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_compat.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/mutex.h> 41 #include <sys/sysproto.h> 42 #include <sys/malloc.h> 43 #include <sys/proc.h> 44 #include <sys/mman.h> 45 #include <sys/mount.h> 46 #include <sys/mbuf.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/vnode.h> 55 56 #include <net/vnet.h> 57 58 #include <security/audit/audit.h> 59 #include <security/mac/mac_framework.h> 60 61 #include <vm/vm.h> 62 #include <vm/vm_object.h> 63 #include <vm/vm_pager.h> 64 65 /* 66 * Structure describing a single sendfile(2) I/O, which may consist of 67 * several underlying pager I/Os. 68 * 69 * The syscall context allocates the structure and initializes 'nios' 70 * to 1. As sendfile_swapin() runs through pages and starts asynchronous 71 * paging operations, it increments 'nios'. 72 * 73 * Every I/O completion calls sendfile_iodone(), which decrements the 'nios', 74 * and the syscall also calls sendfile_iodone() after allocating all mbufs, 75 * linking them and sending to socket. Whoever reaches zero 'nios' is 76 * responsible to * call pru_ready on the socket, to notify it of readyness 77 * of the data. 78 */ 79 struct sf_io { 80 volatile u_int nios; 81 u_int error; 82 int npages; 83 struct socket *so; 84 struct mbuf *m; 85 vm_page_t pa[]; 86 }; 87 88 /* 89 * Structure used to track requests with SF_SYNC flag. 90 */ 91 struct sendfile_sync { 92 struct mtx mtx; 93 struct cv cv; 94 unsigned count; 95 }; 96 97 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)]; 98 99 static void 100 sfstat_init(const void *unused) 101 { 102 103 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t), 104 M_WAITOK); 105 } 106 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL); 107 108 static int 109 sfstat_sysctl(SYSCTL_HANDLER_ARGS) 110 { 111 struct sfstat s; 112 113 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t)); 114 if (req->newptr) 115 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t)); 116 return (SYSCTL_OUT(req, &s, sizeof(s))); 117 } 118 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW, 119 NULL, 0, sfstat_sysctl, "I", "sendfile statistics"); 120 121 /* 122 * Detach mapped page and release resources back to the system. Called 123 * by mbuf(9) code when last reference to a page is freed. 124 */ 125 void 126 sf_ext_free(void *arg1, void *arg2) 127 { 128 struct sf_buf *sf = arg1; 129 struct sendfile_sync *sfs = arg2; 130 vm_page_t pg = sf_buf_page(sf); 131 132 sf_buf_free(sf); 133 134 vm_page_lock(pg); 135 /* 136 * Check for the object going away on us. This can 137 * happen since we don't hold a reference to it. 138 * If so, we're responsible for freeing the page. 139 */ 140 if (vm_page_unwire(pg, PQ_INACTIVE) && pg->object == NULL) 141 vm_page_free(pg); 142 vm_page_unlock(pg); 143 144 if (sfs != NULL) { 145 mtx_lock(&sfs->mtx); 146 KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0")); 147 if (--sfs->count == 0) 148 cv_signal(&sfs->cv); 149 mtx_unlock(&sfs->mtx); 150 } 151 } 152 153 /* 154 * Same as above, but forces the page to be detached from the object 155 * and go into free pool. 156 */ 157 void 158 sf_ext_free_nocache(void *arg1, void *arg2) 159 { 160 struct sf_buf *sf = arg1; 161 struct sendfile_sync *sfs = arg2; 162 vm_page_t pg = sf_buf_page(sf); 163 164 sf_buf_free(sf); 165 166 vm_page_lock(pg); 167 if (vm_page_unwire(pg, PQ_NONE)) { 168 vm_object_t obj; 169 170 /* Try to free the page, but only if it is cheap to. */ 171 if ((obj = pg->object) == NULL) 172 vm_page_free(pg); 173 else if (!vm_page_xbusied(pg) && VM_OBJECT_TRYWLOCK(obj)) { 174 vm_page_free(pg); 175 VM_OBJECT_WUNLOCK(obj); 176 } else 177 vm_page_deactivate(pg); 178 } 179 vm_page_unlock(pg); 180 181 if (sfs != NULL) { 182 mtx_lock(&sfs->mtx); 183 KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0")); 184 if (--sfs->count == 0) 185 cv_signal(&sfs->cv); 186 mtx_unlock(&sfs->mtx); 187 } 188 } 189 190 /* 191 * Helper function to calculate how much data to put into page i of n. 192 * Only first and last pages are special. 193 */ 194 static inline off_t 195 xfsize(int i, int n, off_t off, off_t len) 196 { 197 198 if (i == 0) 199 return (omin(PAGE_SIZE - (off & PAGE_MASK), len)); 200 201 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0) 202 return ((off + len) & PAGE_MASK); 203 204 return (PAGE_SIZE); 205 } 206 207 /* 208 * Helper function to get offset within object for i page. 209 */ 210 static inline vm_ooffset_t 211 vmoff(int i, off_t off) 212 { 213 214 if (i == 0) 215 return ((vm_ooffset_t)off); 216 217 return (trunc_page(off + i * PAGE_SIZE)); 218 } 219 220 /* 221 * Helper function used when allocation of a page or sf_buf failed. 222 * Pretend as if we don't have enough space, subtract xfsize() of 223 * all pages that failed. 224 */ 225 static inline void 226 fixspace(int old, int new, off_t off, int *space) 227 { 228 229 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new)); 230 231 /* Subtract last one. */ 232 *space -= xfsize(old - 1, old, off, *space); 233 old--; 234 235 if (new == old) 236 /* There was only one page. */ 237 return; 238 239 /* Subtract first one. */ 240 if (new == 0) { 241 *space -= xfsize(0, old, off, *space); 242 new++; 243 } 244 245 /* Rest of pages are full sized. */ 246 *space -= (old - new) * PAGE_SIZE; 247 248 KASSERT(*space >= 0, ("%s: space went backwards", __func__)); 249 } 250 251 /* 252 * I/O completion callback. 253 */ 254 static void 255 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error) 256 { 257 struct sf_io *sfio = arg; 258 struct socket *so = sfio->so; 259 260 for (int i = 0; i < count; i++) 261 if (pg[i] != bogus_page) 262 vm_page_xunbusy(pg[i]); 263 264 if (error) 265 sfio->error = error; 266 267 if (!refcount_release(&sfio->nios)) 268 return; 269 270 if (sfio->error) { 271 struct mbuf *m; 272 273 /* 274 * I/O operation failed. The state of data in the socket 275 * is now inconsistent, and all what we can do is to tear 276 * it down. Protocol abort method would tear down protocol 277 * state, free all ready mbufs and detach not ready ones. 278 * We will free the mbufs corresponding to this I/O manually. 279 * 280 * The socket would be marked with EIO and made available 281 * for read, so that application receives EIO on next 282 * syscall and eventually closes the socket. 283 */ 284 so->so_proto->pr_usrreqs->pru_abort(so); 285 so->so_error = EIO; 286 287 m = sfio->m; 288 for (int i = 0; i < sfio->npages; i++) 289 m = m_free(m); 290 } else { 291 CURVNET_SET(so->so_vnet); 292 (void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m, 293 sfio->npages); 294 CURVNET_RESTORE(); 295 } 296 297 SOCK_LOCK(so); 298 sorele(so); 299 free(sfio, M_TEMP); 300 } 301 302 /* 303 * Iterate through pages vector and request paging for non-valid pages. 304 */ 305 static int 306 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len, 307 int npages, int rhpages, int flags) 308 { 309 vm_page_t *pa = sfio->pa; 310 int grabbed, nios; 311 312 nios = 0; 313 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0; 314 315 /* 316 * First grab all the pages and wire them. Note that we grab 317 * only required pages. Readahead pages are dealt with later. 318 */ 319 VM_OBJECT_WLOCK(obj); 320 321 grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off), 322 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages); 323 if (grabbed < npages) { 324 for (int i = grabbed; i < npages; i++) 325 pa[i] = NULL; 326 npages = grabbed; 327 rhpages = 0; 328 } 329 330 for (int i = 0; i < npages;) { 331 int j, a, count, rv; 332 333 /* Skip valid pages. */ 334 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK, 335 xfsize(i, npages, off, len))) { 336 vm_page_xunbusy(pa[i]); 337 SFSTAT_INC(sf_pages_valid); 338 i++; 339 continue; 340 } 341 342 /* 343 * Next page is invalid. Check if it belongs to pager. It 344 * may not be there, which is a regular situation for shmem 345 * pager. For vnode pager this happens only in case of 346 * a sparse file. 347 * 348 * Important feature of vm_pager_has_page() is the hint 349 * stored in 'a', about how many pages we can pagein after 350 * this page in a single I/O. 351 */ 352 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL, 353 &a)) { 354 pmap_zero_page(pa[i]); 355 pa[i]->valid = VM_PAGE_BITS_ALL; 356 MPASS(pa[i]->dirty == 0); 357 vm_page_xunbusy(pa[i]); 358 i++; 359 continue; 360 } 361 362 /* 363 * We want to pagein as many pages as possible, limited only 364 * by the 'a' hint and actual request. 365 */ 366 count = min(a + 1, npages - i); 367 368 /* 369 * We should not pagein into a valid page, thus we first trim 370 * any valid pages off the end of request, and substitute 371 * to bogus_page those, that are in the middle. 372 */ 373 for (j = i + count - 1; j > i; j--) { 374 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 375 xfsize(j, npages, off, len))) { 376 count--; 377 rhpages = 0; 378 } else 379 break; 380 } 381 for (j = i + 1; j < i + count - 1; j++) 382 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 383 xfsize(j, npages, off, len))) { 384 vm_page_xunbusy(pa[j]); 385 SFSTAT_INC(sf_pages_valid); 386 SFSTAT_INC(sf_pages_bogus); 387 pa[j] = bogus_page; 388 } 389 390 refcount_acquire(&sfio->nios); 391 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL, 392 i + count == npages ? &rhpages : NULL, 393 &sendfile_iodone, sfio); 394 KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p", 395 __func__, obj, pa[i])); 396 397 SFSTAT_INC(sf_iocnt); 398 SFSTAT_ADD(sf_pages_read, count); 399 if (i + count == npages) 400 SFSTAT_ADD(sf_rhpages_read, rhpages); 401 402 /* 403 * Restore the valid page pointers. They are already 404 * unbusied, but still wired. 405 */ 406 for (j = i; j < i + count; j++) 407 if (pa[j] == bogus_page) { 408 pa[j] = vm_page_lookup(obj, 409 OFF_TO_IDX(vmoff(j, off))); 410 KASSERT(pa[j], ("%s: page %p[%d] disappeared", 411 __func__, pa, j)); 412 413 } 414 i += count; 415 nios++; 416 } 417 418 VM_OBJECT_WUNLOCK(obj); 419 420 if (nios == 0 && npages != 0) 421 SFSTAT_INC(sf_noiocnt); 422 423 return (nios); 424 } 425 426 static int 427 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res, 428 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size, 429 int *bsize) 430 { 431 struct vattr va; 432 vm_object_t obj; 433 struct vnode *vp; 434 struct shmfd *shmfd; 435 int error; 436 437 vp = *vp_res = NULL; 438 obj = NULL; 439 shmfd = *shmfd_res = NULL; 440 *bsize = 0; 441 442 /* 443 * The file descriptor must be a regular file and have a 444 * backing VM object. 445 */ 446 if (fp->f_type == DTYPE_VNODE) { 447 vp = fp->f_vnode; 448 vn_lock(vp, LK_SHARED | LK_RETRY); 449 if (vp->v_type != VREG) { 450 error = EINVAL; 451 goto out; 452 } 453 *bsize = vp->v_mount->mnt_stat.f_iosize; 454 error = VOP_GETATTR(vp, &va, td->td_ucred); 455 if (error != 0) 456 goto out; 457 *obj_size = va.va_size; 458 obj = vp->v_object; 459 if (obj == NULL) { 460 error = EINVAL; 461 goto out; 462 } 463 } else if (fp->f_type == DTYPE_SHM) { 464 error = 0; 465 shmfd = fp->f_data; 466 obj = shmfd->shm_object; 467 *obj_size = shmfd->shm_size; 468 } else { 469 error = EINVAL; 470 goto out; 471 } 472 473 VM_OBJECT_WLOCK(obj); 474 if ((obj->flags & OBJ_DEAD) != 0) { 475 VM_OBJECT_WUNLOCK(obj); 476 error = EBADF; 477 goto out; 478 } 479 480 /* 481 * Temporarily increase the backing VM object's reference 482 * count so that a forced reclamation of its vnode does not 483 * immediately destroy it. 484 */ 485 vm_object_reference_locked(obj); 486 VM_OBJECT_WUNLOCK(obj); 487 *obj_res = obj; 488 *vp_res = vp; 489 *shmfd_res = shmfd; 490 491 out: 492 if (vp != NULL) 493 VOP_UNLOCK(vp, 0); 494 return (error); 495 } 496 497 static int 498 sendfile_getsock(struct thread *td, int s, struct file **sock_fp, 499 struct socket **so) 500 { 501 cap_rights_t rights; 502 int error; 503 504 *sock_fp = NULL; 505 *so = NULL; 506 507 /* 508 * The socket must be a stream socket and connected. 509 */ 510 error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND), 511 sock_fp, NULL, NULL); 512 if (error != 0) 513 return (error); 514 *so = (*sock_fp)->f_data; 515 if ((*so)->so_type != SOCK_STREAM) 516 return (EINVAL); 517 if (((*so)->so_state & SS_ISCONNECTED) == 0) 518 return (ENOTCONN); 519 return (0); 520 } 521 522 int 523 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, 524 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags, 525 struct thread *td) 526 { 527 struct file *sock_fp; 528 struct vnode *vp; 529 struct vm_object *obj; 530 struct socket *so; 531 struct mbuf *m, *mh, *mhtail; 532 struct sf_buf *sf; 533 struct shmfd *shmfd; 534 struct sendfile_sync *sfs; 535 struct vattr va; 536 off_t off, sbytes, rem, obj_size; 537 int error, softerr, bsize, hdrlen; 538 539 obj = NULL; 540 so = NULL; 541 m = mh = NULL; 542 sfs = NULL; 543 hdrlen = sbytes = 0; 544 softerr = 0; 545 546 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize); 547 if (error != 0) 548 return (error); 549 550 error = sendfile_getsock(td, sockfd, &sock_fp, &so); 551 if (error != 0) 552 goto out; 553 554 #ifdef MAC 555 error = mac_socket_check_send(td->td_ucred, so); 556 if (error != 0) 557 goto out; 558 #endif 559 560 SFSTAT_INC(sf_syscalls); 561 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags)); 562 563 if (flags & SF_SYNC) { 564 sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO); 565 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF); 566 cv_init(&sfs->cv, "sendfile"); 567 } 568 569 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset; 570 571 /* 572 * Protect against multiple writers to the socket. 573 * 574 * XXXRW: Historically this has assumed non-interruptibility, so now 575 * we implement that, but possibly shouldn't. 576 */ 577 (void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR); 578 579 /* 580 * Loop through the pages of the file, starting with the requested 581 * offset. Get a file page (do I/O if necessary), map the file page 582 * into an sf_buf, attach an mbuf header to the sf_buf, and queue 583 * it on the socket. 584 * This is done in two loops. The inner loop turns as many pages 585 * as it can, up to available socket buffer space, without blocking 586 * into mbufs to have it bulk delivered into the socket send buffer. 587 * The outer loop checks the state and available space of the socket 588 * and takes care of the overall progress. 589 */ 590 for (off = offset; rem > 0; ) { 591 struct sf_io *sfio; 592 vm_page_t *pa; 593 struct mbuf *mtail; 594 int nios, space, npages, rhpages; 595 596 mtail = NULL; 597 /* 598 * Check the socket state for ongoing connection, 599 * no errors and space in socket buffer. 600 * If space is low allow for the remainder of the 601 * file to be processed if it fits the socket buffer. 602 * Otherwise block in waiting for sufficient space 603 * to proceed, or if the socket is nonblocking, return 604 * to userland with EAGAIN while reporting how far 605 * we've come. 606 * We wait until the socket buffer has significant free 607 * space to do bulk sends. This makes good use of file 608 * system read ahead and allows packet segmentation 609 * offloading hardware to take over lots of work. If 610 * we were not careful here we would send off only one 611 * sfbuf at a time. 612 */ 613 SOCKBUF_LOCK(&so->so_snd); 614 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2) 615 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2; 616 retry_space: 617 if (so->so_snd.sb_state & SBS_CANTSENDMORE) { 618 error = EPIPE; 619 SOCKBUF_UNLOCK(&so->so_snd); 620 goto done; 621 } else if (so->so_error) { 622 error = so->so_error; 623 so->so_error = 0; 624 SOCKBUF_UNLOCK(&so->so_snd); 625 goto done; 626 } 627 space = sbspace(&so->so_snd); 628 if (space < rem && 629 (space <= 0 || 630 space < so->so_snd.sb_lowat)) { 631 if (so->so_state & SS_NBIO) { 632 SOCKBUF_UNLOCK(&so->so_snd); 633 error = EAGAIN; 634 goto done; 635 } 636 /* 637 * sbwait drops the lock while sleeping. 638 * When we loop back to retry_space the 639 * state may have changed and we retest 640 * for it. 641 */ 642 error = sbwait(&so->so_snd); 643 /* 644 * An error from sbwait usually indicates that we've 645 * been interrupted by a signal. If we've sent anything 646 * then return bytes sent, otherwise return the error. 647 */ 648 if (error != 0) { 649 SOCKBUF_UNLOCK(&so->so_snd); 650 goto done; 651 } 652 goto retry_space; 653 } 654 SOCKBUF_UNLOCK(&so->so_snd); 655 656 /* 657 * At the beginning of the first loop check if any headers 658 * are specified and copy them into mbufs. Reduce space in 659 * the socket buffer by the size of the header mbuf chain. 660 * Clear hdr_uio here and hdrlen at the end of the first loop. 661 */ 662 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) { 663 hdr_uio->uio_td = td; 664 hdr_uio->uio_rw = UIO_WRITE; 665 mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0); 666 hdrlen = m_length(mh, &mhtail); 667 space -= hdrlen; 668 /* 669 * If header consumed all the socket buffer space, 670 * don't waste CPU cycles and jump to the end. 671 */ 672 if (space == 0) { 673 sfio = NULL; 674 nios = 0; 675 goto prepend_header; 676 } 677 hdr_uio = NULL; 678 } 679 680 if (vp != NULL) { 681 error = vn_lock(vp, LK_SHARED); 682 if (error != 0) 683 goto done; 684 error = VOP_GETATTR(vp, &va, td->td_ucred); 685 if (error != 0 || off >= va.va_size) { 686 VOP_UNLOCK(vp, 0); 687 goto done; 688 } 689 if (va.va_size != obj_size) { 690 obj_size = va.va_size; 691 rem = nbytes ? 692 omin(nbytes + offset, obj_size) : obj_size; 693 rem -= off; 694 } 695 } 696 697 if (space > rem) 698 space = rem; 699 700 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE); 701 702 /* 703 * Calculate maximum allowed number of pages for readahead 704 * at this iteration. If SF_USER_READAHEAD was set, we don't 705 * do any heuristics and use exactly the value supplied by 706 * application. Otherwise, we allow readahead up to "rem". 707 * If application wants more, let it be, but there is no 708 * reason to go above MAXPHYS. Also check against "obj_size", 709 * since vm_pager_has_page() can hint beyond EOF. 710 */ 711 if (flags & SF_USER_READAHEAD) { 712 rhpages = SF_READAHEAD(flags); 713 } else { 714 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) - 715 npages; 716 rhpages += SF_READAHEAD(flags); 717 } 718 rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages); 719 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) - 720 npages, rhpages); 721 722 sfio = malloc(sizeof(struct sf_io) + 723 npages * sizeof(vm_page_t), M_TEMP, M_WAITOK); 724 refcount_init(&sfio->nios, 1); 725 sfio->so = so; 726 sfio->error = 0; 727 728 nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages, 729 flags); 730 731 /* 732 * Loop and construct maximum sized mbuf chain to be bulk 733 * dumped into socket buffer. 734 */ 735 pa = sfio->pa; 736 for (int i = 0; i < npages; i++) { 737 struct mbuf *m0; 738 739 /* 740 * If a page wasn't grabbed successfully, then 741 * trim the array. Can happen only with SF_NODISKIO. 742 */ 743 if (pa[i] == NULL) { 744 SFSTAT_INC(sf_busy); 745 fixspace(npages, i, off, &space); 746 npages = i; 747 softerr = EBUSY; 748 break; 749 } 750 751 /* 752 * Get a sendfile buf. When allocating the 753 * first buffer for mbuf chain, we usually 754 * wait as long as necessary, but this wait 755 * can be interrupted. For consequent 756 * buffers, do not sleep, since several 757 * threads might exhaust the buffers and then 758 * deadlock. 759 */ 760 sf = sf_buf_alloc(pa[i], 761 m != NULL ? SFB_NOWAIT : SFB_CATCH); 762 if (sf == NULL) { 763 SFSTAT_INC(sf_allocfail); 764 for (int j = i; j < npages; j++) { 765 vm_page_lock(pa[j]); 766 vm_page_unwire(pa[j], PQ_INACTIVE); 767 vm_page_unlock(pa[j]); 768 } 769 if (m == NULL) 770 softerr = ENOBUFS; 771 fixspace(npages, i, off, &space); 772 npages = i; 773 break; 774 } 775 776 m0 = m_get(M_WAITOK, MT_DATA); 777 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf); 778 m0->m_ext.ext_size = PAGE_SIZE; 779 m0->m_ext.ext_arg1 = sf; 780 m0->m_ext.ext_arg2 = sfs; 781 /* 782 * SF_NOCACHE sets the page as being freed upon send. 783 * However, we ignore it for the last page in 'space', 784 * if the page is truncated, and we got more data to 785 * send (rem > space), or if we have readahead 786 * configured (rhpages > 0). 787 */ 788 if ((flags & SF_NOCACHE) == 0 || 789 (i == npages - 1 && 790 ((off + space) & PAGE_MASK) && 791 (rem > space || rhpages > 0))) 792 m0->m_ext.ext_type = EXT_SFBUF; 793 else 794 m0->m_ext.ext_type = EXT_SFBUF_NOCACHE; 795 m0->m_ext.ext_flags = EXT_FLAG_EMBREF; 796 m0->m_ext.ext_count = 1; 797 m0->m_flags |= (M_EXT | M_RDONLY); 798 if (nios) 799 m0->m_flags |= M_NOTREADY; 800 m0->m_data = (char *)sf_buf_kva(sf) + 801 (vmoff(i, off) & PAGE_MASK); 802 m0->m_len = xfsize(i, npages, off, space); 803 804 if (i == 0) 805 sfio->m = m0; 806 807 /* Append to mbuf chain. */ 808 if (mtail != NULL) 809 mtail->m_next = m0; 810 else 811 m = m0; 812 mtail = m0; 813 814 if (sfs != NULL) { 815 mtx_lock(&sfs->mtx); 816 sfs->count++; 817 mtx_unlock(&sfs->mtx); 818 } 819 } 820 821 if (vp != NULL) 822 VOP_UNLOCK(vp, 0); 823 824 /* Keep track of bytes processed. */ 825 off += space; 826 rem -= space; 827 828 /* Prepend header, if any. */ 829 if (hdrlen) { 830 prepend_header: 831 mhtail->m_next = m; 832 m = mh; 833 mh = NULL; 834 } 835 836 if (m == NULL) { 837 KASSERT(softerr, ("%s: m NULL, no error", __func__)); 838 error = softerr; 839 free(sfio, M_TEMP); 840 goto done; 841 } 842 843 /* Add the buffer chain to the socket buffer. */ 844 KASSERT(m_length(m, NULL) == space + hdrlen, 845 ("%s: mlen %u space %d hdrlen %d", 846 __func__, m_length(m, NULL), space, hdrlen)); 847 848 CURVNET_SET(so->so_vnet); 849 if (nios == 0) { 850 /* 851 * If sendfile_swapin() didn't initiate any I/Os, 852 * which happens if all data is cached in VM, then 853 * we can send data right now without the 854 * PRUS_NOTREADY flag. 855 */ 856 free(sfio, M_TEMP); 857 error = (*so->so_proto->pr_usrreqs->pru_send) 858 (so, 0, m, NULL, NULL, td); 859 } else { 860 sfio->npages = npages; 861 soref(so); 862 error = (*so->so_proto->pr_usrreqs->pru_send) 863 (so, PRUS_NOTREADY, m, NULL, NULL, td); 864 sendfile_iodone(sfio, NULL, 0, 0); 865 } 866 CURVNET_RESTORE(); 867 868 m = NULL; /* pru_send always consumes */ 869 if (error) 870 goto done; 871 sbytes += space + hdrlen; 872 if (hdrlen) 873 hdrlen = 0; 874 if (softerr) { 875 error = softerr; 876 goto done; 877 } 878 } 879 880 /* 881 * Send trailers. Wimp out and use writev(2). 882 */ 883 if (trl_uio != NULL) { 884 sbunlock(&so->so_snd); 885 error = kern_writev(td, sockfd, trl_uio); 886 if (error == 0) 887 sbytes += td->td_retval[0]; 888 goto out; 889 } 890 891 done: 892 sbunlock(&so->so_snd); 893 out: 894 /* 895 * If there was no error we have to clear td->td_retval[0] 896 * because it may have been set by writev. 897 */ 898 if (error == 0) { 899 td->td_retval[0] = 0; 900 } 901 if (sent != NULL) { 902 (*sent) = sbytes; 903 } 904 if (obj != NULL) 905 vm_object_deallocate(obj); 906 if (so) 907 fdrop(sock_fp, td); 908 if (m) 909 m_freem(m); 910 if (mh) 911 m_freem(mh); 912 913 if (sfs != NULL) { 914 mtx_lock(&sfs->mtx); 915 if (sfs->count != 0) 916 cv_wait(&sfs->cv, &sfs->mtx); 917 KASSERT(sfs->count == 0, ("sendfile sync still busy")); 918 cv_destroy(&sfs->cv); 919 mtx_destroy(&sfs->mtx); 920 free(sfs, M_TEMP); 921 } 922 923 if (error == ERESTART) 924 error = EINTR; 925 926 return (error); 927 } 928 929 static int 930 sendfile(struct thread *td, struct sendfile_args *uap, int compat) 931 { 932 struct sf_hdtr hdtr; 933 struct uio *hdr_uio, *trl_uio; 934 struct file *fp; 935 cap_rights_t rights; 936 off_t sbytes; 937 int error; 938 939 /* 940 * File offset must be positive. If it goes beyond EOF 941 * we send only the header/trailer and no payload data. 942 */ 943 if (uap->offset < 0) 944 return (EINVAL); 945 946 sbytes = 0; 947 hdr_uio = trl_uio = NULL; 948 949 if (uap->hdtr != NULL) { 950 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr)); 951 if (error != 0) 952 goto out; 953 if (hdtr.headers != NULL) { 954 error = copyinuio(hdtr.headers, hdtr.hdr_cnt, 955 &hdr_uio); 956 if (error != 0) 957 goto out; 958 #ifdef COMPAT_FREEBSD4 959 /* 960 * In FreeBSD < 5.0 the nbytes to send also included 961 * the header. If compat is specified subtract the 962 * header size from nbytes. 963 */ 964 if (compat) { 965 if (uap->nbytes > hdr_uio->uio_resid) 966 uap->nbytes -= hdr_uio->uio_resid; 967 else 968 uap->nbytes = 0; 969 } 970 #endif 971 } 972 if (hdtr.trailers != NULL) { 973 error = copyinuio(hdtr.trailers, hdtr.trl_cnt, 974 &trl_uio); 975 if (error != 0) 976 goto out; 977 } 978 } 979 980 AUDIT_ARG_FD(uap->fd); 981 982 /* 983 * sendfile(2) can start at any offset within a file so we require 984 * CAP_READ+CAP_SEEK = CAP_PREAD. 985 */ 986 if ((error = fget_read(td, uap->fd, 987 cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) { 988 goto out; 989 } 990 991 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset, 992 uap->nbytes, &sbytes, uap->flags, td); 993 fdrop(fp, td); 994 995 if (uap->sbytes != NULL) 996 copyout(&sbytes, uap->sbytes, sizeof(off_t)); 997 998 out: 999 free(hdr_uio, M_IOV); 1000 free(trl_uio, M_IOV); 1001 return (error); 1002 } 1003 1004 /* 1005 * sendfile(2) 1006 * 1007 * int sendfile(int fd, int s, off_t offset, size_t nbytes, 1008 * struct sf_hdtr *hdtr, off_t *sbytes, int flags) 1009 * 1010 * Send a file specified by 'fd' and starting at 'offset' to a socket 1011 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes == 1012 * 0. Optionally add a header and/or trailer to the socket output. If 1013 * specified, write the total number of bytes sent into *sbytes. 1014 */ 1015 int 1016 sys_sendfile(struct thread *td, struct sendfile_args *uap) 1017 { 1018 1019 return (sendfile(td, uap, 0)); 1020 } 1021 1022 #ifdef COMPAT_FREEBSD4 1023 int 1024 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap) 1025 { 1026 struct sendfile_args args; 1027 1028 args.fd = uap->fd; 1029 args.s = uap->s; 1030 args.offset = uap->offset; 1031 args.nbytes = uap->nbytes; 1032 args.hdtr = uap->hdtr; 1033 args.sbytes = uap->sbytes; 1034 args.flags = uap->flags; 1035 1036 return (sendfile(td, &args, 1)); 1037 } 1038 #endif /* COMPAT_FREEBSD4 */ 1039