION_HASH(_skr, _addr) \
202 (&(_skr)->skr_hash_table[SKMEM_REGION_HASH_INDEX((uintptr_t)_addr, \
203 (_skr)->skr_hash_shift, (_skr)->skr_hash_mask)])
204
205 static ZONE_DEFINE(skr_zone, SKMEM_ZONE_PREFIX ".mem.skr",
206 sizeof(struct skmem_region), ZC_ZFREE_CLEARMEM);
207
208 static unsigned int sg_size; /* size of zone element */
209 static struct skmem_cache *skmem_sg_cache; /* cache for sksegment */
210
211 static uint32_t skmem_seg_size = SKMEM_SEG_SIZE;
212 static uint32_t skmem_md_seg_size = SKMEM_MD_SEG_SIZE;
213 static uint32_t skmem_drv_buf_seg_size = SKMEM_DRV_BUF_SEG_SIZE;
214 static uint32_t skmem_drv_buf_seg_eff_size = SKMEM_DRV_BUF_SEG_SIZE;
215 uint32_t skmem_usr_buf_seg_size = SKMEM_USR_BUF_SEG_SIZE;
216
217 #define SKMEM_TAG_SEGMENT_BMAP "com.apple.skywalk.segment.bmap"
218 static SKMEM_TAG_DEFINE(skmem_tag_segment_bmap, SKMEM_TAG_SEGMENT_BMAP);
219
220 #define SKMEM_TAG_SEGMENT_HASH "com.apple.skywalk.segment.hash"
221 static SKMEM_TAG_DEFINE(skmem_tag_segment_hash, SKMEM_TAG_SEGMENT_HASH);
222
223 #define SKMEM_TAG_REGION_MIB "com.apple.skywalk.region.mib"
224 static SKMEM_TAG_DEFINE(skmem_tag_region_mib, SKMEM_TAG_REGION_MIB);
225
226 #define BMAPSZ 64
227
228 /* 64-bit mask with range */
229 #define BMASK64(_beg, _end) \
230 ((((uint64_t)-1) >> ((BMAPSZ - 1) - (_end))) & ~((1ULL << (_beg)) - 1))
231
232 static int __skmem_region_inited = 0;
233
234 void
skmem_region_init(void)235 skmem_region_init(void)
236 {
237 boolean_t randomize_seg_size;
238
239 _CASSERT(sizeof(bitmap_t) == sizeof(uint64_t));
240 _CASSERT(BMAPSZ == (sizeof(bitmap_t) << 3));
241 _CASSERT((SKMEM_SEG_SIZE % SKMEM_PAGE_SIZE) == 0);
242 _CASSERT(SKMEM_REGION_HASH_LIMIT >= SKMEM_REGION_HASH_INITIAL);
243 ASSERT(!__skmem_region_inited);
244
245 /* enforce the ordering here */
246 _CASSERT(SKMEM_REGION_GUARD_HEAD == 0);
247 _CASSERT(SKMEM_REGION_SCHEMA == 1);
248 _CASSERT(SKMEM_REGION_RING == 2);
249 _CASSERT(SKMEM_REGION_BUF == 3);
250 _CASSERT(SKMEM_REGION_RXBUF == 4);
251 _CASSERT(SKMEM_REGION_TXBUF == 5);
252 _CASSERT(SKMEM_REGION_UMD == 6);
253 _CASSERT(SKMEM_REGION_TXAUSD == 7);
254 _CASSERT(SKMEM_REGION_RXFUSD == 8);
255 _CASSERT(SKMEM_REGION_UBFT == 9);
256 _CASSERT(SKMEM_REGION_USTATS == 10);
257 _CASSERT(SKMEM_REGION_FLOWADV == 11);
258 _CASSERT(SKMEM_REGION_NEXUSADV == 12);
259 _CASSERT(SKMEM_REGION_SYSCTLS == 13);
260 _CASSERT(SKMEM_REGION_GUARD_TAIL == 14);
261 _CASSERT(SKMEM_REGION_KMD == 15);
262 _CASSERT(SKMEM_REGION_RXKMD == 16);
263 _CASSERT(SKMEM_REGION_TXKMD == 17);
264 _CASSERT(SKMEM_REGION_KBFT == 18);
265 _CASSERT(SKMEM_REGION_RXKBFT == 19);
266 _CASSERT(SKMEM_REGION_TXKBFT == 20);
267 _CASSERT(SKMEM_REGION_TXAKSD == 21);
268 _CASSERT(SKMEM_REGION_RXFKSD == 22);
269 _CASSERT(SKMEM_REGION_KSTATS == 23);
270 _CASSERT(SKMEM_REGION_INTRINSIC == 24);
271
272 _CASSERT(SREG_GUARD_HEAD == SKMEM_REGION_GUARD_HEAD);
273 _CASSERT(SREG_SCHEMA == SKMEM_REGION_SCHEMA);
274 _CASSERT(SREG_RING == SKMEM_REGION_RING);
275 _CASSERT(SREG_BUF == SKMEM_REGION_BUF);
276 _CASSERT(SREG_RXBUF == SKMEM_REGION_RXBUF);
277 _CASSERT(SREG_TXBUF == SKMEM_REGION_TXBUF);
278 _CASSERT(SREG_UMD == SKMEM_REGION_UMD);
279 _CASSERT(SREG_TXAUSD == SKMEM_REGION_TXAUSD);
280 _CASSERT(SREG_RXFUSD == SKMEM_REGION_RXFUSD);
281 _CASSERT(SREG_UBFT == SKMEM_REGION_UBFT);
282 _CASSERT(SREG_USTATS == SKMEM_REGION_USTATS);
283 _CASSERT(SREG_FLOWADV == SKMEM_REGION_FLOWADV);
284 _CASSERT(SREG_NEXUSADV == SKMEM_REGION_NEXUSADV);
285 _CASSERT(SREG_SYSCTLS == SKMEM_REGION_SYSCTLS);
286 _CASSERT(SREG_GUARD_TAIL == SKMEM_REGION_GUARD_TAIL);
287 _CASSERT(SREG_KMD == SKMEM_REGION_KMD);
288 _CASSERT(SREG_RXKMD == SKMEM_REGION_RXKMD);
289 _CASSERT(SREG_TXKMD == SKMEM_REGION_TXKMD);
290 _CASSERT(SREG_KBFT == SKMEM_REGION_KBFT);
291 _CASSERT(SREG_RXKBFT == SKMEM_REGION_RXKBFT);
292 _CASSERT(SREG_TXKBFT == SKMEM_REGION_TXKBFT);
293 _CASSERT(SREG_TXAKSD == SKMEM_REGION_TXAKSD);
294 _CASSERT(SREG_RXFKSD == SKMEM_REGION_RXFKSD);
295 _CASSERT(SREG_KSTATS == SKMEM_REGION_KSTATS);
296
297 _CASSERT(SKR_MODE_NOREDIRECT == SREG_MODE_NOREDIRECT);
298 _CASSERT(SKR_MODE_MMAPOK == SREG_MODE_MMAPOK);
299 _CASSERT(SKR_MODE_UREADONLY == SREG_MODE_UREADONLY);
300 _CASSERT(SKR_MODE_KREADONLY == SREG_MODE_KREADONLY);
301 _CASSERT(SKR_MODE_PERSISTENT == SREG_MODE_PERSISTENT);
302 _CASSERT(SKR_MODE_MONOLITHIC == SREG_MODE_MONOLITHIC);
303 _CASSERT(SKR_MODE_NOMAGAZINES == SREG_MODE_NOMAGAZINES);
304 _CASSERT(SKR_MODE_NOCACHE == SREG_MODE_NOCACHE);
305 _CASSERT(SKR_MODE_IODIR_IN == SREG_MODE_IODIR_IN);
306 _CASSERT(SKR_MODE_IODIR_OUT == SREG_MODE_IODIR_OUT);
307 _CASSERT(SKR_MODE_GUARD == SREG_MODE_GUARD);
308 _CASSERT(SKR_MODE_SEGPHYSCONTIG == SREG_MODE_SEGPHYSCONTIG);
309 _CASSERT(SKR_MODE_SHAREOK == SREG_MODE_SHAREOK);
310 _CASSERT(SKR_MODE_PUREDATA == SREG_MODE_PUREDATA);
311 _CASSERT(SKR_MODE_PSEUDO == SREG_MODE_PSEUDO);
312 _CASSERT(SKR_MODE_SLAB == SREG_MODE_SLAB);
313 _CASSERT(SKR_MODE_MIRRORED == SREG_MODE_MIRRORED);
314
315 (void) PE_parse_boot_argn("skmem_seg_size", &skmem_seg_size,
316 sizeof(skmem_seg_size));
317 if (skmem_seg_size < SKMEM_MIN_SEG_SIZE) {
318 skmem_seg_size = SKMEM_MIN_SEG_SIZE;
319 }
320 skmem_seg_size = (uint32_t)P2ROUNDUP(skmem_seg_size,
321 SKMEM_MIN_SEG_SIZE);
322 VERIFY(skmem_seg_size != 0 && (skmem_seg_size % SKMEM_PAGE_SIZE) == 0);
323
324 (void) PE_parse_boot_argn("skmem_md_seg_size", &skmem_md_seg_size,
325 sizeof(skmem_md_seg_size));
326 if (skmem_md_seg_size < skmem_seg_size) {
327 skmem_md_seg_size = skmem_seg_size;
328 }
329 skmem_md_seg_size = (uint32_t)P2ROUNDUP(skmem_md_seg_size,
330 SKMEM_MIN_SEG_SIZE);
331 VERIFY((skmem_md_seg_size % SKMEM_PAGE_SIZE) == 0);
332
333 /*
334 * If set via boot-args, honor it and don't randomize.
335 */
336 randomize_seg_size = !PE_parse_boot_argn("skmem_drv_buf_seg_size",
337 &skmem_drv_buf_seg_size, sizeof(skmem_drv_buf_seg_size));
338 if (skmem_drv_buf_seg_size < skmem_seg_size) {
339 skmem_drv_buf_seg_size = skmem_seg_size;
340 }
341 skmem_drv_buf_seg_size = skmem_drv_buf_seg_eff_size =
342 (uint32_t)P2ROUNDUP(skmem_drv_buf_seg_size, SKMEM_MIN_SEG_SIZE);
343 VERIFY((skmem_drv_buf_seg_size % SKMEM_PAGE_SIZE) == 0);
344
345 /*
346 * Randomize the driver buffer segment size; here we choose
347 * a SKMEM_MIN_SEG_SIZE multiplier to bump up the value to.
348 * Set this as the effective driver buffer segment size.
349 */
350 if (randomize_seg_size) {
351 uint32_t sm;
352 read_frandom(&sm, sizeof(sm));
353 skmem_drv_buf_seg_eff_size +=
354 (SKMEM_MIN_SEG_SIZE * (sm % SKMEM_DRV_BUF_SEG_MULTIPLIER));
355 VERIFY((skmem_drv_buf_seg_eff_size % SKMEM_MIN_SEG_SIZE) == 0);
356 }
357 VERIFY(skmem_drv_buf_seg_eff_size >= skmem_drv_buf_seg_size);
358
359 (void) PE_parse_boot_argn("skmem_usr_buf_seg_size",
360 &skmem_usr_buf_seg_size, sizeof(skmem_usr_buf_seg_size));
361 if (skmem_usr_buf_seg_size < skmem_seg_size) {
362 skmem_usr_buf_seg_size = skmem_seg_size;
363 }
364 skmem_usr_buf_seg_size = (uint32_t)P2ROUNDUP(skmem_usr_buf_seg_size,
365 SKMEM_MIN_SEG_SIZE);
366 VERIFY((skmem_usr_buf_seg_size % SKMEM_PAGE_SIZE) == 0);
367
368 SK_ERR("seg_size %u, md_seg_size %u, drv_buf_seg_size %u [eff %u], "
369 "usr_buf_seg_size %u", skmem_seg_size, skmem_md_seg_size,
370 skmem_drv_buf_seg_size, skmem_drv_buf_seg_eff_size,
371 skmem_usr_buf_seg_size);
372
373 TAILQ_INIT(&skmem_region_head);
374
375 skmem_region_update_tc =
376 thread_call_allocate_with_options(skmem_region_update_func,
377 NULL, THREAD_CALL_PRIORITY_KERNEL, THREAD_CALL_OPTIONS_ONCE);
378 if (skmem_region_update_tc == NULL) {
379 panic("%s: thread_call_allocate failed", __func__);
380 /* NOTREACHED */
381 __builtin_unreachable();
382 }
383
384 sg_size = sizeof(struct sksegment);
385 skmem_sg_cache = skmem_cache_create("sg", sg_size,
386 sizeof(uint64_t), NULL, NULL, NULL, NULL, NULL, 0);
387
388 /* and start the periodic region update machinery */
389 skmem_dispatch(skmem_region_update_tc, NULL,
390 (skmem_region_update_interval * NSEC_PER_SEC));
391
392 __skmem_region_inited = 1;
393 }
394
395 void
skmem_region_fini(void)396 skmem_region_fini(void)
397 {
398 if (__skmem_region_inited) {
399 ASSERT(TAILQ_EMPTY(&skmem_region_head));
400
401 if (skmem_region_update_tc != NULL) {
402 (void) thread_call_cancel_wait(skmem_region_update_tc);
403 (void) thread_call_free(skmem_region_update_tc);
404 skmem_region_update_tc = NULL;
405 }
406
407 if (skmem_sg_cache != NULL) {
408 skmem_cache_destroy(skmem_sg_cache);
409 skmem_sg_cache = NULL;
410 }
411
412 __skmem_region_inited = 0;
413 }
414 }
415
416 /*
417 * Reap internal caches.
418 */
419 void
skmem_region_reap_caches(boolean_t purge)420 skmem_region_reap_caches(boolean_t purge)
421 {
422 skmem_cache_reap_now(skmem_sg_cache, purge);
423 }
424
425 /*
426 * Configure and compute the parameters of a region.
427 */
428 void
skmem_region_params_config(struct skmem_region_params * srp)429 skmem_region_params_config(struct skmem_region_params *srp)
430 {
431 uint32_t cache_line_size = skmem_cpu_cache_line_size();
432 size_t seglim, segsize, segcnt;
433 size_t objsize, objcnt;
434
435 ASSERT(srp->srp_id < SKMEM_REGIONS);
436
437 /*
438 * If magazines layer is disabled system-wide, override
439 * the region parameter here. This will effectively reduce
440 * the number of requested objects computed below. Note that
441 * the region may have already been configured to exclude
442 * magazines in the default skmem_regions[] array.
443 */
444 if (!skmem_allow_magazines()) {
445 srp->srp_cflags |= SKMEM_REGION_CR_NOMAGAZINES;
446 }
447
448 objsize = srp->srp_r_obj_size;
449 ASSERT(objsize != 0);
450 objcnt = srp->srp_r_obj_cnt;
451 ASSERT(objcnt != 0);
452
453 if (srp->srp_cflags & SKMEM_REGION_CR_PSEUDO) {
454 size_t align = srp->srp_align;
455
456 VERIFY(align != 0 && (align % SKMEM_CACHE_ALIGN) == 0);
457 VERIFY(powerof2(align));
458 objsize = MAX(objsize, sizeof(uint64_t));
459 #if KASAN
460 /*
461 * When KASAN is enabled, the zone allocator adjusts the
462 * element size to include the redzone regions, in which
463 * case we assume that the elements won't start on the
464 * alignment boundary and thus need to do some fix-ups.
465 * These include increasing the effective object size
466 * which adds at least 16 bytes to the original size.
467 */
468 objsize += sizeof(uint64_t) + align;
469 #endif /* KASAN */
470 objsize = P2ROUNDUP(objsize, align);
471
472 segsize = objsize;
473 srp->srp_r_seg_size = (uint32_t)segsize;
474 segcnt = objcnt;
475 goto done;
476 } else {
477 /* objects are always aligned at CPU cache line size */
478 srp->srp_align = cache_line_size;
479 }
480
481 /*
482 * Start with default segment size for the region, and compute the
483 * effective segment size (to nearest SKMEM_MIN_SEG_SIZE). If the
484 * object size is greater, then we adjust the segment size to next
485 * multiple of the effective size larger than the object size.
486 */
487 if (srp->srp_r_seg_size == 0) {
488 switch (srp->srp_id) {
489 case SKMEM_REGION_UMD:
490 case SKMEM_REGION_KMD:
491 case SKMEM_REGION_RXKMD:
492 case SKMEM_REGION_TXKMD:
493 srp->srp_r_seg_size = skmem_md_seg_size;
494 break;
495
496 case SKMEM_REGION_BUF:
497 case SKMEM_REGION_RXBUF:
498 case SKMEM_REGION_TXBUF:
499 /*
500 * Use the effective driver buffer segment size,
501 * since it reflects any randomization done at
502 * skmem_region_init() time.
503 */
504 srp->srp_r_seg_size = skmem_drv_buf_seg_eff_size;
505 break;
506
507 default:
508 srp->srp_r_seg_size = skmem_seg_size;
509 break;
510 }
511 } else {
512 srp->srp_r_seg_size = (uint32_t)P2ROUNDUP(srp->srp_r_seg_size,
513 SKMEM_MIN_SEG_SIZE);
514 }
515
516 seglim = srp->srp_r_seg_size;
517 VERIFY(seglim != 0 && (seglim % SKMEM_PAGE_SIZE) == 0);
518
519 SK_DF(SK_VERB_MEM, "%s: seglim %zu objsize %zu objcnt %zu",
520 srp->srp_name, seglim, objsize, objcnt);
521
522 /*
523 * Make sure object size is multiple of CPU cache line
524 * size, and that we can evenly divide the segment size.
525 */
526 if (!((objsize < cache_line_size) && (objsize < seglim) &&
527 ((cache_line_size % objsize) == 0) && ((seglim % objsize) == 0))) {
528 objsize = P2ROUNDUP(objsize, cache_line_size);
529 while (objsize < seglim && (seglim % objsize) != 0) {
530 SK_DF(SK_VERB_MEM, "%s: objsize %zu -> %zu",
531 srp->srp_name, objsize, objsize + cache_line_size);
532 objsize += cache_line_size;
533 }
534 }
535
536 /* segment must be larger than object */
537 while (objsize > seglim) {
538 SK_DF(SK_VERB_MEM, "%s: seglim %zu -> %zu", srp->srp_name,
539 seglim, seglim + SKMEM_MIN_SEG_SIZE);
540 seglim += SKMEM_MIN_SEG_SIZE;
541 }
542
543 /*
544 * Take into account worst-case per-CPU cached
545 * objects if this region is configured for it.
546 */
547 if (!(srp->srp_cflags & SKMEM_REGION_CR_NOMAGAZINES)) {
548 uint32_t magazine_max_objs =
549 skmem_cache_magazine_max((uint32_t)objsize);
550 SK_DF(SK_VERB_MEM, "%s: objcnt %zu -> %zu", srp->srp_name,
551 objcnt, objcnt + magazine_max_objs);
552 objcnt += magazine_max_objs;
553 }
554
555 SK_DF(SK_VERB_MEM, "%s: seglim %zu objsize %zu "
556 "objcnt %zu", srp->srp_name, seglim, objsize, objcnt);
557
558 segsize = P2ROUNDUP(objsize * objcnt, SKMEM_MIN_SEG_SIZE);
559 if (seglim > segsize) {
560 /*
561 * If the segment limit is larger than what we need,
562 * avoid memory wastage by shrinking it.
563 */
564 while (seglim > segsize && seglim > SKMEM_MIN_SEG_SIZE) {
565 VERIFY(seglim >= SKMEM_MIN_SEG_SIZE);
566 SK_DF(SK_VERB_MEM,
567 "%s: segsize %zu (%zu*%zu) seglim [-] %zu -> %zu",
568 srp->srp_name, segsize, objsize, objcnt, seglim,
569 P2ROUNDUP(seglim - SKMEM_MIN_SEG_SIZE,
570 SKMEM_MIN_SEG_SIZE));
571 seglim = P2ROUNDUP(seglim - SKMEM_MIN_SEG_SIZE,
572 SKMEM_MIN_SEG_SIZE);
573 }
574
575 /* adjust segment size */
576 segsize = seglim;
577 } else if (seglim < segsize) {
578 size_t oseglim = seglim;
579 /*
580 * If the segment limit is less than the segment size,
581 * see if increasing it slightly (up to 1.5x the segment
582 * size) would allow us to avoid allocating too many
583 * extra objects (due to excessive segment count).
584 */
585 while (seglim < segsize && (segsize % seglim) != 0) {
586 SK_DF(SK_VERB_MEM,
587 "%s: segsize %zu (%zu*%zu) seglim [+] %zu -> %zu",
588 srp->srp_name, segsize, objsize, objcnt, seglim,
589 (seglim + SKMEM_MIN_SEG_SIZE));
590 seglim += SKMEM_MIN_SEG_SIZE;
591 if (seglim >= (oseglim + (oseglim >> 1))) {
592 break;
593 }
594 }
595
596 /* can't use P2ROUNDUP since seglim may not be power of 2 */
597 segsize = SK_ROUNDUP(segsize, seglim);
598 }
599 ASSERT(segsize != 0 && (segsize % seglim) == 0);
600
601 SK_DF(SK_VERB_MEM, "%s: segsize %zu seglim %zu",
602 srp->srp_name, segsize, seglim);
603
604 /* compute segment count, and recompute segment size */
605 if (srp->srp_cflags & SKMEM_REGION_CR_MONOLITHIC) {
606 segcnt = 1;
607 } else {
608 /*
609 * The adjustments above were done in increments of
610 * SKMEM_MIN_SEG_SIZE. If the object size is greater
611 * than that, ensure that the segment size is a multiple
612 * of the object size.
613 */
614 if (objsize > SKMEM_MIN_SEG_SIZE) {
615 ASSERT(seglim >= objsize);
616 if ((seglim % objsize) != 0) {
617 seglim += (seglim - objsize);
618 }
619 /* recompute segsize; see SK_ROUNDUP comment above */
620 segsize = SK_ROUNDUP(segsize, seglim);
621 }
622
623 segcnt = MAX(1, (segsize / seglim));
624 segsize /= segcnt;
625 }
626
627 SK_DF(SK_VERB_MEM, "%s: segcnt %zu segsize %zu",
628 srp->srp_name, segcnt, segsize);
629
630 /* recompute object count to avoid wastage */
631 objcnt = (segsize * segcnt) / objsize;
632 ASSERT(objcnt != 0);
633 done:
634 srp->srp_c_obj_size = (uint32_t)objsize;
635 srp->srp_c_obj_cnt = (uint32_t)objcnt;
636 srp->srp_c_seg_size = (uint32_t)segsize;
637 srp->srp_seg_cnt = (uint32_t)segcnt;
638
639 SK_DF(SK_VERB_MEM, "%s: objsize %zu objcnt %zu segcnt %zu segsize %zu",
640 srp->srp_name, objsize, objcnt, segcnt, segsize);
641
642 #if SK_LOG
643 if (__improbable(sk_verbose != 0)) {
644 char label[32];
645 (void) snprintf(label, sizeof(label), "REGION_%s:",
646 skmem_region_id2name(srp->srp_id));
647 SK_D("%-16s o:[%4u x %6u -> %4u x %6u]", label,
648 (uint32_t)srp->srp_r_obj_cnt,
649 (uint32_t)srp->srp_r_obj_size,
650 (uint32_t)srp->srp_c_obj_cnt,
651 (uint32_t)srp->srp_c_obj_size);
652 }
653 #endif /* SK_LOG */
654 }
655
656 /*
657 * Create a region.
658 */
659 struct skmem_region *
skmem_region_create(const char * name,struct skmem_region_params * srp,sksegment_ctor_fn_t ctor,sksegment_dtor_fn_t dtor,void * private)660 skmem_region_create(const char *name, struct skmem_region_params *srp,
661 sksegment_ctor_fn_t ctor, sksegment_dtor_fn_t dtor, void *private)
662 {
663 boolean_t pseudo = (srp->srp_cflags & SKMEM_REGION_CR_PSEUDO);
664 uint32_t cflags = srp->srp_cflags;
665 struct skmem_region *skr;
666 uint32_t i;
667
668 ASSERT(srp->srp_id < SKMEM_REGIONS);
669 ASSERT(srp->srp_c_seg_size != 0 &&
670 (pseudo || (srp->srp_c_seg_size % SKMEM_PAGE_SIZE) == 0));
671 ASSERT(srp->srp_seg_cnt != 0);
672 ASSERT(srp->srp_c_obj_cnt == 1 ||
673 (srp->srp_c_seg_size % srp->srp_c_obj_size) == 0);
674 ASSERT(srp->srp_c_obj_size <= srp->srp_c_seg_size);
675
676 skr = zalloc_flags(skr_zone, Z_WAITOK | Z_ZERO);
677 skr->skr_params.srp_r_seg_size = srp->srp_r_seg_size;
678 skr->skr_seg_size = srp->srp_c_seg_size;
679 skr->skr_size = (srp->srp_c_seg_size * srp->srp_seg_cnt);
680 skr->skr_seg_objs = (srp->srp_c_seg_size / srp->srp_c_obj_size);
681
682 if (!pseudo) {
683 skr->skr_seg_max_cnt = srp->srp_seg_cnt;
684
685 /* set alignment to CPU cache line size */
686 skr->skr_params.srp_align = skmem_cpu_cache_line_size();
687
688 /* allocate the allocated-address hash chain */
689 skr->skr_hash_initial = SKMEM_REGION_HASH_INITIAL;
690 skr->skr_hash_limit = SKMEM_REGION_HASH_LIMIT;
691 skr->skr_hash_table = sk_alloc_type_array(struct sksegment_bkt,
692 skr->skr_hash_initial, Z_WAITOK | Z_NOFAIL,
693 skmem_tag_segment_hash);
694 skr->skr_hash_mask = (skr->skr_hash_initial - 1);
695 skr->skr_hash_shift = flsll(srp->srp_c_seg_size) - 1;
696
697 for (i = 0; i < (skr->skr_hash_mask + 1); i++) {
698 TAILQ_INIT(&skr->skr_hash_table[i].sgb_head);
699 }
700 } else {
701 /* this upper bound doesn't apply */
702 skr->skr_seg_max_cnt = 0;
703
704 /* pick up value set by skmem_regions_params_config() */
705 skr->skr_params.srp_align = srp->srp_align;
706 }
707
708 skr->skr_r_obj_size = srp->srp_r_obj_size;
709 skr->skr_r_obj_cnt = srp->srp_r_obj_cnt;
710 skr->skr_c_obj_size = srp->srp_c_obj_size;
711 skr->skr_c_obj_cnt = srp->srp_c_obj_cnt;
712
713 skr->skr_params.srp_md_type = srp->srp_md_type;
714 skr->skr_params.srp_md_subtype = srp->srp_md_subtype;
715 skr->skr_params.srp_max_frags = srp->srp_max_frags;
716
717 skr->skr_seg_ctor = ctor;
718 skr->skr_seg_dtor = dtor;
719 skr->skr_private = private;
720
721 lck_mtx_init(&skr->skr_lock, &skmem_region_lock_grp,
722 &skmem_region_lock_attr);
723
724 TAILQ_INIT(&skr->skr_seg_free);
725 RB_INIT(&skr->skr_seg_tfree);
726
727 skr->skr_id = srp->srp_id;
728 uuid_generate_random(skr->skr_uuid);
729 (void) snprintf(skr->skr_name, sizeof(skr->skr_name),
730 "%s.%s.%s", SKMEM_REGION_PREFIX, srp->srp_name, name);
731
732 SK_DF(SK_VERB_MEM_REGION, "\"%s\": skr 0x%llx ",
733 skr->skr_name, SK_KVA(skr));
734
735 /* sanity check */
736 ASSERT(!(cflags & SKMEM_REGION_CR_GUARD) ||
737 !(cflags & (SKMEM_REGION_CR_KREADONLY | SKMEM_REGION_CR_UREADONLY |
738 SKMEM_REGION_CR_PERSISTENT | SKMEM_REGION_CR_SHAREOK |
739 SKMEM_REGION_CR_IODIR_IN | SKMEM_REGION_CR_IODIR_OUT |
740 SKMEM_REGION_CR_PUREDATA)));
741
742 skr->skr_cflags = cflags;
743 if (cflags & SKMEM_REGION_CR_NOREDIRECT) {
744 skr->skr_mode |= SKR_MODE_NOREDIRECT;
745 }
746 if (cflags & SKMEM_REGION_CR_MMAPOK) {
747 skr->skr_mode |= SKR_MODE_MMAPOK;
748 }
749 if ((cflags & SKMEM_REGION_CR_MMAPOK) &&
750 (cflags & SKMEM_REGION_CR_UREADONLY)) {
751 skr->skr_mode |= SKR_MODE_UREADONLY;
752 }
753 if (cflags & SKMEM_REGION_CR_KREADONLY) {
754 skr->skr_mode |= SKR_MODE_KREADONLY;
755 }
756 if (cflags & SKMEM_REGION_CR_PERSISTENT) {
757 skr->skr_mode |= SKR_MODE_PERSISTENT;
758 }
759 if (cflags & SKMEM_REGION_CR_MONOLITHIC) {
760 skr->skr_mode |= SKR_MODE_MONOLITHIC;
761 }
762 if (cflags & SKMEM_REGION_CR_NOMAGAZINES) {
763 skr->skr_mode |= SKR_MODE_NOMAGAZINES;
764 }
765 if (cflags & SKMEM_REGION_CR_NOCACHE) {
766 skr->skr_mode |= SKR_MODE_NOCACHE;
767 }
768 if (cflags & SKMEM_REGION_CR_SEGPHYSCONTIG) {
769 skr->skr_mode |= SKR_MODE_SEGPHYSCONTIG;
770 }
771 if (cflags & SKMEM_REGION_CR_SHAREOK) {
772 skr->skr_mode |= SKR_MODE_SHAREOK;
773 }
774 if (cflags & SKMEM_REGION_CR_IODIR_IN) {
775 skr->skr_mode |= SKR_MODE_IODIR_IN;
776 }
777 if (cflags & SKMEM_REGION_CR_IODIR_OUT) {
778 skr->skr_mode |= SKR_MODE_IODIR_OUT;
779 }
780 if (cflags & SKMEM_REGION_CR_GUARD) {
781 skr->skr_mode |= SKR_MODE_GUARD;
782 }
783 if (cflags & SKMEM_REGION_CR_PUREDATA) {
784 skr->skr_mode |= SKR_MODE_PUREDATA;
785 }
786 if (cflags & SKMEM_REGION_CR_PSEUDO) {
787 skr->skr_mode |= SKR_MODE_PSEUDO;
788 }
789
790 #if XNU_TARGET_OS_OSX
791 /*
792 * Mark all regions as persistent except for the guard and Intrinsic
793 * regions.
794 * This is to ensure that kernel threads won't be faulting-in while
795 * accessing these memory regions. We have observed various kinds of
796 * kernel panics due to kernel threads faulting on non-wired memory
797 * access when the VM subsystem is not in a state to swap-in the page.
798 */
799 if (!((skr->skr_mode & SKR_MODE_PSEUDO) ||
800 (skr->skr_mode & SKR_MODE_GUARD))) {
801 skr->skr_mode |= SKR_MODE_PERSISTENT;
802 }
803 #endif /* XNU_TARGET_OS_OSX */
804
805 /* SKR_MODE_UREADONLY only takes effect for user task mapping */
806 skr->skr_bufspec.user_writable = !(skr->skr_mode & SKR_MODE_UREADONLY);
807 skr->skr_bufspec.kernel_writable = !(skr->skr_mode & SKR_MODE_KREADONLY);
808 skr->skr_bufspec.purgeable = TRUE;
809 skr->skr_bufspec.inhibitCache = !!(skr->skr_mode & SKR_MODE_NOCACHE);
810 skr->skr_bufspec.physcontig = (skr->skr_mode & SKR_MODE_SEGPHYSCONTIG);
811 skr->skr_bufspec.iodir_in = !!(skr->skr_mode & SKR_MODE_IODIR_IN);
812 skr->skr_bufspec.iodir_out = !!(skr->skr_mode & SKR_MODE_IODIR_OUT);
813 skr->skr_bufspec.puredata = !!(skr->skr_mode & SKR_MODE_PUREDATA);
814 skr->skr_regspec.noRedirect = !!(skr->skr_mode & SKR_MODE_NOREDIRECT);
815
816 /* allocate segment bitmaps */
817 if (!(skr->skr_mode & SKR_MODE_PSEUDO)) {
818 ASSERT(skr->skr_seg_max_cnt != 0);
819 skr->skr_seg_bmap_len = BITMAP_LEN(skr->skr_seg_max_cnt);
820 skr->skr_seg_bmap = sk_alloc_data(BITMAP_SIZE(skr->skr_seg_max_cnt),
821 Z_WAITOK | Z_NOFAIL, skmem_tag_segment_bmap);
822 ASSERT(BITMAP_SIZE(skr->skr_seg_max_cnt) ==
823 (skr->skr_seg_bmap_len * sizeof(*skr->skr_seg_bmap)));
824
825 /* mark all bitmaps as free (bit set) */
826 bitmap_full(skr->skr_seg_bmap, skr->skr_seg_max_cnt);
827 }
828
829 /*
830 * Populate the freelist by allocating all segments for the
831 * region, which will be mapped but not faulted-in, and then
832 * immediately insert each to the freelist. That will in
833 * turn unmap the segment's memory object.
834 */
835 SKR_LOCK(skr);
836 if (skr->skr_mode & SKR_MODE_PSEUDO) {
837 char zone_name[64];
838 (void) snprintf(zone_name, sizeof(zone_name), "%s.reg.%s",
839 SKMEM_ZONE_PREFIX, name);
840 skr->skr_zreg = zone_create(zone_name, skr->skr_c_obj_size,
841 ZC_ZFREE_CLEARMEM | ZC_DESTRUCTIBLE);
842 } else {
843 /* create a backing IOSKRegion object */
844 if ((skr->skr_reg = IOSKRegionCreate(&skr->skr_regspec,
845 (IOSKSize)skr->skr_seg_size,
846 (IOSKCount)skr->skr_seg_max_cnt)) == NULL) {
847 SK_ERR("\%s\": [%u * %u] cflags 0x%b skr_reg failed",
848 skr->skr_name, (uint32_t)skr->skr_seg_size,
849 (uint32_t)skr->skr_seg_max_cnt, skr->skr_cflags,
850 SKMEM_REGION_CR_BITS);
851 goto failed;
852 }
853 }
854
855 ASSERT(skr->skr_seg_objs != 0);
856
857 ++skr->skr_refcnt; /* for caller */
858 SKR_UNLOCK(skr);
859
860 SKMEM_REGION_LOCK();
861 TAILQ_INSERT_TAIL(&skmem_region_head, skr, skr_link);
862 SKMEM_REGION_UNLOCK();
863
864 SK_DF(SK_VERB_MEM_REGION,
865 " [TOTAL] seg (%u*%u) obj (%u*%u) cflags 0x%b",
866 (uint32_t)skr->skr_seg_size, (uint32_t)skr->skr_seg_max_cnt,
867 (uint32_t)skr->skr_c_obj_size, (uint32_t)skr->skr_c_obj_cnt,
868 skr->skr_cflags, SKMEM_REGION_CR_BITS);
869
870 return skr;
871
872 failed:
873 SKR_LOCK_ASSERT_HELD(skr);
874 skmem_region_destroy(skr);
875
876 return NULL;
877 }
878
879 /*
880 * Destroy a region.
881 */
882 static void
skmem_region_destroy(struct skmem_region * skr)883 skmem_region_destroy(struct skmem_region *skr)
884 {
885 struct skmem_region *mskr;
886
887 SKR_LOCK_ASSERT_HELD(skr);
888
889 SK_DF(SK_VERB_MEM_REGION, "\"%s\": skr 0x%llx",
890 skr->skr_name, SK_KVA(skr));
891
892 /*
893 * Panic if we detect there are unfreed segments; the caller
894 * destroying this region is responsible for ensuring that all
895 * allocated segments have been freed prior to getting here.
896 */
897 ASSERT(skr->skr_refcnt == 0);
898 if (skr->skr_seginuse != 0) {
899 panic("%s: '%s' (%p) not empty (%u unfreed)",
900 __func__, skr->skr_name, (void *)skr, skr->skr_seginuse);
901 /* NOTREACHED */
902 __builtin_unreachable();
903 }
904
905 if (skr->skr_link.tqe_next != NULL || skr->skr_link.tqe_prev != NULL) {
906 SKR_UNLOCK(skr);
907 SKMEM_REGION_LOCK();
908 TAILQ_REMOVE(&skmem_region_head, skr, skr_link);
909 SKMEM_REGION_UNLOCK();
910 SKR_LOCK(skr);
911 ASSERT(skr->skr_refcnt == 0);
912 }
913
914 /*
915 * Undo what's done earlier at region creation time.
916 */
917 skmem_region_depopulate(skr);
918 ASSERT(TAILQ_EMPTY(&skr->skr_seg_free));
919 ASSERT(RB_EMPTY(&skr->skr_seg_tfree));
920 ASSERT(skr->skr_seg_free_cnt == 0);
921
922 if (skr->skr_reg != NULL) {
923 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
924 IOSKRegionDestroy(skr->skr_reg);
925 skr->skr_reg = NULL;
926 }
927
928 if (skr->skr_zreg != NULL) {
929 ASSERT(skr->skr_mode & SKR_MODE_PSEUDO);
930 zdestroy(skr->skr_zreg);
931 skr->skr_zreg = NULL;
932 }
933
934 if (skr->skr_seg_bmap != NULL) {
935 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
936 #if (DEBUG || DEVELOPMENT)
937 ASSERT(skr->skr_seg_bmap_len != 0);
938 /* must have been set to vacant (bit set) by now */
939 assert(bitmap_is_full(skr->skr_seg_bmap, skr->skr_seg_max_cnt));
940 #endif /* DEBUG || DEVELOPMENT */
941
942 sk_free_data(skr->skr_seg_bmap, BITMAP_SIZE(skr->skr_seg_max_cnt));
943 skr->skr_seg_bmap = NULL;
944 skr->skr_seg_bmap_len = 0;
945 }
946 ASSERT(skr->skr_seg_bmap_len == 0);
947
948 if (skr->skr_hash_table != NULL) {
949 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
950 #if (DEBUG || DEVELOPMENT)
951 for (uint32_t i = 0; i < (skr->skr_hash_mask + 1); i++) {
952 ASSERT(TAILQ_EMPTY(&skr->skr_hash_table[i].sgb_head));
953 }
954 #endif /* DEBUG || DEVELOPMENT */
955
956 sk_free_type_array(struct sksegment_bkt, skr->skr_hash_mask + 1,
957 skr->skr_hash_table);
958 skr->skr_hash_table = NULL;
959 }
960 if ((mskr = skr->skr_mirror) != NULL) {
961 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
962 skr->skr_mirror = NULL;
963 mskr->skr_mode &= ~SKR_MODE_MIRRORED;
964 }
965 SKR_UNLOCK(skr);
966
967 if (mskr != NULL) {
968 skmem_region_release(mskr);
969 }
970
971 lck_mtx_destroy(&skr->skr_lock, &skmem_region_lock_grp);
972
973 zfree(skr_zone, skr);
974 }
975
976 /*
977 * Mirror mskr (slave) to skr (master).
978 */
979 void
skmem_region_mirror(struct skmem_region * skr,struct skmem_region * mskr)980 skmem_region_mirror(struct skmem_region *skr, struct skmem_region *mskr)
981 {
982 SK_DF(SK_VERB_MEM_REGION, "skr master 0x%llx, slave 0x%llx ",
983 SK_KVA(skr), SK_KVA(mskr));
984
985 SKR_LOCK(skr);
986 ASSERT(!(skr->skr_mode & SKR_MODE_MIRRORED));
987 ASSERT(!(mskr->skr_mode & SKR_MODE_MIRRORED));
988 ASSERT(skr->skr_mirror == NULL);
989
990 /* both regions must share identical parameters */
991 ASSERT(skr->skr_size == mskr->skr_size);
992 ASSERT(skr->skr_seg_size == mskr->skr_seg_size);
993 ASSERT(skr->skr_seg_free_cnt == mskr->skr_seg_free_cnt);
994
995 skr->skr_mirror = mskr;
996 skmem_region_retain(mskr);
997 mskr->skr_mode |= SKR_MODE_MIRRORED;
998 SKR_UNLOCK(skr);
999 }
1000
1001 void
skmem_region_slab_config(struct skmem_region * skr,struct skmem_cache * skm)1002 skmem_region_slab_config(struct skmem_region *skr, struct skmem_cache *skm)
1003 {
1004 SKR_LOCK(skr);
1005 if (skm != NULL) {
1006 ASSERT(!(skr->skr_mode & SKR_MODE_SLAB));
1007 skr->skr_mode |= SKR_MODE_SLAB;
1008 ASSERT(skr->skr_cache == NULL);
1009 skr->skr_cache = skm;
1010 skmem_region_retain_locked(skr);
1011 SKR_UNLOCK(skr);
1012 } else {
1013 ASSERT(skr->skr_mode & SKR_MODE_SLAB);
1014 skr->skr_mode &= ~SKR_MODE_SLAB;
1015 ASSERT(skr->skr_cache != NULL);
1016 skr->skr_cache = NULL;
1017 if (!skmem_region_release_locked(skr)) {
1018 SKR_UNLOCK(skr);
1019 }
1020 }
1021 }
1022
1023 /*
1024 * Common routines for skmem_region_{alloc,mirror_alloc}.
1025 */
1026 static void *
skmem_region_alloc_common(struct skmem_region * skr,struct sksegment * sg)1027 skmem_region_alloc_common(struct skmem_region *skr, struct sksegment *sg)
1028 {
1029 struct sksegment_bkt *sgb;
1030 void *addr;
1031
1032 SKR_LOCK_ASSERT_HELD(skr);
1033
1034 ASSERT(sg->sg_md != NULL);
1035 ASSERT(sg->sg_start != 0 && sg->sg_end != 0);
1036 addr = (void *)sg->sg_start;
1037 sgb = SKMEM_REGION_HASH(skr, addr);
1038 ASSERT(sg->sg_link.tqe_next == NULL);
1039 ASSERT(sg->sg_link.tqe_prev == NULL);
1040 TAILQ_INSERT_HEAD(&sgb->sgb_head, sg, sg_link);
1041
1042 skr->skr_seginuse++;
1043 skr->skr_meminuse += skr->skr_seg_size;
1044 if (sg->sg_state == SKSEG_STATE_MAPPED_WIRED) {
1045 skr->skr_w_meminuse += skr->skr_seg_size;
1046 }
1047 skr->skr_alloc++;
1048
1049 return addr;
1050 }
1051
1052 /*
1053 * Allocate a segment from the region.
1054 */
1055 void *
skmem_region_alloc(struct skmem_region * skr,void ** maddr,struct sksegment ** retsg,struct sksegment ** retsgm,uint32_t skmflag)1056 skmem_region_alloc(struct skmem_region *skr, void **maddr,
1057 struct sksegment **retsg, struct sksegment **retsgm, uint32_t skmflag)
1058 {
1059 struct sksegment *sg = NULL;
1060 struct sksegment *sg1 = NULL;
1061 void *addr = NULL, *addr1 = NULL;
1062 uint32_t retries = 0;
1063
1064 VERIFY(!(skr->skr_mode & SKR_MODE_GUARD));
1065
1066 if (retsg != NULL) {
1067 *retsg = NULL;
1068 }
1069 if (retsgm != NULL) {
1070 *retsgm = NULL;
1071 }
1072
1073 /* SKMEM_NOSLEEP and SKMEM_FAILOK are mutually exclusive */
1074 VERIFY((skmflag & (SKMEM_NOSLEEP | SKMEM_FAILOK)) !=
1075 (SKMEM_NOSLEEP | SKMEM_FAILOK));
1076
1077 SKR_LOCK(skr);
1078 while (sg == NULL) {
1079 /* see if there's a segment in the freelist */
1080 sg = TAILQ_FIRST(&skr->skr_seg_free);
1081 if (sg == NULL) {
1082 /* see if we can grow the freelist */
1083 sg = sksegment_freelist_grow(skr);
1084 if (sg != NULL) {
1085 break;
1086 }
1087
1088 if (skr->skr_mode & SKR_MODE_SLAB) {
1089 SKR_UNLOCK(skr);
1090 /*
1091 * None found; it's possible that the slab
1092 * layer is caching extra amount, so ask
1093 * skmem_cache to reap/purge its caches.
1094 */
1095 skmem_cache_reap_now(skr->skr_cache, TRUE);
1096 SKR_LOCK(skr);
1097 /*
1098 * If we manage to get some freed, try again.
1099 */
1100 if (TAILQ_FIRST(&skr->skr_seg_free) != NULL) {
1101 continue;
1102 }
1103 }
1104
1105 /*
1106 * Give up if this is a non-blocking allocation,
1107 * or if this is a blocking allocation but the
1108 * caller is willing to retry.
1109 */
1110 if (skmflag & (SKMEM_NOSLEEP | SKMEM_FAILOK)) {
1111 break;
1112 }
1113
1114 /* otherwise we wait until one is available */
1115 ++skr->skr_seg_waiters;
1116 (void) msleep(&skr->skr_seg_free, &skr->skr_lock,
1117 (PZERO - 1), skr->skr_name, NULL);
1118 }
1119 }
1120
1121 SKR_LOCK_ASSERT_HELD(skr);
1122
1123 if (sg != NULL) {
1124 retry:
1125 /*
1126 * We have a segment; remove it from the freelist and
1127 * insert it into the allocated-address hash chain.
1128 * Note that this may return NULL if we can't allocate
1129 * the memory descriptor.
1130 */
1131 if (sksegment_freelist_remove(skr, sg, skmflag,
1132 FALSE) == NULL) {
1133 ASSERT(sg->sg_state == SKSEG_STATE_DETACHED);
1134 ASSERT(sg->sg_md == NULL);
1135 ASSERT(sg->sg_start == 0 && sg->sg_end == 0);
1136
1137 /*
1138 * If it's non-blocking allocation, simply just give
1139 * up and let the caller decide when to retry. Else,
1140 * it gets a bit complicated due to the contract we
1141 * have for blocking allocations with the client; the
1142 * most sensible thing to do here is to retry the
1143 * allocation ourselves. Note that we keep using the
1144 * same segment we originally got, since we only need
1145 * the memory descriptor to be allocated for it; thus
1146 * we make sure we don't release the region lock when
1147 * retrying allocation. Doing so is crucial when the
1148 * region is mirrored, since the segment indices on
1149 * both regions need to match.
1150 */
1151 if (skmflag & SKMEM_NOSLEEP) {
1152 SK_ERR("\"%s\": failed to allocate segment "
1153 "(non-sleeping mode)", skr->skr_name);
1154 sg = NULL;
1155 } else {
1156 if (++retries > SKMEM_WDT_MAXTIME) {
1157 panic_plain("\"%s\": failed to "
1158 "allocate segment (sleeping mode) "
1159 "after %u retries\n\n%s",
1160 skr->skr_name, SKMEM_WDT_MAXTIME,
1161 skmem_dump(skr));
1162 /* NOTREACHED */
1163 __builtin_unreachable();
1164 } else {
1165 SK_ERR("\"%s\": failed to allocate "
1166 "segment (sleeping mode): %u "
1167 "retries", skr->skr_name, retries);
1168 }
1169 if (skr->skr_mode & SKR_MODE_SLAB) {
1170 /*
1171 * We can't get any memory descriptor
1172 * for this segment; reap extra cached
1173 * objects from the slab layer and hope
1174 * that we get lucky next time around.
1175 *
1176 * XXX [email protected]: perhaps also
1177 * trigger the zone allocator to do
1178 * its garbage collection here?
1179 */
1180 skmem_cache_reap();
1181 }
1182 delay(1 * USEC_PER_SEC); /* 1 sec */
1183 goto retry;
1184 }
1185 }
1186
1187 if (sg != NULL) {
1188 /* insert to allocated-address hash chain */
1189 addr = skmem_region_alloc_common(skr, sg);
1190 }
1191 }
1192
1193 if (sg == NULL) {
1194 VERIFY(skmflag & (SKMEM_NOSLEEP | SKMEM_FAILOK));
1195 if (skmflag & SKMEM_PANIC) {
1196 VERIFY((skmflag & (SKMEM_NOSLEEP | SKMEM_FAILOK)) ==
1197 SKMEM_NOSLEEP);
1198 /*
1199 * If is a failed non-blocking alloc and the caller
1200 * insists that it must be successful, then panic.
1201 */
1202 panic_plain("\"%s\": skr 0x%p unable to satisfy "
1203 "mandatory allocation\n", skr->skr_name, skr);
1204 /* NOTREACHED */
1205 __builtin_unreachable();
1206 } else {
1207 /*
1208 * Give up if this is a non-blocking allocation,
1209 * or one where the caller is willing to handle
1210 * allocation failures.
1211 */
1212 goto done;
1213 }
1214 }
1215
1216 ASSERT((mach_vm_address_t)addr == sg->sg_start);
1217
1218 #if SK_LOG
1219 SK_DF(SK_VERB_MEM_REGION, "skr 0x%llx sg 0x%llx",
1220 SK_KVA(skr), SK_KVA(sg));
1221 if (skr->skr_mirror == NULL ||
1222 !(skr->skr_mirror->skr_mode & SKR_MODE_MIRRORED)) {
1223 SK_DF(SK_VERB_MEM_REGION, " [%u] [0x%llx-0x%llx)",
1224 sg->sg_index, SK_KVA(sg->sg_start), SK_KVA(sg->sg_end));
1225 } else {
1226 SK_DF(SK_VERB_MEM_REGION, " [%u] [0x%llx-0x%llx) mirrored",
1227 sg->sg_index, SK_KVA(sg), SK_KVA(sg->sg_start),
1228 SK_KVA(sg->sg_end));
1229 }
1230 #endif /* SK_LOG */
1231
1232 /*
1233 * If mirroring, allocate shadow object from slave region.
1234 */
1235 if (skr->skr_mirror != NULL) {
1236 ASSERT(skr->skr_mirror != skr);
1237 ASSERT(!(skr->skr_mode & SKR_MODE_MIRRORED));
1238 ASSERT(skr->skr_mirror->skr_mode & SKR_MODE_MIRRORED);
1239 addr1 = skmem_region_mirror_alloc(skr->skr_mirror, sg, &sg1);
1240 ASSERT(addr1 != NULL);
1241 ASSERT(sg1 != NULL && sg1 != sg);
1242 ASSERT(sg1->sg_index == sg->sg_index);
1243 }
1244
1245 done:
1246 SKR_UNLOCK(skr);
1247
1248 /* return segment metadata to caller if asked (reference not needed) */
1249 if (addr != NULL) {
1250 if (retsg != NULL) {
1251 *retsg = sg;
1252 }
1253 if (retsgm != NULL) {
1254 *retsgm = sg1;
1255 }
1256 }
1257
1258 if (maddr != NULL) {
1259 *maddr = addr1;
1260 }
1261
1262 return addr;
1263 }
1264
1265 /*
1266 * Allocate a segment from a mirror region at the same index. While it
1267 * is somewhat a simplified variant of skmem_region_alloc, keeping it
1268 * separate allows us to avoid further convoluting that routine.
1269 */
1270 static void *
skmem_region_mirror_alloc(struct skmem_region * skr,struct sksegment * sg0,struct sksegment ** retsg)1271 skmem_region_mirror_alloc(struct skmem_region *skr, struct sksegment *sg0,
1272 struct sksegment **retsg)
1273 {
1274 struct sksegment sg_key = { .sg_index = sg0->sg_index };
1275 struct sksegment *sg = NULL;
1276 void *addr = NULL;
1277
1278 ASSERT(skr->skr_mode & SKR_MODE_MIRRORED);
1279 ASSERT(skr->skr_mirror == NULL);
1280 ASSERT(sg0->sg_type == SKSEG_TYPE_ALLOC);
1281
1282 if (retsg != NULL) {
1283 *retsg = NULL;
1284 }
1285
1286 SKR_LOCK(skr);
1287
1288 /*
1289 * See if we can find one in the freelist first. Otherwise,
1290 * create a new segment of the same index and add that to the
1291 * freelist. We would always get a segment since both regions
1292 * are synchronized when it comes to the indices of allocated
1293 * segments.
1294 */
1295 sg = RB_FIND(segtfreehead, &skr->skr_seg_tfree, &sg_key);
1296 if (sg == NULL) {
1297 sg = sksegment_alloc_with_idx(skr, sg0->sg_index);
1298 VERIFY(sg != NULL);
1299 }
1300 VERIFY(sg->sg_index == sg0->sg_index);
1301
1302 /*
1303 * We have a segment; remove it from the freelist and insert
1304 * it into the allocated-address hash chain. This either
1305 * succeeds or panics (SKMEM_PANIC) when a memory descriptor
1306 * can't be allocated.
1307 *
1308 * TODO: consider retrying IOBMD allocation attempts if needed.
1309 */
1310 sg = sksegment_freelist_remove(skr, sg, SKMEM_PANIC, FALSE);
1311 VERIFY(sg != NULL);
1312
1313 /* insert to allocated-address hash chain */
1314 addr = skmem_region_alloc_common(skr, sg);
1315
1316 #if SK_LOG
1317 SK_DF(SK_VERB_MEM_REGION, "skr 0x%llx sg 0x%llx",
1318 SK_KVA(skr), SK_KVA(sg));
1319 SK_DF(SK_VERB_MEM_REGION, " [%u] [0x%llx-0x%llx)",
1320 sg->sg_index, SK_KVA(sg->sg_start), SK_KVA(sg->sg_end));
1321 #endif /* SK_LOG */
1322
1323 SKR_UNLOCK(skr);
1324
1325 /* return segment metadata to caller if asked (reference not needed) */
1326 if (retsg != NULL) {
1327 *retsg = sg;
1328 }
1329
1330 return addr;
1331 }
1332
1333 /*
1334 * Free a segment to the region.
1335 */
1336 void
skmem_region_free(struct skmem_region * skr,void * addr,void * maddr)1337 skmem_region_free(struct skmem_region *skr, void *addr, void *maddr)
1338 {
1339 struct sksegment_bkt *sgb;
1340 struct sksegment *sg, *tsg;
1341
1342 VERIFY(!(skr->skr_mode & SKR_MODE_GUARD));
1343
1344 /*
1345 * Search the hash chain to find a matching segment for the
1346 * given address. If found, remove the segment from the
1347 * hash chain and insert it into the freelist. Otherwise,
1348 * we panic since the caller has given us a bogus address.
1349 */
1350 SKR_LOCK(skr);
1351 sgb = SKMEM_REGION_HASH(skr, addr);
1352 TAILQ_FOREACH_SAFE(sg, &sgb->sgb_head, sg_link, tsg) {
1353 ASSERT(sg->sg_start != 0 && sg->sg_end != 0);
1354 if (sg->sg_start == (mach_vm_address_t)addr) {
1355 TAILQ_REMOVE(&sgb->sgb_head, sg, sg_link);
1356 sg->sg_link.tqe_next = NULL;
1357 sg->sg_link.tqe_prev = NULL;
1358 break;
1359 }
1360 }
1361
1362 ASSERT(sg != NULL);
1363 if (sg->sg_state == SKSEG_STATE_MAPPED_WIRED) {
1364 ASSERT(skr->skr_w_meminuse >= skr->skr_seg_size);
1365 skr->skr_w_meminuse -= skr->skr_seg_size;
1366 }
1367 sksegment_freelist_insert(skr, sg, FALSE);
1368
1369 ASSERT(skr->skr_seginuse != 0);
1370 skr->skr_seginuse--;
1371 skr->skr_meminuse -= skr->skr_seg_size;
1372 skr->skr_free++;
1373
1374 #if SK_LOG
1375 SK_DF(SK_VERB_MEM_REGION, "skr 0x%llx sg 0x%llx",
1376 SK_KVA(skr), SK_KVA(sg));
1377 if (skr->skr_mirror == NULL ||
1378 !(skr->skr_mirror->skr_mode & SKR_MODE_MIRRORED)) {
1379 SK_DF(SK_VERB_MEM_REGION, " [%u] [0x%llx-0x%llx)",
1380 sg->sg_index, SK_KVA(addr),
1381 SK_KVA((uintptr_t)addr + skr->skr_seg_size));
1382 } else {
1383 SK_DF(SK_VERB_MEM_REGION, " [%u] [0x%llx-0x%llx) mirrored",
1384 sg->sg_index, SK_KVA(sg), SK_KVA(addr),
1385 SK_KVA((uintptr_t)addr + skr->skr_seg_size));
1386 }
1387 #endif /* SK_LOG */
1388
1389 /*
1390 * If mirroring, also free shadow object in slave region.
1391 */
1392 if (skr->skr_mirror != NULL) {
1393 ASSERT(maddr != NULL);
1394 ASSERT(skr->skr_mirror != skr);
1395 ASSERT(!(skr->skr_mode & SKR_MODE_MIRRORED));
1396 ASSERT(skr->skr_mirror->skr_mode & SKR_MODE_MIRRORED);
1397 skmem_region_free(skr->skr_mirror, maddr, NULL);
1398 }
1399
1400 /* wake up any blocked threads waiting for a segment */
1401 if (skr->skr_seg_waiters != 0) {
1402 SK_DF(SK_VERB_MEM_REGION,
1403 "sg 0x%llx waking up %u waiters", SK_KVA(sg),
1404 skr->skr_seg_waiters);
1405 skr->skr_seg_waiters = 0;
1406 wakeup(&skr->skr_seg_free);
1407 }
1408 SKR_UNLOCK(skr);
1409 }
1410
1411 __attribute__((always_inline))
1412 static inline void
skmem_region_retain_locked(struct skmem_region * skr)1413 skmem_region_retain_locked(struct skmem_region *skr)
1414 {
1415 SKR_LOCK_ASSERT_HELD(skr);
1416 skr->skr_refcnt++;
1417 ASSERT(skr->skr_refcnt != 0);
1418 }
1419
1420 /*
1421 * Retain a segment.
1422 */
1423 void
skmem_region_retain(struct skmem_region * skr)1424 skmem_region_retain(struct skmem_region *skr)
1425 {
1426 SKR_LOCK(skr);
1427 skmem_region_retain_locked(skr);
1428 SKR_UNLOCK(skr);
1429 }
1430
1431 __attribute__((always_inline))
1432 static inline boolean_t
skmem_region_release_locked(struct skmem_region * skr)1433 skmem_region_release_locked(struct skmem_region *skr)
1434 {
1435 SKR_LOCK_ASSERT_HELD(skr);
1436 ASSERT(skr->skr_refcnt != 0);
1437 if (--skr->skr_refcnt == 0) {
1438 skmem_region_destroy(skr);
1439 return TRUE;
1440 }
1441 return FALSE;
1442 }
1443
1444 /*
1445 * Release (and potentially destroy) a segment.
1446 */
1447 boolean_t
skmem_region_release(struct skmem_region * skr)1448 skmem_region_release(struct skmem_region *skr)
1449 {
1450 boolean_t lastref;
1451
1452 SKR_LOCK(skr);
1453 if (!(lastref = skmem_region_release_locked(skr))) {
1454 SKR_UNLOCK(skr);
1455 }
1456
1457 return lastref;
1458 }
1459
1460 /*
1461 * Depopulate the segment freelist.
1462 */
1463 static void
skmem_region_depopulate(struct skmem_region * skr)1464 skmem_region_depopulate(struct skmem_region *skr)
1465 {
1466 struct sksegment *sg, *tsg;
1467
1468 SK_DF(SK_VERB_MEM_REGION, "\"%s\": skr 0x%llx ",
1469 skr->skr_name, SK_KVA(skr));
1470
1471 SKR_LOCK_ASSERT_HELD(skr);
1472 ASSERT(skr->skr_seg_bmap_len != 0 || (skr->skr_mode & SKR_MODE_PSEUDO));
1473
1474 TAILQ_FOREACH_SAFE(sg, &skr->skr_seg_free, sg_link, tsg) {
1475 struct sksegment *sg0;
1476 uint32_t i;
1477
1478 i = sg->sg_index;
1479 sg0 = sksegment_freelist_remove(skr, sg, 0, TRUE);
1480 VERIFY(sg0 == sg);
1481
1482 sksegment_destroy(skr, sg);
1483 ASSERT(bit_test(skr->skr_seg_bmap[i / BMAPSZ], i % BMAPSZ));
1484 }
1485 }
1486
1487 /*
1488 * Free tree segment compare routine.
1489 */
1490 static int
sksegment_cmp(const struct sksegment * sg1,const struct sksegment * sg2)1491 sksegment_cmp(const struct sksegment *sg1, const struct sksegment *sg2)
1492 {
1493 return sg1->sg_index - sg2->sg_index;
1494 }
1495
1496 /*
1497 * Create a segment.
1498 *
1499 * Upon success, clear the bit for the segment's index in skr_seg_bmap bitmap.
1500 */
1501 static struct sksegment *
sksegment_create(struct skmem_region * skr,uint32_t i)1502 sksegment_create(struct skmem_region *skr, uint32_t i)
1503 {
1504 struct sksegment *sg = NULL;
1505 bitmap_t *bmap;
1506
1507 SKR_LOCK_ASSERT_HELD(skr);
1508
1509 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
1510 ASSERT(i < skr->skr_seg_max_cnt);
1511 ASSERT(skr->skr_reg != NULL);
1512 ASSERT(skr->skr_seg_size == round_page(skr->skr_seg_size));
1513
1514 bmap = &skr->skr_seg_bmap[i / BMAPSZ];
1515 ASSERT(bit_test(*bmap, i % BMAPSZ));
1516
1517 sg = skmem_cache_alloc(skmem_sg_cache, SKMEM_SLEEP);
1518 bzero(sg, sg_size);
1519
1520 sg->sg_region = skr;
1521 sg->sg_index = i;
1522 sg->sg_state = SKSEG_STATE_DETACHED;
1523
1524 /* claim it (clear bit) */
1525 bit_clear(*bmap, i % BMAPSZ);
1526
1527 SK_DF(SK_VERB_MEM_REGION, " [%u] [0x%llx-0x%llx) 0x%b", i,
1528 SK_KVA(sg->sg_start), SK_KVA(sg->sg_end), skr->skr_mode,
1529 SKR_MODE_BITS);
1530
1531 return sg;
1532 }
1533
1534 /*
1535 * Destroy a segment.
1536 *
1537 * Set the bit for the segment's index in skr_seg_bmap bitmap,
1538 * indicating that it is now vacant.
1539 */
1540 static void
sksegment_destroy(struct skmem_region * skr,struct sksegment * sg)1541 sksegment_destroy(struct skmem_region *skr, struct sksegment *sg)
1542 {
1543 uint32_t i = sg->sg_index;
1544 bitmap_t *bmap;
1545
1546 SKR_LOCK_ASSERT_HELD(skr);
1547
1548 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
1549 ASSERT(skr == sg->sg_region);
1550 ASSERT(skr->skr_reg != NULL);
1551 ASSERT(sg->sg_type == SKSEG_TYPE_DESTROYED);
1552 ASSERT(i < skr->skr_seg_max_cnt);
1553
1554 bmap = &skr->skr_seg_bmap[i / BMAPSZ];
1555 ASSERT(!bit_test(*bmap, i % BMAPSZ));
1556
1557 SK_DF(SK_VERB_MEM_REGION, " [%u] [0x%llx-0x%llx) 0x%b",
1558 i, SK_KVA(sg->sg_start), SK_KVA(sg->sg_end),
1559 skr->skr_mode, SKR_MODE_BITS);
1560
1561 /*
1562 * Undo what's done earlier at segment creation time.
1563 */
1564
1565 ASSERT(sg->sg_md == NULL);
1566 ASSERT(sg->sg_start == 0 && sg->sg_end == 0);
1567 ASSERT(sg->sg_state == SKSEG_STATE_DETACHED);
1568
1569 /* release it (set bit) */
1570 bit_set(*bmap, i % BMAPSZ);
1571
1572 skmem_cache_free(skmem_sg_cache, sg);
1573 }
1574
1575 /*
1576 * Insert a segment into freelist (freeing the segment).
1577 */
1578 static void
sksegment_freelist_insert(struct skmem_region * skr,struct sksegment * sg,boolean_t populating)1579 sksegment_freelist_insert(struct skmem_region *skr, struct sksegment *sg,
1580 boolean_t populating)
1581 {
1582 SKR_LOCK_ASSERT_HELD(skr);
1583
1584 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
1585 ASSERT(sg->sg_type != SKSEG_TYPE_FREE);
1586 ASSERT(skr == sg->sg_region);
1587 ASSERT(skr->skr_reg != NULL);
1588 ASSERT(sg->sg_index < skr->skr_seg_max_cnt);
1589
1590 /*
1591 * If the region is being populated, then we're done.
1592 */
1593 if (__improbable(populating)) {
1594 ASSERT(sg->sg_md == NULL);
1595 ASSERT(sg->sg_start == 0 && sg->sg_end == 0);
1596 ASSERT(sg->sg_state == SKSEG_STATE_DETACHED);
1597 } else {
1598 IOSKMemoryBufferRef md;
1599 IOReturn err;
1600
1601 ASSERT(sg->sg_md != NULL);
1602 ASSERT(sg->sg_start != 0 && sg->sg_end != 0);
1603
1604 /*
1605 * Let the client remove the memory from IOMMU, and unwire it.
1606 */
1607 if (skr->skr_seg_dtor != NULL) {
1608 skr->skr_seg_dtor(sg, sg->sg_md, skr->skr_private);
1609 }
1610
1611 ASSERT(sg->sg_state == SKSEG_STATE_MAPPED ||
1612 sg->sg_state == SKSEG_STATE_MAPPED_WIRED);
1613
1614 IOSKRegionClearBufferDebug(skr->skr_reg, sg->sg_index, &md);
1615 VERIFY(sg->sg_md == md);
1616
1617 /* if persistent, unwire this memory now */
1618 if (skr->skr_mode & SKR_MODE_PERSISTENT) {
1619 err = IOSKMemoryUnwire(md);
1620 if (err != kIOReturnSuccess) {
1621 panic("Fail to unwire md %p, err %d", md, err);
1622 }
1623 }
1624
1625 /* mark memory as empty/discarded for consistency */
1626 err = IOSKMemoryDiscard(md);
1627 if (err != kIOReturnSuccess) {
1628 panic("Fail to discard md %p, err %d", md, err);
1629 }
1630
1631 IOSKMemoryDestroy(md);
1632 sg->sg_md = NULL;
1633 sg->sg_start = sg->sg_end = 0;
1634 sg->sg_state = SKSEG_STATE_DETACHED;
1635
1636 ASSERT(skr->skr_memtotal >= skr->skr_seg_size);
1637 skr->skr_memtotal -= skr->skr_seg_size;
1638 }
1639
1640 sg->sg_type = SKSEG_TYPE_FREE;
1641 ASSERT(sg->sg_link.tqe_next == NULL);
1642 ASSERT(sg->sg_link.tqe_prev == NULL);
1643 TAILQ_INSERT_TAIL(&skr->skr_seg_free, sg, sg_link);
1644 ASSERT(sg->sg_node.rbe_left == NULL);
1645 ASSERT(sg->sg_node.rbe_right == NULL);
1646 ASSERT(sg->sg_node.rbe_parent == NULL);
1647 RB_INSERT(segtfreehead, &skr->skr_seg_tfree, sg);
1648 ++skr->skr_seg_free_cnt;
1649 ASSERT(skr->skr_seg_free_cnt <= skr->skr_seg_max_cnt);
1650 }
1651
1652 /*
1653 * Remove a segment from the freelist (allocating the segment).
1654 */
1655 static struct sksegment *
sksegment_freelist_remove(struct skmem_region * skr,struct sksegment * sg,uint32_t skmflag,boolean_t purging)1656 sksegment_freelist_remove(struct skmem_region *skr, struct sksegment *sg,
1657 uint32_t skmflag, boolean_t purging)
1658 {
1659 #pragma unused(skmflag)
1660 mach_vm_address_t segstart;
1661 IOReturn err;
1662
1663 SKR_LOCK_ASSERT_HELD(skr);
1664
1665 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
1666 ASSERT(sg != NULL);
1667 ASSERT(skr == sg->sg_region);
1668 ASSERT(skr->skr_reg != NULL);
1669 ASSERT(sg->sg_type == SKSEG_TYPE_FREE);
1670 ASSERT(sg->sg_index < skr->skr_seg_max_cnt);
1671
1672 #if (DEVELOPMENT || DEBUG)
1673 uint64_t mtbf = skmem_region_get_mtbf();
1674 /*
1675 * MTBF doesn't apply when SKMEM_PANIC is set as caller would assert.
1676 */
1677 if (__improbable(mtbf != 0 && !purging &&
1678 (net_uptime_ms() % mtbf) == 0 &&
1679 !(skmflag & SKMEM_PANIC))) {
1680 SK_ERR("skr \"%s\" 0x%llx sg 0x%llx MTBF failure",
1681 skr->skr_name, SK_KVA(skr), SK_KVA(sg));
1682 net_update_uptime();
1683 return NULL;
1684 }
1685 #endif /* (DEVELOPMENT || DEBUG) */
1686
1687 TAILQ_REMOVE(&skr->skr_seg_free, sg, sg_link);
1688 sg->sg_link.tqe_next = NULL;
1689 sg->sg_link.tqe_prev = NULL;
1690 RB_REMOVE(segtfreehead, &skr->skr_seg_tfree, sg);
1691 sg->sg_node.rbe_left = NULL;
1692 sg->sg_node.rbe_right = NULL;
1693 sg->sg_node.rbe_parent = NULL;
1694
1695 ASSERT(skr->skr_seg_free_cnt != 0);
1696 --skr->skr_seg_free_cnt;
1697
1698 /*
1699 * If the region is being depopulated, then we're done.
1700 */
1701 if (__improbable(purging)) {
1702 ASSERT(sg->sg_md == NULL);
1703 ASSERT(sg->sg_start == 0 && sg->sg_end == 0);
1704 ASSERT(sg->sg_state == SKSEG_STATE_DETACHED);
1705 sg->sg_type = SKSEG_TYPE_DESTROYED;
1706 return sg;
1707 }
1708
1709 ASSERT(sg->sg_md == NULL);
1710 ASSERT(sg->sg_start == 0 && sg->sg_end == 0);
1711 ASSERT(sg->sg_state == SKSEG_STATE_DETACHED);
1712
1713 /* created as non-volatile (mapped) upon success */
1714 if ((sg->sg_md = IOSKMemoryBufferCreate(skr->skr_seg_size,
1715 &skr->skr_bufspec, &segstart)) == NULL) {
1716 ASSERT(sg->sg_type == SKSEG_TYPE_FREE);
1717 if (skmflag & SKMEM_PANIC) {
1718 /* if the caller insists for a success then panic */
1719 panic_plain("\"%s\": skr 0x%p sg 0x%p (idx %u) unable "
1720 "to satisfy mandatory allocation\n", skr->skr_name,
1721 skr, sg, sg->sg_index);
1722 /* NOTREACHED */
1723 __builtin_unreachable();
1724 }
1725 /* reinsert this segment to freelist */
1726 ASSERT(sg->sg_link.tqe_next == NULL);
1727 ASSERT(sg->sg_link.tqe_prev == NULL);
1728 TAILQ_INSERT_HEAD(&skr->skr_seg_free, sg, sg_link);
1729 ASSERT(sg->sg_node.rbe_left == NULL);
1730 ASSERT(sg->sg_node.rbe_right == NULL);
1731 ASSERT(sg->sg_node.rbe_parent == NULL);
1732 RB_INSERT(segtfreehead, &skr->skr_seg_tfree, sg);
1733 ++skr->skr_seg_free_cnt;
1734 return NULL;
1735 }
1736
1737 sg->sg_start = segstart;
1738 sg->sg_end = (segstart + skr->skr_seg_size);
1739 ASSERT(sg->sg_start != 0 && sg->sg_end != 0);
1740
1741 /* mark memory as non-volatile just to be consistent */
1742 err = IOSKMemoryReclaim(sg->sg_md);
1743 if (err != kIOReturnSuccess) {
1744 panic("Fail to reclaim md %p, err %d", sg->sg_md, err);
1745 }
1746
1747 /* if persistent, wire down its memory now */
1748 if (skr->skr_mode & SKR_MODE_PERSISTENT) {
1749 err = IOSKMemoryWire(sg->sg_md);
1750 if (err != kIOReturnSuccess) {
1751 panic("Fail to wire md %p, err %d", sg->sg_md, err);
1752 }
1753 }
1754
1755 err = IOSKRegionSetBuffer(skr->skr_reg, sg->sg_index, sg->sg_md);
1756 if (err != kIOReturnSuccess) {
1757 panic("Fail to set md %p, err %d", sg->sg_md, err);
1758 }
1759
1760 /*
1761 * Let the client wire it and insert to IOMMU, if applicable.
1762 * Try to find out if it's wired and set the right state.
1763 */
1764 if (skr->skr_seg_ctor != NULL) {
1765 skr->skr_seg_ctor(sg, sg->sg_md, skr->skr_private);
1766 }
1767
1768 sg->sg_state = IOSKBufferIsWired(sg->sg_md) ?
1769 SKSEG_STATE_MAPPED_WIRED : SKSEG_STATE_MAPPED;
1770
1771 skr->skr_memtotal += skr->skr_seg_size;
1772
1773 ASSERT(sg->sg_md != NULL);
1774 ASSERT(sg->sg_start != 0 && sg->sg_end != 0);
1775
1776 sg->sg_type = SKSEG_TYPE_ALLOC;
1777 return sg;
1778 }
1779
1780 /*
1781 * Find the first available index and allocate a segment at that index.
1782 */
1783 static struct sksegment *
sksegment_freelist_grow(struct skmem_region * skr)1784 sksegment_freelist_grow(struct skmem_region *skr)
1785 {
1786 struct sksegment *sg = NULL;
1787 uint32_t i, j, idx;
1788
1789 SKR_LOCK_ASSERT_HELD(skr);
1790
1791 ASSERT(!(skr->skr_mode & SKR_MODE_PSEUDO));
1792 ASSERT(skr->skr_seg_bmap_len != 0);
1793 ASSERT(skr->skr_seg_max_cnt != 0);
1794
1795 for (i = 0; i < skr->skr_seg_bmap_len; i++) {
1796 bitmap_t *bmap, mask;
1797 uint32_t end = (BMAPSZ - 1);
1798
1799 if (i == (skr->skr_seg_bmap_len - 1)) {
1800 end = (skr->skr_seg_max_cnt - 1) % BMAPSZ;
1801 }
1802
1803 bmap = &skr->skr_seg_bmap[i];
1804 mask = BMASK64(0, end);
1805
1806 j = ffsll((*bmap) & mask);
1807 if (j == 0) {
1808 continue;
1809 }
1810
1811 --j;
1812 idx = (i * BMAPSZ) + j;
1813
1814 sg = sksegment_alloc_with_idx(skr, idx);
1815
1816 /* we're done */
1817 break;
1818 }
1819
1820 ASSERT((sg != NULL) || (skr->skr_seginuse == skr->skr_seg_max_cnt));
1821 return sg;
1822 }
1823
1824 /*
1825 * Create a single segment at a specific index and add it to the freelist.
1826 */
1827 static struct sksegment *
sksegment_alloc_with_idx(struct skmem_region * skr,uint32_t idx)1828 sksegment_alloc_with_idx(struct skmem_region *skr, uint32_t idx)
1829 {
1830 struct sksegment *sg;
1831
1832 SKR_LOCK_ASSERT_HELD(skr);
1833
1834 if (!bit_test(skr->skr_seg_bmap[idx / BMAPSZ], idx % BMAPSZ)) {
1835 panic("%s: '%s' (%p) idx %u (out of %u) is already allocated",
1836 __func__, skr->skr_name, (void *)skr, idx,
1837 (skr->skr_seg_max_cnt - 1));
1838 /* NOTREACHED */
1839 __builtin_unreachable();
1840 }
1841
1842 /* must not fail, blocking alloc */
1843 sg = sksegment_create(skr, idx);
1844 VERIFY(sg != NULL);
1845 VERIFY(!bit_test(skr->skr_seg_bmap[idx / BMAPSZ], idx % BMAPSZ));
1846
1847 /* populate the freelist */
1848 sksegment_freelist_insert(skr, sg, TRUE);
1849 ASSERT(sg == TAILQ_LAST(&skr->skr_seg_free, segfreehead));
1850 #if (DEVELOPMENT || DEBUG)
1851 struct sksegment sg_key = { .sg_index = sg->sg_index };
1852 ASSERT(sg == RB_FIND(segtfreehead, &skr->skr_seg_tfree, &sg_key));
1853 #endif /* (DEVELOPMENT || DEBUG) */
1854
1855 SK_DF(SK_VERB_MEM_REGION, "sg %u/%u", (idx + 1), skr->skr_seg_max_cnt);
1856
1857 return sg;
1858 }
1859
1860 /*
1861 * Rescale the regions's allocated-address hash table.
1862 */
1863 static void
skmem_region_hash_rescale(struct skmem_region * skr)1864 skmem_region_hash_rescale(struct skmem_region *skr)
1865 {
1866 struct sksegment_bkt *old_table, *new_table;
1867 size_t old_size, new_size;
1868 uint32_t i, moved = 0;
1869
1870 if (skr->skr_mode & SKR_MODE_PSEUDO) {
1871 ASSERT(skr->skr_hash_table == NULL);
1872 /* this is no-op for pseudo region */
1873 return;
1874 }
1875
1876 ASSERT(skr->skr_hash_table != NULL);
1877 /* insist that we are executing in the update thread call context */
1878 ASSERT(sk_is_region_update_protected());
1879
1880 /*
1881 * To get small average lookup time (lookup depth near 1.0), the hash
1882 * table size should be roughly the same (not necessarily equivalent)
1883 * as the region size.
1884 */
1885 new_size = MAX(skr->skr_hash_initial,
1886 (1 << (flsll(3 * skr->skr_seginuse + 4) - 2)));
1887 new_size = MIN(skr->skr_hash_limit, new_size);
1888 old_size = (skr->skr_hash_mask + 1);
1889
1890 if ((old_size >> 1) <= new_size && new_size <= (old_size << 1)) {
1891 return;
1892 }
1893
1894 new_table = sk_alloc_type_array(struct sksegment_bkt, new_size,
1895 Z_NOWAIT, skmem_tag_segment_hash);
1896 if (__improbable(new_table == NULL)) {
1897 return;
1898 }
1899
1900 for (i = 0; i < new_size; i++) {
1901 TAILQ_INIT(&new_table[i].sgb_head);
1902 }
1903
1904 SKR_LOCK(skr);
1905
1906 old_size = (skr->skr_hash_mask + 1);
1907 old_table = skr->skr_hash_table;
1908
1909 skr->skr_hash_mask = (uint32_t)(new_size - 1);
1910 skr->skr_hash_table = new_table;
1911 skr->skr_rescale++;
1912
1913 for (i = 0; i < old_size; i++) {
1914 struct sksegment_bkt *sgb = &old_table[i];
1915 struct sksegment_bkt *new_sgb;
1916 struct sksegment *sg;
1917
1918 while ((sg = TAILQ_FIRST(&sgb->sgb_head)) != NULL) {
1919 TAILQ_REMOVE(&sgb->sgb_head, sg, sg_link);
1920 ASSERT(sg->sg_start != 0 && sg->sg_end != 0);
1921 new_sgb = SKMEM_REGION_HASH(skr, sg->sg_start);
1922 TAILQ_INSERT_TAIL(&new_sgb->sgb_head, sg, sg_link);
1923 ++moved;
1924 }
1925 ASSERT(TAILQ_EMPTY(&sgb->sgb_head));
1926 }
1927
1928 SK_DF(SK_VERB_MEM_REGION,
1929 "skr 0x%llx old_size %u new_size %u [%u moved]", SK_KVA(skr),
1930 (uint32_t)old_size, (uint32_t)new_size, moved);
1931
1932 SKR_UNLOCK(skr);
1933
1934 sk_free_type_array(struct sksegment_bkt, old_size, old_table);
1935 }
1936
1937 /*
1938 * Apply a function to operate on all regions.
1939 */
1940 static void
skmem_region_applyall(void (* func)(struct skmem_region *))1941 skmem_region_applyall(void (*func)(struct skmem_region *))
1942 {
1943 struct skmem_region *skr;
1944
1945 net_update_uptime();
1946
1947 SKMEM_REGION_LOCK();
1948 TAILQ_FOREACH(skr, &skmem_region_head, skr_link) {
1949 func(skr);
1950 }
1951 SKMEM_REGION_UNLOCK();
1952 }
1953
1954 static void
skmem_region_update(struct skmem_region * skr)1955 skmem_region_update(struct skmem_region *skr)
1956 {
1957 SKMEM_REGION_LOCK_ASSERT_HELD();
1958
1959 /* insist that we are executing in the update thread call context */
1960 ASSERT(sk_is_region_update_protected());
1961
1962 SKR_LOCK(skr);
1963 /*
1964 * If there are threads blocked waiting for an available
1965 * segment, wake them up periodically so they can issue
1966 * another skmem_cache_reap() to reclaim resources cached
1967 * by skmem_cache.
1968 */
1969 if (skr->skr_seg_waiters != 0) {
1970 SK_DF(SK_VERB_MEM_REGION,
1971 "waking up %u waiters to reclaim", skr->skr_seg_waiters);
1972 skr->skr_seg_waiters = 0;
1973 wakeup(&skr->skr_seg_free);
1974 }
1975 SKR_UNLOCK(skr);
1976
1977 /*
1978 * Rescale the hash table if needed.
1979 */
1980 skmem_region_hash_rescale(skr);
1981 }
1982
1983 /*
1984 * Thread call callback for update.
1985 */
1986 static void
skmem_region_update_func(thread_call_param_t dummy,thread_call_param_t arg)1987 skmem_region_update_func(thread_call_param_t dummy, thread_call_param_t arg)
1988 {
1989 #pragma unused(dummy, arg)
1990 sk_protect_t protect;
1991
1992 protect = sk_region_update_protect();
1993 skmem_region_applyall(skmem_region_update);
1994 sk_region_update_unprotect(protect);
1995
1996 skmem_dispatch(skmem_region_update_tc, NULL,
1997 (skmem_region_update_interval * NSEC_PER_SEC));
1998 }
1999
2000 boolean_t
skmem_region_for_pp(skmem_region_id_t id)2001 skmem_region_for_pp(skmem_region_id_t id)
2002 {
2003 int i;
2004
2005 for (i = 0; i < SKMEM_PP_REGIONS; i++) {
2006 if (id == skmem_pp_region_ids[i]) {
2007 return TRUE;
2008 }
2009 }
2010 return FALSE;
2011 }
2012
2013 void
skmem_region_get_stats(struct skmem_region * skr,struct sk_stats_region * sreg)2014 skmem_region_get_stats(struct skmem_region *skr, struct sk_stats_region *sreg)
2015 {
2016 bzero(sreg, sizeof(*sreg));
2017
2018 (void) snprintf(sreg->sreg_name, sizeof(sreg->sreg_name),
2019 "%s", skr->skr_name);
2020 uuid_copy(sreg->sreg_uuid, skr->skr_uuid);
2021 sreg->sreg_id = (sk_stats_region_id_t)skr->skr_id;
2022 sreg->sreg_mode = skr->skr_mode;
2023
2024 sreg->sreg_r_seg_size = skr->skr_params.srp_r_seg_size;
2025 sreg->sreg_c_seg_size = skr->skr_seg_size;
2026 sreg->sreg_seg_cnt = skr->skr_seg_max_cnt;
2027 sreg->sreg_seg_objs = skr->skr_seg_objs;
2028 sreg->sreg_r_obj_size = skr->skr_r_obj_size;
2029 sreg->sreg_r_obj_cnt = skr->skr_r_obj_cnt;
2030 sreg->sreg_c_obj_size = skr->skr_c_obj_size;
2031 sreg->sreg_c_obj_cnt = skr->skr_c_obj_cnt;
2032 sreg->sreg_align = skr->skr_align;
2033 sreg->sreg_max_frags = skr->skr_max_frags;
2034
2035 sreg->sreg_meminuse = skr->skr_meminuse;
2036 sreg->sreg_w_meminuse = skr->skr_w_meminuse;
2037 sreg->sreg_memtotal = skr->skr_memtotal;
2038 sreg->sreg_seginuse = skr->skr_seginuse;
2039 sreg->sreg_rescale = skr->skr_rescale;
2040 sreg->sreg_hash_size = (skr->skr_hash_mask + 1);
2041 sreg->sreg_alloc = skr->skr_alloc;
2042 sreg->sreg_free = skr->skr_free;
2043 }
2044
2045 static size_t
skmem_region_mib_get_stats(struct skmem_region * skr,void * out,size_t len)2046 skmem_region_mib_get_stats(struct skmem_region *skr, void *out, size_t len)
2047 {
2048 size_t actual_space = sizeof(struct sk_stats_region);
2049 struct sk_stats_region *sreg = out;
2050
2051 if (out == NULL || len < actual_space) {
2052 goto done;
2053 }
2054
2055 skmem_region_get_stats(skr, sreg);
2056
2057 done:
2058 return actual_space;
2059 }
2060
2061 static int
2062 skmem_region_mib_get_sysctl SYSCTL_HANDLER_ARGS
2063 {
2064 #pragma unused(arg1, arg2, oidp)
2065 struct skmem_region *skr;
2066 size_t actual_space;
2067 size_t buffer_space;
2068 size_t allocated_space;
2069 caddr_t buffer = NULL;
2070 caddr_t scan;
2071 int error = 0;
2072
2073 if (!kauth_cred_issuser(kauth_cred_get())) {
2074 return EPERM;
2075 }
2076
2077 net_update_uptime();
2078 buffer_space = req->oldlen;
2079 if (req->oldptr != USER_ADDR_NULL && buffer_space != 0) {
2080 if (buffer_space > SK_SYSCTL_ALLOC_MAX) {
2081 buffer_space = SK_SYSCTL_ALLOC_MAX;
2082 }
2083 allocated_space = buffer_space;
2084 buffer = sk_alloc_data(allocated_space, Z_WAITOK, skmem_tag_region_mib);
2085 if (__improbable(buffer == NULL)) {
2086 return ENOBUFS;
2087 }
2088 } else if (req->oldptr == USER_ADDR_NULL) {
2089 buffer_space = 0;
2090 }
2091 actual_space = 0;
2092 scan = buffer;
2093
2094 SKMEM_REGION_LOCK();
2095 TAILQ_FOREACH(skr, &skmem_region_head, skr_link) {
2096 size_t size = skmem_region_mib_get_stats(skr, scan, buffer_space);
2097 if (scan != NULL) {
2098 if (buffer_space < size) {
2099 /* supplied buffer too small, stop copying */
2100 error = ENOMEM;
2101 break;
2102 }
2103 scan += size;
2104 buffer_space -= size;
2105 }
2106 actual_space += size;
2107 }
2108 SKMEM_REGION_UNLOCK();
2109
2110 if (actual_space != 0) {
2111 int out_error = SYSCTL_OUT(req, buffer, actual_space);
2112 if (out_error != 0) {
2113 error = out_error;
2114 }
2115 }
2116 if (buffer != NULL) {
2117 sk_free_data(buffer, allocated_space);
2118 }
2119
2120 return error;
2121 }
2122
2123 #if SK_LOG
2124 const char *
skmem_region_id2name(skmem_region_id_t id)2125 skmem_region_id2name(skmem_region_id_t id)
2126 {
2127 const char *name;
2128 switch (id) {
2129 case SKMEM_REGION_SCHEMA:
2130 name = "SCHEMA";
2131 break;
2132
2133 case SKMEM_REGION_RING:
2134 name = "RING";
2135 break;
2136
2137 case SKMEM_REGION_BUF:
2138 name = "BUF";
2139 break;
2140
2141 case SKMEM_REGION_RXBUF:
2142 name = "RXBUF";
2143 break;
2144
2145 case SKMEM_REGION_TXBUF:
2146 name = "TXBUF";
2147 break;
2148
2149 case SKMEM_REGION_UMD:
2150 name = "UMD";
2151 break;
2152
2153 case SKMEM_REGION_TXAUSD:
2154 name = "TXAUSD";
2155 break;
2156
2157 case SKMEM_REGION_RXFUSD:
2158 name = "RXFUSD";
2159 break;
2160
2161 case SKMEM_REGION_USTATS:
2162 name = "USTATS";
2163 break;
2164
2165 case SKMEM_REGION_FLOWADV:
2166 name = "FLOWADV";
2167 break;
2168
2169 case SKMEM_REGION_NEXUSADV:
2170 name = "NEXUSADV";
2171 break;
2172
2173 case SKMEM_REGION_SYSCTLS:
2174 name = "SYSCTLS";
2175 break;
2176
2177 case SKMEM_REGION_GUARD_HEAD:
2178 name = "HEADGUARD";
2179 break;
2180
2181 case SKMEM_REGION_GUARD_TAIL:
2182 name = "TAILGUARD";
2183 break;
2184
2185 case SKMEM_REGION_KMD:
2186 name = "KMD";
2187 break;
2188
2189 case SKMEM_REGION_RXKMD:
2190 name = "RXKMD";
2191 break;
2192
2193 case SKMEM_REGION_TXKMD:
2194 name = "TXKMD";
2195 break;
2196
2197 case SKMEM_REGION_TXAKSD:
2198 name = "TXAKSD";
2199 break;
2200
2201 case SKMEM_REGION_RXFKSD:
2202 name = "RXFKSD";
2203 break;
2204
2205 case SKMEM_REGION_KSTATS:
2206 name = "KSTATS";
2207 break;
2208
2209 case SKMEM_REGION_KBFT:
2210 name = "KBFT";
2211 break;
2212
2213 case SKMEM_REGION_UBFT:
2214 name = "UBFT";
2215 break;
2216
2217 case SKMEM_REGION_RXKBFT:
2218 name = "RXKBFT";
2219 break;
2220
2221 case SKMEM_REGION_TXKBFT:
2222 name = "TXKBFT";
2223 break;
2224
2225 case SKMEM_REGION_INTRINSIC:
2226 name = "INTRINSIC";
2227 break;
2228
2229 default:
2230 name = "UNKNOWN";
2231 break;
2232 }
2233
2234 return name;
2235 }
2236 #endif /* SK_LOG */
2237
2238 #if (DEVELOPMENT || DEBUG)
2239 uint64_t
skmem_region_get_mtbf(void)2240 skmem_region_get_mtbf(void)
2241 {
2242 return skmem_region_mtbf;
2243 }
2244
2245 void
skmem_region_set_mtbf(uint64_t newval)2246 skmem_region_set_mtbf(uint64_t newval)
2247 {
2248 if (newval < SKMEM_REGION_MTBF_MIN) {
2249 if (newval != 0) {
2250 newval = SKMEM_REGION_MTBF_MIN;
2251 }
2252 } else if (newval > SKMEM_REGION_MTBF_MAX) {
2253 newval = SKMEM_REGION_MTBF_MAX;
2254 }
2255
2256 if (skmem_region_mtbf != newval) {
2257 atomic_set_64(&skmem_region_mtbf, newval);
2258 SK_ERR("MTBF set to %llu msec", skmem_region_mtbf);
2259 }
2260 }
2261
2262 static int
skmem_region_mtbf_sysctl(struct sysctl_oid * oidp,void * arg1,int arg2,struct sysctl_req * req)2263 skmem_region_mtbf_sysctl(struct sysctl_oid *oidp, void *arg1, int arg2,
2264 struct sysctl_req *req)
2265 {
2266 #pragma unused(oidp, arg1, arg2)
2267 int changed, error;
2268 uint64_t newval;
2269
2270 _CASSERT(sizeof(skmem_region_mtbf) == sizeof(uint64_t));
2271 if ((error = sysctl_io_number(req, skmem_region_mtbf,
2272 sizeof(uint64_t), &newval, &changed)) == 0) {
2273 if (changed) {
2274 skmem_region_set_mtbf(newval);
2275 }
2276 }
2277 return error;
2278 }
2279 #endif /* (DEVELOPMENT || DEBUG) */
2280