1 /* $NetBSD: genfs_io.c,v 1.61 2015/05/06 15:57:08 hannken 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 1.61 2015/05/06 15:57:08 hannken 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/vnode.h> 42 #include <sys/kmem.h> 43 #include <sys/kauth.h> 44 #include <sys/fstrans.h> 45 #include <sys/buf.h> 46 47 #include <miscfs/genfs/genfs.h> 48 #include <miscfs/genfs/genfs_node.h> 49 #include <miscfs/specfs/specdev.h> 50 51 #include <uvm/uvm.h> 52 #include <uvm/uvm_pager.h> 53 54 static int genfs_do_directio(struct vmspace *, vaddr_t, size_t, struct vnode *, 55 off_t, enum uio_rw); 56 static void genfs_dio_iodone(struct buf *); 57 58 static int genfs_getpages_read(struct vnode *, struct vm_page **, int, off_t, 59 off_t, bool, bool, bool, bool); 60 static int genfs_do_io(struct vnode *, off_t, vaddr_t, size_t, int, enum uio_rw, 61 void (*)(struct buf *)); 62 static void genfs_rel_pages(struct vm_page **, unsigned int); 63 static void genfs_markdirty(struct vnode *); 64 65 int genfs_maxdio = MAXPHYS; 66 67 static void 68 genfs_rel_pages(struct vm_page **pgs, unsigned int npages) 69 { 70 unsigned int i; 71 72 for (i = 0; i < npages; i++) { 73 struct vm_page *pg = pgs[i]; 74 75 if (pg == NULL || pg == PGO_DONTCARE) 76 continue; 77 KASSERT(uvm_page_locked_p(pg)); 78 if (pg->flags & PG_FAKE) { 79 pg->flags |= PG_RELEASED; 80 } 81 } 82 mutex_enter(&uvm_pageqlock); 83 uvm_page_unbusy(pgs, npages); 84 mutex_exit(&uvm_pageqlock); 85 } 86 87 static void 88 genfs_markdirty(struct vnode *vp) 89 { 90 struct genfs_node * const gp = VTOG(vp); 91 92 KASSERT(mutex_owned(vp->v_interlock)); 93 gp->g_dirtygen++; 94 if ((vp->v_iflag & VI_ONWORKLST) == 0) { 95 vn_syncer_add_to_worklist(vp, filedelay); 96 } 97 if ((vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP) { 98 vp->v_iflag |= VI_WRMAPDIRTY; 99 } 100 } 101 102 /* 103 * generic VM getpages routine. 104 * Return PG_BUSY pages for the given range, 105 * reading from backing store if necessary. 106 */ 107 108 int 109 genfs_getpages(void *v) 110 { 111 struct vop_getpages_args /* { 112 struct vnode *a_vp; 113 voff_t a_offset; 114 struct vm_page **a_m; 115 int *a_count; 116 int a_centeridx; 117 vm_prot_t a_access_type; 118 int a_advice; 119 int a_flags; 120 } */ * const ap = v; 121 122 off_t diskeof, memeof; 123 int i, error, npages; 124 const int flags = ap->a_flags; 125 struct vnode * const vp = ap->a_vp; 126 struct uvm_object * const uobj = &vp->v_uobj; 127 const bool async = (flags & PGO_SYNCIO) == 0; 128 const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0; 129 const bool overwrite = (flags & PGO_OVERWRITE) != 0; 130 const bool blockalloc = memwrite && (flags & PGO_NOBLOCKALLOC) == 0; 131 const bool glocked = (flags & PGO_GLOCKHELD) != 0; 132 const bool need_wapbl = blockalloc && vp->v_mount->mnt_wapbl; 133 bool has_trans_wapbl = false; 134 UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist); 135 136 UVMHIST_LOG(ubchist, "vp %p off 0x%x/%x count %d", 137 vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count); 138 139 KASSERT(vp->v_type == VREG || vp->v_type == VDIR || 140 vp->v_type == VLNK || vp->v_type == VBLK); 141 142 startover: 143 error = 0; 144 const voff_t origvsize = vp->v_size; 145 const off_t origoffset = ap->a_offset; 146 const int orignpages = *ap->a_count; 147 148 GOP_SIZE(vp, origvsize, &diskeof, 0); 149 if (flags & PGO_PASTEOF) { 150 off_t newsize; 151 #if defined(DIAGNOSTIC) 152 off_t writeeof; 153 #endif /* defined(DIAGNOSTIC) */ 154 155 newsize = MAX(origvsize, 156 origoffset + (orignpages << PAGE_SHIFT)); 157 GOP_SIZE(vp, newsize, &memeof, GOP_SIZE_MEM); 158 #if defined(DIAGNOSTIC) 159 GOP_SIZE(vp, vp->v_writesize, &writeeof, GOP_SIZE_MEM); 160 if (newsize > round_page(writeeof)) { 161 panic("%s: past eof: %" PRId64 " vs. %" PRId64, 162 __func__, newsize, round_page(writeeof)); 163 } 164 #endif /* defined(DIAGNOSTIC) */ 165 } else { 166 GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM); 167 } 168 KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages); 169 KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0); 170 KASSERT(orignpages > 0); 171 172 /* 173 * Bounds-check the request. 174 */ 175 176 if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) { 177 if ((flags & PGO_LOCKED) == 0) { 178 mutex_exit(uobj->vmobjlock); 179 } 180 UVMHIST_LOG(ubchist, "off 0x%x count %d goes past EOF 0x%x", 181 origoffset, *ap->a_count, memeof,0); 182 error = EINVAL; 183 goto out_err; 184 } 185 186 /* uobj is locked */ 187 188 if ((flags & PGO_NOTIMESTAMP) == 0 && 189 (vp->v_type != VBLK || 190 (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) { 191 int updflags = 0; 192 193 if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) { 194 updflags = GOP_UPDATE_ACCESSED; 195 } 196 if (memwrite) { 197 updflags |= GOP_UPDATE_MODIFIED; 198 } 199 if (updflags != 0) { 200 GOP_MARKUPDATE(vp, updflags); 201 } 202 } 203 204 /* 205 * For PGO_LOCKED requests, just return whatever's in memory. 206 */ 207 208 if (flags & PGO_LOCKED) { 209 int nfound; 210 struct vm_page *pg; 211 212 KASSERT(!glocked); 213 npages = *ap->a_count; 214 #if defined(DEBUG) 215 for (i = 0; i < npages; i++) { 216 pg = ap->a_m[i]; 217 KASSERT(pg == NULL || pg == PGO_DONTCARE); 218 } 219 #endif /* defined(DEBUG) */ 220 nfound = uvn_findpages(uobj, origoffset, &npages, 221 ap->a_m, UFP_NOWAIT|UFP_NOALLOC|(memwrite ? UFP_NORDONLY : 0)); 222 KASSERT(npages == *ap->a_count); 223 if (nfound == 0) { 224 error = EBUSY; 225 goto out_err; 226 } 227 if (!genfs_node_rdtrylock(vp)) { 228 genfs_rel_pages(ap->a_m, npages); 229 230 /* 231 * restore the array. 232 */ 233 234 for (i = 0; i < npages; i++) { 235 pg = ap->a_m[i]; 236 237 if (pg != NULL && pg != PGO_DONTCARE) { 238 ap->a_m[i] = NULL; 239 } 240 KASSERT(ap->a_m[i] == NULL || 241 ap->a_m[i] == PGO_DONTCARE); 242 } 243 } else { 244 genfs_node_unlock(vp); 245 } 246 error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0); 247 if (error == 0 && memwrite) { 248 genfs_markdirty(vp); 249 } 250 goto out_err; 251 } 252 mutex_exit(uobj->vmobjlock); 253 254 /* 255 * find the requested pages and make some simple checks. 256 * leave space in the page array for a whole block. 257 */ 258 259 const int fs_bshift = (vp->v_type != VBLK) ? 260 vp->v_mount->mnt_fs_bshift : DEV_BSHIFT; 261 const int fs_bsize = 1 << fs_bshift; 262 #define blk_mask (fs_bsize - 1) 263 #define trunc_blk(x) ((x) & ~blk_mask) 264 #define round_blk(x) (((x) + blk_mask) & ~blk_mask) 265 266 const int orignmempages = MIN(orignpages, 267 round_page(memeof - origoffset) >> PAGE_SHIFT); 268 npages = orignmempages; 269 const off_t startoffset = trunc_blk(origoffset); 270 const off_t endoffset = MIN( 271 round_page(round_blk(origoffset + (npages << PAGE_SHIFT))), 272 round_page(memeof)); 273 const int ridx = (origoffset - startoffset) >> PAGE_SHIFT; 274 275 const int pgs_size = sizeof(struct vm_page *) * 276 ((endoffset - startoffset) >> PAGE_SHIFT); 277 struct vm_page **pgs, *pgs_onstack[UBC_MAX_PAGES]; 278 279 if (pgs_size > sizeof(pgs_onstack)) { 280 pgs = kmem_zalloc(pgs_size, async ? KM_NOSLEEP : KM_SLEEP); 281 if (pgs == NULL) { 282 pgs = pgs_onstack; 283 error = ENOMEM; 284 goto out_err; 285 } 286 } else { 287 pgs = pgs_onstack; 288 (void)memset(pgs, 0, pgs_size); 289 } 290 291 UVMHIST_LOG(ubchist, "ridx %d npages %d startoff %ld endoff %ld", 292 ridx, npages, startoffset, endoffset); 293 294 if (!has_trans_wapbl) { 295 fstrans_start(vp->v_mount, FSTRANS_SHARED); 296 /* 297 * XXX: This assumes that we come here only via 298 * the mmio path 299 */ 300 if (need_wapbl) { 301 error = WAPBL_BEGIN(vp->v_mount); 302 if (error) { 303 fstrans_done(vp->v_mount); 304 goto out_err_free; 305 } 306 } 307 has_trans_wapbl = true; 308 } 309 310 /* 311 * hold g_glock to prevent a race with truncate. 312 * 313 * check if our idea of v_size is still valid. 314 */ 315 316 KASSERT(!glocked || genfs_node_wrlocked(vp)); 317 if (!glocked) { 318 if (blockalloc) { 319 genfs_node_wrlock(vp); 320 } else { 321 genfs_node_rdlock(vp); 322 } 323 } 324 mutex_enter(uobj->vmobjlock); 325 if (vp->v_size < origvsize) { 326 if (!glocked) { 327 genfs_node_unlock(vp); 328 } 329 if (pgs != pgs_onstack) 330 kmem_free(pgs, pgs_size); 331 goto startover; 332 } 333 334 if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx], 335 async ? UFP_NOWAIT : UFP_ALL) != orignmempages) { 336 if (!glocked) { 337 genfs_node_unlock(vp); 338 } 339 KASSERT(async != 0); 340 genfs_rel_pages(&pgs[ridx], orignmempages); 341 mutex_exit(uobj->vmobjlock); 342 error = EBUSY; 343 goto out_err_free; 344 } 345 346 /* 347 * if the pages are already resident, just return them. 348 */ 349 350 for (i = 0; i < npages; i++) { 351 struct vm_page *pg = pgs[ridx + i]; 352 353 if ((pg->flags & PG_FAKE) || 354 (blockalloc && (pg->flags & PG_RDONLY))) { 355 break; 356 } 357 } 358 if (i == npages) { 359 if (!glocked) { 360 genfs_node_unlock(vp); 361 } 362 UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0); 363 npages += ridx; 364 goto out; 365 } 366 367 /* 368 * if PGO_OVERWRITE is set, don't bother reading the pages. 369 */ 370 371 if (overwrite) { 372 if (!glocked) { 373 genfs_node_unlock(vp); 374 } 375 UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0); 376 377 for (i = 0; i < npages; i++) { 378 struct vm_page *pg = pgs[ridx + i]; 379 380 pg->flags &= ~(PG_RDONLY|PG_CLEAN); 381 } 382 npages += ridx; 383 goto out; 384 } 385 386 /* 387 * the page wasn't resident and we're not overwriting, 388 * so we're going to have to do some i/o. 389 * find any additional pages needed to cover the expanded range. 390 */ 391 392 npages = (endoffset - startoffset) >> PAGE_SHIFT; 393 if (startoffset != origoffset || npages != orignmempages) { 394 int npgs; 395 396 /* 397 * we need to avoid deadlocks caused by locking 398 * additional pages at lower offsets than pages we 399 * already have locked. unlock them all and start over. 400 */ 401 402 genfs_rel_pages(&pgs[ridx], orignmempages); 403 memset(pgs, 0, pgs_size); 404 405 UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x", 406 startoffset, endoffset, 0,0); 407 npgs = npages; 408 if (uvn_findpages(uobj, startoffset, &npgs, pgs, 409 async ? UFP_NOWAIT : UFP_ALL) != npages) { 410 if (!glocked) { 411 genfs_node_unlock(vp); 412 } 413 KASSERT(async != 0); 414 genfs_rel_pages(pgs, npages); 415 mutex_exit(uobj->vmobjlock); 416 error = EBUSY; 417 goto out_err_free; 418 } 419 } 420 421 mutex_exit(uobj->vmobjlock); 422 error = genfs_getpages_read(vp, pgs, npages, startoffset, diskeof, 423 async, memwrite, blockalloc, glocked); 424 if (error == 0 && async) 425 goto out_err_free; 426 if (!glocked) { 427 genfs_node_unlock(vp); 428 } 429 mutex_enter(uobj->vmobjlock); 430 431 /* 432 * we're almost done! release the pages... 433 * for errors, we free the pages. 434 * otherwise we activate them and mark them as valid and clean. 435 * also, unbusy pages that were not actually requested. 436 */ 437 438 if (error) { 439 genfs_rel_pages(pgs, npages); 440 mutex_exit(uobj->vmobjlock); 441 UVMHIST_LOG(ubchist, "returning error %d", error,0,0,0); 442 goto out_err_free; 443 } 444 445 out: 446 UVMHIST_LOG(ubchist, "succeeding, npages %d", npages,0,0,0); 447 error = 0; 448 mutex_enter(&uvm_pageqlock); 449 for (i = 0; i < npages; i++) { 450 struct vm_page *pg = pgs[i]; 451 if (pg == NULL) { 452 continue; 453 } 454 UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x", 455 pg, pg->flags, 0,0); 456 if (pg->flags & PG_FAKE && !overwrite) { 457 pg->flags &= ~(PG_FAKE); 458 pmap_clear_modify(pgs[i]); 459 } 460 KASSERT(!memwrite || !blockalloc || (pg->flags & PG_RDONLY) == 0); 461 if (i < ridx || i >= ridx + orignmempages || async) { 462 UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x", 463 pg, pg->offset,0,0); 464 if (pg->flags & PG_WANTED) { 465 wakeup(pg); 466 } 467 if (pg->flags & PG_FAKE) { 468 KASSERT(overwrite); 469 uvm_pagezero(pg); 470 } 471 if (pg->flags & PG_RELEASED) { 472 uvm_pagefree(pg); 473 continue; 474 } 475 uvm_pageenqueue(pg); 476 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE); 477 UVM_PAGE_OWN(pg, NULL); 478 } 479 } 480 mutex_exit(&uvm_pageqlock); 481 if (memwrite) { 482 genfs_markdirty(vp); 483 } 484 mutex_exit(uobj->vmobjlock); 485 if (ap->a_m != NULL) { 486 memcpy(ap->a_m, &pgs[ridx], 487 orignmempages * sizeof(struct vm_page *)); 488 } 489 490 out_err_free: 491 if (pgs != NULL && pgs != pgs_onstack) 492 kmem_free(pgs, pgs_size); 493 out_err: 494 if (has_trans_wapbl) { 495 if (need_wapbl) 496 WAPBL_END(vp->v_mount); 497 fstrans_done(vp->v_mount); 498 } 499 return error; 500 } 501 502 /* 503 * genfs_getpages_read: Read the pages in with VOP_BMAP/VOP_STRATEGY. 504 */ 505 static int 506 genfs_getpages_read(struct vnode *vp, struct vm_page **pgs, int npages, 507 off_t startoffset, off_t diskeof, 508 bool async, bool memwrite, bool blockalloc, bool glocked) 509 { 510 struct uvm_object * const uobj = &vp->v_uobj; 511 const int fs_bshift = (vp->v_type != VBLK) ? 512 vp->v_mount->mnt_fs_bshift : DEV_BSHIFT; 513 const int dev_bshift = (vp->v_type != VBLK) ? 514 vp->v_mount->mnt_dev_bshift : DEV_BSHIFT; 515 kauth_cred_t const cred = curlwp->l_cred; /* XXXUBC curlwp */ 516 size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes; 517 vaddr_t kva; 518 struct buf *bp, *mbp; 519 bool sawhole = false; 520 int i; 521 int error = 0; 522 523 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 524 525 /* 526 * read the desired page(s). 527 */ 528 529 totalbytes = npages << PAGE_SHIFT; 530 bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0)); 531 tailbytes = totalbytes - bytes; 532 skipbytes = 0; 533 534 kva = uvm_pagermapin(pgs, npages, 535 UVMPAGER_MAPIN_READ | (async ? 0 : UVMPAGER_MAPIN_WAITOK)); 536 if (kva == 0) 537 return EBUSY; 538 539 mbp = getiobuf(vp, true); 540 mbp->b_bufsize = totalbytes; 541 mbp->b_data = (void *)kva; 542 mbp->b_resid = mbp->b_bcount = bytes; 543 mbp->b_cflags = BC_BUSY; 544 if (async) { 545 mbp->b_flags = B_READ | B_ASYNC; 546 mbp->b_iodone = uvm_aio_biodone; 547 } else { 548 mbp->b_flags = B_READ; 549 mbp->b_iodone = NULL; 550 } 551 if (async) 552 BIO_SETPRIO(mbp, BPRIO_TIMELIMITED); 553 else 554 BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL); 555 556 /* 557 * if EOF is in the middle of the range, zero the part past EOF. 558 * skip over pages which are not PG_FAKE since in that case they have 559 * valid data that we need to preserve. 560 */ 561 562 tailstart = bytes; 563 while (tailbytes > 0) { 564 const int len = PAGE_SIZE - (tailstart & PAGE_MASK); 565 566 KASSERT(len <= tailbytes); 567 if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) { 568 memset((void *)(kva + tailstart), 0, len); 569 UVMHIST_LOG(ubchist, "tailbytes %p 0x%x 0x%x", 570 kva, tailstart, len, 0); 571 } 572 tailstart += len; 573 tailbytes -= len; 574 } 575 576 /* 577 * now loop over the pages, reading as needed. 578 */ 579 580 bp = NULL; 581 off_t offset; 582 for (offset = startoffset; 583 bytes > 0; 584 offset += iobytes, bytes -= iobytes) { 585 int run; 586 daddr_t lbn, blkno; 587 int pidx; 588 struct vnode *devvp; 589 590 /* 591 * skip pages which don't need to be read. 592 */ 593 594 pidx = (offset - startoffset) >> PAGE_SHIFT; 595 while ((pgs[pidx]->flags & PG_FAKE) == 0) { 596 size_t b; 597 598 KASSERT((offset & (PAGE_SIZE - 1)) == 0); 599 if ((pgs[pidx]->flags & PG_RDONLY)) { 600 sawhole = true; 601 } 602 b = MIN(PAGE_SIZE, bytes); 603 offset += b; 604 bytes -= b; 605 skipbytes += b; 606 pidx++; 607 UVMHIST_LOG(ubchist, "skipping, new offset 0x%x", 608 offset, 0,0,0); 609 if (bytes == 0) { 610 goto loopdone; 611 } 612 } 613 614 /* 615 * bmap the file to find out the blkno to read from and 616 * how much we can read in one i/o. if bmap returns an error, 617 * skip the rest of the top-level i/o. 618 */ 619 620 lbn = offset >> fs_bshift; 621 error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run); 622 if (error) { 623 UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n", 624 lbn,error,0,0); 625 skipbytes += bytes; 626 bytes = 0; 627 goto loopdone; 628 } 629 630 /* 631 * see how many pages can be read with this i/o. 632 * reduce the i/o size if necessary to avoid 633 * overwriting pages with valid data. 634 */ 635 636 iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset, 637 bytes); 638 if (offset + iobytes > round_page(offset)) { 639 int pcount; 640 641 pcount = 1; 642 while (pidx + pcount < npages && 643 pgs[pidx + pcount]->flags & PG_FAKE) { 644 pcount++; 645 } 646 iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) - 647 (offset - trunc_page(offset))); 648 } 649 650 /* 651 * if this block isn't allocated, zero it instead of 652 * reading it. unless we are going to allocate blocks, 653 * mark the pages we zeroed PG_RDONLY. 654 */ 655 656 if (blkno == (daddr_t)-1) { 657 int holepages = (round_page(offset + iobytes) - 658 trunc_page(offset)) >> PAGE_SHIFT; 659 UVMHIST_LOG(ubchist, "lbn 0x%x -> HOLE", lbn,0,0,0); 660 661 sawhole = true; 662 memset((char *)kva + (offset - startoffset), 0, 663 iobytes); 664 skipbytes += iobytes; 665 666 mutex_enter(uobj->vmobjlock); 667 for (i = 0; i < holepages; i++) { 668 if (memwrite) { 669 pgs[pidx + i]->flags &= ~PG_CLEAN; 670 } 671 if (!blockalloc) { 672 pgs[pidx + i]->flags |= PG_RDONLY; 673 } 674 } 675 mutex_exit(uobj->vmobjlock); 676 continue; 677 } 678 679 /* 680 * allocate a sub-buf for this piece of the i/o 681 * (or just use mbp if there's only 1 piece), 682 * and start it going. 683 */ 684 685 if (offset == startoffset && iobytes == bytes) { 686 bp = mbp; 687 } else { 688 UVMHIST_LOG(ubchist, "vp %p bp %p num now %d", 689 vp, bp, vp->v_numoutput, 0); 690 bp = getiobuf(vp, true); 691 nestiobuf_setup(mbp, bp, offset - startoffset, iobytes); 692 } 693 bp->b_lblkno = 0; 694 695 /* adjust physical blkno for partial blocks */ 696 bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >> 697 dev_bshift); 698 699 UVMHIST_LOG(ubchist, 700 "bp %p offset 0x%x bcount 0x%x blkno 0x%x", 701 bp, offset, bp->b_bcount, bp->b_blkno); 702 703 VOP_STRATEGY(devvp, bp); 704 } 705 706 loopdone: 707 nestiobuf_done(mbp, skipbytes, error); 708 if (async) { 709 UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0); 710 if (!glocked) { 711 genfs_node_unlock(vp); 712 } 713 return 0; 714 } 715 if (bp != NULL) { 716 error = biowait(mbp); 717 } 718 719 /* Remove the mapping (make KVA available as soon as possible) */ 720 uvm_pagermapout(kva, npages); 721 722 /* 723 * if this we encountered a hole then we have to do a little more work. 724 * for read faults, we marked the page PG_RDONLY so that future 725 * write accesses to the page will fault again. 726 * for write faults, we must make sure that the backing store for 727 * the page is completely allocated while the pages are locked. 728 */ 729 730 if (!error && sawhole && blockalloc) { 731 error = GOP_ALLOC(vp, startoffset, 732 npages << PAGE_SHIFT, 0, cred); 733 UVMHIST_LOG(ubchist, "gop_alloc off 0x%x/0x%x -> %d", 734 startoffset, npages << PAGE_SHIFT, error,0); 735 if (!error) { 736 mutex_enter(uobj->vmobjlock); 737 for (i = 0; i < npages; i++) { 738 struct vm_page *pg = pgs[i]; 739 740 if (pg == NULL) { 741 continue; 742 } 743 pg->flags &= ~(PG_CLEAN|PG_RDONLY); 744 UVMHIST_LOG(ubchist, "mark dirty pg %p", 745 pg,0,0,0); 746 } 747 mutex_exit(uobj->vmobjlock); 748 } 749 } 750 751 putiobuf(mbp); 752 return error; 753 } 754 755 /* 756 * generic VM putpages routine. 757 * Write the given range of pages to backing store. 758 * 759 * => "offhi == 0" means flush all pages at or after "offlo". 760 * => object should be locked by caller. we return with the 761 * object unlocked. 762 * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O). 763 * thus, a caller might want to unlock higher level resources 764 * (e.g. vm_map) before calling flush. 765 * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block 766 * => if PGO_ALLPAGES is set, then all pages in the object will be processed. 767 * => NOTE: we rely on the fact that the object's memq is a TAILQ and 768 * that new pages are inserted on the tail end of the list. thus, 769 * we can make a complete pass through the object in one go by starting 770 * at the head and working towards the tail (new pages are put in 771 * front of us). 772 * => NOTE: we are allowed to lock the page queues, so the caller 773 * must not be holding the page queue lock. 774 * 775 * note on "cleaning" object and PG_BUSY pages: 776 * this routine is holding the lock on the object. the only time 777 * that it can run into a PG_BUSY page that it does not own is if 778 * some other process has started I/O on the page (e.g. either 779 * a pagein, or a pageout). if the PG_BUSY page is being paged 780 * in, then it can not be dirty (!PG_CLEAN) because no one has 781 * had a chance to modify it yet. if the PG_BUSY page is being 782 * paged out then it means that someone else has already started 783 * cleaning the page for us (how nice!). in this case, if we 784 * have syncio specified, then after we make our pass through the 785 * object we need to wait for the other PG_BUSY pages to clear 786 * off (i.e. we need to do an iosync). also note that once a 787 * page is PG_BUSY it must stay in its object until it is un-busyed. 788 * 789 * note on page traversal: 790 * we can traverse the pages in an object either by going down the 791 * linked list in "uobj->memq", or we can go over the address range 792 * by page doing hash table lookups for each address. depending 793 * on how many pages are in the object it may be cheaper to do one 794 * or the other. we set "by_list" to true if we are using memq. 795 * if the cost of a hash lookup was equal to the cost of the list 796 * traversal we could compare the number of pages in the start->stop 797 * range to the total number of pages in the object. however, it 798 * seems that a hash table lookup is more expensive than the linked 799 * list traversal, so we multiply the number of pages in the 800 * range by an estimate of the relatively higher cost of the hash lookup. 801 */ 802 803 int 804 genfs_putpages(void *v) 805 { 806 struct vop_putpages_args /* { 807 struct vnode *a_vp; 808 voff_t a_offlo; 809 voff_t a_offhi; 810 int a_flags; 811 } */ * const ap = v; 812 813 return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi, 814 ap->a_flags, NULL); 815 } 816 817 int 818 genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff, 819 int origflags, struct vm_page **busypg) 820 { 821 struct uvm_object * const uobj = &vp->v_uobj; 822 kmutex_t * const slock = uobj->vmobjlock; 823 off_t off; 824 /* Even for strange MAXPHYS, the shift rounds down to a page */ 825 #define maxpages (MAXPHYS >> PAGE_SHIFT) 826 int i, error, npages, nback; 827 int freeflag; 828 struct vm_page *pgs[maxpages], *pg, *nextpg, *tpg, curmp, endmp; 829 bool wasclean, by_list, needs_clean, yld; 830 bool async = (origflags & PGO_SYNCIO) == 0; 831 bool pagedaemon = curlwp == uvm.pagedaemon_lwp; 832 struct lwp * const l = curlwp ? curlwp : &lwp0; 833 struct genfs_node * const gp = VTOG(vp); 834 int flags; 835 int dirtygen; 836 bool modified; 837 bool need_wapbl; 838 bool has_trans; 839 bool cleanall; 840 bool onworklst; 841 842 UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist); 843 844 KASSERT(origflags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)); 845 KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0); 846 KASSERT(startoff < endoff || endoff == 0); 847 848 UVMHIST_LOG(ubchist, "vp %p pages %d off 0x%x len 0x%x", 849 vp, uobj->uo_npages, startoff, endoff - startoff); 850 851 has_trans = false; 852 need_wapbl = (!pagedaemon && vp->v_mount && vp->v_mount->mnt_wapbl && 853 (origflags & PGO_JOURNALLOCKED) == 0); 854 855 retry: 856 modified = false; 857 flags = origflags; 858 KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 || 859 (vp->v_iflag & VI_WRMAPDIRTY) == 0); 860 if (uobj->uo_npages == 0) { 861 if (vp->v_iflag & VI_ONWORKLST) { 862 vp->v_iflag &= ~VI_WRMAPDIRTY; 863 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) 864 vn_syncer_remove_from_worklist(vp); 865 } 866 if (has_trans) { 867 if (need_wapbl) 868 WAPBL_END(vp->v_mount); 869 fstrans_done(vp->v_mount); 870 } 871 mutex_exit(slock); 872 return (0); 873 } 874 875 /* 876 * the vnode has pages, set up to process the request. 877 */ 878 879 if (!has_trans && (flags & PGO_CLEANIT) != 0) { 880 mutex_exit(slock); 881 if (pagedaemon) { 882 error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY); 883 if (error) 884 return error; 885 } else 886 fstrans_start(vp->v_mount, FSTRANS_LAZY); 887 if (need_wapbl) { 888 error = WAPBL_BEGIN(vp->v_mount); 889 if (error) { 890 fstrans_done(vp->v_mount); 891 return error; 892 } 893 } 894 has_trans = true; 895 mutex_enter(slock); 896 goto retry; 897 } 898 899 error = 0; 900 wasclean = (vp->v_numoutput == 0); 901 off = startoff; 902 if (endoff == 0 || flags & PGO_ALLPAGES) { 903 endoff = trunc_page(LLONG_MAX); 904 } 905 by_list = (uobj->uo_npages <= 906 ((endoff - startoff) >> PAGE_SHIFT) * UVM_PAGE_TREE_PENALTY); 907 908 /* 909 * if this vnode is known not to have dirty pages, 910 * don't bother to clean it out. 911 */ 912 913 if ((vp->v_iflag & VI_ONWORKLST) == 0) { 914 #if !defined(DEBUG) 915 if ((flags & (PGO_FREE|PGO_DEACTIVATE)) == 0) { 916 goto skip_scan; 917 } 918 #endif /* !defined(DEBUG) */ 919 flags &= ~PGO_CLEANIT; 920 } 921 922 /* 923 * start the loop. when scanning by list, hold the last page 924 * in the list before we start. pages allocated after we start 925 * will be added to the end of the list, so we can stop at the 926 * current last page. 927 */ 928 929 cleanall = (flags & PGO_CLEANIT) != 0 && wasclean && 930 startoff == 0 && endoff == trunc_page(LLONG_MAX) && 931 (vp->v_iflag & VI_ONWORKLST) != 0; 932 dirtygen = gp->g_dirtygen; 933 freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED; 934 if (by_list) { 935 curmp.flags = PG_MARKER; 936 endmp.flags = PG_MARKER; 937 pg = TAILQ_FIRST(&uobj->memq); 938 TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq.queue); 939 } else { 940 pg = uvm_pagelookup(uobj, off); 941 } 942 nextpg = NULL; 943 while (by_list || off < endoff) { 944 945 /* 946 * if the current page is not interesting, move on to the next. 947 */ 948 949 KASSERT(pg == NULL || pg->uobject == uobj || 950 (pg->flags & PG_MARKER) != 0); 951 KASSERT(pg == NULL || 952 (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 || 953 (pg->flags & (PG_BUSY|PG_MARKER)) != 0); 954 if (by_list) { 955 if (pg == &endmp) { 956 break; 957 } 958 if (pg->flags & PG_MARKER) { 959 pg = TAILQ_NEXT(pg, listq.queue); 960 continue; 961 } 962 if (pg->offset < startoff || pg->offset >= endoff || 963 pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 964 if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 965 wasclean = false; 966 } 967 pg = TAILQ_NEXT(pg, listq.queue); 968 continue; 969 } 970 off = pg->offset; 971 } else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) { 972 if (pg != NULL) { 973 wasclean = false; 974 } 975 off += PAGE_SIZE; 976 if (off < endoff) { 977 pg = uvm_pagelookup(uobj, off); 978 } 979 continue; 980 } 981 982 /* 983 * if the current page needs to be cleaned and it's busy, 984 * wait for it to become unbusy. 985 */ 986 987 yld = (l->l_cpu->ci_schedstate.spc_flags & 988 SPCF_SHOULDYIELD) && !pagedaemon; 989 if (pg->flags & PG_BUSY || yld) { 990 UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0); 991 if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) { 992 UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0); 993 error = EDEADLK; 994 if (busypg != NULL) 995 *busypg = pg; 996 break; 997 } 998 if (pagedaemon) { 999 /* 1000 * someone has taken the page while we 1001 * dropped the lock for fstrans_start. 1002 */ 1003 break; 1004 } 1005 if (by_list) { 1006 TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue); 1007 UVMHIST_LOG(ubchist, "curmp next %p", 1008 TAILQ_NEXT(&curmp, listq.queue), 0,0,0); 1009 } 1010 if (yld) { 1011 mutex_exit(slock); 1012 preempt(); 1013 mutex_enter(slock); 1014 } else { 1015 pg->flags |= PG_WANTED; 1016 UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0); 1017 mutex_enter(slock); 1018 } 1019 if (by_list) { 1020 UVMHIST_LOG(ubchist, "after next %p", 1021 TAILQ_NEXT(&curmp, listq.queue), 0,0,0); 1022 pg = TAILQ_NEXT(&curmp, listq.queue); 1023 TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue); 1024 } else { 1025 pg = uvm_pagelookup(uobj, off); 1026 } 1027 continue; 1028 } 1029 1030 /* 1031 * if we're freeing, remove all mappings of the page now. 1032 * if we're cleaning, check if the page is needs to be cleaned. 1033 */ 1034 1035 if (flags & PGO_FREE) { 1036 pmap_page_protect(pg, VM_PROT_NONE); 1037 } else if (flags & PGO_CLEANIT) { 1038 1039 /* 1040 * if we still have some hope to pull this vnode off 1041 * from the syncer queue, write-protect the page. 1042 */ 1043 1044 if (cleanall && wasclean && 1045 gp->g_dirtygen == dirtygen) { 1046 1047 /* 1048 * uobj pages get wired only by uvm_fault 1049 * where uobj is locked. 1050 */ 1051 1052 if (pg->wire_count == 0) { 1053 pmap_page_protect(pg, 1054 VM_PROT_READ|VM_PROT_EXECUTE); 1055 } else { 1056 cleanall = false; 1057 } 1058 } 1059 } 1060 1061 if (flags & PGO_CLEANIT) { 1062 needs_clean = pmap_clear_modify(pg) || 1063 (pg->flags & PG_CLEAN) == 0; 1064 pg->flags |= PG_CLEAN; 1065 } else { 1066 needs_clean = false; 1067 } 1068 1069 /* 1070 * if we're cleaning, build a cluster. 1071 * the cluster will consist of pages which are currently dirty, 1072 * but they will be returned to us marked clean. 1073 * if not cleaning, just operate on the one page. 1074 */ 1075 1076 if (needs_clean) { 1077 KDASSERT((vp->v_iflag & VI_ONWORKLST)); 1078 wasclean = false; 1079 memset(pgs, 0, sizeof(pgs)); 1080 pg->flags |= PG_BUSY; 1081 UVM_PAGE_OWN(pg, "genfs_putpages"); 1082 1083 /* 1084 * first look backward. 1085 */ 1086 1087 npages = MIN(maxpages >> 1, off >> PAGE_SHIFT); 1088 nback = npages; 1089 uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0], 1090 UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD); 1091 if (nback) { 1092 memmove(&pgs[0], &pgs[npages - nback], 1093 nback * sizeof(pgs[0])); 1094 if (npages - nback < nback) 1095 memset(&pgs[nback], 0, 1096 (npages - nback) * sizeof(pgs[0])); 1097 else 1098 memset(&pgs[npages - nback], 0, 1099 nback * sizeof(pgs[0])); 1100 } 1101 1102 /* 1103 * then plug in our page of interest. 1104 */ 1105 1106 pgs[nback] = pg; 1107 1108 /* 1109 * then look forward to fill in the remaining space in 1110 * the array of pages. 1111 */ 1112 1113 npages = maxpages - nback - 1; 1114 uvn_findpages(uobj, off + PAGE_SIZE, &npages, 1115 &pgs[nback + 1], 1116 UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY); 1117 npages += nback + 1; 1118 } else { 1119 pgs[0] = pg; 1120 npages = 1; 1121 nback = 0; 1122 } 1123 1124 /* 1125 * apply FREE or DEACTIVATE options if requested. 1126 */ 1127 1128 if (flags & (PGO_DEACTIVATE|PGO_FREE)) { 1129 mutex_enter(&uvm_pageqlock); 1130 } 1131 for (i = 0; i < npages; i++) { 1132 tpg = pgs[i]; 1133 KASSERT(tpg->uobject == uobj); 1134 if (by_list && tpg == TAILQ_NEXT(pg, listq.queue)) 1135 pg = tpg; 1136 if (tpg->offset < startoff || tpg->offset >= endoff) 1137 continue; 1138 if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) { 1139 uvm_pagedeactivate(tpg); 1140 } else if (flags & PGO_FREE) { 1141 pmap_page_protect(tpg, VM_PROT_NONE); 1142 if (tpg->flags & PG_BUSY) { 1143 tpg->flags |= freeflag; 1144 if (pagedaemon) { 1145 uvm_pageout_start(1); 1146 uvm_pagedequeue(tpg); 1147 } 1148 } else { 1149 1150 /* 1151 * ``page is not busy'' 1152 * implies that npages is 1 1153 * and needs_clean is false. 1154 */ 1155 1156 nextpg = TAILQ_NEXT(tpg, listq.queue); 1157 uvm_pagefree(tpg); 1158 if (pagedaemon) 1159 uvmexp.pdfreed++; 1160 } 1161 } 1162 } 1163 if (flags & (PGO_DEACTIVATE|PGO_FREE)) { 1164 mutex_exit(&uvm_pageqlock); 1165 } 1166 if (needs_clean) { 1167 modified = true; 1168 1169 /* 1170 * start the i/o. if we're traversing by list, 1171 * keep our place in the list with a marker page. 1172 */ 1173 1174 if (by_list) { 1175 TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp, 1176 listq.queue); 1177 } 1178 mutex_exit(slock); 1179 error = GOP_WRITE(vp, pgs, npages, flags); 1180 mutex_enter(slock); 1181 if (by_list) { 1182 pg = TAILQ_NEXT(&curmp, listq.queue); 1183 TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue); 1184 } 1185 if (error) { 1186 break; 1187 } 1188 if (by_list) { 1189 continue; 1190 } 1191 } 1192 1193 /* 1194 * find the next page and continue if there was no error. 1195 */ 1196 1197 if (by_list) { 1198 if (nextpg) { 1199 pg = nextpg; 1200 nextpg = NULL; 1201 } else { 1202 pg = TAILQ_NEXT(pg, listq.queue); 1203 } 1204 } else { 1205 off += (npages - nback) << PAGE_SHIFT; 1206 if (off < endoff) { 1207 pg = uvm_pagelookup(uobj, off); 1208 } 1209 } 1210 } 1211 if (by_list) { 1212 TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue); 1213 } 1214 1215 if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 && 1216 (vp->v_type != VBLK || 1217 (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) { 1218 GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED); 1219 } 1220 1221 /* 1222 * if we're cleaning and there was nothing to clean, 1223 * take us off the syncer list. if we started any i/o 1224 * and we're doing sync i/o, wait for all writes to finish. 1225 */ 1226 1227 if (cleanall && wasclean && gp->g_dirtygen == dirtygen && 1228 (vp->v_iflag & VI_ONWORKLST) != 0) { 1229 #if defined(DEBUG) 1230 TAILQ_FOREACH(pg, &uobj->memq, listq.queue) { 1231 if ((pg->flags & (PG_FAKE | PG_MARKER)) != 0) { 1232 continue; 1233 } 1234 if ((pg->flags & PG_CLEAN) == 0) { 1235 printf("%s: %p: !CLEAN\n", __func__, pg); 1236 } 1237 if (pmap_is_modified(pg)) { 1238 printf("%s: %p: modified\n", __func__, pg); 1239 } 1240 } 1241 #endif /* defined(DEBUG) */ 1242 vp->v_iflag &= ~VI_WRMAPDIRTY; 1243 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) 1244 vn_syncer_remove_from_worklist(vp); 1245 } 1246 1247 #if !defined(DEBUG) 1248 skip_scan: 1249 #endif /* !defined(DEBUG) */ 1250 1251 /* Wait for output to complete. */ 1252 if (!wasclean && !async && vp->v_numoutput != 0) { 1253 while (vp->v_numoutput != 0) 1254 cv_wait(&vp->v_cv, slock); 1255 } 1256 onworklst = (vp->v_iflag & VI_ONWORKLST) != 0; 1257 mutex_exit(slock); 1258 1259 if ((flags & PGO_RECLAIM) != 0 && onworklst) { 1260 /* 1261 * in the case of PGO_RECLAIM, ensure to make the vnode clean. 1262 * retrying is not a big deal because, in many cases, 1263 * uobj->uo_npages is already 0 here. 1264 */ 1265 mutex_enter(slock); 1266 goto retry; 1267 } 1268 1269 if (has_trans) { 1270 if (need_wapbl) 1271 WAPBL_END(vp->v_mount); 1272 fstrans_done(vp->v_mount); 1273 } 1274 1275 return (error); 1276 } 1277 1278 int 1279 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags) 1280 { 1281 off_t off; 1282 vaddr_t kva; 1283 size_t len; 1284 int error; 1285 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1286 1287 UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x", 1288 vp, pgs, npages, flags); 1289 1290 off = pgs[0]->offset; 1291 kva = uvm_pagermapin(pgs, npages, 1292 UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK); 1293 len = npages << PAGE_SHIFT; 1294 1295 error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE, 1296 uvm_aio_biodone); 1297 1298 return error; 1299 } 1300 1301 int 1302 genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages, int flags) 1303 { 1304 off_t off; 1305 vaddr_t kva; 1306 size_t len; 1307 int error; 1308 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1309 1310 UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x", 1311 vp, pgs, npages, flags); 1312 1313 off = pgs[0]->offset; 1314 kva = uvm_pagermapin(pgs, npages, 1315 UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK); 1316 len = npages << PAGE_SHIFT; 1317 1318 error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE, 1319 uvm_aio_biodone); 1320 1321 return error; 1322 } 1323 1324 /* 1325 * Backend routine for doing I/O to vnode pages. Pages are already locked 1326 * and mapped into kernel memory. Here we just look up the underlying 1327 * device block addresses and call the strategy routine. 1328 */ 1329 1330 static int 1331 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags, 1332 enum uio_rw rw, void (*iodone)(struct buf *)) 1333 { 1334 int s, error; 1335 int fs_bshift, dev_bshift; 1336 off_t eof, offset, startoffset; 1337 size_t bytes, iobytes, skipbytes; 1338 struct buf *mbp, *bp; 1339 const bool async = (flags & PGO_SYNCIO) == 0; 1340 const bool lazy = (flags & PGO_LAZY) == 0; 1341 const bool iowrite = rw == UIO_WRITE; 1342 const int brw = iowrite ? B_WRITE : B_READ; 1343 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1344 1345 UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x", 1346 vp, kva, len, flags); 1347 1348 KASSERT(vp->v_size <= vp->v_writesize); 1349 GOP_SIZE(vp, vp->v_writesize, &eof, 0); 1350 if (vp->v_type != VBLK) { 1351 fs_bshift = vp->v_mount->mnt_fs_bshift; 1352 dev_bshift = vp->v_mount->mnt_dev_bshift; 1353 } else { 1354 fs_bshift = DEV_BSHIFT; 1355 dev_bshift = DEV_BSHIFT; 1356 } 1357 error = 0; 1358 startoffset = off; 1359 bytes = MIN(len, eof - startoffset); 1360 skipbytes = 0; 1361 KASSERT(bytes != 0); 1362 1363 if (iowrite) { 1364 mutex_enter(vp->v_interlock); 1365 vp->v_numoutput += 2; 1366 mutex_exit(vp->v_interlock); 1367 } 1368 mbp = getiobuf(vp, true); 1369 UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x", 1370 vp, mbp, vp->v_numoutput, bytes); 1371 mbp->b_bufsize = len; 1372 mbp->b_data = (void *)kva; 1373 mbp->b_resid = mbp->b_bcount = bytes; 1374 mbp->b_cflags = BC_BUSY | BC_AGE; 1375 if (async) { 1376 mbp->b_flags = brw | B_ASYNC; 1377 mbp->b_iodone = iodone; 1378 } else { 1379 mbp->b_flags = brw; 1380 mbp->b_iodone = NULL; 1381 } 1382 if (curlwp == uvm.pagedaemon_lwp) 1383 BIO_SETPRIO(mbp, BPRIO_TIMELIMITED); 1384 else if (async || lazy) 1385 BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL); 1386 else 1387 BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL); 1388 1389 bp = NULL; 1390 for (offset = startoffset; 1391 bytes > 0; 1392 offset += iobytes, bytes -= iobytes) { 1393 int run; 1394 daddr_t lbn, blkno; 1395 struct vnode *devvp; 1396 1397 /* 1398 * bmap the file to find out the blkno to read from and 1399 * how much we can read in one i/o. if bmap returns an error, 1400 * skip the rest of the top-level i/o. 1401 */ 1402 1403 lbn = offset >> fs_bshift; 1404 error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run); 1405 if (error) { 1406 UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n", 1407 lbn,error,0,0); 1408 skipbytes += bytes; 1409 bytes = 0; 1410 goto loopdone; 1411 } 1412 1413 /* 1414 * see how many pages can be read with this i/o. 1415 * reduce the i/o size if necessary to avoid 1416 * overwriting pages with valid data. 1417 */ 1418 1419 iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset, 1420 bytes); 1421 1422 /* 1423 * if this block isn't allocated, zero it instead of 1424 * reading it. unless we are going to allocate blocks, 1425 * mark the pages we zeroed PG_RDONLY. 1426 */ 1427 1428 if (blkno == (daddr_t)-1) { 1429 if (!iowrite) { 1430 memset((char *)kva + (offset - startoffset), 0, 1431 iobytes); 1432 } 1433 skipbytes += iobytes; 1434 continue; 1435 } 1436 1437 /* 1438 * allocate a sub-buf for this piece of the i/o 1439 * (or just use mbp if there's only 1 piece), 1440 * and start it going. 1441 */ 1442 1443 if (offset == startoffset && iobytes == bytes) { 1444 bp = mbp; 1445 } else { 1446 UVMHIST_LOG(ubchist, "vp %p bp %p num now %d", 1447 vp, bp, vp->v_numoutput, 0); 1448 bp = getiobuf(vp, true); 1449 nestiobuf_setup(mbp, bp, offset - startoffset, iobytes); 1450 } 1451 bp->b_lblkno = 0; 1452 1453 /* adjust physical blkno for partial blocks */ 1454 bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >> 1455 dev_bshift); 1456 1457 UVMHIST_LOG(ubchist, 1458 "bp %p offset 0x%x bcount 0x%x blkno 0x%x", 1459 bp, offset, bp->b_bcount, bp->b_blkno); 1460 1461 VOP_STRATEGY(devvp, bp); 1462 } 1463 1464 loopdone: 1465 if (skipbytes) { 1466 UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0); 1467 } 1468 nestiobuf_done(mbp, skipbytes, error); 1469 if (async) { 1470 UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0); 1471 return (0); 1472 } 1473 UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0); 1474 error = biowait(mbp); 1475 s = splbio(); 1476 (*iodone)(mbp); 1477 splx(s); 1478 UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0); 1479 return (error); 1480 } 1481 1482 int 1483 genfs_compat_getpages(void *v) 1484 { 1485 struct vop_getpages_args /* { 1486 struct vnode *a_vp; 1487 voff_t a_offset; 1488 struct vm_page **a_m; 1489 int *a_count; 1490 int a_centeridx; 1491 vm_prot_t a_access_type; 1492 int a_advice; 1493 int a_flags; 1494 } */ *ap = v; 1495 1496 off_t origoffset; 1497 struct vnode *vp = ap->a_vp; 1498 struct uvm_object *uobj = &vp->v_uobj; 1499 struct vm_page *pg, **pgs; 1500 vaddr_t kva; 1501 int i, error, orignpages, npages; 1502 struct iovec iov; 1503 struct uio uio; 1504 kauth_cred_t cred = curlwp->l_cred; 1505 const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0; 1506 1507 error = 0; 1508 origoffset = ap->a_offset; 1509 orignpages = *ap->a_count; 1510 pgs = ap->a_m; 1511 1512 if (ap->a_flags & PGO_LOCKED) { 1513 uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m, 1514 UFP_NOWAIT|UFP_NOALLOC| (memwrite ? UFP_NORDONLY : 0)); 1515 1516 error = ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0; 1517 if (error == 0 && memwrite) { 1518 genfs_markdirty(vp); 1519 } 1520 return error; 1521 } 1522 if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) { 1523 mutex_exit(uobj->vmobjlock); 1524 return EINVAL; 1525 } 1526 if ((ap->a_flags & PGO_SYNCIO) == 0) { 1527 mutex_exit(uobj->vmobjlock); 1528 return 0; 1529 } 1530 npages = orignpages; 1531 uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL); 1532 mutex_exit(uobj->vmobjlock); 1533 kva = uvm_pagermapin(pgs, npages, 1534 UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK); 1535 for (i = 0; i < npages; i++) { 1536 pg = pgs[i]; 1537 if ((pg->flags & PG_FAKE) == 0) { 1538 continue; 1539 } 1540 iov.iov_base = (char *)kva + (i << PAGE_SHIFT); 1541 iov.iov_len = PAGE_SIZE; 1542 uio.uio_iov = &iov; 1543 uio.uio_iovcnt = 1; 1544 uio.uio_offset = origoffset + (i << PAGE_SHIFT); 1545 uio.uio_rw = UIO_READ; 1546 uio.uio_resid = PAGE_SIZE; 1547 UIO_SETUP_SYSSPACE(&uio); 1548 /* XXX vn_lock */ 1549 error = VOP_READ(vp, &uio, 0, cred); 1550 if (error) { 1551 break; 1552 } 1553 if (uio.uio_resid) { 1554 memset(iov.iov_base, 0, uio.uio_resid); 1555 } 1556 } 1557 uvm_pagermapout(kva, npages); 1558 mutex_enter(uobj->vmobjlock); 1559 mutex_enter(&uvm_pageqlock); 1560 for (i = 0; i < npages; i++) { 1561 pg = pgs[i]; 1562 if (error && (pg->flags & PG_FAKE) != 0) { 1563 pg->flags |= PG_RELEASED; 1564 } else { 1565 pmap_clear_modify(pg); 1566 uvm_pageactivate(pg); 1567 } 1568 } 1569 if (error) { 1570 uvm_page_unbusy(pgs, npages); 1571 } 1572 mutex_exit(&uvm_pageqlock); 1573 if (error == 0 && memwrite) { 1574 genfs_markdirty(vp); 1575 } 1576 mutex_exit(uobj->vmobjlock); 1577 return error; 1578 } 1579 1580 int 1581 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, 1582 int flags) 1583 { 1584 off_t offset; 1585 struct iovec iov; 1586 struct uio uio; 1587 kauth_cred_t cred = curlwp->l_cred; 1588 struct buf *bp; 1589 vaddr_t kva; 1590 int error; 1591 1592 offset = pgs[0]->offset; 1593 kva = uvm_pagermapin(pgs, npages, 1594 UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK); 1595 1596 iov.iov_base = (void *)kva; 1597 iov.iov_len = npages << PAGE_SHIFT; 1598 uio.uio_iov = &iov; 1599 uio.uio_iovcnt = 1; 1600 uio.uio_offset = offset; 1601 uio.uio_rw = UIO_WRITE; 1602 uio.uio_resid = npages << PAGE_SHIFT; 1603 UIO_SETUP_SYSSPACE(&uio); 1604 /* XXX vn_lock */ 1605 error = VOP_WRITE(vp, &uio, 0, cred); 1606 1607 mutex_enter(vp->v_interlock); 1608 vp->v_numoutput++; 1609 mutex_exit(vp->v_interlock); 1610 1611 bp = getiobuf(vp, true); 1612 bp->b_cflags = BC_BUSY | BC_AGE; 1613 bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift; 1614 bp->b_data = (char *)kva; 1615 bp->b_bcount = npages << PAGE_SHIFT; 1616 bp->b_bufsize = npages << PAGE_SHIFT; 1617 bp->b_resid = 0; 1618 bp->b_error = error; 1619 uvm_aio_aiodone(bp); 1620 return (error); 1621 } 1622 1623 /* 1624 * Process a uio using direct I/O. If we reach a part of the request 1625 * which cannot be processed in this fashion for some reason, just return. 1626 * The caller must handle some additional part of the request using 1627 * buffered I/O before trying direct I/O again. 1628 */ 1629 1630 void 1631 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag) 1632 { 1633 struct vmspace *vs; 1634 struct iovec *iov; 1635 vaddr_t va; 1636 size_t len; 1637 const int mask = DEV_BSIZE - 1; 1638 int error; 1639 bool need_wapbl = (vp->v_mount && vp->v_mount->mnt_wapbl && 1640 (ioflag & IO_JOURNALLOCKED) == 0); 1641 1642 /* 1643 * We only support direct I/O to user space for now. 1644 */ 1645 1646 if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) { 1647 return; 1648 } 1649 1650 /* 1651 * If the vnode is mapped, we would need to get the getpages lock 1652 * to stabilize the bmap, but then we would get into trouble while 1653 * locking the pages if the pages belong to this same vnode (or a 1654 * multi-vnode cascade to the same effect). Just fall back to 1655 * buffered I/O if the vnode is mapped to avoid this mess. 1656 */ 1657 1658 if (vp->v_vflag & VV_MAPPED) { 1659 return; 1660 } 1661 1662 if (need_wapbl) { 1663 error = WAPBL_BEGIN(vp->v_mount); 1664 if (error) 1665 return; 1666 } 1667 1668 /* 1669 * Do as much of the uio as possible with direct I/O. 1670 */ 1671 1672 vs = uio->uio_vmspace; 1673 while (uio->uio_resid) { 1674 iov = uio->uio_iov; 1675 if (iov->iov_len == 0) { 1676 uio->uio_iov++; 1677 uio->uio_iovcnt--; 1678 continue; 1679 } 1680 va = (vaddr_t)iov->iov_base; 1681 len = MIN(iov->iov_len, genfs_maxdio); 1682 len &= ~mask; 1683 1684 /* 1685 * If the next chunk is smaller than DEV_BSIZE or extends past 1686 * the current EOF, then fall back to buffered I/O. 1687 */ 1688 1689 if (len == 0 || uio->uio_offset + len > vp->v_size) { 1690 break; 1691 } 1692 1693 /* 1694 * Check alignment. The file offset must be at least 1695 * sector-aligned. The exact constraint on memory alignment 1696 * is very hardware-dependent, but requiring sector-aligned 1697 * addresses there too is safe. 1698 */ 1699 1700 if (uio->uio_offset & mask || va & mask) { 1701 break; 1702 } 1703 error = genfs_do_directio(vs, va, len, vp, uio->uio_offset, 1704 uio->uio_rw); 1705 if (error) { 1706 break; 1707 } 1708 iov->iov_base = (char *)iov->iov_base + len; 1709 iov->iov_len -= len; 1710 uio->uio_offset += len; 1711 uio->uio_resid -= len; 1712 } 1713 1714 if (need_wapbl) 1715 WAPBL_END(vp->v_mount); 1716 } 1717 1718 /* 1719 * Iodone routine for direct I/O. We don't do much here since the request is 1720 * always synchronous, so the caller will do most of the work after biowait(). 1721 */ 1722 1723 static void 1724 genfs_dio_iodone(struct buf *bp) 1725 { 1726 1727 KASSERT((bp->b_flags & B_ASYNC) == 0); 1728 if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) { 1729 mutex_enter(bp->b_objlock); 1730 vwakeup(bp); 1731 mutex_exit(bp->b_objlock); 1732 } 1733 putiobuf(bp); 1734 } 1735 1736 /* 1737 * Process one chunk of a direct I/O request. 1738 */ 1739 1740 static int 1741 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp, 1742 off_t off, enum uio_rw rw) 1743 { 1744 struct vm_map *map; 1745 struct pmap *upm, *kpm __unused; 1746 size_t klen = round_page(uva + len) - trunc_page(uva); 1747 off_t spoff, epoff; 1748 vaddr_t kva, puva; 1749 paddr_t pa; 1750 vm_prot_t prot; 1751 int error, rv __diagused, poff, koff; 1752 const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | PGO_JOURNALLOCKED | 1753 (rw == UIO_WRITE ? PGO_FREE : 0); 1754 1755 /* 1756 * For writes, verify that this range of the file already has fully 1757 * allocated backing store. If there are any holes, just punt and 1758 * make the caller take the buffered write path. 1759 */ 1760 1761 if (rw == UIO_WRITE) { 1762 daddr_t lbn, elbn, blkno; 1763 int bsize, bshift, run; 1764 1765 bshift = vp->v_mount->mnt_fs_bshift; 1766 bsize = 1 << bshift; 1767 lbn = off >> bshift; 1768 elbn = (off + len + bsize - 1) >> bshift; 1769 while (lbn < elbn) { 1770 error = VOP_BMAP(vp, lbn, NULL, &blkno, &run); 1771 if (error) { 1772 return error; 1773 } 1774 if (blkno == (daddr_t)-1) { 1775 return ENOSPC; 1776 } 1777 lbn += 1 + run; 1778 } 1779 } 1780 1781 /* 1782 * Flush any cached pages for parts of the file that we're about to 1783 * access. If we're writing, invalidate pages as well. 1784 */ 1785 1786 spoff = trunc_page(off); 1787 epoff = round_page(off + len); 1788 mutex_enter(vp->v_interlock); 1789 error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags); 1790 if (error) { 1791 return error; 1792 } 1793 1794 /* 1795 * Wire the user pages and remap them into kernel memory. 1796 */ 1797 1798 prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ; 1799 error = uvm_vslock(vs, (void *)uva, len, prot); 1800 if (error) { 1801 return error; 1802 } 1803 1804 map = &vs->vm_map; 1805 upm = vm_map_pmap(map); 1806 kpm = vm_map_pmap(kernel_map); 1807 puva = trunc_page(uva); 1808 kva = uvm_km_alloc(kernel_map, klen, atop(puva) & uvmexp.colormask, 1809 UVM_KMF_VAONLY | UVM_KMF_WAITVA | UVM_KMF_COLORMATCH); 1810 for (poff = 0; poff < klen; poff += PAGE_SIZE) { 1811 rv = pmap_extract(upm, puva + poff, &pa); 1812 KASSERT(rv); 1813 pmap_kenter_pa(kva + poff, pa, prot, PMAP_WIRED); 1814 } 1815 pmap_update(kpm); 1816 1817 /* 1818 * Do the I/O. 1819 */ 1820 1821 koff = uva - trunc_page(uva); 1822 error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw, 1823 genfs_dio_iodone); 1824 1825 /* 1826 * Tear down the kernel mapping. 1827 */ 1828 1829 pmap_kremove(kva, klen); 1830 pmap_update(kpm); 1831 uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY); 1832 1833 /* 1834 * Unwire the user pages. 1835 */ 1836 1837 uvm_vsunlock(vs, (void *)uva, len); 1838 return error; 1839 } 1840