1 /* $NetBSD: puffs_msgif.c,v 1.108 2025/02/01 22:53:00 andvar Exp $ */ 2 3 /* 4 * Copyright (c) 2005, 2006, 2007 Antti Kantee. All Rights Reserved. 5 * 6 * Development of this software was supported by the 7 * Google Summer of Code program and the Ulla Tuominen Foundation. 8 * The Google SoC project was mentored by Bill Studenmund. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS 20 * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 21 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 22 * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 25 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 */ 31 32 #include <sys/cdefs.h> 33 __KERNEL_RCSID(0, "$NetBSD: puffs_msgif.c,v 1.108 2025/02/01 22:53:00 andvar Exp $"); 34 35 #include <sys/param.h> 36 #include <sys/kernel.h> 37 #include <sys/atomic.h> 38 #include <sys/kmem.h> 39 #include <sys/kthread.h> 40 #include <sys/lock.h> 41 #include <sys/mount.h> 42 #include <sys/namei.h> 43 #include <sys/proc.h> 44 #include <sys/vnode.h> 45 #include <sys/atomic.h> 46 #include <sys/compat_stub.h> 47 48 #include <uvm/uvm.h> 49 50 #include <dev/putter/putter_sys.h> 51 52 #include <fs/puffs/puffs_msgif.h> 53 #include <fs/puffs/puffs_sys.h> 54 55 /* 56 * waitq data structures 57 */ 58 59 /* 60 * While a request is going to userspace, park the caller within the 61 * kernel. This is the kernel counterpart of "struct puffs_req". 62 */ 63 struct puffs_msgpark { 64 struct puffs_req *park_preq; /* req followed by buf */ 65 66 size_t park_copylen; /* userspace copylength */ 67 size_t park_maxlen; /* max size in comeback */ 68 69 struct puffs_req *park_creq; /* non-compat preq */ 70 size_t park_creqlen; /* non-compat preq len */ 71 72 parkdone_fn park_done; /* "biodone" a'la puffs */ 73 void *park_donearg; 74 75 int park_flags; 76 int park_refcount; 77 78 kcondvar_t park_cv; 79 kmutex_t park_mtx; 80 81 TAILQ_ENTRY(puffs_msgpark) park_entries; 82 }; 83 #define PARKFLAG_WAITERGONE 0x01 84 #define PARKFLAG_DONE 0x02 85 #define PARKFLAG_ONQUEUE1 0x04 86 #define PARKFLAG_ONQUEUE2 0x08 87 #define PARKFLAG_CALL 0x10 88 #define PARKFLAG_WANTREPLY 0x20 89 #define PARKFLAG_HASERROR 0x40 90 91 static pool_cache_t parkpc; 92 #ifdef PUFFSDEBUG 93 static int totalpark; 94 #endif 95 96 int puffs_sopreq_expire_timeout = PUFFS_SOPREQ_EXPIRE_TIMEOUT; 97 98 static int 99 makepark(void *arg, void *obj, int flags) 100 { 101 struct puffs_msgpark *park = obj; 102 103 mutex_init(&park->park_mtx, MUTEX_DEFAULT, IPL_NONE); 104 cv_init(&park->park_cv, "puffsrpl"); 105 106 return 0; 107 } 108 109 static void 110 nukepark(void *arg, void *obj) 111 { 112 struct puffs_msgpark *park = obj; 113 114 cv_destroy(&park->park_cv); 115 mutex_destroy(&park->park_mtx); 116 } 117 118 void 119 puffs_msgif_init(void) 120 { 121 122 parkpc = pool_cache_init(sizeof(struct puffs_msgpark), 0, 0, 0, 123 "puffprkl", NULL, IPL_NONE, makepark, nukepark, NULL); 124 } 125 126 void 127 puffs_msgif_destroy(void) 128 { 129 130 pool_cache_destroy(parkpc); 131 } 132 133 static struct puffs_msgpark * 134 puffs_msgpark_alloc(int waitok) 135 { 136 struct puffs_msgpark *park; 137 138 KASSERT(curlwp != uvm.pagedaemon_lwp || !waitok); 139 140 park = pool_cache_get(parkpc, waitok ? PR_WAITOK : PR_NOWAIT); 141 if (park == NULL) 142 return park; 143 144 park->park_refcount = 1; 145 park->park_preq = park->park_creq = NULL; 146 park->park_flags = PARKFLAG_WANTREPLY; 147 148 #ifdef PUFFSDEBUG 149 totalpark++; 150 #endif 151 152 return park; 153 } 154 155 static void 156 puffs_msgpark_reference(struct puffs_msgpark *park) 157 { 158 159 KASSERT(mutex_owned(&park->park_mtx)); 160 park->park_refcount++; 161 } 162 163 /* 164 * Release reference to park structure. 165 */ 166 static void 167 puffs_msgpark_release1(struct puffs_msgpark *park, int howmany) 168 { 169 struct puffs_req *preq = park->park_preq; 170 struct puffs_req *creq = park->park_creq; 171 int refcnt; 172 173 KASSERT(mutex_owned(&park->park_mtx)); 174 refcnt = park->park_refcount -= howmany; 175 mutex_exit(&park->park_mtx); 176 177 KASSERT(refcnt >= 0); 178 179 if (refcnt == 0) { 180 if (preq) 181 kmem_free(preq, park->park_maxlen); 182 #if 1 183 if (creq) 184 kmem_free(creq, park->park_creqlen); 185 #endif 186 pool_cache_put(parkpc, park); 187 188 #ifdef PUFFSDEBUG 189 totalpark--; 190 #endif 191 } 192 } 193 #define puffs_msgpark_release(a) puffs_msgpark_release1(a, 1) 194 195 #ifdef PUFFSDEBUG 196 static void 197 parkdump(struct puffs_msgpark *park) 198 { 199 200 DPRINTF(("park %p, preq %p, id %" PRIu64 "\n" 201 "\tcopy %zu, max %zu - done: %p/%p\n" 202 "\tflags 0x%08x, refcount %d, cv/mtx: %p/%p\n", 203 park, park->park_preq, park->park_preq->preq_id, 204 park->park_copylen, park->park_maxlen, 205 park->park_done, park->park_donearg, 206 park->park_flags, park->park_refcount, 207 &park->park_cv, &park->park_mtx)); 208 } 209 210 static void 211 parkqdump(struct puffs_wq *q, int dumpall) 212 { 213 struct puffs_msgpark *park; 214 int total = 0; 215 216 TAILQ_FOREACH(park, q, park_entries) { 217 if (dumpall) 218 parkdump(park); 219 total++; 220 } 221 DPRINTF(("puffs waitqueue at %p dumped, %d total\n", q, total)); 222 223 } 224 #endif /* PUFFSDEBUG */ 225 226 /* 227 * A word about locking in the park structures: the lock protects the 228 * fields of the *park* structure (not preq) and acts as an interlock 229 * in cv operations. The lock is always internal to this module and 230 * callers do not need to worry about it. 231 */ 232 233 int 234 puffs_msgmem_alloc(size_t len, struct puffs_msgpark **ppark, void **mem, 235 int cansleep) 236 { 237 struct puffs_msgpark *park; 238 void *m; 239 240 KASSERT(curlwp != uvm.pagedaemon_lwp || !cansleep); 241 m = kmem_zalloc(len, cansleep ? KM_SLEEP : KM_NOSLEEP); 242 if (m == NULL) { 243 KASSERT(cansleep == 0); 244 return ENOMEM; 245 } 246 247 park = puffs_msgpark_alloc(cansleep); 248 if (park == NULL) { 249 KASSERT(cansleep == 0); 250 kmem_free(m, len); 251 return ENOMEM; 252 } 253 254 park->park_preq = m; 255 park->park_maxlen = park->park_copylen = len; 256 257 *ppark = park; 258 *mem = m; 259 260 return 0; 261 } 262 263 void 264 puffs_msgmem_release(struct puffs_msgpark *park) 265 { 266 267 if (park == NULL) 268 return; 269 270 mutex_enter(&park->park_mtx); 271 puffs_msgpark_release(park); 272 } 273 274 void 275 puffs_msg_setfaf(struct puffs_msgpark *park) 276 { 277 278 KASSERT((park->park_flags & PARKFLAG_CALL) == 0); 279 park->park_flags &= ~PARKFLAG_WANTREPLY; 280 } 281 282 void 283 puffs_msg_setdelta(struct puffs_msgpark *park, size_t delta) 284 { 285 286 KASSERT(delta < park->park_maxlen); /* "<=" wouldn't make sense */ 287 park->park_copylen = park->park_maxlen - delta; 288 } 289 290 void 291 puffs_msg_setinfo(struct puffs_msgpark *park, int opclass, int type, 292 puffs_cookie_t ck) 293 { 294 295 park->park_preq->preq_opclass = PUFFSOP_OPCLASS(opclass); 296 park->park_preq->preq_optype = type; 297 park->park_preq->preq_cookie = ck; 298 } 299 300 void 301 puffs_msg_setcall(struct puffs_msgpark *park, parkdone_fn donefn, void *donearg) 302 { 303 304 KASSERT(park->park_flags & PARKFLAG_WANTREPLY); 305 park->park_done = donefn; 306 park->park_donearg = donearg; 307 park->park_flags |= PARKFLAG_CALL; 308 } 309 310 /* 311 * kernel-user-kernel waitqueues 312 */ 313 314 static uint64_t 315 puffs_getmsgid(struct puffs_mount *pmp) 316 { 317 uint64_t rv; 318 319 mutex_enter(&pmp->pmp_lock); 320 rv = pmp->pmp_nextmsgid++; 321 mutex_exit(&pmp->pmp_lock); 322 323 return rv; 324 } 325 326 /* 327 * A word about reference counting of parks. A reference must be taken 328 * when accessing a park and additionally when it is on a queue. So 329 * when taking it off a queue and releasing the access reference, the 330 * reference count is generally decremented by 2. 331 */ 332 333 void 334 puffs_msg_enqueue(struct puffs_mount *pmp, struct puffs_msgpark *park) 335 { 336 struct lwp *l = curlwp; 337 struct puffs_req *preq, *creq; 338 ssize_t delta; 339 #if 1 340 int ret; 341 #endif 342 343 /* 344 * Some clients reuse a park, so reset some flags. We might 345 * want to provide a caller-side interface for this and add 346 * a few more invariant checks here, but this will do for now. 347 */ 348 park->park_flags &= ~(PARKFLAG_DONE | PARKFLAG_HASERROR); 349 KASSERT((park->park_flags & PARKFLAG_WAITERGONE) == 0); 350 351 preq = park->park_preq; 352 353 #if 1 354 /* check if we do compat adjustments */ 355 if (pmp->pmp_docompat) { 356 MODULE_HOOK_CALL(puffs_out_50_hook, (preq, &creq, &delta), 357 enosys(), ret); 358 if (ret == 0) { 359 park->park_creq = park->park_preq; 360 park->park_creqlen = park->park_maxlen; 361 362 park->park_maxlen += delta; 363 park->park_copylen += delta; 364 park->park_preq = preq = creq; 365 } 366 } 367 #endif 368 369 preq->preq_buflen = park->park_maxlen; 370 KASSERT(preq->preq_id == 0 371 || (preq->preq_opclass & PUFFSOPFLAG_ISRESPONSE)); 372 373 if ((park->park_flags & PARKFLAG_WANTREPLY) == 0) 374 preq->preq_opclass |= PUFFSOPFLAG_FAF; 375 else 376 preq->preq_id = puffs_getmsgid(pmp); 377 378 /* fill in caller information */ 379 preq->preq_pid = l->l_proc->p_pid; 380 preq->preq_lid = l->l_lid; 381 382 /* 383 * To support cv_sig, yet another movie: check if there are signals 384 * pending and we are issuing a non-FAF. If so, return an error 385 * directly UNLESS we are issuing INACTIVE/RECLAIM. In that case, 386 * convert it to a FAF, fire off to the file server and return 387 * an error. Yes, this is bordering disgusting. Barfbags are on me. 388 */ 389 if (__predict_false((park->park_flags & PARKFLAG_WANTREPLY) 390 && (park->park_flags & PARKFLAG_CALL) == 0 391 && (l->l_flag & LW_PENDSIG) != 0 && sigispending(l, 0))) { 392 sigset_t ss; 393 394 /* 395 * see the comment about signals in puffs_msg_wait. 396 */ 397 sigpending1(l, &ss); 398 if (sigismember(&ss, SIGINT) || 399 sigismember(&ss, SIGTERM) || 400 sigismember(&ss, SIGKILL) || 401 sigismember(&ss, SIGHUP) || 402 sigismember(&ss, SIGQUIT)) { 403 park->park_flags |= PARKFLAG_HASERROR; 404 preq->preq_rv = EINTR; 405 if (PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN 406 && (preq->preq_optype == PUFFS_VN_INACTIVE 407 || preq->preq_optype == PUFFS_VN_RECLAIM)) { 408 park->park_preq->preq_opclass |= 409 PUFFSOPFLAG_FAF; 410 park->park_flags &= ~PARKFLAG_WANTREPLY; 411 DPRINTF(("puffs_msg_enqueue: " 412 "converted to FAF %p\n", park)); 413 } else { 414 return; 415 } 416 } 417 } 418 419 mutex_enter(&pmp->pmp_lock); 420 if (pmp->pmp_status != PUFFSTAT_RUNNING) { 421 mutex_exit(&pmp->pmp_lock); 422 park->park_flags |= PARKFLAG_HASERROR; 423 preq->preq_rv = ENXIO; 424 return; 425 } 426 427 #ifdef PUFFSDEBUG 428 parkqdump(&pmp->pmp_msg_touser, puffsdebug > 1); 429 parkqdump(&pmp->pmp_msg_replywait, puffsdebug > 1); 430 #endif 431 432 /* 433 * Note: we don't need to lock park since we have the only 434 * reference to it at this point. 435 */ 436 TAILQ_INSERT_TAIL(&pmp->pmp_msg_touser, park, park_entries); 437 park->park_flags |= PARKFLAG_ONQUEUE1; 438 pmp->pmp_msg_touser_count++; 439 park->park_refcount++; 440 441 cv_broadcast(&pmp->pmp_msg_waiter_cv); 442 mutex_exit(&pmp->pmp_lock); 443 putter_notify(pmp->pmp_pi); 444 445 DPRINTF(("touser: req %" PRIu64 ", preq: %p, park: %p, " 446 "c/t: 0x%x/0x%x, f: 0x%x\n", preq->preq_id, preq, park, 447 preq->preq_opclass, preq->preq_optype, park->park_flags)); 448 } 449 450 int 451 puffs_msg_wait(struct puffs_mount *pmp, struct puffs_msgpark *park) 452 { 453 lwp_t *l = curlwp; 454 proc_t *p = l->l_proc; 455 struct puffs_req *preq = park->park_preq; /* XXX: hmmm */ 456 sigset_t ss; 457 sigset_t oss; 458 int error = 0; 459 int rv; 460 461 /* 462 * block unimportant signals. 463 * 464 * The set of "important" signals here was chosen to be same as 465 * nfs interruptible mount. 466 */ 467 sigfillset(&ss); 468 sigdelset(&ss, SIGINT); 469 sigdelset(&ss, SIGTERM); 470 sigdelset(&ss, SIGKILL); 471 sigdelset(&ss, SIGHUP); 472 sigdelset(&ss, SIGQUIT); 473 mutex_enter(p->p_lock); 474 sigprocmask1(l, SIG_BLOCK, &ss, &oss); 475 mutex_exit(p->p_lock); 476 477 mutex_enter(&pmp->pmp_lock); 478 puffs_mp_reference(pmp); 479 mutex_exit(&pmp->pmp_lock); 480 481 mutex_enter(&park->park_mtx); 482 /* did the response beat us to the wait? */ 483 if (__predict_false((park->park_flags & PARKFLAG_DONE) 484 || (park->park_flags & PARKFLAG_HASERROR))) { 485 rv = park->park_preq->preq_rv; 486 mutex_exit(&park->park_mtx); 487 goto skipwait; 488 } 489 490 if ((park->park_flags & PARKFLAG_WANTREPLY) == 0 491 || (park->park_flags & PARKFLAG_CALL)) { 492 mutex_exit(&park->park_mtx); 493 rv = 0; 494 goto skipwait; 495 } 496 497 error = cv_wait_sig(&park->park_cv, &park->park_mtx); 498 DPRINTF(("puffs_touser: waiter for %p woke up with %d\n", 499 park, error)); 500 if (error) { 501 park->park_flags |= PARKFLAG_WAITERGONE; 502 if (park->park_flags & PARKFLAG_DONE) { 503 rv = preq->preq_rv; 504 mutex_exit(&park->park_mtx); 505 } else { 506 /* 507 * ok, we marked it as going away, but 508 * still need to do queue ops. take locks 509 * in correct order. 510 * 511 * We don't want to release our reference 512 * if it's on replywait queue to avoid error 513 * to file server. putop() code will DTRT. 514 */ 515 mutex_exit(&park->park_mtx); 516 mutex_enter(&pmp->pmp_lock); 517 mutex_enter(&park->park_mtx); 518 519 /* 520 * Still on queue1? We can safely remove it 521 * without any consequences since the file 522 * server hasn't seen it. "else" we need to 523 * wait for the response and just ignore it 524 * to avoid signalling an incorrect error to 525 * the file server. 526 */ 527 if (park->park_flags & PARKFLAG_ONQUEUE1) { 528 TAILQ_REMOVE(&pmp->pmp_msg_touser, 529 park, park_entries); 530 puffs_msgpark_release(park); 531 pmp->pmp_msg_touser_count--; 532 park->park_flags &= ~PARKFLAG_ONQUEUE1; 533 } else { 534 mutex_exit(&park->park_mtx); 535 } 536 mutex_exit(&pmp->pmp_lock); 537 538 rv = EINTR; 539 } 540 } else { 541 rv = preq->preq_rv; 542 mutex_exit(&park->park_mtx); 543 } 544 545 skipwait: 546 mutex_enter(&pmp->pmp_lock); 547 puffs_mp_release(pmp); 548 mutex_exit(&pmp->pmp_lock); 549 550 mutex_enter(p->p_lock); 551 sigprocmask1(l, SIG_SETMASK, &oss, NULL); 552 mutex_exit(p->p_lock); 553 554 return rv; 555 } 556 557 /* 558 * XXX: this suuuucks. Hopefully I'll get rid of this lossage once 559 * the whole setback-nonsense gets fixed. 560 */ 561 int 562 puffs_msg_wait2(struct puffs_mount *pmp, struct puffs_msgpark *park, 563 struct puffs_node *pn1, struct puffs_node *pn2) 564 { 565 struct puffs_req *preq; 566 int rv; 567 568 rv = puffs_msg_wait(pmp, park); 569 570 preq = park->park_preq; 571 if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N1) 572 pn1->pn_stat |= PNODE_DOINACT; 573 if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N2) 574 pn2->pn_stat |= PNODE_DOINACT; 575 576 if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N1) 577 pn1->pn_stat |= PNODE_NOREFS; 578 if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N2) 579 pn2->pn_stat |= PNODE_NOREFS; 580 581 return rv; 582 583 } 584 585 /* 586 * XXX: lazy bum. please, for the love of foie gras, fix me. 587 * This should *NOT* depend on setfaf. Also "memcpy" could 588 * be done more nicely. 589 */ 590 void 591 puffs_msg_sendresp(struct puffs_mount *pmp, struct puffs_req *origpreq, int rv) 592 { 593 struct puffs_msgpark *park; 594 struct puffs_req *preq; 595 596 puffs_msgmem_alloc(sizeof(struct puffs_req), &park, (void *)&preq, 1); 597 puffs_msg_setfaf(park); /* XXXXXX: avoids reqid override */ 598 599 memcpy(preq, origpreq, sizeof(struct puffs_req)); 600 preq->preq_rv = rv; 601 preq->preq_opclass |= PUFFSOPFLAG_ISRESPONSE; 602 603 puffs_msg_enqueue(pmp, park); 604 puffs_msgmem_release(park); 605 } 606 607 /* 608 * Get next request in the outgoing queue. "maxsize" controls the 609 * size the caller can accommodate and "nonblock" signals if this 610 * should block while waiting for input. Handles all locking internally. 611 */ 612 int 613 puffs_msgif_getout(void *ctx, size_t maxsize, int nonblock, 614 uint8_t **data, size_t *dlen, void **parkptr) 615 { 616 struct puffs_mount *pmp = ctx; 617 struct puffs_msgpark *park = NULL; 618 struct puffs_req *preq = NULL; 619 int error; 620 621 error = 0; 622 mutex_enter(&pmp->pmp_lock); 623 puffs_mp_reference(pmp); 624 for (;;) { 625 /* RIP? */ 626 if (pmp->pmp_status != PUFFSTAT_RUNNING) { 627 error = ENXIO; 628 break; 629 } 630 631 /* need platinum yendorian express card? */ 632 if (TAILQ_EMPTY(&pmp->pmp_msg_touser)) { 633 DPRINTF(("puffs_getout: no outgoing op, ")); 634 if (nonblock) { 635 DPRINTF(("returning EWOULDBLOCK\n")); 636 error = EWOULDBLOCK; 637 break; 638 } 639 DPRINTF(("waiting ...\n")); 640 641 error = cv_wait_sig(&pmp->pmp_msg_waiter_cv, 642 &pmp->pmp_lock); 643 if (error) 644 break; 645 else 646 continue; 647 } 648 649 park = TAILQ_FIRST(&pmp->pmp_msg_touser); 650 if (park == NULL) 651 continue; 652 653 mutex_enter(&park->park_mtx); 654 puffs_msgpark_reference(park); 655 656 DPRINTF(("puffs_getout: found park at %p, ", park)); 657 658 /* If it's a goner, don't process any further */ 659 if (park->park_flags & PARKFLAG_WAITERGONE) { 660 DPRINTF(("waitergone!\n")); 661 puffs_msgpark_release(park); 662 continue; 663 } 664 preq = park->park_preq; 665 666 #if 0 667 /* check size */ 668 /* 669 * XXX: this check is not valid for now, we don't know 670 * the size of the caller's input buffer. i.e. this 671 * will most likely go away 672 */ 673 if (maxsize < preq->preq_frhdr.pfr_len) { 674 DPRINTF(("buffer too small\n")); 675 puffs_msgpark_release(park); 676 error = E2BIG; 677 break; 678 } 679 #endif 680 681 DPRINTF(("returning\n")); 682 683 /* 684 * Ok, we found what we came for. Release it from the 685 * outgoing queue but do not unlock. We will unlock 686 * only after we "releaseout" it to avoid complications: 687 * otherwise it is (theoretically) possible for userland 688 * to race us into "put" before we have a change to put 689 * this baby on the receiving queue. 690 */ 691 TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries); 692 KASSERT(park->park_flags & PARKFLAG_ONQUEUE1); 693 park->park_flags &= ~PARKFLAG_ONQUEUE1; 694 mutex_exit(&park->park_mtx); 695 696 pmp->pmp_msg_touser_count--; 697 KASSERT(pmp->pmp_msg_touser_count >= 0); 698 699 break; 700 } 701 puffs_mp_release(pmp); 702 mutex_exit(&pmp->pmp_lock); 703 704 if (error == 0) { 705 *data = (uint8_t *)preq; 706 preq->preq_pth.pth_framelen = park->park_copylen; 707 *dlen = preq->preq_pth.pth_framelen; 708 *parkptr = park; 709 } 710 711 return error; 712 } 713 714 /* 715 * Release outgoing structure. Now, depending on the success of the 716 * outgoing send, it is either going onto the result waiting queue 717 * or the death chamber. 718 */ 719 void 720 puffs_msgif_releaseout(void *ctx, void *parkptr, int status) 721 { 722 struct puffs_mount *pmp = ctx; 723 struct puffs_msgpark *park = parkptr; 724 725 DPRINTF(("puffs_releaseout: returning park %p, errno %d: " , 726 park, status)); 727 mutex_enter(&pmp->pmp_lock); 728 mutex_enter(&park->park_mtx); 729 if (park->park_flags & PARKFLAG_WANTREPLY) { 730 if (status == 0) { 731 DPRINTF(("enqueue replywait\n")); 732 TAILQ_INSERT_TAIL(&pmp->pmp_msg_replywait, park, 733 park_entries); 734 park->park_flags |= PARKFLAG_ONQUEUE2; 735 } else { 736 DPRINTF(("error path!\n")); 737 park->park_preq->preq_rv = status; 738 park->park_flags |= PARKFLAG_DONE; 739 cv_signal(&park->park_cv); 740 } 741 puffs_msgpark_release(park); 742 } else { 743 DPRINTF(("release\n")); 744 puffs_msgpark_release1(park, 2); 745 } 746 mutex_exit(&pmp->pmp_lock); 747 } 748 749 size_t 750 puffs_msgif_waitcount(void *ctx) 751 { 752 struct puffs_mount *pmp = ctx; 753 size_t rv; 754 755 mutex_enter(&pmp->pmp_lock); 756 rv = pmp->pmp_msg_touser_count; 757 mutex_exit(&pmp->pmp_lock); 758 759 return rv; 760 } 761 762 /* 763 * XXX: locking with this one? 764 */ 765 static void 766 puffsop_msg(void *ctx, struct puffs_req *preq) 767 { 768 struct puffs_mount *pmp = ctx; 769 struct putter_hdr *pth = &preq->preq_pth; 770 struct puffs_msgpark *park; 771 int wgone; 772 773 mutex_enter(&pmp->pmp_lock); 774 775 /* Locate waiter */ 776 TAILQ_FOREACH(park, &pmp->pmp_msg_replywait, park_entries) { 777 if (park->park_preq->preq_id == preq->preq_id) 778 break; 779 } 780 if (park == NULL) { 781 DPRINTF(("puffsop_msg: no request: %" PRIu64 "\n", 782 preq->preq_id)); 783 mutex_exit(&pmp->pmp_lock); 784 return; /* XXX send error */ 785 } 786 787 mutex_enter(&park->park_mtx); 788 puffs_msgpark_reference(park); 789 if (pth->pth_framelen > park->park_maxlen) { 790 DPRINTF(("puffsop_msg: invalid buffer length: " 791 "%" PRIu64 " (req %" PRIu64 ", \n", pth->pth_framelen, 792 preq->preq_id)); 793 park->park_preq->preq_rv = EPROTO; 794 cv_signal(&park->park_cv); 795 puffs_msgpark_release1(park, 2); 796 mutex_exit(&pmp->pmp_lock); 797 return; /* XXX: error */ 798 } 799 wgone = park->park_flags & PARKFLAG_WAITERGONE; 800 801 KASSERT(park->park_flags & PARKFLAG_ONQUEUE2); 802 TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries); 803 park->park_flags &= ~PARKFLAG_ONQUEUE2; 804 mutex_exit(&pmp->pmp_lock); 805 806 if (wgone) { 807 DPRINTF(("puffsop_msg: bad service - waiter gone for " 808 "park %p\n", park)); 809 } else { 810 #if 1 811 if (park->park_creq) { 812 struct puffs_req *creq; 813 size_t csize; 814 815 KASSERT(pmp->pmp_docompat); 816 MODULE_HOOK_CALL_VOID(puffs_in_50_hook, 817 (preq, park->park_creq), __nothing); 818 creq = park->park_creq; 819 csize = park->park_creqlen; 820 park->park_creq = park->park_preq; 821 park->park_creqlen = park->park_maxlen; 822 823 park->park_preq = creq; 824 park->park_maxlen = csize; 825 826 memcpy(park->park_creq, preq, pth->pth_framelen); 827 } else { 828 #endif 829 memcpy(park->park_preq, preq, pth->pth_framelen); 830 } 831 832 if (park->park_flags & PARKFLAG_CALL) { 833 DPRINTF(("puffsop_msg: call for %p, arg %p\n", 834 park->park_preq, park->park_donearg)); 835 park->park_done(pmp, preq, park->park_donearg); 836 } 837 } 838 839 if (!wgone) { 840 DPRINTF(("puffs_putop: flagging done for " 841 "park %p\n", park)); 842 cv_signal(&park->park_cv); 843 } 844 845 park->park_flags |= PARKFLAG_DONE; 846 puffs_msgpark_release1(park, 2); 847 } 848 849 /* 850 * Node expiry. We come here after an inactive on an unexpired node. 851 * The expiry has been queued and is done in sop thread. 852 */ 853 static void 854 puffsop_expire(struct puffs_mount *pmp, puffs_cookie_t cookie) 855 { 856 struct vnode *vp; 857 858 KASSERT(PUFFS_USE_FS_TTL(pmp)); 859 860 /* 861 * If it still exists and has no reference, 862 * vrele should cause it to be reclaimed. 863 * Otherwise, we have nothing to do. 864 */ 865 if (puffs_cookie2vnode(pmp, cookie, &vp) == 0) { 866 VPTOPP(vp)->pn_stat &= ~PNODE_SOPEXP; 867 vrele(vp); 868 } 869 870 return; 871 } 872 873 static void 874 puffsop_flush(struct puffs_mount *pmp, struct puffs_flush *pf) 875 { 876 struct vnode *vp; 877 voff_t offlo, offhi; 878 int rv, flags = 0; 879 880 KASSERT(pf->pf_req.preq_pth.pth_framelen == sizeof(struct puffs_flush)); 881 882 /* XXX: slurry */ 883 if (pf->pf_op == PUFFS_INVAL_NAMECACHE_ALL) { 884 cache_purgevfs(PMPTOMP(pmp)); 885 rv = 0; 886 goto out; 887 } 888 889 /* 890 * Get vnode, don't lock it. Namecache is protected by its own lock 891 * and we have a reference to protect against premature harvesting. 892 * 893 * The node we want here might be locked and the op is in 894 * userspace waiting for us to complete ==> deadlock. Another 895 * reason we need to eventually bump locking to userspace, as we 896 * will need to lock the node if we wish to do flushes. 897 */ 898 rv = puffs_cookie2vnode(pmp, pf->pf_cookie, &vp); 899 if (rv) { 900 if (rv == PUFFS_NOSUCHCOOKIE) 901 rv = ENOENT; 902 goto out; 903 } 904 905 switch (pf->pf_op) { 906 #if 0 907 /* not quite ready, yet */ 908 case PUFFS_INVAL_NAMECACHE_NODE: 909 struct componentname *pf_cn; 910 char *name; 911 /* get comfortab^Wcomponentname */ 912 pf_cn = kmem_alloc(componentname); 913 memset(pf_cn, 0, sizeof(struct componentname)); 914 break; 915 916 #endif 917 case PUFFS_INVAL_NAMECACHE_DIR: 918 if (vp->v_type != VDIR) { 919 rv = EINVAL; 920 break; 921 } 922 cache_purge1(vp, NULL, 0, PURGE_CHILDREN); 923 break; 924 925 case PUFFS_INVAL_PAGECACHE_NODE_RANGE: 926 flags = PGO_FREE; 927 /*FALLTHROUGH*/ 928 case PUFFS_FLUSH_PAGECACHE_NODE_RANGE: 929 if (flags == 0) 930 flags = PGO_CLEANIT; 931 932 if (pf->pf_end > vp->v_size || vp->v_type != VREG) { 933 rv = EINVAL; 934 break; 935 } 936 937 offlo = trunc_page(pf->pf_start); 938 offhi = round_page(pf->pf_end); 939 if (offhi != 0 && offlo >= offhi) { 940 rv = EINVAL; 941 break; 942 } 943 944 rw_enter(vp->v_uobj.vmobjlock, RW_WRITER); 945 rv = VOP_PUTPAGES(vp, offlo, offhi, flags); 946 break; 947 948 default: 949 rv = EINVAL; 950 } 951 952 vrele(vp); 953 954 out: 955 puffs_msg_sendresp(pmp, &pf->pf_req, rv); 956 } 957 958 int 959 puffs_msgif_dispatch(void *ctx, struct putter_hdr *pth) 960 { 961 struct puffs_mount *pmp = ctx; 962 struct puffs_req *preq = (struct puffs_req *)pth; 963 struct puffs_sopreq *psopr; 964 965 if (pth->pth_framelen < sizeof(struct puffs_req)) { 966 puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */ 967 return 0; 968 } 969 970 switch (PUFFSOP_OPCLASS(preq->preq_opclass)) { 971 case PUFFSOP_VN: 972 case PUFFSOP_VFS: 973 DPRINTF(("dispatch: vn/vfs message 0x%x\n", preq->preq_optype)); 974 puffsop_msg(pmp, preq); 975 break; 976 977 case PUFFSOP_FLUSH: /* process in sop thread */ 978 { 979 struct puffs_flush *pf; 980 981 DPRINTF(("dispatch: flush 0x%x\n", preq->preq_optype)); 982 983 if (preq->preq_pth.pth_framelen != sizeof(struct puffs_flush)) { 984 puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */ 985 break; 986 } 987 pf = (struct puffs_flush *)preq; 988 989 KASSERT(curlwp != uvm.pagedaemon_lwp); 990 psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP); 991 memcpy(&psopr->psopr_pf, pf, sizeof(*pf)); 992 psopr->psopr_sopreq = PUFFS_SOPREQ_FLUSH; 993 994 mutex_enter(&pmp->pmp_sopmtx); 995 if (pmp->pmp_sopthrcount == 0) { 996 mutex_exit(&pmp->pmp_sopmtx); 997 kmem_free(psopr, sizeof(*psopr)); 998 puffs_msg_sendresp(pmp, preq, ENXIO); 999 } else { 1000 TAILQ_INSERT_TAIL(&pmp->pmp_sopfastreqs, 1001 psopr, psopr_entries); 1002 cv_signal(&pmp->pmp_sopcv); 1003 mutex_exit(&pmp->pmp_sopmtx); 1004 } 1005 break; 1006 } 1007 1008 case PUFFSOP_UNMOUNT: /* process in sop thread */ 1009 { 1010 1011 DPRINTF(("dispatch: unmount 0x%x\n", preq->preq_optype)); 1012 1013 KASSERT(curlwp != uvm.pagedaemon_lwp); 1014 psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP); 1015 psopr->psopr_preq = *preq; 1016 psopr->psopr_sopreq = PUFFS_SOPREQ_UNMOUNT; 1017 1018 mutex_enter(&pmp->pmp_sopmtx); 1019 if (pmp->pmp_sopthrcount == 0) { 1020 mutex_exit(&pmp->pmp_sopmtx); 1021 kmem_free(psopr, sizeof(*psopr)); 1022 puffs_msg_sendresp(pmp, preq, ENXIO); 1023 } else { 1024 TAILQ_INSERT_TAIL(&pmp->pmp_sopfastreqs, 1025 psopr, psopr_entries); 1026 cv_signal(&pmp->pmp_sopcv); 1027 mutex_exit(&pmp->pmp_sopmtx); 1028 } 1029 break; 1030 } 1031 1032 default: 1033 DPRINTF(("dispatch: invalid opclass 0x%x\n", preq->preq_opclass)); 1034 puffs_msg_sendresp(pmp, preq, EOPNOTSUPP); 1035 break; 1036 } 1037 1038 return 0; 1039 } 1040 1041 /* 1042 * Work loop for thread processing all ops from server which 1043 * cannot safely be handled in caller context. This includes 1044 * everything which might need a lock currently "held" by the file 1045 * server, i.e. a long-term kernel lock which will be released only 1046 * once the file server acknowledges a request 1047 */ 1048 #define TIMED_OUT(expire) \ 1049 ((int)((unsigned int)getticks() - (unsigned int)expire) > 0) 1050 void 1051 puffs_sop_thread(void *arg) 1052 { 1053 struct puffs_mount *pmp = arg; 1054 struct mount *mp = PMPTOMP(pmp); 1055 struct puffs_sopreq *psopr; 1056 bool keeprunning; 1057 bool unmountme = false; 1058 int timeo; 1059 1060 timeo = PUFFS_USE_FS_TTL(pmp) ? puffs_sopreq_expire_timeout : 0; 1061 1062 mutex_enter(&pmp->pmp_sopmtx); 1063 for (keeprunning = true; keeprunning; ) { 1064 /* 1065 * We have a fast queue for flush and umount, and a node 1066 * queue for delayes node reclaims. Requests on node queue * are not honoured before clock reaches psopr_at. This 1067 * code assumes that requests are ordered by psopr_at. 1068 */ 1069 do { 1070 psopr = TAILQ_FIRST(&pmp->pmp_sopfastreqs); 1071 if (psopr != NULL) { 1072 TAILQ_REMOVE(&pmp->pmp_sopfastreqs, 1073 psopr, psopr_entries); 1074 break; 1075 } 1076 1077 psopr = TAILQ_FIRST(&pmp->pmp_sopnodereqs); 1078 if ((psopr != NULL) && TIMED_OUT(psopr->psopr_at)) { 1079 TAILQ_REMOVE(&pmp->pmp_sopnodereqs, 1080 psopr, psopr_entries); 1081 break; 1082 } 1083 1084 cv_timedwait(&pmp->pmp_sopcv, &pmp->pmp_sopmtx, timeo); 1085 } while (1 /* CONSTCOND */); 1086 1087 mutex_exit(&pmp->pmp_sopmtx); 1088 1089 switch (psopr->psopr_sopreq) { 1090 case PUFFS_SOPREQSYS_EXIT: 1091 keeprunning = false; 1092 break; 1093 case PUFFS_SOPREQ_FLUSH: 1094 puffsop_flush(pmp, &psopr->psopr_pf); 1095 break; 1096 case PUFFS_SOPREQ_EXPIRE: 1097 puffsop_expire(pmp, psopr->psopr_ck); 1098 break; 1099 case PUFFS_SOPREQ_UNMOUNT: 1100 puffs_msg_sendresp(pmp, &psopr->psopr_preq, 0); 1101 1102 unmountme = true; 1103 keeprunning = false; 1104 1105 /* 1106 * We know the mountpoint is still alive because 1107 * the thread that is us (poetic?) is still alive. 1108 */ 1109 vfs_ref(mp); 1110 break; 1111 } 1112 1113 kmem_free(psopr, sizeof(*psopr)); 1114 mutex_enter(&pmp->pmp_sopmtx); 1115 } 1116 1117 /* 1118 * Purge remaining ops. 1119 */ 1120 while ((psopr = TAILQ_FIRST(&pmp->pmp_sopfastreqs)) != NULL) { 1121 TAILQ_REMOVE(&pmp->pmp_sopfastreqs, psopr, psopr_entries); 1122 mutex_exit(&pmp->pmp_sopmtx); 1123 puffs_msg_sendresp(pmp, &psopr->psopr_preq, ENXIO); 1124 kmem_free(psopr, sizeof(*psopr)); 1125 mutex_enter(&pmp->pmp_sopmtx); 1126 } 1127 1128 while ((psopr = TAILQ_FIRST(&pmp->pmp_sopnodereqs)) != NULL) { 1129 TAILQ_REMOVE(&pmp->pmp_sopnodereqs, psopr, psopr_entries); 1130 mutex_exit(&pmp->pmp_sopmtx); 1131 KASSERT(psopr->psopr_sopreq == PUFFS_SOPREQ_EXPIRE); 1132 kmem_free(psopr, sizeof(*psopr)); 1133 mutex_enter(&pmp->pmp_sopmtx); 1134 } 1135 1136 pmp->pmp_sopthrcount--; 1137 cv_broadcast(&pmp->pmp_sopcv); 1138 mutex_exit(&pmp->pmp_sopmtx); /* not allowed to access fs after this */ 1139 1140 /* 1141 * If unmount was requested, we can now safely do it here, since 1142 * our context is dead from the point-of-view of puffs_unmount() 1143 * and we are just another thread. dounmount() makes internally 1144 * sure that VFS_UNMOUNT() isn't called reentrantly and that it 1145 * is eventually completed. 1146 */ 1147 if (unmountme) { 1148 (void)dounmount(mp, MNT_FORCE, curlwp); 1149 vfs_rele(mp); 1150 } 1151 1152 kthread_exit(0); 1153 } 1154 1155 int 1156 puffs_msgif_close(void *ctx) 1157 { 1158 struct puffs_mount *pmp = ctx; 1159 struct mount *mp = PMPTOMP(pmp); 1160 1161 mutex_enter(&pmp->pmp_lock); 1162 puffs_mp_reference(pmp); 1163 1164 /* 1165 * Free the waiting callers before proceeding any further. 1166 * The syncer might be jogging around in this file system 1167 * currently. If we allow it to go to the userspace of no 1168 * return while trying to get the syncer lock, well ... 1169 */ 1170 puffs_userdead(pmp); 1171 1172 /* 1173 * Make sure someone from puffs_unmount() isn't currently in 1174 * userspace. If we don't take this precautionary step, 1175 * they might notice that the mountpoint has disappeared 1176 * from under them once they return. Especially note that we 1177 * cannot simply test for an unmounter before calling 1178 * dounmount(), since it might be possible that that particular 1179 * invocation of unmount was called without MNT_FORCE. Here we 1180 * *must* make sure unmount succeeds. Also, restart is necessary 1181 * since pmp isn't locked. We might end up with PUTTER_DEAD after 1182 * restart and exit from there. 1183 */ 1184 if (pmp->pmp_unmounting) { 1185 cv_wait(&pmp->pmp_unmounting_cv, &pmp->pmp_lock); 1186 puffs_mp_release(pmp); 1187 mutex_exit(&pmp->pmp_lock); 1188 DPRINTF(("puffs_fop_close: unmount was in progress for pmp %p, " 1189 "restart\n", pmp)); 1190 return ERESTART; 1191 } 1192 1193 /* Won't access pmp from here anymore */ 1194 vfs_ref(mp); 1195 puffs_mp_release(pmp); 1196 mutex_exit(&pmp->pmp_lock); 1197 1198 /* Detach from VFS. */ 1199 (void)dounmount(mp, MNT_FORCE, curlwp); 1200 vfs_rele(mp); 1201 1202 return 0; 1203 } 1204 1205 /* 1206 * We're dead, kaput, RIP, slightly more than merely pining for the 1207 * fjords, belly-up, fallen, lifeless, finished, expired, gone to meet 1208 * our maker, ceased to be, etcetc. YASD. It's a dead FS! 1209 * 1210 * Caller must hold puffs mutex. 1211 */ 1212 void 1213 puffs_userdead(struct puffs_mount *pmp) 1214 { 1215 struct puffs_msgpark *park, *park_next; 1216 1217 /* 1218 * Mark filesystem status as dying so that operations don't 1219 * attempt to march to userspace any longer. 1220 */ 1221 pmp->pmp_status = PUFFSTAT_DYING; 1222 1223 /* signal waiters on REQUEST TO file server queue */ 1224 for (park = TAILQ_FIRST(&pmp->pmp_msg_touser); park; park = park_next) { 1225 1226 mutex_enter(&park->park_mtx); 1227 puffs_msgpark_reference(park); 1228 park_next = TAILQ_NEXT(park, park_entries); 1229 1230 KASSERT(park->park_flags & PARKFLAG_ONQUEUE1); 1231 TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries); 1232 park->park_flags &= ~PARKFLAG_ONQUEUE1; 1233 pmp->pmp_msg_touser_count--; 1234 1235 /* 1236 * Even though waiters on QUEUE1 are removed in touser() 1237 * in case of WAITERGONE, it is still possible for us to 1238 * get raced here due to having to retake locks in said 1239 * touser(). In the race case simply "ignore" the item 1240 * on the queue and move on to the next one. 1241 */ 1242 if (park->park_flags & PARKFLAG_WAITERGONE) { 1243 KASSERT((park->park_flags & PARKFLAG_CALL) == 0); 1244 KASSERT(park->park_flags & PARKFLAG_WANTREPLY); 1245 puffs_msgpark_release(park); 1246 1247 } else { 1248 park->park_preq->preq_rv = ENXIO; 1249 1250 if (park->park_flags & PARKFLAG_CALL) { 1251 park->park_done(pmp, park->park_preq, 1252 park->park_donearg); 1253 puffs_msgpark_release1(park, 2); 1254 } else if ((park->park_flags & PARKFLAG_WANTREPLY)==0) { 1255 puffs_msgpark_release1(park, 2); 1256 } else { 1257 park->park_preq->preq_rv = ENXIO; 1258 cv_signal(&park->park_cv); 1259 puffs_msgpark_release(park); 1260 } 1261 } 1262 } 1263 1264 /* signal waiters on RESPONSE FROM file server queue */ 1265 for (park=TAILQ_FIRST(&pmp->pmp_msg_replywait); park; park=park_next) { 1266 mutex_enter(&park->park_mtx); 1267 puffs_msgpark_reference(park); 1268 park_next = TAILQ_NEXT(park, park_entries); 1269 1270 KASSERT(park->park_flags & PARKFLAG_ONQUEUE2); 1271 KASSERT(park->park_flags & PARKFLAG_WANTREPLY); 1272 1273 TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries); 1274 park->park_flags &= ~PARKFLAG_ONQUEUE2; 1275 1276 if (park->park_flags & PARKFLAG_WAITERGONE) { 1277 KASSERT((park->park_flags & PARKFLAG_CALL) == 0); 1278 puffs_msgpark_release(park); 1279 } else { 1280 park->park_preq->preq_rv = ENXIO; 1281 if (park->park_flags & PARKFLAG_CALL) { 1282 park->park_done(pmp, park->park_preq, 1283 park->park_donearg); 1284 puffs_msgpark_release1(park, 2); 1285 } else { 1286 cv_signal(&park->park_cv); 1287 puffs_msgpark_release(park); 1288 } 1289 } 1290 } 1291 1292 cv_broadcast(&pmp->pmp_msg_waiter_cv); 1293 } 1294
Indexes created Mon Sep 22 21:09:42 GMT 2025