1 /*- 2 * Copyright (c) 2013-2015 Gleb Smirnoff <glebius@FreeBSD.org> 3 * Copyright (c) 1998, David Greenman. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. Neither the name of the University nor the names of its contributors 14 * may be used to endorse or promote products derived from this software 15 * without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 __FBSDID("$FreeBSD$"); 32 33 #include "opt_kern_tls.h" 34 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/capsicum.h> 38 #include <sys/kernel.h> 39 #include <sys/lock.h> 40 #include <sys/ktls.h> 41 #include <sys/mutex.h> 42 #include <sys/malloc.h> 43 #include <sys/mman.h> 44 #include <sys/mount.h> 45 #include <sys/mbuf.h> 46 #include <sys/proc.h> 47 #include <sys/protosw.h> 48 #include <sys/rwlock.h> 49 #include <sys/sf_buf.h> 50 #include <sys/socket.h> 51 #include <sys/socketvar.h> 52 #include <sys/syscallsubr.h> 53 #include <sys/sysctl.h> 54 #include <sys/sysproto.h> 55 #include <sys/vnode.h> 56 57 #include <net/vnet.h> 58 #include <netinet/in.h> 59 #include <netinet/tcp.h> 60 #include <netinet/in_pcb.h> 61 #include <netinet/tcp_var.h> 62 #include <netinet/tcp_log_buf.h> 63 64 #include <security/audit/audit.h> 65 #include <security/mac/mac_framework.h> 66 67 #include <vm/vm.h> 68 #include <vm/vm_object.h> 69 #include <vm/vm_pager.h> 70 71 static MALLOC_DEFINE(M_SENDFILE, "sendfile", "sendfile dynamic memory"); 72 73 #define EXT_FLAG_SYNC EXT_FLAG_VENDOR1 74 #define EXT_FLAG_NOCACHE EXT_FLAG_VENDOR2 75 #define EXT_FLAG_CACHE_LAST EXT_FLAG_VENDOR3 76 77 /* 78 * Structure describing a single sendfile(2) I/O, which may consist of 79 * several underlying pager I/Os. 80 * 81 * The syscall context allocates the structure and initializes 'nios' 82 * to 1. As sendfile_swapin() runs through pages and starts asynchronous 83 * paging operations, it increments 'nios'. 84 * 85 * Every I/O completion calls sendfile_iodone(), which decrements the 'nios', 86 * and the syscall also calls sendfile_iodone() after allocating all mbufs, 87 * linking them and sending to socket. Whoever reaches zero 'nios' is 88 * responsible to * call pru_ready on the socket, to notify it of readyness 89 * of the data. 90 */ 91 struct sf_io { 92 volatile u_int nios; 93 u_int error; 94 int npages; 95 struct socket *so; 96 struct mbuf *m; 97 vm_object_t obj; 98 vm_pindex_t pindex0; 99 #ifdef KERN_TLS 100 struct ktls_session *tls; 101 #endif 102 vm_page_t pa[]; 103 }; 104 105 /* 106 * Structure used to track requests with SF_SYNC flag. 107 */ 108 struct sendfile_sync { 109 struct mtx mtx; 110 struct cv cv; 111 unsigned count; 112 bool waiting; 113 }; 114 115 static void 116 sendfile_sync_destroy(struct sendfile_sync *sfs) 117 { 118 KASSERT(sfs->count == 0, ("sendfile sync %p still busy", sfs)); 119 120 cv_destroy(&sfs->cv); 121 mtx_destroy(&sfs->mtx); 122 free(sfs, M_SENDFILE); 123 } 124 125 static void 126 sendfile_sync_signal(struct sendfile_sync *sfs) 127 { 128 mtx_lock(&sfs->mtx); 129 KASSERT(sfs->count > 0, ("sendfile sync %p not busy", sfs)); 130 if (--sfs->count == 0) { 131 if (!sfs->waiting) { 132 /* The sendfile() waiter was interrupted by a signal. */ 133 sendfile_sync_destroy(sfs); 134 return; 135 } else { 136 cv_signal(&sfs->cv); 137 } 138 } 139 mtx_unlock(&sfs->mtx); 140 } 141 142 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)]; 143 144 static void 145 sfstat_init(const void *unused) 146 { 147 148 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t), 149 M_WAITOK); 150 } 151 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL); 152 153 static int 154 sfstat_sysctl(SYSCTL_HANDLER_ARGS) 155 { 156 struct sfstat s; 157 158 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t)); 159 if (req->newptr) 160 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t)); 161 return (SYSCTL_OUT(req, &s, sizeof(s))); 162 } 163 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, 164 CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, 165 sfstat_sysctl, "I", 166 "sendfile statistics"); 167 168 static void 169 sendfile_free_mext(struct mbuf *m) 170 { 171 struct sf_buf *sf; 172 vm_page_t pg; 173 int flags; 174 175 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF, 176 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m)); 177 178 sf = m->m_ext.ext_arg1; 179 pg = sf_buf_page(sf); 180 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0; 181 182 sf_buf_free(sf); 183 vm_page_release(pg, flags); 184 185 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) { 186 struct sendfile_sync *sfs = m->m_ext.ext_arg2; 187 sendfile_sync_signal(sfs); 188 } 189 } 190 191 static void 192 sendfile_free_mext_pg(struct mbuf *m) 193 { 194 vm_page_t pg; 195 int flags, i; 196 bool cache_last; 197 198 M_ASSERTEXTPG(m); 199 200 cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST; 201 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0; 202 203 for (i = 0; i < m->m_epg_npgs; i++) { 204 if (cache_last && i == m->m_epg_npgs - 1) 205 flags = 0; 206 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); 207 vm_page_release(pg, flags); 208 } 209 210 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) { 211 struct sendfile_sync *sfs = m->m_ext.ext_arg1; 212 sendfile_sync_signal(sfs); 213 } 214 } 215 216 /* 217 * Helper function to calculate how much data to put into page i of n. 218 * Only first and last pages are special. 219 */ 220 static inline off_t 221 xfsize(int i, int n, off_t off, off_t len) 222 { 223 224 if (i == 0) 225 return (omin(PAGE_SIZE - (off & PAGE_MASK), len)); 226 227 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0) 228 return ((off + len) & PAGE_MASK); 229 230 return (PAGE_SIZE); 231 } 232 233 /* 234 * Helper function to get offset within object for i page. 235 */ 236 static inline vm_ooffset_t 237 vmoff(int i, off_t off) 238 { 239 240 if (i == 0) 241 return ((vm_ooffset_t)off); 242 243 return (trunc_page(off + i * PAGE_SIZE)); 244 } 245 246 /* 247 * Helper function used when allocation of a page or sf_buf failed. 248 * Pretend as if we don't have enough space, subtract xfsize() of 249 * all pages that failed. 250 */ 251 static inline void 252 fixspace(int old, int new, off_t off, int *space) 253 { 254 255 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new)); 256 257 /* Subtract last one. */ 258 *space -= xfsize(old - 1, old, off, *space); 259 old--; 260 261 if (new == old) 262 /* There was only one page. */ 263 return; 264 265 /* Subtract first one. */ 266 if (new == 0) { 267 *space -= xfsize(0, old, off, *space); 268 new++; 269 } 270 271 /* Rest of pages are full sized. */ 272 *space -= (old - new) * PAGE_SIZE; 273 274 KASSERT(*space >= 0, ("%s: space went backwards", __func__)); 275 } 276 277 /* 278 * Wait for all in-flight ios to complete, we must not unwire pages 279 * under them. 280 */ 281 static void 282 sendfile_iowait(struct sf_io *sfio, const char *wmesg) 283 { 284 while (atomic_load_int(&sfio->nios) != 1) 285 pause(wmesg, 1); 286 } 287 288 /* 289 * I/O completion callback. 290 */ 291 static void 292 sendfile_iodone(void *arg, vm_page_t *pa, int count, int error) 293 { 294 struct sf_io *sfio = arg; 295 struct socket *so; 296 int i; 297 298 if (error != 0) 299 sfio->error = error; 300 301 /* 302 * Restore the valid page pointers. They are already 303 * unbusied, but still wired. 304 * 305 * XXXKIB since pages are only wired, and we do not 306 * own the object lock, other users might have 307 * invalidated them in meantime. Similarly, after we 308 * unbusied the swapped-in pages, they can become 309 * invalid under us. 310 */ 311 MPASS(count == 0 || pa[0] != bogus_page); 312 for (i = 0; i < count; i++) { 313 if (pa[i] == bogus_page) { 314 sfio->pa[(pa[0]->pindex - sfio->pindex0) + i] = 315 pa[i] = vm_page_relookup(sfio->obj, 316 pa[0]->pindex + i); 317 KASSERT(pa[i] != NULL, 318 ("%s: page %p[%d] disappeared", 319 __func__, pa, i)); 320 } else { 321 vm_page_xunbusy_unchecked(pa[i]); 322 } 323 } 324 325 if (!refcount_release(&sfio->nios)) 326 return; 327 328 #ifdef INVARIANTS 329 for (i = 1; i < sfio->npages; i++) { 330 if (sfio->pa[i] == NULL) 331 break; 332 KASSERT(vm_page_wired(sfio->pa[i]), 333 ("sfio %p page %d %p not wired", sfio, i, sfio->pa[i])); 334 if (i == 0) 335 continue; 336 KASSERT(sfio->pa[0]->object == sfio->pa[i]->object, 337 ("sfio %p page %d %p wrong owner %p %p", sfio, i, 338 sfio->pa[i], sfio->pa[0]->object, sfio->pa[i]->object)); 339 KASSERT(sfio->pa[0]->pindex + i == sfio->pa[i]->pindex, 340 ("sfio %p page %d %p wrong index %jx %jx", sfio, i, 341 sfio->pa[i], (uintmax_t)sfio->pa[0]->pindex, 342 (uintmax_t)sfio->pa[i]->pindex)); 343 } 344 #endif 345 346 vm_object_pip_wakeup(sfio->obj); 347 348 if (sfio->m == NULL) { 349 /* 350 * Either I/O operation failed, or we failed to allocate 351 * buffers, or we bailed out on first busy page, or we 352 * succeeded filling the request without any I/Os. Anyway, 353 * pru_send hadn't been executed - nothing had been sent 354 * to the socket yet. 355 */ 356 MPASS((curthread->td_pflags & TDP_KTHREAD) == 0); 357 free(sfio, M_SENDFILE); 358 return; 359 } 360 361 #if defined(KERN_TLS) && defined(INVARIANTS) 362 if ((sfio->m->m_flags & M_EXTPG) != 0) 363 KASSERT(sfio->tls == sfio->m->m_epg_tls, 364 ("TLS session mismatch")); 365 else 366 KASSERT(sfio->tls == NULL, 367 ("non-ext_pgs mbuf with TLS session")); 368 #endif 369 so = sfio->so; 370 CURVNET_SET(so->so_vnet); 371 if (__predict_false(sfio->error)) { 372 /* 373 * I/O operation failed. The state of data in the socket 374 * is now inconsistent, and all what we can do is to tear 375 * it down. Protocol abort method would tear down protocol 376 * state, free all ready mbufs and detach not ready ones. 377 * We will free the mbufs corresponding to this I/O manually. 378 * 379 * The socket would be marked with EIO and made available 380 * for read, so that application receives EIO on next 381 * syscall and eventually closes the socket. 382 */ 383 so->so_proto->pr_abort(so); 384 so->so_error = EIO; 385 386 mb_free_notready(sfio->m, sfio->npages); 387 #ifdef KERN_TLS 388 } else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) { 389 /* 390 * I/O operation is complete, but we still need to 391 * encrypt. We cannot do this in the interrupt thread 392 * of the disk controller, so forward the mbufs to a 393 * different thread. 394 * 395 * Donate the socket reference from sfio to rather 396 * than explicitly invoking soref(). 397 */ 398 ktls_enqueue(sfio->m, so, sfio->npages); 399 goto out_with_ref; 400 #endif 401 } else 402 (void)so->so_proto->pr_ready(so, sfio->m, sfio->npages); 403 404 sorele(so); 405 #ifdef KERN_TLS 406 out_with_ref: 407 #endif 408 CURVNET_RESTORE(); 409 free(sfio, M_SENDFILE); 410 } 411 412 /* 413 * Iterate through pages vector and request paging for non-valid pages. 414 */ 415 static int 416 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off, 417 off_t len, int rhpages, int flags) 418 { 419 vm_page_t *pa; 420 int a, count, count1, grabbed, i, j, npages, rv; 421 422 pa = sfio->pa; 423 npages = sfio->npages; 424 *nios = 0; 425 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0; 426 sfio->pindex0 = OFF_TO_IDX(off); 427 428 /* 429 * First grab all the pages and wire them. Note that we grab 430 * only required pages. Readahead pages are dealt with later. 431 */ 432 grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off), 433 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages); 434 if (grabbed < npages) { 435 for (int i = grabbed; i < npages; i++) 436 pa[i] = NULL; 437 npages = grabbed; 438 rhpages = 0; 439 } 440 441 for (i = 0; i < npages;) { 442 /* Skip valid pages. */ 443 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK, 444 xfsize(i, npages, off, len))) { 445 vm_page_xunbusy(pa[i]); 446 SFSTAT_INC(sf_pages_valid); 447 i++; 448 continue; 449 } 450 451 /* 452 * Next page is invalid. Check if it belongs to pager. It 453 * may not be there, which is a regular situation for shmem 454 * pager. For vnode pager this happens only in case of 455 * a sparse file. 456 * 457 * Important feature of vm_pager_has_page() is the hint 458 * stored in 'a', about how many pages we can pagein after 459 * this page in a single I/O. 460 */ 461 VM_OBJECT_RLOCK(obj); 462 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL, 463 &a)) { 464 VM_OBJECT_RUNLOCK(obj); 465 pmap_zero_page(pa[i]); 466 vm_page_valid(pa[i]); 467 MPASS(pa[i]->dirty == 0); 468 vm_page_xunbusy(pa[i]); 469 i++; 470 continue; 471 } 472 VM_OBJECT_RUNLOCK(obj); 473 474 /* 475 * We want to pagein as many pages as possible, limited only 476 * by the 'a' hint and actual request. 477 */ 478 count = min(a + 1, npages - i); 479 480 /* 481 * We should not pagein into a valid page because 482 * there might be still unfinished write tracked by 483 * e.g. a buffer, thus we substitute any valid pages 484 * with the bogus one. 485 * 486 * We must not leave around xbusy pages which are not 487 * part of the run passed to vm_pager_getpages(), 488 * otherwise pager might deadlock waiting for the busy 489 * status of the page, e.g. if it constitues the 490 * buffer needed to validate other page. 491 * 492 * First trim the end of the run consisting of the 493 * valid pages, then replace the rest of the valid 494 * with bogus. 495 */ 496 count1 = count; 497 for (j = i + count - 1; j > i; j--) { 498 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 499 xfsize(j, npages, off, len))) { 500 vm_page_xunbusy(pa[j]); 501 SFSTAT_INC(sf_pages_valid); 502 count--; 503 } else { 504 break; 505 } 506 } 507 508 /* 509 * The last page in the run pa[i + count - 1] is 510 * guaranteed to be invalid by the trim above, so it 511 * is not replaced with bogus, thus -1 in the loop end 512 * condition. 513 */ 514 MPASS(pa[i + count - 1]->valid != VM_PAGE_BITS_ALL); 515 for (j = i + 1; j < i + count - 1; j++) { 516 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 517 xfsize(j, npages, off, len))) { 518 vm_page_xunbusy(pa[j]); 519 SFSTAT_INC(sf_pages_valid); 520 SFSTAT_INC(sf_pages_bogus); 521 pa[j] = bogus_page; 522 } 523 } 524 525 refcount_acquire(&sfio->nios); 526 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL, 527 i + count == npages ? &rhpages : NULL, 528 &sendfile_iodone, sfio); 529 if (__predict_false(rv != VM_PAGER_OK)) { 530 sendfile_iowait(sfio, "sferrio"); 531 532 /* 533 * Do remaining pages recovery before returning EIO. 534 * Pages from 0 to npages are wired. 535 * Pages from (i + count1) to npages are busied. 536 */ 537 for (j = 0; j < npages; j++) { 538 if (j >= i + count1) 539 vm_page_xunbusy(pa[j]); 540 KASSERT(pa[j] != NULL && pa[j] != bogus_page, 541 ("%s: page %p[%d] I/O recovery failure", 542 __func__, pa, j)); 543 vm_page_unwire(pa[j], PQ_INACTIVE); 544 pa[j] = NULL; 545 } 546 return (EIO); 547 } 548 549 SFSTAT_INC(sf_iocnt); 550 SFSTAT_ADD(sf_pages_read, count); 551 if (i + count == npages) 552 SFSTAT_ADD(sf_rhpages_read, rhpages); 553 554 i += count1; 555 (*nios)++; 556 } 557 558 if (*nios == 0 && npages != 0) 559 SFSTAT_INC(sf_noiocnt); 560 561 return (0); 562 } 563 564 static int 565 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res, 566 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size, 567 int *bsize) 568 { 569 vm_object_t obj; 570 struct vnode *vp; 571 struct shmfd *shmfd; 572 int error; 573 574 error = 0; 575 vp = *vp_res = NULL; 576 obj = NULL; 577 shmfd = *shmfd_res = NULL; 578 *bsize = 0; 579 580 /* 581 * The file descriptor must be a regular file and have a 582 * backing VM object. 583 */ 584 if (fp->f_type == DTYPE_VNODE) { 585 vp = fp->f_vnode; 586 vn_lock(vp, LK_SHARED | LK_RETRY); 587 if (vp->v_type != VREG) { 588 error = EINVAL; 589 goto out; 590 } 591 *bsize = vp->v_mount->mnt_stat.f_iosize; 592 obj = vp->v_object; 593 if (obj == NULL) { 594 error = EINVAL; 595 goto out; 596 } 597 598 /* 599 * Use the pager size when available to simplify synchronization 600 * with filesystems, which otherwise must atomically update both 601 * the vnode pager size and file size. 602 */ 603 if (obj->type == OBJT_VNODE) { 604 VM_OBJECT_RLOCK(obj); 605 *obj_size = obj->un_pager.vnp.vnp_size; 606 } else { 607 error = vn_getsize_locked(vp, obj_size, td->td_ucred); 608 if (error != 0) 609 goto out; 610 VM_OBJECT_RLOCK(obj); 611 } 612 } else if (fp->f_type == DTYPE_SHM) { 613 shmfd = fp->f_data; 614 obj = shmfd->shm_object; 615 VM_OBJECT_RLOCK(obj); 616 *obj_size = shmfd->shm_size; 617 } else { 618 error = EINVAL; 619 goto out; 620 } 621 622 if ((obj->flags & OBJ_DEAD) != 0) { 623 VM_OBJECT_RUNLOCK(obj); 624 error = EBADF; 625 goto out; 626 } 627 628 /* 629 * Temporarily increase the backing VM object's reference 630 * count so that a forced reclamation of its vnode does not 631 * immediately destroy it. 632 */ 633 vm_object_reference_locked(obj); 634 VM_OBJECT_RUNLOCK(obj); 635 *obj_res = obj; 636 *vp_res = vp; 637 *shmfd_res = shmfd; 638 639 out: 640 if (vp != NULL) 641 VOP_UNLOCK(vp); 642 return (error); 643 } 644 645 static int 646 sendfile_getsock(struct thread *td, int s, struct file **sock_fp, 647 struct socket **so) 648 { 649 int error; 650 651 *sock_fp = NULL; 652 *so = NULL; 653 654 /* 655 * The socket must be a stream socket and connected. 656 */ 657 error = getsock(td, s, &cap_send_rights, sock_fp); 658 if (error != 0) 659 return (error); 660 *so = (*sock_fp)->f_data; 661 if ((*so)->so_type != SOCK_STREAM) 662 return (EINVAL); 663 /* 664 * SCTP one-to-one style sockets currently don't work with 665 * sendfile(). So indicate EINVAL for now. 666 */ 667 if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP) 668 return (EINVAL); 669 return (0); 670 } 671 672 int 673 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, 674 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags, 675 struct thread *td) 676 { 677 struct file *sock_fp; 678 struct vnode *vp; 679 struct vm_object *obj; 680 vm_page_t pga; 681 struct socket *so; 682 #ifdef KERN_TLS 683 struct ktls_session *tls; 684 #endif 685 struct mbuf *m, *mh, *mhtail; 686 struct sf_buf *sf; 687 struct shmfd *shmfd; 688 struct sendfile_sync *sfs; 689 struct vattr va; 690 off_t off, sbytes, rem, obj_size, nobj_size; 691 int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr; 692 #ifdef KERN_TLS 693 int tls_enq_cnt; 694 #endif 695 bool use_ext_pgs; 696 697 obj = NULL; 698 so = NULL; 699 m = mh = NULL; 700 sfs = NULL; 701 #ifdef KERN_TLS 702 tls = NULL; 703 #endif 704 hdrlen = sbytes = 0; 705 softerr = 0; 706 use_ext_pgs = false; 707 708 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize); 709 if (error != 0) 710 return (error); 711 712 error = sendfile_getsock(td, sockfd, &sock_fp, &so); 713 if (error != 0) 714 goto out; 715 716 #ifdef MAC 717 error = mac_socket_check_send(td->td_ucred, so); 718 if (error != 0) 719 goto out; 720 #endif 721 722 SFSTAT_INC(sf_syscalls); 723 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags)); 724 725 if (flags & SF_SYNC) { 726 sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO); 727 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF); 728 cv_init(&sfs->cv, "sendfile"); 729 sfs->waiting = true; 730 } 731 732 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset; 733 734 /* 735 * Protect against multiple writers to the socket. 736 * 737 * XXXRW: Historically this has assumed non-interruptibility, so now 738 * we implement that, but possibly shouldn't. 739 */ 740 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT | SBL_NOINTR); 741 if (error != 0) 742 goto out; 743 #ifdef KERN_TLS 744 tls = ktls_hold(so->so_snd.sb_tls_info); 745 #endif 746 747 /* 748 * Loop through the pages of the file, starting with the requested 749 * offset. Get a file page (do I/O if necessary), map the file page 750 * into an sf_buf, attach an mbuf header to the sf_buf, and queue 751 * it on the socket. 752 * This is done in two loops. The inner loop turns as many pages 753 * as it can, up to available socket buffer space, without blocking 754 * into mbufs to have it bulk delivered into the socket send buffer. 755 * The outer loop checks the state and available space of the socket 756 * and takes care of the overall progress. 757 */ 758 for (off = offset; rem > 0; ) { 759 struct sf_io *sfio; 760 vm_page_t *pa; 761 struct mbuf *m0, *mtail; 762 int nios, space, npages, rhpages; 763 764 mtail = NULL; 765 /* 766 * Check the socket state for ongoing connection, 767 * no errors and space in socket buffer. 768 * If space is low allow for the remainder of the 769 * file to be processed if it fits the socket buffer. 770 * Otherwise block in waiting for sufficient space 771 * to proceed, or if the socket is nonblocking, return 772 * to userland with EAGAIN while reporting how far 773 * we've come. 774 * We wait until the socket buffer has significant free 775 * space to do bulk sends. This makes good use of file 776 * system read ahead and allows packet segmentation 777 * offloading hardware to take over lots of work. If 778 * we were not careful here we would send off only one 779 * sfbuf at a time. 780 */ 781 SOCKBUF_LOCK(&so->so_snd); 782 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2) 783 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2; 784 retry_space: 785 if (so->so_snd.sb_state & SBS_CANTSENDMORE) { 786 error = EPIPE; 787 SOCKBUF_UNLOCK(&so->so_snd); 788 goto done; 789 } else if (so->so_error) { 790 error = so->so_error; 791 so->so_error = 0; 792 SOCKBUF_UNLOCK(&so->so_snd); 793 goto done; 794 } 795 if ((so->so_state & SS_ISCONNECTED) == 0) { 796 SOCKBUF_UNLOCK(&so->so_snd); 797 error = ENOTCONN; 798 goto done; 799 } 800 801 space = sbspace(&so->so_snd); 802 if (space < rem && 803 (space <= 0 || 804 space < so->so_snd.sb_lowat)) { 805 if (so->so_state & SS_NBIO) { 806 SOCKBUF_UNLOCK(&so->so_snd); 807 error = EAGAIN; 808 goto done; 809 } 810 /* 811 * sbwait drops the lock while sleeping. 812 * When we loop back to retry_space the 813 * state may have changed and we retest 814 * for it. 815 */ 816 error = sbwait(so, SO_SND); 817 /* 818 * An error from sbwait usually indicates that we've 819 * been interrupted by a signal. If we've sent anything 820 * then return bytes sent, otherwise return the error. 821 */ 822 if (error != 0) { 823 SOCKBUF_UNLOCK(&so->so_snd); 824 goto done; 825 } 826 goto retry_space; 827 } 828 SOCKBUF_UNLOCK(&so->so_snd); 829 830 /* 831 * At the beginning of the first loop check if any headers 832 * are specified and copy them into mbufs. Reduce space in 833 * the socket buffer by the size of the header mbuf chain. 834 * Clear hdr_uio here and hdrlen at the end of the first loop. 835 */ 836 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) { 837 hdr_uio->uio_td = td; 838 hdr_uio->uio_rw = UIO_WRITE; 839 #ifdef KERN_TLS 840 if (tls != NULL) 841 mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 842 tls->params.max_frame_len, M_EXTPG); 843 else 844 #endif 845 mh = m_uiotombuf(hdr_uio, M_WAITOK, 846 space, 0, 0); 847 hdrlen = m_length(mh, &mhtail); 848 space -= hdrlen; 849 /* 850 * If header consumed all the socket buffer space, 851 * don't waste CPU cycles and jump to the end. 852 */ 853 if (space == 0) { 854 sfio = NULL; 855 nios = 0; 856 goto prepend_header; 857 } 858 hdr_uio = NULL; 859 } 860 861 if (vp != NULL) { 862 error = vn_lock(vp, LK_SHARED); 863 if (error != 0) 864 goto done; 865 866 /* 867 * Check to see if the file size has changed. 868 */ 869 if (obj->type == OBJT_VNODE) { 870 VM_OBJECT_RLOCK(obj); 871 nobj_size = obj->un_pager.vnp.vnp_size; 872 VM_OBJECT_RUNLOCK(obj); 873 } else { 874 error = VOP_GETATTR(vp, &va, td->td_ucred); 875 if (error != 0) { 876 VOP_UNLOCK(vp); 877 goto done; 878 } 879 nobj_size = va.va_size; 880 } 881 if (off >= nobj_size) { 882 VOP_UNLOCK(vp); 883 goto done; 884 } 885 if (nobj_size != obj_size) { 886 obj_size = nobj_size; 887 rem = nbytes ? omin(nbytes + offset, obj_size) : 888 obj_size; 889 rem -= off; 890 } 891 } 892 893 if (space > rem) 894 space = rem; 895 else if (space > PAGE_SIZE) { 896 /* 897 * Use page boundaries when possible for large 898 * requests. 899 */ 900 if (off & PAGE_MASK) 901 space -= (PAGE_SIZE - (off & PAGE_MASK)); 902 space = trunc_page(space); 903 if (off & PAGE_MASK) 904 space += (PAGE_SIZE - (off & PAGE_MASK)); 905 } 906 907 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE); 908 909 /* 910 * Calculate maximum allowed number of pages for readahead 911 * at this iteration. If SF_USER_READAHEAD was set, we don't 912 * do any heuristics and use exactly the value supplied by 913 * application. Otherwise, we allow readahead up to "rem". 914 * If application wants more, let it be, but there is no 915 * reason to go above maxphys. Also check against "obj_size", 916 * since vm_pager_has_page() can hint beyond EOF. 917 */ 918 if (flags & SF_USER_READAHEAD) { 919 rhpages = SF_READAHEAD(flags); 920 } else { 921 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) - 922 npages; 923 rhpages += SF_READAHEAD(flags); 924 } 925 rhpages = min(howmany(maxphys, PAGE_SIZE), rhpages); 926 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) - 927 npages, rhpages); 928 929 sfio = malloc(sizeof(struct sf_io) + 930 npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK); 931 refcount_init(&sfio->nios, 1); 932 sfio->obj = obj; 933 sfio->error = 0; 934 sfio->m = NULL; 935 sfio->npages = npages; 936 #ifdef KERN_TLS 937 /* 938 * This doesn't use ktls_hold() because sfio->m will 939 * also have a reference on 'tls' that will be valid 940 * for all of sfio's lifetime. 941 */ 942 sfio->tls = tls; 943 #endif 944 vm_object_pip_add(obj, 1); 945 error = sendfile_swapin(obj, sfio, &nios, off, space, rhpages, 946 flags); 947 if (error != 0) { 948 if (vp != NULL) 949 VOP_UNLOCK(vp); 950 sendfile_iodone(sfio, NULL, 0, error); 951 goto done; 952 } 953 954 /* 955 * Loop and construct maximum sized mbuf chain to be bulk 956 * dumped into socket buffer. 957 */ 958 pa = sfio->pa; 959 960 /* 961 * Use unmapped mbufs if enabled for TCP. Unmapped 962 * bufs are restricted to TCP as that is what has been 963 * tested. In particular, unmapped mbufs have not 964 * been tested with UNIX-domain sockets. 965 * 966 * TLS frames always require unmapped mbufs. 967 */ 968 if ((mb_use_ext_pgs && 969 so->so_proto->pr_protocol == IPPROTO_TCP) 970 #ifdef KERN_TLS 971 || tls != NULL 972 #endif 973 ) { 974 use_ext_pgs = true; 975 #ifdef KERN_TLS 976 if (tls != NULL) 977 max_pgs = num_pages(tls->params.max_frame_len); 978 else 979 #endif 980 max_pgs = MBUF_PEXT_MAX_PGS; 981 982 /* Start at last index, to wrap on first use. */ 983 ext_pgs_idx = max_pgs - 1; 984 } 985 986 for (int i = 0; i < npages; i++) { 987 /* 988 * If a page wasn't grabbed successfully, then 989 * trim the array. Can happen only with SF_NODISKIO. 990 */ 991 if (pa[i] == NULL) { 992 SFSTAT_INC(sf_busy); 993 fixspace(npages, i, off, &space); 994 sfio->npages = i; 995 softerr = EBUSY; 996 break; 997 } 998 pga = pa[i]; 999 if (pga == bogus_page) 1000 pga = vm_page_relookup(obj, sfio->pindex0 + i); 1001 1002 if (use_ext_pgs) { 1003 off_t xfs; 1004 1005 ext_pgs_idx++; 1006 if (ext_pgs_idx == max_pgs) { 1007 m0 = mb_alloc_ext_pgs(M_WAITOK, 1008 sendfile_free_mext_pg); 1009 1010 if (flags & SF_NOCACHE) { 1011 m0->m_ext.ext_flags |= 1012 EXT_FLAG_NOCACHE; 1013 1014 /* 1015 * See comment below regarding 1016 * ignoring SF_NOCACHE for the 1017 * last page. 1018 */ 1019 if ((npages - i <= max_pgs) && 1020 ((off + space) & PAGE_MASK) && 1021 (rem > space || rhpages > 0)) 1022 m0->m_ext.ext_flags |= 1023 EXT_FLAG_CACHE_LAST; 1024 } 1025 if (sfs != NULL) { 1026 m0->m_ext.ext_flags |= 1027 EXT_FLAG_SYNC; 1028 m0->m_ext.ext_arg1 = sfs; 1029 mtx_lock(&sfs->mtx); 1030 sfs->count++; 1031 mtx_unlock(&sfs->mtx); 1032 } 1033 ext_pgs_idx = 0; 1034 1035 /* Append to mbuf chain. */ 1036 if (mtail != NULL) 1037 mtail->m_next = m0; 1038 else 1039 m = m0; 1040 mtail = m0; 1041 m0->m_epg_1st_off = 1042 vmoff(i, off) & PAGE_MASK; 1043 } 1044 if (nios) { 1045 mtail->m_flags |= M_NOTREADY; 1046 m0->m_epg_nrdy++; 1047 } 1048 1049 m0->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga); 1050 m0->m_epg_npgs++; 1051 xfs = xfsize(i, npages, off, space); 1052 m0->m_epg_last_len = xfs; 1053 MBUF_EXT_PGS_ASSERT_SANITY(m0); 1054 mtail->m_len += xfs; 1055 mtail->m_ext.ext_size += PAGE_SIZE; 1056 continue; 1057 } 1058 1059 /* 1060 * Get a sendfile buf. When allocating the 1061 * first buffer for mbuf chain, we usually 1062 * wait as long as necessary, but this wait 1063 * can be interrupted. For consequent 1064 * buffers, do not sleep, since several 1065 * threads might exhaust the buffers and then 1066 * deadlock. 1067 */ 1068 sf = sf_buf_alloc(pga, 1069 m != NULL ? SFB_NOWAIT : SFB_CATCH); 1070 if (sf == NULL) { 1071 SFSTAT_INC(sf_allocfail); 1072 sendfile_iowait(sfio, "sfnosf"); 1073 for (int j = i; j < npages; j++) { 1074 vm_page_unwire(pa[j], PQ_INACTIVE); 1075 pa[j] = NULL; 1076 } 1077 if (m == NULL) 1078 softerr = ENOBUFS; 1079 fixspace(npages, i, off, &space); 1080 sfio->npages = i; 1081 break; 1082 } 1083 1084 m0 = m_get(M_WAITOK, MT_DATA); 1085 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf); 1086 m0->m_ext.ext_size = PAGE_SIZE; 1087 m0->m_ext.ext_arg1 = sf; 1088 m0->m_ext.ext_type = EXT_SFBUF; 1089 m0->m_ext.ext_flags = EXT_FLAG_EMBREF; 1090 m0->m_ext.ext_free = sendfile_free_mext; 1091 /* 1092 * SF_NOCACHE sets the page as being freed upon send. 1093 * However, we ignore it for the last page in 'space', 1094 * if the page is truncated, and we got more data to 1095 * send (rem > space), or if we have readahead 1096 * configured (rhpages > 0). 1097 */ 1098 if ((flags & SF_NOCACHE) && 1099 (i != npages - 1 || 1100 !((off + space) & PAGE_MASK) || 1101 !(rem > space || rhpages > 0))) 1102 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE; 1103 if (sfs != NULL) { 1104 m0->m_ext.ext_flags |= EXT_FLAG_SYNC; 1105 m0->m_ext.ext_arg2 = sfs; 1106 mtx_lock(&sfs->mtx); 1107 sfs->count++; 1108 mtx_unlock(&sfs->mtx); 1109 } 1110 m0->m_ext.ext_count = 1; 1111 m0->m_flags |= (M_EXT | M_RDONLY); 1112 if (nios) 1113 m0->m_flags |= M_NOTREADY; 1114 m0->m_data = (char *)sf_buf_kva(sf) + 1115 (vmoff(i, off) & PAGE_MASK); 1116 m0->m_len = xfsize(i, npages, off, space); 1117 1118 /* Append to mbuf chain. */ 1119 if (mtail != NULL) 1120 mtail->m_next = m0; 1121 else 1122 m = m0; 1123 mtail = m0; 1124 } 1125 1126 if (vp != NULL) 1127 VOP_UNLOCK(vp); 1128 1129 /* Keep track of bytes processed. */ 1130 off += space; 1131 rem -= space; 1132 1133 /* 1134 * Prepend header, if any. Save pointer to first mbuf 1135 * with a page. 1136 */ 1137 if (hdrlen) { 1138 prepend_header: 1139 m0 = mhtail->m_next = m; 1140 m = mh; 1141 mh = NULL; 1142 } else 1143 m0 = m; 1144 1145 if (m == NULL) { 1146 KASSERT(softerr, ("%s: m NULL, no error", __func__)); 1147 error = softerr; 1148 sendfile_iodone(sfio, NULL, 0, 0); 1149 goto done; 1150 } 1151 1152 /* Add the buffer chain to the socket buffer. */ 1153 KASSERT(m_length(m, NULL) == space + hdrlen, 1154 ("%s: mlen %u space %d hdrlen %d", 1155 __func__, m_length(m, NULL), space, hdrlen)); 1156 1157 CURVNET_SET(so->so_vnet); 1158 #ifdef KERN_TLS 1159 if (tls != NULL) 1160 ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP); 1161 #endif 1162 if (nios == 0) { 1163 /* 1164 * If sendfile_swapin() didn't initiate any I/Os, 1165 * which happens if all data is cached in VM, or if 1166 * the header consumed all socket buffer space and 1167 * sfio is NULL, then we can send data right now 1168 * without the PRUS_NOTREADY flag. 1169 */ 1170 if (sfio != NULL) 1171 sendfile_iodone(sfio, NULL, 0, 0); 1172 #ifdef KERN_TLS 1173 if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) { 1174 error = so->so_proto->pr_send(so, 1175 PRUS_NOTREADY, m, NULL, NULL, td); 1176 if (error != 0) { 1177 m_freem(m); 1178 } else { 1179 soref(so); 1180 ktls_enqueue(m, so, tls_enq_cnt); 1181 } 1182 } else 1183 #endif 1184 error = so->so_proto->pr_send(so, 0, m, NULL, 1185 NULL, td); 1186 } else { 1187 sfio->so = so; 1188 sfio->m = m0; 1189 soref(so); 1190 error = so->so_proto->pr_send(so, PRUS_NOTREADY, m, 1191 NULL, NULL, td); 1192 sendfile_iodone(sfio, NULL, 0, error); 1193 } 1194 #ifdef TCP_REQUEST_TRK 1195 if (so->so_proto->pr_protocol == IPPROTO_TCP) { 1196 /* log the sendfile call to the TCP log, if enabled */ 1197 tcp_log_sendfile(so, offset, nbytes, flags); 1198 } 1199 #endif 1200 CURVNET_RESTORE(); 1201 1202 m = NULL; 1203 if (error) 1204 goto done; 1205 sbytes += space + hdrlen; 1206 if (hdrlen) 1207 hdrlen = 0; 1208 if (softerr) { 1209 error = softerr; 1210 goto done; 1211 } 1212 } 1213 1214 /* 1215 * Send trailers. Wimp out and use writev(2). 1216 */ 1217 if (trl_uio != NULL) { 1218 SOCK_IO_SEND_UNLOCK(so); 1219 error = kern_writev(td, sockfd, trl_uio); 1220 if (error == 0) 1221 sbytes += td->td_retval[0]; 1222 goto out; 1223 } 1224 1225 done: 1226 SOCK_IO_SEND_UNLOCK(so); 1227 out: 1228 /* 1229 * If there was no error we have to clear td->td_retval[0] 1230 * because it may have been set by writev. 1231 */ 1232 if (error == 0) { 1233 td->td_retval[0] = 0; 1234 } 1235 if (sent != NULL) { 1236 (*sent) = sbytes; 1237 } 1238 if (obj != NULL) 1239 vm_object_deallocate(obj); 1240 if (so) 1241 fdrop(sock_fp, td); 1242 if (m) 1243 m_freem(m); 1244 if (mh) 1245 m_freem(mh); 1246 1247 if (sfs != NULL) { 1248 mtx_lock(&sfs->mtx); 1249 if (sfs->count != 0) 1250 error = cv_wait_sig(&sfs->cv, &sfs->mtx); 1251 if (sfs->count == 0) { 1252 sendfile_sync_destroy(sfs); 1253 } else { 1254 sfs->waiting = false; 1255 mtx_unlock(&sfs->mtx); 1256 } 1257 } 1258 #ifdef KERN_TLS 1259 if (tls != NULL) 1260 ktls_free(tls); 1261 #endif 1262 1263 if (error == ERESTART) 1264 error = EINTR; 1265 1266 return (error); 1267 } 1268 1269 static int 1270 sendfile(struct thread *td, struct sendfile_args *uap, int compat) 1271 { 1272 struct sf_hdtr hdtr; 1273 struct uio *hdr_uio, *trl_uio; 1274 struct file *fp; 1275 off_t sbytes; 1276 int error; 1277 1278 /* 1279 * File offset must be positive. If it goes beyond EOF 1280 * we send only the header/trailer and no payload data. 1281 */ 1282 if (uap->offset < 0) 1283 return (EINVAL); 1284 1285 sbytes = 0; 1286 hdr_uio = trl_uio = NULL; 1287 1288 if (uap->hdtr != NULL) { 1289 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr)); 1290 if (error != 0) 1291 goto out; 1292 if (hdtr.headers != NULL) { 1293 error = copyinuio(hdtr.headers, hdtr.hdr_cnt, 1294 &hdr_uio); 1295 if (error != 0) 1296 goto out; 1297 #ifdef COMPAT_FREEBSD4 1298 /* 1299 * In FreeBSD < 5.0 the nbytes to send also included 1300 * the header. If compat is specified subtract the 1301 * header size from nbytes. 1302 */ 1303 if (compat) { 1304 if (uap->nbytes > hdr_uio->uio_resid) 1305 uap->nbytes -= hdr_uio->uio_resid; 1306 else 1307 uap->nbytes = 0; 1308 } 1309 #endif 1310 } 1311 if (hdtr.trailers != NULL) { 1312 error = copyinuio(hdtr.trailers, hdtr.trl_cnt, 1313 &trl_uio); 1314 if (error != 0) 1315 goto out; 1316 } 1317 } 1318 1319 AUDIT_ARG_FD(uap->fd); 1320 1321 /* 1322 * sendfile(2) can start at any offset within a file so we require 1323 * CAP_READ+CAP_SEEK = CAP_PREAD. 1324 */ 1325 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0) 1326 goto out; 1327 1328 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset, 1329 uap->nbytes, &sbytes, uap->flags, td); 1330 fdrop(fp, td); 1331 1332 if (uap->sbytes != NULL) 1333 copyout(&sbytes, uap->sbytes, sizeof(off_t)); 1334 1335 out: 1336 free(hdr_uio, M_IOV); 1337 free(trl_uio, M_IOV); 1338 return (error); 1339 } 1340 1341 /* 1342 * sendfile(2) 1343 * 1344 * int sendfile(int fd, int s, off_t offset, size_t nbytes, 1345 * struct sf_hdtr *hdtr, off_t *sbytes, int flags) 1346 * 1347 * Send a file specified by 'fd' and starting at 'offset' to a socket 1348 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes == 1349 * 0. Optionally add a header and/or trailer to the socket output. If 1350 * specified, write the total number of bytes sent into *sbytes. 1351 */ 1352 int 1353 sys_sendfile(struct thread *td, struct sendfile_args *uap) 1354 { 1355 1356 return (sendfile(td, uap, 0)); 1357 } 1358 1359 #ifdef COMPAT_FREEBSD4 1360 int 1361 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap) 1362 { 1363 struct sendfile_args args; 1364 1365 args.fd = uap->fd; 1366 args.s = uap->s; 1367 args.offset = uap->offset; 1368 args.nbytes = uap->nbytes; 1369 args.hdtr = uap->hdtr; 1370 args.sbytes = uap->sbytes; 1371 args.flags = uap->flags; 1372 1373 return (sendfile(td, &args, 1)); 1374 } 1375 #endif /* COMPAT_FREEBSD4 */ 1376