xref: /lighttpd1.4/src/stat_cache.c (revision f0786a75)
1 #include "first.h"
2 
3 #include "stat_cache.h"
4 #include "log.h"
5 #include "fdevent.h"
6 #include "http_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 
stat_cache_sptree_find(splay_tree ** const sptree,const char * const name,uint32_t len)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 #ifndef IN_EXCL_UNLINK /*(not defined in some very old glibc headers)*/
140 #define IN_EXCL_UNLINK 0x04000000
141 #endif
142 
143 /*(translate FAM API to inotify; this is specific to stat_cache.c use of FAM)*/
144 #define fam fd /*(translate struct stat_cache_fam scf->fam -> scf->fd)*/
145 typedef int FAMRequest; /*(fr)*/
146 #define FAMClose(fd) \
147         close(*(fd))
148 #define FAMCancelMonitor(fd, wd) \
149         inotify_rm_watch(*(fd), *(wd))
150 #define fam_watch_mask ( IN_ATTRIB | IN_CREATE | IN_DELETE | IN_DELETE_SELF \
151                        | IN_MODIFY | IN_MOVE_SELF | IN_MOVED_FROM \
152                        | IN_EXCL_UNLINK | IN_ONLYDIR )
153                      /*(note: follows symlinks; not providing IN_DONT_FOLLOW)*/
154 #define FAMMonitorDirectory(fd, fn, wd, userData) \
155         ((*(wd) = inotify_add_watch(*(fd), (fn), (fam_watch_mask))) < 0)
156 typedef enum FAMCodes { /*(copied from fam.h to define arbitrary enum values)*/
157     FAMChanged=1,
158     FAMDeleted=2,
159     FAMCreated=5,
160     FAMMoved=6,
161 } FAMCodes;
162 
163 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
164 #undef HAVE_SYS_INOTIFY_H
165 
166 #include <sys/event.h>
167 #include <sys/time.h>
168 
169 /*(translate FAM API to inotify; this is specific to stat_cache.c use of FAM)*/
170 #define fam fd /*(translate struct stat_cache_fam scf->fam -> scf->fd)*/
171 typedef int FAMRequest; /*(fr)*/
172 #define FAMClose(fd) \
173         (-1 != (*(fd)) ? close(*(fd)) : 0)
FAMCancelMonitor(const int * const fd,int * const wd)174 static int FAMCancelMonitor (const int * const fd, int * const wd)
175 {
176     if (-1 == *fd) return 0;
177     if (-1 == *wd) return 0;
178     struct timespec t0 = { 0, 0 };
179     struct kevent kev;
180     EV_SET(&kev, *wd, EVFILT_VNODE, EV_DELETE, 0, 0, 0);
181     int rc = kevent(*fd, &kev, 1, NULL, 0, &t0);
182     close(*wd);
183     *wd = -1;
184     return rc;
185 }
FAMMonitorDirectory(int * const fd,char * const fn,int * const wd,void * const userData)186 static int FAMMonitorDirectory (int * const fd, char * const fn, int * const wd, void * const userData)
187 {
188     *wd = fdevent_open_dirname(fn, 1); /*(note: follows symlinks)*/
189     if (-1 == *wd) return -1;
190     struct timespec t0 = { 0, 0 };
191     struct kevent kev;
192     unsigned short kev_flags = EV_ADD | EV_ENABLE | EV_CLEAR;
193     unsigned int kev_fflags = NOTE_ATTRIB | NOTE_EXTEND | NOTE_LINK | NOTE_WRITE
194                             | NOTE_DELETE | NOTE_REVOKE | NOTE_RENAME;
195     EV_SET(&kev, *wd, EVFILT_VNODE, kev_flags, kev_fflags, 0, userData);
196     return kevent(*fd, &kev, 1, NULL, 0, &t0);
197 }
198 typedef enum FAMCodes { /*(copied from fam.h to define arbitrary enum values)*/
199     FAMChanged=1,
200     FAMDeleted=2,
201     FAMCreated=5,
202     FAMMoved=6,
203 } FAMCodes;
204 
205 #else
206 
207 #include <fam.h>
208 
209 #ifdef HAVE_FAMNOEXISTS
210 #ifndef LIGHTTPD_STATIC
211 #ifdef HAVE_DLFCN_H
212 #include <dlfcn.h>
213 #endif
214 #endif
215 #endif
216 
217 #endif
218 
219 typedef struct fam_dir_entry {
220 	buffer name;
221 	int refcnt;
222 	FAMRequest req;
223 	unix_time64_t stat_ts;
224 	dev_t st_dev;
225 	ino_t st_ino;
226 	struct fam_dir_entry *fam_parent;
227 } fam_dir_entry;
228 
229 typedef struct stat_cache_fam {
230 	splay_tree *dirs; /* indexed by path; node data is fam_dir_entry */
231   #ifdef HAVE_SYS_INOTIFY_H
232 	splay_tree *wds;  /* indexed by inotify watch descriptor */
233   #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
234   #else
235 	FAMConnection fam;
236   #endif
237 	log_error_st *errh;
238 	fdevents *ev;
239 	fdnode *fdn;
240 	int fd;
241 } stat_cache_fam;
242 
243 __attribute_returns_nonnull__
fam_dir_entry_init(const char * name,size_t len)244 static fam_dir_entry * fam_dir_entry_init(const char *name, size_t len)
245 {
246     fam_dir_entry * const fam_dir = ck_calloc(1, sizeof(*fam_dir));
247     buffer_copy_string_len(&fam_dir->name, name, len);
248     fam_dir->refcnt = 0;
249   #if defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
250     fam_dir->req = -1;
251   #endif
252 
253     return fam_dir;
254 }
255 
fam_dir_entry_free(fam_dir_entry * fam_dir)256 static void fam_dir_entry_free(fam_dir_entry *fam_dir)
257 {
258     if (!fam_dir) return;
259     /*(fam_dir->fam_parent might be invalid pointer here; ignore)*/
260     free(fam_dir->name.ptr);
261   #if defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
262     if (-1 != fam_dir->req)
263         close(fam_dir->req);
264   #endif
265     free(fam_dir);
266 }
267 
fam_dir_invalidate_node(fam_dir_entry * fam_dir)268 static void fam_dir_invalidate_node(fam_dir_entry *fam_dir)
269 {
270     fam_dir->stat_ts = 0;
271     if (fam_dir->fam_parent) {
272         --fam_dir->fam_parent->refcnt;
273         fam_dir->fam_parent = NULL;
274     }
275 }
276 
277 /*
278  * walk though splay_tree and collect contents of dir tree.
279  * remove tagged entries in a second loop
280  */
281 
fam_dir_tag_refcnt(splay_tree * t,int * keys,int * ndx)282 static void fam_dir_tag_refcnt(splay_tree *t, int *keys, int *ndx)
283 {
284     if (*ndx == 512) return; /*(must match num array entries in keys[])*/
285     if (t->left)  fam_dir_tag_refcnt(t->left,  keys, ndx);
286     if (t->right) fam_dir_tag_refcnt(t->right, keys, ndx);
287     if (*ndx == 512) return; /*(must match num array entries in keys[])*/
288 
289     fam_dir_entry * const fam_dir = t->data;
290     if (0 == fam_dir->refcnt) {
291         fam_dir_invalidate_node(fam_dir);
292         keys[(*ndx)++] = t->key;
293     }
294 }
295 
296 __attribute_noinline__
fam_dir_periodic_cleanup(void)297 static void fam_dir_periodic_cleanup(void) {
298     stat_cache_fam * const scf = sc.scf;
299     int max_ndx, i;
300     int keys[512]; /* 2k size on stack */
301   #if defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
302     struct kevent kevl[512]; /* 32k size on stack to batch kevent EV_DELETE */
303   #endif
304     do {
305         if (!scf->dirs) break;
306         max_ndx = 0;
307         fam_dir_tag_refcnt(scf->dirs, keys, &max_ndx);
308         for (i = 0; i < max_ndx; ++i) {
309             const int ndx = keys[i];
310             splay_tree *node = scf->dirs = splaytree_splay(scf->dirs, ndx);
311             if (node && node->key == ndx) {
312                 fam_dir_entry *fam_dir = node->data;
313                 scf->dirs = splaytree_delete(scf->dirs, ndx);
314               #ifdef HAVE_SYS_INOTIFY_H
315                 scf->wds = splaytree_delete(scf->wds, fam_dir->req);
316               #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
317                 /* batch process kevent removal; defer cancel */
318                 EV_SET(kevl+i, fam_dir->req, EVFILT_VNODE, EV_DELETE, 0, 0, 0);
319                 fam_dir->req = -1; /*(make FAMCancelMonitor() a no-op)*/
320               #endif
321                 FAMCancelMonitor(&scf->fam, &fam_dir->req);
322                 fam_dir_entry_free(fam_dir);
323             }
324         }
325       #if defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
326         /* batch process: kevent() to submit EV_DELETE, then close dir fds */
327         if (0 == max_ndx) break;
328         struct timespec t0 = { 0, 0 };
329         kevent(scf->fd, kevl, max_ndx, NULL, 0, &t0);
330         for (i = 0; i < max_ndx; ++i)
331             close((int)kevl[i].ident);
332       #endif
333     } while (max_ndx == sizeof(keys)/sizeof(int));
334 }
335 
fam_dir_invalidate_tree(splay_tree * t,const char * name,size_t len)336 static void fam_dir_invalidate_tree(splay_tree *t, const char *name, size_t len)
337 {
338   #ifdef __clang_analyzer__
339     force_assert(name);
340   #endif
341     /*force_assert(t);*/
342     if (t->left)  fam_dir_invalidate_tree(t->left,  name, len);
343     if (t->right) fam_dir_invalidate_tree(t->right, name, len);
344 
345     fam_dir_entry * const fam_dir = t->data;
346   #ifdef __clang_analyzer__
347     force_assert(fam_dir);
348   #endif
349     const buffer * const b = &fam_dir->name;
350     size_t blen = buffer_clen(b);
351     if (blen > len && b->ptr[len] == '/' && 0 == memcmp(b->ptr, name, len))
352         fam_dir_invalidate_node(fam_dir);
353 }
354 
355 /* declarations */
356 static void stat_cache_delete_tree(const char *name, uint32_t len);
357 static void stat_cache_invalidate_dir_tree(const char *name, size_t len);
358 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);
359 
stat_cache_handle_fdevent_in(stat_cache_fam * scf)360 static void stat_cache_handle_fdevent_in(stat_cache_fam *scf)
361 {
362   #ifdef HAVE_SYS_INOTIFY_H
363     /*(inotify pads in->len to align struct following in->name[])*/
364     char buf[4096]
365       __attribute__ ((__aligned__(__alignof__(struct inotify_event))));
366     int rd;
367     do {
368         rd = (int)read(scf->fd, buf, sizeof(buf));
369         if (rd <= 0) {
370             if (-1 == rd && errno != EINTR && errno != EAGAIN) {
371                 log_perror(scf->errh, __FILE__, __LINE__, "inotify error");
372                 /* TODO: could flush cache, close scf->fd, and re-open inotify*/
373             }
374             break;
375         }
376         for (int i = 0; i < rd; ) {
377             struct inotify_event * const in =
378               (struct inotify_event *)((uintptr_t)buf + i);
379             uint32_t len = in->len;
380             if (len > sizeof(buf)) break; /*(should not happen)*/
381             i += sizeof(struct inotify_event) + len;
382             if (i > rd) break; /*(should not happen (partial record))*/
383             if (in->mask & IN_CREATE)
384                 continue; /*(see comment below for FAMCreated)*/
385             if (in->mask & IN_Q_OVERFLOW) {
386                 log_error(scf->errh, __FILE__, __LINE__,
387                           "inotify queue overflow");
388                 continue;
389             }
390             /* ignore events which may have been pending for
391              * paths recently cancelled via FAMCancelMonitor() */
392             scf->wds = splaytree_splay(scf->wds, in->wd);
393             if (!scf->wds || scf->wds->key != in->wd)
394                 continue;
395             fam_dir_entry *fam_dir = scf->wds->data;
396             if (NULL == fam_dir)        /*(should not happen)*/
397                 continue;
398             if (fam_dir->req != in->wd) /*(should not happen)*/
399                 continue;
400             /*(specific to use here in stat_cache.c)*/
401             int code = 0;
402             if (in->mask & (IN_ATTRIB | IN_MODIFY))
403                 code = FAMChanged;
404             else if (in->mask & (IN_DELETE | IN_DELETE_SELF | IN_UNMOUNT))
405                 code = FAMDeleted;
406             else if (in->mask & (IN_MOVE_SELF | IN_MOVED_FROM))
407                 code = FAMMoved;
408 
409             if (len) {
410                 do { --len; } while (len && in->name[len-1] == '\0');
411             }
412             stat_cache_handle_fdevent_fn(scf, fam_dir, in->name, len, code);
413         }
414     } while (rd + sizeof(struct inotify_event) + NAME_MAX + 1 > sizeof(buf));
415   #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
416     struct kevent kevl[256];
417     struct timespec t0 = { 0, 0 };
418     int n;
419     do {
420         n = kevent(scf->fd, NULL, 0, kevl, sizeof(kevl)/sizeof(*kevl), &t0);
421         if (n <= 0) break;
422         for (int i = 0; i < n; ++i) {
423             const struct kevent * const kev = kevl+i;
424             /* ignore events which may have been pending for
425              * paths recently cancelled via FAMCancelMonitor() */
426             int ndx = (int)(intptr_t)kev->udata;
427             scf->dirs = splaytree_splay(scf->dirs, ndx);
428             if (!scf->dirs || scf->dirs->key != ndx)
429                 continue;
430             fam_dir_entry *fam_dir = scf->dirs->data;
431             if (fam_dir->req != (int)kev->ident)
432                 continue;
433             /*(specific to use here in stat_cache.c)*/
434             /* note: stat_cache only monitors on directories,
435              *       so events here are only on directories
436              * note: changes are treated as FAMDeleted since
437              *       it is unknown which file in dir was changed
438              *       This is not efficient, but this stat_cache mechanism also
439              *       should not be used on frequently modified directories. */
440             int code = 0;
441             if (kev->fflags & (NOTE_WRITE|NOTE_ATTRIB|NOTE_EXTEND|NOTE_LINK))
442                 code = FAMDeleted; /*(not FAMChanged; see comment above)*/
443             else if (kev->fflags & (NOTE_DELETE|NOTE_REVOKE))
444                 code = FAMDeleted;
445             else if (kev->fflags & NOTE_RENAME)
446                 code = FAMMoved;
447             if (kev->flags & EV_ERROR) /*(not expected; treat as FAMDeleted)*/
448                 code = FAMDeleted;
449             stat_cache_handle_fdevent_fn(scf, fam_dir, NULL, 0, code);
450         }
451     } while (n == sizeof(kevl)/sizeof(*kevl));
452   #else
453     for (int i = 0, ndx; i || (i = FAMPending(&scf->fam)) > 0; --i) {
454         FAMEvent fe;
455         if (FAMNextEvent(&scf->fam, &fe) < 0) break;
456 
457         /* ignore events which may have been pending for
458          * paths recently cancelled via FAMCancelMonitor() */
459         ndx = (int)(intptr_t)fe.userdata;
460         scf->dirs = splaytree_splay(scf->dirs, ndx);
461         if (!scf->dirs || scf->dirs->key != ndx) {
462             continue;
463         }
464         fam_dir_entry *fam_dir = scf->dirs->data;
465         if (FAMREQUEST_GETREQNUM(&fam_dir->req)
466             != FAMREQUEST_GETREQNUM(&fe.fr)) {
467             continue;
468         }
469 
470         uint32_t fnlen = (fe.code != FAMCreated && fe.filename[0] != '/')
471           ? (uint32_t)strlen(fe.filename)
472           : 0;
473         stat_cache_handle_fdevent_fn(scf, fam_dir, fe.filename, fnlen, fe.code);
474     }
475   #endif
476 }
477 
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)478 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)
479 {
480         if (fnlen) {
481             buffer * const n = &fam_dir->name;
482             fam_dir_entry *fam_link;
483             uint32_t len;
484             switch (code) {
485             case FAMCreated:
486                 /* file created in monitored dir modifies dir and
487                  * we should get a separate FAMChanged event for dir.
488                  * Therefore, ignore file FAMCreated event here.
489                  * Also, if FAMNoExists() is used, might get spurious
490                  * FAMCreated events as changes are made e.g. in monitored
491                  * sub-sub-sub dirs and the library discovers new (already
492                  * existing) dir entries */
493                 return;
494             case FAMChanged:
495                 /* file changed in monitored dir does not modify dir */
496             case FAMDeleted:
497             case FAMMoved:
498                 /* file deleted or moved in monitored dir modifies dir,
499                  * but FAM provides separate notification for that */
500 
501                 /* temporarily append filename to dir in fam_dir->name to
502                  * construct path, then delete stat_cache entry (if any)*/
503                 len = buffer_clen(n);
504                 buffer_append_path_len(n, fn, fnlen);
505                 /* (alternatively, could chose to stat() and update)*/
506                 stat_cache_invalidate_entry(BUF_PTR_LEN(n));
507 
508                 fam_link = /*(check if might be symlink to monitored dir)*/
509                 stat_cache_sptree_find(&scf->dirs, BUF_PTR_LEN(n));
510                 if (fam_link && !buffer_is_equal(&fam_link->name, n))
511                     fam_link = NULL;
512 
513                 buffer_truncate(n, len);
514 
515                 if (fam_link) {
516                     /* replaced symlink changes containing dir */
517                     stat_cache_invalidate_entry(n->ptr, len);
518                     /* handle symlink to dir as deleted dir below */
519                     code = FAMDeleted;
520                     fam_dir = fam_link;
521                     break;
522                 }
523                 return;
524             default:
525                 return;
526             }
527         }
528 
529         switch(code) {
530         case FAMChanged:
531             stat_cache_invalidate_entry(BUF_PTR_LEN(&fam_dir->name));
532             break;
533         case FAMDeleted:
534         case FAMMoved:
535             stat_cache_delete_tree(BUF_PTR_LEN(&fam_dir->name));
536             fam_dir_invalidate_node(fam_dir);
537             if (scf->dirs)
538                 fam_dir_invalidate_tree(scf->dirs,
539                                         BUF_PTR_LEN(&fam_dir->name));
540             fam_dir_periodic_cleanup();
541             break;
542         default:
543             break;
544         }
545 }
546 
stat_cache_handle_fdevent(void * ctx,int revent)547 static handler_t stat_cache_handle_fdevent(void *ctx, int revent)
548 {
549 	stat_cache_fam * const scf = ctx; /* sc.scf */
550 
551 	if (revent & FDEVENT_IN) {
552 		stat_cache_handle_fdevent_in(scf);
553 	}
554 
555 	if (revent & (FDEVENT_HUP|FDEVENT_RDHUP)) {
556 		/* fam closed the connection */
557 		log_error(scf->errh, __FILE__, __LINE__,
558 		  "FAM connection closed; disabling stat_cache.");
559 		/* (although effectively STAT_CACHE_ENGINE_NONE,
560 		 *  do not change here so that periodic jobs clean up memory)*/
561 		/*sc.stat_cache_engine = STAT_CACHE_ENGINE_NONE; */
562 		fdevent_fdnode_event_del(scf->ev, scf->fdn);
563 		fdevent_unregister(scf->ev, scf->fdn);
564 		scf->fdn = NULL;
565 
566 		FAMClose(&scf->fam);
567 		scf->fd = -1;
568 	}
569 
570 	return HANDLER_GO_ON;
571 }
572 
stat_cache_init_fam(fdevents * ev,log_error_st * errh)573 static stat_cache_fam * stat_cache_init_fam(fdevents *ev, log_error_st *errh) {
574 	stat_cache_fam *scf = ck_calloc(1, sizeof(*scf));
575 	scf->fd = -1;
576 	scf->ev = ev;
577 	scf->errh = errh;
578 
579   #ifdef HAVE_SYS_INOTIFY_H
580    #if !defined(IN_NONBLOCK) || !defined(IN_CLOEXEC)
581 	scf->fd = inotify_init();
582 	if (scf->fd >= 0 && 0 != fdevent_fcntl_set_nb_cloexec(scf->fd)) {
583 		close(scf->fd);
584 		scf->fd = -1;
585 	}
586    #else
587 	scf->fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC);
588    #endif
589 	if (scf->fd < 0) {
590 		log_perror(errh, __FILE__, __LINE__, "inotify_init1()");
591 		free(scf);
592 		return NULL;
593 	}
594   #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
595    #ifdef __NetBSD__
596 	scf->fd = kqueue1(O_NONBLOCK|O_CLOEXEC|O_NOSIGPIPE);
597    #else
598 	scf->fd = kqueue();
599 	if (scf->fd >= 0) fdevent_setfd_cloexec(scf->fd);
600    #endif
601 	if (scf->fd < 0) {
602 		log_perror(errh, __FILE__, __LINE__, "kqueue()");
603 		free(scf);
604 		return NULL;
605 	}
606   #else
607 	/* setup FAM */
608 	if (0 != FAMOpen2(&scf->fam, "lighttpd")) {
609 		log_error(errh, __FILE__, __LINE__,
610 		  "could not open a fam connection, dying.");
611 		free(scf);
612 		return NULL;
613 	}
614       #ifdef HAVE_FAMNOEXISTS
615       #ifdef LIGHTTPD_STATIC
616 	FAMNoExists(&scf->fam);
617       #else
618 	int (*FAMNoExists_fn)(FAMConnection *);
619 	FAMNoExists_fn =
620 	  (int (*)(FAMConnection *))(intptr_t)dlsym(RTLD_DEFAULT,"FAMNoExists");
621 	if (FAMNoExists_fn) FAMNoExists_fn(&scf->fam);
622       #endif
623       #endif
624 
625 	scf->fd = FAMCONNECTION_GETFD(&scf->fam);
626 	fdevent_setfd_cloexec(scf->fd);
627   #endif
628 	scf->fdn = fdevent_register(scf->ev, scf->fd, stat_cache_handle_fdevent, scf);
629 	fdevent_fdnode_event_set(scf->ev, scf->fdn, FDEVENT_IN | FDEVENT_RDHUP);
630 
631 	return scf;
632 }
633 
stat_cache_free_fam(stat_cache_fam * scf)634 static void stat_cache_free_fam(stat_cache_fam *scf) {
635 	if (NULL == scf) return;
636 
637       #ifdef HAVE_SYS_INOTIFY_H
638 	while (scf->wds) {
639 		splay_tree *node = scf->wds;
640 		scf->wds = splaytree_delete(scf->wds, node->key);
641 	}
642       #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
643 	/*(quicker cleanup to close kqueue() before cancel per entry)*/
644 	close(scf->fd);
645 	scf->fd = -1;
646       #endif
647 	while (scf->dirs) {
648 		/*(skip entry invalidation and FAMCancelMonitor())*/
649 		splay_tree *node = scf->dirs;
650 		fam_dir_entry_free((fam_dir_entry *)node->data);
651 		scf->dirs = splaytree_delete(scf->dirs, node->key);
652 	}
653 
654 	if (-1 != scf->fd) {
655 		/*scf->fdn already cleaned up in fdevent_free()*/
656 		FAMClose(&scf->fam);
657 		/*scf->fd = -1;*/
658 	}
659 
660 	free(scf);
661 }
662 
fam_dir_monitor(stat_cache_fam * scf,char * fn,uint32_t dirlen,struct stat * st)663 static fam_dir_entry * fam_dir_monitor(stat_cache_fam *scf, char *fn, uint32_t dirlen, struct stat *st)
664 {
665     if (NULL == scf->fdn) return NULL; /* FAM connection closed; do nothing */
666     const int fn_is_dir = S_ISDIR(st->st_mode);
667     /*force_assert(0 != dirlen);*/
668     /*force_assert(fn[0] == '/');*/
669     /* consistency: ensure fn does not end in '/' unless root "/"
670      * FAM events will not end in '/', so easier to match this way */
671     if (fn[dirlen-1] == '/') --dirlen;
672     if (0 == dirlen) dirlen = 1; /* root dir ("/") */
673     /* Note: paths are expected to be normalized before calling stat_cache,
674      * e.g. without repeated '/' */
675     if (!fn_is_dir) {
676         while (fn[--dirlen] != '/') ;
677         if (0 == dirlen) dirlen = 1; /*(should not happen for file)*/
678     }
679     int dir_ndx = splaytree_djbhash(fn, dirlen);
680     fam_dir_entry *fam_dir = NULL;
681 
682     scf->dirs = splaytree_splay(scf->dirs, dir_ndx);
683     if (NULL != scf->dirs && scf->dirs->key == dir_ndx) {
684         fam_dir = scf->dirs->data;
685         if (!buffer_eq_slen(&fam_dir->name, fn, dirlen)) {
686             /* hash collision; preserve existing
687              * do not monitor new to avoid cache thrashing */
688             return NULL;
689         }
690         /* directory already registered */
691     }
692 
693     const unix_time64_t cur_ts = log_monotonic_secs;
694     struct stat lst;
695     int ck_dir = fn_is_dir;
696     if (!fn_is_dir && (NULL==fam_dir || cur_ts - fam_dir->stat_ts >= 16)) {
697         ck_dir = 1;
698         /*(temporarily modify fn)*/
699         fn[dirlen] = '\0';
700         if (0 != lstat(fn, &lst)) {
701             fn[dirlen] = '/';
702             return NULL;
703         }
704         if (!S_ISLNK(lst.st_mode)) {
705             st = &lst;
706         }
707         else if (0 != stat(fn, st)) { /*st passed in now is stat() of dir*/
708             fn[dirlen] = '/';
709             return NULL;
710         }
711         fn[dirlen] = '/';
712     }
713 
714     int ck_lnk = (NULL == fam_dir);
715     if (ck_dir && NULL != fam_dir) {
716         /* check stat() matches device and inode, just in case an external event
717          * not being monitored occurs (e.g. rename of unmonitored parent dir)*/
718         if (st->st_dev != fam_dir->st_dev || st->st_ino != fam_dir->st_ino) {
719             ck_lnk = 1;
720             /*(modifies scf->dirs but no need to re-splay for dir_ndx since
721              * fam_dir is not NULL and so splaytree_insert not called below)*/
722             if (scf->dirs) fam_dir_invalidate_tree(scf->dirs, fn, dirlen);
723             if (!fn_is_dir) /*(if dir, caller is updating stat_cache_entry)*/
724                 stat_cache_update_entry(fn, dirlen, st, NULL);
725             /*(must not delete tree since caller is holding a valid node)*/
726             stat_cache_invalidate_dir_tree(fn, dirlen);
727           #ifdef HAVE_SYS_INOTIFY_H
728             scf->wds = splaytree_delete(scf->wds, fam_dir->req);
729           #endif
730             if (0 != FAMCancelMonitor(&scf->fam, &fam_dir->req)
731                 || 0 != FAMMonitorDirectory(&scf->fam, fam_dir->name.ptr,
732                                             &fam_dir->req,
733                                             (void *)(intptr_t)dir_ndx)) {
734                 fam_dir->stat_ts = 0; /* invalidate */
735                 return NULL;
736             }
737             fam_dir->st_dev = st->st_dev;
738             fam_dir->st_ino = st->st_ino;
739           #ifdef HAVE_SYS_INOTIFY_H
740             scf->wds = splaytree_insert(scf->wds, fam_dir->req, fam_dir);
741           #endif
742         }
743         fam_dir->stat_ts = cur_ts;
744     }
745 
746     if (NULL == fam_dir) {
747         fam_dir = fam_dir_entry_init(fn, dirlen);
748 
749         if (0 != FAMMonitorDirectory(&scf->fam,fam_dir->name.ptr,&fam_dir->req,
750                                      (void *)(intptr_t)dir_ndx)) {
751           #if defined(HAVE_SYS_INOTIFY_H) \
752            || (defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE)
753             log_perror(scf->errh, __FILE__, __LINE__,
754               "monitoring dir failed: %s file: %s",
755               fam_dir->name.ptr, fn);
756           #else
757             log_error(scf->errh, __FILE__, __LINE__,
758               "monitoring dir failed: %s file: %s %s",
759               fam_dir->name.ptr, fn, FamErrlist[FAMErrno]);
760           #endif
761             fam_dir_entry_free(fam_dir);
762             return NULL;
763         }
764 
765         scf->dirs = splaytree_insert(scf->dirs, dir_ndx, fam_dir);
766       #ifdef HAVE_SYS_INOTIFY_H
767         scf->wds = splaytree_insert(scf->wds, fam_dir->req, fam_dir);
768       #endif
769         fam_dir->stat_ts= cur_ts;
770         fam_dir->st_dev = st->st_dev;
771         fam_dir->st_ino = st->st_ino;
772     }
773 
774     if (ck_lnk) {
775         if (fn_is_dir) {
776             /*(temporarily modify fn)*/
777             char e = fn[dirlen];
778             fn[dirlen] = '\0';
779             if (0 != lstat(fn, &lst)) {
780                 fn[dirlen] = e;
781                 return NULL;
782             }
783             fn[dirlen] = e;
784         }
785         if (fam_dir->fam_parent) {
786             --fam_dir->fam_parent->refcnt;
787             fam_dir->fam_parent = NULL;
788         }
789         if (S_ISLNK(lst.st_mode)) {
790             fam_dir->fam_parent = fam_dir_monitor(scf, fn, dirlen, &lst);
791         }
792     }
793 
794     ++fam_dir->refcnt;
795     return fam_dir;
796 }
797 
798 #endif
799 
800 
801 __attribute_malloc__
802 __attribute_noinline__
803 __attribute_returns_nonnull__
stat_cache_entry_init(void)804 static stat_cache_entry * stat_cache_entry_init(void) {
805     stat_cache_entry *sce = ck_calloc(1, sizeof(*sce));
806     sce->fd = -1;
807     sce->refcnt = 1;
808     return sce;
809 }
810 
stat_cache_entry_free(void * data)811 static void stat_cache_entry_free(void *data) {
812     stat_cache_entry *sce = data;
813     if (!sce) return;
814 
815     if (--sce->refcnt) return;
816 
817   #ifdef HAVE_FAM_H
818     /*(decrement refcnt only;
819      * defer cancelling FAM monitor on dir even if refcnt reaches zero)*/
820     if (sce->fam_dir) --((fam_dir_entry *)sce->fam_dir)->refcnt;
821   #endif
822 
823     free(sce->name.ptr);
824     free(sce->etag.ptr);
825     if (sce->content_type.size) free(sce->content_type.ptr);
826     if (sce->fd >= 0) close(sce->fd);
827 
828     free(sce);
829 }
830 
stat_cache_entry_refchg(void * data,int mod)831 void stat_cache_entry_refchg(void *data, int mod) {
832     /*(expect mod == -1 or mod == 1)*/
833     stat_cache_entry * const sce = data;
834     if (mod < 0 && 1 == sce->refcnt)
835         stat_cache_entry_free(data);
836     else
837         sce->refcnt += mod;
838 }
839 
840 #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR)
841 
842 static const char *attrname = "Content-Type";
843 static char attrval[128];
844 static buffer attrb = { attrval, 0, 0 };
845 
stat_cache_attr_get(const char * name)846 static int stat_cache_attr_get(const char *name) {
847   #if defined(HAVE_XATTR)
848    #if defined(HAVE_SYS_XATTR_H)
849     ssize_t attrlen;
850     if (0 < (attrlen = getxattr(name, attrname,
851                                 attrval, sizeof(attrval)-1)))
852    #else
853     int attrlen = sizeof(attrval)-1;
854     if (0 == attr_get(name, attrname, attrval, &attrlen, 0))
855    #endif
856   #elif defined(HAVE_EXTATTR)
857     ssize_t attrlen;
858     if (0 < (attrlen = extattr_get_file(name, EXTATTR_NAMESPACE_USER, attrname,
859                                         attrval, sizeof(attrval)-1)))
860   #endif
861     {
862         attrval[attrlen] = '\0';
863         attrb.used = (uint32_t)(attrlen + 1);
864         return 1;
865     }
866     return 0;
867 }
868 
869 #endif
870 
stat_cache_init(fdevents * ev,log_error_st * errh)871 int stat_cache_init(fdevents *ev, log_error_st *errh) {
872   #ifdef HAVE_FAM_H
873     if (sc.stat_cache_engine == STAT_CACHE_ENGINE_FAM) {
874         sc.scf = stat_cache_init_fam(ev, errh);
875         if (NULL == sc.scf) return 0;
876     }
877   #else
878     UNUSED(ev);
879     UNUSED(errh);
880   #endif
881 
882     return 1;
883 }
884 
stat_cache_free(void)885 void stat_cache_free(void) {
886     splay_tree *sptree = sc.files;
887     while (sptree) {
888         stat_cache_entry_free(sptree->data);
889         sptree = splaytree_delete(sptree, sptree->key);
890     }
891     sc.files = NULL;
892 
893   #ifdef HAVE_FAM_H
894     stat_cache_free_fam(sc.scf);
895     sc.scf = NULL;
896   #endif
897 
898   #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR)
899     attrname = "Content-Type";
900   #endif
901 
902     sc.stat_cache_engine = STAT_CACHE_ENGINE_SIMPLE; /*(default)*/
903 }
904 
stat_cache_xattrname(const char * name)905 void stat_cache_xattrname (const char *name) {
906   #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR)
907     attrname = name;
908   #else
909     UNUSED(name);
910   #endif
911 }
912 
stat_cache_choose_engine(const buffer * stat_cache_string,log_error_st * errh)913 int stat_cache_choose_engine (const buffer *stat_cache_string, log_error_st *errh) {
914     if (buffer_is_blank(stat_cache_string))
915         sc.stat_cache_engine = STAT_CACHE_ENGINE_SIMPLE;
916     else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("simple")))
917         sc.stat_cache_engine = STAT_CACHE_ENGINE_SIMPLE;
918 #ifdef HAVE_SYS_INOTIFY_H
919     else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("inotify")))
920         sc.stat_cache_engine = STAT_CACHE_ENGINE_INOTIFY;
921         /*(STAT_CACHE_ENGINE_FAM == STAT_CACHE_ENGINE_INOTIFY)*/
922 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
923     else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("kqueue")))
924         sc.stat_cache_engine = STAT_CACHE_ENGINE_KQUEUE;
925         /*(STAT_CACHE_ENGINE_FAM == STAT_CACHE_ENGINE_KQUEUE)*/
926 #endif
927 #ifdef HAVE_FAM_H
928     else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("fam")))
929         sc.stat_cache_engine = STAT_CACHE_ENGINE_FAM;
930 #endif
931     else if (buffer_eq_slen(stat_cache_string, CONST_STR_LEN("disable"))
932              || buffer_eq_slen(stat_cache_string, CONST_STR_LEN("none")))
933         sc.stat_cache_engine = STAT_CACHE_ENGINE_NONE;
934     else {
935         log_error(errh, __FILE__, __LINE__,
936           "server.stat-cache-engine can be one of \"disable\", \"simple\","
937 #ifdef HAVE_SYS_INOTIFY_H
938           " \"inotify\","
939 #elif defined HAVE_SYS_EVENT_H && defined HAVE_KQUEUE
940           " \"kqueue\","
941 #endif
942 #ifdef HAVE_FAM_H
943           " \"fam\","
944 #endif
945           " but not: %s", stat_cache_string->ptr);
946         return -1;
947     }
948     return 0;
949 }
950 
stat_cache_mimetype_by_ext(const array * const mimetypes,const char * const name,const uint32_t nlen)951 const buffer * stat_cache_mimetype_by_ext(const array * const mimetypes, const char * const name, const uint32_t nlen)
952 {
953     const char * const end = name + nlen; /*(end of string)*/
954     const uint32_t used = mimetypes->used;
955     if (used < 16) {
956         for (uint32_t i = 0; i < used; ++i) {
957             /* suffix match */
958             const data_string *ds = (data_string *)mimetypes->data[i];
959             const size_t klen = buffer_clen(&ds->key);
960             if (klen <= nlen && buffer_eq_icase_ssn(end-klen, ds->key.ptr, klen))
961                 return &ds->value;
962         }
963     }
964     else {
965         const char *s;
966         const data_string *ds;
967         if (nlen) {
968             for (s = end-1; s != name && *s != '/'; --s) ; /*(like memrchr())*/
969             if (*s == '/') ++s;
970         }
971         else {
972             s = name;
973         }
974         /* search for basename, then longest .ext2.ext1, then .ext1, then "" */
975         ds = (const data_string *)array_get_element_klen(mimetypes, s, end - s);
976         if (NULL != ds) return &ds->value;
977         while (++s < end) {
978             while (*s != '.' && ++s != end) ;
979             if (s == end) break;
980             /* search ".ext" then "ext" */
981             ds = (const data_string *)array_get_element_klen(mimetypes, s, end - s);
982             if (NULL != ds) return &ds->value;
983             /* repeat search without leading '.' to handle situation where
984              * admin configured mimetype.assign keys without leading '.' */
985             if (++s < end) {
986                 if (*s == '.') { --s; continue; }
987                 ds = (const data_string *)array_get_element_klen(mimetypes, s, end - s);
988                 if (NULL != ds) return &ds->value;
989             }
990         }
991         /* search for ""; catchall */
992         ds = (const data_string *)array_get_element_klen(mimetypes, CONST_STR_LEN(""));
993         if (NULL != ds) return &ds->value;
994     }
995 
996     return NULL;
997 }
998 
999 #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR)
1000 
stat_cache_mimetype_by_xattr(const char * const name)1001 const buffer * stat_cache_mimetype_by_xattr(const char * const name)
1002 {
1003     return stat_cache_attr_get(name) ? &attrb : NULL;
1004 }
1005 
stat_cache_content_type_get_by_xattr(stat_cache_entry * sce,const array * mimetypes,int use_xattr)1006 const buffer * stat_cache_content_type_get_by_xattr(stat_cache_entry *sce, const array *mimetypes, int use_xattr)
1007 {
1008     /*(invalid caching if user config has multiple, different
1009      * r->conf.mimetypes for same extension (not expected))*/
1010     if (!buffer_is_blank(&sce->content_type)) return &sce->content_type;
1011 
1012     if (!S_ISREG(sce->st.st_mode)) return NULL;
1013 
1014     /* cache mimetype */
1015     const buffer *mtype =
1016       (use_xattr) ? stat_cache_mimetype_by_xattr(sce->name.ptr) : NULL;
1017     if (NULL == mtype)
1018         mtype = stat_cache_mimetype_by_ext(mimetypes, BUF_PTR_LEN(&sce->name));
1019     if (NULL != mtype) {
1020         if (sce->content_type.size) {
1021             buffer_copy_buffer(&sce->content_type, mtype);
1022         }
1023         else if (mtype == &attrb) {
1024             sce->content_type.ptr = NULL;
1025             buffer_copy_buffer(&sce->content_type, mtype);
1026         }
1027         else {
1028             /*(copy pointers from mimetypes array; avoid allocation)*/
1029             sce->content_type.ptr = mtype->ptr;
1030             sce->content_type.used = mtype->used;
1031             /*(leave sce->content_type.size = 0 to flag not-allocated)*/
1032         }
1033     }
1034     else
1035         buffer_clear(&sce->content_type);
1036 
1037     return &sce->content_type;
1038 }
1039 
1040 #else
1041 
stat_cache_content_type_get_by_ext(stat_cache_entry * sce,const array * mimetypes)1042 const buffer * stat_cache_content_type_get_by_ext(stat_cache_entry *sce, const array *mimetypes)
1043 {
1044     /*(invalid caching if user config has multiple, different
1045      * r->conf.mimetypes for same extension (not expected))*/
1046     if (!buffer_is_blank(&sce->content_type)) return &sce->content_type;
1047 
1048     if (!S_ISREG(sce->st.st_mode)) return NULL;
1049 
1050     /* cache mimetype */
1051     const buffer * const mtype =
1052       stat_cache_mimetype_by_ext(mimetypes, BUF_PTR_LEN(&sce->name));
1053     if (NULL != mtype) {
1054         /*(copy pointers from mimetypes array; avoid allocation)*/
1055         sce->content_type.ptr = mtype->ptr;
1056         sce->content_type.used = mtype->used;
1057         /*(leave sce->content_type.size = 0 to flag not-allocated)*/
1058     }
1059     else
1060         buffer_clear(&sce->content_type);
1061 
1062     return &sce->content_type;
1063 }
1064 
1065 #endif
1066 
stat_cache_etag_get(stat_cache_entry * sce,int flags)1067 const buffer * stat_cache_etag_get(stat_cache_entry *sce, int flags) {
1068     /*(invalid caching if user cfg has multiple, different r->conf.etag_flags
1069      * for same path (not expected, since etag flags should be by filesystem))*/
1070     if (!buffer_is_blank(&sce->etag)) return &sce->etag;
1071 
1072     if (S_ISREG(sce->st.st_mode) || S_ISDIR(sce->st.st_mode)) {
1073         if (0 == flags) return NULL;
1074         http_etag_create(&sce->etag, &sce->st, flags);
1075         return &sce->etag;
1076     }
1077 
1078     return NULL;
1079 }
1080 
1081 __attribute_pure__
stat_cache_stat_eq(const struct stat * const sta,const struct stat * const stb)1082 static int stat_cache_stat_eq(const struct stat * const sta, const struct stat * const stb) {
1083     return
1084       #ifdef st_mtime /* use high-precision timestamp if available */
1085       #if defined(__APPLE__) && defined(__MACH__)
1086         sta->st_mtimespec.tv_nsec == stb->st_mtimespec.tv_nsec
1087       #else
1088         sta->st_mtim.tv_nsec == stb->st_mtim.tv_nsec
1089       #endif
1090       #else
1091         1
1092       #endif
1093         && sta->st_mtime == stb->st_mtime
1094         && sta->st_size  == stb->st_size
1095         && sta->st_ino   == stb->st_ino
1096         && sta->st_dev   == stb->st_dev;
1097 }
1098 
stat_cache_update_entry(const char * name,uint32_t len,const struct stat * st,const buffer * etagb)1099 void stat_cache_update_entry(const char *name, uint32_t len,
1100                              const struct stat *st, const buffer *etagb)
1101 {
1102     if (sc.stat_cache_engine == STAT_CACHE_ENGINE_NONE) return;
1103     force_assert(0 != len);
1104     if (name[len-1] == '/') { if (0 == --len) len = 1; }
1105     splay_tree **sptree = &sc.files;
1106     stat_cache_entry *sce =
1107       stat_cache_sptree_find(sptree, name, len);
1108     if (sce && buffer_is_equal_string(&sce->name, name, len)) {
1109         if (!stat_cache_stat_eq(&sce->st, st)) {
1110             /* etagb might be NULL to clear etag (invalidate) */
1111             buffer_clear(&sce->etag);
1112             if (etagb)
1113                 buffer_copy_string_len(&sce->etag, BUF_PTR_LEN(etagb));
1114           #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR)
1115             buffer_clear(&sce->content_type);
1116           #endif
1117             if (sce->fd >= 0) {
1118                 if (1 == sce->refcnt) {
1119                     close(sce->fd);
1120                     sce->fd = -1;
1121                 }
1122                 else {
1123                     --sce->refcnt; /* stat_cache_entry_free(sce); */
1124                     (*sptree)->data = sce = stat_cache_entry_init();
1125                     buffer_copy_string_len(&sce->name, name, len);
1126                 }
1127             }
1128             sce->st = *st;
1129         }
1130         sce->stat_ts = log_monotonic_secs;
1131     }
1132 }
1133 
stat_cache_delete_entry(const char * name,uint32_t len)1134 void stat_cache_delete_entry(const char *name, uint32_t len)
1135 {
1136     if (sc.stat_cache_engine == STAT_CACHE_ENGINE_NONE) return;
1137     force_assert(0 != len);
1138     if (name[len-1] == '/') { if (0 == --len) len = 1; }
1139     splay_tree **sptree = &sc.files;
1140     stat_cache_entry *sce = stat_cache_sptree_find(sptree, name, len);
1141     if (sce && buffer_is_equal_string(&sce->name, name, len)) {
1142         stat_cache_entry_free(sce);
1143         *sptree = splaytree_delete(*sptree, (*sptree)->key);
1144     }
1145 }
1146 
stat_cache_invalidate_entry(const char * name,uint32_t len)1147 void stat_cache_invalidate_entry(const char *name, uint32_t len)
1148 {
1149     splay_tree **sptree = &sc.files;
1150     stat_cache_entry *sce = stat_cache_sptree_find(sptree, name, len);
1151     if (sce && buffer_is_equal_string(&sce->name, name, len)) {
1152         sce->stat_ts = 0;
1153       #ifdef HAVE_FAM_H
1154         if (sce->fam_dir != NULL) {
1155             --((fam_dir_entry *)sce->fam_dir)->refcnt;
1156             sce->fam_dir = NULL;
1157         }
1158       #endif
1159     }
1160 }
1161 
1162 #ifdef HAVE_FAM_H
1163 
stat_cache_invalidate_dir_tree_walk(splay_tree * t,const char * name,size_t len)1164 static void stat_cache_invalidate_dir_tree_walk(splay_tree *t,
1165                                                 const char *name, size_t len)
1166 {
1167     if (t->left)  stat_cache_invalidate_dir_tree_walk(t->left,  name, len);
1168     if (t->right) stat_cache_invalidate_dir_tree_walk(t->right, name, len);
1169 
1170     const buffer * const b = &((stat_cache_entry *)t->data)->name;
1171     const size_t blen = buffer_clen(b);
1172     if (blen > len && b->ptr[len] == '/' && 0 == memcmp(b->ptr, name, len)) {
1173         stat_cache_entry *sce = t->data;
1174         sce->stat_ts = 0;
1175         if (sce->fam_dir != NULL) {
1176             --((fam_dir_entry *)sce->fam_dir)->refcnt;
1177             sce->fam_dir = NULL;
1178         }
1179     }
1180 }
1181 
stat_cache_invalidate_dir_tree(const char * name,size_t len)1182 static void stat_cache_invalidate_dir_tree(const char *name, size_t len)
1183 {
1184     splay_tree * const sptree = sc.files;
1185     if (sptree) stat_cache_invalidate_dir_tree_walk(sptree, name, len);
1186 }
1187 
1188 #endif
1189 
1190 /*
1191  * walk though splay_tree and collect contents of dir tree.
1192  * remove tagged entries in a second loop
1193  */
1194 
stat_cache_tag_dir_tree(splay_tree * t,const char * name,size_t len,int * keys,int * ndx)1195 static void stat_cache_tag_dir_tree(splay_tree *t, const char *name, size_t len,
1196                                     int *keys, int *ndx)
1197 {
1198     if (*ndx == 8192) return; /*(must match num array entries in keys[])*/
1199     if (t->left)  stat_cache_tag_dir_tree(t->left,  name, len, keys, ndx);
1200     if (t->right) stat_cache_tag_dir_tree(t->right, name, len, keys, ndx);
1201     if (*ndx == 8192) return; /*(must match num array entries in keys[])*/
1202 
1203     const buffer * const b = &((stat_cache_entry *)t->data)->name;
1204     const size_t blen = buffer_clen(b);
1205     if (blen > len && b->ptr[len] == '/' && 0 == memcmp(b->ptr, name, len))
1206         keys[(*ndx)++] = t->key;
1207 }
1208 
1209 __attribute_noinline__
stat_cache_prune_dir_tree(const char * name,size_t len)1210 static void stat_cache_prune_dir_tree(const char *name, size_t len)
1211 {
1212     splay_tree *sptree = sc.files;
1213     int max_ndx, i;
1214     int keys[8192]; /* 32k size on stack */
1215     do {
1216         if (!sptree) break;
1217         max_ndx = 0;
1218         stat_cache_tag_dir_tree(sptree, name, len, keys, &max_ndx);
1219         for (i = 0; i < max_ndx; ++i) {
1220             const int ndx = keys[i];
1221             splay_tree *node = sptree = splaytree_splay(sptree, ndx);
1222             if (node && node->key == ndx) {
1223                 stat_cache_entry_free(node->data);
1224                 sptree = splaytree_delete(sptree, ndx);
1225             }
1226         }
1227     } while (max_ndx == sizeof(keys)/sizeof(int));
1228     sc.files = sptree;
1229 }
1230 
stat_cache_delete_tree(const char * name,uint32_t len)1231 static void stat_cache_delete_tree(const char *name, uint32_t len)
1232 {
1233     stat_cache_delete_entry(name, len);
1234     stat_cache_prune_dir_tree(name, len);
1235 }
1236 
stat_cache_delete_dir(const char * name,uint32_t len)1237 void stat_cache_delete_dir(const char *name, uint32_t len)
1238 {
1239     force_assert(0 != len);
1240     if (name[len-1] == '/') { if (0 == --len) len = 1; }
1241     stat_cache_delete_tree(name, len);
1242   #ifdef HAVE_FAM_H
1243     if (sc.stat_cache_engine == STAT_CACHE_ENGINE_FAM) {
1244         splay_tree **sptree = &sc.scf->dirs;
1245         fam_dir_entry *fam_dir = stat_cache_sptree_find(sptree, name, len);
1246         if (fam_dir && buffer_eq_slen(&fam_dir->name, name, len))
1247             fam_dir_invalidate_node(fam_dir);
1248         if (*sptree) fam_dir_invalidate_tree(*sptree, name, len);
1249         fam_dir_periodic_cleanup();
1250     }
1251   #endif
1252 }
1253 
1254 /***
1255  *
1256  *
1257  *
1258  * returns:
1259  *  - HANDLER_FINISHED on cache-miss (don't forget to reopen the file)
1260  *  - HANDLER_ERROR on stat() failed -> see errno for problem
1261  */
1262 
stat_cache_get_entry(const buffer * const name)1263 stat_cache_entry * stat_cache_get_entry(const buffer * const name) {
1264 	stat_cache_entry *sce = NULL;
1265 
1266 	/* consistency: ensure lookup name does not end in '/' unless root "/"
1267 	 * (but use full path given with stat(), even with trailing '/') */
1268 	int final_slash = 0;
1269 	size_t len = buffer_clen(name);
1270 	force_assert(0 != len);
1271 	if (name->ptr[len-1] == '/') { final_slash = 1; if (0 == --len) len = 1; }
1272 	/* Note: paths are expected to be normalized before calling stat_cache,
1273 	 * e.g. without repeated '/' */
1274 
1275 	if (name->ptr[0] != '/') {
1276 		errno = EINVAL;
1277 		return NULL;
1278 	}
1279 
1280 	/*
1281 	 * check if the directory for this file has changed
1282 	 */
1283 
1284 	const unix_time64_t cur_ts = log_monotonic_secs;
1285 
1286 	const int file_ndx = splaytree_djbhash(name->ptr, len);
1287 	splay_tree *sptree = sc.files = splaytree_splay(sc.files, file_ndx);
1288 
1289 	if (sptree && (sptree->key == file_ndx)) {
1290 		/* we have seen this file already and
1291 		 * don't stat() it again in the same second */
1292 
1293 		sce = sptree->data;
1294 
1295 		/* check if the name is the same, we might have a collision */
1296 
1297 		if (buffer_is_equal_string(&sce->name, name->ptr, len)) {
1298 			if (sc.stat_cache_engine == STAT_CACHE_ENGINE_SIMPLE) {
1299 				if (sce->stat_ts == cur_ts) {
1300 					if (final_slash && !S_ISDIR(sce->st.st_mode)) {
1301 						errno = ENOTDIR;
1302 						return NULL;
1303 					}
1304 					return sce;
1305 				}
1306 			}
1307 		      #ifdef HAVE_FAM_H
1308 			else if (sc.stat_cache_engine == STAT_CACHE_ENGINE_FAM
1309 				 && sce->fam_dir) { /* entry is in monitored dir */
1310 				/* re-stat() periodically, even if monitoring for changes
1311 				 * (due to limitations in stat_cache.c use of FAM)
1312 				 * (gaps due to not continually monitoring an entire tree) */
1313 				if (cur_ts - sce->stat_ts < 16) {
1314 					if (final_slash && !S_ISDIR(sce->st.st_mode)) {
1315 						errno = ENOTDIR;
1316 						return NULL;
1317 					}
1318 					return sce;
1319 				}
1320 			}
1321 		      #endif
1322 		} else {
1323 			/* collision, forget about the entry */
1324 			sce = NULL;
1325 		}
1326 	}
1327 
1328 	struct stat st;
1329 	if (-1 == stat(name->ptr, &st)) {
1330 		return NULL;
1331 	}
1332 
1333 	if (NULL == sce) {
1334 
1335 		/* fix broken stat/open for symlinks to reg files with appended slash on freebsd,osx */
1336 		if (final_slash && S_ISREG(st.st_mode)) {
1337 			errno = ENOTDIR;
1338 			return NULL;
1339 		}
1340 
1341 		sce = stat_cache_entry_init();
1342 		buffer_copy_string_len(&sce->name, name->ptr, len);
1343 
1344 		/* already splayed file_ndx */
1345 		if (NULL != sptree && sptree->key == file_ndx) {
1346 			/* hash collision: replace old entry */
1347 			stat_cache_entry_free(sptree->data);
1348 			sptree->data = sce;
1349 		} else {
1350 			/*sptree =*/ sc.files = splaytree_insert(sptree, file_ndx, sce);
1351 		}
1352 
1353 	} else {
1354 
1355 		buffer_clear(&sce->etag);
1356 	      #if defined(HAVE_XATTR) || defined(HAVE_EXTATTR)
1357 		buffer_clear(&sce->content_type);
1358 	      #endif
1359 
1360 		/* close fd if file changed */
1361 		if (sce->fd >= 0 && !stat_cache_stat_eq(&sce->st, &st)) {
1362 			if (1 == sce->refcnt) {
1363 				close(sce->fd);
1364 				sce->fd = -1;
1365 			}
1366 			else {
1367 				--sce->refcnt; /* stat_cache_entry_free(sce); */
1368 				sptree->data = sce = stat_cache_entry_init();
1369 				buffer_copy_string_len(&sce->name, name->ptr, len);
1370 			}
1371 		}
1372 	}
1373 
1374 	sce->st = st; /*(copy prior to calling fam_dir_monitor())*/
1375 
1376 #ifdef HAVE_FAM_H
1377 	if (sc.stat_cache_engine == STAT_CACHE_ENGINE_FAM) {
1378 		if (sce->fam_dir) --((fam_dir_entry *)sce->fam_dir)->refcnt;
1379 		sce->fam_dir =
1380 		  fam_dir_monitor(sc.scf, name->ptr, len, &st);
1381 	      #if 0 /*(performed below)*/
1382 		if (NULL != sce->fam_dir) {
1383 			/*(may have been invalidated by dir change)*/
1384 			sce->stat_ts = cur_ts;
1385 		}
1386 	      #endif
1387 	}
1388 #endif
1389 
1390 	sce->stat_ts = cur_ts;
1391 	return sce;
1392 }
1393 
stat_cache_get_entry_open(const buffer * const name,const int symlinks)1394 stat_cache_entry * stat_cache_get_entry_open(const buffer * const name, const int symlinks) {
1395     stat_cache_entry * const sce = stat_cache_get_entry(name);
1396     if (NULL == sce) return NULL;
1397     if (sce->fd >= 0) return sce;
1398     if (sce->st.st_size > 0) {
1399         sce->fd = stat_cache_open_rdonly_fstat(name, &sce->st, symlinks);
1400         buffer_clear(&sce->etag);
1401     }
1402     return sce; /* (note: sce->fd might still be -1 if open() failed) */
1403 }
1404 
stat_cache_path_stat(const buffer * const name)1405 const stat_cache_st * stat_cache_path_stat (const buffer * const name) {
1406     const stat_cache_entry * const sce = stat_cache_get_entry(name);
1407     return sce ? &sce->st : NULL;
1408 }
1409 
stat_cache_path_isdir(const buffer * name)1410 int stat_cache_path_isdir(const buffer *name) {
1411     const stat_cache_entry * const sce = stat_cache_get_entry(name);
1412     return (sce && (S_ISDIR(sce->st.st_mode) ? 1 : (errno = ENOTDIR, 0)));
1413 }
1414 
stat_cache_path_contains_symlink(const buffer * name,log_error_st * errh)1415 int stat_cache_path_contains_symlink(const buffer *name, log_error_st *errh) {
1416     /* caller should check for symlinks only if we should block symlinks. */
1417 
1418     /* catch the obvious symlinks
1419      *
1420      * this is not a secure check as we still have a race-condition between
1421      * the stat() and the open. We can only solve this by
1422      * 1. open() the file
1423      * 2. fstat() the fd
1424      *
1425      * and keeping the file open for the rest of the time. But this can
1426      * only be done at network level.
1427      * */
1428 
1429   #ifdef HAVE_LSTAT
1430     /* we assume "/" can not be symlink,
1431      * so skip the symlink stuff if path is "/" */
1432     size_t len = buffer_clen(name);
1433     force_assert(0 != len);
1434     force_assert(name->ptr[0] == '/');
1435     if (1 == len) return 0;
1436    #ifndef PATH_MAX
1437    #define PATH_MAX 4096
1438    #endif
1439     if (len >= PATH_MAX) return -1;
1440 
1441     char buf[PATH_MAX];
1442     memcpy(buf, name->ptr, len);
1443     char *s_cur = buf+len;
1444     do {
1445         *s_cur = '\0';
1446         struct stat st;
1447         if (0 == lstat(buf, &st)) {
1448             if (S_ISLNK(st.st_mode)) return 1;
1449         }
1450         else {
1451             log_perror(errh, __FILE__, __LINE__, "lstat failed for: %s", buf);
1452             return -1;
1453         }
1454     } while ((s_cur = strrchr(buf, '/')) > buf); /*(&buf[0]==buf; NULL < buf)*/
1455   #endif
1456 
1457     return 0;
1458 }
1459 
stat_cache_open_rdonly_fstat(const buffer * name,struct stat * st,int symlinks)1460 int stat_cache_open_rdonly_fstat (const buffer *name, struct stat *st, int symlinks) {
1461 	/*(Note: O_NOFOLLOW affects only the final path segment, the target file,
1462 	 * not any intermediate symlinks along the path)*/
1463 	const int fd = fdevent_open_cloexec(name->ptr, symlinks, O_RDONLY, 0);
1464 	if (fd >= 0) {
1465 		if (0 == fstat(fd, st)) {
1466 			return fd;
1467 		} else {
1468 			const int errnum = errno;
1469 			close(fd);
1470 			errno = errnum;
1471 		}
1472 	}
1473 	return -1;
1474 }
1475 
1476 /**
1477  * remove stat() from cache which haven't been stat()ed for
1478  * more than 2 seconds
1479  *
1480  *
1481  * walk though the stat-cache, collect the ids which are too old
1482  * and remove them in a second loop
1483  */
1484 
stat_cache_tag_old_entries(splay_tree * const t,int * const keys,int * const ndx,const time_t max_age,const unix_time64_t cur_ts)1485 static void stat_cache_tag_old_entries(splay_tree * const t, int * const keys, int * const ndx, const time_t max_age, const unix_time64_t cur_ts) {
1486     if (*ndx == 8192) return; /*(must match num array entries in keys[])*/
1487     if (t->left)
1488         stat_cache_tag_old_entries(t->left, keys, ndx, max_age, cur_ts);
1489     if (t->right)
1490         stat_cache_tag_old_entries(t->right, keys, ndx, max_age, cur_ts);
1491     if (*ndx == 8192) return; /*(must match num array entries in keys[])*/
1492 
1493     const stat_cache_entry * const sce = t->data;
1494     if (cur_ts - sce->stat_ts > max_age)
1495         keys[(*ndx)++] = t->key;
1496 }
1497 
stat_cache_periodic_cleanup(const time_t max_age,const unix_time64_t cur_ts)1498 static void stat_cache_periodic_cleanup(const time_t max_age, const unix_time64_t cur_ts) {
1499     splay_tree *sptree = sc.files;
1500     int max_ndx, i;
1501     int keys[8192]; /* 32k size on stack */
1502     do {
1503         if (!sptree) break;
1504         max_ndx = 0;
1505         stat_cache_tag_old_entries(sptree, keys, &max_ndx, max_age, cur_ts);
1506         for (i = 0; i < max_ndx; ++i) {
1507             int ndx = keys[i];
1508             sptree = splaytree_splay(sptree, ndx);
1509             if (sptree && sptree->key == ndx) {
1510                 stat_cache_entry_free(sptree->data);
1511                 sptree = splaytree_delete(sptree, ndx);
1512             }
1513         }
1514     } while (max_ndx == sizeof(keys)/sizeof(int));
1515     sc.files = sptree;
1516 }
1517 
stat_cache_trigger_cleanup(void)1518 void stat_cache_trigger_cleanup(void) {
1519 	time_t max_age = 2;
1520 
1521       #ifdef HAVE_FAM_H
1522 	if (STAT_CACHE_ENGINE_FAM == sc.stat_cache_engine) {
1523 		if (log_monotonic_secs & 0x1F) return;
1524 		/* once every 32 seconds (0x1F == 31) */
1525 		max_age = 32;
1526 		fam_dir_periodic_cleanup();
1527 		/* By doing this before stat_cache_periodic_cleanup(),
1528 		 * entries used within the next max_age secs will remain
1529 		 * monitored, instead of effectively flushing and
1530 		 * rebuilding the FAM monitoring every max_age seconds */
1531 	}
1532       #endif
1533 
1534 	stat_cache_periodic_cleanup(max_age, log_monotonic_secs);
1535 }
1536