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 "opt_kern_tls.h" 31 32 #include <sys/param.h> 33 #include <sys/systm.h> 34 #include <sys/capsicum.h> 35 #include <sys/inotify.h> 36 #include <sys/kernel.h> 37 #include <sys/lock.h> 38 #include <sys/ktls.h> 39 #include <sys/mutex.h> 40 #include <sys/malloc.h> 41 #include <sys/mman.h> 42 #include <sys/mount.h> 43 #include <sys/mbuf.h> 44 #include <sys/proc.h> 45 #include <sys/protosw.h> 46 #include <sys/rwlock.h> 47 #include <sys/sf_buf.h> 48 #include <sys/socket.h> 49 #include <sys/socketvar.h> 50 #include <sys/syscallsubr.h> 51 #include <sys/sysctl.h> 52 #include <sys/sysproto.h> 53 #include <sys/vnode.h> 54 55 #include <net/vnet.h> 56 #include <netinet/in.h> 57 #include <netinet/tcp.h> 58 #include <netinet/in_pcb.h> 59 #include <netinet/tcp_var.h> 60 #include <netinet/tcp_log_buf.h> 61 62 #include <security/audit/audit.h> 63 #include <security/mac/mac_framework.h> 64 65 #include <vm/vm.h> 66 #include <vm/vm_object.h> 67 #include <vm/vm_pager.h> 68 69 static MALLOC_DEFINE(M_SENDFILE, "sendfile", "sendfile dynamic memory"); 70 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 pr_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 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)]; 103 104 static void 105 sfstat_init(const void *unused) 106 { 107 108 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t), 109 M_WAITOK); 110 } 111 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL); 112 113 static int 114 sfstat_sysctl(SYSCTL_HANDLER_ARGS) 115 { 116 struct sfstat s; 117 118 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t)); 119 if (req->newptr) 120 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t)); 121 return (SYSCTL_OUT(req, &s, sizeof(s))); 122 } 123 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, 124 CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, 125 sfstat_sysctl, "I", 126 "sendfile statistics"); 127 128 static void 129 sendfile_free_mext(struct mbuf *m) 130 { 131 struct sf_buf *sf; 132 vm_page_t pg; 133 int flags; 134 135 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF, 136 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m)); 137 138 sf = m->m_ext.ext_arg1; 139 pg = sf_buf_page(sf); 140 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0; 141 142 sf_buf_free(sf); 143 vm_page_release(pg, flags); 144 } 145 146 static void 147 sendfile_free_mext_pg(struct mbuf *m) 148 { 149 vm_page_t pg; 150 int flags, i; 151 bool cache_last; 152 153 M_ASSERTEXTPG(m); 154 155 cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST; 156 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0; 157 158 for (i = 0; i < m->m_epg_npgs; i++) { 159 if (cache_last && i == m->m_epg_npgs - 1) 160 flags = 0; 161 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); 162 vm_page_release(pg, flags); 163 } 164 } 165 166 /* 167 * Helper function to calculate how much data to put into page i of n. 168 * Only first and last pages are special. 169 */ 170 static inline off_t 171 xfsize(int i, int n, off_t off, off_t len) 172 { 173 174 if (i == 0) 175 return (omin(PAGE_SIZE - (off & PAGE_MASK), len)); 176 177 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0) 178 return ((off + len) & PAGE_MASK); 179 180 return (PAGE_SIZE); 181 } 182 183 /* 184 * Helper function to get offset within object for i page. 185 */ 186 static inline vm_ooffset_t 187 vmoff(int i, off_t off) 188 { 189 190 if (i == 0) 191 return ((vm_ooffset_t)off); 192 193 return (trunc_page(off + i * PAGE_SIZE)); 194 } 195 196 /* 197 * Helper function used when allocation of a page or sf_buf failed. 198 * Pretend as if we don't have enough space, subtract xfsize() of 199 * all pages that failed. 200 */ 201 static inline void 202 fixspace(int old, int new, off_t off, int *space) 203 { 204 205 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new)); 206 207 /* Subtract last one. */ 208 *space -= xfsize(old - 1, old, off, *space); 209 old--; 210 211 if (new == old) 212 /* There was only one page. */ 213 return; 214 215 /* Subtract first one. */ 216 if (new == 0) { 217 *space -= xfsize(0, old, off, *space); 218 new++; 219 } 220 221 /* Rest of pages are full sized. */ 222 *space -= (old - new) * PAGE_SIZE; 223 224 KASSERT(*space >= 0, ("%s: space went backwards", __func__)); 225 } 226 227 /* 228 * Wait for all in-flight ios to complete, we must not unwire pages 229 * under them. 230 */ 231 static void 232 sendfile_iowait(struct sf_io *sfio, const char *wmesg) 233 { 234 while (atomic_load_int(&sfio->nios) != 1) 235 pause(wmesg, 1); 236 } 237 238 /* 239 * I/O completion callback. 240 * 241 * When called via I/O path, the curvnet is not set and should be obtained 242 * from the socket. When called synchronously from vn_sendfile(), usually 243 * to report error or just release the reference (all pages are valid), then 244 * curvnet shall be already set. 245 */ 246 static void 247 sendfile_iodone(void *arg, vm_page_t *pa, int count, int error) 248 { 249 struct sf_io *sfio = arg; 250 struct socket *so; 251 int i; 252 253 if (error != 0) 254 sfio->error = error; 255 256 /* 257 * Restore the valid page pointers. They are already 258 * unbusied, but still wired. 259 * 260 * XXXKIB since pages are only wired, and we do not 261 * own the object lock, other users might have 262 * invalidated them in meantime. Similarly, after we 263 * unbusied the swapped-in pages, they can become 264 * invalid under us. 265 */ 266 MPASS(count == 0 || pa[0] != bogus_page); 267 for (i = 0; i < count; i++) { 268 if (pa[i] == bogus_page) { 269 sfio->pa[(pa[0]->pindex - sfio->pindex0) + i] = 270 pa[i] = vm_page_relookup(sfio->obj, 271 pa[0]->pindex + i); 272 KASSERT(pa[i] != NULL, 273 ("%s: page %p[%d] disappeared", 274 __func__, pa, i)); 275 } else { 276 vm_page_xunbusy_unchecked(pa[i]); 277 } 278 } 279 280 if (!refcount_release(&sfio->nios)) 281 return; 282 283 #ifdef INVARIANTS 284 for (i = 1; i < sfio->npages; i++) { 285 if (sfio->pa[i] == NULL) 286 break; 287 KASSERT(vm_page_wired(sfio->pa[i]), 288 ("sfio %p page %d %p not wired", sfio, i, sfio->pa[i])); 289 if (i == 0) 290 continue; 291 KASSERT(sfio->pa[0]->object == sfio->pa[i]->object, 292 ("sfio %p page %d %p wrong owner %p %p", sfio, i, 293 sfio->pa[i], sfio->pa[0]->object, sfio->pa[i]->object)); 294 KASSERT(sfio->pa[0]->pindex + i == sfio->pa[i]->pindex, 295 ("sfio %p page %d %p wrong index %jx %jx", sfio, i, 296 sfio->pa[i], (uintmax_t)sfio->pa[0]->pindex, 297 (uintmax_t)sfio->pa[i]->pindex)); 298 } 299 #endif 300 301 vm_object_pip_wakeup(sfio->obj); 302 303 if (sfio->m == NULL) { 304 /* 305 * Either I/O operation failed, or we failed to allocate 306 * buffers, or we bailed out on first busy page, or we 307 * succeeded filling the request without any I/Os. Anyway, 308 * pr_send() hadn't been executed - nothing had been sent 309 * to the socket yet. 310 */ 311 MPASS((curthread->td_pflags & TDP_KTHREAD) == 0); 312 free(sfio, M_SENDFILE); 313 return; 314 } 315 316 #if defined(KERN_TLS) && defined(INVARIANTS) 317 if ((sfio->m->m_flags & M_EXTPG) != 0) 318 KASSERT(sfio->tls == sfio->m->m_epg_tls, 319 ("TLS session mismatch")); 320 else 321 KASSERT(sfio->tls == NULL, 322 ("non-ext_pgs mbuf with TLS session")); 323 #endif 324 so = sfio->so; 325 CURVNET_SET_QUIET(so->so_vnet); 326 if (__predict_false(sfio->error)) { 327 /* 328 * I/O operation failed. The state of data in the socket 329 * is now inconsistent, and all what we can do is to tear 330 * it down. Protocol abort method would tear down protocol 331 * state, free all ready mbufs and detach not ready ones. 332 * We will free the mbufs corresponding to this I/O manually. 333 * 334 * The socket would be marked with EIO and made available 335 * for read, so that application receives EIO on next 336 * syscall and eventually closes the socket. 337 */ 338 so->so_proto->pr_abort(so); 339 so->so_error = EIO; 340 341 mb_free_notready(sfio->m, sfio->npages); 342 #ifdef KERN_TLS 343 } else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) { 344 /* 345 * I/O operation is complete, but we still need to 346 * encrypt. We cannot do this in the interrupt thread 347 * of the disk controller, so forward the mbufs to a 348 * different thread. 349 * 350 * Donate the socket reference from sfio to rather 351 * than explicitly invoking soref(). 352 */ 353 ktls_enqueue(sfio->m, so, sfio->npages); 354 goto out_with_ref; 355 #endif 356 } else 357 (void)so->so_proto->pr_ready(so, sfio->m, sfio->npages); 358 359 sorele(so); 360 #ifdef KERN_TLS 361 out_with_ref: 362 #endif 363 CURVNET_RESTORE(); 364 free(sfio, M_SENDFILE); 365 } 366 367 /* 368 * Iterate through pages vector and request paging for non-valid pages. 369 */ 370 static int 371 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off, 372 off_t len, int rhpages, int flags) 373 { 374 vm_page_t *pa; 375 int a, count, count1, grabbed, i, j, npages, rv; 376 377 pa = sfio->pa; 378 npages = sfio->npages; 379 *nios = 0; 380 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0; 381 sfio->pindex0 = OFF_TO_IDX(off); 382 383 /* 384 * First grab all the pages and wire them. Note that we grab 385 * only required pages. Readahead pages are dealt with later. 386 */ 387 grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off), 388 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages); 389 if (grabbed < npages) { 390 for (int i = grabbed; i < npages; i++) 391 pa[i] = NULL; 392 npages = grabbed; 393 rhpages = 0; 394 } 395 396 for (i = 0; i < npages;) { 397 /* Skip valid pages. */ 398 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK, 399 xfsize(i, npages, off, len))) { 400 vm_page_xunbusy(pa[i]); 401 SFSTAT_INC(sf_pages_valid); 402 i++; 403 continue; 404 } 405 406 /* 407 * Next page is invalid. Check if it belongs to pager. It 408 * may not be there, which is a regular situation for shmem 409 * pager. For vnode pager this happens only in case of 410 * a sparse file. 411 * 412 * Important feature of vm_pager_has_page() is the hint 413 * stored in 'a', about how many pages we can pagein after 414 * this page in a single I/O. 415 */ 416 VM_OBJECT_RLOCK(obj); 417 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL, 418 &a)) { 419 VM_OBJECT_RUNLOCK(obj); 420 pmap_zero_page(pa[i]); 421 vm_page_valid(pa[i]); 422 MPASS(pa[i]->dirty == 0); 423 vm_page_xunbusy(pa[i]); 424 i++; 425 continue; 426 } 427 VM_OBJECT_RUNLOCK(obj); 428 429 /* 430 * We want to pagein as many pages as possible, limited only 431 * by the 'a' hint and actual request. 432 */ 433 count = min(a + 1, npages - i); 434 435 /* 436 * We should not pagein into a valid page because 437 * there might be still unfinished write tracked by 438 * e.g. a buffer, thus we substitute any valid pages 439 * with the bogus one. 440 * 441 * We must not leave around xbusy pages which are not 442 * part of the run passed to vm_pager_getpages(), 443 * otherwise pager might deadlock waiting for the busy 444 * status of the page, e.g. if it constitues the 445 * buffer needed to validate other page. 446 * 447 * First trim the end of the run consisting of the 448 * valid pages, then replace the rest of the valid 449 * with bogus. 450 */ 451 count1 = count; 452 for (j = i + count - 1; j > i; j--) { 453 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 454 xfsize(j, npages, off, len))) { 455 vm_page_xunbusy(pa[j]); 456 SFSTAT_INC(sf_pages_valid); 457 count--; 458 } else { 459 break; 460 } 461 } 462 463 /* 464 * The last page in the run pa[i + count - 1] is 465 * guaranteed to be invalid by the trim above, so it 466 * is not replaced with bogus, thus -1 in the loop end 467 * condition. 468 */ 469 MPASS(pa[i + count - 1]->valid != VM_PAGE_BITS_ALL); 470 for (j = i + 1; j < i + count - 1; j++) { 471 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 472 xfsize(j, npages, off, len))) { 473 vm_page_xunbusy(pa[j]); 474 SFSTAT_INC(sf_pages_valid); 475 SFSTAT_INC(sf_pages_bogus); 476 pa[j] = bogus_page; 477 } 478 } 479 480 refcount_acquire(&sfio->nios); 481 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL, 482 i + count == npages ? &rhpages : NULL, 483 &sendfile_iodone, sfio); 484 if (__predict_false(rv != VM_PAGER_OK)) { 485 sendfile_iowait(sfio, "sferrio"); 486 487 /* 488 * Do remaining pages recovery before returning EIO. 489 * Pages from 0 to npages are wired. 490 * Pages from (i + count1) to npages are busied. 491 */ 492 for (j = 0; j < npages; j++) { 493 if (j >= i + count1) 494 vm_page_xunbusy(pa[j]); 495 KASSERT(pa[j] != NULL && pa[j] != bogus_page, 496 ("%s: page %p[%d] I/O recovery failure", 497 __func__, pa, j)); 498 vm_page_unwire(pa[j], PQ_INACTIVE); 499 pa[j] = NULL; 500 } 501 return (EIO); 502 } 503 504 SFSTAT_INC(sf_iocnt); 505 SFSTAT_ADD(sf_pages_read, count); 506 if (i + count == npages) 507 SFSTAT_ADD(sf_rhpages_read, rhpages); 508 509 i += count1; 510 (*nios)++; 511 } 512 513 if (*nios == 0 && npages != 0) 514 SFSTAT_INC(sf_noiocnt); 515 516 return (0); 517 } 518 519 static int 520 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res, 521 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size, 522 int *bsize) 523 { 524 vm_object_t obj; 525 struct vnode *vp; 526 struct shmfd *shmfd; 527 int error; 528 529 error = 0; 530 vp = *vp_res = NULL; 531 obj = NULL; 532 shmfd = *shmfd_res = NULL; 533 *bsize = 0; 534 535 /* 536 * The file descriptor must be a regular file and have a 537 * backing VM object. 538 */ 539 if (fp->f_type == DTYPE_VNODE) { 540 vp = fp->f_vnode; 541 vn_lock(vp, LK_SHARED | LK_RETRY); 542 if (vp->v_type != VREG) { 543 error = EINVAL; 544 goto out; 545 } 546 *bsize = vp->v_mount->mnt_stat.f_iosize; 547 obj = vp->v_object; 548 if (obj == NULL) { 549 error = EINVAL; 550 goto out; 551 } 552 553 /* 554 * Use the pager size when available to simplify synchronization 555 * with filesystems, which otherwise must atomically update both 556 * the vnode pager size and file size. 557 */ 558 if (obj->type == OBJT_VNODE) { 559 VM_OBJECT_RLOCK(obj); 560 *obj_size = obj->un_pager.vnp.vnp_size; 561 } else { 562 error = vn_getsize_locked(vp, obj_size, td->td_ucred); 563 if (error != 0) 564 goto out; 565 VM_OBJECT_RLOCK(obj); 566 } 567 } else if (fp->f_type == DTYPE_SHM) { 568 shmfd = fp->f_data; 569 obj = shmfd->shm_object; 570 VM_OBJECT_RLOCK(obj); 571 *obj_size = shmfd->shm_size; 572 } else { 573 error = EINVAL; 574 goto out; 575 } 576 577 if ((obj->flags & OBJ_DEAD) != 0) { 578 VM_OBJECT_RUNLOCK(obj); 579 error = EBADF; 580 goto out; 581 } 582 583 /* 584 * Temporarily increase the backing VM object's reference 585 * count so that a forced reclamation of its vnode does not 586 * immediately destroy it. 587 */ 588 vm_object_reference_locked(obj); 589 VM_OBJECT_RUNLOCK(obj); 590 *obj_res = obj; 591 *vp_res = vp; 592 *shmfd_res = shmfd; 593 594 out: 595 if (vp != NULL) 596 VOP_UNLOCK(vp); 597 return (error); 598 } 599 600 static int 601 sendfile_getsock(struct thread *td, int s, struct file **sock_fp, 602 struct socket **so) 603 { 604 int error; 605 606 *sock_fp = NULL; 607 *so = NULL; 608 609 /* 610 * The socket must be a stream socket and connected. 611 */ 612 error = getsock(td, s, &cap_send_rights, sock_fp); 613 if (error != 0) 614 return (error); 615 *so = (*sock_fp)->f_data; 616 if ((*so)->so_type != SOCK_STREAM) 617 return (EINVAL); 618 /* 619 * SCTP one-to-one style sockets currently don't work with 620 * sendfile(). So indicate EINVAL for now. 621 */ 622 if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP) 623 return (EINVAL); 624 return (0); 625 } 626 627 /* 628 * Check socket state and wait (or EAGAIN) for needed amount of space. 629 */ 630 int 631 sendfile_wait_generic(struct socket *so, off_t need, int *space) 632 { 633 int error; 634 635 MPASS(need > 0); 636 MPASS(space != NULL); 637 638 /* 639 * XXXGL: the hack with sb_lowat originates from d99b0dd2c5297. To 640 * achieve high performance sending with sendfile(2) a non-blocking 641 * socket needs a fairly high low watermark. Otherwise, the socket 642 * will be reported as writable too early, and sendfile(2) will send 643 * just a few bytes each time. It is important to understand that 644 * we are changing sb_lowat not for the current invocation of the 645 * syscall, but for the *next* syscall. So there is no way to 646 * workaround the problem, e.g. provide a special version of sbspace(). 647 * Since this hack has been in the kernel for a long time, we 648 * anticipate that there is a lot of software that will suffer if we 649 * remove it. See also b21104487324. 650 */ 651 error = 0; 652 SOCK_SENDBUF_LOCK(so); 653 if (so->so_snd.sb_flags & SB_AUTOLOWAT) { 654 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2) 655 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2; 656 if (so->so_snd.sb_lowat < PAGE_SIZE && 657 so->so_snd.sb_hiwat >= PAGE_SIZE) 658 so->so_snd.sb_lowat = PAGE_SIZE; 659 } 660 retry_space: 661 if (so->so_snd.sb_state & SBS_CANTSENDMORE) { 662 error = EPIPE; 663 goto done; 664 } else if (so->so_error) { 665 error = so->so_error; 666 so->so_error = 0; 667 goto done; 668 } 669 if ((so->so_state & SS_ISCONNECTED) == 0) { 670 error = ENOTCONN; 671 goto done; 672 } 673 674 *space = sbspace(&so->so_snd); 675 if (*space < need && (*space <= 0 || *space < so->so_snd.sb_lowat)) { 676 if (so->so_state & SS_NBIO) { 677 error = EAGAIN; 678 goto done; 679 } 680 /* 681 * sbwait() drops the lock while sleeping. When we loop back 682 * to retry_space the state may have changed and we retest 683 * for it. 684 */ 685 error = sbwait(so, SO_SND); 686 /* 687 * An error from sbwait() usually indicates that we've been 688 * interrupted by a signal. If we've sent anything then return 689 * bytes sent, otherwise return the error. 690 */ 691 if (error != 0) 692 goto done; 693 goto retry_space; 694 } 695 done: 696 SOCK_SENDBUF_UNLOCK(so); 697 698 return (error); 699 } 700 701 int 702 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, 703 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags, 704 struct thread *td) 705 { 706 struct file *sock_fp; 707 struct vnode *vp; 708 struct vm_object *obj; 709 vm_page_t pga; 710 struct socket *so; 711 const struct protosw *pr; 712 #ifdef KERN_TLS 713 struct ktls_session *tls; 714 #endif 715 struct mbuf *m, *mh, *mhtail; 716 struct sf_buf *sf; 717 struct shmfd *shmfd; 718 struct vattr va; 719 off_t off, sbytes, rem, obj_size, nobj_size; 720 int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr; 721 #ifdef KERN_TLS 722 int tls_enq_cnt; 723 #endif 724 bool use_ext_pgs; 725 726 obj = NULL; 727 so = NULL; 728 m = mh = NULL; 729 #ifdef KERN_TLS 730 tls = NULL; 731 #endif 732 hdrlen = sbytes = 0; 733 softerr = 0; 734 use_ext_pgs = false; 735 736 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize); 737 if (error != 0) 738 goto out; 739 740 error = sendfile_getsock(td, sockfd, &sock_fp, &so); 741 if (error != 0) 742 goto out; 743 pr = so->so_proto; 744 745 #ifdef MAC 746 error = mac_socket_check_send(td->td_ucred, so); 747 if (error != 0) 748 goto out; 749 #endif 750 751 SFSTAT_INC(sf_syscalls); 752 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags)); 753 754 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset; 755 756 /* 757 * Protect against multiple writers to the socket. 758 * 759 * XXXRW: Historically this has assumed non-interruptibility, so now 760 * we implement that, but possibly shouldn't. 761 */ 762 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT | SBL_NOINTR); 763 if (error != 0) 764 goto out; 765 CURVNET_SET(so->so_vnet); 766 #ifdef KERN_TLS 767 tls = ktls_hold(so->so_snd.sb_tls_info); 768 #endif 769 770 /* 771 * Loop through the pages of the file, starting with the requested 772 * offset. Get a file page (do I/O if necessary), map the file page 773 * into an sf_buf, attach an mbuf header to the sf_buf, and queue 774 * it on the socket. 775 * This is done in two loops. The inner loop turns as many pages 776 * as it can, up to available socket buffer space, without blocking 777 * into mbufs to have it bulk delivered into the socket send buffer. 778 * The outer loop checks the state and available space of the socket 779 * and takes care of the overall progress. 780 */ 781 for (off = offset; rem > 0; ) { 782 struct sf_io *sfio; 783 vm_page_t *pa; 784 struct mbuf *m0, *mtail; 785 int nios, space, npages, rhpages; 786 787 mtail = NULL; 788 if ((error = pr->pr_sendfile_wait(so, rem, &space)) != 0) 789 goto done; 790 /* 791 * At the beginning of the first loop check if any headers 792 * are specified and copy them into mbufs. Reduce space in 793 * the socket buffer by the size of the header mbuf chain. 794 * Clear hdr_uio here and hdrlen at the end of the first loop. 795 */ 796 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) { 797 hdr_uio->uio_td = td; 798 hdr_uio->uio_rw = UIO_WRITE; 799 #ifdef KERN_TLS 800 if (tls != NULL) 801 mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 802 tls->params.max_frame_len, M_EXTPG); 803 else 804 #endif 805 mh = m_uiotombuf(hdr_uio, M_WAITOK, 806 space, 0, 0); 807 hdrlen = m_length(mh, &mhtail); 808 space -= hdrlen; 809 /* 810 * If header consumed all the socket buffer space, 811 * don't waste CPU cycles and jump to the end. 812 */ 813 if (space == 0) { 814 sfio = NULL; 815 nios = 0; 816 goto prepend_header; 817 } 818 hdr_uio = NULL; 819 } 820 821 if (vp != NULL) { 822 error = vn_lock(vp, LK_SHARED); 823 if (error != 0) 824 goto done; 825 826 /* 827 * Check to see if the file size has changed. 828 */ 829 if (obj->type == OBJT_VNODE) { 830 VM_OBJECT_RLOCK(obj); 831 nobj_size = obj->un_pager.vnp.vnp_size; 832 VM_OBJECT_RUNLOCK(obj); 833 } else { 834 error = VOP_GETATTR(vp, &va, td->td_ucred); 835 if (error != 0) { 836 VOP_UNLOCK(vp); 837 goto done; 838 } 839 nobj_size = va.va_size; 840 } 841 if (off >= nobj_size) { 842 VOP_UNLOCK(vp); 843 goto done; 844 } 845 if (nobj_size != obj_size) { 846 obj_size = nobj_size; 847 rem = nbytes ? omin(nbytes + offset, obj_size) : 848 obj_size; 849 rem -= off; 850 } 851 } 852 853 if (space > rem) 854 space = rem; 855 else if (space > PAGE_SIZE) { 856 /* 857 * Use page boundaries when possible for large 858 * requests. 859 */ 860 if (off & PAGE_MASK) 861 space -= (PAGE_SIZE - (off & PAGE_MASK)); 862 space = trunc_page(space); 863 if (off & PAGE_MASK) 864 space += (PAGE_SIZE - (off & PAGE_MASK)); 865 } 866 867 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE); 868 869 /* 870 * Calculate maximum allowed number of pages for readahead 871 * at this iteration. If SF_USER_READAHEAD was set, we don't 872 * do any heuristics and use exactly the value supplied by 873 * application. Otherwise, we allow readahead up to "rem". 874 * If application wants more, let it be, but there is no 875 * reason to go above maxphys. Also check against "obj_size", 876 * since vm_pager_has_page() can hint beyond EOF. 877 */ 878 if (flags & SF_USER_READAHEAD) { 879 rhpages = SF_READAHEAD(flags); 880 } else { 881 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) - 882 npages; 883 rhpages += SF_READAHEAD(flags); 884 } 885 rhpages = min(howmany(maxphys, PAGE_SIZE), rhpages); 886 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) - 887 npages, rhpages); 888 889 sfio = malloc(sizeof(struct sf_io) + 890 npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK); 891 refcount_init(&sfio->nios, 1); 892 sfio->obj = obj; 893 sfio->error = 0; 894 sfio->m = NULL; 895 sfio->npages = npages; 896 #ifdef KERN_TLS 897 /* 898 * This doesn't use ktls_hold() because sfio->m will 899 * also have a reference on 'tls' that will be valid 900 * for all of sfio's lifetime. 901 */ 902 sfio->tls = tls; 903 #endif 904 vm_object_pip_add(obj, 1); 905 error = sendfile_swapin(obj, sfio, &nios, off, space, rhpages, 906 flags); 907 if (error != 0) { 908 if (vp != NULL) 909 VOP_UNLOCK(vp); 910 sendfile_iodone(sfio, NULL, 0, error); 911 goto done; 912 } 913 914 /* 915 * Loop and construct maximum sized mbuf chain to be bulk 916 * dumped into socket buffer. 917 */ 918 pa = sfio->pa; 919 920 /* 921 * Use unmapped mbufs if enabled for TCP. Unmapped 922 * bufs are restricted to TCP as that is what has been 923 * tested. In particular, unmapped mbufs have not 924 * been tested with UNIX-domain sockets. 925 * 926 * TLS frames always require unmapped mbufs. 927 */ 928 if ((mb_use_ext_pgs && pr->pr_protocol == IPPROTO_TCP) 929 #ifdef KERN_TLS 930 || tls != NULL 931 #endif 932 ) { 933 use_ext_pgs = true; 934 #ifdef KERN_TLS 935 if (tls != NULL) 936 max_pgs = num_pages(tls->params.max_frame_len); 937 else 938 #endif 939 max_pgs = MBUF_PEXT_MAX_PGS; 940 941 /* Start at last index, to wrap on first use. */ 942 ext_pgs_idx = max_pgs - 1; 943 } 944 945 for (int i = 0; i < npages; i++) { 946 /* 947 * If a page wasn't grabbed successfully, then 948 * trim the array. Can happen only with SF_NODISKIO. 949 */ 950 if (pa[i] == NULL) { 951 SFSTAT_INC(sf_busy); 952 fixspace(npages, i, off, &space); 953 sfio->npages = i; 954 softerr = EBUSY; 955 break; 956 } 957 pga = pa[i]; 958 if (pga == bogus_page) 959 pga = vm_page_relookup(obj, sfio->pindex0 + i); 960 961 if (use_ext_pgs) { 962 off_t xfs; 963 964 ext_pgs_idx++; 965 if (ext_pgs_idx == max_pgs) { 966 m0 = mb_alloc_ext_pgs(M_WAITOK, 967 sendfile_free_mext_pg, M_RDONLY); 968 969 if (flags & SF_NOCACHE) { 970 m0->m_ext.ext_flags |= 971 EXT_FLAG_NOCACHE; 972 973 /* 974 * See comment below regarding 975 * ignoring SF_NOCACHE for the 976 * last page. 977 */ 978 if ((npages - i <= max_pgs) && 979 ((off + space) & PAGE_MASK) && 980 (rem > space || rhpages > 0)) 981 m0->m_ext.ext_flags |= 982 EXT_FLAG_CACHE_LAST; 983 } 984 ext_pgs_idx = 0; 985 986 /* Append to mbuf chain. */ 987 if (mtail != NULL) 988 mtail->m_next = m0; 989 else 990 m = m0; 991 mtail = m0; 992 m0->m_epg_1st_off = 993 vmoff(i, off) & PAGE_MASK; 994 } 995 if (nios) { 996 mtail->m_flags |= M_NOTREADY; 997 m0->m_epg_nrdy++; 998 } 999 1000 m0->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga); 1001 m0->m_epg_npgs++; 1002 xfs = xfsize(i, npages, off, space); 1003 m0->m_epg_last_len = xfs; 1004 MBUF_EXT_PGS_ASSERT_SANITY(m0); 1005 mtail->m_len += xfs; 1006 mtail->m_ext.ext_size += PAGE_SIZE; 1007 continue; 1008 } 1009 1010 /* 1011 * Get a sendfile buf. When allocating the 1012 * first buffer for mbuf chain, we usually 1013 * wait as long as necessary, but this wait 1014 * can be interrupted. For consequent 1015 * buffers, do not sleep, since several 1016 * threads might exhaust the buffers and then 1017 * deadlock. 1018 */ 1019 sf = sf_buf_alloc(pga, 1020 m != NULL ? SFB_NOWAIT : SFB_CATCH); 1021 if (sf == NULL) { 1022 SFSTAT_INC(sf_allocfail); 1023 sendfile_iowait(sfio, "sfnosf"); 1024 for (int j = i; j < npages; j++) { 1025 vm_page_unwire(pa[j], PQ_INACTIVE); 1026 pa[j] = NULL; 1027 } 1028 if (m == NULL) 1029 softerr = ENOBUFS; 1030 fixspace(npages, i, off, &space); 1031 sfio->npages = i; 1032 break; 1033 } 1034 1035 m0 = m_get(M_WAITOK, MT_DATA); 1036 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf); 1037 m0->m_ext.ext_size = PAGE_SIZE; 1038 m0->m_ext.ext_arg1 = sf; 1039 m0->m_ext.ext_type = EXT_SFBUF; 1040 m0->m_ext.ext_flags = EXT_FLAG_EMBREF; 1041 m0->m_ext.ext_free = sendfile_free_mext; 1042 /* 1043 * SF_NOCACHE sets the page as being freed upon send. 1044 * However, we ignore it for the last page in 'space', 1045 * if the page is truncated, and we got more data to 1046 * send (rem > space), or if we have readahead 1047 * configured (rhpages > 0). 1048 */ 1049 if ((flags & SF_NOCACHE) && 1050 (i != npages - 1 || 1051 !((off + space) & PAGE_MASK) || 1052 !(rem > space || rhpages > 0))) 1053 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE; 1054 m0->m_ext.ext_count = 1; 1055 m0->m_flags |= (M_EXT | M_RDONLY); 1056 if (nios) 1057 m0->m_flags |= M_NOTREADY; 1058 m0->m_data = (char *)sf_buf_kva(sf) + 1059 (vmoff(i, off) & PAGE_MASK); 1060 m0->m_len = xfsize(i, npages, off, space); 1061 1062 /* Append to mbuf chain. */ 1063 if (mtail != NULL) 1064 mtail->m_next = m0; 1065 else 1066 m = m0; 1067 mtail = m0; 1068 } 1069 1070 if (vp != NULL) 1071 VOP_UNLOCK(vp); 1072 1073 /* Keep track of bytes processed. */ 1074 off += space; 1075 rem -= space; 1076 1077 /* 1078 * Prepend header, if any. Save pointer to first mbuf 1079 * with a page. 1080 */ 1081 if (hdrlen) { 1082 prepend_header: 1083 m0 = mhtail->m_next = m; 1084 m = mh; 1085 mh = NULL; 1086 } else 1087 m0 = m; 1088 1089 if (m == NULL) { 1090 KASSERT(softerr, ("%s: m NULL, no error", __func__)); 1091 error = softerr; 1092 sendfile_iodone(sfio, NULL, 0, 0); 1093 goto done; 1094 } 1095 1096 /* Add the buffer chain to the socket buffer. */ 1097 KASSERT(m_length(m, NULL) == space + hdrlen, 1098 ("%s: mlen %u space %d hdrlen %d", 1099 __func__, m_length(m, NULL), space, hdrlen)); 1100 1101 #ifdef KERN_TLS 1102 if (tls != NULL) 1103 ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP); 1104 #endif 1105 if (nios == 0) { 1106 /* 1107 * If sendfile_swapin() didn't initiate any I/Os, 1108 * which happens if all data is cached in VM, or if 1109 * the header consumed all socket buffer space and 1110 * sfio is NULL, then we can send data right now 1111 * without the PRUS_NOTREADY flag. 1112 */ 1113 if (sfio != NULL) 1114 sendfile_iodone(sfio, NULL, 0, 0); 1115 #ifdef KERN_TLS 1116 if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) { 1117 error = pr->pr_send(so, PRUS_NOTREADY, m, NULL, 1118 NULL, td); 1119 if (error != 0) { 1120 m_freem(m); 1121 } else { 1122 soref(so); 1123 ktls_enqueue(m, so, tls_enq_cnt); 1124 } 1125 } else 1126 #endif 1127 error = pr->pr_send(so, 0, m, NULL, NULL, td); 1128 } else { 1129 sfio->so = so; 1130 sfio->m = m0; 1131 soref(so); 1132 error = pr->pr_send(so, PRUS_NOTREADY, m, NULL, NULL, 1133 td); 1134 sendfile_iodone(sfio, NULL, 0, error); 1135 } 1136 #ifdef TCP_REQUEST_TRK 1137 if (so->so_proto->pr_protocol == IPPROTO_TCP) { 1138 /* log the sendfile call to the TCP log, if enabled */ 1139 tcp_log_sendfile(so, offset, nbytes, flags); 1140 } 1141 #endif 1142 m = NULL; 1143 if (error) 1144 goto done; 1145 sbytes += space + hdrlen; 1146 if (hdrlen) 1147 hdrlen = 0; 1148 if (softerr) { 1149 error = softerr; 1150 goto done; 1151 } 1152 } 1153 1154 /* 1155 * Send trailers. Wimp out and use writev(2). 1156 */ 1157 if (trl_uio != NULL) { 1158 SOCK_IO_SEND_UNLOCK(so); 1159 CURVNET_RESTORE(); 1160 error = kern_writev(td, sockfd, trl_uio); 1161 if (error == 0) 1162 sbytes += td->td_retval[0]; 1163 goto out; 1164 } 1165 1166 done: 1167 SOCK_IO_SEND_UNLOCK(so); 1168 CURVNET_RESTORE(); 1169 out: 1170 /* 1171 * If there was no error we have to clear td->td_retval[0] 1172 * because it may have been set by writev. 1173 */ 1174 if (error == 0) { 1175 td->td_retval[0] = 0; 1176 if (sbytes > 0 && vp != NULL) 1177 INOTIFY(vp, IN_ACCESS); 1178 } 1179 if (sent != NULL) { 1180 (*sent) = sbytes; 1181 } 1182 if (obj != NULL) 1183 vm_object_deallocate(obj); 1184 if (so) 1185 fdrop(sock_fp, td); 1186 if (m) 1187 m_freem(m); 1188 if (mh) 1189 m_freem(mh); 1190 #ifdef KERN_TLS 1191 if (tls != NULL) 1192 ktls_free(tls); 1193 #endif 1194 if (error == ERESTART) 1195 error = EINTR; 1196 1197 return (error); 1198 } 1199 1200 static int 1201 sendfile(struct thread *td, struct sendfile_args *uap, int compat) 1202 { 1203 struct sf_hdtr hdtr; 1204 struct uio *hdr_uio, *trl_uio; 1205 struct file *fp; 1206 off_t sbytes; 1207 int error; 1208 1209 /* 1210 * File offset must be positive. If it goes beyond EOF 1211 * we send only the header/trailer and no payload data. 1212 */ 1213 if (uap->offset < 0) 1214 return (EINVAL); 1215 1216 sbytes = 0; 1217 hdr_uio = trl_uio = NULL; 1218 1219 if (uap->hdtr != NULL) { 1220 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr)); 1221 if (error != 0) 1222 goto out; 1223 if (hdtr.headers != NULL) { 1224 error = copyinuio(hdtr.headers, hdtr.hdr_cnt, 1225 &hdr_uio); 1226 if (error != 0) 1227 goto out; 1228 #ifdef COMPAT_FREEBSD4 1229 /* 1230 * In FreeBSD < 5.0 the nbytes to send also included 1231 * the header. If compat is specified subtract the 1232 * header size from nbytes. 1233 */ 1234 if (compat) { 1235 if (uap->nbytes > hdr_uio->uio_resid) 1236 uap->nbytes -= hdr_uio->uio_resid; 1237 else 1238 uap->nbytes = 0; 1239 } 1240 #endif 1241 } 1242 if (hdtr.trailers != NULL) { 1243 error = copyinuio(hdtr.trailers, hdtr.trl_cnt, 1244 &trl_uio); 1245 if (error != 0) 1246 goto out; 1247 } 1248 } 1249 1250 AUDIT_ARG_FD(uap->fd); 1251 1252 /* 1253 * sendfile(2) can start at any offset within a file so we require 1254 * CAP_READ+CAP_SEEK = CAP_PREAD. 1255 */ 1256 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0) 1257 goto out; 1258 1259 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset, 1260 uap->nbytes, &sbytes, uap->flags, td); 1261 fdrop(fp, td); 1262 1263 if (uap->sbytes != NULL) 1264 (void)copyout(&sbytes, uap->sbytes, sizeof(off_t)); 1265 1266 out: 1267 freeuio(hdr_uio); 1268 freeuio(trl_uio); 1269 return (error); 1270 } 1271 1272 /* 1273 * sendfile(2) 1274 * 1275 * int sendfile(int fd, int s, off_t offset, size_t nbytes, 1276 * struct sf_hdtr *hdtr, off_t *sbytes, int flags) 1277 * 1278 * Send a file specified by 'fd' and starting at 'offset' to a socket 1279 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes == 1280 * 0. Optionally add a header and/or trailer to the socket output. If 1281 * specified, write the total number of bytes sent into *sbytes. 1282 */ 1283 int 1284 sys_sendfile(struct thread *td, struct sendfile_args *uap) 1285 { 1286 1287 return (sendfile(td, uap, 0)); 1288 } 1289 1290 #ifdef COMPAT_FREEBSD4 1291 int 1292 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap) 1293 { 1294 struct sendfile_args args; 1295 1296 args.fd = uap->fd; 1297 args.s = uap->s; 1298 args.offset = uap->offset; 1299 args.nbytes = uap->nbytes; 1300 args.hdtr = uap->hdtr; 1301 args.sbytes = uap->sbytes; 1302 args.flags = uap->flags; 1303 1304 return (sendfile(td, &args, 1)); 1305 } 1306 #endif /* COMPAT_FREEBSD4 */ 1307