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 <sys/thread2.h> 89 #include <sys/sysref2.h> 90 91 static int nvtruncbuf_bp_trunc_cmp(struct buf *bp, void *data); 92 static int nvtruncbuf_bp_trunc(struct buf *bp, void *data); 93 static int nvtruncbuf_bp_metasync_cmp(struct buf *bp, void *data); 94 static int nvtruncbuf_bp_metasync(struct buf *bp, void *data); 95 96 /* 97 * Truncate a file's buffer and pages to a specified length. The 98 * byte-granular length of the file is specified along with the block 99 * size of the buffer containing that offset. 100 * 101 * If the last buffer straddles the length its contents will be zero-filled 102 * as appropriate. All buffers and pages after the last buffer will be 103 * destroyed. The last buffer itself will be destroyed only if the length 104 * is exactly aligned with it. 105 * 106 * UFS typically passes the old block size prior to the actual truncation, 107 * then later resizes the block based on the new file size. NFS uses a 108 * fixed block size and doesn't care. HAMMER uses a block size based on 109 * the offset which is fixed for any particular offset. 110 * 111 * When zero-filling we must bdwrite() to avoid a window of opportunity 112 * where the kernel might throw away a clean buffer and the filesystem 113 * then attempts to bread() it again before completing (or as part of) 114 * the extension. The filesystem is still responsible for zero-filling 115 * any remainder when writing to the media in the strategy function when 116 * it is able to do so without the page being mapped. The page may still 117 * be mapped by userland here. 118 * 119 * When modifying a buffer we must clear any cached raw disk offset. 120 * bdwrite() will call BMAP on it again. Some filesystems, like HAMMER, 121 * never overwrite existing data blocks. 122 */ 123 124 struct truncbuf_info { 125 struct vnode *vp; 126 off_t truncloffset; /* truncation point */ 127 int clean; /* clean tree, else dirty tree */ 128 }; 129 130 int 131 nvtruncbuf(struct vnode *vp, off_t length, int blksize, int boff) 132 { 133 struct truncbuf_info info; 134 off_t truncboffset; 135 const char *filename; 136 struct buf *bp; 137 int count; 138 int error; 139 140 /* 141 * Round up to the *next* block, then destroy the buffers in question. 142 * Since we are only removing some of the buffers we must rely on the 143 * scan count to determine whether a loop is necessary. 144 * 145 * Destroy any pages beyond the last buffer. 146 */ 147 if (boff < 0) 148 boff = (int)(length % blksize); 149 if (boff) 150 info.truncloffset = length + (blksize - boff); 151 else 152 info.truncloffset = length; 153 info.vp = vp; 154 lwkt_gettoken(&vp->v_token); 155 do { 156 info.clean = 1; 157 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 158 nvtruncbuf_bp_trunc_cmp, 159 nvtruncbuf_bp_trunc, &info); 160 info.clean = 0; 161 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 162 nvtruncbuf_bp_trunc_cmp, 163 nvtruncbuf_bp_trunc, &info); 164 } while(count); 165 166 nvnode_pager_setsize(vp, length, blksize, boff); 167 168 /* 169 * Zero-fill the area beyond the file EOF that still fits within 170 * the last buffer. We must mark the buffer as dirty even though 171 * the modified area is beyond EOF to avoid races where the kernel 172 * might flush the buffer before the filesystem is able to reallocate 173 * the block. 174 * 175 * The VFS is responsible for dealing with the actual truncation. 176 */ 177 if (boff) { 178 truncboffset = length - boff; 179 error = bread(vp, truncboffset, blksize, &bp); 180 if (error == 0) { 181 bzero(bp->b_data + boff, blksize - boff); 182 if (bp->b_flags & B_DELWRI) { 183 if (bp->b_dirtyoff > boff) 184 bp->b_dirtyoff = boff; 185 if (bp->b_dirtyend > boff) 186 bp->b_dirtyend = boff; 187 } 188 bp->b_bio2.bio_offset = NOOFFSET; 189 bdwrite(bp); 190 } 191 } else { 192 error = 0; 193 } 194 195 /* 196 * For safety, fsync any remaining metadata if the file is not being 197 * truncated to 0. Since the metadata does not represent the entire 198 * dirty list we have to rely on the hit count to ensure that we get 199 * all of it. 200 * 201 * This is typically applicable only to UFS. NFS and HAMMER do 202 * not store indirect blocks in the per-vnode buffer cache. 203 */ 204 if (length > 0) { 205 do { 206 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 207 nvtruncbuf_bp_metasync_cmp, 208 nvtruncbuf_bp_metasync, &info); 209 } while (count); 210 } 211 212 /* 213 * It is possible to have in-progress I/O from buffers that were 214 * not part of the truncation. This should not happen if we 215 * are truncating to 0-length. 216 */ 217 bio_track_wait(&vp->v_track_write, 0, 0); 218 219 /* 220 * Debugging only 221 */ 222 spin_lock(&vp->v_spin); 223 filename = TAILQ_FIRST(&vp->v_namecache) ? 224 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"; 225 spin_unlock(&vp->v_spin); 226 227 /* 228 * Make sure no buffers were instantiated while we were trying 229 * to clean out the remaining VM pages. This could occur due 230 * to busy dirty VM pages being flushed out to disk. 231 */ 232 do { 233 info.clean = 1; 234 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 235 nvtruncbuf_bp_trunc_cmp, 236 nvtruncbuf_bp_trunc, &info); 237 info.clean = 0; 238 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 239 nvtruncbuf_bp_trunc_cmp, 240 nvtruncbuf_bp_trunc, &info); 241 if (count) { 242 kprintf("Warning: vtruncbuf(): Had to re-clean %d " 243 "left over buffers in %s\n", count, filename); 244 } 245 } while(count); 246 247 lwkt_reltoken(&vp->v_token); 248 249 return (error); 250 } 251 252 /* 253 * The callback buffer is beyond the new file EOF and must be destroyed. 254 * Note that the compare function must conform to the RB_SCAN's requirements. 255 */ 256 static 257 int 258 nvtruncbuf_bp_trunc_cmp(struct buf *bp, void *data) 259 { 260 struct truncbuf_info *info = data; 261 262 if (bp->b_loffset >= info->truncloffset) 263 return(0); 264 return(-1); 265 } 266 267 static 268 int 269 nvtruncbuf_bp_trunc(struct buf *bp, void *data) 270 { 271 struct truncbuf_info *info = data; 272 273 /* 274 * Do not try to use a buffer we cannot immediately lock, 275 * but sleep anyway to prevent a livelock. The code will 276 * loop until all buffers can be acted upon. 277 */ 278 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 279 atomic_add_int(&bp->b_refs, 1); 280 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 281 BUF_UNLOCK(bp); 282 atomic_subtract_int(&bp->b_refs, 1); 283 } else if ((info->clean && (bp->b_flags & B_DELWRI)) || 284 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0) || 285 bp->b_vp != info->vp || 286 nvtruncbuf_bp_trunc_cmp(bp, data)) { 287 BUF_UNLOCK(bp); 288 } else { 289 bremfree(bp); 290 bp->b_flags |= (B_INVAL | B_RELBUF | B_NOCACHE); 291 brelse(bp); 292 } 293 return(1); 294 } 295 296 /* 297 * Fsync all meta-data after truncating a file to be non-zero. Only metadata 298 * blocks (with a negative loffset) are scanned. 299 * Note that the compare function must conform to the RB_SCAN's requirements. 300 */ 301 static int 302 nvtruncbuf_bp_metasync_cmp(struct buf *bp, void *data __unused) 303 { 304 if (bp->b_loffset < 0) 305 return(0); 306 return(1); 307 } 308 309 static int 310 nvtruncbuf_bp_metasync(struct buf *bp, void *data) 311 { 312 struct truncbuf_info *info = data; 313 314 /* 315 * Do not try to use a buffer we cannot immediately lock, 316 * but sleep anyway to prevent a livelock. The code will 317 * loop until all buffers can be acted upon. 318 */ 319 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 320 atomic_add_int(&bp->b_refs, 1); 321 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 322 BUF_UNLOCK(bp); 323 atomic_subtract_int(&bp->b_refs, 1); 324 } else if ((bp->b_flags & B_DELWRI) == 0 || 325 bp->b_vp != info->vp || 326 nvtruncbuf_bp_metasync_cmp(bp, data)) { 327 BUF_UNLOCK(bp); 328 } else { 329 bremfree(bp); 330 bawrite(bp); 331 } 332 return(1); 333 } 334 335 /* 336 * Extend a file's buffer and pages to a new, larger size. The block size 337 * at both the old and new length must be passed, but buffer cache operations 338 * will only be performed on the old block. The new nlength/nblksize will 339 * be used to properly set the VM object size. 340 * 341 * To make this explicit we require the old length to passed even though 342 * we can acquire it from vp->v_filesize, which also avoids potential 343 * corruption if the filesystem and vp get desynchronized somehow. 344 * 345 * If the caller intends to immediately write into the newly extended 346 * space pass trivial == 1. If trivial is 0 the original buffer will be 347 * zero-filled as necessary to clean out any junk in the extended space. 348 * 349 * When zero-filling we must bdwrite() to avoid a window of opportunity 350 * where the kernel might throw away a clean buffer and the filesystem 351 * then attempts to bread() it again before completing (or as part of) 352 * the extension. The filesystem is still responsible for zero-filling 353 * any remainder when writing to the media in the strategy function when 354 * it is able to do so without the page being mapped. The page may still 355 * be mapped by userland here. 356 * 357 * When modifying a buffer we must clear any cached raw disk offset. 358 * bdwrite() will call BMAP on it again. Some filesystems, like HAMMER, 359 * never overwrite existing data blocks. 360 */ 361 int 362 nvextendbuf(struct vnode *vp, off_t olength, off_t nlength, 363 int oblksize, int nblksize, int oboff, int nboff, int trivial) 364 { 365 off_t truncboffset; 366 struct buf *bp; 367 int error; 368 369 error = 0; 370 nvnode_pager_setsize(vp, nlength, nblksize, nboff); 371 if (trivial == 0) { 372 if (oboff < 0) 373 oboff = (int)(olength % oblksize); 374 truncboffset = olength - oboff; 375 376 if (oboff) { 377 error = bread(vp, truncboffset, oblksize, &bp); 378 if (error == 0) { 379 bzero(bp->b_data + oboff, oblksize - oboff); 380 bp->b_bio2.bio_offset = NOOFFSET; 381 bdwrite(bp); 382 } 383 } 384 } 385 return (error); 386 } 387 388 /* 389 * Set vp->v_filesize and vp->v_object->size, destroy pages beyond 390 * the last buffer when truncating. 391 * 392 * This function does not do any zeroing or invalidating of partially 393 * overlapping pages. Zeroing is the responsibility of nvtruncbuf(). 394 * However, it does unmap VM pages from the user address space on a 395 * page-granular (verses buffer cache granular) basis. 396 * 397 * If boff is passed as -1 the base offset of the buffer cache buffer is 398 * calculated from length and blksize. Filesystems such as UFS which deal 399 * with fragments have to specify a boff >= 0 since the base offset cannot 400 * be calculated from length and blksize. 401 * 402 * For UFS blksize is the 'new' blocksize, used only to determine how large 403 * the VM object must become. 404 */ 405 void 406 nvnode_pager_setsize(struct vnode *vp, off_t length, int blksize, int boff) 407 { 408 vm_pindex_t nobjsize; 409 vm_pindex_t oobjsize; 410 vm_pindex_t pi; 411 vm_object_t object; 412 vm_page_t m; 413 off_t truncboffset; 414 415 /* 416 * Degenerate conditions 417 */ 418 if ((object = vp->v_object) == NULL) 419 return; 420 vm_object_hold(object); 421 if (length == vp->v_filesize) { 422 vm_object_drop(object); 423 return; 424 } 425 426 /* 427 * Calculate the size of the VM object, coverage includes 428 * the buffer straddling EOF. If EOF is buffer-aligned 429 * we don't bother. 430 * 431 * Buffers do not have to be page-aligned. Make sure 432 * nobjsize is beyond the last page of the buffer. 433 */ 434 if (boff < 0) 435 boff = (int)(length % blksize); 436 truncboffset = length - boff; 437 oobjsize = object->size; 438 if (boff) 439 nobjsize = OFF_TO_IDX(truncboffset + blksize + PAGE_MASK); 440 else 441 nobjsize = OFF_TO_IDX(truncboffset + PAGE_MASK); 442 object->size = nobjsize; 443 444 if (length < vp->v_filesize) { 445 /* 446 * File has shrunk, toss any cached pages beyond 447 * the end of the buffer (blksize aligned) for the 448 * new EOF. 449 */ 450 vp->v_filesize = length; 451 if (nobjsize < oobjsize) { 452 vm_object_page_remove(object, nobjsize, oobjsize, 453 FALSE); 454 } 455 456 /* 457 * Unmap any pages (page aligned) beyond the new EOF. 458 * The pages remain part of the (last) buffer and are not 459 * invalidated. 460 */ 461 pi = OFF_TO_IDX(length + PAGE_MASK); 462 while (pi < nobjsize) { 463 m = vm_page_lookup_busy_wait(object, pi, FALSE, "vmpg"); 464 if (m) { 465 vm_page_protect(m, VM_PROT_NONE); 466 vm_page_wakeup(m); 467 } 468 ++pi; 469 } 470 } else { 471 /* 472 * File has expanded. 473 */ 474 vp->v_filesize = length; 475 } 476 vm_object_drop(object); 477 } 478