1 /*
2 * Copyright (c) 2015-2021 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 /*
30 * Copyright (C) 2014 Giuseppe Lettieri. All rights reserved.
31 *
32 * Redistribution and use in source and binary forms, with or without
33 * modification, are permitted provided that the following conditions
34 * are met:
35 * 1. Redistributions of source code must retain the above copyright
36 * notice, this list of conditions and the following disclaimer.
37 * 2. Redistributions in binary form must reproduce the above copyright
38 * notice, this list of conditions and the following disclaimer in the
39 * documentation and/or other materials provided with the distribution.
40 *
41 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
42 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
43 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
44 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
45 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
46 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
47 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
48 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
49 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
50 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
51 * SUCH DAMAGE.
52 */
53
54 #include <skywalk/os_skywalk_private.h>
55 #include <skywalk/nexus/upipe/nx_user_pipe.h>
56
57 #define NX_UPIPE_RINGSIZE 128 /* default ring size */
58 #define NX_UPIPE_MAXRINGS NX_MAX_NUM_RING_PAIR
59 #define NX_UPIPE_MINSLOTS 2 /* XXX same as above */
60 #define NX_UPIPE_MAXSLOTS 4096 /* XXX same as above */
61 #define NX_UPIPE_BUFSIZE (2 * 1024)
62 #define NX_UPIPE_MINBUFSIZE 1024
63 #define NX_UPIPE_MAXBUFSIZE (16 * 1024)
64 #define NX_UPIPE_MHINTS NEXUS_MHINTS_NORMAL
65
66 static int nx_upipe_na_alloc(struct nexus_adapter *, uint32_t);
67 static struct nexus_upipe_adapter *nx_upipe_find(struct nexus_adapter *,
68 uint32_t);
69 static int nx_upipe_na_add(struct nexus_adapter *,
70 struct nexus_upipe_adapter *);
71 static void nx_upipe_na_remove(struct nexus_adapter *,
72 struct nexus_upipe_adapter *);
73 static int nx_upipe_na_txsync(struct __kern_channel_ring *,
74 struct proc *, uint32_t);
75 static int nx_upipe_na_txsync_locked(struct __kern_channel_ring *,
76 struct proc *, uint32_t, int *, boolean_t);
77 static int nx_upipe_na_rxsync(struct __kern_channel_ring *,
78 struct proc *, uint32_t);
79 static int nx_upipe_na_krings_create(struct nexus_adapter *,
80 struct kern_channel *);
81 static int nx_upipe_na_activate(struct nexus_adapter *, na_activate_mode_t);
82 static void nx_upipe_na_krings_delete(struct nexus_adapter *,
83 struct kern_channel *, boolean_t);
84 static void nx_upipe_na_dtor(struct nexus_adapter *);
85
86 static void nx_upipe_dom_init(struct nxdom *);
87 static void nx_upipe_dom_terminate(struct nxdom *);
88 static void nx_upipe_dom_fini(struct nxdom *);
89 static int nx_upipe_dom_bind_port(struct kern_nexus *, nexus_port_t *,
90 struct nxbind *, void *);
91 static int nx_upipe_dom_unbind_port(struct kern_nexus *, nexus_port_t);
92 static int nx_upipe_dom_connect(struct kern_nexus_domain_provider *,
93 struct kern_nexus *, struct kern_channel *, struct chreq *,
94 struct kern_channel *, struct nxbind *, struct proc *);
95 static void nx_upipe_dom_disconnect(struct kern_nexus_domain_provider *,
96 struct kern_nexus *, struct kern_channel *);
97 static void nx_upipe_dom_defunct(struct kern_nexus_domain_provider *,
98 struct kern_nexus *, struct kern_channel *, struct proc *);
99 static void nx_upipe_dom_defunct_finalize(struct kern_nexus_domain_provider *,
100 struct kern_nexus *, struct kern_channel *, boolean_t);
101
102 static int nx_upipe_prov_init(struct kern_nexus_domain_provider *);
103 static int nx_upipe_prov_params_adjust(
104 const struct kern_nexus_domain_provider *, const struct nxprov_params *,
105 struct nxprov_adjusted_params *);
106 static int nx_upipe_prov_params(struct kern_nexus_domain_provider *,
107 const uint32_t, const struct nxprov_params *, struct nxprov_params *,
108 struct skmem_region_params[SKMEM_REGIONS], uint32_t);
109 static int nx_upipe_prov_mem_new(struct kern_nexus_domain_provider *,
110 struct kern_nexus *, struct nexus_adapter *);
111 static void nx_upipe_prov_fini(struct kern_nexus_domain_provider *);
112 static int nx_upipe_prov_nx_ctor(struct kern_nexus *);
113 static void nx_upipe_prov_nx_dtor(struct kern_nexus *);
114
115 static struct nexus_upipe_adapter *na_upipe_alloc(zalloc_flags_t);
116 static void na_upipe_free(struct nexus_adapter *);
117
118 static struct nx_upipe *nx_upipe_alloc(zalloc_flags_t);
119 static void nx_upipe_free(struct nx_upipe *);
120
121 #if (DEVELOPMENT || DEBUG)
122 static uint32_t nx_upipe_mhints = 0;
123 SYSCTL_NODE(_kern_skywalk, OID_AUTO, upipe, CTLFLAG_RW | CTLFLAG_LOCKED,
124 0, "Skywalk upipe tuning");
125 SYSCTL_UINT(_kern_skywalk_upipe, OID_AUTO, nx_mhints,
126 CTLFLAG_RW | CTLFLAG_LOCKED, &nx_upipe_mhints, 0,
127 "upipe nexus memory usage hints");
128 #endif /* (DEVELOPMENT || DEBUG) */
129
130 struct nxdom nx_upipe_dom_s = {
131 .nxdom_prov_head =
132 STAILQ_HEAD_INITIALIZER(nx_upipe_dom_s.nxdom_prov_head),
133 .nxdom_type = NEXUS_TYPE_USER_PIPE,
134 .nxdom_md_type = NEXUS_META_TYPE_QUANTUM,
135 .nxdom_md_subtype = NEXUS_META_SUBTYPE_PAYLOAD,
136 .nxdom_name = "upipe",
137 .nxdom_ports = {
138 .nb_def = 2,
139 .nb_min = 2,
140 .nb_max = 2,
141 },
142 .nxdom_tx_rings = {
143 .nb_def = 1,
144 .nb_min = 1,
145 .nb_max = NX_UPIPE_MAXRINGS,
146 },
147 .nxdom_rx_rings = {
148 .nb_def = 1,
149 .nb_min = 1,
150 .nb_max = NX_UPIPE_MAXRINGS,
151 },
152 .nxdom_tx_slots = {
153 .nb_def = NX_UPIPE_RINGSIZE,
154 .nb_min = NX_UPIPE_MINSLOTS,
155 .nb_max = NX_UPIPE_MAXSLOTS,
156 },
157 .nxdom_rx_slots = {
158 .nb_def = NX_UPIPE_RINGSIZE,
159 .nb_min = NX_UPIPE_MINSLOTS,
160 .nb_max = NX_UPIPE_MAXSLOTS,
161 },
162 .nxdom_buf_size = {
163 .nb_def = NX_UPIPE_BUFSIZE,
164 .nb_min = NX_UPIPE_MINBUFSIZE,
165 .nb_max = NX_UPIPE_MAXBUFSIZE,
166 },
167 .nxdom_large_buf_size = {
168 .nb_def = 0,
169 .nb_min = 0,
170 .nb_max = 0,
171 },
172 .nxdom_meta_size = {
173 .nb_def = NX_METADATA_OBJ_MIN_SZ,
174 .nb_min = NX_METADATA_OBJ_MIN_SZ,
175 .nb_max = NX_METADATA_USR_MAX_SZ,
176 },
177 .nxdom_stats_size = {
178 .nb_def = 0,
179 .nb_min = 0,
180 .nb_max = NX_STATS_MAX_SZ,
181 },
182 .nxdom_pipes = {
183 .nb_def = 0,
184 .nb_min = 0,
185 .nb_max = NX_UPIPE_MAXPIPES,
186 },
187 .nxdom_mhints = {
188 .nb_def = NX_UPIPE_MHINTS,
189 .nb_min = NEXUS_MHINTS_NORMAL,
190 .nb_max = (NEXUS_MHINTS_NORMAL | NEXUS_MHINTS_WILLNEED |
191 NEXUS_MHINTS_LOWLATENCY | NEXUS_MHINTS_HIUSE),
192 },
193 .nxdom_flowadv_max = {
194 .nb_def = 0,
195 .nb_min = 0,
196 .nb_max = NX_FLOWADV_MAX,
197 },
198 .nxdom_nexusadv_size = {
199 .nb_def = 0,
200 .nb_min = 0,
201 .nb_max = NX_NEXUSADV_MAX_SZ,
202 },
203 .nxdom_capabilities = {
204 .nb_def = NXPCAP_USER_CHANNEL,
205 .nb_min = NXPCAP_USER_CHANNEL,
206 .nb_max = NXPCAP_USER_CHANNEL,
207 },
208 .nxdom_qmap = {
209 .nb_def = NEXUS_QMAP_TYPE_INVALID,
210 .nb_min = NEXUS_QMAP_TYPE_INVALID,
211 .nb_max = NEXUS_QMAP_TYPE_INVALID,
212 },
213 .nxdom_max_frags = {
214 .nb_def = NX_PBUF_FRAGS_DEFAULT,
215 .nb_min = NX_PBUF_FRAGS_MIN,
216 .nb_max = NX_PBUF_FRAGS_DEFAULT,
217 },
218 .nxdom_init = nx_upipe_dom_init,
219 .nxdom_terminate = nx_upipe_dom_terminate,
220 .nxdom_fini = nx_upipe_dom_fini,
221 .nxdom_find_port = NULL,
222 .nxdom_port_is_reserved = NULL,
223 .nxdom_bind_port = nx_upipe_dom_bind_port,
224 .nxdom_unbind_port = nx_upipe_dom_unbind_port,
225 .nxdom_connect = nx_upipe_dom_connect,
226 .nxdom_disconnect = nx_upipe_dom_disconnect,
227 .nxdom_defunct = nx_upipe_dom_defunct,
228 .nxdom_defunct_finalize = nx_upipe_dom_defunct_finalize,
229 };
230
231 static struct kern_nexus_domain_provider nx_upipe_prov_s = {
232 .nxdom_prov_name = NEXUS_PROVIDER_USER_PIPE,
233 .nxdom_prov_flags = NXDOMPROVF_DEFAULT,
234 .nxdom_prov_cb = {
235 .dp_cb_init = nx_upipe_prov_init,
236 .dp_cb_fini = nx_upipe_prov_fini,
237 .dp_cb_params = nx_upipe_prov_params,
238 .dp_cb_mem_new = nx_upipe_prov_mem_new,
239 .dp_cb_config = NULL,
240 .dp_cb_nx_ctor = nx_upipe_prov_nx_ctor,
241 .dp_cb_nx_dtor = nx_upipe_prov_nx_dtor,
242 .dp_cb_nx_mem_info = NULL,
243 .dp_cb_nx_mib_get = NULL,
244 .dp_cb_nx_stop = NULL,
245 },
246 };
247
248 static SKMEM_TYPE_DEFINE(na_upipe_zone, struct nexus_upipe_adapter);
249
250 static SKMEM_TYPE_DEFINE(nx_upipe_zone, struct nx_upipe);
251
252 #define SKMEM_TAG_PIPES "com.apple.skywalk.pipes"
253 static SKMEM_TAG_DEFINE(skmem_tag_pipes, SKMEM_TAG_PIPES);
254
255 static void
nx_upipe_dom_init(struct nxdom * nxdom)256 nx_upipe_dom_init(struct nxdom *nxdom)
257 {
258 SK_LOCK_ASSERT_HELD();
259 ASSERT(!(nxdom->nxdom_flags & NEXUSDOMF_INITIALIZED));
260
261 (void) nxdom_prov_add(nxdom, &nx_upipe_prov_s);
262 }
263
264 static void
nx_upipe_dom_terminate(struct nxdom * nxdom)265 nx_upipe_dom_terminate(struct nxdom *nxdom)
266 {
267 struct kern_nexus_domain_provider *nxdom_prov, *tnxdp;
268
269 STAILQ_FOREACH_SAFE(nxdom_prov, &nxdom->nxdom_prov_head,
270 nxdom_prov_link, tnxdp) {
271 (void) nxdom_prov_del(nxdom_prov);
272 }
273 }
274
275 static void
nx_upipe_dom_fini(struct nxdom * nxdom)276 nx_upipe_dom_fini(struct nxdom *nxdom)
277 {
278 #pragma unused(nxdom)
279 }
280
281 static int
nx_upipe_prov_init(struct kern_nexus_domain_provider * nxdom_prov)282 nx_upipe_prov_init(struct kern_nexus_domain_provider *nxdom_prov)
283 {
284 #pragma unused(nxdom_prov)
285 SK_D("initializing %s", nxdom_prov->nxdom_prov_name);
286 return 0;
287 }
288
289 static int
nx_upipe_prov_params_adjust(const struct kern_nexus_domain_provider * nxdom_prov,const struct nxprov_params * nxp,struct nxprov_adjusted_params * adj)290 nx_upipe_prov_params_adjust(const struct kern_nexus_domain_provider *nxdom_prov,
291 const struct nxprov_params *nxp, struct nxprov_adjusted_params *adj)
292 {
293 #pragma unused(nxdom_prov, nxp)
294 /*
295 * User pipe requires double the amount of rings.
296 * The ring counts must also be symmetrical.
297 */
298 if (*(adj->adj_tx_rings) != *(adj->adj_rx_rings)) {
299 SK_ERR("rings: tx (%u) != rx (%u)", *(adj->adj_tx_rings),
300 *(adj->adj_rx_rings));
301 return EINVAL;
302 }
303 *(adj->adj_tx_rings) *= 2;
304 *(adj->adj_rx_rings) *= 2;
305 return 0;
306 }
307
308 static int
nx_upipe_prov_params(struct kern_nexus_domain_provider * nxdom_prov,const uint32_t req,const struct nxprov_params * nxp0,struct nxprov_params * nxp,struct skmem_region_params srp[SKMEM_REGIONS],uint32_t pp_region_config_flags)309 nx_upipe_prov_params(struct kern_nexus_domain_provider *nxdom_prov,
310 const uint32_t req, const struct nxprov_params *nxp0,
311 struct nxprov_params *nxp, struct skmem_region_params srp[SKMEM_REGIONS],
312 uint32_t pp_region_config_flags)
313 {
314 struct nxdom *nxdom = nxdom_prov->nxdom_prov_dom;
315 int err;
316
317 err = nxprov_params_adjust(nxdom_prov, req, nxp0, nxp, srp,
318 nxdom, nxdom, nxdom, pp_region_config_flags,
319 nx_upipe_prov_params_adjust);
320 #if (DEVELOPMENT || DEBUG)
321 /* sysctl override */
322 if ((err == 0) && (nx_upipe_mhints != 0)) {
323 nxp->nxp_mhints = nx_upipe_mhints;
324 }
325 #endif /* (DEVELOPMENT || DEBUG) */
326 return err;
327 }
328
329 static int
nx_upipe_prov_mem_new(struct kern_nexus_domain_provider * nxdom_prov,struct kern_nexus * nx,struct nexus_adapter * na)330 nx_upipe_prov_mem_new(struct kern_nexus_domain_provider *nxdom_prov,
331 struct kern_nexus *nx, struct nexus_adapter *na)
332 {
333 #pragma unused(nxdom_prov)
334 int err = 0;
335
336 SK_DF(SK_VERB_USER_PIPE,
337 "nx 0x%llx (\"%s\":\"%s\") na \"%s\" (0x%llx)", SK_KVA(nx),
338 NX_DOM(nx)->nxdom_name, nxdom_prov->nxdom_prov_name, na->na_name,
339 SK_KVA(na));
340
341 ASSERT(na->na_arena == NULL);
342 ASSERT(NX_USER_CHANNEL_PROV(nx));
343 /*
344 * The underlying nexus adapters already share the same memory
345 * allocator, and thus we don't care about storing the pp in
346 * the nexus.
347 *
348 * This means that clients calling kern_nexus_get_pbufpool()
349 * will get NULL, but this is fine since we don't expose the
350 * user pipe to external kernel clients.
351 */
352 na->na_arena = skmem_arena_create_for_nexus(na,
353 NX_PROV(nx)->nxprov_region_params, NULL, NULL, FALSE,
354 FALSE, NULL, &err);
355 ASSERT(na->na_arena != NULL || err != 0);
356
357 return err;
358 }
359
360 static void
nx_upipe_prov_fini(struct kern_nexus_domain_provider * nxdom_prov)361 nx_upipe_prov_fini(struct kern_nexus_domain_provider *nxdom_prov)
362 {
363 #pragma unused(nxdom_prov)
364 SK_D("destroying %s", nxdom_prov->nxdom_prov_name);
365 }
366
367 static int
nx_upipe_prov_nx_ctor(struct kern_nexus * nx)368 nx_upipe_prov_nx_ctor(struct kern_nexus *nx)
369 {
370 SK_LOCK_ASSERT_HELD();
371 ASSERT(nx->nx_arg == NULL);
372
373 SK_D("nexus 0x%llx (%s)", SK_KVA(nx), NX_DOM_PROV(nx)->nxdom_prov_name);
374
375 nx->nx_arg = nx_upipe_alloc(Z_WAITOK);
376 SK_D("create new upipe 0x%llx for nexus 0x%llx",
377 SK_KVA(NX_UPIPE_PRIVATE(nx)), SK_KVA(nx));
378
379 return 0;
380 }
381
382 static void
nx_upipe_prov_nx_dtor(struct kern_nexus * nx)383 nx_upipe_prov_nx_dtor(struct kern_nexus *nx)
384 {
385 struct nx_upipe *u = NX_UPIPE_PRIVATE(nx);
386
387 SK_LOCK_ASSERT_HELD();
388
389 SK_D("nexus 0x%llx (%s) upipe 0x%llx", SK_KVA(nx),
390 NX_DOM_PROV(nx)->nxdom_prov_name, SK_KVA(u));
391
392 if (u->nup_cli_nxb != NULL) {
393 nxb_free(u->nup_cli_nxb);
394 u->nup_cli_nxb = NULL;
395 }
396 if (u->nup_srv_nxb != NULL) {
397 nxb_free(u->nup_srv_nxb);
398 u->nup_srv_nxb = NULL;
399 }
400
401 SK_DF(SK_VERB_USER_PIPE, "marking upipe 0x%llx as free", SK_KVA(u));
402 nx_upipe_free(u);
403 nx->nx_arg = NULL;
404 }
405
406 static struct nexus_upipe_adapter *
na_upipe_alloc(zalloc_flags_t how)407 na_upipe_alloc(zalloc_flags_t how)
408 {
409 struct nexus_upipe_adapter *pna;
410
411 _CASSERT(offsetof(struct nexus_upipe_adapter, pna_up) == 0);
412
413 pna = zalloc_flags(na_upipe_zone, how | Z_ZERO);
414 if (pna) {
415 pna->pna_up.na_type = NA_USER_PIPE;
416 pna->pna_up.na_free = na_upipe_free;
417 }
418 return pna;
419 }
420
421 static void
na_upipe_free(struct nexus_adapter * na)422 na_upipe_free(struct nexus_adapter *na)
423 {
424 struct nexus_upipe_adapter *pna = (struct nexus_upipe_adapter *)na;
425
426 ASSERT(pna->pna_up.na_refcount == 0);
427 SK_DF(SK_VERB_MEM, "pna 0x%llx FREE", SK_KVA(pna));
428 bzero(pna, sizeof(*pna));
429 zfree(na_upipe_zone, pna);
430 }
431
432 static int
nx_upipe_dom_bind_port(struct kern_nexus * nx,nexus_port_t * nx_port,struct nxbind * nxb0,void * info)433 nx_upipe_dom_bind_port(struct kern_nexus *nx, nexus_port_t *nx_port,
434 struct nxbind *nxb0, void *info)
435 {
436 #pragma unused(info)
437 struct nx_upipe *u = NX_UPIPE_PRIVATE(nx);
438 struct nxbind *nxb = NULL;
439 int error = 0;
440
441 ASSERT(nx_port != NULL);
442 ASSERT(nxb0 != NULL);
443
444 switch (*nx_port) {
445 case NEXUS_PORT_USER_PIPE_CLIENT:
446 case NEXUS_PORT_USER_PIPE_SERVER:
447 if ((*nx_port == NEXUS_PORT_USER_PIPE_CLIENT &&
448 u->nup_cli_nxb != NULL) ||
449 (*nx_port == NEXUS_PORT_USER_PIPE_SERVER &&
450 u->nup_srv_nxb != NULL)) {
451 error = EEXIST;
452 break;
453 }
454
455 nxb = nxb_alloc(Z_WAITOK);
456 nxb_move(nxb0, nxb);
457 if (*nx_port == NEXUS_PORT_USER_PIPE_CLIENT) {
458 u->nup_cli_nxb = nxb;
459 } else {
460 u->nup_srv_nxb = nxb;
461 }
462
463 ASSERT(error == 0);
464 break;
465
466 default:
467 error = EDOM;
468 break;
469 }
470
471 return error;
472 }
473
474 static int
nx_upipe_dom_unbind_port(struct kern_nexus * nx,nexus_port_t nx_port)475 nx_upipe_dom_unbind_port(struct kern_nexus *nx, nexus_port_t nx_port)
476 {
477 struct nx_upipe *u = NX_UPIPE_PRIVATE(nx);
478 struct nxbind *nxb = NULL;
479 int error = 0;
480
481 ASSERT(nx_port != NEXUS_PORT_ANY);
482
483 switch (nx_port) {
484 case NEXUS_PORT_USER_PIPE_CLIENT:
485 case NEXUS_PORT_USER_PIPE_SERVER:
486 if ((nx_port == NEXUS_PORT_USER_PIPE_CLIENT &&
487 u->nup_cli_nxb == NULL) ||
488 (nx_port == NEXUS_PORT_USER_PIPE_SERVER &&
489 u->nup_srv_nxb == NULL)) {
490 error = ENOENT;
491 break;
492 }
493
494 if (nx_port == NEXUS_PORT_USER_PIPE_CLIENT) {
495 nxb = u->nup_cli_nxb;
496 u->nup_cli_nxb = NULL;
497 } else {
498 nxb = u->nup_srv_nxb;
499 u->nup_srv_nxb = NULL;
500 }
501 nxb_free(nxb);
502 ASSERT(error == 0);
503 break;
504
505 default:
506 error = EDOM;
507 break;
508 }
509
510 return error;
511 }
512
513 static int
nx_upipe_dom_connect(struct kern_nexus_domain_provider * nxdom_prov,struct kern_nexus * nx,struct kern_channel * ch,struct chreq * chr,struct kern_channel * ch0,struct nxbind * nxb,struct proc * p)514 nx_upipe_dom_connect(struct kern_nexus_domain_provider *nxdom_prov,
515 struct kern_nexus *nx, struct kern_channel *ch, struct chreq *chr,
516 struct kern_channel *ch0, struct nxbind *nxb, struct proc *p)
517 {
518 #pragma unused(nxdom_prov)
519 nexus_port_t port = chr->cr_port;
520 int err = 0;
521
522 SK_LOCK_ASSERT_HELD();
523
524 ASSERT(NX_DOM_PROV(nx) == nxdom_prov);
525 ASSERT(nx->nx_prov->nxprov_params->nxp_type ==
526 nxdom_prov->nxdom_prov_dom->nxdom_type &&
527 nx->nx_prov->nxprov_params->nxp_type == NEXUS_TYPE_USER_PIPE);
528
529 /*
530 * XXX: channel in user packet pool mode is not supported for
531 * user-pipe for now.
532 */
533 if (chr->cr_mode & CHMODE_USER_PACKET_POOL) {
534 SK_ERR("User packet pool mode not supported for upipe");
535 err = ENOTSUP;
536 goto done;
537 }
538
539 if (chr->cr_mode & CHMODE_EVENT_RING) {
540 SK_ERR("event ring is not supported for upipe");
541 err = ENOTSUP;
542 goto done;
543 }
544
545 if (chr->cr_mode & CHMODE_LOW_LATENCY) {
546 SK_ERR("low latency is not supported for upipe");
547 err = ENOTSUP;
548 goto done;
549 }
550
551 if (port == NEXUS_PORT_USER_PIPE_SERVER) {
552 chr->cr_real_endpoint = CH_ENDPOINT_USER_PIPE_MASTER;
553 } else if (port == NEXUS_PORT_USER_PIPE_CLIENT) {
554 chr->cr_real_endpoint = CH_ENDPOINT_USER_PIPE_SLAVE;
555 } else {
556 err = EINVAL;
557 goto done;
558 }
559
560 chr->cr_endpoint = chr->cr_real_endpoint;
561 chr->cr_ring_set = RING_SET_DEFAULT;
562 chr->cr_pipe_id = 0;
563 (void) snprintf(chr->cr_name, sizeof(chr->cr_name), "upipe:%llu:%.*s",
564 nx->nx_id, (int)nx->nx_prov->nxprov_params->nxp_namelen,
565 nx->nx_prov->nxprov_params->nxp_name);
566
567 err = na_connect(nx, ch, chr, ch0, nxb, p);
568 done:
569 return err;
570 }
571
572 static void
nx_upipe_dom_disconnect(struct kern_nexus_domain_provider * nxdom_prov,struct kern_nexus * nx,struct kern_channel * ch)573 nx_upipe_dom_disconnect(struct kern_nexus_domain_provider *nxdom_prov,
574 struct kern_nexus *nx, struct kern_channel *ch)
575 {
576 #pragma unused(nxdom_prov)
577 SK_LOCK_ASSERT_HELD();
578
579 SK_D("channel 0x%llx -!- nexus 0x%llx (%s:\"%s\":%u:%d)", SK_KVA(ch),
580 SK_KVA(nx), nxdom_prov->nxdom_prov_name, ch->ch_na->na_name,
581 ch->ch_info->cinfo_nx_port, (int)ch->ch_info->cinfo_ch_ring_id);
582
583 na_disconnect(nx, ch);
584 /*
585 * Set NXF_REJECT on the nexus which would cause any channel on the
586 * peer adapter to cease to function.
587 */
588 if (NX_PROV(nx)->nxprov_params->nxp_reject_on_close) {
589 os_atomic_or(&nx->nx_flags, NXF_REJECT, relaxed);
590 }
591 }
592
593 static void
nx_upipe_dom_defunct(struct kern_nexus_domain_provider * nxdom_prov,struct kern_nexus * nx,struct kern_channel * ch,struct proc * p)594 nx_upipe_dom_defunct(struct kern_nexus_domain_provider *nxdom_prov,
595 struct kern_nexus *nx, struct kern_channel *ch, struct proc *p)
596 {
597 #pragma unused(nxdom_prov, nx)
598 struct nexus_adapter *na = ch->ch_na;
599 struct nexus_upipe_adapter *pna = (struct nexus_upipe_adapter *)na;
600 ring_id_t qfirst = ch->ch_first[NR_TX];
601 ring_id_t qlast = ch->ch_last[NR_TX];
602 uint32_t i;
603
604 LCK_MTX_ASSERT(&ch->ch_lock, LCK_MTX_ASSERT_OWNED);
605 ASSERT(!(ch->ch_flags & CHANF_KERNEL));
606 ASSERT(na->na_type == NA_USER_PIPE);
607
608 /*
609 * Inform the peer receiver thread in nx_upipe_na_rxsync() or the
610 * peer transmit thread in nx_upipe_na_txsync() about
611 * this endpoint going defunct. We utilize the TX ring's
612 * lock for serialization, since that is what's being used
613 * by the receiving endpoint.
614 */
615 for (i = qfirst; i < qlast; i++) {
616 /*
617 * For maintaining lock ordering between the two channels of
618 * user pipe.
619 */
620 if (pna->pna_role == CH_ENDPOINT_USER_PIPE_MASTER) {
621 (void) kr_enter(&NAKR(na, NR_TX)[i], TRUE);
622 (void) kr_enter(NAKR(na, NR_RX)[i].ckr_pipe, TRUE);
623 } else {
624 (void) kr_enter(NAKR(na, NR_RX)[i].ckr_pipe, TRUE);
625 (void) kr_enter(&NAKR(na, NR_TX)[i], TRUE);
626 }
627 }
628
629 na_ch_rings_defunct(ch, p);
630
631 for (i = qfirst; i < qlast; i++) {
632 if (pna->pna_role == CH_ENDPOINT_USER_PIPE_MASTER) {
633 (void) kr_exit(NAKR(na, NR_RX)[i].ckr_pipe);
634 (void) kr_exit(&NAKR(na, NR_TX)[i]);
635 } else {
636 (void) kr_exit(&NAKR(na, NR_TX)[i]);
637 (void) kr_exit(NAKR(na, NR_RX)[i].ckr_pipe);
638 }
639 }
640 }
641
642 static void
nx_upipe_dom_defunct_finalize(struct kern_nexus_domain_provider * nxdom_prov,struct kern_nexus * nx,struct kern_channel * ch,boolean_t locked)643 nx_upipe_dom_defunct_finalize(struct kern_nexus_domain_provider *nxdom_prov,
644 struct kern_nexus *nx, struct kern_channel *ch, boolean_t locked)
645 {
646 #pragma unused(nxdom_prov)
647 struct nexus_upipe_adapter *pna =
648 (struct nexus_upipe_adapter *)ch->ch_na;
649
650 if (!locked) {
651 SK_LOCK_ASSERT_NOTHELD();
652 SK_LOCK();
653 LCK_MTX_ASSERT(&ch->ch_lock, LCK_MTX_ASSERT_NOTOWNED);
654 } else {
655 SK_LOCK_ASSERT_HELD();
656 LCK_MTX_ASSERT(&ch->ch_lock, LCK_MTX_ASSERT_OWNED);
657 }
658
659 ASSERT(!(ch->ch_flags & CHANF_KERNEL));
660 ASSERT(ch->ch_na->na_type == NA_USER_PIPE);
661
662 /*
663 * At this point, we know that the arena shared by the master and
664 * slave adapters has no more valid mappings on the channels opened
665 * to them. We need to invoke na_defunct() on both adapters to
666 * release any remaining slots attached to their rings.
667 *
668 * Note that the 'ch' that we pass in here is irrelevant as we
669 * don't support user packet pool for user pipe.
670 */
671 na_defunct(nx, ch, &pna->pna_up, locked);
672 if (pna->pna_peer != NULL) {
673 na_defunct(nx, ch, &pna->pna_peer->pna_up, locked);
674 }
675
676 /*
677 * And if their parent adapter (the memory owner) is a pseudo
678 * nexus adapter that we initially created in nx_upipe_na_find(),
679 * invoke na_defunct() on it now to do the final teardown on
680 * the arena.
681 */
682 if (pna->pna_parent->na_type == NA_PSEUDO) {
683 na_defunct(nx, ch, pna->pna_parent, locked);
684 }
685
686 SK_D("%s(%d): ch 0x%llx -/- nx 0x%llx (%s:\"%s\":%u:%d)",
687 ch->ch_name, ch->ch_pid, SK_KVA(ch), SK_KVA(nx),
688 nxdom_prov->nxdom_prov_name, ch->ch_na->na_name,
689 ch->ch_info->cinfo_nx_port, (int)ch->ch_info->cinfo_ch_ring_id);
690
691 if (!locked) {
692 LCK_MTX_ASSERT(&ch->ch_lock, LCK_MTX_ASSERT_NOTOWNED);
693 SK_UNLOCK();
694 } else {
695 LCK_MTX_ASSERT(&ch->ch_lock, LCK_MTX_ASSERT_OWNED);
696 SK_LOCK_ASSERT_HELD();
697 }
698 }
699
700 /* allocate the pipe array in the parent adapter */
701 static int
nx_upipe_na_alloc(struct nexus_adapter * na,uint32_t npipes)702 nx_upipe_na_alloc(struct nexus_adapter *na, uint32_t npipes)
703 {
704 struct nexus_upipe_adapter **npa;
705
706 if (npipes <= na->na_max_pipes) {
707 /* we already have more entries that requested */
708 return 0;
709 }
710 if (npipes < na->na_next_pipe || npipes > NX_UPIPE_MAXPIPES) {
711 return EINVAL;
712 }
713
714 npa = sk_realloc_type_array(struct nexus_upipe_adapter *,
715 na->na_max_pipes, npipes, na->na_pipes, Z_WAITOK, skmem_tag_pipes);
716 if (npa == NULL) {
717 return ENOMEM;
718 }
719
720 na->na_pipes = npa;
721 na->na_max_pipes = npipes;
722
723 return 0;
724 }
725
726 /* deallocate the parent array in the parent adapter */
727 void
nx_upipe_na_dealloc(struct nexus_adapter * na)728 nx_upipe_na_dealloc(struct nexus_adapter *na)
729 {
730 if (na->na_pipes) {
731 if (na->na_next_pipe > 0) {
732 SK_ERR("freeing not empty pipe array for %s "
733 "(%u dangling pipes)!", na->na_name,
734 na->na_next_pipe);
735 }
736 sk_free_type_array(struct nexus_upipe_adapter *,
737 na->na_max_pipes, na->na_pipes);
738 na->na_pipes = NULL;
739 na->na_max_pipes = 0;
740 na->na_next_pipe = 0;
741 }
742 }
743
744 /* find a pipe endpoint with the given id among the parent's pipes */
745 static struct nexus_upipe_adapter *
nx_upipe_find(struct nexus_adapter * parent,uint32_t pipe_id)746 nx_upipe_find(struct nexus_adapter *parent, uint32_t pipe_id)
747 {
748 uint32_t i;
749 struct nexus_upipe_adapter *na;
750
751 for (i = 0; i < parent->na_next_pipe; i++) {
752 na = parent->na_pipes[i];
753 if (na->pna_id == pipe_id) {
754 return na;
755 }
756 }
757 return NULL;
758 }
759
760 /* add a new pipe endpoint to the parent array */
761 static int
nx_upipe_na_add(struct nexus_adapter * parent,struct nexus_upipe_adapter * na)762 nx_upipe_na_add(struct nexus_adapter *parent, struct nexus_upipe_adapter *na)
763 {
764 if (parent->na_next_pipe >= parent->na_max_pipes) {
765 uint32_t npipes = parent->na_max_pipes ?
766 2 * parent->na_max_pipes : 2;
767 int error = nx_upipe_na_alloc(parent, npipes);
768 if (error) {
769 return error;
770 }
771 }
772
773 parent->na_pipes[parent->na_next_pipe] = na;
774 na->pna_parent_slot = parent->na_next_pipe;
775 parent->na_next_pipe++;
776 return 0;
777 }
778
779 /* remove the given pipe endpoint from the parent array */
780 static void
nx_upipe_na_remove(struct nexus_adapter * parent,struct nexus_upipe_adapter * na)781 nx_upipe_na_remove(struct nexus_adapter *parent, struct nexus_upipe_adapter *na)
782 {
783 uint32_t n;
784 n = --parent->na_next_pipe;
785 if (n != na->pna_parent_slot) {
786 struct nexus_upipe_adapter **p =
787 &parent->na_pipes[na->pna_parent_slot];
788 *p = parent->na_pipes[n];
789 (*p)->pna_parent_slot = na->pna_parent_slot;
790 }
791 parent->na_pipes[n] = NULL;
792 }
793
794 static int
nx_upipe_na_txsync(struct __kern_channel_ring * txkring,struct proc * p,uint32_t flags)795 nx_upipe_na_txsync(struct __kern_channel_ring *txkring, struct proc *p,
796 uint32_t flags)
797 {
798 struct __kern_channel_ring *rxkring = txkring->ckr_pipe;
799 volatile uint64_t *tx_tsync, *tx_tnote, *rx_tsync;
800 int sent = 0, ret = 0;
801
802 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_TX,
803 "%s(%d) kr \"%s\" (0x%llx) krflags 0x%b ring %u "
804 "flags 0x%x -> kr \"%s\" (0x%llx) krflags 0x%b ring %u",
805 sk_proc_name_address(p), sk_proc_pid(p), txkring->ckr_name,
806 SK_KVA(txkring), txkring->ckr_flags, CKRF_BITS,
807 txkring->ckr_ring_id, flags, rxkring->ckr_name, SK_KVA(rxkring),
808 rxkring->ckr_flags, CKRF_BITS, rxkring->ckr_ring_id);
809
810 /*
811 * Serialize write access to the transmit ring, since another
812 * thread coming down for rxsync might pick up pending slots.
813 */
814 ASSERT(txkring->ckr_owner == current_thread());
815
816 /*
817 * Record the time of sync and grab sync time of other side;
818 * use atomic store and load since we're not holding the
819 * lock used by the receive ring. This allows us to avoid
820 * the potentially costly os_atomic_thread_fence(seq_cst).
821 */
822 /* deconst */
823 tx_tsync = __DECONST(uint64_t *, &txkring->ckr_ring->ring_sync_time);
824 os_atomic_store(tx_tsync, txkring->ckr_sync_time, release);
825
826 /*
827 * Read from the peer's kring, not its user ring; the peer's channel
828 * may be defunct, in which case it's unsafe to access its user ring.
829 */
830 rx_tsync = __DECONST(uint64_t *, &rxkring->ckr_sync_time);
831 tx_tnote = __DECONST(uint64_t *, &txkring->ckr_ring->ring_notify_time);
832 *tx_tnote = os_atomic_add_orig(rx_tsync, 0, relaxed);
833
834 if (__probable(txkring->ckr_rhead != txkring->ckr_khead)) {
835 sent = nx_upipe_na_txsync_locked(txkring, p, flags,
836 &ret, FALSE);
837 }
838
839 if (sent != 0) {
840 (void) rxkring->ckr_na_notify(rxkring, p, 0);
841 }
842
843 return ret;
844 }
845
846 int
nx_upipe_na_txsync_locked(struct __kern_channel_ring * txkring,struct proc * p,uint32_t flags,int * ret,boolean_t rx)847 nx_upipe_na_txsync_locked(struct __kern_channel_ring *txkring, struct proc *p,
848 uint32_t flags, int *ret, boolean_t rx)
849 {
850 #pragma unused(p, flags, rx)
851 struct __kern_channel_ring *rxkring = txkring->ckr_pipe;
852 const slot_idx_t lim_tx = txkring->ckr_lim;
853 const slot_idx_t lim_rx = rxkring->ckr_lim;
854 slot_idx_t j, k;
855 int n, m, b, sent = 0;
856 uint32_t byte_count = 0;
857 int limit; /* max # of slots to transfer */
858
859 *ret = 0;
860
861 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_TX,
862 "%s(%d) kr \"%s\", kh %3u kt %3u | "
863 "rh %3u rt %3u [pre%s]", sk_proc_name_address(p),
864 sk_proc_pid(p), txkring->ckr_name, txkring->ckr_khead,
865 txkring->ckr_ktail, txkring->ckr_rhead,
866 txkring->ckr_rtail, rx ? "*" : "");
867 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_TX,
868 "%s(%d) kr \"%s\", kh %3u kt %3u | "
869 "rh %3u rt %3u [pre%s]", sk_proc_name_address(p),
870 sk_proc_pid(p), rxkring->ckr_name, rxkring->ckr_khead,
871 rxkring->ckr_ktail, rxkring->ckr_rhead,
872 rxkring->ckr_rtail, rx ? "*" : "");
873
874 if (__improbable(KR_DROP(txkring) || KR_DROP(rxkring))) {
875 *ret = ENXIO;
876 goto done;
877 }
878
879 j = rxkring->ckr_ktail; /* RX */
880 k = txkring->ckr_khead; /* TX */
881
882 /* # of new tx slots */
883 n = txkring->ckr_rhead - txkring->ckr_khead;
884 if (n < 0) {
885 n += txkring->ckr_num_slots;
886 }
887 limit = n;
888
889 /* # of rx busy (unclaimed) slots */
890 b = j - rxkring->ckr_khead;
891 if (b < 0) {
892 b += rxkring->ckr_num_slots;
893 }
894
895 /* # of rx avail free slots (subtract busy from max) */
896 m = lim_rx - b;
897 if (m < limit) {
898 limit = m;
899 }
900
901 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_TX,
902 "%s(%d) kr \"%s\" -> new %u, kr \"%s\" "
903 "-> free %u", sk_proc_name_address(p), sk_proc_pid(p),
904 txkring->ckr_name, n, rxkring->ckr_name, m);
905
906 /* rxring is full, or nothing to send? */
907 if (__improbable((sent = limit) == 0)) {
908 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_TX,
909 "%s(%d) kr \"%s\" -> %s%s",
910 sk_proc_name_address(p), sk_proc_pid(p), (n > m) ?
911 rxkring->ckr_name : txkring->ckr_name, ((n > m) ?
912 "no room avail" : "no new slots"),
913 (rx ? " (lost race, ok)" : ""));
914 goto done;
915 }
916
917 ASSERT(limit > 0);
918 while (limit--) {
919 struct __kern_slot_desc *ksd_tx = KR_KSD(txkring, k);
920 struct __user_slot_desc *usd_tx = KR_USD(txkring, k);
921 struct __kern_slot_desc *ksd_rx = KR_KSD(rxkring, j);
922 struct __user_slot_desc *usd_rx = KR_USD(rxkring, j);
923 struct __kern_quantum *kqum;
924
925 kqum = ksd_tx->sd_qum;
926 /*
927 * Packets failing internalization should be dropped in
928 * TX sync prologue.
929 */
930 ASSERT((kqum->qum_qflags & (QUM_F_INTERNALIZED |
931 QUM_F_FINALIZED)) == (QUM_F_INTERNALIZED |
932 QUM_F_FINALIZED));
933
934 byte_count += kqum->qum_len;
935
936 /*
937 * Swap the slots.
938 *
939 * XXX: [email protected] -- this bypasses the slot attach/detach
940 * interface, and needs to be changed when upipe adopts the
941 * packet APIs. SD_SWAP() will perform a block copy of the
942 * swap, and will readjust the kernel slot descriptor's sd_user
943 * accordingly.
944 */
945 SD_SWAP(ksd_rx, usd_rx, ksd_tx, usd_tx);
946
947 j = SLOT_NEXT(j, lim_rx);
948 k = SLOT_NEXT(k, lim_tx);
949 }
950
951 kr_update_stats(rxkring, sent, byte_count);
952 if (__improbable(kr_stat_enable != 0)) {
953 txkring->ckr_stats = rxkring->ckr_stats;
954 }
955
956 /*
957 * Make sure the slots are updated before ckr_ktail reach global
958 * visibility, since we are not holding rx ring's kr_enter().
959 */
960 os_atomic_thread_fence(seq_cst);
961
962 rxkring->ckr_ktail = j;
963 txkring->ckr_khead = k;
964 txkring->ckr_ktail = SLOT_PREV(k, lim_tx);
965
966 done:
967 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_TX,
968 "%s(%d) kr \"%s\", kh %3u kt %3u | "
969 "rh %3u rt %3u [post%s]", sk_proc_name_address(p),
970 sk_proc_pid(p), txkring->ckr_name, txkring->ckr_khead,
971 txkring->ckr_ktail, txkring->ckr_rhead,
972 txkring->ckr_rtail, rx ? "*" : "");
973 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_TX,
974 "%s(%d) kr \"%s\", kh %3u kt %3u | "
975 "rh %3u rt %3u [post%s]", sk_proc_name_address(p),
976 sk_proc_pid(p), rxkring->ckr_name, rxkring->ckr_khead,
977 rxkring->ckr_ktail, rxkring->ckr_rhead,
978 rxkring->ckr_rtail, rx ? "*" : "");
979
980 return sent;
981 }
982
983 static int
nx_upipe_na_rxsync(struct __kern_channel_ring * rxkring,struct proc * p,uint32_t flags)984 nx_upipe_na_rxsync(struct __kern_channel_ring *rxkring, struct proc *p,
985 uint32_t flags)
986 {
987 #pragma unused(p)
988 struct __kern_channel_ring *txkring = rxkring->ckr_pipe;
989 volatile uint64_t *rx_tsync, *rx_tnote, *tx_tsync;
990 const slot_idx_t lim_rx = rxkring->ckr_lim;
991 int n; /* new slots from transmit side */
992 int m, b, ret = 0;
993 uint32_t r;
994
995 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_RX,
996 "%s(%d) kr \"%s\" (0x%llx) krflags 0x%b ring %u "
997 "flags 0x%x <- kr \"%s\" (0x%llx) krflags 0x%b ring %u",
998 sk_proc_name_address(p), sk_proc_pid(p), rxkring->ckr_name,
999 SK_KVA(rxkring), rxkring->ckr_flags, CKRF_BITS,
1000 rxkring->ckr_ring_id, flags, txkring->ckr_name, SK_KVA(txkring),
1001 txkring->ckr_flags, CKRF_BITS, txkring->ckr_ring_id);
1002
1003 ASSERT(rxkring->ckr_owner == current_thread());
1004
1005 /* reclaim and get # of rx reclaimed slots */
1006 r = kr_reclaim(rxkring);
1007
1008 /* # of rx busy (unclaimed) slots */
1009 b = rxkring->ckr_ktail - rxkring->ckr_khead;
1010 if (b < 0) {
1011 b += rxkring->ckr_num_slots;
1012 }
1013
1014 /* # of rx avail free slots (subtract busy from max) */
1015 m = lim_rx - b;
1016
1017 /*
1018 * Check if there's any new slots on transmit ring; do this
1019 * first without acquiring that ring's ckr_qlock, and use
1020 * the memory barrier (paired with second one in txsync.)
1021 * If we missed the race we'd just pay the cost of acquiring
1022 * ckr_qlock and potentially returning from "internal txsync"
1023 * without anything to process, which is okay.
1024 */
1025 os_atomic_thread_fence(seq_cst);
1026 n = txkring->ckr_rhead - txkring->ckr_khead;
1027 if (n < 0) {
1028 n += txkring->ckr_num_slots;
1029 }
1030
1031 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_RX,
1032 "%s(%d) kr \"%s\" <- free %u, kr \"%s\" <- new %u",
1033 sk_proc_name_address(p), sk_proc_pid(p),
1034 rxkring->ckr_name, m, txkring->ckr_name, n);
1035
1036 /*
1037 * Record the time of sync and grab sync time of other side;
1038 * use atomic store and load since we're not holding the
1039 * lock used by the receive ring. This allows us to avoid
1040 * the potentially costly os_atomic_thread_fence(seq_cst).
1041 */
1042 /* deconst */
1043 rx_tsync = __DECONST(uint64_t *, &rxkring->ckr_ring->ring_sync_time);
1044 os_atomic_store(rx_tsync, rxkring->ckr_sync_time, release);
1045
1046 /*
1047 * Read from the peer's kring, not its user ring; the peer's channel
1048 * may be defunct, in which case it's unsafe to access its user ring.
1049 */
1050 tx_tsync = __DECONST(uint64_t *, &txkring->ckr_sync_time);
1051 rx_tnote = __DECONST(uint64_t *, &rxkring->ckr_ring->ring_notify_time);
1052 *rx_tnote = os_atomic_add_orig(tx_tsync, 0, relaxed);
1053
1054 /*
1055 * If we have slots to pick up from the transmit side and and we
1056 * have space available, perform an equivalent of "internal txsync".
1057 *
1058 * Acquire write access to the transmit (peer) ring,
1059 * Serialize write access to it, since another thread
1060 * coming down for txsync might add new slots.
1061 * If we fail to get the kring lock, then don't worry because
1062 * there's already a transmit sync in progress to move packets.
1063 */
1064 if (__probable(n != 0 && m != 0 && (flags & NA_SYNCF_MONITOR) == 0)) {
1065 (void) kr_enter(txkring, TRUE);
1066 n = nx_upipe_na_txsync_locked(txkring, p, flags, &ret, TRUE);
1067 kr_exit(txkring);
1068 } else {
1069 n = 0;
1070 }
1071
1072 /*
1073 * If we have reclaimed some slots or transferred new slots
1074 * from the transmit side, notify the other end. Also notify
1075 * ourselves to pick up newly transferred ones, if any.
1076 */
1077 if (__probable(r != 0 || n != 0)) {
1078 SK_DF(SK_VERB_USER_PIPE | SK_VERB_SYNC | SK_VERB_RX,
1079 "%s(%d) kr \"%s\", kh %3u kt %3u | "
1080 "rh %3u rt %3u [rel %u new %u]",
1081 sk_proc_name_address(p), sk_proc_pid(p), rxkring->ckr_name,
1082 rxkring->ckr_khead, rxkring->ckr_ktail,
1083 rxkring->ckr_rhead, rxkring->ckr_rtail, r, n);
1084
1085 (void) txkring->ckr_na_notify(txkring, p, 0);
1086 }
1087
1088 return ret;
1089 }
1090
1091 static int
nx_upipe_na_rings_create(struct nexus_adapter * na,struct kern_channel * ch)1092 nx_upipe_na_rings_create(struct nexus_adapter *na, struct kern_channel *ch)
1093 {
1094 struct nexus_upipe_adapter *pna = (struct nexus_upipe_adapter *)na;
1095 struct nexus_adapter *ona = &pna->pna_peer->pna_up;
1096 int error = 0;
1097 enum txrx t;
1098 uint32_t i;
1099
1100 /*
1101 * Create krings and all the rings for this end;
1102 * we'll update ckr_save_ring pointers below.
1103 */
1104 error = na_rings_mem_setup(na, FALSE, ch);
1105 if (error != 0) {
1106 goto err;
1107 }
1108
1109 /* update our hidden ring pointers */
1110 for_rx_tx(t) {
1111 for (i = 0; i < na_get_nrings(na, t); i++) {
1112 NAKR(na, t)[i].ckr_save_ring =
1113 NAKR(na, t)[i].ckr_ring;
1114 }
1115 }
1116
1117 /* now, create krings and rings of the other end */
1118 error = na_rings_mem_setup(ona, FALSE, ch);
1119 if (error != 0) {
1120 na_rings_mem_teardown(na, ch, FALSE); /* this end */
1121 goto err;
1122 }
1123
1124 for_rx_tx(t) {
1125 for (i = 0; i < na_get_nrings(ona, t); i++) {
1126 NAKR(ona, t)[i].ckr_save_ring =
1127 NAKR(ona, t)[i].ckr_ring;
1128 }
1129 }
1130
1131 /* cross link the krings */
1132 for_rx_tx(t) {
1133 /* swap NR_TX <-> NR_RX (skip host ring) */
1134 enum txrx r = sk_txrx_swap(t);
1135 for (i = 0; i < na_get_nrings(na, t); i++) {
1136 NAKR(na, t)[i].ckr_pipe =
1137 NAKR(&pna->pna_peer->pna_up, r) + i;
1138 NAKR(&pna->pna_peer->pna_up, r)[i].ckr_pipe =
1139 NAKR(na, t) + i;
1140 }
1141 }
1142 err:
1143 return error;
1144 }
1145
1146 /*
1147 * Pipe endpoints are created and destroyed together, so that endopoints do not
1148 * have to check for the existence of their peer at each ?xsync.
1149 *
1150 * To play well with the existing nexus adapter infrastructure (refcounts etc.),
1151 * we adopt the following strategy:
1152 *
1153 * 1) The first endpoint that is created also creates the other endpoint and
1154 * grabs a reference to it.
1155 *
1156 * state A) user1 --> endpoint1 --> endpoint2
1157 *
1158 * 2) If, starting from state A, endpoint2 is then registered, endpoint1 gives
1159 * its reference to the user:
1160 *
1161 * state B) user1 --> endpoint1 endpoint2 <--- user2
1162 *
1163 * 3) Assume that, starting from state B endpoint2 is closed. In the unregister
1164 * callback endpoint2 notes that endpoint1 is still active and adds a reference
1165 * from endpoint1 to itself. When user2 then releases her own reference,
1166 * endpoint2 is not destroyed and we are back to state A. A symmetrical state
1167 * would be reached if endpoint1 were released instead.
1168 *
1169 * 4) If, starting from state A, endpoint1 is closed, the destructor notes that
1170 * it owns a reference to endpoint2 and releases it.
1171 *
1172 * Something similar goes on for the creation and destruction of the krings.
1173 */
1174
1175
1176 /*
1177 * nx_upipe_na_krings_create.
1178 *
1179 * There are two cases:
1180 *
1181 * 1) state is
1182 *
1183 * usr1 --> e1 --> e2
1184 *
1185 * and we are e1. We have to create both sets
1186 * of krings.
1187 *
1188 * 2) state is
1189 *
1190 * usr1 --> e1 --> e2
1191 *
1192 * and we are e2. e1 is certainly registered and our
1193 * krings already exist, but they may be hidden.
1194 */
1195 static int
nx_upipe_na_krings_create(struct nexus_adapter * na,struct kern_channel * ch)1196 nx_upipe_na_krings_create(struct nexus_adapter *na, struct kern_channel *ch)
1197 {
1198 struct nexus_upipe_adapter *pna = (struct nexus_upipe_adapter *)na;
1199 int error = 0;
1200 enum txrx t;
1201 uint32_t i;
1202
1203 /*
1204 * Verify symmetrical ring counts; validated
1205 * at nexus provider registration time.
1206 */
1207 ASSERT(na_get_nrings(na, NR_TX) == na_get_nrings(na, NR_RX));
1208
1209 if (pna->pna_peer_ref) {
1210 /* case 1) above */
1211 SK_DF(SK_VERB_USER_PIPE,
1212 "0x%llx: case 1, create everything", SK_KVA(na));
1213 error = nx_upipe_na_rings_create(na, ch);
1214 } else {
1215 /* case 2) above */
1216 /* recover the hidden rings */
1217 SK_DF(SK_VERB_USER_PIPE,
1218 "0x%llx: case 2, hidden rings", SK_KVA(na));
1219 for_rx_tx(t) {
1220 for (i = 0; i < na_get_nrings(na, t); i++) {
1221 NAKR(na, t)[i].ckr_ring =
1222 NAKR(na, t)[i].ckr_save_ring;
1223 }
1224 }
1225 }
1226
1227 ASSERT(error == 0 || (na->na_tx_rings == NULL &&
1228 na->na_rx_rings == NULL && na->na_slot_ctxs == NULL));
1229 ASSERT(error == 0 || (pna->pna_peer->pna_up.na_tx_rings == NULL &&
1230 pna->pna_peer->pna_up.na_rx_rings == NULL &&
1231 pna->pna_peer->pna_up.na_slot_ctxs == NULL));
1232
1233 return error;
1234 }
1235
1236 /*
1237 * nx_upipe_na_activate.
1238 *
1239 * There are two cases on registration (onoff==1)
1240 *
1241 * 1.a) state is
1242 *
1243 * usr1 --> e1 --> e2
1244 *
1245 * and we are e1. Nothing special to do.
1246 *
1247 * 1.b) state is
1248 *
1249 * usr1 --> e1 --> e2 <-- usr2
1250 *
1251 * and we are e2. Drop the ref e1 is holding.
1252 *
1253 * There are two additional cases on unregister (onoff==0)
1254 *
1255 * 2.a) state is
1256 *
1257 * usr1 --> e1 --> e2
1258 *
1259 * and we are e1. Nothing special to do, e2 will
1260 * be cleaned up by the destructor of e1.
1261 *
1262 * 2.b) state is
1263 *
1264 * usr1 --> e1 e2 <-- usr2
1265 *
1266 * and we are either e1 or e2. Add a ref from the
1267 * other end and hide our rings.
1268 */
1269 static int
nx_upipe_na_activate(struct nexus_adapter * na,na_activate_mode_t mode)1270 nx_upipe_na_activate(struct nexus_adapter *na, na_activate_mode_t mode)
1271 {
1272 struct nexus_upipe_adapter *pna = (struct nexus_upipe_adapter *)na;
1273
1274 SK_LOCK_ASSERT_HELD();
1275
1276 SK_DF(SK_VERB_USER_PIPE, "na \"%s\" (0x%llx) %s", na->na_name,
1277 SK_KVA(na), na_activate_mode2str(mode));
1278
1279 switch (mode) {
1280 case NA_ACTIVATE_MODE_ON:
1281 os_atomic_or(&na->na_flags, NAF_ACTIVE, relaxed);
1282 break;
1283
1284 case NA_ACTIVATE_MODE_DEFUNCT:
1285 break;
1286
1287 case NA_ACTIVATE_MODE_OFF:
1288 os_atomic_andnot(&na->na_flags, NAF_ACTIVE, relaxed);
1289 break;
1290
1291 default:
1292 VERIFY(0);
1293 /* NOTREACHED */
1294 __builtin_unreachable();
1295 }
1296
1297 if (pna->pna_peer_ref) {
1298 SK_DF(SK_VERB_USER_PIPE,
1299 "0x%llx: case 1.a or 2.a, nothing to do", SK_KVA(na));
1300 return 0;
1301 }
1302
1303 switch (mode) {
1304 case NA_ACTIVATE_MODE_ON:
1305 SK_DF(SK_VERB_USER_PIPE,
1306 "0x%llx: case 1.b, drop peer", SK_KVA(na));
1307 if (pna->pna_peer->pna_peer_ref) {
1308 pna->pna_peer->pna_peer_ref = FALSE;
1309 (void) na_release_locked(na);
1310 }
1311 break;
1312
1313 case NA_ACTIVATE_MODE_OFF:
1314 SK_DF(SK_VERB_USER_PIPE,
1315 "0x%llx: case 2.b, grab peer", SK_KVA(na));
1316 if (!pna->pna_peer->pna_peer_ref) {
1317 na_retain_locked(na);
1318 pna->pna_peer->pna_peer_ref = TRUE;
1319 }
1320 break;
1321
1322 default:
1323 break;
1324 }
1325
1326 return 0;
1327 }
1328
1329 /*
1330 * nx_upipe_na_krings_delete.
1331 *
1332 * There are two cases:
1333 *
1334 * 1) state is
1335 *
1336 * usr1 --> e1 --> e2
1337 *
1338 * and we are e1 (e2 is not bound, so krings_delete cannot be
1339 * called on it);
1340 *
1341 * 2) state is
1342 *
1343 * usr1 --> e1 e2 <-- usr2
1344 *
1345 * and we are either e1 or e2.
1346 *
1347 * In the former case we have to also delete the krings of e2;
1348 * in the latter case we do nothing (note that our krings
1349 * have already been hidden in the unregister callback).
1350 */
1351 static void
nx_upipe_na_krings_delete(struct nexus_adapter * na,struct kern_channel * ch,boolean_t defunct)1352 nx_upipe_na_krings_delete(struct nexus_adapter *na, struct kern_channel *ch,
1353 boolean_t defunct)
1354 {
1355 struct nexus_upipe_adapter *pna = (struct nexus_upipe_adapter *)na;
1356 struct nexus_adapter *ona; /* na of the other end */
1357 uint32_t i;
1358 enum txrx t;
1359
1360 SK_LOCK_ASSERT_HELD();
1361
1362 if (!pna->pna_peer_ref) {
1363 SK_DF(SK_VERB_USER_PIPE,
1364 "0x%llx: case 2, kept alive by peer", SK_KVA(na));
1365 /*
1366 * If adapter is defunct (note the explicit test against
1367 * NAF_DEFUNCT, and not the "defunct" parameter passed in
1368 * by the caller), then the peer's channel has gone defunct.
1369 * We get here because this channel was not defuncted, and
1370 * that this is the last active reference to the adapter.
1371 * At this point we tear everything down, since the caller
1372 * will proceed to destroying the memory regions.
1373 */
1374 if (na->na_flags & NAF_DEFUNCT) {
1375 na_rings_mem_teardown(na, ch, defunct);
1376 }
1377 return;
1378 }
1379
1380 /* case 1) above */
1381 SK_DF(SK_VERB_USER_PIPE,
1382 "0x%llx: case 1, deleting everyhing", SK_KVA(na));
1383
1384 ASSERT(na->na_channels == 0 || (na->na_flags & NAF_DEFUNCT));
1385
1386 /* restore the ring to be deleted on the peer */
1387 ona = &pna->pna_peer->pna_up;
1388 if (ona->na_tx_rings == NULL) {
1389 /*
1390 * Already deleted, we must be on an
1391 * cleanup-after-error path
1392 * Just delete this end
1393 */
1394 na_rings_mem_teardown(na, ch, defunct);
1395 return;
1396 }
1397
1398 /* delete the memory rings */
1399 na_rings_mem_teardown(na, ch, defunct);
1400
1401 if (!defunct) {
1402 for_rx_tx(t) {
1403 for (i = 0; i < na_get_nrings(ona, t); i++) {
1404 NAKR(ona, t)[i].ckr_ring =
1405 NAKR(ona, t)[i].ckr_save_ring;
1406 }
1407 }
1408 }
1409
1410 /* Delete the memory rings */
1411 na_rings_mem_teardown(ona, ch, defunct);
1412 }
1413
1414 static void
nx_upipe_na_dtor(struct nexus_adapter * na)1415 nx_upipe_na_dtor(struct nexus_adapter *na)
1416 {
1417 struct nexus_upipe_adapter *pna = (struct nexus_upipe_adapter *)na;
1418 struct nx_upipe *u = NX_UPIPE_PRIVATE(na->na_nx);
1419
1420 SK_LOCK_ASSERT_HELD();
1421
1422 SK_DF(SK_VERB_USER_PIPE, "0x%llx", SK_KVA(na));
1423 if (pna->pna_peer_ref) {
1424 SK_DF(SK_VERB_USER_PIPE,
1425 "0x%llx: clean up peer 0x%llx", SK_KVA(na),
1426 SK_KVA(&pna->pna_peer->pna_up));
1427 pna->pna_peer_ref = FALSE;
1428 (void) na_release_locked(&pna->pna_peer->pna_up);
1429 }
1430 if (pna->pna_role == CH_ENDPOINT_USER_PIPE_MASTER) {
1431 nx_upipe_na_remove(pna->pna_parent, pna);
1432 }
1433 (void) na_release_locked(pna->pna_parent);
1434 pna->pna_parent = NULL;
1435
1436 /* release reference to parent adapter held by nx_upipe_na_find() */
1437 ASSERT(u->nup_pna_users != 0);
1438 if (--u->nup_pna_users == 0) {
1439 ASSERT(u->nup_pna != NULL);
1440 SK_DF(SK_VERB_USER_PIPE, "release parent: \"%s\" (0x%llx)",
1441 u->nup_pna->na_name, SK_KVA(u->nup_pna));
1442 na_release_locked(u->nup_pna);
1443 u->nup_pna = NULL;
1444 }
1445 }
1446
1447 int
nx_upipe_na_find(struct kern_nexus * nx,struct kern_channel * ch,struct chreq * chr,struct nxbind * nxb,struct proc * p,struct nexus_adapter ** na,boolean_t create)1448 nx_upipe_na_find(struct kern_nexus *nx, struct kern_channel *ch,
1449 struct chreq *chr, struct nxbind *nxb, struct proc *p,
1450 struct nexus_adapter **na, boolean_t create)
1451 {
1452 #pragma unused(ch, p)
1453 struct nx_upipe *u = NX_UPIPE_PRIVATE(nx);
1454 struct nxprov_params *nxp = NX_PROV(nx)->nxprov_params;
1455 struct nexus_adapter *pna = NULL; /* parent adapter */
1456 boolean_t anon = NX_ANONYMOUS_PROV(nx);
1457 struct nexus_upipe_adapter *mna, *sna, *req;
1458 ch_endpoint_t ep = chr->cr_endpoint;
1459 uint32_t pipe_id;
1460 int error;
1461
1462 SK_LOCK_ASSERT_HELD();
1463 *na = NULL;
1464
1465 #if SK_LOG
1466 uuid_string_t uuidstr;
1467 SK_D("name \"%s\" spec_uuid \"%s\" port %d mode 0x%b pipe_id %u "
1468 "ring_id %d ring_set %u ep_type %u:%u create %u%s",
1469 chr->cr_name, sk_uuid_unparse(chr->cr_spec_uuid, uuidstr),
1470 (int)chr->cr_port, chr->cr_mode, CHMODE_BITS,
1471 chr->cr_pipe_id, (int)chr->cr_ring_id, chr->cr_ring_set,
1472 chr->cr_real_endpoint, chr->cr_endpoint, create,
1473 (ep != CH_ENDPOINT_USER_PIPE_MASTER &&
1474 ep != CH_ENDPOINT_USER_PIPE_SLAVE) ? " (skipped)" : "");
1475 #endif /* SK_LOG */
1476
1477 if (ep != CH_ENDPOINT_USER_PIPE_MASTER &&
1478 ep != CH_ENDPOINT_USER_PIPE_SLAVE) {
1479 return 0;
1480 }
1481
1482 /*
1483 * Check client credentials.
1484 */
1485 if (chr->cr_port == NEXUS_PORT_USER_PIPE_SERVER) {
1486 if (!anon && (u->nup_srv_nxb == NULL || nxb == NULL ||
1487 !nxb_is_equal(u->nup_srv_nxb, nxb))) {
1488 return EACCES;
1489 }
1490 } else {
1491 ASSERT(chr->cr_port == NEXUS_PORT_USER_PIPE_CLIENT);
1492 if (!anon && (u->nup_cli_nxb == NULL || nxb == NULL ||
1493 !nxb_is_equal(u->nup_cli_nxb, nxb))) {
1494 return EACCES;
1495 }
1496 }
1497
1498 /*
1499 * First, try to find a previously-created parent adapter
1500 * for this nexus; else, create one and store it in the
1501 * nexus. We'll release this at nexus destructor time.
1502 */
1503 if ((pna = u->nup_pna) != NULL) {
1504 na_retain_locked(pna); /* for us */
1505 SK_DF(SK_VERB_USER_PIPE, "found parent: \"%s\" (0x%llx)",
1506 pna->na_name, SK_KVA(pna));
1507 } else {
1508 /* callee will hold a reference for us upon success */
1509 error = na_pseudo_create(nx, chr, &pna);
1510 if (error != 0) {
1511 SK_ERR("parent create failed: %d", error);
1512 return error;
1513 }
1514 /* hold an extra reference for nx_upipe */
1515 u->nup_pna = pna;
1516 na_retain_locked(pna);
1517 SK_DF(SK_VERB_USER_PIPE, "created parent: \"%s\" (0x%llx)",
1518 pna->na_name, SK_KVA(pna));
1519 }
1520
1521 /* next, lookup the pipe id in the parent list */
1522 req = NULL;
1523 pipe_id = chr->cr_pipe_id;
1524 mna = nx_upipe_find(pna, pipe_id);
1525 if (mna != NULL) {
1526 if (mna->pna_role == ep) {
1527 SK_DF(SK_VERB_USER_PIPE,
1528 "found pipe_id %u directly at slot %u",
1529 pipe_id, mna->pna_parent_slot);
1530 req = mna;
1531 } else {
1532 SK_DF(SK_VERB_USER_PIPE,
1533 "found pipe_id %u indirectly at slot %u",
1534 pipe_id, mna->pna_parent_slot);
1535 req = mna->pna_peer;
1536 }
1537 /*
1538 * The pipe we have found already holds a ref to the parent,
1539 * so we need to drop the one we got from above.
1540 */
1541 (void) na_release_locked(pna);
1542 goto found;
1543 }
1544 SK_DF(SK_VERB_USER_PIPE,
1545 "pipe_id %u not found, create %u", pipe_id, create);
1546 if (!create) {
1547 error = ENODEV;
1548 goto put_out;
1549 }
1550 /*
1551 * We create both master and slave.
1552 * The endpoint we were asked for holds a reference to
1553 * the other one.
1554 */
1555 mna = na_upipe_alloc(Z_WAITOK);
1556
1557 ASSERT(mna->pna_up.na_type == NA_USER_PIPE);
1558 ASSERT(mna->pna_up.na_free == na_upipe_free);
1559
1560 (void) snprintf(mna->pna_up.na_name, sizeof(mna->pna_up.na_name),
1561 "%s{%u", pna->na_name, pipe_id);
1562 uuid_generate_random(mna->pna_up.na_uuid);
1563
1564 mna->pna_id = pipe_id;
1565 mna->pna_role = CH_ENDPOINT_USER_PIPE_MASTER;
1566 mna->pna_parent = pna;
1567 mna->pna_up.na_txsync = nx_upipe_na_txsync;
1568 mna->pna_up.na_rxsync = nx_upipe_na_rxsync;
1569 mna->pna_up.na_activate = nx_upipe_na_activate;
1570 mna->pna_up.na_dtor = nx_upipe_na_dtor;
1571 mna->pna_up.na_krings_create = nx_upipe_na_krings_create;
1572 mna->pna_up.na_krings_delete = nx_upipe_na_krings_delete;
1573 mna->pna_up.na_arena = pna->na_arena;
1574 skmem_arena_retain((&mna->pna_up)->na_arena);
1575 os_atomic_or(&mna->pna_up.na_flags, NAF_MEM_LOANED, relaxed);
1576 *(nexus_meta_type_t *)(uintptr_t)&mna->pna_up.na_md_type =
1577 pna->na_md_type;
1578 *(nexus_meta_subtype_t *)(uintptr_t)&mna->pna_up.na_md_subtype =
1579 pna->na_md_subtype;
1580
1581 *(nexus_stats_type_t *)(uintptr_t)&mna->pna_up.na_stats_type =
1582 NEXUS_STATS_TYPE_INVALID;
1583 *(uint32_t *)(uintptr_t)&mna->pna_up.na_flowadv_max =
1584 nxp->nxp_flowadv_max;
1585 ASSERT(mna->pna_up.na_flowadv_max == 0 ||
1586 skmem_arena_nexus(mna->pna_up.na_arena)->arn_flowadv_obj != NULL);
1587
1588 /*
1589 * Parent adapter parameters must match the nexus provider's by the
1590 * time we get here, since na_find() above shouldn't return
1591 * one otherwise.
1592 */
1593 na_set_nrings(&mna->pna_up, NR_TX, nxp->nxp_tx_rings);
1594 na_set_nrings(&mna->pna_up, NR_RX, nxp->nxp_rx_rings);
1595 na_set_nslots(&mna->pna_up, NR_TX, nxp->nxp_tx_slots);
1596 na_set_nslots(&mna->pna_up, NR_RX, nxp->nxp_rx_slots);
1597 ASSERT(na_get_nrings(&mna->pna_up, NR_TX) == na_get_nrings(pna, NR_TX));
1598 ASSERT(na_get_nrings(&mna->pna_up, NR_RX) == na_get_nrings(pna, NR_RX));
1599 ASSERT(na_get_nslots(&mna->pna_up, NR_TX) == na_get_nslots(pna, NR_TX));
1600 ASSERT(na_get_nslots(&mna->pna_up, NR_RX) == na_get_nslots(pna, NR_RX));
1601
1602 na_attach_common(&mna->pna_up, nx, &nx_upipe_prov_s);
1603
1604 /* register the master with the parent */
1605 error = nx_upipe_na_add(pna, mna);
1606 if (error != 0) {
1607 goto free_mna;
1608 }
1609
1610 /* create the slave */
1611 sna = na_upipe_alloc(Z_WAITOK);
1612
1613 /* most fields are the same, copy from master and then fix */
1614 bcopy(mna, sna, sizeof(*sna));
1615 skmem_arena_retain((&sna->pna_up)->na_arena);
1616 os_atomic_or(&sna->pna_up.na_flags, NAF_MEM_LOANED, relaxed);
1617
1618 ASSERT(sna->pna_up.na_type == NA_USER_PIPE);
1619 ASSERT(sna->pna_up.na_free == na_upipe_free);
1620
1621 (void) snprintf(sna->pna_up.na_name, sizeof(sna->pna_up.na_name),
1622 "%s}%d", pna->na_name, pipe_id);
1623 uuid_generate_random(sna->pna_up.na_uuid);
1624
1625 sna->pna_role = CH_ENDPOINT_USER_PIPE_SLAVE;
1626 na_attach_common(&sna->pna_up, nx, &nx_upipe_prov_s);
1627
1628 /* join the two endpoints */
1629 mna->pna_peer = sna;
1630 sna->pna_peer = mna;
1631
1632 /*
1633 * We already have a reference to the parent, but we
1634 * need another one for the other endpoint we created
1635 */
1636 na_retain_locked(pna);
1637
1638 if ((chr->cr_mode & CHMODE_DEFUNCT_OK) != 0) {
1639 os_atomic_or(&pna->na_flags, NAF_DEFUNCT_OK, relaxed);
1640 }
1641
1642 if (ep == CH_ENDPOINT_USER_PIPE_MASTER) {
1643 req = mna;
1644 mna->pna_peer_ref = TRUE;
1645 na_retain_locked(&sna->pna_up);
1646 } else {
1647 req = sna;
1648 sna->pna_peer_ref = TRUE;
1649 na_retain_locked(&mna->pna_up);
1650 }
1651
1652 /* parent adapter now has two users (mna and sna) */
1653 u->nup_pna_users += 2;
1654
1655 #if SK_LOG
1656 SK_DF(SK_VERB_USER_PIPE, "created master 0x%llx and slave 0x%llx",
1657 SK_KVA(mna), SK_KVA(sna));
1658 SK_DF(SK_VERB_USER_PIPE, "mna: \"%s\"", mna->pna_up.na_name);
1659 SK_DF(SK_VERB_USER_PIPE, " UUID: %s",
1660 sk_uuid_unparse(mna->pna_up.na_uuid, uuidstr));
1661 SK_DF(SK_VERB_USER_PIPE, " nx: 0x%llx (\"%s\":\"%s\")",
1662 SK_KVA(mna->pna_up.na_nx), NX_DOM(mna->pna_up.na_nx)->nxdom_name,
1663 NX_DOM_PROV(mna->pna_up.na_nx)->nxdom_prov_name);
1664 SK_DF(SK_VERB_USER_PIPE, " flags: 0x%b",
1665 mna->pna_up.na_flags, NAF_BITS);
1666 SK_DF(SK_VERB_USER_PIPE, " flowadv_max: %u",
1667 mna->pna_up.na_flowadv_max);
1668 SK_DF(SK_VERB_USER_PIPE, " rings: tx %u rx %u",
1669 na_get_nrings(&mna->pna_up, NR_TX),
1670 na_get_nrings(&mna->pna_up, NR_RX));
1671 SK_DF(SK_VERB_USER_PIPE, " slots: tx %u rx %u",
1672 na_get_nslots(&mna->pna_up, NR_TX),
1673 na_get_nslots(&mna->pna_up, NR_RX));
1674 SK_DF(SK_VERB_USER_PIPE, " next_pipe: %u", mna->pna_up.na_next_pipe);
1675 SK_DF(SK_VERB_USER_PIPE, " max_pipes: %u", mna->pna_up.na_max_pipes);
1676 SK_DF(SK_VERB_USER_PIPE, " parent: \"%s\"",
1677 mna->pna_parent->na_name);
1678 SK_DF(SK_VERB_USER_PIPE, " id: %u", mna->pna_id);
1679 SK_DF(SK_VERB_USER_PIPE, " role: %u", mna->pna_role);
1680 SK_DF(SK_VERB_USER_PIPE, " peer_ref: %u", mna->pna_peer_ref);
1681 SK_DF(SK_VERB_USER_PIPE, " parent_slot: %u", mna->pna_parent_slot);
1682 SK_DF(SK_VERB_USER_PIPE, "sna: \"%s\"", sna->pna_up.na_name);
1683 SK_DF(SK_VERB_USER_PIPE, " UUID: %s",
1684 sk_uuid_unparse(sna->pna_up.na_uuid, uuidstr));
1685 SK_DF(SK_VERB_USER_PIPE, " nx: 0x%llx (\"%s\":\"%s\")",
1686 SK_KVA(sna->pna_up.na_nx), NX_DOM(sna->pna_up.na_nx)->nxdom_name,
1687 NX_DOM_PROV(sna->pna_up.na_nx)->nxdom_prov_name);
1688 SK_DF(SK_VERB_USER_PIPE, " flags: 0x%b",
1689 sna->pna_up.na_flags, NAF_BITS);
1690 SK_DF(SK_VERB_USER_PIPE, " flowadv_max: %u",
1691 sna->pna_up.na_flowadv_max);
1692 SK_DF(SK_VERB_USER_PIPE, " rings: tx %u rx %u",
1693 na_get_nrings(&sna->pna_up, NR_TX),
1694 na_get_nrings(&sna->pna_up, NR_RX));
1695 SK_DF(SK_VERB_USER_PIPE, " slots: tx %u rx %u",
1696 na_get_nslots(&sna->pna_up, NR_TX),
1697 na_get_nslots(&sna->pna_up, NR_RX));
1698 SK_DF(SK_VERB_USER_PIPE, " next_pipe: %u", sna->pna_up.na_next_pipe);
1699 SK_DF(SK_VERB_USER_PIPE, " max_pipes: %u", sna->pna_up.na_max_pipes);
1700 SK_DF(SK_VERB_USER_PIPE, " parent: \"%s\"",
1701 sna->pna_parent->na_name);
1702 SK_DF(SK_VERB_USER_PIPE, " id: %u", sna->pna_id);
1703 SK_DF(SK_VERB_USER_PIPE, " role: %u", sna->pna_role);
1704 SK_DF(SK_VERB_USER_PIPE, " peer_ref: %u", sna->pna_peer_ref);
1705 SK_DF(SK_VERB_USER_PIPE, " parent_slot: %u", sna->pna_parent_slot);
1706 #endif /* SK_LOG */
1707
1708 found:
1709
1710 SK_DF(SK_VERB_USER_PIPE, "pipe_id %u role %s at 0x%llx", pipe_id,
1711 (req->pna_role == CH_ENDPOINT_USER_PIPE_MASTER ?
1712 "master" : "slave"), SK_KVA(req));
1713 if ((chr->cr_mode & CHMODE_DEFUNCT_OK) == 0) {
1714 os_atomic_andnot(&pna->na_flags, NAF_DEFUNCT_OK, relaxed);
1715 }
1716 *na = &req->pna_up;
1717 na_retain_locked(*na);
1718
1719 /*
1720 * Keep the reference to the parent; it will be released
1721 * by the adapter's destructor.
1722 */
1723 return 0;
1724
1725 free_mna:
1726 if (mna->pna_up.na_arena != NULL) {
1727 skmem_arena_release((&mna->pna_up)->na_arena);
1728 mna->pna_up.na_arena = NULL;
1729 }
1730 NA_FREE(&mna->pna_up);
1731 put_out:
1732 (void) na_release_locked(pna);
1733 return error;
1734 }
1735
1736 static struct nx_upipe *
nx_upipe_alloc(zalloc_flags_t how)1737 nx_upipe_alloc(zalloc_flags_t how)
1738 {
1739 struct nx_upipe *u;
1740
1741 SK_LOCK_ASSERT_HELD();
1742
1743 u = zalloc_flags(nx_upipe_zone, how | Z_ZERO);
1744 if (u) {
1745 SK_DF(SK_VERB_MEM, "upipe 0x%llx ALLOC", SK_KVA(u));
1746 }
1747 return u;
1748 }
1749
1750 static void
nx_upipe_free(struct nx_upipe * u)1751 nx_upipe_free(struct nx_upipe *u)
1752 {
1753 ASSERT(u->nup_pna == NULL);
1754 ASSERT(u->nup_pna_users == 0);
1755 ASSERT(u->nup_cli_nxb == NULL);
1756 ASSERT(u->nup_srv_nxb == NULL);
1757
1758 SK_DF(SK_VERB_MEM, "upipe 0x%llx FREE", SK_KVA(u));
1759 zfree(nx_upipe_zone, u);
1760 }
1761