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 }; 110 111 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)]; 112 113 static void 114 sfstat_init(const void *unused) 115 { 116 117 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t), 118 M_WAITOK); 119 } 120 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL); 121 122 static int 123 sfstat_sysctl(SYSCTL_HANDLER_ARGS) 124 { 125 struct sfstat s; 126 127 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t)); 128 if (req->newptr) 129 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t)); 130 return (SYSCTL_OUT(req, &s, sizeof(s))); 131 } 132 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, 133 CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0, 134 sfstat_sysctl, "I", 135 "sendfile statistics"); 136 137 static void 138 sendfile_free_mext(struct mbuf *m) 139 { 140 struct sf_buf *sf; 141 vm_page_t pg; 142 int flags; 143 144 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF, 145 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m)); 146 147 sf = m->m_ext.ext_arg1; 148 pg = sf_buf_page(sf); 149 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0; 150 151 sf_buf_free(sf); 152 vm_page_release(pg, flags); 153 154 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) { 155 struct sendfile_sync *sfs = m->m_ext.ext_arg2; 156 157 mtx_lock(&sfs->mtx); 158 KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0")); 159 if (--sfs->count == 0) 160 cv_signal(&sfs->cv); 161 mtx_unlock(&sfs->mtx); 162 } 163 } 164 165 static void 166 sendfile_free_mext_pg(struct mbuf *m) 167 { 168 struct mbuf_ext_pgs *ext_pgs; 169 vm_page_t pg; 170 int flags, i; 171 bool cache_last; 172 173 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_PGS, 174 ("%s: m %p !M_EXT or !EXT_PGS", __func__, m)); 175 176 cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST; 177 ext_pgs = &m->m_ext_pgs; 178 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0; 179 180 for (i = 0; i < ext_pgs->npgs; i++) { 181 if (cache_last && i == ext_pgs->npgs - 1) 182 flags = 0; 183 pg = PHYS_TO_VM_PAGE(ext_pgs->m_epg_pa[i]); 184 vm_page_release(pg, flags); 185 } 186 187 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) { 188 struct sendfile_sync *sfs = m->m_ext.ext_arg1; 189 190 mtx_lock(&sfs->mtx); 191 KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0")); 192 if (--sfs->count == 0) 193 cv_signal(&sfs->cv); 194 mtx_unlock(&sfs->mtx); 195 } 196 } 197 198 /* 199 * Helper function to calculate how much data to put into page i of n. 200 * Only first and last pages are special. 201 */ 202 static inline off_t 203 xfsize(int i, int n, off_t off, off_t len) 204 { 205 206 if (i == 0) 207 return (omin(PAGE_SIZE - (off & PAGE_MASK), len)); 208 209 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0) 210 return ((off + len) & PAGE_MASK); 211 212 return (PAGE_SIZE); 213 } 214 215 /* 216 * Helper function to get offset within object for i page. 217 */ 218 static inline vm_ooffset_t 219 vmoff(int i, off_t off) 220 { 221 222 if (i == 0) 223 return ((vm_ooffset_t)off); 224 225 return (trunc_page(off + i * PAGE_SIZE)); 226 } 227 228 /* 229 * Helper function used when allocation of a page or sf_buf failed. 230 * Pretend as if we don't have enough space, subtract xfsize() of 231 * all pages that failed. 232 */ 233 static inline void 234 fixspace(int old, int new, off_t off, int *space) 235 { 236 237 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new)); 238 239 /* Subtract last one. */ 240 *space -= xfsize(old - 1, old, off, *space); 241 old--; 242 243 if (new == old) 244 /* There was only one page. */ 245 return; 246 247 /* Subtract first one. */ 248 if (new == 0) { 249 *space -= xfsize(0, old, off, *space); 250 new++; 251 } 252 253 /* Rest of pages are full sized. */ 254 *space -= (old - new) * PAGE_SIZE; 255 256 KASSERT(*space >= 0, ("%s: space went backwards", __func__)); 257 } 258 259 /* 260 * Wait for all in-flight ios to complete, we must not unwire pages 261 * under them. 262 */ 263 static void 264 sendfile_iowait(struct sf_io *sfio, const char *wmesg) 265 { 266 while (atomic_load_int(&sfio->nios) != 1) 267 pause(wmesg, 1); 268 } 269 270 /* 271 * I/O completion callback. 272 */ 273 static void 274 sendfile_iodone(void *arg, vm_page_t *pa, int count, int error) 275 { 276 struct sf_io *sfio = arg; 277 struct socket *so; 278 int i; 279 280 if (error != 0) { 281 sfio->error = error; 282 /* 283 * Restore of the pg[] elements is done by 284 * sendfile_swapin(). 285 */ 286 } else { 287 /* 288 * Restore the valid page pointers. They are already 289 * unbusied, but still wired. For error != 0 case, 290 * sendfile_swapin() handles unbusy. 291 * 292 * XXXKIB since pages are only wired, and we do not 293 * own the object lock, other users might have 294 * invalidated them in meantime. Similarly, after we 295 * unbusied the swapped-in pages, they can become 296 * invalid under us. 297 */ 298 MPASS(count == 0 || pa[0] != bogus_page); 299 for (i = 0; i < count; i++) { 300 if (pa[i] == bogus_page) { 301 sfio->pa[(pa[0]->pindex - sfio->pindex0) + i] = 302 pa[i] = vm_page_relookup(sfio->obj, 303 pa[0]->pindex + i); 304 KASSERT(pa[i] != NULL, 305 ("%s: page %p[%d] disappeared", 306 __func__, pa, i)); 307 } else { 308 vm_page_xunbusy_unchecked(pa[i]); 309 } 310 } 311 } 312 313 if (!refcount_release(&sfio->nios)) 314 return; 315 316 #ifdef INVARIANTS 317 for (i = 1; i < sfio->npages; i++) { 318 if (sfio->pa[i] == NULL) 319 break; 320 KASSERT(vm_page_wired(sfio->pa[i]), 321 ("sfio %p page %d %p not wired", sfio, i, sfio->pa[i])); 322 if (i == 0) 323 continue; 324 KASSERT(sfio->pa[0]->object == sfio->pa[i]->object, 325 ("sfio %p page %d %p wrong owner %p %p", sfio, i, 326 sfio->pa[i], sfio->pa[0]->object, sfio->pa[i]->object)); 327 KASSERT(sfio->pa[0]->pindex + i == sfio->pa[i]->pindex, 328 ("sfio %p page %d %p wrong index %jx %jx", sfio, i, 329 sfio->pa[i], (uintmax_t)sfio->pa[0]->pindex, 330 (uintmax_t)sfio->pa[i]->pindex)); 331 } 332 #endif 333 334 vm_object_pip_wakeup(sfio->obj); 335 336 if (sfio->m == NULL) { 337 /* 338 * Either I/O operation failed, or we failed to allocate 339 * buffers, or we bailed out on first busy page, or we 340 * succeeded filling the request without any I/Os. Anyway, 341 * pru_send hadn't been executed - nothing had been sent 342 * to the socket yet. 343 */ 344 MPASS((curthread->td_pflags & TDP_KTHREAD) == 0); 345 free(sfio, M_SENDFILE); 346 return; 347 } 348 349 #if defined(KERN_TLS) && defined(INVARIANTS) 350 if ((sfio->m->m_flags & M_EXT) != 0 && 351 sfio->m->m_ext.ext_type == EXT_PGS) 352 KASSERT(sfio->tls == sfio->m->m_ext_pgs.tls, 353 ("TLS session mismatch")); 354 else 355 KASSERT(sfio->tls == NULL, 356 ("non-ext_pgs mbuf with TLS session")); 357 #endif 358 so = sfio->so; 359 CURVNET_SET(so->so_vnet); 360 if (__predict_false(sfio->error)) { 361 /* 362 * I/O operation failed. The state of data in the socket 363 * is now inconsistent, and all what we can do is to tear 364 * it down. Protocol abort method would tear down protocol 365 * state, free all ready mbufs and detach not ready ones. 366 * We will free the mbufs corresponding to this I/O manually. 367 * 368 * The socket would be marked with EIO and made available 369 * for read, so that application receives EIO on next 370 * syscall and eventually closes the socket. 371 */ 372 so->so_proto->pr_usrreqs->pru_abort(so); 373 so->so_error = EIO; 374 375 mb_free_notready(sfio->m, sfio->npages); 376 #ifdef KERN_TLS 377 } else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) { 378 /* 379 * I/O operation is complete, but we still need to 380 * encrypt. We cannot do this in the interrupt thread 381 * of the disk controller, so forward the mbufs to a 382 * different thread. 383 * 384 * Donate the socket reference from sfio to rather 385 * than explicitly invoking soref(). 386 */ 387 ktls_enqueue(sfio->m, so, sfio->npages); 388 goto out_with_ref; 389 #endif 390 } else 391 (void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m, 392 sfio->npages); 393 394 SOCK_LOCK(so); 395 sorele(so); 396 #ifdef KERN_TLS 397 out_with_ref: 398 #endif 399 CURVNET_RESTORE(); 400 free(sfio, M_SENDFILE); 401 } 402 403 /* 404 * Iterate through pages vector and request paging for non-valid pages. 405 */ 406 static int 407 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off, 408 off_t len, int npages, int rhpages, int flags) 409 { 410 vm_page_t *pa; 411 int a, count, count1, grabbed, i, j, rv; 412 413 pa = sfio->pa; 414 *nios = 0; 415 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0; 416 sfio->pindex0 = OFF_TO_IDX(off); 417 418 /* 419 * First grab all the pages and wire them. Note that we grab 420 * only required pages. Readahead pages are dealt with later. 421 */ 422 grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off), 423 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages); 424 if (grabbed < npages) { 425 for (int i = grabbed; i < npages; i++) 426 pa[i] = NULL; 427 npages = grabbed; 428 rhpages = 0; 429 } 430 431 for (i = 0; i < npages;) { 432 /* Skip valid pages. */ 433 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK, 434 xfsize(i, npages, off, len))) { 435 vm_page_xunbusy(pa[i]); 436 SFSTAT_INC(sf_pages_valid); 437 i++; 438 continue; 439 } 440 441 /* 442 * Next page is invalid. Check if it belongs to pager. It 443 * may not be there, which is a regular situation for shmem 444 * pager. For vnode pager this happens only in case of 445 * a sparse file. 446 * 447 * Important feature of vm_pager_has_page() is the hint 448 * stored in 'a', about how many pages we can pagein after 449 * this page in a single I/O. 450 */ 451 VM_OBJECT_RLOCK(obj); 452 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL, 453 &a)) { 454 VM_OBJECT_RUNLOCK(obj); 455 pmap_zero_page(pa[i]); 456 vm_page_valid(pa[i]); 457 MPASS(pa[i]->dirty == 0); 458 vm_page_xunbusy(pa[i]); 459 i++; 460 continue; 461 } 462 VM_OBJECT_RUNLOCK(obj); 463 464 /* 465 * We want to pagein as many pages as possible, limited only 466 * by the 'a' hint and actual request. 467 */ 468 count = min(a + 1, npages - i); 469 470 /* 471 * We should not pagein into a valid page because 472 * there might be still unfinished write tracked by 473 * e.g. a buffer, thus we substitute any valid pages 474 * with the bogus one. 475 * 476 * We must not leave around xbusy pages which are not 477 * part of the run passed to vm_pager_getpages(), 478 * otherwise pager might deadlock waiting for the busy 479 * status of the page, e.g. if it constitues the 480 * buffer needed to validate other page. 481 * 482 * First trim the end of the run consisting of the 483 * valid pages, then replace the rest of the valid 484 * with bogus. 485 */ 486 count1 = count; 487 for (j = i + count - 1; j > i; j--) { 488 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 489 xfsize(j, npages, off, len))) { 490 vm_page_xunbusy(pa[j]); 491 SFSTAT_INC(sf_pages_valid); 492 count--; 493 } else { 494 break; 495 } 496 } 497 498 /* 499 * The last page in the run pa[i + count - 1] is 500 * guaranteed to be invalid by the trim above, so it 501 * is not replaced with bogus, thus -1 in the loop end 502 * condition. 503 */ 504 MPASS(pa[i + count - 1]->valid != VM_PAGE_BITS_ALL); 505 for (j = i + 1; j < i + count - 1; j++) { 506 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, 507 xfsize(j, npages, off, len))) { 508 vm_page_xunbusy(pa[j]); 509 SFSTAT_INC(sf_pages_valid); 510 SFSTAT_INC(sf_pages_bogus); 511 pa[j] = bogus_page; 512 } 513 } 514 515 refcount_acquire(&sfio->nios); 516 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL, 517 i + count == npages ? &rhpages : NULL, 518 &sendfile_iodone, sfio); 519 if (__predict_false(rv != VM_PAGER_OK)) { 520 sendfile_iowait(sfio, "sferrio"); 521 522 /* 523 * Perform full pages recovery before returning EIO. 524 * Pages from 0 to npages are wired. 525 * Pages from (i + 1) to (i + count - 1) may be 526 * substituted to bogus page, and not busied. 527 * Pages from (i + count) to (i + count1 - 1) are 528 * not busied. 529 * Rest of the pages from i to npages are busied. 530 */ 531 for (j = 0; j < npages; j++) { 532 if (j >= i + count && j < i + count1) 533 ; 534 else if (j > i && j < i + count - 1 && 535 pa[j] == bogus_page) 536 pa[j] = vm_page_relookup(obj, 537 OFF_TO_IDX(vmoff(j, off))); 538 else if (j >= i) 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 } 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_ext_pgs *ext_pgs; 678 struct mbuf *m, *mh, *mhtail; 679 struct sf_buf *sf; 680 struct shmfd *shmfd; 681 struct sendfile_sync *sfs; 682 struct vattr va; 683 off_t off, sbytes, rem, obj_size; 684 int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr; 685 #ifdef KERN_TLS 686 int tls_enq_cnt; 687 #endif 688 bool use_ext_pgs; 689 690 obj = NULL; 691 so = NULL; 692 m = mh = NULL; 693 sfs = NULL; 694 #ifdef KERN_TLS 695 tls = NULL; 696 #endif 697 hdrlen = sbytes = 0; 698 softerr = 0; 699 use_ext_pgs = false; 700 701 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize); 702 if (error != 0) 703 return (error); 704 705 error = sendfile_getsock(td, sockfd, &sock_fp, &so); 706 if (error != 0) 707 goto out; 708 709 #ifdef MAC 710 error = mac_socket_check_send(td->td_ucred, so); 711 if (error != 0) 712 goto out; 713 #endif 714 715 SFSTAT_INC(sf_syscalls); 716 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags)); 717 718 if (flags & SF_SYNC) { 719 sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO); 720 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF); 721 cv_init(&sfs->cv, "sendfile"); 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_NOMAP); 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 #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, npages, 920 rhpages, 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 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 if (m0->m_ext.ext_type == 1003 EXT_PGS) 1004 m0->m_ext.ext_arg1 = 1005 sfs; 1006 else 1007 m0->m_ext.ext_arg2 = 1008 sfs; 1009 mtx_lock(&sfs->mtx); 1010 sfs->count++; 1011 mtx_unlock(&sfs->mtx); 1012 } 1013 ext_pgs = &m0->m_ext_pgs; 1014 ext_pgs_idx = 0; 1015 1016 /* Append to mbuf chain. */ 1017 if (mtail != NULL) 1018 mtail->m_next = m0; 1019 else 1020 m = m0; 1021 mtail = m0; 1022 ext_pgs->first_pg_off = 1023 vmoff(i, off) & PAGE_MASK; 1024 } 1025 if (nios) { 1026 mtail->m_flags |= M_NOTREADY; 1027 ext_pgs->nrdy++; 1028 } 1029 1030 ext_pgs->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga); 1031 ext_pgs->npgs++; 1032 xfs = xfsize(i, npages, off, space); 1033 ext_pgs->last_pg_len = xfs; 1034 MBUF_EXT_PGS_ASSERT_SANITY(ext_pgs); 1035 mtail->m_len += xfs; 1036 mtail->m_ext.ext_size += PAGE_SIZE; 1037 continue; 1038 } 1039 1040 /* 1041 * Get a sendfile buf. When allocating the 1042 * first buffer for mbuf chain, we usually 1043 * wait as long as necessary, but this wait 1044 * can be interrupted. For consequent 1045 * buffers, do not sleep, since several 1046 * threads might exhaust the buffers and then 1047 * deadlock. 1048 */ 1049 sf = sf_buf_alloc(pga, 1050 m != NULL ? SFB_NOWAIT : SFB_CATCH); 1051 if (sf == NULL) { 1052 SFSTAT_INC(sf_allocfail); 1053 sendfile_iowait(sfio, "sfnosf"); 1054 for (int j = i; j < npages; j++) 1055 vm_page_unwire(pa[j], PQ_INACTIVE); 1056 if (m == NULL) 1057 softerr = ENOBUFS; 1058 fixspace(npages, i, off, &space); 1059 npages = i; 1060 break; 1061 } 1062 1063 m0 = m_get(M_WAITOK, MT_DATA); 1064 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf); 1065 m0->m_ext.ext_size = PAGE_SIZE; 1066 m0->m_ext.ext_arg1 = sf; 1067 m0->m_ext.ext_type = EXT_SFBUF; 1068 m0->m_ext.ext_flags = EXT_FLAG_EMBREF; 1069 m0->m_ext.ext_free = sendfile_free_mext; 1070 /* 1071 * SF_NOCACHE sets the page as being freed upon send. 1072 * However, we ignore it for the last page in 'space', 1073 * if the page is truncated, and we got more data to 1074 * send (rem > space), or if we have readahead 1075 * configured (rhpages > 0). 1076 */ 1077 if ((flags & SF_NOCACHE) && 1078 (i != npages - 1 || 1079 !((off + space) & PAGE_MASK) || 1080 !(rem > space || rhpages > 0))) 1081 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE; 1082 if (sfs != NULL) { 1083 m0->m_ext.ext_flags |= EXT_FLAG_SYNC; 1084 if (m0->m_ext.ext_type == EXT_PGS) 1085 m0->m_ext.ext_arg1 = sfs; 1086 else 1087 m0->m_ext.ext_arg2 = sfs; 1088 m0->m_ext.ext_arg2 = sfs; 1089 mtx_lock(&sfs->mtx); 1090 sfs->count++; 1091 mtx_unlock(&sfs->mtx); 1092 } 1093 m0->m_ext.ext_count = 1; 1094 m0->m_flags |= (M_EXT | M_RDONLY); 1095 if (nios) 1096 m0->m_flags |= M_NOTREADY; 1097 m0->m_data = (char *)sf_buf_kva(sf) + 1098 (vmoff(i, off) & PAGE_MASK); 1099 m0->m_len = xfsize(i, npages, off, space); 1100 1101 /* Append to mbuf chain. */ 1102 if (mtail != NULL) 1103 mtail->m_next = m0; 1104 else 1105 m = m0; 1106 mtail = m0; 1107 } 1108 1109 if (vp != NULL) 1110 VOP_UNLOCK(vp); 1111 1112 /* Keep track of bytes processed. */ 1113 off += space; 1114 rem -= space; 1115 1116 /* 1117 * Prepend header, if any. Save pointer to first mbuf 1118 * with a page. 1119 */ 1120 if (hdrlen) { 1121 prepend_header: 1122 m0 = mhtail->m_next = m; 1123 m = mh; 1124 mh = NULL; 1125 } else 1126 m0 = m; 1127 1128 if (m == NULL) { 1129 KASSERT(softerr, ("%s: m NULL, no error", __func__)); 1130 error = softerr; 1131 sendfile_iodone(sfio, NULL, 0, 0); 1132 goto done; 1133 } 1134 1135 /* Add the buffer chain to the socket buffer. */ 1136 KASSERT(m_length(m, NULL) == space + hdrlen, 1137 ("%s: mlen %u space %d hdrlen %d", 1138 __func__, m_length(m, NULL), space, hdrlen)); 1139 1140 CURVNET_SET(so->so_vnet); 1141 #ifdef KERN_TLS 1142 if (tls != NULL) 1143 ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP); 1144 #endif 1145 if (nios == 0) { 1146 /* 1147 * If sendfile_swapin() didn't initiate any I/Os, 1148 * which happens if all data is cached in VM, or if 1149 * the header consumed all socket buffer space and 1150 * sfio is NULL, then we can send data right now 1151 * without the PRUS_NOTREADY flag. 1152 */ 1153 if (sfio != NULL) 1154 sendfile_iodone(sfio, NULL, 0, 0); 1155 #ifdef KERN_TLS 1156 if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) { 1157 error = (*so->so_proto->pr_usrreqs->pru_send) 1158 (so, PRUS_NOTREADY, m, NULL, NULL, td); 1159 soref(so); 1160 ktls_enqueue(m, so, tls_enq_cnt); 1161 } else 1162 #endif 1163 error = (*so->so_proto->pr_usrreqs->pru_send) 1164 (so, 0, m, NULL, NULL, td); 1165 } else { 1166 sfio->so = so; 1167 sfio->m = m0; 1168 sfio->npages = npages; 1169 soref(so); 1170 error = (*so->so_proto->pr_usrreqs->pru_send) 1171 (so, PRUS_NOTREADY, m, NULL, NULL, td); 1172 sendfile_iodone(sfio, NULL, 0, 0); 1173 } 1174 CURVNET_RESTORE(); 1175 1176 m = NULL; /* pru_send always consumes */ 1177 if (error) 1178 goto done; 1179 sbytes += space + hdrlen; 1180 if (hdrlen) 1181 hdrlen = 0; 1182 if (softerr) { 1183 error = softerr; 1184 goto done; 1185 } 1186 } 1187 1188 /* 1189 * Send trailers. Wimp out and use writev(2). 1190 */ 1191 if (trl_uio != NULL) { 1192 sbunlock(&so->so_snd); 1193 error = kern_writev(td, sockfd, trl_uio); 1194 if (error == 0) 1195 sbytes += td->td_retval[0]; 1196 goto out; 1197 } 1198 1199 done: 1200 sbunlock(&so->so_snd); 1201 out: 1202 /* 1203 * If there was no error we have to clear td->td_retval[0] 1204 * because it may have been set by writev. 1205 */ 1206 if (error == 0) { 1207 td->td_retval[0] = 0; 1208 } 1209 if (sent != NULL) { 1210 (*sent) = sbytes; 1211 } 1212 if (obj != NULL) 1213 vm_object_deallocate(obj); 1214 if (so) 1215 fdrop(sock_fp, td); 1216 if (m) 1217 m_freem(m); 1218 if (mh) 1219 m_freem(mh); 1220 1221 if (sfs != NULL) { 1222 mtx_lock(&sfs->mtx); 1223 if (sfs->count != 0) 1224 cv_wait(&sfs->cv, &sfs->mtx); 1225 KASSERT(sfs->count == 0, ("sendfile sync still busy")); 1226 cv_destroy(&sfs->cv); 1227 mtx_destroy(&sfs->mtx); 1228 free(sfs, M_SENDFILE); 1229 } 1230 #ifdef KERN_TLS 1231 if (tls != NULL) 1232 ktls_free(tls); 1233 #endif 1234 1235 if (error == ERESTART) 1236 error = EINTR; 1237 1238 return (error); 1239 } 1240 1241 static int 1242 sendfile(struct thread *td, struct sendfile_args *uap, int compat) 1243 { 1244 struct sf_hdtr hdtr; 1245 struct uio *hdr_uio, *trl_uio; 1246 struct file *fp; 1247 off_t sbytes; 1248 int error; 1249 1250 /* 1251 * File offset must be positive. If it goes beyond EOF 1252 * we send only the header/trailer and no payload data. 1253 */ 1254 if (uap->offset < 0) 1255 return (EINVAL); 1256 1257 sbytes = 0; 1258 hdr_uio = trl_uio = NULL; 1259 1260 if (uap->hdtr != NULL) { 1261 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr)); 1262 if (error != 0) 1263 goto out; 1264 if (hdtr.headers != NULL) { 1265 error = copyinuio(hdtr.headers, hdtr.hdr_cnt, 1266 &hdr_uio); 1267 if (error != 0) 1268 goto out; 1269 #ifdef COMPAT_FREEBSD4 1270 /* 1271 * In FreeBSD < 5.0 the nbytes to send also included 1272 * the header. If compat is specified subtract the 1273 * header size from nbytes. 1274 */ 1275 if (compat) { 1276 if (uap->nbytes > hdr_uio->uio_resid) 1277 uap->nbytes -= hdr_uio->uio_resid; 1278 else 1279 uap->nbytes = 0; 1280 } 1281 #endif 1282 } 1283 if (hdtr.trailers != NULL) { 1284 error = copyinuio(hdtr.trailers, hdtr.trl_cnt, 1285 &trl_uio); 1286 if (error != 0) 1287 goto out; 1288 } 1289 } 1290 1291 AUDIT_ARG_FD(uap->fd); 1292 1293 /* 1294 * sendfile(2) can start at any offset within a file so we require 1295 * CAP_READ+CAP_SEEK = CAP_PREAD. 1296 */ 1297 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0) 1298 goto out; 1299 1300 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset, 1301 uap->nbytes, &sbytes, uap->flags, td); 1302 fdrop(fp, td); 1303 1304 if (uap->sbytes != NULL) 1305 copyout(&sbytes, uap->sbytes, sizeof(off_t)); 1306 1307 out: 1308 free(hdr_uio, M_IOV); 1309 free(trl_uio, M_IOV); 1310 return (error); 1311 } 1312 1313 /* 1314 * sendfile(2) 1315 * 1316 * int sendfile(int fd, int s, off_t offset, size_t nbytes, 1317 * struct sf_hdtr *hdtr, off_t *sbytes, int flags) 1318 * 1319 * Send a file specified by 'fd' and starting at 'offset' to a socket 1320 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes == 1321 * 0. Optionally add a header and/or trailer to the socket output. If 1322 * specified, write the total number of bytes sent into *sbytes. 1323 */ 1324 int 1325 sys_sendfile(struct thread *td, struct sendfile_args *uap) 1326 { 1327 1328 return (sendfile(td, uap, 0)); 1329 } 1330 1331 #ifdef COMPAT_FREEBSD4 1332 int 1333 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap) 1334 { 1335 struct sendfile_args args; 1336 1337 args.fd = uap->fd; 1338 args.s = uap->s; 1339 args.offset = uap->offset; 1340 args.nbytes = uap->nbytes; 1341 args.hdtr = uap->hdtr; 1342 args.sbytes = uap->sbytes; 1343 args.flags = uap->flags; 1344 1345 return (sendfile(td, &args, 1)); 1346 } 1347 #endif /* COMPAT_FREEBSD4 */ 1348