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