1 /*- 2 * Copyright (c) 2013-2015 Gleb Smirnoff <glebius@FreeBSD.org> 3 * Copyright (c) 1998, David Greenman. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. Neither the name of the University nor the names of its contributors 14 * may be used to endorse or promote products derived from this software 15 * without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 __FBSDID("$FreeBSD$"); 32 33 #include "opt_kern_tls.h" 34 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/capsicum.h> 38 #include <sys/kernel.h> 39 #include <sys/lock.h> 40 #include <sys/ktls.h> 41 #include <sys/mutex.h> 42 #include <sys/malloc.h> 43 #include <sys/mman.h> 44 #include <sys/mount.h> 45 #include <sys/mbuf.h> 46 #include <sys/proc.h> 47 #include <sys/protosw.h> 48 #include <sys/rwlock.h> 49 #include <sys/sf_buf.h> 50 #include <sys/socket.h> 51 #include <sys/socketvar.h> 52 #include <sys/syscallsubr.h> 53 #include <sys/sysctl.h> 54 #include <sys/sysproto.h> 55 #include <sys/vnode.h> 56 57 #include <net/vnet.h> 58 #include <netinet/in.h> 59 #include <netinet/tcp.h> 60 61 #include <security/audit/audit.h> 62 #include <security/mac/mac_framework.h> 63 64 #include <vm/vm.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_pager.h> 67 68 static MALLOC_DEFINE(M_SENDFILE, "sendfile", "sendfile dynamic memory"); 69 70 #define EXT_FLAG_SYNC EXT_FLAG_VENDOR1 71 #define EXT_FLAG_NOCACHE EXT_FLAG_VENDOR2 72 #define EXT_FLAG_CACHE_LAST EXT_FLAG_VENDOR3 73 74 /* 75 * Structure describing a single sendfile(2) I/O, which may consist of 76 * several underlying pager I/Os. 77 * 78 * The syscall context allocates the structure and initializes 'nios' 79 * to 1. As sendfile_swapin() runs through pages and starts asynchronous 80 * paging operations, it increments 'nios'. 81 * 82 * Every I/O completion calls sendfile_iodone(), which decrements the 'nios', 83 * and the syscall also calls sendfile_iodone() after allocating all mbufs, 84 * linking them and sending to socket. Whoever reaches zero 'nios' is 85 * responsible to * call pru_ready on the socket, to notify it of readyness 86 * of the data. 87 */ 88 struct sf_io { 89 volatile u_int nios; 90 u_int error; 91 int npages; 92 struct socket *so; 93 struct mbuf *m; 94 vm_object_t obj; 95 vm_pindex_t pindex0; 96 #ifdef KERN_TLS 97 struct ktls_session *tls; 98 #endif 99 vm_page_t pa[]; 100 }; 101 102 /* 103 * Structure used to track requests with SF_SYNC flag. 104 */ 105 struct sendfile_sync { 106 struct mtx mtx; 107 struct cv cv; 108 unsigned count; 109 bool waiting; 110 }; 111 112 static void 113 sendfile_sync_destroy(struct sendfile_sync *sfs) 114 { 115 KASSERT(sfs->count == 0, ("sendfile sync %p still busy", sfs)); 116 117 cv_destroy(&sfs->cv); 118 mtx_destroy(&sfs->mtx); 119 free(sfs, M_SENDFILE); 120 } 121 122 static void 123 sendfile_sync_signal(struct sendfile_sync *sfs) 124 { 125 mtx_lock(&sfs->mtx); 126 KASSERT(sfs->count > 0, ("sendfile sync %p not busy", sfs)); 127 if (--sfs->count == 0) { 128 if (!sfs->waiting) { 129 /* The sendfile() waiter was interrupted by a signal. */ 130 sendfile_sync_destroy(sfs); 131 return; 132 } else { 133 cv_signal(&sfs->cv); 134 } 135 } 136 mtx_unlock(&sfs->mtx); 137 } 138 139 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)]; 140 141 static void 142 sfstat_init(const void *unused) 143 { 144 145 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t), 146 M_WAITOK); 147 } 148 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL); 149 150 static int 151 sfstat_sysctl(SYSCTL_HANDLER_ARGS) 152 { 153 struct sfstat s; 154 155 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t)); 156 if (req->newptr) 157 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t)); 158 return (SYSCTL_OUT(req, &s, sizeof(s))); 159 } 160 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, 161 CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0, 162 sfstat_sysctl, "I", 163 "sendfile statistics"); 164 165 static void 166 sendfile_free_mext(struct mbuf *m) 167 { 168 struct sf_buf *sf; 169 vm_page_t pg; 170 int flags; 171 172 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF, 173 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m)); 174 175 sf = m->m_ext.ext_arg1; 176 pg = sf_buf_page(sf); 177 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0; 178 179 sf_buf_free(sf); 180 vm_page_release(pg, flags); 181 182 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) { 183 struct sendfile_sync *sfs = m->m_ext.ext_arg2; 184 sendfile_sync_signal(sfs); 185 } 186 } 187 188 static void 189 sendfile_free_mext_pg(struct mbuf *m) 190 { 191 vm_page_t pg; 192 int flags, i; 193 bool cache_last; 194 195 M_ASSERTEXTPG(m); 196 197 cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST; 198 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0; 199 200 for (i = 0; i < m->m_epg_npgs; i++) { 201 if (cache_last && i == m->m_epg_npgs - 1) 202 flags = 0; 203 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); 204 vm_page_release(pg, flags); 205 } 206 207 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) { 208 struct sendfile_sync *sfs = m->m_ext.ext_arg1; 209 sendfile_sync_signal(sfs); 210 } 211 } 212 213 /* 214 * Helper function to calculate how much data to put into page i of n. 215 * Only first and last pages are special. 216 */ 217 static inline off_t 218 xfsize(int i, int n, off_t off, off_t len) 219 { 220 221 if (i == 0) 222 return (omin(PAGE_SIZE - (off & PAGE_MASK), len)); 223 224 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0) 225 return ((off + len) & PAGE_MASK); 226 227 return (PAGE_SIZE); 228 } 229 230 /* 231 * Helper function to get offset within object for i page. 232 */ 233 static inline vm_ooffset_t 234 vmoff(int i, off_t off) 235 { 236 237 if (i == 0) 238 return ((vm_ooffset_t)off); 239 240 return (trunc_page(off + i * PAGE_SIZE)); 241 } 242 243 /* 244 * Helper function used when allocation of a page or sf_buf failed. 245 * Pretend as if we don't have enough space, subtract xfsize() of 246 * all pages that failed. 247 */ 248 static inline void 249 fixspace(int old, int new, off_t off, int *space) 250 { 251 252 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new)); 253 254 /* Subtract last one. */ 255 *space -= xfsize(old - 1, old, off, *space); 256 old--; 257 258 if (new == old) 259 /* There was only one page. */ 260 return; 261 262 /* Subtract first one. */ 263 if (new == 0) { 264 *space -= xfsize(0, old, off, *space); 265 new++; 266 } 267 268 /* Rest of pages are full sized. */ 269 *space -= (old - new) * PAGE_SIZE; 270 271 KASSERT(*space >= 0, ("%s: space went backwards", __func__)); 272 } 273 274 /* 275 * Wait for all in-flight ios to complete, we must not unwire pages 276 * under them. 277 */ 278 static void 279 sendfile_iowait(struct sf_io *sfio, const char *wmesg) 280 { 281 while (atomic_load_int(&sfio->nios) != 1) 282 pause(wmesg, 1); 283 } 284 285 /* 286 * I/O completion callback. 287 */ 288 static void 289 sendfile_iodone(void *arg, vm_page_t *pa, int count, int error) 290 { 291 struct sf_io *sfio = arg; 292 struct socket *so; 293 int i; 294 295 if (error != 0) { 296 sfio->error = error; 297 /* 298 * Restore of the pg[] elements is done by 299 * sendfile_swapin(). 300 */ 301 } else { 302 /* 303 * Restore the valid page pointers. They are already 304 * unbusied, but still wired. For error != 0 case, 305 * sendfile_swapin() handles unbusy. 306 * 307 * XXXKIB since pages are only wired, and we do not 308 * own the object lock, other users might have 309 * invalidated them in meantime. Similarly, after we 310 * unbusied the swapped-in pages, they can become 311 * invalid under us. 312 */ 313 MPASS(count == 0 || pa[0] != bogus_page); 314 for (i = 0; i < count; i++) { 315 if (pa[i] == bogus_page) { 316 sfio->pa[(pa[0]->pindex - sfio->pindex0) + i] = 317 pa[i] = vm_page_relookup(sfio->obj, 318 pa[0]->pindex + i); 319 KASSERT(pa[i] != NULL, 320 ("%s: page %p[%d] disappeared", 321 __func__, pa, i)); 322 } else { 323 vm_page_xunbusy_unchecked(pa[i]); 324 } 325 } 326 } 327 328 if (!refcount_release(&sfio->nios)) 329 return; 330 331 #ifdef INVARIANTS 332 for (i = 1; i < sfio->npages; i++) { 333 if (sfio->pa[i] == NULL) 334 break; 335 KASSERT(vm_page_wired(sfio->pa[i]), 336 ("sfio %p page %d %p not wired", sfio, i, sfio->pa[i])); 337 if (i == 0) 338 continue; 339 KASSERT(sfio->pa[0]->object == sfio->pa[i]->object, 340 ("sfio %p page %d %p wrong owner %p %p", sfio, i, 341 sfio->pa[i], sfio->pa[0]->object, sfio->pa[i]->object)); 342 KASSERT(sfio->pa[0]->pindex + i == sfio->pa[i]->pindex, 343 ("sfio %p page %d %p wrong index %jx %jx", sfio, i, 344 sfio->pa[i], (uintmax_t)sfio->pa[0]->pindex, 345 (uintmax_t)sfio->pa[i]->pindex)); 346 } 347 #endif 348 349 vm_object_pip_wakeup(sfio->obj); 350 351 if (sfio->m == NULL) { 352 /* 353 * Either I/O operation failed, or we failed to allocate 354 * buffers, or we bailed out on first busy page, or we 355 * succeeded filling the request without any I/Os. Anyway, 356 * pru_send hadn't been executed - nothing had been sent 357 * to the socket yet. 358 */ 359 MPASS((curthread->td_pflags & TDP_KTHREAD) == 0); 360 free(sfio, M_SENDFILE); 361 return; 362 } 363 364 #if defined(KERN_TLS) && defined(INVARIANTS) 365 if ((sfio->m->m_flags & M_EXTPG) != 0) 366 KASSERT(sfio->tls == sfio->m->m_epg_tls, 367 ("TLS session mismatch")); 368 else 369 KASSERT(sfio->tls == NULL, 370 ("non-ext_pgs mbuf with TLS session")); 371 #endif 372 so = sfio->so; 373 CURVNET_SET(so->so_vnet); 374 if (__predict_false(sfio->error)) { 375 /* 376 * I/O operation failed. The state of data in the socket 377 * is now inconsistent, and all what we can do is to tear 378 * it down. Protocol abort method would tear down protocol 379 * state, free all ready mbufs and detach not ready ones. 380 * We will free the mbufs corresponding to this I/O manually. 381 * 382 * The socket would be marked with EIO and made available 383 * for read, so that application receives EIO on next 384 * syscall and eventually closes the socket. 385 */ 386 so->so_proto->pr_usrreqs->pru_abort(so); 387 so->so_error = EIO; 388 389 mb_free_notready(sfio->m, sfio->npages); 390 #ifdef KERN_TLS 391 } else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) { 392 /* 393 * I/O operation is complete, but we still need to 394 * encrypt. We cannot do this in the interrupt thread 395 * of the disk controller, so forward the mbufs to a 396 * different thread. 397 * 398 * Donate the socket reference from sfio to rather 399 * than explicitly invoking soref(). 400 */ 401 ktls_enqueue(sfio->m, so, sfio->npages); 402 goto out_with_ref; 403 #endif 404 } else 405 (void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m, 406 sfio->npages); 407 408 SOCK_LOCK(so); 409 sorele(so); 410 #ifdef KERN_TLS 411 out_with_ref: 412 #endif 413 CURVNET_RESTORE(); 414 free(sfio, M_SENDFILE); 415 } 416 417 /* 418 * Iterate through pages vector and request paging for non-valid pages. 419 */ 420 static int 421 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off, 422 off_t len, int npages, int rhpages, int flags) 423 { 424 vm_page_t *pa; 425 int a, count, count1, grabbed, i, j, rv; 426 427 pa = sfio->pa; 428 *nios = 0; 429 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0; 430 sfio->pindex0 = OFF_TO_IDX(off); 431 432 /* 433 * First grab all the pages and wire them. Note that we grab 434 * only required pages. Readahead pages are dealt with later. 435 */ 436 grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off), 437 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages); 438 if (grabbed < npages) { 439 for (int i = grabbed; i < npages; i++) 440 pa[i] = NULL; 441 npages = grabbed; 442 rhpages = 0; 443 } 444 445 for (i = 0; i < npages;) { 446 /* Skip valid pages. */ 447 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK, 448 xfsize(i, npages, off, len))) { 449 vm_page_xunbusy(pa[i]); 450 SFSTAT_INC(sf_pages_valid); 451 i++; 452 continue; 453 } 454 455 /* 456 * Next page is invalid. Check if it belongs to pager. It 457 * may not be there, which is a regular situation for shmem 458 * pager. For vnode pager this happens only in case of 459 * a sparse file. 460 * 461 * Important feature of vm_pager_has_page() is the hint 462 * stored in 'a', about how many pages we can pagein after 463 * this page in a single I/O. 464 */ 465 VM_OBJECT_RLOCK(obj); 466 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL, 467 &a)) { 468 VM_OBJECT_RUNLOCK(obj); 469 pmap_zero_page(pa[i]); 470 vm_page_valid(pa[i]); 471 MPASS(pa[i]->dirty == 0); 472 vm_page_xunbusy(pa[i]); 473 i++; 474 continue; 475 } 476 VM_OBJECT_RUNLOCK(obj); 477 478 /* 479 * We want to pagein as many pages as possible, limited only 480 * by the 'a' hint and actual request. 481 */ 482 count = min(a + 1, npages - i); 483 484 /* 485 * We should not pagein into a valid page because 486 * there might be still unfinished write tracked by 487 * e.g. a buffer, thus we substitute any valid pages 488 * with the bogus one. 489 * 490 * We must not leave around xbusy pages which are not 491 * part of the run passed to vm_pager_getpages(), 492 * otherwise pager might deadlock waiting for the busy 493 * status of the page, e.g. if it constitues the 494 * buffer needed to validate other page. 495 * 496 * First trim the end of the run consisting of the 497 * valid pages, then replace the rest of the valid 498 * with bogus. 499 */ 500 count1 = count; 501 for (j = i + count - 1; j > i; j--) { 502 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 503 xfsize(j, npages, off, len))) { 504 vm_page_xunbusy(pa[j]); 505 SFSTAT_INC(sf_pages_valid); 506 count--; 507 } else { 508 break; 509 } 510 } 511 512 /* 513 * The last page in the run pa[i + count - 1] is 514 * guaranteed to be invalid by the trim above, so it 515 * is not replaced with bogus, thus -1 in the loop end 516 * condition. 517 */ 518 MPASS(pa[i + count - 1]->valid != VM_PAGE_BITS_ALL); 519 for (j = i + 1; j < i + count - 1; j++) { 520 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 521 xfsize(j, npages, off, len))) { 522 vm_page_xunbusy(pa[j]); 523 SFSTAT_INC(sf_pages_valid); 524 SFSTAT_INC(sf_pages_bogus); 525 pa[j] = bogus_page; 526 } 527 } 528 529 refcount_acquire(&sfio->nios); 530 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL, 531 i + count == npages ? &rhpages : NULL, 532 &sendfile_iodone, sfio); 533 if (__predict_false(rv != VM_PAGER_OK)) { 534 sendfile_iowait(sfio, "sferrio"); 535 536 /* 537 * Perform full pages recovery before returning EIO. 538 * Pages from 0 to npages are wired. 539 * Pages from (i + 1) to (i + count - 1) may be 540 * substituted to bogus page, and not busied. 541 * Pages from (i + count) to (i + count1 - 1) are 542 * not busied. 543 * Rest of the pages from i to npages are busied. 544 */ 545 for (j = 0; j < npages; j++) { 546 if (j >= i + count && j < i + count1) 547 ; 548 else if (j > i && j < i + count - 1 && 549 pa[j] == bogus_page) 550 pa[j] = vm_page_relookup(obj, 551 OFF_TO_IDX(vmoff(j, off))); 552 else if (j >= i) 553 vm_page_xunbusy(pa[j]); 554 KASSERT(pa[j] != NULL && pa[j] != bogus_page, 555 ("%s: page %p[%d] I/O recovery failure", 556 __func__, pa, j)); 557 vm_page_unwire(pa[j], PQ_INACTIVE); 558 } 559 return (EIO); 560 } 561 562 SFSTAT_INC(sf_iocnt); 563 SFSTAT_ADD(sf_pages_read, count); 564 if (i + count == npages) 565 SFSTAT_ADD(sf_rhpages_read, rhpages); 566 567 i += count1; 568 (*nios)++; 569 } 570 571 if (*nios == 0 && npages != 0) 572 SFSTAT_INC(sf_noiocnt); 573 574 return (0); 575 } 576 577 static int 578 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res, 579 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size, 580 int *bsize) 581 { 582 struct vattr va; 583 vm_object_t obj; 584 struct vnode *vp; 585 struct shmfd *shmfd; 586 int error; 587 588 vp = *vp_res = NULL; 589 obj = NULL; 590 shmfd = *shmfd_res = NULL; 591 *bsize = 0; 592 593 /* 594 * The file descriptor must be a regular file and have a 595 * backing VM object. 596 */ 597 if (fp->f_type == DTYPE_VNODE) { 598 vp = fp->f_vnode; 599 vn_lock(vp, LK_SHARED | LK_RETRY); 600 if (vp->v_type != VREG) { 601 error = EINVAL; 602 goto out; 603 } 604 *bsize = vp->v_mount->mnt_stat.f_iosize; 605 error = VOP_GETATTR(vp, &va, td->td_ucred); 606 if (error != 0) 607 goto out; 608 *obj_size = va.va_size; 609 obj = vp->v_object; 610 if (obj == NULL) { 611 error = EINVAL; 612 goto out; 613 } 614 } else if (fp->f_type == DTYPE_SHM) { 615 error = 0; 616 shmfd = fp->f_data; 617 obj = shmfd->shm_object; 618 *obj_size = shmfd->shm_size; 619 } else { 620 error = EINVAL; 621 goto out; 622 } 623 624 VM_OBJECT_WLOCK(obj); 625 if ((obj->flags & OBJ_DEAD) != 0) { 626 VM_OBJECT_WUNLOCK(obj); 627 error = EBADF; 628 goto out; 629 } 630 631 /* 632 * Temporarily increase the backing VM object's reference 633 * count so that a forced reclamation of its vnode does not 634 * immediately destroy it. 635 */ 636 vm_object_reference_locked(obj); 637 VM_OBJECT_WUNLOCK(obj); 638 *obj_res = obj; 639 *vp_res = vp; 640 *shmfd_res = shmfd; 641 642 out: 643 if (vp != NULL) 644 VOP_UNLOCK(vp); 645 return (error); 646 } 647 648 static int 649 sendfile_getsock(struct thread *td, int s, struct file **sock_fp, 650 struct socket **so) 651 { 652 int error; 653 654 *sock_fp = NULL; 655 *so = NULL; 656 657 /* 658 * The socket must be a stream socket and connected. 659 */ 660 error = getsock_cap(td, s, &cap_send_rights, 661 sock_fp, NULL, NULL); 662 if (error != 0) 663 return (error); 664 *so = (*sock_fp)->f_data; 665 if ((*so)->so_type != SOCK_STREAM) 666 return (EINVAL); 667 /* 668 * SCTP one-to-one style sockets currently don't work with 669 * sendfile(). So indicate EINVAL for now. 670 */ 671 if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP) 672 return (EINVAL); 673 if (SOLISTENING(*so)) 674 return (ENOTCONN); 675 return (0); 676 } 677 678 int 679 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, 680 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags, 681 struct thread *td) 682 { 683 struct file *sock_fp; 684 struct vnode *vp; 685 struct vm_object *obj; 686 vm_page_t pga; 687 struct socket *so; 688 #ifdef KERN_TLS 689 struct ktls_session *tls; 690 #endif 691 struct mbuf *m, *mh, *mhtail; 692 struct sf_buf *sf; 693 struct shmfd *shmfd; 694 struct sendfile_sync *sfs; 695 struct vattr va; 696 off_t off, sbytes, rem, obj_size; 697 int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr; 698 #ifdef KERN_TLS 699 int tls_enq_cnt; 700 #endif 701 bool use_ext_pgs; 702 703 obj = NULL; 704 so = NULL; 705 m = mh = NULL; 706 sfs = NULL; 707 #ifdef KERN_TLS 708 tls = NULL; 709 #endif 710 hdrlen = sbytes = 0; 711 softerr = 0; 712 use_ext_pgs = false; 713 714 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize); 715 if (error != 0) 716 return (error); 717 718 error = sendfile_getsock(td, sockfd, &sock_fp, &so); 719 if (error != 0) 720 goto out; 721 722 #ifdef MAC 723 error = mac_socket_check_send(td->td_ucred, so); 724 if (error != 0) 725 goto out; 726 #endif 727 728 SFSTAT_INC(sf_syscalls); 729 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags)); 730 731 if (flags & SF_SYNC) { 732 sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO); 733 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF); 734 cv_init(&sfs->cv, "sendfile"); 735 sfs->waiting = true; 736 } 737 738 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset; 739 740 /* 741 * Protect against multiple writers to the socket. 742 * 743 * XXXRW: Historically this has assumed non-interruptibility, so now 744 * we implement that, but possibly shouldn't. 745 */ 746 (void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR); 747 #ifdef KERN_TLS 748 tls = ktls_hold(so->so_snd.sb_tls_info); 749 #endif 750 751 /* 752 * Loop through the pages of the file, starting with the requested 753 * offset. Get a file page (do I/O if necessary), map the file page 754 * into an sf_buf, attach an mbuf header to the sf_buf, and queue 755 * it on the socket. 756 * This is done in two loops. The inner loop turns as many pages 757 * as it can, up to available socket buffer space, without blocking 758 * into mbufs to have it bulk delivered into the socket send buffer. 759 * The outer loop checks the state and available space of the socket 760 * and takes care of the overall progress. 761 */ 762 for (off = offset; rem > 0; ) { 763 struct sf_io *sfio; 764 vm_page_t *pa; 765 struct mbuf *m0, *mtail; 766 int nios, space, npages, rhpages; 767 768 mtail = NULL; 769 /* 770 * Check the socket state for ongoing connection, 771 * no errors and space in socket buffer. 772 * If space is low allow for the remainder of the 773 * file to be processed if it fits the socket buffer. 774 * Otherwise block in waiting for sufficient space 775 * to proceed, or if the socket is nonblocking, return 776 * to userland with EAGAIN while reporting how far 777 * we've come. 778 * We wait until the socket buffer has significant free 779 * space to do bulk sends. This makes good use of file 780 * system read ahead and allows packet segmentation 781 * offloading hardware to take over lots of work. If 782 * we were not careful here we would send off only one 783 * sfbuf at a time. 784 */ 785 SOCKBUF_LOCK(&so->so_snd); 786 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2) 787 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2; 788 retry_space: 789 if (so->so_snd.sb_state & SBS_CANTSENDMORE) { 790 error = EPIPE; 791 SOCKBUF_UNLOCK(&so->so_snd); 792 goto done; 793 } else if (so->so_error) { 794 error = so->so_error; 795 so->so_error = 0; 796 SOCKBUF_UNLOCK(&so->so_snd); 797 goto done; 798 } 799 if ((so->so_state & SS_ISCONNECTED) == 0) { 800 SOCKBUF_UNLOCK(&so->so_snd); 801 error = ENOTCONN; 802 goto done; 803 } 804 805 space = sbspace(&so->so_snd); 806 if (space < rem && 807 (space <= 0 || 808 space < so->so_snd.sb_lowat)) { 809 if (so->so_state & SS_NBIO) { 810 SOCKBUF_UNLOCK(&so->so_snd); 811 error = EAGAIN; 812 goto done; 813 } 814 /* 815 * sbwait drops the lock while sleeping. 816 * When we loop back to retry_space the 817 * state may have changed and we retest 818 * for it. 819 */ 820 error = sbwait(&so->so_snd); 821 /* 822 * An error from sbwait usually indicates that we've 823 * been interrupted by a signal. If we've sent anything 824 * then return bytes sent, otherwise return the error. 825 */ 826 if (error != 0) { 827 SOCKBUF_UNLOCK(&so->so_snd); 828 goto done; 829 } 830 goto retry_space; 831 } 832 SOCKBUF_UNLOCK(&so->so_snd); 833 834 /* 835 * At the beginning of the first loop check if any headers 836 * are specified and copy them into mbufs. Reduce space in 837 * the socket buffer by the size of the header mbuf chain. 838 * Clear hdr_uio here and hdrlen at the end of the first loop. 839 */ 840 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) { 841 hdr_uio->uio_td = td; 842 hdr_uio->uio_rw = UIO_WRITE; 843 #ifdef KERN_TLS 844 if (tls != NULL) 845 mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 846 tls->params.max_frame_len, M_EXTPG); 847 else 848 #endif 849 mh = m_uiotombuf(hdr_uio, M_WAITOK, 850 space, 0, 0); 851 hdrlen = m_length(mh, &mhtail); 852 space -= hdrlen; 853 /* 854 * If header consumed all the socket buffer space, 855 * don't waste CPU cycles and jump to the end. 856 */ 857 if (space == 0) { 858 sfio = NULL; 859 nios = 0; 860 goto prepend_header; 861 } 862 hdr_uio = NULL; 863 } 864 865 if (vp != NULL) { 866 error = vn_lock(vp, LK_SHARED); 867 if (error != 0) 868 goto done; 869 error = VOP_GETATTR(vp, &va, td->td_ucred); 870 if (error != 0 || off >= va.va_size) { 871 VOP_UNLOCK(vp); 872 goto done; 873 } 874 if (va.va_size != obj_size) { 875 obj_size = va.va_size; 876 rem = nbytes ? 877 omin(nbytes + offset, obj_size) : obj_size; 878 rem -= off; 879 } 880 } 881 882 if (space > rem) 883 space = rem; 884 else if (space > PAGE_SIZE) { 885 /* 886 * Use page boundaries when possible for large 887 * requests. 888 */ 889 if (off & PAGE_MASK) 890 space -= (PAGE_SIZE - (off & PAGE_MASK)); 891 space = trunc_page(space); 892 if (off & PAGE_MASK) 893 space += (PAGE_SIZE - (off & PAGE_MASK)); 894 } 895 896 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE); 897 898 /* 899 * Calculate maximum allowed number of pages for readahead 900 * at this iteration. If SF_USER_READAHEAD was set, we don't 901 * do any heuristics and use exactly the value supplied by 902 * application. Otherwise, we allow readahead up to "rem". 903 * If application wants more, let it be, but there is no 904 * reason to go above MAXPHYS. Also check against "obj_size", 905 * since vm_pager_has_page() can hint beyond EOF. 906 */ 907 if (flags & SF_USER_READAHEAD) { 908 rhpages = SF_READAHEAD(flags); 909 } else { 910 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) - 911 npages; 912 rhpages += SF_READAHEAD(flags); 913 } 914 rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages); 915 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) - 916 npages, rhpages); 917 918 sfio = malloc(sizeof(struct sf_io) + 919 npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK); 920 refcount_init(&sfio->nios, 1); 921 sfio->obj = obj; 922 sfio->error = 0; 923 sfio->m = NULL; 924 #ifdef KERN_TLS 925 /* 926 * This doesn't use ktls_hold() because sfio->m will 927 * also have a reference on 'tls' that will be valid 928 * for all of sfio's lifetime. 929 */ 930 sfio->tls = tls; 931 #endif 932 vm_object_pip_add(obj, 1); 933 error = sendfile_swapin(obj, sfio, &nios, off, space, npages, 934 rhpages, flags); 935 if (error != 0) { 936 if (vp != NULL) 937 VOP_UNLOCK(vp); 938 sendfile_iodone(sfio, NULL, 0, error); 939 goto done; 940 } 941 942 /* 943 * Loop and construct maximum sized mbuf chain to be bulk 944 * dumped into socket buffer. 945 */ 946 pa = sfio->pa; 947 948 /* 949 * Use unmapped mbufs if enabled for TCP. Unmapped 950 * bufs are restricted to TCP as that is what has been 951 * tested. In particular, unmapped mbufs have not 952 * been tested with UNIX-domain sockets. 953 * 954 * TLS frames always require unmapped mbufs. 955 */ 956 if ((mb_use_ext_pgs && 957 so->so_proto->pr_protocol == IPPROTO_TCP) 958 #ifdef KERN_TLS 959 || tls != NULL 960 #endif 961 ) { 962 use_ext_pgs = true; 963 #ifdef KERN_TLS 964 if (tls != NULL) 965 max_pgs = num_pages(tls->params.max_frame_len); 966 else 967 #endif 968 max_pgs = MBUF_PEXT_MAX_PGS; 969 970 /* Start at last index, to wrap on first use. */ 971 ext_pgs_idx = max_pgs - 1; 972 } 973 974 for (int i = 0; i < npages; i++) { 975 /* 976 * If a page wasn't grabbed successfully, then 977 * trim the array. Can happen only with SF_NODISKIO. 978 */ 979 if (pa[i] == NULL) { 980 SFSTAT_INC(sf_busy); 981 fixspace(npages, i, off, &space); 982 npages = i; 983 softerr = EBUSY; 984 break; 985 } 986 pga = pa[i]; 987 if (pga == bogus_page) 988 pga = vm_page_relookup(obj, sfio->pindex0 + i); 989 990 if (use_ext_pgs) { 991 off_t xfs; 992 993 ext_pgs_idx++; 994 if (ext_pgs_idx == max_pgs) { 995 m0 = mb_alloc_ext_pgs(M_WAITOK, 996 sendfile_free_mext_pg); 997 998 if (flags & SF_NOCACHE) { 999 m0->m_ext.ext_flags |= 1000 EXT_FLAG_NOCACHE; 1001 1002 /* 1003 * See comment below regarding 1004 * ignoring SF_NOCACHE for the 1005 * last page. 1006 */ 1007 if ((npages - i <= max_pgs) && 1008 ((off + space) & PAGE_MASK) && 1009 (rem > space || rhpages > 0)) 1010 m0->m_ext.ext_flags |= 1011 EXT_FLAG_CACHE_LAST; 1012 } 1013 if (sfs != NULL) { 1014 m0->m_ext.ext_flags |= 1015 EXT_FLAG_SYNC; 1016 m0->m_ext.ext_arg1 = sfs; 1017 mtx_lock(&sfs->mtx); 1018 sfs->count++; 1019 mtx_unlock(&sfs->mtx); 1020 } 1021 ext_pgs_idx = 0; 1022 1023 /* Append to mbuf chain. */ 1024 if (mtail != NULL) 1025 mtail->m_next = m0; 1026 else 1027 m = m0; 1028 mtail = m0; 1029 m0->m_epg_1st_off = 1030 vmoff(i, off) & PAGE_MASK; 1031 } 1032 if (nios) { 1033 mtail->m_flags |= M_NOTREADY; 1034 m0->m_epg_nrdy++; 1035 } 1036 1037 m0->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga); 1038 m0->m_epg_npgs++; 1039 xfs = xfsize(i, npages, off, space); 1040 m0->m_epg_last_len = xfs; 1041 MBUF_EXT_PGS_ASSERT_SANITY(m0); 1042 mtail->m_len += xfs; 1043 mtail->m_ext.ext_size += PAGE_SIZE; 1044 continue; 1045 } 1046 1047 /* 1048 * Get a sendfile buf. When allocating the 1049 * first buffer for mbuf chain, we usually 1050 * wait as long as necessary, but this wait 1051 * can be interrupted. For consequent 1052 * buffers, do not sleep, since several 1053 * threads might exhaust the buffers and then 1054 * deadlock. 1055 */ 1056 sf = sf_buf_alloc(pga, 1057 m != NULL ? SFB_NOWAIT : SFB_CATCH); 1058 if (sf == NULL) { 1059 SFSTAT_INC(sf_allocfail); 1060 sendfile_iowait(sfio, "sfnosf"); 1061 for (int j = i; j < npages; j++) 1062 vm_page_unwire(pa[j], PQ_INACTIVE); 1063 if (m == NULL) 1064 softerr = ENOBUFS; 1065 fixspace(npages, i, off, &space); 1066 npages = i; 1067 break; 1068 } 1069 1070 m0 = m_get(M_WAITOK, MT_DATA); 1071 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf); 1072 m0->m_ext.ext_size = PAGE_SIZE; 1073 m0->m_ext.ext_arg1 = sf; 1074 m0->m_ext.ext_type = EXT_SFBUF; 1075 m0->m_ext.ext_flags = EXT_FLAG_EMBREF; 1076 m0->m_ext.ext_free = sendfile_free_mext; 1077 /* 1078 * SF_NOCACHE sets the page as being freed upon send. 1079 * However, we ignore it for the last page in 'space', 1080 * if the page is truncated, and we got more data to 1081 * send (rem > space), or if we have readahead 1082 * configured (rhpages > 0). 1083 */ 1084 if ((flags & SF_NOCACHE) && 1085 (i != npages - 1 || 1086 !((off + space) & PAGE_MASK) || 1087 !(rem > space || rhpages > 0))) 1088 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE; 1089 if (sfs != NULL) { 1090 m0->m_ext.ext_flags |= EXT_FLAG_SYNC; 1091 m0->m_ext.ext_arg2 = sfs; 1092 mtx_lock(&sfs->mtx); 1093 sfs->count++; 1094 mtx_unlock(&sfs->mtx); 1095 } 1096 m0->m_ext.ext_count = 1; 1097 m0->m_flags |= (M_EXT | M_RDONLY); 1098 if (nios) 1099 m0->m_flags |= M_NOTREADY; 1100 m0->m_data = (char *)sf_buf_kva(sf) + 1101 (vmoff(i, off) & PAGE_MASK); 1102 m0->m_len = xfsize(i, npages, off, space); 1103 1104 /* Append to mbuf chain. */ 1105 if (mtail != NULL) 1106 mtail->m_next = m0; 1107 else 1108 m = m0; 1109 mtail = m0; 1110 } 1111 1112 if (vp != NULL) 1113 VOP_UNLOCK(vp); 1114 1115 /* Keep track of bytes processed. */ 1116 off += space; 1117 rem -= space; 1118 1119 /* 1120 * Prepend header, if any. Save pointer to first mbuf 1121 * with a page. 1122 */ 1123 if (hdrlen) { 1124 prepend_header: 1125 m0 = mhtail->m_next = m; 1126 m = mh; 1127 mh = NULL; 1128 } else 1129 m0 = m; 1130 1131 if (m == NULL) { 1132 KASSERT(softerr, ("%s: m NULL, no error", __func__)); 1133 error = softerr; 1134 sendfile_iodone(sfio, NULL, 0, 0); 1135 goto done; 1136 } 1137 1138 /* Add the buffer chain to the socket buffer. */ 1139 KASSERT(m_length(m, NULL) == space + hdrlen, 1140 ("%s: mlen %u space %d hdrlen %d", 1141 __func__, m_length(m, NULL), space, hdrlen)); 1142 1143 CURVNET_SET(so->so_vnet); 1144 #ifdef KERN_TLS 1145 if (tls != NULL) 1146 ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP); 1147 #endif 1148 if (nios == 0) { 1149 /* 1150 * If sendfile_swapin() didn't initiate any I/Os, 1151 * which happens if all data is cached in VM, or if 1152 * the header consumed all socket buffer space and 1153 * sfio is NULL, then we can send data right now 1154 * without the PRUS_NOTREADY flag. 1155 */ 1156 if (sfio != NULL) 1157 sendfile_iodone(sfio, NULL, 0, 0); 1158 #ifdef KERN_TLS 1159 if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) { 1160 error = (*so->so_proto->pr_usrreqs->pru_send) 1161 (so, PRUS_NOTREADY, m, NULL, NULL, td); 1162 soref(so); 1163 ktls_enqueue(m, so, tls_enq_cnt); 1164 } else 1165 #endif 1166 error = (*so->so_proto->pr_usrreqs->pru_send) 1167 (so, 0, m, NULL, NULL, td); 1168 } else { 1169 sfio->so = so; 1170 sfio->m = m0; 1171 sfio->npages = npages; 1172 soref(so); 1173 error = (*so->so_proto->pr_usrreqs->pru_send) 1174 (so, PRUS_NOTREADY, m, NULL, NULL, td); 1175 sendfile_iodone(sfio, NULL, 0, 0); 1176 } 1177 CURVNET_RESTORE(); 1178 1179 m = NULL; /* pru_send always consumes */ 1180 if (error) 1181 goto done; 1182 sbytes += space + hdrlen; 1183 if (hdrlen) 1184 hdrlen = 0; 1185 if (softerr) { 1186 error = softerr; 1187 goto done; 1188 } 1189 } 1190 1191 /* 1192 * Send trailers. Wimp out and use writev(2). 1193 */ 1194 if (trl_uio != NULL) { 1195 sbunlock(&so->so_snd); 1196 error = kern_writev(td, sockfd, trl_uio); 1197 if (error == 0) 1198 sbytes += td->td_retval[0]; 1199 goto out; 1200 } 1201 1202 done: 1203 sbunlock(&so->so_snd); 1204 out: 1205 /* 1206 * If there was no error we have to clear td->td_retval[0] 1207 * because it may have been set by writev. 1208 */ 1209 if (error == 0) { 1210 td->td_retval[0] = 0; 1211 } 1212 if (sent != NULL) { 1213 (*sent) = sbytes; 1214 } 1215 if (obj != NULL) 1216 vm_object_deallocate(obj); 1217 if (so) 1218 fdrop(sock_fp, td); 1219 if (m) 1220 m_freem(m); 1221 if (mh) 1222 m_freem(mh); 1223 1224 if (sfs != NULL) { 1225 mtx_lock(&sfs->mtx); 1226 if (sfs->count != 0) 1227 error = cv_wait_sig(&sfs->cv, &sfs->mtx); 1228 if (sfs->count == 0) { 1229 sendfile_sync_destroy(sfs); 1230 } else { 1231 sfs->waiting = false; 1232 mtx_unlock(&sfs->mtx); 1233 } 1234 } 1235 #ifdef KERN_TLS 1236 if (tls != NULL) 1237 ktls_free(tls); 1238 #endif 1239 1240 if (error == ERESTART) 1241 error = EINTR; 1242 1243 return (error); 1244 } 1245 1246 static int 1247 sendfile(struct thread *td, struct sendfile_args *uap, int compat) 1248 { 1249 struct sf_hdtr hdtr; 1250 struct uio *hdr_uio, *trl_uio; 1251 struct file *fp; 1252 off_t sbytes; 1253 int error; 1254 1255 /* 1256 * File offset must be positive. If it goes beyond EOF 1257 * we send only the header/trailer and no payload data. 1258 */ 1259 if (uap->offset < 0) 1260 return (EINVAL); 1261 1262 sbytes = 0; 1263 hdr_uio = trl_uio = NULL; 1264 1265 if (uap->hdtr != NULL) { 1266 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr)); 1267 if (error != 0) 1268 goto out; 1269 if (hdtr.headers != NULL) { 1270 error = copyinuio(hdtr.headers, hdtr.hdr_cnt, 1271 &hdr_uio); 1272 if (error != 0) 1273 goto out; 1274 #ifdef COMPAT_FREEBSD4 1275 /* 1276 * In FreeBSD < 5.0 the nbytes to send also included 1277 * the header. If compat is specified subtract the 1278 * header size from nbytes. 1279 */ 1280 if (compat) { 1281 if (uap->nbytes > hdr_uio->uio_resid) 1282 uap->nbytes -= hdr_uio->uio_resid; 1283 else 1284 uap->nbytes = 0; 1285 } 1286 #endif 1287 } 1288 if (hdtr.trailers != NULL) { 1289 error = copyinuio(hdtr.trailers, hdtr.trl_cnt, 1290 &trl_uio); 1291 if (error != 0) 1292 goto out; 1293 } 1294 } 1295 1296 AUDIT_ARG_FD(uap->fd); 1297 1298 /* 1299 * sendfile(2) can start at any offset within a file so we require 1300 * CAP_READ+CAP_SEEK = CAP_PREAD. 1301 */ 1302 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0) 1303 goto out; 1304 1305 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset, 1306 uap->nbytes, &sbytes, uap->flags, td); 1307 fdrop(fp, td); 1308 1309 if (uap->sbytes != NULL) 1310 copyout(&sbytes, uap->sbytes, sizeof(off_t)); 1311 1312 out: 1313 free(hdr_uio, M_IOV); 1314 free(trl_uio, M_IOV); 1315 return (error); 1316 } 1317 1318 /* 1319 * sendfile(2) 1320 * 1321 * int sendfile(int fd, int s, off_t offset, size_t nbytes, 1322 * struct sf_hdtr *hdtr, off_t *sbytes, int flags) 1323 * 1324 * Send a file specified by 'fd' and starting at 'offset' to a socket 1325 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes == 1326 * 0. Optionally add a header and/or trailer to the socket output. If 1327 * specified, write the total number of bytes sent into *sbytes. 1328 */ 1329 int 1330 sys_sendfile(struct thread *td, struct sendfile_args *uap) 1331 { 1332 1333 return (sendfile(td, uap, 0)); 1334 } 1335 1336 #ifdef COMPAT_FREEBSD4 1337 int 1338 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap) 1339 { 1340 struct sendfile_args args; 1341 1342 args.fd = uap->fd; 1343 args.s = uap->s; 1344 args.offset = uap->offset; 1345 args.nbytes = uap->nbytes; 1346 args.hdtr = uap->hdtr; 1347 args.sbytes = uap->sbytes; 1348 args.flags = uap->flags; 1349 1350 return (sendfile(td, &args, 1)); 1351 } 1352 #endif /* COMPAT_FREEBSD4 */ 1353