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 * 4. 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 file *sock_fp; 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_offset_t 211 vmoff(int i, off_t off) 212 { 213 214 if (i == 0) 215 return ((vm_offset_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; 259 260 for (int i = 0; i < count; i++) 261 vm_page_xunbusy(pg[i]); 262 263 if (error) 264 sfio->error = error; 265 266 if (!refcount_release(&sfio->nios)) 267 return; 268 269 so = sfio->sock_fp->f_data; 270 271 if (sfio->error) { 272 struct mbuf *m; 273 274 /* 275 * I/O operation failed. The state of data in the socket 276 * is now inconsistent, and all what we can do is to tear 277 * it down. Protocol abort method would tear down protocol 278 * state, free all ready mbufs and detach not ready ones. 279 * We will free the mbufs corresponding to this I/O manually. 280 * 281 * The socket would be marked with EIO and made available 282 * for read, so that application receives EIO on next 283 * syscall and eventually closes the socket. 284 */ 285 so->so_proto->pr_usrreqs->pru_abort(so); 286 so->so_error = EIO; 287 288 m = sfio->m; 289 for (int i = 0; i < sfio->npages; i++) 290 m = m_free(m); 291 } else { 292 CURVNET_SET(so->so_vnet); 293 (void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m, 294 sfio->npages); 295 CURVNET_RESTORE(); 296 } 297 298 /* XXXGL: curthread */ 299 fdrop(sfio->sock_fp, curthread); 300 free(sfio, M_TEMP); 301 } 302 303 /* 304 * Iterate through pages vector and request paging for non-valid pages. 305 */ 306 static int 307 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len, 308 int npages, int rhpages, int flags) 309 { 310 vm_page_t *pa = sfio->pa; 311 int nios; 312 313 nios = 0; 314 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0; 315 316 /* 317 * First grab all the pages and wire them. Note that we grab 318 * only required pages. Readahead pages are dealt with later. 319 */ 320 VM_OBJECT_WLOCK(obj); 321 for (int i = 0; i < npages; i++) { 322 pa[i] = vm_page_grab(obj, OFF_TO_IDX(vmoff(i, off)), 323 VM_ALLOC_WIRED | VM_ALLOC_NORMAL | flags); 324 if (pa[i] == NULL) { 325 npages = i; 326 rhpages = 0; 327 break; 328 } 329 } 330 331 for (int i = 0; i < npages;) { 332 int j, a, count, rv; 333 334 /* Skip valid pages. */ 335 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK, 336 xfsize(i, npages, off, len))) { 337 vm_page_xunbusy(pa[i]); 338 SFSTAT_INC(sf_pages_valid); 339 i++; 340 continue; 341 } 342 343 /* 344 * Now 'i' points to first invalid page, iterate further 345 * to make 'j' point at first valid after a bunch of 346 * invalid ones. 347 */ 348 for (j = i + 1; j < npages; j++) 349 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 350 xfsize(j, npages, off, len))) { 351 SFSTAT_INC(sf_pages_valid); 352 break; 353 } 354 355 /* 356 * Now we got region of invalid pages between 'i' and 'j'. 357 * Check that they belong to pager. They may not be there, 358 * which is a regular situation for shmem pager. For vnode 359 * pager this happens only in case of sparse file. 360 * 361 * Important feature of vm_pager_has_page() is the hint 362 * stored in 'a', about how many pages we can pagein after 363 * this page in a single I/O. 364 */ 365 while (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), 366 NULL, &a) && i < j) { 367 pmap_zero_page(pa[i]); 368 pa[i]->valid = VM_PAGE_BITS_ALL; 369 pa[i]->dirty = 0; 370 vm_page_xunbusy(pa[i]); 371 i++; 372 } 373 if (i == j) 374 continue; 375 376 /* 377 * We want to pagein as many pages as possible, limited only 378 * by the 'a' hint and actual request. 379 * 380 * We should not pagein into already valid page, thus if 381 * 'j' didn't reach last page, trim by that page. 382 * 383 * When the pagein fulfils the request, also specify readahead. 384 */ 385 if (j < npages) 386 a = min(a, j - i - 1); 387 count = min(a + 1, npages - i); 388 389 refcount_acquire(&sfio->nios); 390 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL, 391 i + count == npages ? &rhpages : NULL, 392 &sendfile_iodone, sfio); 393 KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p", 394 __func__, obj, pa[i])); 395 396 SFSTAT_INC(sf_iocnt); 397 SFSTAT_ADD(sf_pages_read, count); 398 if (i + count == npages) 399 SFSTAT_ADD(sf_rhpages_read, rhpages); 400 401 #ifdef INVARIANTS 402 for (j = i; j < i + count && j < npages; j++) 403 KASSERT(pa[j] == vm_page_lookup(obj, 404 OFF_TO_IDX(vmoff(j, off))), 405 ("pa[j] %p lookup %p\n", pa[j], 406 vm_page_lookup(obj, OFF_TO_IDX(vmoff(j, off))))); 407 #endif 408 i += count; 409 nios++; 410 } 411 412 VM_OBJECT_WUNLOCK(obj); 413 414 if (nios == 0 && npages != 0) 415 SFSTAT_INC(sf_noiocnt); 416 417 return (nios); 418 } 419 420 static int 421 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res, 422 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size, 423 int *bsize) 424 { 425 struct vattr va; 426 vm_object_t obj; 427 struct vnode *vp; 428 struct shmfd *shmfd; 429 int error; 430 431 vp = *vp_res = NULL; 432 obj = NULL; 433 shmfd = *shmfd_res = NULL; 434 *bsize = 0; 435 436 /* 437 * The file descriptor must be a regular file and have a 438 * backing VM object. 439 */ 440 if (fp->f_type == DTYPE_VNODE) { 441 vp = fp->f_vnode; 442 vn_lock(vp, LK_SHARED | LK_RETRY); 443 if (vp->v_type != VREG) { 444 error = EINVAL; 445 goto out; 446 } 447 *bsize = vp->v_mount->mnt_stat.f_iosize; 448 error = VOP_GETATTR(vp, &va, td->td_ucred); 449 if (error != 0) 450 goto out; 451 *obj_size = va.va_size; 452 obj = vp->v_object; 453 if (obj == NULL) { 454 error = EINVAL; 455 goto out; 456 } 457 } else if (fp->f_type == DTYPE_SHM) { 458 error = 0; 459 shmfd = fp->f_data; 460 obj = shmfd->shm_object; 461 *obj_size = shmfd->shm_size; 462 } else { 463 error = EINVAL; 464 goto out; 465 } 466 467 VM_OBJECT_WLOCK(obj); 468 if ((obj->flags & OBJ_DEAD) != 0) { 469 VM_OBJECT_WUNLOCK(obj); 470 error = EBADF; 471 goto out; 472 } 473 474 /* 475 * Temporarily increase the backing VM object's reference 476 * count so that a forced reclamation of its vnode does not 477 * immediately destroy it. 478 */ 479 vm_object_reference_locked(obj); 480 VM_OBJECT_WUNLOCK(obj); 481 *obj_res = obj; 482 *vp_res = vp; 483 *shmfd_res = shmfd; 484 485 out: 486 if (vp != NULL) 487 VOP_UNLOCK(vp, 0); 488 return (error); 489 } 490 491 static int 492 sendfile_getsock(struct thread *td, int s, struct file **sock_fp, 493 struct socket **so) 494 { 495 cap_rights_t rights; 496 int error; 497 498 *sock_fp = NULL; 499 *so = NULL; 500 501 /* 502 * The socket must be a stream socket and connected. 503 */ 504 error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND), 505 sock_fp, NULL); 506 if (error != 0) 507 return (error); 508 *so = (*sock_fp)->f_data; 509 if ((*so)->so_type != SOCK_STREAM) 510 return (EINVAL); 511 if (((*so)->so_state & SS_ISCONNECTED) == 0) 512 return (ENOTCONN); 513 return (0); 514 } 515 516 int 517 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, 518 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags, 519 int kflags, struct thread *td) 520 { 521 struct file *sock_fp; 522 struct vnode *vp; 523 struct vm_object *obj; 524 struct socket *so; 525 struct mbuf *m, *mh, *mhtail; 526 struct sf_buf *sf; 527 struct shmfd *shmfd; 528 struct sendfile_sync *sfs; 529 struct vattr va; 530 off_t off, sbytes, rem, obj_size; 531 int error, softerr, bsize, hdrlen; 532 533 obj = NULL; 534 so = NULL; 535 m = mh = NULL; 536 sfs = NULL; 537 sbytes = 0; 538 softerr = 0; 539 540 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize); 541 if (error != 0) 542 return (error); 543 544 error = sendfile_getsock(td, sockfd, &sock_fp, &so); 545 if (error != 0) 546 goto out; 547 548 #ifdef MAC 549 error = mac_socket_check_send(td->td_ucred, so); 550 if (error != 0) 551 goto out; 552 #endif 553 554 SFSTAT_INC(sf_syscalls); 555 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags)); 556 557 if (flags & SF_SYNC) { 558 sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO); 559 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF); 560 cv_init(&sfs->cv, "sendfile"); 561 } 562 563 /* If headers are specified copy them into mbufs. */ 564 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) { 565 hdr_uio->uio_td = td; 566 hdr_uio->uio_rw = UIO_WRITE; 567 /* 568 * In FBSD < 5.0 the nbytes to send also included 569 * the header. If compat is specified subtract the 570 * header size from nbytes. 571 */ 572 if (kflags & SFK_COMPAT) { 573 if (nbytes > hdr_uio->uio_resid) 574 nbytes -= hdr_uio->uio_resid; 575 else 576 nbytes = 0; 577 } 578 mh = m_uiotombuf(hdr_uio, M_WAITOK, 0, 0, 0); 579 hdrlen = m_length(mh, &mhtail); 580 } else 581 hdrlen = 0; 582 583 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset; 584 585 /* 586 * Protect against multiple writers to the socket. 587 * 588 * XXXRW: Historically this has assumed non-interruptibility, so now 589 * we implement that, but possibly shouldn't. 590 */ 591 (void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR); 592 593 /* 594 * Loop through the pages of the file, starting with the requested 595 * offset. Get a file page (do I/O if necessary), map the file page 596 * into an sf_buf, attach an mbuf header to the sf_buf, and queue 597 * it on the socket. 598 * This is done in two loops. The inner loop turns as many pages 599 * as it can, up to available socket buffer space, without blocking 600 * into mbufs to have it bulk delivered into the socket send buffer. 601 * The outer loop checks the state and available space of the socket 602 * and takes care of the overall progress. 603 */ 604 for (off = offset; rem > 0; ) { 605 struct sf_io *sfio; 606 vm_page_t *pa; 607 struct mbuf *mtail; 608 int nios, space, npages, rhpages; 609 610 mtail = NULL; 611 /* 612 * Check the socket state for ongoing connection, 613 * no errors and space in socket buffer. 614 * If space is low allow for the remainder of the 615 * file to be processed if it fits the socket buffer. 616 * Otherwise block in waiting for sufficient space 617 * to proceed, or if the socket is nonblocking, return 618 * to userland with EAGAIN while reporting how far 619 * we've come. 620 * We wait until the socket buffer has significant free 621 * space to do bulk sends. This makes good use of file 622 * system read ahead and allows packet segmentation 623 * offloading hardware to take over lots of work. If 624 * we were not careful here we would send off only one 625 * sfbuf at a time. 626 */ 627 SOCKBUF_LOCK(&so->so_snd); 628 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2) 629 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2; 630 retry_space: 631 if (so->so_snd.sb_state & SBS_CANTSENDMORE) { 632 error = EPIPE; 633 SOCKBUF_UNLOCK(&so->so_snd); 634 goto done; 635 } else if (so->so_error) { 636 error = so->so_error; 637 so->so_error = 0; 638 SOCKBUF_UNLOCK(&so->so_snd); 639 goto done; 640 } 641 space = sbspace(&so->so_snd); 642 if (space < rem && 643 (space <= 0 || 644 space < so->so_snd.sb_lowat)) { 645 if (so->so_state & SS_NBIO) { 646 SOCKBUF_UNLOCK(&so->so_snd); 647 error = EAGAIN; 648 goto done; 649 } 650 /* 651 * sbwait drops the lock while sleeping. 652 * When we loop back to retry_space the 653 * state may have changed and we retest 654 * for it. 655 */ 656 error = sbwait(&so->so_snd); 657 /* 658 * An error from sbwait usually indicates that we've 659 * been interrupted by a signal. If we've sent anything 660 * then return bytes sent, otherwise return the error. 661 */ 662 if (error != 0) { 663 SOCKBUF_UNLOCK(&so->so_snd); 664 goto done; 665 } 666 goto retry_space; 667 } 668 SOCKBUF_UNLOCK(&so->so_snd); 669 670 /* 671 * Reduce space in the socket buffer by the size of 672 * the header mbuf chain. 673 * hdrlen is set to 0 after the first loop. 674 */ 675 space -= hdrlen; 676 677 if (vp != NULL) { 678 error = vn_lock(vp, LK_SHARED); 679 if (error != 0) 680 goto done; 681 error = VOP_GETATTR(vp, &va, td->td_ucred); 682 if (error != 0 || off >= va.va_size) { 683 VOP_UNLOCK(vp, 0); 684 goto done; 685 } 686 if (va.va_size != obj_size) { 687 if (nbytes == 0) 688 rem += va.va_size - obj_size; 689 else if (offset + nbytes > va.va_size) 690 rem -= (offset + nbytes - va.va_size); 691 obj_size = va.va_size; 692 } 693 } 694 695 if (space > rem) 696 space = rem; 697 698 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE); 699 700 /* 701 * Calculate maximum allowed number of pages for readahead 702 * at this iteration. First, we allow readahead up to "rem". 703 * If application wants more, let it be, but there is no 704 * reason to go above MAXPHYS. Also check against "obj_size", 705 * since vm_pager_has_page() can hint beyond EOF. 706 */ 707 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) - npages; 708 rhpages += SF_READAHEAD(flags); 709 rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages); 710 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) - 711 npages, rhpages); 712 713 sfio = malloc(sizeof(struct sf_io) + 714 npages * sizeof(vm_page_t), M_TEMP, M_WAITOK); 715 refcount_init(&sfio->nios, 1); 716 sfio->error = 0; 717 718 nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages, 719 flags); 720 721 /* 722 * Loop and construct maximum sized mbuf chain to be bulk 723 * dumped into socket buffer. 724 */ 725 pa = sfio->pa; 726 for (int i = 0; i < npages; i++) { 727 struct mbuf *m0; 728 729 /* 730 * If a page wasn't grabbed successfully, then 731 * trim the array. Can happen only with SF_NODISKIO. 732 */ 733 if (pa[i] == NULL) { 734 SFSTAT_INC(sf_busy); 735 fixspace(npages, i, off, &space); 736 npages = i; 737 softerr = EBUSY; 738 break; 739 } 740 741 /* 742 * Get a sendfile buf. When allocating the 743 * first buffer for mbuf chain, we usually 744 * wait as long as necessary, but this wait 745 * can be interrupted. For consequent 746 * buffers, do not sleep, since several 747 * threads might exhaust the buffers and then 748 * deadlock. 749 */ 750 sf = sf_buf_alloc(pa[i], 751 m != NULL ? SFB_NOWAIT : SFB_CATCH); 752 if (sf == NULL) { 753 SFSTAT_INC(sf_allocfail); 754 for (int j = i; j < npages; j++) { 755 vm_page_lock(pa[j]); 756 vm_page_unwire(pa[j], PQ_INACTIVE); 757 vm_page_unlock(pa[j]); 758 } 759 if (m == NULL) 760 softerr = ENOBUFS; 761 fixspace(npages, i, off, &space); 762 npages = i; 763 break; 764 } 765 766 m0 = m_get(M_WAITOK, MT_DATA); 767 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf); 768 m0->m_ext.ext_size = PAGE_SIZE; 769 m0->m_ext.ext_arg1 = sf; 770 m0->m_ext.ext_arg2 = sfs; 771 /* 772 * SF_NOCACHE sets the page as being freed upon send. 773 * However, we ignore it for the last page in 'space', 774 * if the page is truncated, and we got more data to 775 * send (rem > space), or if we have readahead 776 * configured (rhpages > 0). 777 */ 778 if ((flags & SF_NOCACHE) == 0 || 779 (i == npages - 1 && 780 ((off + space) & PAGE_MASK) && 781 (rem > space || rhpages > 0))) 782 m0->m_ext.ext_type = EXT_SFBUF; 783 else 784 m0->m_ext.ext_type = EXT_SFBUF_NOCACHE; 785 m0->m_ext.ext_flags = EXT_FLAG_EMBREF; 786 m0->m_ext.ext_count = 1; 787 m0->m_flags |= (M_EXT | M_RDONLY); 788 if (nios) 789 m0->m_flags |= M_NOTREADY; 790 m0->m_data = (char *)sf_buf_kva(sf) + 791 (vmoff(i, off) & PAGE_MASK); 792 m0->m_len = xfsize(i, npages, off, space); 793 794 if (i == 0) 795 sfio->m = m0; 796 797 /* Append to mbuf chain. */ 798 if (mtail != NULL) 799 mtail->m_next = m0; 800 else 801 m = m0; 802 mtail = m0; 803 804 if (sfs != NULL) { 805 mtx_lock(&sfs->mtx); 806 sfs->count++; 807 mtx_unlock(&sfs->mtx); 808 } 809 } 810 811 if (vp != NULL) 812 VOP_UNLOCK(vp, 0); 813 814 /* Keep track of bytes processed. */ 815 off += space; 816 rem -= space; 817 818 /* Prepend header, if any. */ 819 if (hdrlen) { 820 mhtail->m_next = m; 821 m = mh; 822 mh = NULL; 823 } 824 825 if (m == NULL) { 826 KASSERT(softerr, ("%s: m NULL, no error", __func__)); 827 error = softerr; 828 free(sfio, M_TEMP); 829 goto done; 830 } 831 832 /* Add the buffer chain to the socket buffer. */ 833 KASSERT(m_length(m, NULL) == space + hdrlen, 834 ("%s: mlen %u space %d hdrlen %d", 835 __func__, m_length(m, NULL), space, hdrlen)); 836 837 CURVNET_SET(so->so_vnet); 838 if (nios == 0) { 839 /* 840 * If sendfile_swapin() didn't initiate any I/Os, 841 * which happens if all data is cached in VM, then 842 * we can send data right now without the 843 * PRUS_NOTREADY flag. 844 */ 845 free(sfio, M_TEMP); 846 error = (*so->so_proto->pr_usrreqs->pru_send) 847 (so, 0, m, NULL, NULL, td); 848 } else { 849 sfio->sock_fp = sock_fp; 850 sfio->npages = npages; 851 fhold(sock_fp); 852 error = (*so->so_proto->pr_usrreqs->pru_send) 853 (so, PRUS_NOTREADY, m, NULL, NULL, td); 854 sendfile_iodone(sfio, NULL, 0, 0); 855 } 856 CURVNET_RESTORE(); 857 858 m = NULL; /* pru_send always consumes */ 859 if (error) 860 goto done; 861 sbytes += space + hdrlen; 862 if (hdrlen) 863 hdrlen = 0; 864 if (softerr) { 865 error = softerr; 866 goto done; 867 } 868 } 869 870 /* 871 * Send trailers. Wimp out and use writev(2). 872 */ 873 if (trl_uio != NULL) { 874 sbunlock(&so->so_snd); 875 error = kern_writev(td, sockfd, trl_uio); 876 if (error == 0) 877 sbytes += td->td_retval[0]; 878 goto out; 879 } 880 881 done: 882 sbunlock(&so->so_snd); 883 out: 884 /* 885 * If there was no error we have to clear td->td_retval[0] 886 * because it may have been set by writev. 887 */ 888 if (error == 0) { 889 td->td_retval[0] = 0; 890 } 891 if (sent != NULL) { 892 (*sent) = sbytes; 893 } 894 if (obj != NULL) 895 vm_object_deallocate(obj); 896 if (so) 897 fdrop(sock_fp, td); 898 if (m) 899 m_freem(m); 900 if (mh) 901 m_freem(mh); 902 903 if (sfs != NULL) { 904 mtx_lock(&sfs->mtx); 905 if (sfs->count != 0) 906 cv_wait(&sfs->cv, &sfs->mtx); 907 KASSERT(sfs->count == 0, ("sendfile sync still busy")); 908 cv_destroy(&sfs->cv); 909 mtx_destroy(&sfs->mtx); 910 free(sfs, M_TEMP); 911 } 912 913 if (error == ERESTART) 914 error = EINTR; 915 916 return (error); 917 } 918 919 static int 920 sendfile(struct thread *td, struct sendfile_args *uap, int compat) 921 { 922 struct sf_hdtr hdtr; 923 struct uio *hdr_uio, *trl_uio; 924 struct file *fp; 925 cap_rights_t rights; 926 off_t sbytes; 927 int error; 928 929 /* 930 * File offset must be positive. If it goes beyond EOF 931 * we send only the header/trailer and no payload data. 932 */ 933 if (uap->offset < 0) 934 return (EINVAL); 935 936 hdr_uio = trl_uio = NULL; 937 938 if (uap->hdtr != NULL) { 939 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr)); 940 if (error != 0) 941 goto out; 942 if (hdtr.headers != NULL) { 943 error = copyinuio(hdtr.headers, hdtr.hdr_cnt, 944 &hdr_uio); 945 if (error != 0) 946 goto out; 947 } 948 if (hdtr.trailers != NULL) { 949 error = copyinuio(hdtr.trailers, hdtr.trl_cnt, 950 &trl_uio); 951 if (error != 0) 952 goto out; 953 } 954 } 955 956 AUDIT_ARG_FD(uap->fd); 957 958 /* 959 * sendfile(2) can start at any offset within a file so we require 960 * CAP_READ+CAP_SEEK = CAP_PREAD. 961 */ 962 if ((error = fget_read(td, uap->fd, 963 cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) { 964 goto out; 965 } 966 967 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset, 968 uap->nbytes, &sbytes, uap->flags, compat ? SFK_COMPAT : 0, td); 969 fdrop(fp, td); 970 971 if (uap->sbytes != NULL) 972 copyout(&sbytes, uap->sbytes, sizeof(off_t)); 973 974 out: 975 free(hdr_uio, M_IOV); 976 free(trl_uio, M_IOV); 977 return (error); 978 } 979 980 /* 981 * sendfile(2) 982 * 983 * int sendfile(int fd, int s, off_t offset, size_t nbytes, 984 * struct sf_hdtr *hdtr, off_t *sbytes, int flags) 985 * 986 * Send a file specified by 'fd' and starting at 'offset' to a socket 987 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes == 988 * 0. Optionally add a header and/or trailer to the socket output. If 989 * specified, write the total number of bytes sent into *sbytes. 990 */ 991 int 992 sys_sendfile(struct thread *td, struct sendfile_args *uap) 993 { 994 995 return (sendfile(td, uap, 0)); 996 } 997 998 #ifdef COMPAT_FREEBSD4 999 int 1000 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap) 1001 { 1002 struct sendfile_args args; 1003 1004 args.fd = uap->fd; 1005 args.s = uap->s; 1006 args.offset = uap->offset; 1007 args.nbytes = uap->nbytes; 1008 args.hdtr = uap->hdtr; 1009 args.sbytes = uap->sbytes; 1010 args.flags = uap->flags; 1011 1012 return (sendfile(td, &args, 1)); 1013 } 1014 #endif /* COMPAT_FREEBSD4 */ 1015