1 /* $NetBSD: genfs_io.c,v 1.8 2008/06/04 12:41:40 ad Exp $ */ 2 3 /* 4 * Copyright (c) 1982, 1986, 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Neither the name of the University nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 */ 32 33 #include <sys/cdefs.h> 34 __KERNEL_RCSID(0, "$NetBSD: genfs_io.c,v 1.8 2008/06/04 12:41:40 ad Exp $"); 35 36 #include <sys/param.h> 37 #include <sys/systm.h> 38 #include <sys/proc.h> 39 #include <sys/kernel.h> 40 #include <sys/mount.h> 41 #include <sys/namei.h> 42 #include <sys/vnode.h> 43 #include <sys/fcntl.h> 44 #include <sys/kmem.h> 45 #include <sys/poll.h> 46 #include <sys/mman.h> 47 #include <sys/file.h> 48 #include <sys/kauth.h> 49 #include <sys/fstrans.h> 50 51 #include <miscfs/genfs/genfs.h> 52 #include <miscfs/genfs/genfs_node.h> 53 #include <miscfs/specfs/specdev.h> 54 55 #include <uvm/uvm.h> 56 #include <uvm/uvm_pager.h> 57 58 static int genfs_do_directio(struct vmspace *, vaddr_t, size_t, struct vnode *, 59 off_t, enum uio_rw); 60 static void genfs_dio_iodone(struct buf *); 61 62 static int genfs_do_io(struct vnode *, off_t, vaddr_t, size_t, int, enum uio_rw, 63 void (*)(struct buf *)); 64 static inline void genfs_rel_pages(struct vm_page **, int); 65 66 #define MAX_READ_PAGES 16 /* XXXUBC 16 */ 67 68 int genfs_maxdio = MAXPHYS; 69 70 static inline void 71 genfs_rel_pages(struct vm_page **pgs, int npages) 72 { 73 int i; 74 75 for (i = 0; i < npages; i++) { 76 struct vm_page *pg = pgs[i]; 77 78 if (pg == NULL || pg == PGO_DONTCARE) 79 continue; 80 if (pg->flags & PG_FAKE) { 81 pg->flags |= PG_RELEASED; 82 } 83 } 84 mutex_enter(&uvm_pageqlock); 85 uvm_page_unbusy(pgs, npages); 86 mutex_exit(&uvm_pageqlock); 87 } 88 89 /* 90 * generic VM getpages routine. 91 * Return PG_BUSY pages for the given range, 92 * reading from backing store if necessary. 93 */ 94 95 int 96 genfs_getpages(void *v) 97 { 98 struct vop_getpages_args /* { 99 struct vnode *a_vp; 100 voff_t a_offset; 101 struct vm_page **a_m; 102 int *a_count; 103 int a_centeridx; 104 vm_prot_t a_access_type; 105 int a_advice; 106 int a_flags; 107 } */ *ap = v; 108 109 off_t newsize, diskeof, memeof; 110 off_t offset, origoffset, startoffset, endoffset; 111 daddr_t lbn, blkno; 112 int i, error, npages, orignpages, npgs, run, ridx, pidx, pcount; 113 int fs_bshift, fs_bsize, dev_bshift; 114 int flags = ap->a_flags; 115 size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes; 116 vaddr_t kva; 117 struct buf *bp, *mbp; 118 struct vnode *vp = ap->a_vp; 119 struct vnode *devvp; 120 struct genfs_node *gp = VTOG(vp); 121 struct uvm_object *uobj = &vp->v_uobj; 122 struct vm_page *pg, **pgs, *pgs_onstack[MAX_READ_PAGES]; 123 int pgs_size; 124 kauth_cred_t cred = curlwp->l_cred; /* XXXUBC curlwp */ 125 bool async = (flags & PGO_SYNCIO) == 0; 126 bool write = (ap->a_access_type & VM_PROT_WRITE) != 0; 127 bool sawhole = false; 128 bool has_trans = false; 129 bool overwrite = (flags & PGO_OVERWRITE) != 0; 130 bool blockalloc = write && (flags & PGO_NOBLOCKALLOC) == 0; 131 voff_t origvsize; 132 UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist); 133 134 UVMHIST_LOG(ubchist, "vp %p off 0x%x/%x count %d", 135 vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count); 136 137 KASSERT(vp->v_type == VREG || vp->v_type == VDIR || 138 vp->v_type == VLNK || vp->v_type == VBLK); 139 140 /* XXXUBC temp limit */ 141 if (*ap->a_count > MAX_READ_PAGES) { 142 panic("genfs_getpages: too many pages"); 143 } 144 145 pgs = pgs_onstack; 146 pgs_size = sizeof(pgs_onstack); 147 148 startover: 149 error = 0; 150 origvsize = vp->v_size; 151 origoffset = ap->a_offset; 152 orignpages = *ap->a_count; 153 GOP_SIZE(vp, origvsize, &diskeof, 0); 154 if (flags & PGO_PASTEOF) { 155 #if defined(DIAGNOSTIC) 156 off_t writeeof; 157 #endif /* defined(DIAGNOSTIC) */ 158 159 newsize = MAX(origvsize, 160 origoffset + (orignpages << PAGE_SHIFT)); 161 GOP_SIZE(vp, newsize, &memeof, GOP_SIZE_MEM); 162 #if defined(DIAGNOSTIC) 163 GOP_SIZE(vp, vp->v_writesize, &writeeof, GOP_SIZE_MEM); 164 if (newsize > round_page(writeeof)) { 165 panic("%s: past eof", __func__); 166 } 167 #endif /* defined(DIAGNOSTIC) */ 168 } else { 169 GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM); 170 } 171 KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages); 172 KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0); 173 KASSERT(orignpages > 0); 174 175 /* 176 * Bounds-check the request. 177 */ 178 179 if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) { 180 if ((flags & PGO_LOCKED) == 0) { 181 mutex_exit(&uobj->vmobjlock); 182 } 183 UVMHIST_LOG(ubchist, "off 0x%x count %d goes past EOF 0x%x", 184 origoffset, *ap->a_count, memeof,0); 185 error = EINVAL; 186 goto out_err; 187 } 188 189 /* uobj is locked */ 190 191 if ((flags & PGO_NOTIMESTAMP) == 0 && 192 (vp->v_type != VBLK || 193 (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) { 194 int updflags = 0; 195 196 if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) { 197 updflags = GOP_UPDATE_ACCESSED; 198 } 199 if (write) { 200 updflags |= GOP_UPDATE_MODIFIED; 201 } 202 if (updflags != 0) { 203 GOP_MARKUPDATE(vp, updflags); 204 } 205 } 206 207 if (write) { 208 gp->g_dirtygen++; 209 if ((vp->v_iflag & VI_ONWORKLST) == 0) { 210 vn_syncer_add_to_worklist(vp, filedelay); 211 } 212 if ((vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP) { 213 vp->v_iflag |= VI_WRMAPDIRTY; 214 } 215 } 216 217 /* 218 * For PGO_LOCKED requests, just return whatever's in memory. 219 */ 220 221 if (flags & PGO_LOCKED) { 222 int nfound; 223 224 npages = *ap->a_count; 225 #if defined(DEBUG) 226 for (i = 0; i < npages; i++) { 227 pg = ap->a_m[i]; 228 KASSERT(pg == NULL || pg == PGO_DONTCARE); 229 } 230 #endif /* defined(DEBUG) */ 231 nfound = uvn_findpages(uobj, origoffset, &npages, 232 ap->a_m, UFP_NOWAIT|UFP_NOALLOC|(write ? UFP_NORDONLY : 0)); 233 KASSERT(npages == *ap->a_count); 234 if (nfound == 0) { 235 error = EBUSY; 236 goto out_err; 237 } 238 if (!rw_tryenter(&gp->g_glock, RW_READER)) { 239 genfs_rel_pages(ap->a_m, npages); 240 241 /* 242 * restore the array. 243 */ 244 245 for (i = 0; i < npages; i++) { 246 pg = ap->a_m[i]; 247 248 if (pg != NULL || pg != PGO_DONTCARE) { 249 ap->a_m[i] = NULL; 250 } 251 } 252 } else { 253 rw_exit(&gp->g_glock); 254 } 255 error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0); 256 goto out_err; 257 } 258 mutex_exit(&uobj->vmobjlock); 259 260 /* 261 * find the requested pages and make some simple checks. 262 * leave space in the page array for a whole block. 263 */ 264 265 if (vp->v_type != VBLK) { 266 fs_bshift = vp->v_mount->mnt_fs_bshift; 267 dev_bshift = vp->v_mount->mnt_dev_bshift; 268 } else { 269 fs_bshift = DEV_BSHIFT; 270 dev_bshift = DEV_BSHIFT; 271 } 272 fs_bsize = 1 << fs_bshift; 273 274 orignpages = MIN(orignpages, 275 round_page(memeof - origoffset) >> PAGE_SHIFT); 276 npages = orignpages; 277 startoffset = origoffset & ~(fs_bsize - 1); 278 endoffset = round_page((origoffset + (npages << PAGE_SHIFT) + 279 fs_bsize - 1) & ~(fs_bsize - 1)); 280 endoffset = MIN(endoffset, round_page(memeof)); 281 ridx = (origoffset - startoffset) >> PAGE_SHIFT; 282 283 pgs_size = sizeof(struct vm_page *) * 284 ((endoffset - startoffset) >> PAGE_SHIFT); 285 if (pgs_size > sizeof(pgs_onstack)) { 286 pgs = kmem_zalloc(pgs_size, async ? KM_NOSLEEP : KM_SLEEP); 287 if (pgs == NULL) { 288 pgs = pgs_onstack; 289 error = ENOMEM; 290 goto out_err; 291 } 292 } else { 293 /* pgs == pgs_onstack */ 294 memset(pgs, 0, pgs_size); 295 } 296 UVMHIST_LOG(ubchist, "ridx %d npages %d startoff %ld endoff %ld", 297 ridx, npages, startoffset, endoffset); 298 299 if (!has_trans) { 300 fstrans_start(vp->v_mount, FSTRANS_SHARED); 301 has_trans = true; 302 } 303 304 /* 305 * hold g_glock to prevent a race with truncate. 306 * 307 * check if our idea of v_size is still valid. 308 */ 309 310 if (blockalloc) { 311 rw_enter(&gp->g_glock, RW_WRITER); 312 } else { 313 rw_enter(&gp->g_glock, RW_READER); 314 } 315 mutex_enter(&uobj->vmobjlock); 316 if (vp->v_size < origvsize) { 317 rw_exit(&gp->g_glock); 318 if (pgs != pgs_onstack) 319 kmem_free(pgs, pgs_size); 320 goto startover; 321 } 322 323 if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx], 324 async ? UFP_NOWAIT : UFP_ALL) != orignpages) { 325 rw_exit(&gp->g_glock); 326 KASSERT(async != 0); 327 genfs_rel_pages(&pgs[ridx], orignpages); 328 mutex_exit(&uobj->vmobjlock); 329 error = EBUSY; 330 goto out_err; 331 } 332 333 /* 334 * if the pages are already resident, just return them. 335 */ 336 337 for (i = 0; i < npages; i++) { 338 struct vm_page *pg1 = pgs[ridx + i]; 339 340 if ((pg1->flags & PG_FAKE) || 341 (blockalloc && (pg1->flags & PG_RDONLY))) { 342 break; 343 } 344 } 345 if (i == npages) { 346 rw_exit(&gp->g_glock); 347 UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0); 348 npages += ridx; 349 goto out; 350 } 351 352 /* 353 * if PGO_OVERWRITE is set, don't bother reading the pages. 354 */ 355 356 if (overwrite) { 357 rw_exit(&gp->g_glock); 358 UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0); 359 360 for (i = 0; i < npages; i++) { 361 struct vm_page *pg1 = pgs[ridx + i]; 362 363 pg1->flags &= ~(PG_RDONLY|PG_CLEAN); 364 } 365 npages += ridx; 366 goto out; 367 } 368 369 /* 370 * the page wasn't resident and we're not overwriting, 371 * so we're going to have to do some i/o. 372 * find any additional pages needed to cover the expanded range. 373 */ 374 375 npages = (endoffset - startoffset) >> PAGE_SHIFT; 376 if (startoffset != origoffset || npages != orignpages) { 377 378 /* 379 * we need to avoid deadlocks caused by locking 380 * additional pages at lower offsets than pages we 381 * already have locked. unlock them all and start over. 382 */ 383 384 genfs_rel_pages(&pgs[ridx], orignpages); 385 memset(pgs, 0, pgs_size); 386 387 UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x", 388 startoffset, endoffset, 0,0); 389 npgs = npages; 390 if (uvn_findpages(uobj, startoffset, &npgs, pgs, 391 async ? UFP_NOWAIT : UFP_ALL) != npages) { 392 rw_exit(&gp->g_glock); 393 KASSERT(async != 0); 394 genfs_rel_pages(pgs, npages); 395 mutex_exit(&uobj->vmobjlock); 396 error = EBUSY; 397 goto out_err; 398 } 399 } 400 mutex_exit(&uobj->vmobjlock); 401 402 /* 403 * read the desired page(s). 404 */ 405 406 totalbytes = npages << PAGE_SHIFT; 407 bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0)); 408 tailbytes = totalbytes - bytes; 409 skipbytes = 0; 410 411 kva = uvm_pagermapin(pgs, npages, 412 UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK); 413 414 mbp = getiobuf(vp, true); 415 mbp->b_bufsize = totalbytes; 416 mbp->b_data = (void *)kva; 417 mbp->b_resid = mbp->b_bcount = bytes; 418 mbp->b_cflags = BC_BUSY; 419 if (async) { 420 mbp->b_flags = B_READ | B_ASYNC; 421 mbp->b_iodone = uvm_aio_biodone; 422 } else { 423 mbp->b_flags = B_READ; 424 mbp->b_iodone = NULL; 425 } 426 if (async) 427 BIO_SETPRIO(mbp, BPRIO_TIMELIMITED); 428 else 429 BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL); 430 431 /* 432 * if EOF is in the middle of the range, zero the part past EOF. 433 * skip over pages which are not PG_FAKE since in that case they have 434 * valid data that we need to preserve. 435 */ 436 437 tailstart = bytes; 438 while (tailbytes > 0) { 439 const int len = PAGE_SIZE - (tailstart & PAGE_MASK); 440 441 KASSERT(len <= tailbytes); 442 if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) { 443 memset((void *)(kva + tailstart), 0, len); 444 UVMHIST_LOG(ubchist, "tailbytes %p 0x%x 0x%x", 445 kva, tailstart, len, 0); 446 } 447 tailstart += len; 448 tailbytes -= len; 449 } 450 451 /* 452 * now loop over the pages, reading as needed. 453 */ 454 455 bp = NULL; 456 for (offset = startoffset; 457 bytes > 0; 458 offset += iobytes, bytes -= iobytes) { 459 460 /* 461 * skip pages which don't need to be read. 462 */ 463 464 pidx = (offset - startoffset) >> PAGE_SHIFT; 465 while ((pgs[pidx]->flags & PG_FAKE) == 0) { 466 size_t b; 467 468 KASSERT((offset & (PAGE_SIZE - 1)) == 0); 469 if ((pgs[pidx]->flags & PG_RDONLY)) { 470 sawhole = true; 471 } 472 b = MIN(PAGE_SIZE, bytes); 473 offset += b; 474 bytes -= b; 475 skipbytes += b; 476 pidx++; 477 UVMHIST_LOG(ubchist, "skipping, new offset 0x%x", 478 offset, 0,0,0); 479 if (bytes == 0) { 480 goto loopdone; 481 } 482 } 483 484 /* 485 * bmap the file to find out the blkno to read from and 486 * how much we can read in one i/o. if bmap returns an error, 487 * skip the rest of the top-level i/o. 488 */ 489 490 lbn = offset >> fs_bshift; 491 error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run); 492 if (error) { 493 UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n", 494 lbn, error,0,0); 495 skipbytes += bytes; 496 goto loopdone; 497 } 498 499 /* 500 * see how many pages can be read with this i/o. 501 * reduce the i/o size if necessary to avoid 502 * overwriting pages with valid data. 503 */ 504 505 iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset, 506 bytes); 507 if (offset + iobytes > round_page(offset)) { 508 pcount = 1; 509 while (pidx + pcount < npages && 510 pgs[pidx + pcount]->flags & PG_FAKE) { 511 pcount++; 512 } 513 iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) - 514 (offset - trunc_page(offset))); 515 } 516 517 /* 518 * if this block isn't allocated, zero it instead of 519 * reading it. unless we are going to allocate blocks, 520 * mark the pages we zeroed PG_RDONLY. 521 */ 522 523 if (blkno < 0) { 524 int holepages = (round_page(offset + iobytes) - 525 trunc_page(offset)) >> PAGE_SHIFT; 526 UVMHIST_LOG(ubchist, "lbn 0x%x -> HOLE", lbn,0,0,0); 527 528 sawhole = true; 529 memset((char *)kva + (offset - startoffset), 0, 530 iobytes); 531 skipbytes += iobytes; 532 533 for (i = 0; i < holepages; i++) { 534 if (write) { 535 pgs[pidx + i]->flags &= ~PG_CLEAN; 536 } 537 if (!blockalloc) { 538 pgs[pidx + i]->flags |= PG_RDONLY; 539 } 540 } 541 continue; 542 } 543 544 /* 545 * allocate a sub-buf for this piece of the i/o 546 * (or just use mbp if there's only 1 piece), 547 * and start it going. 548 */ 549 550 if (offset == startoffset && iobytes == bytes) { 551 bp = mbp; 552 } else { 553 bp = getiobuf(vp, true); 554 nestiobuf_setup(mbp, bp, offset - startoffset, iobytes); 555 } 556 bp->b_lblkno = 0; 557 558 /* adjust physical blkno for partial blocks */ 559 bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >> 560 dev_bshift); 561 562 UVMHIST_LOG(ubchist, 563 "bp %p offset 0x%x bcount 0x%x blkno 0x%x", 564 bp, offset, iobytes, bp->b_blkno); 565 566 VOP_STRATEGY(devvp, bp); 567 } 568 569 loopdone: 570 nestiobuf_done(mbp, skipbytes, error); 571 if (async) { 572 UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0); 573 rw_exit(&gp->g_glock); 574 error = 0; 575 goto out_err; 576 } 577 if (bp != NULL) { 578 error = biowait(mbp); 579 } 580 putiobuf(mbp); 581 uvm_pagermapout(kva, npages); 582 583 /* 584 * if this we encountered a hole then we have to do a little more work. 585 * for read faults, we marked the page PG_RDONLY so that future 586 * write accesses to the page will fault again. 587 * for write faults, we must make sure that the backing store for 588 * the page is completely allocated while the pages are locked. 589 */ 590 591 if (!error && sawhole && blockalloc) { 592 error = GOP_ALLOC(vp, startoffset, npages << PAGE_SHIFT, 0, 593 cred); 594 UVMHIST_LOG(ubchist, "gop_alloc off 0x%x/0x%x -> %d", 595 startoffset, npages << PAGE_SHIFT, error,0); 596 if (!error) { 597 for (i = 0; i < npages; i++) { 598 if (pgs[i] == NULL) { 599 continue; 600 } 601 pgs[i]->flags &= ~(PG_CLEAN|PG_RDONLY); 602 UVMHIST_LOG(ubchist, "mark dirty pg %p", 603 pgs[i],0,0,0); 604 } 605 } 606 } 607 rw_exit(&gp->g_glock); 608 mutex_enter(&uobj->vmobjlock); 609 610 /* 611 * we're almost done! release the pages... 612 * for errors, we free the pages. 613 * otherwise we activate them and mark them as valid and clean. 614 * also, unbusy pages that were not actually requested. 615 */ 616 617 if (error) { 618 for (i = 0; i < npages; i++) { 619 if (pgs[i] == NULL) { 620 continue; 621 } 622 UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x", 623 pgs[i], pgs[i]->flags, 0,0); 624 if (pgs[i]->flags & PG_FAKE) { 625 pgs[i]->flags |= PG_RELEASED; 626 } 627 } 628 mutex_enter(&uvm_pageqlock); 629 uvm_page_unbusy(pgs, npages); 630 mutex_exit(&uvm_pageqlock); 631 mutex_exit(&uobj->vmobjlock); 632 UVMHIST_LOG(ubchist, "returning error %d", error,0,0,0); 633 goto out_err; 634 } 635 636 out: 637 UVMHIST_LOG(ubchist, "succeeding, npages %d", npages,0,0,0); 638 error = 0; 639 mutex_enter(&uvm_pageqlock); 640 for (i = 0; i < npages; i++) { 641 pg = pgs[i]; 642 if (pg == NULL) { 643 continue; 644 } 645 UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x", 646 pg, pg->flags, 0,0); 647 if (pg->flags & PG_FAKE && !overwrite) { 648 pg->flags &= ~(PG_FAKE); 649 pmap_clear_modify(pgs[i]); 650 } 651 KASSERT(!write || !blockalloc || (pg->flags & PG_RDONLY) == 0); 652 if (i < ridx || i >= ridx + orignpages || async) { 653 UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x", 654 pg, pg->offset,0,0); 655 if (pg->flags & PG_WANTED) { 656 wakeup(pg); 657 } 658 if (pg->flags & PG_FAKE) { 659 KASSERT(overwrite); 660 uvm_pagezero(pg); 661 } 662 if (pg->flags & PG_RELEASED) { 663 uvm_pagefree(pg); 664 continue; 665 } 666 uvm_pageenqueue(pg); 667 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE); 668 UVM_PAGE_OWN(pg, NULL); 669 } 670 } 671 mutex_exit(&uvm_pageqlock); 672 mutex_exit(&uobj->vmobjlock); 673 if (ap->a_m != NULL) { 674 memcpy(ap->a_m, &pgs[ridx], 675 orignpages * sizeof(struct vm_page *)); 676 } 677 678 out_err: 679 if (pgs != pgs_onstack) 680 kmem_free(pgs, pgs_size); 681 if (has_trans) 682 fstrans_done(vp->v_mount); 683 return (error); 684 } 685 686 /* 687 * generic VM putpages routine. 688 * Write the given range of pages to backing store. 689 * 690 * => "offhi == 0" means flush all pages at or after "offlo". 691 * => object should be locked by caller. we return with the 692 * object unlocked. 693 * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O). 694 * thus, a caller might want to unlock higher level resources 695 * (e.g. vm_map) before calling flush. 696 * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block 697 * => if PGO_ALLPAGES is set, then all pages in the object will be processed. 698 * => NOTE: we rely on the fact that the object's memq is a TAILQ and 699 * that new pages are inserted on the tail end of the list. thus, 700 * we can make a complete pass through the object in one go by starting 701 * at the head and working towards the tail (new pages are put in 702 * front of us). 703 * => NOTE: we are allowed to lock the page queues, so the caller 704 * must not be holding the page queue lock. 705 * 706 * note on "cleaning" object and PG_BUSY pages: 707 * this routine is holding the lock on the object. the only time 708 * that it can run into a PG_BUSY page that it does not own is if 709 * some other process has started I/O on the page (e.g. either 710 * a pagein, or a pageout). if the PG_BUSY page is being paged 711 * in, then it can not be dirty (!PG_CLEAN) because no one has 712 * had a chance to modify it yet. if the PG_BUSY page is being 713 * paged out then it means that someone else has already started 714 * cleaning the page for us (how nice!). in this case, if we 715 * have syncio specified, then after we make our pass through the 716 * object we need to wait for the other PG_BUSY pages to clear 717 * off (i.e. we need to do an iosync). also note that once a 718 * page is PG_BUSY it must stay in its object until it is un-busyed. 719 * 720 * note on page traversal: 721 * we can traverse the pages in an object either by going down the 722 * linked list in "uobj->memq", or we can go over the address range 723 * by page doing hash table lookups for each address. depending 724 * on how many pages are in the object it may be cheaper to do one 725 * or the other. we set "by_list" to true if we are using memq. 726 * if the cost of a hash lookup was equal to the cost of the list 727 * traversal we could compare the number of pages in the start->stop 728 * range to the total number of pages in the object. however, it 729 * seems that a hash table lookup is more expensive than the linked 730 * list traversal, so we multiply the number of pages in the 731 * range by an estimate of the relatively higher cost of the hash lookup. 732 */ 733 734 int 735 genfs_putpages(void *v) 736 { 737 struct vop_putpages_args /* { 738 struct vnode *a_vp; 739 voff_t a_offlo; 740 voff_t a_offhi; 741 int a_flags; 742 } */ *ap = v; 743 744 return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi, 745 ap->a_flags, NULL); 746 } 747 748 int 749 genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff, 750 int origflags, struct vm_page **busypg) 751 { 752 struct uvm_object *uobj = &vp->v_uobj; 753 kmutex_t *slock = &uobj->vmobjlock; 754 off_t off; 755 /* Even for strange MAXPHYS, the shift rounds down to a page */ 756 #define maxpages (MAXPHYS >> PAGE_SHIFT) 757 int i, error, npages, nback; 758 int freeflag; 759 struct vm_page *pgs[maxpages], *pg, *nextpg, *tpg, curmp, endmp; 760 bool wasclean, by_list, needs_clean, yld; 761 bool async = (origflags & PGO_SYNCIO) == 0; 762 bool pagedaemon = curlwp == uvm.pagedaemon_lwp; 763 struct lwp *l = curlwp ? curlwp : &lwp0; 764 struct genfs_node *gp = VTOG(vp); 765 int flags; 766 int dirtygen; 767 bool modified; 768 bool has_trans; 769 bool cleanall; 770 bool onworklst; 771 772 UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist); 773 774 KASSERT(origflags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)); 775 KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0); 776 KASSERT(startoff < endoff || endoff == 0); 777 778 UVMHIST_LOG(ubchist, "vp %p pages %d off 0x%x len 0x%x", 779 vp, uobj->uo_npages, startoff, endoff - startoff); 780 781 has_trans = false; 782 783 retry: 784 modified = false; 785 flags = origflags; 786 KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 || 787 (vp->v_iflag & VI_WRMAPDIRTY) == 0); 788 if (uobj->uo_npages == 0) { 789 if (vp->v_iflag & VI_ONWORKLST) { 790 vp->v_iflag &= ~VI_WRMAPDIRTY; 791 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) 792 vn_syncer_remove_from_worklist(vp); 793 } 794 if (has_trans) 795 fstrans_done(vp->v_mount); 796 mutex_exit(slock); 797 return (0); 798 } 799 800 /* 801 * the vnode has pages, set up to process the request. 802 */ 803 804 if (!has_trans && (flags & PGO_CLEANIT) != 0) { 805 mutex_exit(slock); 806 if (pagedaemon) { 807 error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY); 808 if (error) 809 return error; 810 } else 811 fstrans_start(vp->v_mount, FSTRANS_LAZY); 812 has_trans = true; 813 mutex_enter(slock); 814 goto retry; 815 } 816 817 error = 0; 818 wasclean = (vp->v_numoutput == 0); 819 off = startoff; 820 if (endoff == 0 || flags & PGO_ALLPAGES) { 821 endoff = trunc_page(LLONG_MAX); 822 } 823 by_list = (uobj->uo_npages <= 824 ((endoff - startoff) >> PAGE_SHIFT) * UVM_PAGE_HASH_PENALTY); 825 826 #if !defined(DEBUG) 827 /* 828 * if this vnode is known not to have dirty pages, 829 * don't bother to clean it out. 830 */ 831 832 if ((vp->v_iflag & VI_ONWORKLST) == 0) { 833 if ((flags & (PGO_FREE|PGO_DEACTIVATE)) == 0) { 834 goto skip_scan; 835 } 836 flags &= ~PGO_CLEANIT; 837 } 838 #endif /* !defined(DEBUG) */ 839 840 /* 841 * start the loop. when scanning by list, hold the last page 842 * in the list before we start. pages allocated after we start 843 * will be added to the end of the list, so we can stop at the 844 * current last page. 845 */ 846 847 cleanall = (flags & PGO_CLEANIT) != 0 && wasclean && 848 startoff == 0 && endoff == trunc_page(LLONG_MAX) && 849 (vp->v_iflag & VI_ONWORKLST) != 0; 850 dirtygen = gp->g_dirtygen; 851 freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED; 852 if (by_list) { 853 curmp.uobject = uobj; 854 curmp.offset = (voff_t)-1; 855 curmp.flags = PG_BUSY; 856 endmp.uobject = uobj; 857 endmp.offset = (voff_t)-1; 858 endmp.flags = PG_BUSY; 859 pg = TAILQ_FIRST(&uobj->memq); 860 TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq.queue); 861 uvm_lwp_hold(l); 862 } else { 863 pg = uvm_pagelookup(uobj, off); 864 } 865 nextpg = NULL; 866 while (by_list || off < endoff) { 867 868 /* 869 * if the current page is not interesting, move on to the next. 870 */ 871 872 KASSERT(pg == NULL || pg->uobject == uobj); 873 KASSERT(pg == NULL || 874 (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 || 875 (pg->flags & PG_BUSY) != 0); 876 if (by_list) { 877 if (pg == &endmp) { 878 break; 879 } 880 if (pg->offset < startoff || pg->offset >= endoff || 881 pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 882 if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 883 wasclean = false; 884 } 885 pg = TAILQ_NEXT(pg, listq.queue); 886 continue; 887 } 888 off = pg->offset; 889 } else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 890 if (pg != NULL) { 891 wasclean = false; 892 } 893 off += PAGE_SIZE; 894 if (off < endoff) { 895 pg = uvm_pagelookup(uobj, off); 896 } 897 continue; 898 } 899 900 /* 901 * if the current page needs to be cleaned and it's busy, 902 * wait for it to become unbusy. 903 */ 904 905 yld = (l->l_cpu->ci_schedstate.spc_flags & 906 SPCF_SHOULDYIELD) && !pagedaemon; 907 if (pg->flags & PG_BUSY || yld) { 908 UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0); 909 if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) { 910 UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0); 911 error = EDEADLK; 912 if (busypg != NULL) 913 *busypg = pg; 914 break; 915 } 916 if (pagedaemon) { 917 /* 918 * someone has taken the page while we 919 * dropped the lock for fstrans_start. 920 */ 921 break; 922 } 923 if (by_list) { 924 TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue); 925 UVMHIST_LOG(ubchist, "curmp next %p", 926 TAILQ_NEXT(&curmp, listq.queue), 0,0,0); 927 } 928 if (yld) { 929 mutex_exit(slock); 930 preempt(); 931 mutex_enter(slock); 932 } else { 933 pg->flags |= PG_WANTED; 934 UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0); 935 mutex_enter(slock); 936 } 937 if (by_list) { 938 UVMHIST_LOG(ubchist, "after next %p", 939 TAILQ_NEXT(&curmp, listq.queue), 0,0,0); 940 pg = TAILQ_NEXT(&curmp, listq.queue); 941 TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue); 942 } else { 943 pg = uvm_pagelookup(uobj, off); 944 } 945 continue; 946 } 947 948 /* 949 * if we're freeing, remove all mappings of the page now. 950 * if we're cleaning, check if the page is needs to be cleaned. 951 */ 952 953 if (flags & PGO_FREE) { 954 pmap_page_protect(pg, VM_PROT_NONE); 955 } else if (flags & PGO_CLEANIT) { 956 957 /* 958 * if we still have some hope to pull this vnode off 959 * from the syncer queue, write-protect the page. 960 */ 961 962 if (cleanall && wasclean && 963 gp->g_dirtygen == dirtygen) { 964 965 /* 966 * uobj pages get wired only by uvm_fault 967 * where uobj is locked. 968 */ 969 970 if (pg->wire_count == 0) { 971 pmap_page_protect(pg, 972 VM_PROT_READ|VM_PROT_EXECUTE); 973 } else { 974 cleanall = false; 975 } 976 } 977 } 978 979 if (flags & PGO_CLEANIT) { 980 needs_clean = pmap_clear_modify(pg) || 981 (pg->flags & PG_CLEAN) == 0; 982 pg->flags |= PG_CLEAN; 983 } else { 984 needs_clean = false; 985 } 986 987 /* 988 * if we're cleaning, build a cluster. 989 * the cluster will consist of pages which are currently dirty, 990 * but they will be returned to us marked clean. 991 * if not cleaning, just operate on the one page. 992 */ 993 994 if (needs_clean) { 995 KDASSERT((vp->v_iflag & VI_ONWORKLST)); 996 wasclean = false; 997 memset(pgs, 0, sizeof(pgs)); 998 pg->flags |= PG_BUSY; 999 UVM_PAGE_OWN(pg, "genfs_putpages"); 1000 1001 /* 1002 * first look backward. 1003 */ 1004 1005 npages = MIN(maxpages >> 1, off >> PAGE_SHIFT); 1006 nback = npages; 1007 uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0], 1008 UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD); 1009 if (nback) { 1010 memmove(&pgs[0], &pgs[npages - nback], 1011 nback * sizeof(pgs[0])); 1012 if (npages - nback < nback) 1013 memset(&pgs[nback], 0, 1014 (npages - nback) * sizeof(pgs[0])); 1015 else 1016 memset(&pgs[npages - nback], 0, 1017 nback * sizeof(pgs[0])); 1018 } 1019 1020 /* 1021 * then plug in our page of interest. 1022 */ 1023 1024 pgs[nback] = pg; 1025 1026 /* 1027 * then look forward to fill in the remaining space in 1028 * the array of pages. 1029 */ 1030 1031 npages = maxpages - nback - 1; 1032 uvn_findpages(uobj, off + PAGE_SIZE, &npages, 1033 &pgs[nback + 1], 1034 UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY); 1035 npages += nback + 1; 1036 } else { 1037 pgs[0] = pg; 1038 npages = 1; 1039 nback = 0; 1040 } 1041 1042 /* 1043 * apply FREE or DEACTIVATE options if requested. 1044 */ 1045 1046 if (flags & (PGO_DEACTIVATE|PGO_FREE)) { 1047 mutex_enter(&uvm_pageqlock); 1048 } 1049 for (i = 0; i < npages; i++) { 1050 tpg = pgs[i]; 1051 KASSERT(tpg->uobject == uobj); 1052 if (by_list && tpg == TAILQ_NEXT(pg, listq.queue)) 1053 pg = tpg; 1054 if (tpg->offset < startoff || tpg->offset >= endoff) 1055 continue; 1056 if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) { 1057 uvm_pagedeactivate(tpg); 1058 } else if (flags & PGO_FREE) { 1059 pmap_page_protect(tpg, VM_PROT_NONE); 1060 if (tpg->flags & PG_BUSY) { 1061 tpg->flags |= freeflag; 1062 if (pagedaemon) { 1063 uvm_pageout_start(1); 1064 uvm_pagedequeue(tpg); 1065 } 1066 } else { 1067 1068 /* 1069 * ``page is not busy'' 1070 * implies that npages is 1 1071 * and needs_clean is false. 1072 */ 1073 1074 nextpg = TAILQ_NEXT(tpg, listq.queue); 1075 uvm_pagefree(tpg); 1076 if (pagedaemon) 1077 uvmexp.pdfreed++; 1078 } 1079 } 1080 } 1081 if (flags & (PGO_DEACTIVATE|PGO_FREE)) { 1082 mutex_exit(&uvm_pageqlock); 1083 } 1084 if (needs_clean) { 1085 modified = true; 1086 1087 /* 1088 * start the i/o. if we're traversing by list, 1089 * keep our place in the list with a marker page. 1090 */ 1091 1092 if (by_list) { 1093 TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp, 1094 listq.queue); 1095 } 1096 mutex_exit(slock); 1097 error = GOP_WRITE(vp, pgs, npages, flags); 1098 mutex_enter(slock); 1099 if (by_list) { 1100 pg = TAILQ_NEXT(&curmp, listq.queue); 1101 TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue); 1102 } 1103 if (error) { 1104 break; 1105 } 1106 if (by_list) { 1107 continue; 1108 } 1109 } 1110 1111 /* 1112 * find the next page and continue if there was no error. 1113 */ 1114 1115 if (by_list) { 1116 if (nextpg) { 1117 pg = nextpg; 1118 nextpg = NULL; 1119 } else { 1120 pg = TAILQ_NEXT(pg, listq.queue); 1121 } 1122 } else { 1123 off += (npages - nback) << PAGE_SHIFT; 1124 if (off < endoff) { 1125 pg = uvm_pagelookup(uobj, off); 1126 } 1127 } 1128 } 1129 if (by_list) { 1130 TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue); 1131 uvm_lwp_rele(l); 1132 } 1133 1134 if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 && 1135 (vp->v_type != VBLK || 1136 (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) { 1137 GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED); 1138 } 1139 1140 /* 1141 * if we're cleaning and there was nothing to clean, 1142 * take us off the syncer list. if we started any i/o 1143 * and we're doing sync i/o, wait for all writes to finish. 1144 */ 1145 1146 if (cleanall && wasclean && gp->g_dirtygen == dirtygen && 1147 (vp->v_iflag & VI_ONWORKLST) != 0) { 1148 #if defined(DEBUG) 1149 TAILQ_FOREACH(pg, &uobj->memq, listq.queue) { 1150 if ((pg->flags & PG_CLEAN) == 0) { 1151 printf("%s: %p: !CLEAN\n", __func__, pg); 1152 } 1153 if (pmap_is_modified(pg)) { 1154 printf("%s: %p: modified\n", __func__, pg); 1155 } 1156 } 1157 #endif /* defined(DEBUG) */ 1158 vp->v_iflag &= ~VI_WRMAPDIRTY; 1159 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) 1160 vn_syncer_remove_from_worklist(vp); 1161 } 1162 1163 #if !defined(DEBUG) 1164 skip_scan: 1165 #endif /* !defined(DEBUG) */ 1166 1167 /* Wait for output to complete. */ 1168 if (!wasclean && !async && vp->v_numoutput != 0) { 1169 while (vp->v_numoutput != 0) 1170 cv_wait(&vp->v_cv, slock); 1171 } 1172 onworklst = (vp->v_iflag & VI_ONWORKLST) != 0; 1173 mutex_exit(slock); 1174 1175 if ((flags & PGO_RECLAIM) != 0 && onworklst) { 1176 /* 1177 * in the case of PGO_RECLAIM, ensure to make the vnode clean. 1178 * retrying is not a big deal because, in many cases, 1179 * uobj->uo_npages is already 0 here. 1180 */ 1181 mutex_enter(slock); 1182 goto retry; 1183 } 1184 1185 if (has_trans) 1186 fstrans_done(vp->v_mount); 1187 1188 return (error); 1189 } 1190 1191 int 1192 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags) 1193 { 1194 off_t off; 1195 vaddr_t kva; 1196 size_t len; 1197 int error; 1198 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1199 1200 UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x", 1201 vp, pgs, npages, flags); 1202 1203 off = pgs[0]->offset; 1204 kva = uvm_pagermapin(pgs, npages, 1205 UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK); 1206 len = npages << PAGE_SHIFT; 1207 1208 error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE, 1209 uvm_aio_biodone); 1210 1211 return error; 1212 } 1213 1214 int 1215 genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages, int flags) 1216 { 1217 off_t off; 1218 vaddr_t kva; 1219 size_t len; 1220 int error; 1221 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1222 1223 UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x", 1224 vp, pgs, npages, flags); 1225 1226 off = pgs[0]->offset; 1227 kva = uvm_pagermapin(pgs, npages, 1228 UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK); 1229 len = npages << PAGE_SHIFT; 1230 1231 error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE, 1232 uvm_aio_biodone); 1233 1234 return error; 1235 } 1236 1237 /* 1238 * Backend routine for doing I/O to vnode pages. Pages are already locked 1239 * and mapped into kernel memory. Here we just look up the underlying 1240 * device block addresses and call the strategy routine. 1241 */ 1242 1243 static int 1244 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags, 1245 enum uio_rw rw, void (*iodone)(struct buf *)) 1246 { 1247 int s, error, run; 1248 int fs_bshift, dev_bshift; 1249 off_t eof, offset, startoffset; 1250 size_t bytes, iobytes, skipbytes; 1251 daddr_t lbn, blkno; 1252 struct buf *mbp, *bp; 1253 struct vnode *devvp; 1254 bool async = (flags & PGO_SYNCIO) == 0; 1255 bool write = rw == UIO_WRITE; 1256 int brw = write ? B_WRITE : B_READ; 1257 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1258 1259 UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x", 1260 vp, kva, len, flags); 1261 1262 KASSERT(vp->v_size <= vp->v_writesize); 1263 GOP_SIZE(vp, vp->v_writesize, &eof, 0); 1264 if (vp->v_type != VBLK) { 1265 fs_bshift = vp->v_mount->mnt_fs_bshift; 1266 dev_bshift = vp->v_mount->mnt_dev_bshift; 1267 } else { 1268 fs_bshift = DEV_BSHIFT; 1269 dev_bshift = DEV_BSHIFT; 1270 } 1271 error = 0; 1272 startoffset = off; 1273 bytes = MIN(len, eof - startoffset); 1274 skipbytes = 0; 1275 KASSERT(bytes != 0); 1276 1277 if (write) { 1278 mutex_enter(&vp->v_interlock); 1279 vp->v_numoutput += 2; 1280 mutex_exit(&vp->v_interlock); 1281 } 1282 mbp = getiobuf(vp, true); 1283 UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x", 1284 vp, mbp, vp->v_numoutput, bytes); 1285 mbp->b_bufsize = len; 1286 mbp->b_data = (void *)kva; 1287 mbp->b_resid = mbp->b_bcount = bytes; 1288 mbp->b_cflags = BC_BUSY | BC_AGE; 1289 if (async) { 1290 mbp->b_flags = brw | B_ASYNC; 1291 mbp->b_iodone = iodone; 1292 } else { 1293 mbp->b_flags = brw; 1294 mbp->b_iodone = NULL; 1295 } 1296 if (curlwp == uvm.pagedaemon_lwp) 1297 BIO_SETPRIO(mbp, BPRIO_TIMELIMITED); 1298 else if (async) 1299 BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL); 1300 else 1301 BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL); 1302 1303 bp = NULL; 1304 for (offset = startoffset; 1305 bytes > 0; 1306 offset += iobytes, bytes -= iobytes) { 1307 lbn = offset >> fs_bshift; 1308 error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run); 1309 if (error) { 1310 UVMHIST_LOG(ubchist, "VOP_BMAP() -> %d", error,0,0,0); 1311 skipbytes += bytes; 1312 bytes = 0; 1313 break; 1314 } 1315 1316 iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset, 1317 bytes); 1318 if (blkno == (daddr_t)-1) { 1319 if (!write) { 1320 memset((char *)kva + (offset - startoffset), 0, 1321 iobytes); 1322 } 1323 skipbytes += iobytes; 1324 continue; 1325 } 1326 1327 /* if it's really one i/o, don't make a second buf */ 1328 if (offset == startoffset && iobytes == bytes) { 1329 bp = mbp; 1330 } else { 1331 UVMHIST_LOG(ubchist, "vp %p bp %p num now %d", 1332 vp, bp, vp->v_numoutput, 0); 1333 bp = getiobuf(vp, true); 1334 nestiobuf_setup(mbp, bp, offset - startoffset, iobytes); 1335 } 1336 bp->b_lblkno = 0; 1337 1338 /* adjust physical blkno for partial blocks */ 1339 bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >> 1340 dev_bshift); 1341 UVMHIST_LOG(ubchist, 1342 "vp %p offset 0x%x bcount 0x%x blkno 0x%x", 1343 vp, offset, bp->b_bcount, bp->b_blkno); 1344 1345 VOP_STRATEGY(devvp, bp); 1346 } 1347 if (skipbytes) { 1348 UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0); 1349 } 1350 nestiobuf_done(mbp, skipbytes, error); 1351 if (async) { 1352 UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0); 1353 return (0); 1354 } 1355 UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0); 1356 error = biowait(mbp); 1357 s = splbio(); 1358 (*iodone)(mbp); 1359 splx(s); 1360 UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0); 1361 return (error); 1362 } 1363 1364 /* 1365 * VOP_PUTPAGES() for vnodes which never have pages. 1366 */ 1367 1368 int 1369 genfs_null_putpages(void *v) 1370 { 1371 struct vop_putpages_args /* { 1372 struct vnode *a_vp; 1373 voff_t a_offlo; 1374 voff_t a_offhi; 1375 int a_flags; 1376 } */ *ap = v; 1377 struct vnode *vp = ap->a_vp; 1378 1379 KASSERT(vp->v_uobj.uo_npages == 0); 1380 mutex_exit(&vp->v_interlock); 1381 return (0); 1382 } 1383 1384 int 1385 genfs_compat_getpages(void *v) 1386 { 1387 struct vop_getpages_args /* { 1388 struct vnode *a_vp; 1389 voff_t a_offset; 1390 struct vm_page **a_m; 1391 int *a_count; 1392 int a_centeridx; 1393 vm_prot_t a_access_type; 1394 int a_advice; 1395 int a_flags; 1396 } */ *ap = v; 1397 1398 off_t origoffset; 1399 struct vnode *vp = ap->a_vp; 1400 struct uvm_object *uobj = &vp->v_uobj; 1401 struct vm_page *pg, **pgs; 1402 vaddr_t kva; 1403 int i, error, orignpages, npages; 1404 struct iovec iov; 1405 struct uio uio; 1406 kauth_cred_t cred = curlwp->l_cred; 1407 bool write = (ap->a_access_type & VM_PROT_WRITE) != 0; 1408 1409 error = 0; 1410 origoffset = ap->a_offset; 1411 orignpages = *ap->a_count; 1412 pgs = ap->a_m; 1413 1414 if (write && (vp->v_iflag & VI_ONWORKLST) == 0) { 1415 vn_syncer_add_to_worklist(vp, filedelay); 1416 } 1417 if (ap->a_flags & PGO_LOCKED) { 1418 uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m, 1419 UFP_NOWAIT|UFP_NOALLOC| (write ? UFP_NORDONLY : 0)); 1420 1421 return (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0); 1422 } 1423 if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) { 1424 mutex_exit(&uobj->vmobjlock); 1425 return (EINVAL); 1426 } 1427 if ((ap->a_flags & PGO_SYNCIO) == 0) { 1428 mutex_exit(&uobj->vmobjlock); 1429 return 0; 1430 } 1431 npages = orignpages; 1432 uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL); 1433 mutex_exit(&uobj->vmobjlock); 1434 kva = uvm_pagermapin(pgs, npages, 1435 UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK); 1436 for (i = 0; i < npages; i++) { 1437 pg = pgs[i]; 1438 if ((pg->flags & PG_FAKE) == 0) { 1439 continue; 1440 } 1441 iov.iov_base = (char *)kva + (i << PAGE_SHIFT); 1442 iov.iov_len = PAGE_SIZE; 1443 uio.uio_iov = &iov; 1444 uio.uio_iovcnt = 1; 1445 uio.uio_offset = origoffset + (i << PAGE_SHIFT); 1446 uio.uio_rw = UIO_READ; 1447 uio.uio_resid = PAGE_SIZE; 1448 UIO_SETUP_SYSSPACE(&uio); 1449 /* XXX vn_lock */ 1450 error = VOP_READ(vp, &uio, 0, cred); 1451 if (error) { 1452 break; 1453 } 1454 if (uio.uio_resid) { 1455 memset(iov.iov_base, 0, uio.uio_resid); 1456 } 1457 } 1458 uvm_pagermapout(kva, npages); 1459 mutex_enter(&uobj->vmobjlock); 1460 mutex_enter(&uvm_pageqlock); 1461 for (i = 0; i < npages; i++) { 1462 pg = pgs[i]; 1463 if (error && (pg->flags & PG_FAKE) != 0) { 1464 pg->flags |= PG_RELEASED; 1465 } else { 1466 pmap_clear_modify(pg); 1467 uvm_pageactivate(pg); 1468 } 1469 } 1470 if (error) { 1471 uvm_page_unbusy(pgs, npages); 1472 } 1473 mutex_exit(&uvm_pageqlock); 1474 mutex_exit(&uobj->vmobjlock); 1475 return (error); 1476 } 1477 1478 int 1479 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, 1480 int flags) 1481 { 1482 off_t offset; 1483 struct iovec iov; 1484 struct uio uio; 1485 kauth_cred_t cred = curlwp->l_cred; 1486 struct buf *bp; 1487 vaddr_t kva; 1488 int error; 1489 1490 offset = pgs[0]->offset; 1491 kva = uvm_pagermapin(pgs, npages, 1492 UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK); 1493 1494 iov.iov_base = (void *)kva; 1495 iov.iov_len = npages << PAGE_SHIFT; 1496 uio.uio_iov = &iov; 1497 uio.uio_iovcnt = 1; 1498 uio.uio_offset = offset; 1499 uio.uio_rw = UIO_WRITE; 1500 uio.uio_resid = npages << PAGE_SHIFT; 1501 UIO_SETUP_SYSSPACE(&uio); 1502 /* XXX vn_lock */ 1503 error = VOP_WRITE(vp, &uio, 0, cred); 1504 1505 mutex_enter(&vp->v_interlock); 1506 vp->v_numoutput++; 1507 mutex_exit(&vp->v_interlock); 1508 1509 bp = getiobuf(vp, true); 1510 bp->b_cflags = BC_BUSY | BC_AGE; 1511 bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift; 1512 bp->b_data = (char *)kva; 1513 bp->b_bcount = npages << PAGE_SHIFT; 1514 bp->b_bufsize = npages << PAGE_SHIFT; 1515 bp->b_resid = 0; 1516 bp->b_error = error; 1517 uvm_aio_aiodone(bp); 1518 return (error); 1519 } 1520 1521 /* 1522 * Process a uio using direct I/O. If we reach a part of the request 1523 * which cannot be processed in this fashion for some reason, just return. 1524 * The caller must handle some additional part of the request using 1525 * buffered I/O before trying direct I/O again. 1526 */ 1527 1528 void 1529 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag) 1530 { 1531 struct vmspace *vs; 1532 struct iovec *iov; 1533 vaddr_t va; 1534 size_t len; 1535 const int mask = DEV_BSIZE - 1; 1536 int error; 1537 1538 /* 1539 * We only support direct I/O to user space for now. 1540 */ 1541 1542 if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) { 1543 return; 1544 } 1545 1546 /* 1547 * If the vnode is mapped, we would need to get the getpages lock 1548 * to stabilize the bmap, but then we would get into trouble whil e 1549 * locking the pages if the pages belong to this same vnode (or a 1550 * multi-vnode cascade to the same effect). Just fall back to 1551 * buffered I/O if the vnode is mapped to avoid this mess. 1552 */ 1553 1554 if (vp->v_vflag & VV_MAPPED) { 1555 return; 1556 } 1557 1558 /* 1559 * Do as much of the uio as possible with direct I/O. 1560 */ 1561 1562 vs = uio->uio_vmspace; 1563 while (uio->uio_resid) { 1564 iov = uio->uio_iov; 1565 if (iov->iov_len == 0) { 1566 uio->uio_iov++; 1567 uio->uio_iovcnt--; 1568 continue; 1569 } 1570 va = (vaddr_t)iov->iov_base; 1571 len = MIN(iov->iov_len, genfs_maxdio); 1572 len &= ~mask; 1573 1574 /* 1575 * If the next chunk is smaller than DEV_BSIZE or extends past 1576 * the current EOF, then fall back to buffered I/O. 1577 */ 1578 1579 if (len == 0 || uio->uio_offset + len > vp->v_size) { 1580 return; 1581 } 1582 1583 /* 1584 * Check alignment. The file offset must be at least 1585 * sector-aligned. The exact constraint on memory alignment 1586 * is very hardware-dependent, but requiring sector-aligned 1587 * addresses there too is safe. 1588 */ 1589 1590 if (uio->uio_offset & mask || va & mask) { 1591 return; 1592 } 1593 error = genfs_do_directio(vs, va, len, vp, uio->uio_offset, 1594 uio->uio_rw); 1595 if (error) { 1596 break; 1597 } 1598 iov->iov_base = (char *)iov->iov_base + len; 1599 iov->iov_len -= len; 1600 uio->uio_offset += len; 1601 uio->uio_resid -= len; 1602 } 1603 } 1604 1605 /* 1606 * Iodone routine for direct I/O. We don't do much here since the request is 1607 * always synchronous, so the caller will do most of the work after biowait(). 1608 */ 1609 1610 static void 1611 genfs_dio_iodone(struct buf *bp) 1612 { 1613 1614 KASSERT((bp->b_flags & B_ASYNC) == 0); 1615 if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) { 1616 mutex_enter(bp->b_objlock); 1617 vwakeup(bp); 1618 mutex_exit(bp->b_objlock); 1619 } 1620 putiobuf(bp); 1621 } 1622 1623 /* 1624 * Process one chunk of a direct I/O request. 1625 */ 1626 1627 static int 1628 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp, 1629 off_t off, enum uio_rw rw) 1630 { 1631 struct vm_map *map; 1632 struct pmap *upm, *kpm; 1633 size_t klen = round_page(uva + len) - trunc_page(uva); 1634 off_t spoff, epoff; 1635 vaddr_t kva, puva; 1636 paddr_t pa; 1637 vm_prot_t prot; 1638 int error, rv, poff, koff; 1639 const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | 1640 (rw == UIO_WRITE ? PGO_FREE : 0); 1641 1642 /* 1643 * For writes, verify that this range of the file already has fully 1644 * allocated backing store. If there are any holes, just punt and 1645 * make the caller take the buffered write path. 1646 */ 1647 1648 if (rw == UIO_WRITE) { 1649 daddr_t lbn, elbn, blkno; 1650 int bsize, bshift, run; 1651 1652 bshift = vp->v_mount->mnt_fs_bshift; 1653 bsize = 1 << bshift; 1654 lbn = off >> bshift; 1655 elbn = (off + len + bsize - 1) >> bshift; 1656 while (lbn < elbn) { 1657 error = VOP_BMAP(vp, lbn, NULL, &blkno, &run); 1658 if (error) { 1659 return error; 1660 } 1661 if (blkno == (daddr_t)-1) { 1662 return ENOSPC; 1663 } 1664 lbn += 1 + run; 1665 } 1666 } 1667 1668 /* 1669 * Flush any cached pages for parts of the file that we're about to 1670 * access. If we're writing, invalidate pages as well. 1671 */ 1672 1673 spoff = trunc_page(off); 1674 epoff = round_page(off + len); 1675 mutex_enter(&vp->v_interlock); 1676 error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags); 1677 if (error) { 1678 return error; 1679 } 1680 1681 /* 1682 * Wire the user pages and remap them into kernel memory. 1683 */ 1684 1685 prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ; 1686 error = uvm_vslock(vs, (void *)uva, len, prot); 1687 if (error) { 1688 return error; 1689 } 1690 1691 map = &vs->vm_map; 1692 upm = vm_map_pmap(map); 1693 kpm = vm_map_pmap(kernel_map); 1694 kva = uvm_km_alloc(kernel_map, klen, 0, 1695 UVM_KMF_VAONLY | UVM_KMF_WAITVA); 1696 puva = trunc_page(uva); 1697 for (poff = 0; poff < klen; poff += PAGE_SIZE) { 1698 rv = pmap_extract(upm, puva + poff, &pa); 1699 KASSERT(rv); 1700 pmap_enter(kpm, kva + poff, pa, prot, prot | PMAP_WIRED); 1701 } 1702 pmap_update(kpm); 1703 1704 /* 1705 * Do the I/O. 1706 */ 1707 1708 koff = uva - trunc_page(uva); 1709 error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw, 1710 genfs_dio_iodone); 1711 1712 /* 1713 * Tear down the kernel mapping. 1714 */ 1715 1716 pmap_remove(kpm, kva, kva + klen); 1717 pmap_update(kpm); 1718 uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY); 1719 1720 /* 1721 * Unwire the user pages. 1722 */ 1723 1724 uvm_vsunlock(vs, (void *)uva, len); 1725 return error; 1726 } 1727 1728