1 #include "first.h" 2 3 #include "stat_cache.h" 4 #include "log.h" 5 #include "fdevent.h" 6 #include "etag.h" 7 #include "algo_splaytree.h" 8 9 #include <sys/types.h> 10 #include <sys/stat.h> 11 12 #include <stdlib.h> 13 #include <string.h> 14 #include <errno.h> 15 #include <unistd.h> 16 #include <fcntl.h> 17 18 #if defined(HAVE_SYS_XATTR_H) 19 # include <sys/xattr.h> 20 #elif defined(HAVE_ATTR_ATTRIBUTES_H) 21 # include <attr/attributes.h> 22 #endif 23 24 #ifdef HAVE_SYS_EXTATTR_H 25 # include <sys/extattr.h> 26 #endif 27 28 #ifndef HAVE_LSTAT 29 #define lstat stat 30 #ifndef S_ISLNK 31 #define S_ISLNK(mode) (0) 32 #endif 33 #endif 34 35 /* 36 * stat-cache 37 * 38 * - a splay-tree is used as we can use the caching effect of it 39 */ 40 41 enum { 42 STAT_CACHE_ENGINE_SIMPLE = 0 /*(default)*/ 43 ,STAT_CACHE_ENGINE_NONE = 1 44 ,STAT_CACHE_ENGINE_FAM = 2 /* same as STAT_CACHE_ENGINE_INOTIFY */ 45 ,STAT_CACHE_ENGINE_INOTIFY = 2 /* same as STAT_CACHE_ENGINE_FAM */ 46 ,STAT_CACHE_ENGINE_KQUEUE = 2 /* same as STAT_CACHE_ENGINE_FAM */ 47 }; 48 49 struct stat_cache_fam; /* declaration */ 50 51 typedef struct stat_cache { 52 int stat_cache_engine; 53 splay_tree *files; /* nodes of tree are (stat_cache_entry *) */ 54 struct stat_cache_fam *scf; 55 } stat_cache; 56 57 static stat_cache sc; 58 59 60 static void * stat_cache_sptree_find(splay_tree ** const sptree, 61 const char * const name, 62 uint32_t len) 63 { 64 const int ndx = splaytree_djbhash(name, len); 65 *sptree = splaytree_splay(*sptree, ndx); 66 return (*sptree && (*sptree)->key == ndx) ? (*sptree)->data : NULL; 67 } 68 69 70 #if defined(HAVE_SYS_INOTIFY_H) \ 71 || (defined(HAVE_SYS_EVENT_H) && defined(HAVE_KQUEUE)) 72 #ifndef HAVE_FAM_H 73 #define HAVE_FAM_H 74 #endif 75 #endif 76 77 #ifdef HAVE_FAM_H 78 79 /* monitor changes in directories using FAM 80 * 81 * This implementation employing FAM monitors directories as they are used, 82 * and maintains a reference count for cache use within stat_cache.c. 83 * A periodic job runs in lighttpd every 32 seconds, expiring entries unused 84 * in last 64 seconds out of the cache and cancelling FAM monitoring. Items 85 * within the cache are checked against the filesystem upon use if last stat() 86 * was greater than or equal to 16 seconds ago. 87 * 88 * This implementation does not monitor every directory in a tree, and therefore 89 * the cache may get out-of-sync with the filesystem. Delays in receiving and 90 * processing events from FAM might also lead to stale cache entries. 91 * 92 * For many websites, a large number of files are seldom, if ever, modified, 93 * and a common practice with images is to create a new file with a new name 94 * when a new version is needed, in order for client browsers and CDNs to better 95 * cache the content. Given this, most use will see little difference in 96 * performance between server.stat-cache-engine = "fam" and "simple" (default). 97 * The default server.stat-cache-engine = "simple" calls stat() on a target once 98 * per second, and reuses that information until the next second. For use where 99 * changes must be immediately visible, server.stat-cache-engine = "disable" 100 * should be used. 101 * 102 * When considering use of server.stat-cache-engine = "fam", there are a few 103 * additional limitations for this cache implementation using FAM. 104 * - symlinks to files located outside of the current directory do not result 105 * in changes to that file being monitored (unless that file is in a directory 106 * which is monitored as a result of a different request). symlinks can be 107 * chained and can be circular. This implementation *does not* readlink() or 108 * realpath() to resolve the chains to find and monitor the ultimate target 109 * directory. While symlinks to files located outside the current directory 110 * are not monitored, symlinks to directories *are* monitored, though chains 111 * of symlinks to directories do not result in monitoring of the directories 112 * containing intermediate symlinks to the target directory. 113 * - directory rename of a directory which is not currently being monitored will 114 * result in stale information in the cache if there is a subdirectory that is 115 * being monitored. 116 * Even though lighttpd will not receive FAM events in the above cases, lighttpd 117 * does re-validate the information in the cache upon use if the cache entry has 118 * not been checked in 16 seconds, so that is the upper limit for use of stale 119 * data. 120 * 121 * Use of server.stat-cache-engine = "fam" is discouraged for extremely volatile 122 * directories such as temporary directories (e.g. /tmp and maybe /var/tmp) due 123 * to the overhead of processing the additional noise generated from changes. 124 * Related, server.stat-cache-engine = "fam" is not recommended on trees of 125 * untrusted files where a malicious user could generate an excess of change 126 * events. 127 * 128 * Internal note: lighttpd walks the caches to prune trees in stat_cache when an 129 * event is received for a directory (or symlink to a directory) which has been 130 * deleted or renamed. The splaytree data structure is suboptimal for frequent 131 * changes of large directories trees where there have been a large number of 132 * different files recently accessed and part of the stat_cache. 133 */ 134 135 #if defined(HAVE_SYS_INOTIFY_H) \ 136 && !(defined(HAVE_SYS_EVENT_H) && defined(HAVE_KQUEUE)) 137 138 #include <sys/inotify.h> 139 140 /*(translate FAM API to inotify; this is specific to stat_cache.c use of FAM)*/ 141 #define fam fd /*(translate struct stat_cache_fam scf->fam -> scf->fd)*/ 142 typedef int FAMRequest; /*(fr)*/ 143 #define FAMClose(fd) \ 144 close(*(fd)) 145 #define FAMCancelMonitor(fd, wd) \ 146 inotify_rm_watch(*(fd), *(wd)) 147 #define fam_watch_mask ( IN_ATTRIB | IN_CREATE | IN_DELETE | IN_DELETE_SELF \ 148 | IN_MODIFY | IN_MOVE_SELF | IN_MOVED_FROM \ 149 | IN_EXCL_UNLINK | IN_ONLYDIR ) 150 /*(note: follows symlinks; not providing IN_DONT_FOLLOW)*/ 151 #define FAMMonitorDirectory(fd, fn, wd, userData) \ 152 ((*(wd) = inotify_add_watch(*(fd), (fn), (fam_watch_mask))) < 0) 153 typedef enum FAMCodes { /*(copied from fam.h to define arbitrary enum values)*/ 154 FAMChanged=1, 155 FAMDeleted=2, 156 FAMCreated=5, 157 FAMMoved=6, 158 } FAMCodes; 159 160 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 161 #undef HAVE_SYS_INOTIFY_H 162 163 #include <sys/event.h> 164 #include <sys/time.h> 165 166 /*(translate FAM API to inotify; this is specific to stat_cache.c use of FAM)*/ 167 #define fam fd /*(translate struct stat_cache_fam scf->fam -> scf->fd)*/ 168 typedef int FAMRequest; /*(fr)*/ 169 #define FAMClose(fd) \ 170 (-1 != (*(fd)) ? close(*(fd)) : 0) 171 static int FAMCancelMonitor (const int * const fd, int * const wd) 172 { 173 if (-1 == *fd) return 0; 174 if (-1 == *wd) return 0; 175 struct timespec t0 = { 0, 0 }; 176 struct kevent kev; 177 EV_SET(&kev, *wd, EVFILT_VNODE, EV_DELETE, 0, 0, 0); 178 int rc = kevent(*fd, &kev, 1, NULL, 0, &t0); 179 close(*wd); 180 *wd = -1; 181 return rc; 182 } 183 static int FAMMonitorDirectory (int * const fd, char * const fn, int * const wd, void * const userData) 184 { 185 *wd = fdevent_open_dirname(fn, 1); /*(note: follows symlinks)*/ 186 if (-1 == *wd) return -1; 187 struct timespec t0 = { 0, 0 }; 188 struct kevent kev; 189 unsigned short kev_flags = EV_ADD | EV_ENABLE | EV_CLEAR; 190 unsigned int kev_fflags = NOTE_ATTRIB | NOTE_EXTEND | NOTE_LINK | NOTE_WRITE 191 | NOTE_DELETE | NOTE_REVOKE | NOTE_RENAME; 192 EV_SET(&kev, *wd, EVFILT_VNODE, kev_flags, kev_fflags, 0, userData); 193 return kevent(*fd, &kev, 1, NULL, 0, &t0); 194 } 195 typedef enum FAMCodes { /*(copied from fam.h to define arbitrary enum values)*/ 196 FAMChanged=1, 197 FAMDeleted=2, 198 FAMCreated=5, 199 FAMMoved=6, 200 } FAMCodes; 201 202 #else 203 204 #include <fam.h> 205 206 #ifdef HAVE_FAMNOEXISTS 207 #ifndef LIGHTTPD_STATIC 208 #include <dlfcn.h> 209 #endif 210 #endif 211 212 #endif 213 214 typedef struct fam_dir_entry { 215 buffer *name; 216 int refcnt; 217 FAMRequest req; 218 time_t stat_ts; 219 dev_t st_dev; 220 ino_t st_ino; 221 struct fam_dir_entry *fam_parent; 222 } fam_dir_entry; 223 224 typedef struct stat_cache_fam { 225 splay_tree *dirs; /* indexed by path; node data is fam_dir_entry */ 226 #ifdef HAVE_SYS_INOTIFY_H 227 splay_tree *wds; /* indexed by inotify watch descriptor */ 228 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 229 #else 230 FAMConnection fam; 231 #endif 232 log_error_st *errh; 233 fdevents *ev; 234 fdnode *fdn; 235 int fd; 236 } stat_cache_fam; 237 238 static fam_dir_entry * fam_dir_entry_init(const char *name, size_t len) 239 { 240 fam_dir_entry * const fam_dir = calloc(1, sizeof(*fam_dir)); 241 force_assert(NULL != fam_dir); 242 243 fam_dir->name = buffer_init(); 244 buffer_copy_string_len(fam_dir->name, name, len); 245 fam_dir->refcnt = 0; 246 #if defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 247 fam_dir->req = -1; 248 #endif 249 250 return fam_dir; 251 } 252 253 static void fam_dir_entry_free(fam_dir_entry *fam_dir) 254 { 255 if (!fam_dir) return; 256 /*(fam_dir->parent might be invalid pointer here; ignore)*/ 257 buffer_free(fam_dir->name); 258 #if defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 259 if (-1 != fam_dir->req) 260 close(fam_dir->req); 261 #endif 262 free(fam_dir); 263 } 264 265 static void fam_dir_invalidate_node(fam_dir_entry *fam_dir) 266 { 267 fam_dir->stat_ts = 0; 268 if (fam_dir->fam_parent) { 269 --fam_dir->fam_parent->refcnt; 270 fam_dir->fam_parent = NULL; 271 } 272 } 273 274 /* 275 * walk though splay_tree and collect contents of dir tree. 276 * remove tagged entries in a second loop 277 */ 278 279 static void fam_dir_tag_refcnt(splay_tree *t, int *keys, int *ndx) 280 { 281 if (*ndx == 512) return; /*(must match num array entries in keys[])*/ 282 if (t->left) fam_dir_tag_refcnt(t->left, keys, ndx); 283 if (t->right) fam_dir_tag_refcnt(t->right, keys, ndx); 284 if (*ndx == 512) return; /*(must match num array entries in keys[])*/ 285 286 fam_dir_entry * const fam_dir = t->data; 287 if (0 == fam_dir->refcnt) { 288 fam_dir_invalidate_node(fam_dir); 289 keys[(*ndx)++] = t->key; 290 } 291 } 292 293 __attribute_noinline__ 294 static void fam_dir_periodic_cleanup() { 295 stat_cache_fam * const scf = sc.scf; 296 int max_ndx, i; 297 int keys[512]; /* 2k size on stack */ 298 #if defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 299 struct kevent kevl[512]; /* 32k size on stack to batch kevent EV_DELETE */ 300 #endif 301 do { 302 if (!scf->dirs) break; 303 max_ndx = 0; 304 fam_dir_tag_refcnt(scf->dirs, keys, &max_ndx); 305 for (i = 0; i < max_ndx; ++i) { 306 const int ndx = keys[i]; 307 splay_tree *node = scf->dirs = splaytree_splay(scf->dirs, ndx); 308 if (node && node->key == ndx) { 309 fam_dir_entry *fam_dir = node->data; 310 scf->dirs = splaytree_delete(scf->dirs, ndx); 311 #ifdef HAVE_SYS_INOTIFY_H 312 scf->wds = splaytree_delete(scf->wds, fam_dir->req); 313 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 314 /* batch process kevent removal; defer cancel */ 315 EV_SET(kevl+i, fam_dir->req, EVFILT_VNODE, EV_DELETE, 0, 0, 0); 316 fam_dir->req = -1; /*(make FAMCancelMonitor() a no-op)*/ 317 #endif 318 FAMCancelMonitor(&scf->fam, &fam_dir->req); 319 fam_dir_entry_free(fam_dir); 320 } 321 } 322 #if defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 323 /* batch process: kevent() to submit EV_DELETE, then close dir fds */ 324 if (0 == max_ndx) break; 325 struct timespec t0 = { 0, 0 }; 326 kevent(scf->fd, kevl, max_ndx, NULL, 0, &t0); 327 for (i = 0; i < max_ndx; ++i) 328 close((int)kevl[i].ident); 329 #endif 330 } while (max_ndx == sizeof(keys)/sizeof(int)); 331 } 332 333 static void fam_dir_invalidate_tree(splay_tree *t, const char *name, size_t len) 334 { 335 #ifdef __clang_analyzer__ 336 force_assert(name); 337 #endif 338 /*force_assert(t);*/ 339 if (t->left) fam_dir_invalidate_tree(t->left, name, len); 340 if (t->right) fam_dir_invalidate_tree(t->right, name, len); 341 342 fam_dir_entry * const fam_dir = t->data; 343 #ifdef __clang_analyzer__ 344 force_assert(fam_dir); 345 #endif 346 buffer *b = fam_dir->name; 347 size_t blen = buffer_string_length(b); 348 if (blen > len && b->ptr[len] == '/' && 0 == memcmp(b->ptr, name, len)) 349 fam_dir_invalidate_node(fam_dir); 350 } 351 352 /* declarations */ 353 static void stat_cache_delete_tree(const char *name, uint32_t len); 354 static void stat_cache_invalidate_dir_tree(const char *name, size_t len); 355 static void stat_cache_handle_fdevent_fn(stat_cache_fam * const scf, fam_dir_entry * const fam_dir, const char * const fn, const uint32_t fnlen, int code); 356 357 static void stat_cache_handle_fdevent_in(stat_cache_fam *scf) 358 { 359 #ifdef HAVE_SYS_INOTIFY_H 360 /*(inotify pads in->len to align struct following in->name[])*/ 361 char buf[4096] 362 __attribute__ ((__aligned__(__alignof__(struct inotify_event)))); 363 int rd; 364 do { 365 rd = (int)read(scf->fd, buf, sizeof(buf)); 366 if (rd <= 0) { 367 if (-1 == rd && errno != EINTR && errno != EAGAIN) { 368 log_perror(scf->errh, __FILE__, __LINE__, "inotify error"); 369 /* TODO: could flush cache, close scf->fd, and re-open inotify*/ 370 } 371 break; 372 } 373 for (int i = 0; i < rd; ) { 374 struct inotify_event * const in = 375 (struct inotify_event *)((uintptr_t)buf + i); 376 uint32_t len = in->len; 377 if (len > sizeof(buf)) break; /*(should not happen)*/ 378 i += sizeof(struct inotify_event) + len; 379 if (i > rd) break; /*(should not happen (partial record))*/ 380 if (in->mask & IN_CREATE) 381 continue; /*(see comment below for FAMCreated)*/ 382 if (in->mask & IN_Q_OVERFLOW) { 383 log_error(scf->errh, __FILE__, __LINE__, 384 "inotify queue overflow"); 385 continue; 386 } 387 /* ignore events which may have been pending for 388 * paths recently cancelled via FAMCancelMonitor() */ 389 scf->wds = splaytree_splay(scf->wds, in->wd); 390 if (!scf->wds || scf->wds->key != in->wd) 391 continue; 392 fam_dir_entry *fam_dir = scf->wds->data; 393 if (NULL == fam_dir) /*(should not happen)*/ 394 continue; 395 if (fam_dir->req != in->wd) /*(should not happen)*/ 396 continue; 397 /*(specific to use here in stat_cache.c)*/ 398 int code = 0; 399 if (in->mask & (IN_ATTRIB | IN_MODIFY)) 400 code = FAMChanged; 401 else if (in->mask & (IN_DELETE | IN_DELETE_SELF | IN_UNMOUNT)) 402 code = FAMDeleted; 403 else if (in->mask & (IN_MOVE_SELF | IN_MOVED_FROM)) 404 code = FAMMoved; 405 406 if (len) { 407 do { --len; } while (len && in->name[len-1] == '\0'); 408 } 409 stat_cache_handle_fdevent_fn(scf, fam_dir, in->name, len, code); 410 } 411 } while (rd + sizeof(struct inotify_event) + NAME_MAX + 1 > sizeof(buf)); 412 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 413 struct kevent kevl[256]; 414 struct timespec t0 = { 0, 0 }; 415 int n; 416 do { 417 n = kevent(scf->fd, NULL, 0, kevl, sizeof(kevl)/sizeof(*kevl), &t0); 418 if (n <= 0) break; 419 for (int i = 0; i < n; ++i) { 420 const struct kevent * const kev = kevl+i; 421 /* ignore events which may have been pending for 422 * paths recently cancelled via FAMCancelMonitor() */ 423 int ndx = (int)(intptr_t)kev->udata; 424 scf->dirs = splaytree_splay(scf->dirs, ndx); 425 if (!scf->dirs || scf->dirs->key != ndx) 426 continue; 427 fam_dir_entry *fam_dir = scf->dirs->data; 428 if (fam_dir->req != (int)kev->ident) 429 continue; 430 /*(specific to use here in stat_cache.c)*/ 431 /* note: stat_cache only monitors on directories, 432 * so events here are only on directories 433 * note: changes are treated as FAMDeleted since 434 * it is unknown which file in dir was changed 435 * This is not efficient, but this stat_cache mechanism also 436 * should not be used on frequently modified directories. */ 437 int code = 0; 438 if (kev->fflags & (NOTE_WRITE|NOTE_ATTRIB|NOTE_EXTEND|NOTE_LINK)) 439 code = FAMDeleted; /*(not FAMChanged; see comment above)*/ 440 else if (kev->fflags & (NOTE_DELETE|NOTE_REVOKE)) 441 code = FAMDeleted; 442 else if (kev->fflags & NOTE_RENAME) 443 code = FAMMoved; 444 if (kev->flags & EV_ERROR) /*(not expected; treat as FAMDeleted)*/ 445 code = FAMDeleted; 446 stat_cache_handle_fdevent_fn(scf, fam_dir, NULL, 0, code); 447 } 448 } while (n == sizeof(kevl)/sizeof(*kevl)); 449 #else 450 for (int i = 0, ndx; i || (i = FAMPending(&scf->fam)) > 0; --i) { 451 FAMEvent fe; 452 if (FAMNextEvent(&scf->fam, &fe) < 0) break; 453 454 /* ignore events which may have been pending for 455 * paths recently cancelled via FAMCancelMonitor() */ 456 ndx = (int)(intptr_t)fe.userdata; 457 scf->dirs = splaytree_splay(scf->dirs, ndx); 458 if (!scf->dirs || scf->dirs->key != ndx) { 459 continue; 460 } 461 fam_dir_entry *fam_dir = scf->dirs->data; 462 if (FAMREQUEST_GETREQNUM(&fam_dir->req) 463 != FAMREQUEST_GETREQNUM(&fe.fr)) { 464 continue; 465 } 466 467 uint32_t fnlen = (fe.code != FAMCreated && fe.filename[0] != '/') 468 ? (uint32_t)strlen(fe.filename) 469 : 0; 470 stat_cache_handle_fdevent_fn(scf, fam_dir, fe.filename, fnlen, fe.code); 471 } 472 #endif 473 } 474 475 static void stat_cache_handle_fdevent_fn(stat_cache_fam * const scf, fam_dir_entry *fam_dir, const char * const fn, const uint32_t fnlen, int code) 476 { 477 if (fnlen) { 478 buffer * const n = fam_dir->name; 479 fam_dir_entry *fam_link; 480 uint32_t len; 481 switch (code) { 482 case FAMCreated: 483 /* file created in monitored dir modifies dir and 484 * we should get a separate FAMChanged event for dir. 485 * Therefore, ignore file FAMCreated event here. 486 * Also, if FAMNoExists() is used, might get spurious 487 * FAMCreated events as changes are made e.g. in monitored 488 * sub-sub-sub dirs and the library discovers new (already 489 * existing) dir entries */ 490 return; 491 case FAMChanged: 492 /* file changed in monitored dir does not modify dir */ 493 case FAMDeleted: 494 case FAMMoved: 495 /* file deleted or moved in monitored dir modifies dir, 496 * but FAM provides separate notification for that */ 497 498 /* temporarily append filename to dir in fam_dir->name to 499 * construct path, then delete stat_cache entry (if any)*/ 500 len = buffer_string_length(n); 501 buffer_append_string_len(n, CONST_STR_LEN("/")); 502 buffer_append_string_len(n, fn, fnlen); 503 /* (alternatively, could chose to stat() and update)*/ 504 stat_cache_invalidate_entry(CONST_BUF_LEN(n)); 505 506 fam_link = /*(check if might be symlink to monitored dir)*/ 507 stat_cache_sptree_find(&scf->dirs, CONST_BUF_LEN(n)); 508 if (fam_link && !buffer_is_equal(fam_link->name, n)) 509 fam_link = NULL; 510 511 buffer_string_set_length(n, len); 512 513 if (fam_link) { 514 /* replaced symlink changes containing dir */ 515 stat_cache_invalidate_entry(CONST_BUF_LEN(n)); 516 /* handle symlink to dir as deleted dir below */ 517 code = FAMDeleted; 518 fam_dir = fam_link; 519 break; 520 } 521 return; 522 default: 523 return; 524 } 525 } 526 527 switch(code) { 528 case FAMChanged: 529 stat_cache_invalidate_entry(CONST_BUF_LEN(fam_dir->name)); 530 break; 531 case FAMDeleted: 532 case FAMMoved: 533 stat_cache_delete_tree(CONST_BUF_LEN(fam_dir->name)); 534 fam_dir_invalidate_node(fam_dir); 535 if (scf->dirs) 536 fam_dir_invalidate_tree(scf->dirs,CONST_BUF_LEN(fam_dir->name)); 537 fam_dir_periodic_cleanup(); 538 break; 539 default: 540 break; 541 } 542 } 543 544 static handler_t stat_cache_handle_fdevent(void *ctx, int revent) 545 { 546 stat_cache_fam * const scf = ctx; /* sc.scf */ 547 548 if (revent & FDEVENT_IN) { 549 stat_cache_handle_fdevent_in(scf); 550 } 551 552 if (revent & (FDEVENT_HUP|FDEVENT_RDHUP)) { 553 /* fam closed the connection */ 554 log_error(scf->errh, __FILE__, __LINE__, 555 "FAM connection closed; disabling stat_cache."); 556 /* (although effectively STAT_CACHE_ENGINE_NONE, 557 * do not change here so that periodic jobs clean up memory)*/ 558 /*sc.stat_cache_engine = STAT_CACHE_ENGINE_NONE; */ 559 fdevent_fdnode_event_del(scf->ev, scf->fdn); 560 fdevent_unregister(scf->ev, scf->fd); 561 scf->fdn = NULL; 562 563 FAMClose(&scf->fam); 564 scf->fd = -1; 565 } 566 567 return HANDLER_GO_ON; 568 } 569 570 static stat_cache_fam * stat_cache_init_fam(fdevents *ev, log_error_st *errh) { 571 stat_cache_fam *scf = calloc(1, sizeof(*scf)); 572 force_assert(scf); 573 scf->fd = -1; 574 scf->ev = ev; 575 scf->errh = errh; 576 577 #ifdef HAVE_SYS_INOTIFY_H 578 scf->fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC); 579 if (scf->fd < 0) { 580 log_perror(errh, __FILE__, __LINE__, "inotify_init1()"); 581 free(scf); 582 return NULL; 583 } 584 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 585 #ifdef __NetBSD__ 586 scf->fd = kqueue1(O_NONBLOCK|O_CLOEXEC|O_NOSIGPIPE); 587 #else 588 scf->fd = kqueue(); 589 if (scf->fd >= 0) fdevent_setfd_cloexec(scf->fd); 590 #endif 591 if (scf->fd < 0) { 592 log_perror(errh, __FILE__, __LINE__, "kqueue()"); 593 free(scf); 594 return NULL; 595 } 596 #else 597 /* setup FAM */ 598 if (0 != FAMOpen2(&scf->fam, "lighttpd")) { 599 log_error(errh, __FILE__, __LINE__, 600 "could not open a fam connection, dying."); 601 free(scf); 602 return NULL; 603 } 604 #ifdef HAVE_FAMNOEXISTS 605 #ifdef LIGHTTPD_STATIC 606 FAMNoExists(&scf->fam); 607 #else 608 int (*FAMNoExists_fn)(FAMConnection *); 609 FAMNoExists_fn = 610 (int (*)(FAMConnection *))(intptr_t)dlsym(RTLD_DEFAULT,"FAMNoExists"); 611 if (FAMNoExists_fn) FAMNoExists_fn(&scf->fam); 612 #endif 613 #endif 614 615 scf->fd = FAMCONNECTION_GETFD(&scf->fam); 616 fdevent_setfd_cloexec(scf->fd); 617 #endif 618 scf->fdn = fdevent_register(scf->ev, scf->fd, stat_cache_handle_fdevent, scf); 619 fdevent_fdnode_event_set(scf->ev, scf->fdn, FDEVENT_IN | FDEVENT_RDHUP); 620 621 return scf; 622 } 623 624 static void stat_cache_free_fam(stat_cache_fam *scf) { 625 if (NULL == scf) return; 626 627 #ifdef HAVE_SYS_INOTIFY_H 628 while (scf->wds) { 629 splay_tree *node = scf->wds; 630 scf->wds = splaytree_delete(scf->wds, node->key); 631 } 632 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 633 /*(quicker cleanup to close kqueue() before cancel per entry)*/ 634 close(scf->fd); 635 scf->fd = -1; 636 #endif 637 while (scf->dirs) { 638 /*(skip entry invalidation and FAMCancelMonitor())*/ 639 splay_tree *node = scf->dirs; 640 fam_dir_entry_free((fam_dir_entry *)node->data); 641 scf->dirs = splaytree_delete(scf->dirs, node->key); 642 } 643 644 if (-1 != scf->fd) { 645 /*scf->fdn already cleaned up in fdevent_free()*/ 646 FAMClose(&scf->fam); 647 /*scf->fd = -1;*/ 648 } 649 650 free(scf); 651 } 652 653 static fam_dir_entry * fam_dir_monitor(stat_cache_fam *scf, char *fn, uint32_t dirlen, struct stat *st) 654 { 655 if (NULL == scf->fdn) return NULL; /* FAM connection closed; do nothing */ 656 const int fn_is_dir = S_ISDIR(st->st_mode); 657 /*force_assert(0 != dirlen);*/ 658 /*force_assert(fn[0] == '/');*/ 659 /* consistency: ensure fn does not end in '/' unless root "/" 660 * FAM events will not end in '/', so easier to match this way */ 661 if (fn[dirlen-1] == '/') --dirlen; 662 if (0 == dirlen) dirlen = 1; /* root dir ("/") */ 663 /* Note: paths are expected to be normalized before calling stat_cache, 664 * e.g. without repeated '/' */ 665 if (!fn_is_dir) { 666 while (fn[--dirlen] != '/') ; 667 if (0 == dirlen) dirlen = 1; /*(should not happen for file)*/ 668 } 669 int dir_ndx = splaytree_djbhash(fn, dirlen); 670 fam_dir_entry *fam_dir = NULL; 671 672 scf->dirs = splaytree_splay(scf->dirs, dir_ndx); 673 if (NULL != scf->dirs && scf->dirs->key == dir_ndx) { 674 fam_dir = scf->dirs->data; 675 if (!buffer_is_equal_string(fam_dir->name, fn, dirlen)) { 676 /* hash collision; preserve existing 677 * do not monitor new to avoid cache thrashing */ 678 return NULL; 679 } 680 /* directory already registered */ 681 } 682 683 const time_t cur_ts = log_epoch_secs; 684 struct stat lst; 685 int ck_dir = fn_is_dir; 686 if (!fn_is_dir && (NULL==fam_dir || cur_ts - fam_dir->stat_ts >= 16)) { 687 ck_dir = 1; 688 /*(temporarily modify fn)*/ 689 fn[dirlen] = '\0'; 690 if (0 != lstat(fn, &lst)) { 691 fn[dirlen] = '/'; 692 return NULL; 693 } 694 if (!S_ISLNK(lst.st_mode)) { 695 st = &lst; 696 } 697 else if (0 != stat(fn, st)) { /*st passed in now is stat() of dir*/ 698 fn[dirlen] = '/'; 699 return NULL; 700 } 701 fn[dirlen] = '/'; 702 } 703 704 int ck_lnk = (NULL == fam_dir); 705 if (ck_dir && NULL != fam_dir) { 706 /* check stat() matches device and inode, just in case an external event 707 * not being monitored occurs (e.g. rename of unmonitored parent dir)*/ 708 if (st->st_dev != fam_dir->st_dev || st->st_ino != fam_dir->st_ino) { 709 ck_lnk = 1; 710 /*(modifies scf->dirs but no need to re-splay for dir_ndx since 711 * fam_dir is not NULL and so splaytree_insert not called below)*/ 712 if (scf->dirs) fam_dir_invalidate_tree(scf->dirs, fn, dirlen); 713 if (!fn_is_dir) /*(if dir, caller is updating stat_cache_entry)*/ 714 stat_cache_update_entry(fn, dirlen, st, NULL); 715 /*(must not delete tree since caller is holding a valid node)*/ 716 stat_cache_invalidate_dir_tree(fn, dirlen); 717 #ifdef HAVE_SYS_INOTIFY_H 718 scf->wds = splaytree_delete(scf->wds, fam_dir->req); 719 #endif 720 if (0 != FAMCancelMonitor(&scf->fam, &fam_dir->req) 721 || 0 != FAMMonitorDirectory(&scf->fam, fam_dir->name->ptr, 722 &fam_dir->req, 723 (void *)(intptr_t)dir_ndx)) { 724 fam_dir->stat_ts = 0; /* invalidate */ 725 return NULL; 726 } 727 fam_dir->st_dev = st->st_dev; 728 fam_dir->st_ino = st->st_ino; 729 #ifdef HAVE_SYS_INOTIFY_H 730 scf->wds = splaytree_insert(scf->wds, fam_dir->req, fam_dir); 731 #endif 732 } 733 fam_dir->stat_ts = cur_ts; 734 } 735 736 if (NULL == fam_dir) { 737 fam_dir = fam_dir_entry_init(fn, dirlen); 738 739 if (0 != FAMMonitorDirectory(&scf->fam,fam_dir->name->ptr,&fam_dir->req, 740 (void *)(intptr_t)dir_ndx)) { 741 #if defined(HAVE_SYS_INOTIFY_H) \ 742 || (defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE) 743 log_perror(scf->errh, __FILE__, __LINE__, 744 "monitoring dir failed: %s file: %s", 745 fam_dir->name->ptr, fn); 746 #else 747 log_error(scf->errh, __FILE__, __LINE__, 748 "monitoring dir failed: %s file: %s %s", 749 fam_dir->name->ptr, fn, FamErrlist[FAMErrno]); 750 #endif 751 fam_dir_entry_free(fam_dir); 752 return NULL; 753 } 754 755 scf->dirs = splaytree_insert(scf->dirs, dir_ndx, fam_dir); 756 #ifdef HAVE_SYS_INOTIFY_H 757 scf->wds = splaytree_insert(scf->wds, fam_dir->req, fam_dir); 758 #endif 759 fam_dir->stat_ts= cur_ts; 760 fam_dir->st_dev = st->st_dev; 761 fam_dir->st_ino = st->st_ino; 762 } 763 764 if (ck_lnk) { 765 if (fn_is_dir) { 766 /*(temporarily modify fn)*/ 767 char e = fn[dirlen]; 768 fn[dirlen] = '\0'; 769 if (0 != lstat(fn, &lst)) { 770 fn[dirlen] = e; 771 return NULL; 772 } 773 fn[dirlen] = e; 774 } 775 if (fam_dir->fam_parent) { 776 --fam_dir->fam_parent->refcnt; 777 fam_dir->fam_parent = NULL; 778 } 779 if (S_ISLNK(lst.st_mode)) { 780 fam_dir->fam_parent = fam_dir_monitor(scf, fn, dirlen, &lst); 781 } 782 } 783 784 ++fam_dir->refcnt; 785 return fam_dir; 786 } 787 788 #endif 789 790 791 static stat_cache_entry * stat_cache_entry_init(void) { 792 stat_cache_entry *sce = calloc(1, sizeof(*sce)); 793 force_assert(NULL != sce); 794 sce->fd = -1; 795 sce->refcnt = 1; 796 return sce; 797 } 798 799 static void stat_cache_entry_free(void *data) { 800 stat_cache_entry *sce = data; 801 if (!sce) return; 802 803 if (--sce->refcnt) return; 804 805 #ifdef HAVE_FAM_H 806 /*(decrement refcnt only; 807 * defer cancelling FAM monitor on dir even if refcnt reaches zero)*/ 808 if (sce->fam_dir) --((fam_dir_entry *)sce->fam_dir)->refcnt; 809 #endif 810 811 free(sce->name.ptr); 812 free(sce->etag.ptr); 813 if (sce->content_type.size) free(sce->content_type.ptr); 814 if (sce->fd >= 0) close(sce->fd); 815 816 free(sce); 817 } 818 819 void stat_cache_entry_refchg(void *data, int mod) { 820 /*(expect mod == -1 or mod == 1)*/ 821 stat_cache_entry * const sce = data; 822 if (mod < 0 && 1 == sce->refcnt) 823 stat_cache_entry_free(data); 824 else 825 sce->refcnt += mod; 826 } 827 828 #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR) 829 830 static const char *attrname = "Content-Type"; 831 static char attrval[128]; 832 static buffer attrb = { attrval, 0, 0 }; 833 834 static int stat_cache_attr_get(const char *name) { 835 #if defined(HAVE_XATTR) 836 #if defined(HAVE_SYS_XATTR_H) 837 ssize_t attrlen; 838 if (0 < (attrlen = getxattr(name, attrname, 839 attrval, sizeof(attrval)-1))) 840 #else 841 int attrlen = sizeof(attrval)-1; 842 if (0 == attr_get(name, attrname, attrval, &attrlen, 0)) 843 #endif 844 #elif defined(HAVE_EXTATTR) 845 ssize_t attrlen; 846 if (0 < (attrlen = extattr_get_file(name, EXTATTR_NAMESPACE_USER, attrname, 847 attrval, sizeof(attrval)-1))) 848 #endif 849 { 850 attrval[attrlen] = '\0'; 851 attrb.used = (uint32_t)(attrlen + 1); 852 return 1; 853 } 854 return 0; 855 } 856 857 #endif 858 859 int stat_cache_init(fdevents *ev, log_error_st *errh) { 860 #ifdef HAVE_FAM_H 861 if (sc.stat_cache_engine == STAT_CACHE_ENGINE_FAM) { 862 sc.scf = stat_cache_init_fam(ev, errh); 863 if (NULL == sc.scf) return 0; 864 } 865 #else 866 UNUSED(ev); 867 UNUSED(errh); 868 #endif 869 870 return 1; 871 } 872 873 void stat_cache_free(void) { 874 splay_tree *sptree = sc.files; 875 while (sptree) { 876 stat_cache_entry_free(sptree->data); 877 sptree = splaytree_delete(sptree, sptree->key); 878 } 879 sc.files = NULL; 880 881 #ifdef HAVE_FAM_H 882 stat_cache_free_fam(sc.scf); 883 sc.scf = NULL; 884 #endif 885 886 #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR) 887 attrname = "Content-Type"; 888 #endif 889 890 sc.stat_cache_engine = STAT_CACHE_ENGINE_SIMPLE; /*(default)*/ 891 } 892 893 void stat_cache_xattrname (const char *name) { 894 #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR) 895 attrname = name; 896 #else 897 UNUSED(name); 898 #endif 899 } 900 901 int stat_cache_choose_engine (const buffer *stat_cache_string, log_error_st *errh) { 902 if (buffer_string_is_empty(stat_cache_string)) 903 sc.stat_cache_engine = STAT_CACHE_ENGINE_SIMPLE; 904 else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("simple"))) 905 sc.stat_cache_engine = STAT_CACHE_ENGINE_SIMPLE; 906 #ifdef HAVE_SYS_INOTIFY_H 907 else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("inotify"))) 908 sc.stat_cache_engine = STAT_CACHE_ENGINE_INOTIFY; 909 /*(STAT_CACHE_ENGINE_FAM == STAT_CACHE_ENGINE_INOTIFY)*/ 910 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 911 else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("kqueue"))) 912 sc.stat_cache_engine = STAT_CACHE_ENGINE_KQUEUE; 913 /*(STAT_CACHE_ENGINE_FAM == STAT_CACHE_ENGINE_KQUEUE)*/ 914 #endif 915 #ifdef HAVE_FAM_H 916 else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("fam"))) 917 sc.stat_cache_engine = STAT_CACHE_ENGINE_FAM; 918 #endif 919 else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("disable")) 920 || buffer_eq_slen(stat_cache_string, CONST_STR_LEN("none"))) 921 sc.stat_cache_engine = STAT_CACHE_ENGINE_NONE; 922 else { 923 log_error(errh, __FILE__, __LINE__, 924 "server.stat-cache-engine can be one of \"disable\", \"simple\"," 925 #ifdef HAVE_SYS_INOTIFY_H 926 " \"inotify\"," 927 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE 928 " \"kqueue\"," 929 #endif 930 #ifdef HAVE_FAM_H 931 " \"fam\"," 932 #endif 933 " but not: %s", stat_cache_string->ptr); 934 return -1; 935 } 936 return 0; 937 } 938 939 const buffer * stat_cache_mimetype_by_ext(const array * const mimetypes, const char * const name, const uint32_t nlen) 940 { 941 const char * const end = name + nlen; /*(end of string)*/ 942 const uint32_t used = mimetypes->used; 943 if (used < 16) { 944 for (uint32_t i = 0; i < used; ++i) { 945 /* suffix match */ 946 const data_string *ds = (data_string *)mimetypes->data[i]; 947 const size_t klen = buffer_string_length(&ds->key); 948 if (klen <= nlen && buffer_eq_icase_ssn(end-klen, ds->key.ptr, klen)) 949 return &ds->value; 950 } 951 } 952 else { 953 const char *s; 954 const data_string *ds; 955 if (nlen) { 956 for (s = end-1; s != name && *s != '/'; --s) ; /*(like memrchr())*/ 957 if (*s == '/') ++s; 958 } 959 else { 960 s = name; 961 } 962 /* search for basename, then longest .ext2.ext1, then .ext1, then "" */ 963 ds = (const data_string *)array_get_element_klen(mimetypes, s, end - s); 964 if (NULL != ds) return &ds->value; 965 while (++s < end) { 966 while (*s != '.' && ++s != end) ; 967 if (s == end) break; 968 /* search ".ext" then "ext" */ 969 ds = (const data_string *)array_get_element_klen(mimetypes, s, end - s); 970 if (NULL != ds) return &ds->value; 971 /* repeat search without leading '.' to handle situation where 972 * admin configured mimetype.assign keys without leading '.' */ 973 if (++s < end) { 974 if (*s == '.') { --s; continue; } 975 ds = (const data_string *)array_get_element_klen(mimetypes, s, end - s); 976 if (NULL != ds) return &ds->value; 977 } 978 } 979 /* search for ""; catchall */ 980 ds = (const data_string *)array_get_element_klen(mimetypes, CONST_STR_LEN("")); 981 if (NULL != ds) return &ds->value; 982 } 983 984 return NULL; 985 } 986 987 #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR) 988 989 const buffer * stat_cache_mimetype_by_xattr(const char * const name) 990 { 991 return stat_cache_attr_get(name) ? &attrb : NULL; 992 } 993 994 const buffer * stat_cache_content_type_get_by_xattr(stat_cache_entry *sce, const array *mimetypes, int use_xattr) 995 { 996 /*(invalid caching if user config has multiple, different 997 * r->conf.mimetypes for same extension (not expected))*/ 998 if (!buffer_string_is_empty(&sce->content_type)) return &sce->content_type; 999 1000 if (!S_ISREG(sce->st.st_mode)) return NULL; 1001 1002 /* cache mimetype */ 1003 const buffer *mtype = 1004 (use_xattr) ? stat_cache_mimetype_by_xattr(sce->name.ptr) : NULL; 1005 if (NULL == mtype) 1006 mtype = stat_cache_mimetype_by_ext(mimetypes,CONST_BUF_LEN(&sce->name)); 1007 if (NULL != mtype) { 1008 if (sce->content_type.size) { 1009 buffer_copy_buffer(&sce->content_type, mtype); 1010 } 1011 else if (mtype == &attrb) { 1012 sce->content_type.ptr = NULL; 1013 buffer_copy_buffer(&sce->content_type, mtype); 1014 } 1015 else { 1016 /*(copy pointers from mimetypes array; avoid allocation)*/ 1017 sce->content_type.ptr = mtype->ptr; 1018 sce->content_type.used = mtype->used; 1019 /*(leave sce->content_type.size = 0 to flag not-allocated)*/ 1020 } 1021 } 1022 else 1023 buffer_clear(&sce->content_type); 1024 1025 return &sce->content_type; 1026 } 1027 1028 #else 1029 1030 const buffer * stat_cache_content_type_get_by_ext(stat_cache_entry *sce, const array *mimetypes) 1031 { 1032 /*(invalid caching if user config has multiple, different 1033 * r->conf.mimetypes for same extension (not expected))*/ 1034 if (!buffer_string_is_empty(&sce->content_type)) return &sce->content_type; 1035 1036 if (!S_ISREG(sce->st.st_mode)) return NULL; 1037 1038 /* cache mimetype */ 1039 const buffer * const mtype = 1040 stat_cache_mimetype_by_ext(mimetypes, CONST_BUF_LEN(&sce->name)); 1041 if (NULL != mtype) { 1042 /*(copy pointers from mimetypes array; avoid allocation)*/ 1043 sce->content_type.ptr = mtype->ptr; 1044 sce->content_type.used = mtype->used; 1045 /*(leave sce->content_type.size = 0 to flag not-allocated)*/ 1046 } 1047 else 1048 buffer_clear(&sce->content_type); 1049 1050 return &sce->content_type; 1051 } 1052 1053 #endif 1054 1055 const buffer * stat_cache_etag_get(stat_cache_entry *sce, int flags) { 1056 /*(invalid caching if user cfg has multiple, different r->conf.etag_flags 1057 * for same path (not expected, since etag flags should be by filesystem))*/ 1058 if (!buffer_string_is_empty(&sce->etag)) return &sce->etag; 1059 1060 if (S_ISREG(sce->st.st_mode) || S_ISDIR(sce->st.st_mode)) { 1061 if (0 == flags) return NULL; 1062 etag_create(&sce->etag, &sce->st, flags); 1063 return &sce->etag; 1064 } 1065 1066 return NULL; 1067 } 1068 1069 __attribute_pure__ 1070 static int stat_cache_stat_eq(const struct stat * const sta, const struct stat * const stb) { 1071 return 1072 #ifdef st_mtime /* use high-precision timestamp if available */ 1073 #if defined(__APPLE__) && defined(__MACH__) 1074 sta->st_mtimespec.tv_nsec == stb->st_mtimespec.tv_nsec 1075 #else 1076 sta->st_mtim.tv_nsec == stb->st_mtim.tv_nsec 1077 #endif 1078 #else 1079 1 1080 #endif 1081 && sta->st_mtime == stb->st_mtime 1082 && sta->st_size == stb->st_size 1083 && sta->st_ino == stb->st_ino 1084 && sta->st_dev == stb->st_dev; 1085 } 1086 1087 void stat_cache_update_entry(const char *name, uint32_t len, 1088 struct stat *st, buffer *etagb) 1089 { 1090 if (sc.stat_cache_engine == STAT_CACHE_ENGINE_NONE) return; 1091 force_assert(0 != len); 1092 if (name[len-1] == '/') { if (0 == --len) len = 1; } 1093 splay_tree **sptree = &sc.files; 1094 stat_cache_entry *sce = 1095 stat_cache_sptree_find(sptree, name, len); 1096 if (sce && buffer_is_equal_string(&sce->name, name, len)) { 1097 if (!stat_cache_stat_eq(&sce->st, st)) { 1098 /* etagb might be NULL to clear etag (invalidate) */ 1099 buffer_copy_string_len(&sce->etag, CONST_BUF_LEN(etagb)); 1100 #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR) 1101 buffer_clear(&sce->content_type); 1102 #endif 1103 if (sce->fd >= 0) { 1104 if (1 == sce->refcnt) { 1105 close(sce->fd); 1106 sce->fd = -1; 1107 } 1108 else { 1109 --sce->refcnt; /* stat_cache_entry_free(sce); */ 1110 (*sptree)->data = sce = stat_cache_entry_init(); 1111 buffer_copy_string_len(&sce->name, name, len); 1112 } 1113 } 1114 sce->st = *st; 1115 } 1116 sce->stat_ts = log_epoch_secs; 1117 } 1118 } 1119 1120 void stat_cache_delete_entry(const char *name, uint32_t len) 1121 { 1122 if (sc.stat_cache_engine == STAT_CACHE_ENGINE_NONE) return; 1123 force_assert(0 != len); 1124 if (name[len-1] == '/') { if (0 == --len) len = 1; } 1125 splay_tree **sptree = &sc.files; 1126 stat_cache_entry *sce = stat_cache_sptree_find(sptree, name, len); 1127 if (sce && buffer_is_equal_string(&sce->name, name, len)) { 1128 stat_cache_entry_free(sce); 1129 *sptree = splaytree_delete(*sptree, (*sptree)->key); 1130 } 1131 } 1132 1133 void stat_cache_invalidate_entry(const char *name, uint32_t len) 1134 { 1135 splay_tree **sptree = &sc.files; 1136 stat_cache_entry *sce = stat_cache_sptree_find(sptree, name, len); 1137 if (sce && buffer_is_equal_string(&sce->name, name, len)) { 1138 sce->stat_ts = 0; 1139 #ifdef HAVE_FAM_H 1140 if (sce->fam_dir != NULL) { 1141 --((fam_dir_entry *)sce->fam_dir)->refcnt; 1142 sce->fam_dir = NULL; 1143 } 1144 #endif 1145 } 1146 } 1147 1148 #ifdef HAVE_FAM_H 1149 1150 static void stat_cache_invalidate_dir_tree_walk(splay_tree *t, 1151 const char *name, size_t len) 1152 { 1153 if (t->left) stat_cache_invalidate_dir_tree_walk(t->left, name, len); 1154 if (t->right) stat_cache_invalidate_dir_tree_walk(t->right, name, len); 1155 1156 buffer *b = &((stat_cache_entry *)t->data)->name; 1157 size_t blen = buffer_string_length(b); 1158 if (blen > len && b->ptr[len] == '/' && 0 == memcmp(b->ptr, name, len)) { 1159 stat_cache_entry *sce = t->data; 1160 sce->stat_ts = 0; 1161 if (sce->fam_dir != NULL) { 1162 --((fam_dir_entry *)sce->fam_dir)->refcnt; 1163 sce->fam_dir = NULL; 1164 } 1165 } 1166 } 1167 1168 static void stat_cache_invalidate_dir_tree(const char *name, size_t len) 1169 { 1170 splay_tree * const sptree = sc.files; 1171 if (sptree) stat_cache_invalidate_dir_tree_walk(sptree, name, len); 1172 } 1173 1174 #endif 1175 1176 /* 1177 * walk though splay_tree and collect contents of dir tree. 1178 * remove tagged entries in a second loop 1179 */ 1180 1181 static void stat_cache_tag_dir_tree(splay_tree *t, const char *name, size_t len, 1182 int *keys, int *ndx) 1183 { 1184 if (*ndx == 8192) return; /*(must match num array entries in keys[])*/ 1185 if (t->left) stat_cache_tag_dir_tree(t->left, name, len, keys, ndx); 1186 if (t->right) stat_cache_tag_dir_tree(t->right, name, len, keys, ndx); 1187 if (*ndx == 8192) return; /*(must match num array entries in keys[])*/ 1188 1189 buffer *b = &((stat_cache_entry *)t->data)->name; 1190 size_t blen = buffer_string_length(b); 1191 if (blen > len && b->ptr[len] == '/' && 0 == memcmp(b->ptr, name, len)) 1192 keys[(*ndx)++] = t->key; 1193 } 1194 1195 __attribute_noinline__ 1196 static void stat_cache_prune_dir_tree(const char *name, size_t len) 1197 { 1198 splay_tree *sptree = sc.files; 1199 int max_ndx, i; 1200 int keys[8192]; /* 32k size on stack */ 1201 do { 1202 if (!sptree) break; 1203 max_ndx = 0; 1204 stat_cache_tag_dir_tree(sptree, name, len, keys, &max_ndx); 1205 for (i = 0; i < max_ndx; ++i) { 1206 const int ndx = keys[i]; 1207 splay_tree *node = sptree = splaytree_splay(sptree, ndx); 1208 if (node && node->key == ndx) { 1209 stat_cache_entry_free(node->data); 1210 sptree = splaytree_delete(sptree, ndx); 1211 } 1212 } 1213 } while (max_ndx == sizeof(keys)/sizeof(int)); 1214 sc.files = sptree; 1215 } 1216 1217 static void stat_cache_delete_tree(const char *name, uint32_t len) 1218 { 1219 stat_cache_delete_entry(name, len); 1220 stat_cache_prune_dir_tree(name, len); 1221 } 1222 1223 void stat_cache_delete_dir(const char *name, uint32_t len) 1224 { 1225 force_assert(0 != len); 1226 if (name[len-1] == '/') { if (0 == --len) len = 1; } 1227 stat_cache_delete_tree(name, len); 1228 #ifdef HAVE_FAM_H 1229 if (sc.stat_cache_engine == STAT_CACHE_ENGINE_FAM) { 1230 splay_tree **sptree = &sc.scf->dirs; 1231 fam_dir_entry *fam_dir = stat_cache_sptree_find(sptree, name, len); 1232 if (fam_dir && buffer_is_equal_string(fam_dir->name, name, len)) 1233 fam_dir_invalidate_node(fam_dir); 1234 if (*sptree) fam_dir_invalidate_tree(*sptree, name, len); 1235 fam_dir_periodic_cleanup(); 1236 } 1237 #endif 1238 } 1239 1240 /*** 1241 * 1242 * 1243 * 1244 * returns: 1245 * - HANDLER_FINISHED on cache-miss (don't forget to reopen the file) 1246 * - HANDLER_ERROR on stat() failed -> see errno for problem 1247 */ 1248 1249 stat_cache_entry * stat_cache_get_entry(const buffer *name) { 1250 stat_cache_entry *sce = NULL; 1251 struct stat st; 1252 int file_ndx; 1253 1254 /* consistency: ensure lookup name does not end in '/' unless root "/" 1255 * (but use full path given with stat(), even with trailing '/') */ 1256 int final_slash = 0; 1257 size_t len = buffer_string_length(name); 1258 force_assert(0 != len); 1259 if (name->ptr[len-1] == '/') { final_slash = 1; if (0 == --len) len = 1; } 1260 /* Note: paths are expected to be normalized before calling stat_cache, 1261 * e.g. without repeated '/' */ 1262 1263 if (name->ptr[0] != '/') { 1264 errno = EINVAL; 1265 return NULL; 1266 } 1267 1268 /* 1269 * check if the directory for this file has changed 1270 */ 1271 1272 const time_t cur_ts = log_epoch_secs; 1273 1274 file_ndx = splaytree_djbhash(name->ptr, len); 1275 splay_tree *sptree = sc.files = splaytree_splay(sc.files, file_ndx); 1276 1277 if (sptree && (sptree->key == file_ndx)) { 1278 /* we have seen this file already and 1279 * don't stat() it again in the same second */ 1280 1281 sce = sptree->data; 1282 1283 /* check if the name is the same, we might have a collision */ 1284 1285 if (buffer_is_equal_string(&sce->name, name->ptr, len)) { 1286 if (sc.stat_cache_engine == STAT_CACHE_ENGINE_SIMPLE) { 1287 if (sce->stat_ts == cur_ts) { 1288 if (final_slash && !S_ISDIR(sce->st.st_mode)) { 1289 errno = ENOTDIR; 1290 return NULL; 1291 } 1292 return sce; 1293 } 1294 } 1295 #ifdef HAVE_FAM_H 1296 else if (sc.stat_cache_engine == STAT_CACHE_ENGINE_FAM 1297 && sce->fam_dir) { /* entry is in monitored dir */ 1298 /* re-stat() periodically, even if monitoring for changes 1299 * (due to limitations in stat_cache.c use of FAM) 1300 * (gaps due to not continually monitoring an entire tree) */ 1301 if (cur_ts - sce->stat_ts < 16) { 1302 if (final_slash && !S_ISDIR(sce->st.st_mode)) { 1303 errno = ENOTDIR; 1304 return NULL; 1305 } 1306 return sce; 1307 } 1308 } 1309 #endif 1310 } else { 1311 /* collision, forget about the entry */ 1312 sce = NULL; 1313 } 1314 } 1315 1316 if (-1 == stat(name->ptr, &st)) { 1317 return NULL; 1318 } 1319 1320 if (S_ISREG(st.st_mode)) { 1321 /* fix broken stat/open for symlinks to reg files with appended slash on freebsd,osx */ 1322 if (name->ptr[buffer_string_length(name) - 1] == '/') { 1323 errno = ENOTDIR; 1324 return NULL; 1325 } 1326 } 1327 1328 if (NULL == sce) { 1329 1330 sce = stat_cache_entry_init(); 1331 buffer_copy_string_len(&sce->name, name->ptr, len); 1332 1333 /* already splayed file_ndx */ 1334 if (NULL != sptree && sptree->key == file_ndx) { 1335 /* hash collision: replace old entry */ 1336 stat_cache_entry_free(sptree->data); 1337 sptree->data = sce; 1338 } else { 1339 sptree = sc.files = splaytree_insert(sptree, file_ndx, sce); 1340 } 1341 1342 } else { 1343 1344 buffer_clear(&sce->etag); 1345 #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR) 1346 buffer_clear(&sce->content_type); 1347 #endif 1348 1349 } 1350 1351 if (sce->fd >= 0) { 1352 /* close fd if file changed */ 1353 if (!stat_cache_stat_eq(&sce->st, &st)) { 1354 if (1 == sce->refcnt) { 1355 close(sce->fd); 1356 sce->fd = -1; 1357 } 1358 else { 1359 --sce->refcnt; /* stat_cache_entry_free(sce); */ 1360 sptree->data = sce = stat_cache_entry_init(); 1361 buffer_copy_string_len(&sce->name, name->ptr, len); 1362 } 1363 } 1364 } 1365 1366 sce->st = st; /*(copy prior to calling fam_dir_monitor())*/ 1367 1368 #ifdef HAVE_FAM_H 1369 if (sc.stat_cache_engine == STAT_CACHE_ENGINE_FAM) { 1370 if (sce->fam_dir) --((fam_dir_entry *)sce->fam_dir)->refcnt; 1371 sce->fam_dir = 1372 fam_dir_monitor(sc.scf, CONST_BUF_LEN(name), &st); 1373 #if 0 /*(performed below)*/ 1374 if (NULL != sce->fam_dir) { 1375 /*(may have been invalidated by dir change)*/ 1376 sce->stat_ts = cur_ts; 1377 } 1378 #endif 1379 } 1380 #endif 1381 1382 sce->stat_ts = cur_ts; 1383 return sce; 1384 } 1385 1386 stat_cache_entry * stat_cache_get_entry_open(const buffer * const name, const int symlinks) { 1387 stat_cache_entry * const sce = stat_cache_get_entry(name); 1388 if (NULL == sce) return NULL; 1389 if (sce->fd >= 0) return sce; 1390 if (sce->st.st_size > 0) 1391 sce->fd = stat_cache_open_rdonly_fstat(name, &sce->st, symlinks); 1392 return sce; /* (note: sce->fd might still be -1 if open() failed) */ 1393 } 1394 1395 const stat_cache_st * stat_cache_path_stat (const buffer * const name) { 1396 const stat_cache_entry * const sce = stat_cache_get_entry(name); 1397 return sce ? &sce->st : NULL; 1398 } 1399 1400 int stat_cache_path_isdir(const buffer *name) { 1401 const stat_cache_entry * const sce = stat_cache_get_entry(name); 1402 return (sce && (S_ISDIR(sce->st.st_mode) ? 1 : (errno = ENOTDIR, 0))); 1403 } 1404 1405 int stat_cache_path_contains_symlink(const buffer *name, log_error_st *errh) { 1406 /* caller should check for symlinks only if we should block symlinks. */ 1407 1408 /* catch the obvious symlinks 1409 * 1410 * this is not a secure check as we still have a race-condition between 1411 * the stat() and the open. We can only solve this by 1412 * 1. open() the file 1413 * 2. fstat() the fd 1414 * 1415 * and keeping the file open for the rest of the time. But this can 1416 * only be done at network level. 1417 * */ 1418 1419 #ifdef HAVE_LSTAT 1420 /* we assume "/" can not be symlink, 1421 * so skip the symlink stuff if path is "/" */ 1422 size_t len = buffer_string_length(name); 1423 force_assert(0 != len); 1424 force_assert(name->ptr[0] == '/'); 1425 if (1 == len) return 0; 1426 #ifndef PATH_MAX 1427 #define PATH_MAX 4096 1428 #endif 1429 if (len >= PATH_MAX) return -1; 1430 1431 char buf[PATH_MAX]; 1432 memcpy(buf, name->ptr, len); 1433 char *s_cur = buf+len; 1434 do { 1435 *s_cur = '\0'; 1436 struct stat st; 1437 if (0 == lstat(buf, &st)) { 1438 if (S_ISLNK(st.st_mode)) return 1; 1439 } 1440 else { 1441 log_perror(errh, __FILE__, __LINE__, "lstat failed for: %s", buf); 1442 return -1; 1443 } 1444 } while ((s_cur = strrchr(buf, '/')) > buf); /*(&buf[0]==buf; NULL < buf)*/ 1445 #endif 1446 1447 return 0; 1448 } 1449 1450 int stat_cache_open_rdonly_fstat (const buffer *name, struct stat *st, int symlinks) { 1451 /*(Note: O_NOFOLLOW affects only the final path segment, the target file, 1452 * not any intermediate symlinks along the path)*/ 1453 const int fd = fdevent_open_cloexec(name->ptr, symlinks, O_RDONLY, 0); 1454 if (fd >= 0) { 1455 if (0 == fstat(fd, st)) { 1456 return fd; 1457 } else { 1458 const int errnum = errno; 1459 close(fd); 1460 errno = errnum; 1461 } 1462 } 1463 return -1; 1464 } 1465 1466 /** 1467 * remove stat() from cache which haven't been stat()ed for 1468 * more than 2 seconds 1469 * 1470 * 1471 * walk though the stat-cache, collect the ids which are too old 1472 * and remove them in a second loop 1473 */ 1474 1475 static void stat_cache_tag_old_entries(splay_tree * const t, int * const keys, int * const ndx, const time_t max_age, const time_t cur_ts) { 1476 if (*ndx == 8192) return; /*(must match num array entries in keys[])*/ 1477 if (t->left) 1478 stat_cache_tag_old_entries(t->left, keys, ndx, max_age, cur_ts); 1479 if (t->right) 1480 stat_cache_tag_old_entries(t->right, keys, ndx, max_age, cur_ts); 1481 if (*ndx == 8192) return; /*(must match num array entries in keys[])*/ 1482 1483 const stat_cache_entry * const sce = t->data; 1484 if (cur_ts - sce->stat_ts > max_age) 1485 keys[(*ndx)++] = t->key; 1486 } 1487 1488 static void stat_cache_periodic_cleanup(const time_t max_age, const time_t cur_ts) { 1489 splay_tree *sptree = sc.files; 1490 int max_ndx, i; 1491 int keys[8192]; /* 32k size on stack */ 1492 do { 1493 if (!sptree) break; 1494 max_ndx = 0; 1495 stat_cache_tag_old_entries(sptree, keys, &max_ndx, max_age, cur_ts); 1496 for (i = 0; i < max_ndx; ++i) { 1497 int ndx = keys[i]; 1498 sptree = splaytree_splay(sptree, ndx); 1499 if (sptree && sptree->key == ndx) { 1500 stat_cache_entry_free(sptree->data); 1501 sptree = splaytree_delete(sptree, ndx); 1502 } 1503 } 1504 } while (max_ndx == sizeof(keys)/sizeof(int)); 1505 sc.files = sptree; 1506 } 1507 1508 void stat_cache_trigger_cleanup(void) { 1509 time_t max_age = 2; 1510 1511 #ifdef HAVE_FAM_H 1512 if (STAT_CACHE_ENGINE_FAM == sc.stat_cache_engine) { 1513 if (log_epoch_secs & 0x1F) return; 1514 /* once every 32 seconds (0x1F == 31) */ 1515 max_age = 32; 1516 fam_dir_periodic_cleanup(); 1517 /* By doing this before stat_cache_periodic_cleanup(), 1518 * entries used within the next max_age secs will remain 1519 * monitored, instead of effectively flushing and 1520 * rebuilding the FAM monitoring every max_age seconds */ 1521 } 1522 #endif 1523 1524 stat_cache_periodic_cleanup(max_age, log_epoch_secs); 1525 } 1526