1 /* $NetBSD: genfs_io.c,v 1.5 2008/01/18 11:01:23 yamt 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 1.5 2008/01/18 11:01:23 yamt 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 retry: 782 modified = false; 783 has_trans = false; 784 flags = origflags; 785 KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 || 786 (vp->v_iflag & VI_WRMAPDIRTY) == 0); 787 if (uobj->uo_npages == 0) { 788 if (vp->v_iflag & VI_ONWORKLST) { 789 vp->v_iflag &= ~VI_WRMAPDIRTY; 790 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) 791 vn_syncer_remove_from_worklist(vp); 792 } 793 mutex_exit(slock); 794 return (0); 795 } 796 797 /* 798 * the vnode has pages, set up to process the request. 799 */ 800 801 if ((flags & PGO_CLEANIT) != 0) { 802 mutex_exit(slock); 803 if (pagedaemon) { 804 error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY); 805 if (error) 806 return error; 807 } else 808 fstrans_start(vp->v_mount, FSTRANS_LAZY); 809 has_trans = true; 810 mutex_enter(slock); 811 } 812 813 error = 0; 814 wasclean = (vp->v_numoutput == 0); 815 off = startoff; 816 if (endoff == 0 || flags & PGO_ALLPAGES) { 817 endoff = trunc_page(LLONG_MAX); 818 } 819 by_list = (uobj->uo_npages <= 820 ((endoff - startoff) >> PAGE_SHIFT) * UVM_PAGE_HASH_PENALTY); 821 822 #if !defined(DEBUG) 823 /* 824 * if this vnode is known not to have dirty pages, 825 * don't bother to clean it out. 826 */ 827 828 if ((vp->v_iflag & VI_ONWORKLST) == 0) { 829 if ((flags & (PGO_FREE|PGO_DEACTIVATE)) == 0) { 830 goto skip_scan; 831 } 832 flags &= ~PGO_CLEANIT; 833 } 834 #endif /* !defined(DEBUG) */ 835 836 /* 837 * start the loop. when scanning by list, hold the last page 838 * in the list before we start. pages allocated after we start 839 * will be added to the end of the list, so we can stop at the 840 * current last page. 841 */ 842 843 cleanall = (flags & PGO_CLEANIT) != 0 && wasclean && 844 startoff == 0 && endoff == trunc_page(LLONG_MAX) && 845 (vp->v_iflag & VI_ONWORKLST) != 0; 846 dirtygen = gp->g_dirtygen; 847 freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED; 848 if (by_list) { 849 curmp.uobject = uobj; 850 curmp.offset = (voff_t)-1; 851 curmp.flags = PG_BUSY; 852 endmp.uobject = uobj; 853 endmp.offset = (voff_t)-1; 854 endmp.flags = PG_BUSY; 855 pg = TAILQ_FIRST(&uobj->memq); 856 TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq); 857 uvm_lwp_hold(l); 858 } else { 859 pg = uvm_pagelookup(uobj, off); 860 } 861 nextpg = NULL; 862 while (by_list || off < endoff) { 863 864 /* 865 * if the current page is not interesting, move on to the next. 866 */ 867 868 KASSERT(pg == NULL || pg->uobject == uobj); 869 KASSERT(pg == NULL || 870 (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 || 871 (pg->flags & PG_BUSY) != 0); 872 if (by_list) { 873 if (pg == &endmp) { 874 break; 875 } 876 if (pg->offset < startoff || pg->offset >= endoff || 877 pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 878 if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 879 wasclean = false; 880 } 881 pg = TAILQ_NEXT(pg, listq); 882 continue; 883 } 884 off = pg->offset; 885 } else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 886 if (pg != NULL) { 887 wasclean = false; 888 } 889 off += PAGE_SIZE; 890 if (off < endoff) { 891 pg = uvm_pagelookup(uobj, off); 892 } 893 continue; 894 } 895 896 /* 897 * if the current page needs to be cleaned and it's busy, 898 * wait for it to become unbusy. 899 */ 900 901 yld = (l->l_cpu->ci_schedstate.spc_flags & 902 SPCF_SHOULDYIELD) && !pagedaemon; 903 if (pg->flags & PG_BUSY || yld) { 904 UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0); 905 if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) { 906 UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0); 907 error = EDEADLK; 908 if (busypg != NULL) 909 *busypg = pg; 910 break; 911 } 912 if (pagedaemon) { 913 /* 914 * someone has taken the page while we 915 * dropped the lock for fstrans_start. 916 */ 917 break; 918 } 919 if (by_list) { 920 TAILQ_INSERT_BEFORE(pg, &curmp, listq); 921 UVMHIST_LOG(ubchist, "curmp next %p", 922 TAILQ_NEXT(&curmp, listq), 0,0,0); 923 } 924 if (yld) { 925 mutex_exit(slock); 926 preempt(); 927 mutex_enter(slock); 928 } else { 929 pg->flags |= PG_WANTED; 930 UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0); 931 mutex_enter(slock); 932 } 933 if (by_list) { 934 UVMHIST_LOG(ubchist, "after next %p", 935 TAILQ_NEXT(&curmp, listq), 0,0,0); 936 pg = TAILQ_NEXT(&curmp, listq); 937 TAILQ_REMOVE(&uobj->memq, &curmp, listq); 938 } else { 939 pg = uvm_pagelookup(uobj, off); 940 } 941 continue; 942 } 943 944 /* 945 * if we're freeing, remove all mappings of the page now. 946 * if we're cleaning, check if the page is needs to be cleaned. 947 */ 948 949 if (flags & PGO_FREE) { 950 pmap_page_protect(pg, VM_PROT_NONE); 951 } else if (flags & PGO_CLEANIT) { 952 953 /* 954 * if we still have some hope to pull this vnode off 955 * from the syncer queue, write-protect the page. 956 */ 957 958 if (cleanall && wasclean && 959 gp->g_dirtygen == dirtygen) { 960 961 /* 962 * uobj pages get wired only by uvm_fault 963 * where uobj is locked. 964 */ 965 966 if (pg->wire_count == 0) { 967 pmap_page_protect(pg, 968 VM_PROT_READ|VM_PROT_EXECUTE); 969 } else { 970 cleanall = false; 971 } 972 } 973 } 974 975 if (flags & PGO_CLEANIT) { 976 needs_clean = pmap_clear_modify(pg) || 977 (pg->flags & PG_CLEAN) == 0; 978 pg->flags |= PG_CLEAN; 979 } else { 980 needs_clean = false; 981 } 982 983 /* 984 * if we're cleaning, build a cluster. 985 * the cluster will consist of pages which are currently dirty, 986 * but they will be returned to us marked clean. 987 * if not cleaning, just operate on the one page. 988 */ 989 990 if (needs_clean) { 991 KDASSERT((vp->v_iflag & VI_ONWORKLST)); 992 wasclean = false; 993 memset(pgs, 0, sizeof(pgs)); 994 pg->flags |= PG_BUSY; 995 UVM_PAGE_OWN(pg, "genfs_putpages"); 996 997 /* 998 * first look backward. 999 */ 1000 1001 npages = MIN(maxpages >> 1, off >> PAGE_SHIFT); 1002 nback = npages; 1003 uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0], 1004 UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD); 1005 if (nback) { 1006 memmove(&pgs[0], &pgs[npages - nback], 1007 nback * sizeof(pgs[0])); 1008 if (npages - nback < nback) 1009 memset(&pgs[nback], 0, 1010 (npages - nback) * sizeof(pgs[0])); 1011 else 1012 memset(&pgs[npages - nback], 0, 1013 nback * sizeof(pgs[0])); 1014 } 1015 1016 /* 1017 * then plug in our page of interest. 1018 */ 1019 1020 pgs[nback] = pg; 1021 1022 /* 1023 * then look forward to fill in the remaining space in 1024 * the array of pages. 1025 */ 1026 1027 npages = maxpages - nback - 1; 1028 uvn_findpages(uobj, off + PAGE_SIZE, &npages, 1029 &pgs[nback + 1], 1030 UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY); 1031 npages += nback + 1; 1032 } else { 1033 pgs[0] = pg; 1034 npages = 1; 1035 nback = 0; 1036 } 1037 1038 /* 1039 * apply FREE or DEACTIVATE options if requested. 1040 */ 1041 1042 if (flags & (PGO_DEACTIVATE|PGO_FREE)) { 1043 mutex_enter(&uvm_pageqlock); 1044 } 1045 for (i = 0; i < npages; i++) { 1046 tpg = pgs[i]; 1047 KASSERT(tpg->uobject == uobj); 1048 if (by_list && tpg == TAILQ_NEXT(pg, listq)) 1049 pg = tpg; 1050 if (tpg->offset < startoff || tpg->offset >= endoff) 1051 continue; 1052 if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) { 1053 uvm_pagedeactivate(tpg); 1054 } else if (flags & PGO_FREE) { 1055 pmap_page_protect(tpg, VM_PROT_NONE); 1056 if (tpg->flags & PG_BUSY) { 1057 tpg->flags |= freeflag; 1058 if (pagedaemon) { 1059 uvm_pageout_start(1); 1060 uvm_pagedequeue(tpg); 1061 } 1062 } else { 1063 1064 /* 1065 * ``page is not busy'' 1066 * implies that npages is 1 1067 * and needs_clean is false. 1068 */ 1069 1070 nextpg = TAILQ_NEXT(tpg, listq); 1071 uvm_pagefree(tpg); 1072 if (pagedaemon) 1073 uvmexp.pdfreed++; 1074 } 1075 } 1076 } 1077 if (flags & (PGO_DEACTIVATE|PGO_FREE)) { 1078 mutex_exit(&uvm_pageqlock); 1079 } 1080 if (needs_clean) { 1081 modified = true; 1082 1083 /* 1084 * start the i/o. if we're traversing by list, 1085 * keep our place in the list with a marker page. 1086 */ 1087 1088 if (by_list) { 1089 TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp, 1090 listq); 1091 } 1092 mutex_exit(slock); 1093 error = GOP_WRITE(vp, pgs, npages, flags); 1094 mutex_enter(slock); 1095 if (by_list) { 1096 pg = TAILQ_NEXT(&curmp, listq); 1097 TAILQ_REMOVE(&uobj->memq, &curmp, listq); 1098 } 1099 if (error) { 1100 break; 1101 } 1102 if (by_list) { 1103 continue; 1104 } 1105 } 1106 1107 /* 1108 * find the next page and continue if there was no error. 1109 */ 1110 1111 if (by_list) { 1112 if (nextpg) { 1113 pg = nextpg; 1114 nextpg = NULL; 1115 } else { 1116 pg = TAILQ_NEXT(pg, listq); 1117 } 1118 } else { 1119 off += (npages - nback) << PAGE_SHIFT; 1120 if (off < endoff) { 1121 pg = uvm_pagelookup(uobj, off); 1122 } 1123 } 1124 } 1125 if (by_list) { 1126 TAILQ_REMOVE(&uobj->memq, &endmp, listq); 1127 uvm_lwp_rele(l); 1128 } 1129 1130 if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 && 1131 (vp->v_type != VBLK || 1132 (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) { 1133 GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED); 1134 } 1135 1136 /* 1137 * if we're cleaning and there was nothing to clean, 1138 * take us off the syncer list. if we started any i/o 1139 * and we're doing sync i/o, wait for all writes to finish. 1140 */ 1141 1142 if (cleanall && wasclean && gp->g_dirtygen == dirtygen && 1143 (vp->v_iflag & VI_ONWORKLST) != 0) { 1144 #if defined(DEBUG) 1145 TAILQ_FOREACH(pg, &uobj->memq, listq) { 1146 if ((pg->flags & PG_CLEAN) == 0) { 1147 printf("%s: %p: !CLEAN\n", __func__, pg); 1148 } 1149 if (pmap_is_modified(pg)) { 1150 printf("%s: %p: modified\n", __func__, pg); 1151 } 1152 } 1153 #endif /* defined(DEBUG) */ 1154 vp->v_iflag &= ~VI_WRMAPDIRTY; 1155 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) 1156 vn_syncer_remove_from_worklist(vp); 1157 } 1158 1159 #if !defined(DEBUG) 1160 skip_scan: 1161 #endif /* !defined(DEBUG) */ 1162 1163 /* Wait for output to complete. */ 1164 if (!wasclean && !async && vp->v_numoutput != 0) { 1165 while (vp->v_numoutput != 0) 1166 cv_wait(&vp->v_cv, slock); 1167 } 1168 onworklst = (vp->v_iflag & VI_ONWORKLST) != 0; 1169 mutex_exit(slock); 1170 1171 if (has_trans) 1172 fstrans_done(vp->v_mount); 1173 1174 if ((flags & PGO_RECLAIM) != 0 && onworklst) { 1175 /* 1176 * in the case of PGO_RECLAIM, ensure to make the vnode clean. 1177 * retrying is not a big deal because, in many cases, 1178 * uobj->uo_npages is already 0 here. 1179 */ 1180 mutex_enter(slock); 1181 goto retry; 1182 } 1183 1184 return (error); 1185 } 1186 1187 int 1188 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags) 1189 { 1190 off_t off; 1191 vaddr_t kva; 1192 size_t len; 1193 int error; 1194 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1195 1196 UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x", 1197 vp, pgs, npages, flags); 1198 1199 off = pgs[0]->offset; 1200 kva = uvm_pagermapin(pgs, npages, 1201 UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK); 1202 len = npages << PAGE_SHIFT; 1203 1204 error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE, 1205 uvm_aio_biodone); 1206 1207 return error; 1208 } 1209 1210 /* 1211 * Backend routine for doing I/O to vnode pages. Pages are already locked 1212 * and mapped into kernel memory. Here we just look up the underlying 1213 * device block addresses and call the strategy routine. 1214 */ 1215 1216 static int 1217 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags, 1218 enum uio_rw rw, void (*iodone)(struct buf *)) 1219 { 1220 int s, error, run; 1221 int fs_bshift, dev_bshift; 1222 off_t eof, offset, startoffset; 1223 size_t bytes, iobytes, skipbytes; 1224 daddr_t lbn, blkno; 1225 struct buf *mbp, *bp; 1226 struct vnode *devvp; 1227 bool async = (flags & PGO_SYNCIO) == 0; 1228 bool write = rw == UIO_WRITE; 1229 int brw = write ? B_WRITE : B_READ; 1230 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1231 1232 UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x", 1233 vp, kva, len, flags); 1234 1235 KASSERT(vp->v_size <= vp->v_writesize); 1236 GOP_SIZE(vp, vp->v_writesize, &eof, 0); 1237 if (vp->v_type != VBLK) { 1238 fs_bshift = vp->v_mount->mnt_fs_bshift; 1239 dev_bshift = vp->v_mount->mnt_dev_bshift; 1240 } else { 1241 fs_bshift = DEV_BSHIFT; 1242 dev_bshift = DEV_BSHIFT; 1243 } 1244 error = 0; 1245 startoffset = off; 1246 bytes = MIN(len, eof - startoffset); 1247 skipbytes = 0; 1248 KASSERT(bytes != 0); 1249 1250 if (write) { 1251 mutex_enter(&vp->v_interlock); 1252 vp->v_numoutput += 2; 1253 mutex_exit(&vp->v_interlock); 1254 } 1255 mbp = getiobuf(vp, true); 1256 UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x", 1257 vp, mbp, vp->v_numoutput, bytes); 1258 mbp->b_bufsize = len; 1259 mbp->b_data = (void *)kva; 1260 mbp->b_resid = mbp->b_bcount = bytes; 1261 mbp->b_cflags = BC_BUSY | BC_AGE; 1262 if (async) { 1263 mbp->b_flags = brw | B_ASYNC; 1264 mbp->b_iodone = iodone; 1265 } else { 1266 mbp->b_flags = brw; 1267 mbp->b_iodone = NULL; 1268 } 1269 if (curlwp == uvm.pagedaemon_lwp) 1270 BIO_SETPRIO(mbp, BPRIO_TIMELIMITED); 1271 else if (async) 1272 BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL); 1273 else 1274 BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL); 1275 1276 bp = NULL; 1277 for (offset = startoffset; 1278 bytes > 0; 1279 offset += iobytes, bytes -= iobytes) { 1280 lbn = offset >> fs_bshift; 1281 error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run); 1282 if (error) { 1283 UVMHIST_LOG(ubchist, "VOP_BMAP() -> %d", error,0,0,0); 1284 skipbytes += bytes; 1285 bytes = 0; 1286 break; 1287 } 1288 1289 iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset, 1290 bytes); 1291 if (blkno == (daddr_t)-1) { 1292 if (!write) { 1293 memset((char *)kva + (offset - startoffset), 0, 1294 iobytes); 1295 } 1296 skipbytes += iobytes; 1297 continue; 1298 } 1299 1300 /* if it's really one i/o, don't make a second buf */ 1301 if (offset == startoffset && iobytes == bytes) { 1302 bp = mbp; 1303 } else { 1304 UVMHIST_LOG(ubchist, "vp %p bp %p num now %d", 1305 vp, bp, vp->v_numoutput, 0); 1306 bp = getiobuf(vp, true); 1307 nestiobuf_setup(mbp, bp, offset - startoffset, iobytes); 1308 } 1309 bp->b_lblkno = 0; 1310 1311 /* adjust physical blkno for partial blocks */ 1312 bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >> 1313 dev_bshift); 1314 UVMHIST_LOG(ubchist, 1315 "vp %p offset 0x%x bcount 0x%x blkno 0x%x", 1316 vp, offset, bp->b_bcount, bp->b_blkno); 1317 1318 VOP_STRATEGY(devvp, bp); 1319 } 1320 if (skipbytes) { 1321 UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0); 1322 } 1323 nestiobuf_done(mbp, skipbytes, error); 1324 if (async) { 1325 UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0); 1326 return (0); 1327 } 1328 UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0); 1329 error = biowait(mbp); 1330 s = splbio(); 1331 (*iodone)(mbp); 1332 splx(s); 1333 UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0); 1334 return (error); 1335 } 1336 1337 /* 1338 * VOP_PUTPAGES() for vnodes which never have pages. 1339 */ 1340 1341 int 1342 genfs_null_putpages(void *v) 1343 { 1344 struct vop_putpages_args /* { 1345 struct vnode *a_vp; 1346 voff_t a_offlo; 1347 voff_t a_offhi; 1348 int a_flags; 1349 } */ *ap = v; 1350 struct vnode *vp = ap->a_vp; 1351 1352 KASSERT(vp->v_uobj.uo_npages == 0); 1353 mutex_exit(&vp->v_interlock); 1354 return (0); 1355 } 1356 1357 int 1358 genfs_compat_getpages(void *v) 1359 { 1360 struct vop_getpages_args /* { 1361 struct vnode *a_vp; 1362 voff_t a_offset; 1363 struct vm_page **a_m; 1364 int *a_count; 1365 int a_centeridx; 1366 vm_prot_t a_access_type; 1367 int a_advice; 1368 int a_flags; 1369 } */ *ap = v; 1370 1371 off_t origoffset; 1372 struct vnode *vp = ap->a_vp; 1373 struct uvm_object *uobj = &vp->v_uobj; 1374 struct vm_page *pg, **pgs; 1375 vaddr_t kva; 1376 int i, error, orignpages, npages; 1377 struct iovec iov; 1378 struct uio uio; 1379 kauth_cred_t cred = curlwp->l_cred; 1380 bool write = (ap->a_access_type & VM_PROT_WRITE) != 0; 1381 1382 error = 0; 1383 origoffset = ap->a_offset; 1384 orignpages = *ap->a_count; 1385 pgs = ap->a_m; 1386 1387 if (write && (vp->v_iflag & VI_ONWORKLST) == 0) { 1388 vn_syncer_add_to_worklist(vp, filedelay); 1389 } 1390 if (ap->a_flags & PGO_LOCKED) { 1391 uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m, 1392 UFP_NOWAIT|UFP_NOALLOC| (write ? UFP_NORDONLY : 0)); 1393 1394 return (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0); 1395 } 1396 if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) { 1397 mutex_exit(&uobj->vmobjlock); 1398 return (EINVAL); 1399 } 1400 if ((ap->a_flags & PGO_SYNCIO) == 0) { 1401 mutex_exit(&uobj->vmobjlock); 1402 return 0; 1403 } 1404 npages = orignpages; 1405 uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL); 1406 mutex_exit(&uobj->vmobjlock); 1407 kva = uvm_pagermapin(pgs, npages, 1408 UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK); 1409 for (i = 0; i < npages; i++) { 1410 pg = pgs[i]; 1411 if ((pg->flags & PG_FAKE) == 0) { 1412 continue; 1413 } 1414 iov.iov_base = (char *)kva + (i << PAGE_SHIFT); 1415 iov.iov_len = PAGE_SIZE; 1416 uio.uio_iov = &iov; 1417 uio.uio_iovcnt = 1; 1418 uio.uio_offset = origoffset + (i << PAGE_SHIFT); 1419 uio.uio_rw = UIO_READ; 1420 uio.uio_resid = PAGE_SIZE; 1421 UIO_SETUP_SYSSPACE(&uio); 1422 /* XXX vn_lock */ 1423 error = VOP_READ(vp, &uio, 0, cred); 1424 if (error) { 1425 break; 1426 } 1427 if (uio.uio_resid) { 1428 memset(iov.iov_base, 0, uio.uio_resid); 1429 } 1430 } 1431 uvm_pagermapout(kva, npages); 1432 mutex_enter(&uobj->vmobjlock); 1433 mutex_enter(&uvm_pageqlock); 1434 for (i = 0; i < npages; i++) { 1435 pg = pgs[i]; 1436 if (error && (pg->flags & PG_FAKE) != 0) { 1437 pg->flags |= PG_RELEASED; 1438 } else { 1439 pmap_clear_modify(pg); 1440 uvm_pageactivate(pg); 1441 } 1442 } 1443 if (error) { 1444 uvm_page_unbusy(pgs, npages); 1445 } 1446 mutex_exit(&uvm_pageqlock); 1447 mutex_exit(&uobj->vmobjlock); 1448 return (error); 1449 } 1450 1451 int 1452 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, 1453 int flags) 1454 { 1455 off_t offset; 1456 struct iovec iov; 1457 struct uio uio; 1458 kauth_cred_t cred = curlwp->l_cred; 1459 struct buf *bp; 1460 vaddr_t kva; 1461 int error; 1462 1463 offset = pgs[0]->offset; 1464 kva = uvm_pagermapin(pgs, npages, 1465 UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK); 1466 1467 iov.iov_base = (void *)kva; 1468 iov.iov_len = npages << PAGE_SHIFT; 1469 uio.uio_iov = &iov; 1470 uio.uio_iovcnt = 1; 1471 uio.uio_offset = offset; 1472 uio.uio_rw = UIO_WRITE; 1473 uio.uio_resid = npages << PAGE_SHIFT; 1474 UIO_SETUP_SYSSPACE(&uio); 1475 /* XXX vn_lock */ 1476 error = VOP_WRITE(vp, &uio, 0, cred); 1477 1478 mutex_enter(&vp->v_interlock); 1479 vp->v_numoutput++; 1480 mutex_exit(&vp->v_interlock); 1481 1482 bp = getiobuf(vp, true); 1483 bp->b_cflags = BC_BUSY | BC_AGE; 1484 bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift; 1485 bp->b_data = (char *)kva; 1486 bp->b_bcount = npages << PAGE_SHIFT; 1487 bp->b_bufsize = npages << PAGE_SHIFT; 1488 bp->b_resid = 0; 1489 bp->b_error = error; 1490 uvm_aio_aiodone(bp); 1491 return (error); 1492 } 1493 1494 /* 1495 * Process a uio using direct I/O. If we reach a part of the request 1496 * which cannot be processed in this fashion for some reason, just return. 1497 * The caller must handle some additional part of the request using 1498 * buffered I/O before trying direct I/O again. 1499 */ 1500 1501 void 1502 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag) 1503 { 1504 struct vmspace *vs; 1505 struct iovec *iov; 1506 vaddr_t va; 1507 size_t len; 1508 const int mask = DEV_BSIZE - 1; 1509 int error; 1510 1511 /* 1512 * We only support direct I/O to user space for now. 1513 */ 1514 1515 if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) { 1516 return; 1517 } 1518 1519 /* 1520 * If the vnode is mapped, we would need to get the getpages lock 1521 * to stabilize the bmap, but then we would get into trouble whil e 1522 * locking the pages if the pages belong to this same vnode (or a 1523 * multi-vnode cascade to the same effect). Just fall back to 1524 * buffered I/O if the vnode is mapped to avoid this mess. 1525 */ 1526 1527 if (vp->v_vflag & VV_MAPPED) { 1528 return; 1529 } 1530 1531 /* 1532 * Do as much of the uio as possible with direct I/O. 1533 */ 1534 1535 vs = uio->uio_vmspace; 1536 while (uio->uio_resid) { 1537 iov = uio->uio_iov; 1538 if (iov->iov_len == 0) { 1539 uio->uio_iov++; 1540 uio->uio_iovcnt--; 1541 continue; 1542 } 1543 va = (vaddr_t)iov->iov_base; 1544 len = MIN(iov->iov_len, genfs_maxdio); 1545 len &= ~mask; 1546 1547 /* 1548 * If the next chunk is smaller than DEV_BSIZE or extends past 1549 * the current EOF, then fall back to buffered I/O. 1550 */ 1551 1552 if (len == 0 || uio->uio_offset + len > vp->v_size) { 1553 return; 1554 } 1555 1556 /* 1557 * Check alignment. The file offset must be at least 1558 * sector-aligned. The exact constraint on memory alignment 1559 * is very hardware-dependent, but requiring sector-aligned 1560 * addresses there too is safe. 1561 */ 1562 1563 if (uio->uio_offset & mask || va & mask) { 1564 return; 1565 } 1566 error = genfs_do_directio(vs, va, len, vp, uio->uio_offset, 1567 uio->uio_rw); 1568 if (error) { 1569 break; 1570 } 1571 iov->iov_base = (char *)iov->iov_base + len; 1572 iov->iov_len -= len; 1573 uio->uio_offset += len; 1574 uio->uio_resid -= len; 1575 } 1576 } 1577 1578 /* 1579 * Iodone routine for direct I/O. We don't do much here since the request is 1580 * always synchronous, so the caller will do most of the work after biowait(). 1581 */ 1582 1583 static void 1584 genfs_dio_iodone(struct buf *bp) 1585 { 1586 1587 KASSERT((bp->b_flags & B_ASYNC) == 0); 1588 if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) { 1589 mutex_enter(bp->b_objlock); 1590 vwakeup(bp); 1591 mutex_exit(bp->b_objlock); 1592 } 1593 putiobuf(bp); 1594 } 1595 1596 /* 1597 * Process one chunk of a direct I/O request. 1598 */ 1599 1600 static int 1601 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp, 1602 off_t off, enum uio_rw rw) 1603 { 1604 struct vm_map *map; 1605 struct pmap *upm, *kpm; 1606 size_t klen = round_page(uva + len) - trunc_page(uva); 1607 off_t spoff, epoff; 1608 vaddr_t kva, puva; 1609 paddr_t pa; 1610 vm_prot_t prot; 1611 int error, rv, poff, koff; 1612 const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | 1613 (rw == UIO_WRITE ? PGO_FREE : 0); 1614 1615 /* 1616 * For writes, verify that this range of the file already has fully 1617 * allocated backing store. If there are any holes, just punt and 1618 * make the caller take the buffered write path. 1619 */ 1620 1621 if (rw == UIO_WRITE) { 1622 daddr_t lbn, elbn, blkno; 1623 int bsize, bshift, run; 1624 1625 bshift = vp->v_mount->mnt_fs_bshift; 1626 bsize = 1 << bshift; 1627 lbn = off >> bshift; 1628 elbn = (off + len + bsize - 1) >> bshift; 1629 while (lbn < elbn) { 1630 error = VOP_BMAP(vp, lbn, NULL, &blkno, &run); 1631 if (error) { 1632 return error; 1633 } 1634 if (blkno == (daddr_t)-1) { 1635 return ENOSPC; 1636 } 1637 lbn += 1 + run; 1638 } 1639 } 1640 1641 /* 1642 * Flush any cached pages for parts of the file that we're about to 1643 * access. If we're writing, invalidate pages as well. 1644 */ 1645 1646 spoff = trunc_page(off); 1647 epoff = round_page(off + len); 1648 mutex_enter(&vp->v_interlock); 1649 error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags); 1650 if (error) { 1651 return error; 1652 } 1653 1654 /* 1655 * Wire the user pages and remap them into kernel memory. 1656 */ 1657 1658 prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ; 1659 error = uvm_vslock(vs, (void *)uva, len, prot); 1660 if (error) { 1661 return error; 1662 } 1663 1664 map = &vs->vm_map; 1665 upm = vm_map_pmap(map); 1666 kpm = vm_map_pmap(kernel_map); 1667 kva = uvm_km_alloc(kernel_map, klen, 0, 1668 UVM_KMF_VAONLY | UVM_KMF_WAITVA); 1669 puva = trunc_page(uva); 1670 for (poff = 0; poff < klen; poff += PAGE_SIZE) { 1671 rv = pmap_extract(upm, puva + poff, &pa); 1672 KASSERT(rv); 1673 pmap_enter(kpm, kva + poff, pa, prot, prot | PMAP_WIRED); 1674 } 1675 pmap_update(kpm); 1676 1677 /* 1678 * Do the I/O. 1679 */ 1680 1681 koff = uva - trunc_page(uva); 1682 error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw, 1683 genfs_dio_iodone); 1684 1685 /* 1686 * Tear down the kernel mapping. 1687 */ 1688 1689 pmap_remove(kpm, kva, kva + klen); 1690 pmap_update(kpm); 1691 uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY); 1692 1693 /* 1694 * Unwire the user pages. 1695 */ 1696 1697 uvm_vsunlock(vs, (void *)uva, len); 1698 return error; 1699 } 1700 1701