1 /* 2 * Copyright (c) 2010 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 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 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * Implements new VFS/VM coherency functions. For conforming VFSs 37 * we treat the backing VM object slightly differently. Instead of 38 * maintaining a number of pages to exactly fit the size of the file 39 * we instead maintain pages to fit the entire contents of the last 40 * buffer cache buffer used by the file. 41 * 42 * For VFSs like NFS and HAMMER which use (generally speaking) fixed 43 * sized buffers this greatly reduces the complexity of VFS/VM interactions. 44 * 45 * Truncations no longer invalidate pages covered by the buffer cache 46 * beyond the file EOF which still fit within the file's last buffer. 47 * We simply unmap them and do not allow userland to fault them in. 48 * 49 * The VFS is no longer responsible for zero-filling buffers during a 50 * truncation, the last buffer will be automatically zero-filled by 51 * nvtruncbuf(). 52 * 53 * This code is intended to (eventually) replace vtruncbuf() and 54 * vnode_pager_setsize(). 55 */ 56 57 #include <sys/param.h> 58 #include <sys/systm.h> 59 #include <sys/buf.h> 60 #include <sys/conf.h> 61 #include <sys/fcntl.h> 62 #include <sys/file.h> 63 #include <sys/kernel.h> 64 #include <sys/malloc.h> 65 #include <sys/mount.h> 66 #include <sys/proc.h> 67 #include <sys/socket.h> 68 #include <sys/stat.h> 69 #include <sys/sysctl.h> 70 #include <sys/unistd.h> 71 #include <sys/vmmeter.h> 72 #include <sys/vnode.h> 73 74 #include <machine/limits.h> 75 76 #include <vm/vm.h> 77 #include <vm/vm_object.h> 78 #include <vm/vm_extern.h> 79 #include <vm/vm_kern.h> 80 #include <vm/pmap.h> 81 #include <vm/vm_map.h> 82 #include <vm/vm_page.h> 83 #include <vm/vm_pager.h> 84 #include <vm/vnode_pager.h> 85 #include <vm/vm_zone.h> 86 87 #include <sys/buf2.h> 88 #include <vm/vm_page2.h> 89 90 static int nvtruncbuf_bp_trunc_cmp(struct buf *bp, void *data); 91 static int nvtruncbuf_bp_trunc(struct buf *bp, void *data); 92 static int nvtruncbuf_bp_metasync_cmp(struct buf *bp, void *data); 93 static int nvtruncbuf_bp_metasync(struct buf *bp, void *data); 94 95 /* 96 * Truncate a file's buffer and pages to a specified length. The 97 * byte-granular length of the file is specified along with the block 98 * size of the buffer containing that offset. 99 * 100 * If the last buffer straddles the length its contents will be zero-filled 101 * as appropriate. All buffers and pages after the last buffer will be 102 * destroyed. The last buffer itself will be destroyed only if the length 103 * is exactly aligned with it. 104 * 105 * UFS typically passes the old block size prior to the actual truncation, 106 * then later resizes the block based on the new file size. NFS uses a 107 * fixed block size and doesn't care. HAMMER uses a block size based on 108 * the offset which is fixed for any particular offset. 109 * 110 * When zero-filling we must bdwrite() to avoid a window of opportunity 111 * where the kernel might throw away a clean buffer and the filesystem 112 * then attempts to bread() it again before completing (or as part of) 113 * the extension. The filesystem is still responsible for zero-filling 114 * any remainder when writing to the media in the strategy function when 115 * it is able to do so without the page being mapped. The page may still 116 * be mapped by userland here. 117 * 118 * When modifying a buffer we must clear any cached raw disk offset. 119 * bdwrite() will call BMAP on it again. Some filesystems, like HAMMER, 120 * never overwrite existing data blocks. 121 */ 122 123 struct truncbuf_info { 124 struct vnode *vp; 125 off_t truncloffset; /* truncation point */ 126 int clean; /* clean tree, else dirty tree */ 127 }; 128 129 int 130 nvtruncbuf(struct vnode *vp, off_t length, int blksize, int boff, int trivial) 131 { 132 struct truncbuf_info info; 133 off_t truncboffset; 134 const char *filename; 135 struct buf *bp; 136 int count; 137 int error; 138 139 /* 140 * Round up to the *next* block, then destroy the buffers in question. 141 * Since we are only removing some of the buffers we must rely on the 142 * scan count to determine whether a loop is necessary. 143 * 144 * Destroy any pages beyond the last buffer. 145 */ 146 if (boff < 0) 147 boff = (int)(length % blksize); 148 if (boff) 149 info.truncloffset = length + (blksize - boff); 150 else 151 info.truncloffset = length; 152 info.vp = vp; 153 lwkt_gettoken(&vp->v_token); 154 do { 155 info.clean = 1; 156 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 157 nvtruncbuf_bp_trunc_cmp, 158 nvtruncbuf_bp_trunc, &info); 159 info.clean = 0; 160 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 161 nvtruncbuf_bp_trunc_cmp, 162 nvtruncbuf_bp_trunc, &info); 163 } while(count); 164 165 nvnode_pager_setsize(vp, length, blksize, boff); 166 167 /* 168 * Zero-fill the area beyond the file EOF that still fits within 169 * the last buffer. We must mark the buffer as dirty even though 170 * the modified area is beyond EOF to avoid races where the kernel 171 * might flush the buffer before the filesystem is able to reallocate 172 * the block. 173 * 174 * The VFS is responsible for dealing with the actual truncation. 175 * 176 * Only do this if trivial is zero, otherwise it is up to the 177 * VFS to handle the block straddling the EOF. 178 */ 179 if (boff && trivial == 0) { 180 truncboffset = length - boff; 181 error = bread_kvabio(vp, truncboffset, blksize, &bp); 182 if (error == 0) { 183 bkvasync(bp); 184 bzero(bp->b_data + boff, blksize - boff); 185 if (bp->b_flags & B_DELWRI) { 186 if (bp->b_dirtyoff > boff) 187 bp->b_dirtyoff = boff; 188 if (bp->b_dirtyend > boff) 189 bp->b_dirtyend = boff; 190 } 191 bp->b_bio2.bio_offset = NOOFFSET; 192 bdwrite(bp); 193 } else { 194 kprintf("nvtruncbuf: bread error %d @0x%016jx\n", 195 error, truncboffset); 196 bp->b_flags |= B_INVAL | B_RELBUF; 197 brelse(bp); 198 } 199 } else { 200 error = 0; 201 } 202 203 /* 204 * For safety, fsync any remaining metadata if the file is not being 205 * truncated to 0. Since the metadata does not represent the entire 206 * dirty list we have to rely on the hit count to ensure that we get 207 * all of it. 208 * 209 * This is typically applicable only to UFS. NFS and HAMMER do 210 * not store indirect blocks in the per-vnode buffer cache. 211 */ 212 if (length > 0) { 213 do { 214 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 215 nvtruncbuf_bp_metasync_cmp, 216 nvtruncbuf_bp_metasync, &info); 217 } while (count); 218 } 219 220 /* 221 * It is possible to have in-progress I/O from buffers that were 222 * not part of the truncation. This should not happen if we 223 * are truncating to 0-length. 224 */ 225 bio_track_wait(&vp->v_track_write, 0, 0); 226 227 /* 228 * Debugging only 229 */ 230 spin_lock(&vp->v_spin); 231 filename = TAILQ_FIRST(&vp->v_namecache) ? 232 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"; 233 spin_unlock(&vp->v_spin); 234 235 /* 236 * Make sure no buffers were instantiated while we were trying 237 * to clean out the remaining VM pages. This could occur due 238 * to busy dirty VM pages being flushed out to disk. 239 */ 240 do { 241 info.clean = 1; 242 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 243 nvtruncbuf_bp_trunc_cmp, 244 nvtruncbuf_bp_trunc, &info); 245 info.clean = 0; 246 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 247 nvtruncbuf_bp_trunc_cmp, 248 nvtruncbuf_bp_trunc, &info); 249 if (count) { 250 kprintf("Warning: vtruncbuf(): Had to re-clean %d " 251 "left over buffers in %s\n", count, filename); 252 } 253 } while(count); 254 255 lwkt_reltoken(&vp->v_token); 256 257 return (error); 258 } 259 260 /* 261 * The callback buffer is beyond the new file EOF and must be destroyed. 262 * Note that the compare function must conform to the RB_SCAN's requirements. 263 */ 264 static 265 int 266 nvtruncbuf_bp_trunc_cmp(struct buf *bp, void *data) 267 { 268 struct truncbuf_info *info = data; 269 270 if (bp->b_loffset >= info->truncloffset) 271 return(0); 272 return(-1); 273 } 274 275 static 276 int 277 nvtruncbuf_bp_trunc(struct buf *bp, void *data) 278 { 279 struct truncbuf_info *info = data; 280 281 /* 282 * Do not try to use a buffer we cannot immediately lock, 283 * but sleep anyway to prevent a livelock. The code will 284 * loop until all buffers can be acted upon. 285 */ 286 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 287 atomic_add_int(&bp->b_refs, 1); 288 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 289 BUF_UNLOCK(bp); 290 atomic_subtract_int(&bp->b_refs, 1); 291 } else if ((info->clean && (bp->b_flags & B_DELWRI)) || 292 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0) || 293 bp->b_vp != info->vp || 294 nvtruncbuf_bp_trunc_cmp(bp, data)) { 295 BUF_UNLOCK(bp); 296 } else { 297 bremfree(bp); 298 bp->b_flags |= (B_INVAL | B_RELBUF | B_NOCACHE); 299 brelse(bp); 300 } 301 lwkt_yield(); 302 return(1); 303 } 304 305 /* 306 * Fsync all meta-data after truncating a file to be non-zero. Only metadata 307 * blocks (with a negative loffset) are scanned. 308 * Note that the compare function must conform to the RB_SCAN's requirements. 309 */ 310 static int 311 nvtruncbuf_bp_metasync_cmp(struct buf *bp, void *data __unused) 312 { 313 if (bp->b_loffset < 0) 314 return(0); 315 lwkt_yield(); 316 return(1); 317 } 318 319 static int 320 nvtruncbuf_bp_metasync(struct buf *bp, void *data) 321 { 322 struct truncbuf_info *info = data; 323 324 /* 325 * Do not try to use a buffer we cannot immediately lock, 326 * but sleep anyway to prevent a livelock. The code will 327 * loop until all buffers can be acted upon. 328 */ 329 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 330 atomic_add_int(&bp->b_refs, 1); 331 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 332 BUF_UNLOCK(bp); 333 atomic_subtract_int(&bp->b_refs, 1); 334 } else if ((bp->b_flags & B_DELWRI) == 0 || 335 bp->b_vp != info->vp || 336 nvtruncbuf_bp_metasync_cmp(bp, data)) { 337 BUF_UNLOCK(bp); 338 } else { 339 bremfree(bp); 340 bawrite(bp); 341 } 342 lwkt_yield(); 343 return(1); 344 } 345 346 /* 347 * Extend a file's buffer and pages to a new, larger size. The block size 348 * at both the old and new length must be passed, but buffer cache operations 349 * will only be performed on the old block. The new nlength/nblksize will 350 * be used to properly set the VM object size. 351 * 352 * To make this explicit we require the old length to passed even though 353 * we can acquire it from vp->v_filesize, which also avoids potential 354 * corruption if the filesystem and vp get desynchronized somehow. 355 * 356 * If the caller intends to immediately write into the newly extended 357 * space pass trivial == 1. If trivial is 0 the original buffer will be 358 * zero-filled as necessary to clean out any junk in the extended space. 359 * If non-zero the original buffer (straddling EOF) is not touched. 360 * 361 * When zero-filling we must bdwrite() to avoid a window of opportunity 362 * where the kernel might throw away a clean buffer and the filesystem 363 * then attempts to bread() it again before completing (or as part of) 364 * the extension. The filesystem is still responsible for zero-filling 365 * any remainder when writing to the media in the strategy function when 366 * it is able to do so without the page being mapped. The page may still 367 * be mapped by userland here. 368 * 369 * When modifying a buffer we must clear any cached raw disk offset. 370 * bdwrite() will call BMAP on it again. Some filesystems, like HAMMER, 371 * never overwrite existing data blocks. 372 */ 373 int 374 nvextendbuf(struct vnode *vp, off_t olength, off_t nlength, 375 int oblksize, int nblksize, int oboff, int nboff, int trivial) 376 { 377 off_t truncboffset; 378 struct buf *bp; 379 int error; 380 381 error = 0; 382 nvnode_pager_setsize(vp, nlength, nblksize, nboff); 383 if (trivial == 0) { 384 if (oboff < 0) 385 oboff = (int)(olength % oblksize); 386 truncboffset = olength - oboff; 387 388 if (oboff) { 389 error = bread_kvabio(vp, truncboffset, oblksize, &bp); 390 if (error == 0) { 391 bkvasync(bp); 392 bzero(bp->b_data + oboff, oblksize - oboff); 393 bp->b_bio2.bio_offset = NOOFFSET; 394 bdwrite(bp); 395 } else { 396 kprintf("nvextendbuf: bread EOF @ %016jx " 397 "error %d\n", 398 truncboffset, error); 399 bp->b_flags |= B_INVAL | B_RELBUF; 400 brelse(bp); 401 } 402 } 403 } 404 return (error); 405 } 406 407 /* 408 * Set vp->v_filesize and vp->v_object->size, destroy pages beyond 409 * the last buffer when truncating. 410 * 411 * This function does not do any zeroing or invalidating of partially 412 * overlapping pages. Zeroing is the responsibility of nvtruncbuf(). 413 * However, it does unmap VM pages from the user address space on a 414 * page-granular (verses buffer cache granular) basis. 415 * 416 * If boff is passed as -1 the base offset of the buffer cache buffer is 417 * calculated from length and blksize. Filesystems such as UFS which deal 418 * with fragments have to specify a boff >= 0 since the base offset cannot 419 * be calculated from length and blksize. 420 * 421 * For UFS blksize is the 'new' blocksize, used only to determine how large 422 * the VM object must become. 423 */ 424 void 425 nvnode_pager_setsize(struct vnode *vp, off_t length, int blksize, int boff) 426 { 427 vm_pindex_t nobjsize; 428 vm_pindex_t oobjsize; 429 vm_pindex_t pi; 430 vm_object_t object; 431 vm_page_t m; 432 off_t truncboffset; 433 434 /* 435 * Degenerate conditions 436 */ 437 if ((object = vp->v_object) == NULL) 438 return; 439 vm_object_hold(object); 440 if (length == vp->v_filesize) { 441 vm_object_drop(object); 442 return; 443 } 444 445 /* 446 * Calculate the size of the VM object, coverage includes 447 * the buffer straddling EOF. If EOF is buffer-aligned 448 * we don't bother. 449 * 450 * Buffers do not have to be page-aligned. Make sure 451 * nobjsize is beyond the last page of the buffer. 452 */ 453 if (boff < 0) 454 boff = (int)(length % blksize); 455 truncboffset = length - boff; 456 oobjsize = object->size; 457 if (boff) 458 nobjsize = OFF_TO_IDX(truncboffset + blksize + PAGE_MASK); 459 else 460 nobjsize = OFF_TO_IDX(truncboffset + PAGE_MASK); 461 object->size = nobjsize; 462 463 if (length < vp->v_filesize) { 464 /* 465 * File has shrunk, toss any cached pages beyond 466 * the end of the buffer (blksize aligned) for the 467 * new EOF. 468 */ 469 vp->v_filesize = length; 470 if (nobjsize < oobjsize) { 471 vm_object_page_remove(object, nobjsize, oobjsize, 472 FALSE); 473 } 474 475 /* 476 * Unmap any pages (page aligned) beyond the new EOF. 477 * The pages remain part of the (last) buffer and are not 478 * invalidated. 479 */ 480 pi = OFF_TO_IDX(length + PAGE_MASK); 481 while (pi < nobjsize) { 482 m = vm_page_lookup_busy_wait(object, pi, FALSE, "vmpg"); 483 if (m) { 484 vm_page_protect(m, VM_PROT_NONE); 485 vm_page_wakeup(m); 486 } 487 ++pi; 488 lwkt_yield(); 489 } 490 } else { 491 /* 492 * File has expanded. 493 */ 494 vp->v_filesize = length; 495 } 496 vm_object_drop(object); 497 } 498