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 } 544 return (EIO); 545 } 546 547 SFSTAT_INC(sf_iocnt); 548 SFSTAT_ADD(sf_pages_read, count); 549 if (i + count == npages) 550 SFSTAT_ADD(sf_rhpages_read, rhpages); 551 552 i += count1; 553 (*nios)++; 554 } 555 556 if (*nios == 0 && npages != 0) 557 SFSTAT_INC(sf_noiocnt); 558 559 return (0); 560 } 561 562 static int 563 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res, 564 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size, 565 int *bsize) 566 { 567 struct vattr va; 568 vm_object_t obj; 569 struct vnode *vp; 570 struct shmfd *shmfd; 571 int error; 572 573 vp = *vp_res = NULL; 574 obj = NULL; 575 shmfd = *shmfd_res = NULL; 576 *bsize = 0; 577 578 /* 579 * The file descriptor must be a regular file and have a 580 * backing VM object. 581 */ 582 if (fp->f_type == DTYPE_VNODE) { 583 vp = fp->f_vnode; 584 vn_lock(vp, LK_SHARED | LK_RETRY); 585 if (vp->v_type != VREG) { 586 error = EINVAL; 587 goto out; 588 } 589 *bsize = vp->v_mount->mnt_stat.f_iosize; 590 error = VOP_GETATTR(vp, &va, td->td_ucred); 591 if (error != 0) 592 goto out; 593 *obj_size = va.va_size; 594 obj = vp->v_object; 595 if (obj == NULL) { 596 error = EINVAL; 597 goto out; 598 } 599 } else if (fp->f_type == DTYPE_SHM) { 600 error = 0; 601 shmfd = fp->f_data; 602 obj = shmfd->shm_object; 603 *obj_size = shmfd->shm_size; 604 } else { 605 error = EINVAL; 606 goto out; 607 } 608 609 VM_OBJECT_WLOCK(obj); 610 if ((obj->flags & OBJ_DEAD) != 0) { 611 VM_OBJECT_WUNLOCK(obj); 612 error = EBADF; 613 goto out; 614 } 615 616 /* 617 * Temporarily increase the backing VM object's reference 618 * count so that a forced reclamation of its vnode does not 619 * immediately destroy it. 620 */ 621 vm_object_reference_locked(obj); 622 VM_OBJECT_WUNLOCK(obj); 623 *obj_res = obj; 624 *vp_res = vp; 625 *shmfd_res = shmfd; 626 627 out: 628 if (vp != NULL) 629 VOP_UNLOCK(vp); 630 return (error); 631 } 632 633 static int 634 sendfile_getsock(struct thread *td, int s, struct file **sock_fp, 635 struct socket **so) 636 { 637 int error; 638 639 *sock_fp = NULL; 640 *so = NULL; 641 642 /* 643 * The socket must be a stream socket and connected. 644 */ 645 error = getsock_cap(td, s, &cap_send_rights, 646 sock_fp, NULL, NULL); 647 if (error != 0) 648 return (error); 649 *so = (*sock_fp)->f_data; 650 if ((*so)->so_type != SOCK_STREAM) 651 return (EINVAL); 652 /* 653 * SCTP one-to-one style sockets currently don't work with 654 * sendfile(). So indicate EINVAL for now. 655 */ 656 if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP) 657 return (EINVAL); 658 if (SOLISTENING(*so)) 659 return (ENOTCONN); 660 return (0); 661 } 662 663 int 664 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, 665 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags, 666 struct thread *td) 667 { 668 struct file *sock_fp; 669 struct vnode *vp; 670 struct vm_object *obj; 671 vm_page_t pga; 672 struct socket *so; 673 #ifdef KERN_TLS 674 struct ktls_session *tls; 675 #endif 676 struct mbuf *m, *mh, *mhtail; 677 struct sf_buf *sf; 678 struct shmfd *shmfd; 679 struct sendfile_sync *sfs; 680 struct vattr va; 681 off_t off, sbytes, rem, obj_size; 682 int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr; 683 #ifdef KERN_TLS 684 int tls_enq_cnt; 685 #endif 686 bool use_ext_pgs; 687 688 obj = NULL; 689 so = NULL; 690 m = mh = NULL; 691 sfs = NULL; 692 #ifdef KERN_TLS 693 tls = NULL; 694 #endif 695 hdrlen = sbytes = 0; 696 softerr = 0; 697 use_ext_pgs = false; 698 699 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize); 700 if (error != 0) 701 return (error); 702 703 error = sendfile_getsock(td, sockfd, &sock_fp, &so); 704 if (error != 0) 705 goto out; 706 707 #ifdef MAC 708 error = mac_socket_check_send(td->td_ucred, so); 709 if (error != 0) 710 goto out; 711 #endif 712 713 SFSTAT_INC(sf_syscalls); 714 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags)); 715 716 if (flags & SF_SYNC) { 717 sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO); 718 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF); 719 cv_init(&sfs->cv, "sendfile"); 720 sfs->waiting = true; 721 } 722 723 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset; 724 725 /* 726 * Protect against multiple writers to the socket. 727 * 728 * XXXRW: Historically this has assumed non-interruptibility, so now 729 * we implement that, but possibly shouldn't. 730 */ 731 (void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR); 732 #ifdef KERN_TLS 733 tls = ktls_hold(so->so_snd.sb_tls_info); 734 #endif 735 736 /* 737 * Loop through the pages of the file, starting with the requested 738 * offset. Get a file page (do I/O if necessary), map the file page 739 * into an sf_buf, attach an mbuf header to the sf_buf, and queue 740 * it on the socket. 741 * This is done in two loops. The inner loop turns as many pages 742 * as it can, up to available socket buffer space, without blocking 743 * into mbufs to have it bulk delivered into the socket send buffer. 744 * The outer loop checks the state and available space of the socket 745 * and takes care of the overall progress. 746 */ 747 for (off = offset; rem > 0; ) { 748 struct sf_io *sfio; 749 vm_page_t *pa; 750 struct mbuf *m0, *mtail; 751 int nios, space, npages, rhpages; 752 753 mtail = NULL; 754 /* 755 * Check the socket state for ongoing connection, 756 * no errors and space in socket buffer. 757 * If space is low allow for the remainder of the 758 * file to be processed if it fits the socket buffer. 759 * Otherwise block in waiting for sufficient space 760 * to proceed, or if the socket is nonblocking, return 761 * to userland with EAGAIN while reporting how far 762 * we've come. 763 * We wait until the socket buffer has significant free 764 * space to do bulk sends. This makes good use of file 765 * system read ahead and allows packet segmentation 766 * offloading hardware to take over lots of work. If 767 * we were not careful here we would send off only one 768 * sfbuf at a time. 769 */ 770 SOCKBUF_LOCK(&so->so_snd); 771 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2) 772 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2; 773 retry_space: 774 if (so->so_snd.sb_state & SBS_CANTSENDMORE) { 775 error = EPIPE; 776 SOCKBUF_UNLOCK(&so->so_snd); 777 goto done; 778 } else if (so->so_error) { 779 error = so->so_error; 780 so->so_error = 0; 781 SOCKBUF_UNLOCK(&so->so_snd); 782 goto done; 783 } 784 if ((so->so_state & SS_ISCONNECTED) == 0) { 785 SOCKBUF_UNLOCK(&so->so_snd); 786 error = ENOTCONN; 787 goto done; 788 } 789 790 space = sbspace(&so->so_snd); 791 if (space < rem && 792 (space <= 0 || 793 space < so->so_snd.sb_lowat)) { 794 if (so->so_state & SS_NBIO) { 795 SOCKBUF_UNLOCK(&so->so_snd); 796 error = EAGAIN; 797 goto done; 798 } 799 /* 800 * sbwait drops the lock while sleeping. 801 * When we loop back to retry_space the 802 * state may have changed and we retest 803 * for it. 804 */ 805 error = sbwait(&so->so_snd); 806 /* 807 * An error from sbwait usually indicates that we've 808 * been interrupted by a signal. If we've sent anything 809 * then return bytes sent, otherwise return the error. 810 */ 811 if (error != 0) { 812 SOCKBUF_UNLOCK(&so->so_snd); 813 goto done; 814 } 815 goto retry_space; 816 } 817 SOCKBUF_UNLOCK(&so->so_snd); 818 819 /* 820 * At the beginning of the first loop check if any headers 821 * are specified and copy them into mbufs. Reduce space in 822 * the socket buffer by the size of the header mbuf chain. 823 * Clear hdr_uio here and hdrlen at the end of the first loop. 824 */ 825 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) { 826 hdr_uio->uio_td = td; 827 hdr_uio->uio_rw = UIO_WRITE; 828 #ifdef KERN_TLS 829 if (tls != NULL) 830 mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 831 tls->params.max_frame_len, M_EXTPG); 832 else 833 #endif 834 mh = m_uiotombuf(hdr_uio, M_WAITOK, 835 space, 0, 0); 836 hdrlen = m_length(mh, &mhtail); 837 space -= hdrlen; 838 /* 839 * If header consumed all the socket buffer space, 840 * don't waste CPU cycles and jump to the end. 841 */ 842 if (space == 0) { 843 sfio = NULL; 844 nios = 0; 845 goto prepend_header; 846 } 847 hdr_uio = NULL; 848 } 849 850 if (vp != NULL) { 851 error = vn_lock(vp, LK_SHARED); 852 if (error != 0) 853 goto done; 854 error = VOP_GETATTR(vp, &va, td->td_ucred); 855 if (error != 0 || off >= va.va_size) { 856 VOP_UNLOCK(vp); 857 goto done; 858 } 859 if (va.va_size != obj_size) { 860 obj_size = va.va_size; 861 rem = nbytes ? 862 omin(nbytes + offset, obj_size) : obj_size; 863 rem -= off; 864 } 865 } 866 867 if (space > rem) 868 space = rem; 869 else if (space > PAGE_SIZE) { 870 /* 871 * Use page boundaries when possible for large 872 * requests. 873 */ 874 if (off & PAGE_MASK) 875 space -= (PAGE_SIZE - (off & PAGE_MASK)); 876 space = trunc_page(space); 877 if (off & PAGE_MASK) 878 space += (PAGE_SIZE - (off & PAGE_MASK)); 879 } 880 881 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE); 882 883 /* 884 * Calculate maximum allowed number of pages for readahead 885 * at this iteration. If SF_USER_READAHEAD was set, we don't 886 * do any heuristics and use exactly the value supplied by 887 * application. Otherwise, we allow readahead up to "rem". 888 * If application wants more, let it be, but there is no 889 * reason to go above maxphys. Also check against "obj_size", 890 * since vm_pager_has_page() can hint beyond EOF. 891 */ 892 if (flags & SF_USER_READAHEAD) { 893 rhpages = SF_READAHEAD(flags); 894 } else { 895 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) - 896 npages; 897 rhpages += SF_READAHEAD(flags); 898 } 899 rhpages = min(howmany(maxphys, PAGE_SIZE), rhpages); 900 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) - 901 npages, rhpages); 902 903 sfio = malloc(sizeof(struct sf_io) + 904 npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK); 905 refcount_init(&sfio->nios, 1); 906 sfio->obj = obj; 907 sfio->error = 0; 908 sfio->m = NULL; 909 sfio->npages = npages; 910 #ifdef KERN_TLS 911 /* 912 * This doesn't use ktls_hold() because sfio->m will 913 * also have a reference on 'tls' that will be valid 914 * for all of sfio's lifetime. 915 */ 916 sfio->tls = tls; 917 #endif 918 vm_object_pip_add(obj, 1); 919 error = sendfile_swapin(obj, sfio, &nios, off, space, rhpages, 920 flags); 921 if (error != 0) { 922 if (vp != NULL) 923 VOP_UNLOCK(vp); 924 sendfile_iodone(sfio, NULL, 0, error); 925 goto done; 926 } 927 928 /* 929 * Loop and construct maximum sized mbuf chain to be bulk 930 * dumped into socket buffer. 931 */ 932 pa = sfio->pa; 933 934 /* 935 * Use unmapped mbufs if enabled for TCP. Unmapped 936 * bufs are restricted to TCP as that is what has been 937 * tested. In particular, unmapped mbufs have not 938 * been tested with UNIX-domain sockets. 939 * 940 * TLS frames always require unmapped mbufs. 941 */ 942 if ((mb_use_ext_pgs && 943 so->so_proto->pr_protocol == IPPROTO_TCP) 944 #ifdef KERN_TLS 945 || tls != NULL 946 #endif 947 ) { 948 use_ext_pgs = true; 949 #ifdef KERN_TLS 950 if (tls != NULL) 951 max_pgs = num_pages(tls->params.max_frame_len); 952 else 953 #endif 954 max_pgs = MBUF_PEXT_MAX_PGS; 955 956 /* Start at last index, to wrap on first use. */ 957 ext_pgs_idx = max_pgs - 1; 958 } 959 960 for (int i = 0; i < npages; i++) { 961 /* 962 * If a page wasn't grabbed successfully, then 963 * trim the array. Can happen only with SF_NODISKIO. 964 */ 965 if (pa[i] == NULL) { 966 SFSTAT_INC(sf_busy); 967 fixspace(npages, i, off, &space); 968 sfio->npages = i; 969 softerr = EBUSY; 970 break; 971 } 972 pga = pa[i]; 973 if (pga == bogus_page) 974 pga = vm_page_relookup(obj, sfio->pindex0 + i); 975 976 if (use_ext_pgs) { 977 off_t xfs; 978 979 ext_pgs_idx++; 980 if (ext_pgs_idx == max_pgs) { 981 m0 = mb_alloc_ext_pgs(M_WAITOK, 982 sendfile_free_mext_pg); 983 984 if (flags & SF_NOCACHE) { 985 m0->m_ext.ext_flags |= 986 EXT_FLAG_NOCACHE; 987 988 /* 989 * See comment below regarding 990 * ignoring SF_NOCACHE for the 991 * last page. 992 */ 993 if ((npages - i <= max_pgs) && 994 ((off + space) & PAGE_MASK) && 995 (rem > space || rhpages > 0)) 996 m0->m_ext.ext_flags |= 997 EXT_FLAG_CACHE_LAST; 998 } 999 if (sfs != NULL) { 1000 m0->m_ext.ext_flags |= 1001 EXT_FLAG_SYNC; 1002 m0->m_ext.ext_arg1 = sfs; 1003 mtx_lock(&sfs->mtx); 1004 sfs->count++; 1005 mtx_unlock(&sfs->mtx); 1006 } 1007 ext_pgs_idx = 0; 1008 1009 /* Append to mbuf chain. */ 1010 if (mtail != NULL) 1011 mtail->m_next = m0; 1012 else 1013 m = m0; 1014 mtail = m0; 1015 m0->m_epg_1st_off = 1016 vmoff(i, off) & PAGE_MASK; 1017 } 1018 if (nios) { 1019 mtail->m_flags |= M_NOTREADY; 1020 m0->m_epg_nrdy++; 1021 } 1022 1023 m0->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga); 1024 m0->m_epg_npgs++; 1025 xfs = xfsize(i, npages, off, space); 1026 m0->m_epg_last_len = xfs; 1027 MBUF_EXT_PGS_ASSERT_SANITY(m0); 1028 mtail->m_len += xfs; 1029 mtail->m_ext.ext_size += PAGE_SIZE; 1030 continue; 1031 } 1032 1033 /* 1034 * Get a sendfile buf. When allocating the 1035 * first buffer for mbuf chain, we usually 1036 * wait as long as necessary, but this wait 1037 * can be interrupted. For consequent 1038 * buffers, do not sleep, since several 1039 * threads might exhaust the buffers and then 1040 * deadlock. 1041 */ 1042 sf = sf_buf_alloc(pga, 1043 m != NULL ? SFB_NOWAIT : SFB_CATCH); 1044 if (sf == NULL) { 1045 SFSTAT_INC(sf_allocfail); 1046 sendfile_iowait(sfio, "sfnosf"); 1047 for (int j = i; j < npages; j++) { 1048 vm_page_unwire(pa[j], PQ_INACTIVE); 1049 pa[j] = NULL; 1050 } 1051 if (m == NULL) 1052 softerr = ENOBUFS; 1053 fixspace(npages, i, off, &space); 1054 sfio->npages = i; 1055 break; 1056 } 1057 1058 m0 = m_get(M_WAITOK, MT_DATA); 1059 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf); 1060 m0->m_ext.ext_size = PAGE_SIZE; 1061 m0->m_ext.ext_arg1 = sf; 1062 m0->m_ext.ext_type = EXT_SFBUF; 1063 m0->m_ext.ext_flags = EXT_FLAG_EMBREF; 1064 m0->m_ext.ext_free = sendfile_free_mext; 1065 /* 1066 * SF_NOCACHE sets the page as being freed upon send. 1067 * However, we ignore it for the last page in 'space', 1068 * if the page is truncated, and we got more data to 1069 * send (rem > space), or if we have readahead 1070 * configured (rhpages > 0). 1071 */ 1072 if ((flags & SF_NOCACHE) && 1073 (i != npages - 1 || 1074 !((off + space) & PAGE_MASK) || 1075 !(rem > space || rhpages > 0))) 1076 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE; 1077 if (sfs != NULL) { 1078 m0->m_ext.ext_flags |= EXT_FLAG_SYNC; 1079 m0->m_ext.ext_arg2 = sfs; 1080 mtx_lock(&sfs->mtx); 1081 sfs->count++; 1082 mtx_unlock(&sfs->mtx); 1083 } 1084 m0->m_ext.ext_count = 1; 1085 m0->m_flags |= (M_EXT | M_RDONLY); 1086 if (nios) 1087 m0->m_flags |= M_NOTREADY; 1088 m0->m_data = (char *)sf_buf_kva(sf) + 1089 (vmoff(i, off) & PAGE_MASK); 1090 m0->m_len = xfsize(i, npages, off, space); 1091 1092 /* Append to mbuf chain. */ 1093 if (mtail != NULL) 1094 mtail->m_next = m0; 1095 else 1096 m = m0; 1097 mtail = m0; 1098 } 1099 1100 if (vp != NULL) 1101 VOP_UNLOCK(vp); 1102 1103 /* Keep track of bytes processed. */ 1104 off += space; 1105 rem -= space; 1106 1107 /* 1108 * Prepend header, if any. Save pointer to first mbuf 1109 * with a page. 1110 */ 1111 if (hdrlen) { 1112 prepend_header: 1113 m0 = mhtail->m_next = m; 1114 m = mh; 1115 mh = NULL; 1116 } else 1117 m0 = m; 1118 1119 if (m == NULL) { 1120 KASSERT(softerr, ("%s: m NULL, no error", __func__)); 1121 error = softerr; 1122 sendfile_iodone(sfio, NULL, 0, 0); 1123 goto done; 1124 } 1125 1126 /* Add the buffer chain to the socket buffer. */ 1127 KASSERT(m_length(m, NULL) == space + hdrlen, 1128 ("%s: mlen %u space %d hdrlen %d", 1129 __func__, m_length(m, NULL), space, hdrlen)); 1130 1131 CURVNET_SET(so->so_vnet); 1132 #ifdef KERN_TLS 1133 if (tls != NULL) 1134 ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP); 1135 #endif 1136 if (nios == 0) { 1137 /* 1138 * If sendfile_swapin() didn't initiate any I/Os, 1139 * which happens if all data is cached in VM, or if 1140 * the header consumed all socket buffer space and 1141 * sfio is NULL, then we can send data right now 1142 * without the PRUS_NOTREADY flag. 1143 */ 1144 if (sfio != NULL) 1145 sendfile_iodone(sfio, NULL, 0, 0); 1146 #ifdef KERN_TLS 1147 if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) { 1148 error = (*so->so_proto->pr_usrreqs->pru_send) 1149 (so, PRUS_NOTREADY, m, NULL, NULL, td); 1150 soref(so); 1151 ktls_enqueue(m, so, tls_enq_cnt); 1152 } else 1153 #endif 1154 error = (*so->so_proto->pr_usrreqs->pru_send) 1155 (so, 0, m, NULL, NULL, td); 1156 } else { 1157 sfio->so = so; 1158 sfio->m = m0; 1159 soref(so); 1160 error = (*so->so_proto->pr_usrreqs->pru_send) 1161 (so, PRUS_NOTREADY, m, NULL, NULL, td); 1162 sendfile_iodone(sfio, NULL, 0, 0); 1163 } 1164 CURVNET_RESTORE(); 1165 1166 m = NULL; /* pru_send always consumes */ 1167 if (error) 1168 goto done; 1169 sbytes += space + hdrlen; 1170 if (hdrlen) 1171 hdrlen = 0; 1172 if (softerr) { 1173 error = softerr; 1174 goto done; 1175 } 1176 } 1177 1178 /* 1179 * Send trailers. Wimp out and use writev(2). 1180 */ 1181 if (trl_uio != NULL) { 1182 sbunlock(&so->so_snd); 1183 error = kern_writev(td, sockfd, trl_uio); 1184 if (error == 0) 1185 sbytes += td->td_retval[0]; 1186 goto out; 1187 } 1188 1189 done: 1190 sbunlock(&so->so_snd); 1191 out: 1192 /* 1193 * If there was no error we have to clear td->td_retval[0] 1194 * because it may have been set by writev. 1195 */ 1196 if (error == 0) { 1197 td->td_retval[0] = 0; 1198 } 1199 if (sent != NULL) { 1200 (*sent) = sbytes; 1201 } 1202 if (obj != NULL) 1203 vm_object_deallocate(obj); 1204 if (so) 1205 fdrop(sock_fp, td); 1206 if (m) 1207 m_freem(m); 1208 if (mh) 1209 m_freem(mh); 1210 1211 if (sfs != NULL) { 1212 mtx_lock(&sfs->mtx); 1213 if (sfs->count != 0) 1214 error = cv_wait_sig(&sfs->cv, &sfs->mtx); 1215 if (sfs->count == 0) { 1216 sendfile_sync_destroy(sfs); 1217 } else { 1218 sfs->waiting = false; 1219 mtx_unlock(&sfs->mtx); 1220 } 1221 } 1222 #ifdef KERN_TLS 1223 if (tls != NULL) 1224 ktls_free(tls); 1225 #endif 1226 1227 if (error == ERESTART) 1228 error = EINTR; 1229 1230 return (error); 1231 } 1232 1233 static int 1234 sendfile(struct thread *td, struct sendfile_args *uap, int compat) 1235 { 1236 struct sf_hdtr hdtr; 1237 struct uio *hdr_uio, *trl_uio; 1238 struct file *fp; 1239 off_t sbytes; 1240 int error; 1241 1242 /* 1243 * File offset must be positive. If it goes beyond EOF 1244 * we send only the header/trailer and no payload data. 1245 */ 1246 if (uap->offset < 0) 1247 return (EINVAL); 1248 1249 sbytes = 0; 1250 hdr_uio = trl_uio = NULL; 1251 1252 if (uap->hdtr != NULL) { 1253 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr)); 1254 if (error != 0) 1255 goto out; 1256 if (hdtr.headers != NULL) { 1257 error = copyinuio(hdtr.headers, hdtr.hdr_cnt, 1258 &hdr_uio); 1259 if (error != 0) 1260 goto out; 1261 #ifdef COMPAT_FREEBSD4 1262 /* 1263 * In FreeBSD < 5.0 the nbytes to send also included 1264 * the header. If compat is specified subtract the 1265 * header size from nbytes. 1266 */ 1267 if (compat) { 1268 if (uap->nbytes > hdr_uio->uio_resid) 1269 uap->nbytes -= hdr_uio->uio_resid; 1270 else 1271 uap->nbytes = 0; 1272 } 1273 #endif 1274 } 1275 if (hdtr.trailers != NULL) { 1276 error = copyinuio(hdtr.trailers, hdtr.trl_cnt, 1277 &trl_uio); 1278 if (error != 0) 1279 goto out; 1280 } 1281 } 1282 1283 AUDIT_ARG_FD(uap->fd); 1284 1285 /* 1286 * sendfile(2) can start at any offset within a file so we require 1287 * CAP_READ+CAP_SEEK = CAP_PREAD. 1288 */ 1289 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0) 1290 goto out; 1291 1292 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset, 1293 uap->nbytes, &sbytes, uap->flags, td); 1294 fdrop(fp, td); 1295 1296 if (uap->sbytes != NULL) 1297 copyout(&sbytes, uap->sbytes, sizeof(off_t)); 1298 1299 out: 1300 free(hdr_uio, M_IOV); 1301 free(trl_uio, M_IOV); 1302 return (error); 1303 } 1304 1305 /* 1306 * sendfile(2) 1307 * 1308 * int sendfile(int fd, int s, off_t offset, size_t nbytes, 1309 * struct sf_hdtr *hdtr, off_t *sbytes, int flags) 1310 * 1311 * Send a file specified by 'fd' and starting at 'offset' to a socket 1312 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes == 1313 * 0. Optionally add a header and/or trailer to the socket output. If 1314 * specified, write the total number of bytes sent into *sbytes. 1315 */ 1316 int 1317 sys_sendfile(struct thread *td, struct sendfile_args *uap) 1318 { 1319 1320 return (sendfile(td, uap, 0)); 1321 } 1322 1323 #ifdef COMPAT_FREEBSD4 1324 int 1325 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap) 1326 { 1327 struct sendfile_args args; 1328 1329 args.fd = uap->fd; 1330 args.s = uap->s; 1331 args.offset = uap->offset; 1332 args.nbytes = uap->nbytes; 1333 args.hdtr = uap->hdtr; 1334 args.sbytes = uap->sbytes; 1335 args.flags = uap->flags; 1336 1337 return (sendfile(td, &args, 1)); 1338 } 1339 #endif /* COMPAT_FREEBSD4 */ 1340