puffs_msgif.c revision 1.101.10.7
1 /* $NetBSD: puffs_msgif.c,v 1.101.10.7 2019/01/13 10:49:50 pgoyette 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 1.101.10.7 2019/01/13 10:49:50 pgoyette 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 /* Routines to call the compat hooks */ 327 /* Out-going */ 328 MODULE_CALL_HOOK_DECL(puffs_50_out_hook, 329 (struct puffs_req *oreq, struct puffs_req **creqp, ssize_t *deltap)); 330 MODULE_CALL_HOOK(puffs_50_out_hook, 331 (struct puffs_req *oreq, struct puffs_req **creqp, ssize_t *deltap), 332 (oreq, creqp, deltap), enosys()); 333 334 /* Incoming */ 335 MODULE_CALL_HOOK_DECL(puffs_50_in_hook, 336 (struct puffs_req *oreq, struct puffs_req *creqp)); 337 MODULE_CALL_HOOK(puffs_50_in_hook, 338 (struct puffs_req *oreq, struct puffs_req *creqp), 339 (oreq, creqp), enosys()); 340 341 /* 342 * A word about reference counting of parks. A reference must be taken 343 * when accessing a park and additionally when it is on a queue. So 344 * when taking it off a queue and releasing the access reference, the 345 * reference count is generally decremented by 2. 346 */ 347 348 void 349 puffs_msg_enqueue(struct puffs_mount *pmp, struct puffs_msgpark *park) 350 { 351 struct lwp *l = curlwp; 352 struct puffs_req *preq, *creq; 353 ssize_t delta; 354 355 /* 356 * Some clients reuse a park, so reset some flags. We might 357 * want to provide a caller-side interface for this and add 358 * a few more invariant checks here, but this will do for now. 359 */ 360 park->park_flags &= ~(PARKFLAG_DONE | PARKFLAG_HASERROR); 361 KASSERT((park->park_flags & PARKFLAG_WAITERGONE) == 0); 362 363 preq = park->park_preq; 364 365 #if 1 366 /* check if we do compat adjustments */ 367 if (pmp->pmp_docompat && 368 puffs_50_out_hook_call(preq, &creq, &delta) == 0) { 369 park->park_creq = park->park_preq; 370 park->park_creqlen = park->park_maxlen; 371 372 park->park_maxlen += delta; 373 park->park_copylen += delta; 374 park->park_preq = preq = creq; 375 } 376 #endif 377 378 preq->preq_buflen = park->park_maxlen; 379 KASSERT(preq->preq_id == 0 380 || (preq->preq_opclass & PUFFSOPFLAG_ISRESPONSE)); 381 382 if ((park->park_flags & PARKFLAG_WANTREPLY) == 0) 383 preq->preq_opclass |= PUFFSOPFLAG_FAF; 384 else 385 preq->preq_id = puffs_getmsgid(pmp); 386 387 /* fill in caller information */ 388 preq->preq_pid = l->l_proc->p_pid; 389 preq->preq_lid = l->l_lid; 390 391 /* 392 * To support cv_sig, yet another movie: check if there are signals 393 * pending and we are issueing a non-FAF. If so, return an error 394 * directly UNLESS we are issueing INACTIVE/RECLAIM. In that case, 395 * convert it to a FAF, fire off to the file server and return 396 * an error. Yes, this is bordering disgusting. Barfbags are on me. 397 */ 398 if (__predict_false((park->park_flags & PARKFLAG_WANTREPLY) 399 && (park->park_flags & PARKFLAG_CALL) == 0 400 && (l->l_flag & LW_PENDSIG) != 0 && sigispending(l, 0))) { 401 sigset_t ss; 402 403 /* 404 * see the comment about signals in puffs_msg_wait. 405 */ 406 sigpending1(l, &ss); 407 if (sigismember(&ss, SIGINT) || 408 sigismember(&ss, SIGTERM) || 409 sigismember(&ss, SIGKILL) || 410 sigismember(&ss, SIGHUP) || 411 sigismember(&ss, SIGQUIT)) { 412 park->park_flags |= PARKFLAG_HASERROR; 413 preq->preq_rv = EINTR; 414 if (PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN 415 && (preq->preq_optype == PUFFS_VN_INACTIVE 416 || preq->preq_optype == PUFFS_VN_RECLAIM)) { 417 park->park_preq->preq_opclass |= 418 PUFFSOPFLAG_FAF; 419 park->park_flags &= ~PARKFLAG_WANTREPLY; 420 DPRINTF(("puffs_msg_enqueue: " 421 "converted to FAF %p\n", park)); 422 } else { 423 return; 424 } 425 } 426 } 427 428 mutex_enter(&pmp->pmp_lock); 429 if (pmp->pmp_status != PUFFSTAT_RUNNING) { 430 mutex_exit(&pmp->pmp_lock); 431 park->park_flags |= PARKFLAG_HASERROR; 432 preq->preq_rv = ENXIO; 433 return; 434 } 435 436 #ifdef PUFFSDEBUG 437 parkqdump(&pmp->pmp_msg_touser, puffsdebug > 1); 438 parkqdump(&pmp->pmp_msg_replywait, puffsdebug > 1); 439 #endif 440 441 /* 442 * Note: we don't need to lock park since we have the only 443 * reference to it at this point. 444 */ 445 TAILQ_INSERT_TAIL(&pmp->pmp_msg_touser, park, park_entries); 446 park->park_flags |= PARKFLAG_ONQUEUE1; 447 pmp->pmp_msg_touser_count++; 448 park->park_refcount++; 449 450 cv_broadcast(&pmp->pmp_msg_waiter_cv); 451 mutex_exit(&pmp->pmp_lock); 452 putter_notify(pmp->pmp_pi); 453 454 DPRINTF(("touser: req %" PRIu64 ", preq: %p, park: %p, " 455 "c/t: 0x%x/0x%x, f: 0x%x\n", preq->preq_id, preq, park, 456 preq->preq_opclass, preq->preq_optype, park->park_flags)); 457 } 458 459 int 460 puffs_msg_wait(struct puffs_mount *pmp, struct puffs_msgpark *park) 461 { 462 lwp_t *l = curlwp; 463 proc_t *p = l->l_proc; 464 struct puffs_req *preq = park->park_preq; /* XXX: hmmm */ 465 sigset_t ss; 466 sigset_t oss; 467 int error = 0; 468 int rv; 469 470 /* 471 * block unimportant signals. 472 * 473 * The set of "important" signals here was chosen to be same as 474 * nfs interruptible mount. 475 */ 476 sigfillset(&ss); 477 sigdelset(&ss, SIGINT); 478 sigdelset(&ss, SIGTERM); 479 sigdelset(&ss, SIGKILL); 480 sigdelset(&ss, SIGHUP); 481 sigdelset(&ss, SIGQUIT); 482 mutex_enter(p->p_lock); 483 sigprocmask1(l, SIG_BLOCK, &ss, &oss); 484 mutex_exit(p->p_lock); 485 486 mutex_enter(&pmp->pmp_lock); 487 puffs_mp_reference(pmp); 488 mutex_exit(&pmp->pmp_lock); 489 490 mutex_enter(&park->park_mtx); 491 /* did the response beat us to the wait? */ 492 if (__predict_false((park->park_flags & PARKFLAG_DONE) 493 || (park->park_flags & PARKFLAG_HASERROR))) { 494 rv = park->park_preq->preq_rv; 495 mutex_exit(&park->park_mtx); 496 goto skipwait; 497 } 498 499 if ((park->park_flags & PARKFLAG_WANTREPLY) == 0 500 || (park->park_flags & PARKFLAG_CALL)) { 501 mutex_exit(&park->park_mtx); 502 rv = 0; 503 goto skipwait; 504 } 505 506 error = cv_wait_sig(&park->park_cv, &park->park_mtx); 507 DPRINTF(("puffs_touser: waiter for %p woke up with %d\n", 508 park, error)); 509 if (error) { 510 park->park_flags |= PARKFLAG_WAITERGONE; 511 if (park->park_flags & PARKFLAG_DONE) { 512 rv = preq->preq_rv; 513 mutex_exit(&park->park_mtx); 514 } else { 515 /* 516 * ok, we marked it as going away, but 517 * still need to do queue ops. take locks 518 * in correct order. 519 * 520 * We don't want to release our reference 521 * if it's on replywait queue to avoid error 522 * to file server. putop() code will DTRT. 523 */ 524 mutex_exit(&park->park_mtx); 525 mutex_enter(&pmp->pmp_lock); 526 mutex_enter(&park->park_mtx); 527 528 /* 529 * Still on queue1? We can safely remove it 530 * without any consequences since the file 531 * server hasn't seen it. "else" we need to 532 * wait for the response and just ignore it 533 * to avoid signalling an incorrect error to 534 * the file server. 535 */ 536 if (park->park_flags & PARKFLAG_ONQUEUE1) { 537 TAILQ_REMOVE(&pmp->pmp_msg_touser, 538 park, park_entries); 539 puffs_msgpark_release(park); 540 pmp->pmp_msg_touser_count--; 541 park->park_flags &= ~PARKFLAG_ONQUEUE1; 542 } else { 543 mutex_exit(&park->park_mtx); 544 } 545 mutex_exit(&pmp->pmp_lock); 546 547 rv = EINTR; 548 } 549 } else { 550 rv = preq->preq_rv; 551 mutex_exit(&park->park_mtx); 552 } 553 554 skipwait: 555 mutex_enter(&pmp->pmp_lock); 556 puffs_mp_release(pmp); 557 mutex_exit(&pmp->pmp_lock); 558 559 mutex_enter(p->p_lock); 560 sigprocmask1(l, SIG_SETMASK, &oss, NULL); 561 mutex_exit(p->p_lock); 562 563 return rv; 564 } 565 566 /* 567 * XXX: this suuuucks. Hopefully I'll get rid of this lossage once 568 * the whole setback-nonsense gets fixed. 569 */ 570 int 571 puffs_msg_wait2(struct puffs_mount *pmp, struct puffs_msgpark *park, 572 struct puffs_node *pn1, struct puffs_node *pn2) 573 { 574 struct puffs_req *preq; 575 int rv; 576 577 rv = puffs_msg_wait(pmp, park); 578 579 preq = park->park_preq; 580 if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N1) 581 pn1->pn_stat |= PNODE_DOINACT; 582 if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N2) 583 pn2->pn_stat |= PNODE_DOINACT; 584 585 if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N1) 586 pn1->pn_stat |= PNODE_NOREFS; 587 if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N2) 588 pn2->pn_stat |= PNODE_NOREFS; 589 590 return rv; 591 592 } 593 594 /* 595 * XXX: lazy bum. please, for the love of foie gras, fix me. 596 * This should *NOT* depend on setfaf. Also "memcpy" could 597 * be done more nicely. 598 */ 599 void 600 puffs_msg_sendresp(struct puffs_mount *pmp, struct puffs_req *origpreq, int rv) 601 { 602 struct puffs_msgpark *park; 603 struct puffs_req *preq; 604 605 puffs_msgmem_alloc(sizeof(struct puffs_req), &park, (void *)&preq, 1); 606 puffs_msg_setfaf(park); /* XXXXXX: avoids reqid override */ 607 608 memcpy(preq, origpreq, sizeof(struct puffs_req)); 609 preq->preq_rv = rv; 610 preq->preq_opclass |= PUFFSOPFLAG_ISRESPONSE; 611 612 puffs_msg_enqueue(pmp, park); 613 puffs_msgmem_release(park); 614 } 615 616 /* 617 * Get next request in the outgoing queue. "maxsize" controls the 618 * size the caller can accommodate and "nonblock" signals if this 619 * should block while waiting for input. Handles all locking internally. 620 */ 621 int 622 puffs_msgif_getout(void *ctx, size_t maxsize, int nonblock, 623 uint8_t **data, size_t *dlen, void **parkptr) 624 { 625 struct puffs_mount *pmp = ctx; 626 struct puffs_msgpark *park = NULL; 627 struct puffs_req *preq = NULL; 628 int error; 629 630 error = 0; 631 mutex_enter(&pmp->pmp_lock); 632 puffs_mp_reference(pmp); 633 for (;;) { 634 /* RIP? */ 635 if (pmp->pmp_status != PUFFSTAT_RUNNING) { 636 error = ENXIO; 637 break; 638 } 639 640 /* need platinum yendorian express card? */ 641 if (TAILQ_EMPTY(&pmp->pmp_msg_touser)) { 642 DPRINTF(("puffs_getout: no outgoing op, ")); 643 if (nonblock) { 644 DPRINTF(("returning EWOULDBLOCK\n")); 645 error = EWOULDBLOCK; 646 break; 647 } 648 DPRINTF(("waiting ...\n")); 649 650 error = cv_wait_sig(&pmp->pmp_msg_waiter_cv, 651 &pmp->pmp_lock); 652 if (error) 653 break; 654 else 655 continue; 656 } 657 658 park = TAILQ_FIRST(&pmp->pmp_msg_touser); 659 if (park == NULL) 660 continue; 661 662 mutex_enter(&park->park_mtx); 663 puffs_msgpark_reference(park); 664 665 DPRINTF(("puffs_getout: found park at %p, ", park)); 666 667 /* If it's a goner, don't process any furher */ 668 if (park->park_flags & PARKFLAG_WAITERGONE) { 669 DPRINTF(("waitergone!\n")); 670 puffs_msgpark_release(park); 671 continue; 672 } 673 preq = park->park_preq; 674 675 #if 0 676 /* check size */ 677 /* 678 * XXX: this check is not valid for now, we don't know 679 * the size of the caller's input buffer. i.e. this 680 * will most likely go away 681 */ 682 if (maxsize < preq->preq_frhdr.pfr_len) { 683 DPRINTF(("buffer too small\n")); 684 puffs_msgpark_release(park); 685 error = E2BIG; 686 break; 687 } 688 #endif 689 690 DPRINTF(("returning\n")); 691 692 /* 693 * Ok, we found what we came for. Release it from the 694 * outgoing queue but do not unlock. We will unlock 695 * only after we "releaseout" it to avoid complications: 696 * otherwise it is (theoretically) possible for userland 697 * to race us into "put" before we have a change to put 698 * this baby on the receiving queue. 699 */ 700 TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries); 701 KASSERT(park->park_flags & PARKFLAG_ONQUEUE1); 702 park->park_flags &= ~PARKFLAG_ONQUEUE1; 703 mutex_exit(&park->park_mtx); 704 705 pmp->pmp_msg_touser_count--; 706 KASSERT(pmp->pmp_msg_touser_count >= 0); 707 708 break; 709 } 710 puffs_mp_release(pmp); 711 mutex_exit(&pmp->pmp_lock); 712 713 if (error == 0) { 714 *data = (uint8_t *)preq; 715 preq->preq_pth.pth_framelen = park->park_copylen; 716 *dlen = preq->preq_pth.pth_framelen; 717 *parkptr = park; 718 } 719 720 return error; 721 } 722 723 /* 724 * Release outgoing structure. Now, depending on the success of the 725 * outgoing send, it is either going onto the result waiting queue 726 * or the death chamber. 727 */ 728 void 729 puffs_msgif_releaseout(void *ctx, void *parkptr, int status) 730 { 731 struct puffs_mount *pmp = ctx; 732 struct puffs_msgpark *park = parkptr; 733 734 DPRINTF(("puffs_releaseout: returning park %p, errno %d: " , 735 park, status)); 736 mutex_enter(&pmp->pmp_lock); 737 mutex_enter(&park->park_mtx); 738 if (park->park_flags & PARKFLAG_WANTREPLY) { 739 if (status == 0) { 740 DPRINTF(("enqueue replywait\n")); 741 TAILQ_INSERT_TAIL(&pmp->pmp_msg_replywait, park, 742 park_entries); 743 park->park_flags |= PARKFLAG_ONQUEUE2; 744 } else { 745 DPRINTF(("error path!\n")); 746 park->park_preq->preq_rv = status; 747 park->park_flags |= PARKFLAG_DONE; 748 cv_signal(&park->park_cv); 749 } 750 puffs_msgpark_release(park); 751 } else { 752 DPRINTF(("release\n")); 753 puffs_msgpark_release1(park, 2); 754 } 755 mutex_exit(&pmp->pmp_lock); 756 } 757 758 size_t 759 puffs_msgif_waitcount(void *ctx) 760 { 761 struct puffs_mount *pmp = ctx; 762 size_t rv; 763 764 mutex_enter(&pmp->pmp_lock); 765 rv = pmp->pmp_msg_touser_count; 766 mutex_exit(&pmp->pmp_lock); 767 768 return rv; 769 } 770 771 /* 772 * XXX: locking with this one? 773 */ 774 static void 775 puffsop_msg(void *ctx, struct puffs_req *preq) 776 { 777 struct puffs_mount *pmp = ctx; 778 struct putter_hdr *pth = &preq->preq_pth; 779 struct puffs_msgpark *park; 780 int wgone; 781 782 mutex_enter(&pmp->pmp_lock); 783 784 /* Locate waiter */ 785 TAILQ_FOREACH(park, &pmp->pmp_msg_replywait, park_entries) { 786 if (park->park_preq->preq_id == preq->preq_id) 787 break; 788 } 789 if (park == NULL) { 790 DPRINTF(("puffsop_msg: no request: %" PRIu64 "\n", 791 preq->preq_id)); 792 mutex_exit(&pmp->pmp_lock); 793 return; /* XXX send error */ 794 } 795 796 mutex_enter(&park->park_mtx); 797 puffs_msgpark_reference(park); 798 if (pth->pth_framelen > park->park_maxlen) { 799 DPRINTF(("puffsop_msg: invalid buffer length: " 800 "%" PRIu64 " (req %" PRIu64 ", \n", pth->pth_framelen, 801 preq->preq_id)); 802 park->park_preq->preq_rv = EPROTO; 803 cv_signal(&park->park_cv); 804 puffs_msgpark_release1(park, 2); 805 mutex_exit(&pmp->pmp_lock); 806 return; /* XXX: error */ 807 } 808 wgone = park->park_flags & PARKFLAG_WAITERGONE; 809 810 KASSERT(park->park_flags & PARKFLAG_ONQUEUE2); 811 TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries); 812 park->park_flags &= ~PARKFLAG_ONQUEUE2; 813 mutex_exit(&pmp->pmp_lock); 814 815 if (wgone) { 816 DPRINTF(("puffsop_msg: bad service - waiter gone for " 817 "park %p\n", park)); 818 } else { 819 #if 1 820 if (park->park_creq) { 821 struct puffs_req *creq; 822 size_t csize; 823 824 KASSERT(pmp->pmp_docompat); 825 (void)puffs_50_in_hook_call(preq, 826 park->park_creq); 827 creq = park->park_creq; 828 csize = park->park_creqlen; 829 park->park_creq = park->park_preq; 830 park->park_creqlen = park->park_maxlen; 831 832 park->park_preq = creq; 833 park->park_maxlen = csize; 834 835 memcpy(park->park_creq, preq, pth->pth_framelen); 836 } else { 837 #endif 838 memcpy(park->park_preq, preq, pth->pth_framelen); 839 } 840 841 if (park->park_flags & PARKFLAG_CALL) { 842 DPRINTF(("puffsop_msg: call for %p, arg %p\n", 843 park->park_preq, park->park_donearg)); 844 park->park_done(pmp, preq, park->park_donearg); 845 } 846 } 847 848 if (!wgone) { 849 DPRINTF(("puffs_putop: flagging done for " 850 "park %p\n", park)); 851 cv_signal(&park->park_cv); 852 } 853 854 park->park_flags |= PARKFLAG_DONE; 855 puffs_msgpark_release1(park, 2); 856 } 857 858 /* 859 * Node expiry. We come here after an inactive on an unexpired node. 860 * The expiry has been queued and is done in sop thread. 861 */ 862 static void 863 puffsop_expire(struct puffs_mount *pmp, puffs_cookie_t cookie) 864 { 865 struct vnode *vp; 866 867 KASSERT(PUFFS_USE_FS_TTL(pmp)); 868 869 /* 870 * If it still exists and has no reference, 871 * vrele should cause it to be reclaimed. 872 * Otherwise, we have nothing to do. 873 */ 874 if (puffs_cookie2vnode(pmp, cookie, &vp) == 0) { 875 VPTOPP(vp)->pn_stat &= ~PNODE_SOPEXP; 876 vrele(vp); 877 } 878 879 return; 880 } 881 882 static void 883 puffsop_flush(struct puffs_mount *pmp, struct puffs_flush *pf) 884 { 885 struct vnode *vp; 886 voff_t offlo, offhi; 887 int rv, flags = 0; 888 889 KASSERT(pf->pf_req.preq_pth.pth_framelen == sizeof(struct puffs_flush)); 890 891 /* XXX: slurry */ 892 if (pf->pf_op == PUFFS_INVAL_NAMECACHE_ALL) { 893 cache_purgevfs(PMPTOMP(pmp)); 894 rv = 0; 895 goto out; 896 } 897 898 /* 899 * Get vnode, don't lock it. Namecache is protected by its own lock 900 * and we have a reference to protect against premature harvesting. 901 * 902 * The node we want here might be locked and the op is in 903 * userspace waiting for us to complete ==> deadlock. Another 904 * reason we need to eventually bump locking to userspace, as we 905 * will need to lock the node if we wish to do flushes. 906 */ 907 rv = puffs_cookie2vnode(pmp, pf->pf_cookie, &vp); 908 if (rv) { 909 if (rv == PUFFS_NOSUCHCOOKIE) 910 rv = ENOENT; 911 goto out; 912 } 913 914 switch (pf->pf_op) { 915 #if 0 916 /* not quite ready, yet */ 917 case PUFFS_INVAL_NAMECACHE_NODE: 918 struct componentname *pf_cn; 919 char *name; 920 /* get comfortab^Wcomponentname */ 921 pf_cn = kmem_alloc(componentname); 922 memset(pf_cn, 0, sizeof(struct componentname)); 923 break; 924 925 #endif 926 case PUFFS_INVAL_NAMECACHE_DIR: 927 if (vp->v_type != VDIR) { 928 rv = EINVAL; 929 break; 930 } 931 cache_purge1(vp, NULL, 0, PURGE_CHILDREN); 932 break; 933 934 case PUFFS_INVAL_PAGECACHE_NODE_RANGE: 935 flags = PGO_FREE; 936 /*FALLTHROUGH*/ 937 case PUFFS_FLUSH_PAGECACHE_NODE_RANGE: 938 if (flags == 0) 939 flags = PGO_CLEANIT; 940 941 if (pf->pf_end > vp->v_size || vp->v_type != VREG) { 942 rv = EINVAL; 943 break; 944 } 945 946 offlo = trunc_page(pf->pf_start); 947 offhi = round_page(pf->pf_end); 948 if (offhi != 0 && offlo >= offhi) { 949 rv = EINVAL; 950 break; 951 } 952 953 mutex_enter(vp->v_uobj.vmobjlock); 954 rv = VOP_PUTPAGES(vp, offlo, offhi, flags); 955 break; 956 957 default: 958 rv = EINVAL; 959 } 960 961 vrele(vp); 962 963 out: 964 puffs_msg_sendresp(pmp, &pf->pf_req, rv); 965 } 966 967 int 968 puffs_msgif_dispatch(void *ctx, struct putter_hdr *pth) 969 { 970 struct puffs_mount *pmp = ctx; 971 struct puffs_req *preq = (struct puffs_req *)pth; 972 struct puffs_sopreq *psopr; 973 974 if (pth->pth_framelen < sizeof(struct puffs_req)) { 975 puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */ 976 return 0; 977 } 978 979 switch (PUFFSOP_OPCLASS(preq->preq_opclass)) { 980 case PUFFSOP_VN: 981 case PUFFSOP_VFS: 982 DPRINTF(("dispatch: vn/vfs message 0x%x\n", preq->preq_optype)); 983 puffsop_msg(pmp, preq); 984 break; 985 986 case PUFFSOP_FLUSH: /* process in sop thread */ 987 { 988 struct puffs_flush *pf; 989 990 DPRINTF(("dispatch: flush 0x%x\n", preq->preq_optype)); 991 992 if (preq->preq_pth.pth_framelen != sizeof(struct puffs_flush)) { 993 puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */ 994 break; 995 } 996 pf = (struct puffs_flush *)preq; 997 998 KASSERT(curlwp != uvm.pagedaemon_lwp); 999 psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP); 1000 memcpy(&psopr->psopr_pf, pf, sizeof(*pf)); 1001 psopr->psopr_sopreq = PUFFS_SOPREQ_FLUSH; 1002 1003 mutex_enter(&pmp->pmp_sopmtx); 1004 if (pmp->pmp_sopthrcount == 0) { 1005 mutex_exit(&pmp->pmp_sopmtx); 1006 kmem_free(psopr, sizeof(*psopr)); 1007 puffs_msg_sendresp(pmp, preq, ENXIO); 1008 } else { 1009 TAILQ_INSERT_TAIL(&pmp->pmp_sopfastreqs, 1010 psopr, psopr_entries); 1011 cv_signal(&pmp->pmp_sopcv); 1012 mutex_exit(&pmp->pmp_sopmtx); 1013 } 1014 break; 1015 } 1016 1017 case PUFFSOP_UNMOUNT: /* process in sop thread */ 1018 { 1019 1020 DPRINTF(("dispatch: unmount 0x%x\n", preq->preq_optype)); 1021 1022 KASSERT(curlwp != uvm.pagedaemon_lwp); 1023 psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP); 1024 psopr->psopr_preq = *preq; 1025 psopr->psopr_sopreq = PUFFS_SOPREQ_UNMOUNT; 1026 1027 mutex_enter(&pmp->pmp_sopmtx); 1028 if (pmp->pmp_sopthrcount == 0) { 1029 mutex_exit(&pmp->pmp_sopmtx); 1030 kmem_free(psopr, sizeof(*psopr)); 1031 puffs_msg_sendresp(pmp, preq, ENXIO); 1032 } else { 1033 TAILQ_INSERT_TAIL(&pmp->pmp_sopfastreqs, 1034 psopr, psopr_entries); 1035 cv_signal(&pmp->pmp_sopcv); 1036 mutex_exit(&pmp->pmp_sopmtx); 1037 } 1038 break; 1039 } 1040 1041 default: 1042 DPRINTF(("dispatch: invalid opclass 0x%x\n", preq->preq_opclass)); 1043 puffs_msg_sendresp(pmp, preq, EOPNOTSUPP); 1044 break; 1045 } 1046 1047 return 0; 1048 } 1049 1050 /* 1051 * Work loop for thread processing all ops from server which 1052 * cannot safely be handled in caller context. This includes 1053 * everything which might need a lock currently "held" by the file 1054 * server, i.e. a long-term kernel lock which will be released only 1055 * once the file server acknowledges a request 1056 */ 1057 #define TIMED_OUT(expire) \ 1058 ((int)((unsigned int)hardclock_ticks - (unsigned int)expire) > 0) 1059 void 1060 puffs_sop_thread(void *arg) 1061 { 1062 struct puffs_mount *pmp = arg; 1063 struct mount *mp = PMPTOMP(pmp); 1064 struct puffs_sopreq *psopr; 1065 bool keeprunning; 1066 bool unmountme = false; 1067 int timeo; 1068 1069 timeo = PUFFS_USE_FS_TTL(pmp) ? puffs_sopreq_expire_timeout : 0; 1070 1071 mutex_enter(&pmp->pmp_sopmtx); 1072 for (keeprunning = true; keeprunning; ) { 1073 /* 1074 * We have a fast queue for flush and umount, and a node 1075 * queue for delayes node reclaims. Requests on node queue * are not honoured before clock reaches psopr_at. This 1076 * code assumes that requests are ordered by psopr_at. 1077 */ 1078 do { 1079 psopr = TAILQ_FIRST(&pmp->pmp_sopfastreqs); 1080 if (psopr != NULL) { 1081 TAILQ_REMOVE(&pmp->pmp_sopfastreqs, 1082 psopr, psopr_entries); 1083 break; 1084 } 1085 1086 psopr = TAILQ_FIRST(&pmp->pmp_sopnodereqs); 1087 if ((psopr != NULL) && TIMED_OUT(psopr->psopr_at)) { 1088 TAILQ_REMOVE(&pmp->pmp_sopnodereqs, 1089 psopr, psopr_entries); 1090 break; 1091 } 1092 1093 cv_timedwait(&pmp->pmp_sopcv, &pmp->pmp_sopmtx, timeo); 1094 } while (1 /* CONSTCOND */); 1095 1096 mutex_exit(&pmp->pmp_sopmtx); 1097 1098 switch (psopr->psopr_sopreq) { 1099 case PUFFS_SOPREQSYS_EXIT: 1100 keeprunning = false; 1101 break; 1102 case PUFFS_SOPREQ_FLUSH: 1103 puffsop_flush(pmp, &psopr->psopr_pf); 1104 break; 1105 case PUFFS_SOPREQ_EXPIRE: 1106 puffsop_expire(pmp, psopr->psopr_ck); 1107 break; 1108 case PUFFS_SOPREQ_UNMOUNT: 1109 puffs_msg_sendresp(pmp, &psopr->psopr_preq, 0); 1110 1111 unmountme = true; 1112 keeprunning = false; 1113 1114 /* 1115 * We know the mountpoint is still alive because 1116 * the thread that is us (poetic?) is still alive. 1117 */ 1118 vfs_ref(mp); 1119 break; 1120 } 1121 1122 kmem_free(psopr, sizeof(*psopr)); 1123 mutex_enter(&pmp->pmp_sopmtx); 1124 } 1125 1126 /* 1127 * Purge remaining ops. 1128 */ 1129 while ((psopr = TAILQ_FIRST(&pmp->pmp_sopfastreqs)) != NULL) { 1130 TAILQ_REMOVE(&pmp->pmp_sopfastreqs, psopr, psopr_entries); 1131 mutex_exit(&pmp->pmp_sopmtx); 1132 puffs_msg_sendresp(pmp, &psopr->psopr_preq, ENXIO); 1133 kmem_free(psopr, sizeof(*psopr)); 1134 mutex_enter(&pmp->pmp_sopmtx); 1135 } 1136 1137 while ((psopr = TAILQ_FIRST(&pmp->pmp_sopnodereqs)) != NULL) { 1138 TAILQ_REMOVE(&pmp->pmp_sopnodereqs, psopr, psopr_entries); 1139 mutex_exit(&pmp->pmp_sopmtx); 1140 KASSERT(psopr->psopr_sopreq == PUFFS_SOPREQ_EXPIRE); 1141 kmem_free(psopr, sizeof(*psopr)); 1142 mutex_enter(&pmp->pmp_sopmtx); 1143 } 1144 1145 pmp->pmp_sopthrcount--; 1146 cv_broadcast(&pmp->pmp_sopcv); 1147 mutex_exit(&pmp->pmp_sopmtx); /* not allowed to access fs after this */ 1148 1149 /* 1150 * If unmount was requested, we can now safely do it here, since 1151 * our context is dead from the point-of-view of puffs_unmount() 1152 * and we are just another thread. dounmount() makes internally 1153 * sure that VFS_UNMOUNT() isn't called reentrantly and that it 1154 * is eventually completed. 1155 */ 1156 if (unmountme) { 1157 (void)dounmount(mp, MNT_FORCE, curlwp); 1158 vfs_rele(mp); 1159 } 1160 1161 kthread_exit(0); 1162 } 1163 1164 int 1165 puffs_msgif_close(void *ctx) 1166 { 1167 struct puffs_mount *pmp = ctx; 1168 struct mount *mp = PMPTOMP(pmp); 1169 1170 mutex_enter(&pmp->pmp_lock); 1171 puffs_mp_reference(pmp); 1172 1173 /* 1174 * Free the waiting callers before proceeding any further. 1175 * The syncer might be jogging around in this file system 1176 * currently. If we allow it to go to the userspace of no 1177 * return while trying to get the syncer lock, well ... 1178 */ 1179 puffs_userdead(pmp); 1180 1181 /* 1182 * Make sure someone from puffs_unmount() isn't currently in 1183 * userspace. If we don't take this precautionary step, 1184 * they might notice that the mountpoint has disappeared 1185 * from under them once they return. Especially note that we 1186 * cannot simply test for an unmounter before calling 1187 * dounmount(), since it might be possible that that particular 1188 * invocation of unmount was called without MNT_FORCE. Here we 1189 * *must* make sure unmount succeeds. Also, restart is necessary 1190 * since pmp isn't locked. We might end up with PUTTER_DEAD after 1191 * restart and exit from there. 1192 */ 1193 if (pmp->pmp_unmounting) { 1194 cv_wait(&pmp->pmp_unmounting_cv, &pmp->pmp_lock); 1195 puffs_mp_release(pmp); 1196 mutex_exit(&pmp->pmp_lock); 1197 DPRINTF(("puffs_fop_close: unmount was in progress for pmp %p, " 1198 "restart\n", pmp)); 1199 return ERESTART; 1200 } 1201 1202 /* Won't access pmp from here anymore */ 1203 vfs_ref(mp); 1204 puffs_mp_release(pmp); 1205 mutex_exit(&pmp->pmp_lock); 1206 1207 /* Detach from VFS. */ 1208 (void)dounmount(mp, MNT_FORCE, curlwp); 1209 vfs_rele(mp); 1210 1211 return 0; 1212 } 1213 1214 /* 1215 * We're dead, kaput, RIP, slightly more than merely pining for the 1216 * fjords, belly-up, fallen, lifeless, finished, expired, gone to meet 1217 * our maker, ceased to be, etcetc. YASD. It's a dead FS! 1218 * 1219 * Caller must hold puffs mutex. 1220 */ 1221 void 1222 puffs_userdead(struct puffs_mount *pmp) 1223 { 1224 struct puffs_msgpark *park, *park_next; 1225 1226 /* 1227 * Mark filesystem status as dying so that operations don't 1228 * attempt to march to userspace any longer. 1229 */ 1230 pmp->pmp_status = PUFFSTAT_DYING; 1231 1232 /* signal waiters on REQUEST TO file server queue */ 1233 for (park = TAILQ_FIRST(&pmp->pmp_msg_touser); park; park = park_next) { 1234 1235 mutex_enter(&park->park_mtx); 1236 puffs_msgpark_reference(park); 1237 park_next = TAILQ_NEXT(park, park_entries); 1238 1239 KASSERT(park->park_flags & PARKFLAG_ONQUEUE1); 1240 TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries); 1241 park->park_flags &= ~PARKFLAG_ONQUEUE1; 1242 pmp->pmp_msg_touser_count--; 1243 1244 /* 1245 * Even though waiters on QUEUE1 are removed in touser() 1246 * in case of WAITERGONE, it is still possible for us to 1247 * get raced here due to having to retake locks in said 1248 * touser(). In the race case simply "ignore" the item 1249 * on the queue and move on to the next one. 1250 */ 1251 if (park->park_flags & PARKFLAG_WAITERGONE) { 1252 KASSERT((park->park_flags & PARKFLAG_CALL) == 0); 1253 KASSERT(park->park_flags & PARKFLAG_WANTREPLY); 1254 puffs_msgpark_release(park); 1255 1256 } else { 1257 park->park_preq->preq_rv = ENXIO; 1258 1259 if (park->park_flags & PARKFLAG_CALL) { 1260 park->park_done(pmp, park->park_preq, 1261 park->park_donearg); 1262 puffs_msgpark_release1(park, 2); 1263 } else if ((park->park_flags & PARKFLAG_WANTREPLY)==0) { 1264 puffs_msgpark_release1(park, 2); 1265 } else { 1266 park->park_preq->preq_rv = ENXIO; 1267 cv_signal(&park->park_cv); 1268 puffs_msgpark_release(park); 1269 } 1270 } 1271 } 1272 1273 /* signal waiters on RESPONSE FROM file server queue */ 1274 for (park=TAILQ_FIRST(&pmp->pmp_msg_replywait); park; park=park_next) { 1275 mutex_enter(&park->park_mtx); 1276 puffs_msgpark_reference(park); 1277 park_next = TAILQ_NEXT(park, park_entries); 1278 1279 KASSERT(park->park_flags & PARKFLAG_ONQUEUE2); 1280 KASSERT(park->park_flags & PARKFLAG_WANTREPLY); 1281 1282 TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries); 1283 park->park_flags &= ~PARKFLAG_ONQUEUE2; 1284 1285 if (park->park_flags & PARKFLAG_WAITERGONE) { 1286 KASSERT((park->park_flags & PARKFLAG_CALL) == 0); 1287 puffs_msgpark_release(park); 1288 } else { 1289 park->park_preq->preq_rv = ENXIO; 1290 if (park->park_flags & PARKFLAG_CALL) { 1291 park->park_done(pmp, park->park_preq, 1292 park->park_donearg); 1293 puffs_msgpark_release1(park, 2); 1294 } else { 1295 cv_signal(&park->park_cv); 1296 puffs_msgpark_release(park); 1297 } 1298 } 1299 } 1300 1301 cv_broadcast(&pmp->pmp_msg_waiter_cv); 1302 } 1303
Indexes created Sat Oct 25 16:10:12 GMT 2025