| Nick Piggin | 31e6b01 | 2011-01-07 17:49:52 +1100 | [diff] [blame] | 1 | Path walking and name lookup locking | 
 | 2 | ==================================== | 
 | 3 |  | 
 | 4 | Path resolution is the finding a dentry corresponding to a path name string, by | 
 | 5 | performing a path walk. Typically, for every open(), stat() etc., the path name | 
 | 6 | will be resolved. Paths are resolved by walking the namespace tree, starting | 
 | 7 | with the first component of the pathname (eg. root or cwd) with a known dentry, | 
 | 8 | then finding the child of that dentry, which is named the next component in the | 
 | 9 | path string. Then repeating the lookup from the child dentry and finding its | 
 | 10 | child with the next element, and so on. | 
 | 11 |  | 
 | 12 | Since it is a frequent operation for workloads like multiuser environments and | 
 | 13 | web servers, it is important to optimize this code. | 
 | 14 |  | 
 | 15 | Path walking synchronisation history: | 
 | 16 | Prior to 2.5.10, dcache_lock was acquired in d_lookup (dcache hash lookup) and | 
 | 17 | thus in every component during path look-up. Since 2.5.10 onwards, fast-walk | 
 | 18 | algorithm changed this by holding the dcache_lock at the beginning and walking | 
 | 19 | as many cached path component dentries as possible. This significantly | 
 | 20 | decreases the number of acquisition of dcache_lock. However it also increases | 
 | 21 | the lock hold time significantly and affects performance in large SMP machines. | 
 | 22 | Since 2.5.62 kernel, dcache has been using a new locking model that uses RCU to | 
 | 23 | make dcache look-up lock-free. | 
 | 24 |  | 
 | 25 | All the above algorithms required taking a lock and reference count on the | 
 | 26 | dentry that was looked up, so that may be used as the basis for walking the | 
 | 27 | next path element. This is inefficient and unscalable. It is inefficient | 
 | 28 | because of the locks and atomic operations required for every dentry element | 
 | 29 | slows things down. It is not scalable because many parallel applications that | 
 | 30 | are path-walk intensive tend to do path lookups starting from a common dentry | 
 | 31 | (usually, the root "/" or current working directory). So contention on these | 
 | 32 | common path elements causes lock and cacheline queueing. | 
 | 33 |  | 
 | 34 | Since 2.6.38, RCU is used to make a significant part of the entire path walk | 
 | 35 | (including dcache look-up) completely "store-free" (so, no locks, atomics, or | 
 | 36 | even stores into cachelines of common dentries). This is known as "rcu-walk" | 
 | 37 | path walking. | 
 | 38 |  | 
 | 39 | Path walking overview | 
 | 40 | ===================== | 
 | 41 |  | 
 | 42 | A name string specifies a start (root directory, cwd, fd-relative) and a | 
 | 43 | sequence of elements (directory entry names), which together refer to a path in | 
 | 44 | the namespace. A path is represented as a (dentry, vfsmount) tuple. The name | 
| Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 45 | elements are sub-strings, separated by '/'. | 
| Nick Piggin | 31e6b01 | 2011-01-07 17:49:52 +1100 | [diff] [blame] | 46 |  | 
 | 47 | Name lookups will want to find a particular path that a name string refers to | 
 | 48 | (usually the final element, or parent of final element). This is done by taking | 
 | 49 | the path given by the name's starting point (which we know in advance -- eg. | 
 | 50 | current->fs->cwd or current->fs->root) as the first parent of the lookup. Then | 
 | 51 | iteratively for each subsequent name element, look up the child of the current | 
 | 52 | parent with the given name and if it is not the desired entry, make it the | 
 | 53 | parent for the next lookup. | 
 | 54 |  | 
 | 55 | A parent, of course, must be a directory, and we must have appropriate | 
 | 56 | permissions on the parent inode to be able to walk into it. | 
 | 57 |  | 
 | 58 | Turning the child into a parent for the next lookup requires more checks and | 
 | 59 | procedures. Symlinks essentially substitute the symlink name for the target | 
 | 60 | name in the name string, and require some recursive path walking.  Mount points | 
 | 61 | must be followed into (thus changing the vfsmount that subsequent path elements | 
 | 62 | refer to), switching from the mount point path to the root of the particular | 
 | 63 | mounted vfsmount. These behaviours are variously modified depending on the | 
 | 64 | exact path walking flags. | 
 | 65 |  | 
 | 66 | Path walking then must, broadly, do several particular things: | 
 | 67 | - find the start point of the walk; | 
 | 68 | - perform permissions and validity checks on inodes; | 
 | 69 | - perform dcache hash name lookups on (parent, name element) tuples; | 
 | 70 | - traverse mount points; | 
 | 71 | - traverse symlinks; | 
 | 72 | - lookup and create missing parts of the path on demand. | 
 | 73 |  | 
 | 74 | Safe store-free look-up of dcache hash table | 
 | 75 | ============================================ | 
 | 76 |  | 
 | 77 | Dcache name lookup | 
 | 78 | ------------------ | 
 | 79 | In order to lookup a dcache (parent, name) tuple, we take a hash on the tuple | 
 | 80 | and use that to select a bucket in the dcache-hash table. The list of entries | 
 | 81 | in that bucket is then walked, and we do a full comparison of each entry | 
 | 82 | against our (parent, name) tuple. | 
 | 83 |  | 
 | 84 | The hash lists are RCU protected, so list walking is not serialised with | 
 | 85 | concurrent updates (insertion, deletion from the hash). This is a standard RCU | 
 | 86 | list application with the exception of renames, which will be covered below. | 
 | 87 |  | 
 | 88 | Parent and name members of a dentry, as well as its membership in the dcache | 
 | 89 | hash, and its inode are protected by the per-dentry d_lock spinlock. A | 
 | 90 | reference is taken on the dentry (while the fields are verified under d_lock), | 
 | 91 | and this stabilises its d_inode pointer and actual inode. This gives a stable | 
 | 92 | point to perform the next step of our path walk against. | 
 | 93 |  | 
 | 94 | These members are also protected by d_seq seqlock, although this offers | 
 | 95 | read-only protection and no durability of results, so care must be taken when | 
 | 96 | using d_seq for synchronisation (see seqcount based lookups, below). | 
 | 97 |  | 
 | 98 | Renames | 
 | 99 | ------- | 
 | 100 | Back to the rename case. In usual RCU protected lists, the only operations that | 
 | 101 | will happen to an object is insertion, and then eventually removal from the | 
 | 102 | list. The object will not be reused until an RCU grace period is complete. | 
 | 103 | This ensures the RCU list traversal primitives can run over the object without | 
 | 104 | problems (see RCU documentation for how this works). | 
 | 105 |  | 
 | 106 | However when a dentry is renamed, its hash value can change, requiring it to be | 
 | 107 | moved to a new hash list. Allocating and inserting a new alias would be | 
 | 108 | expensive and also problematic for directory dentries. Latency would be far to | 
 | 109 | high to wait for a grace period after removing the dentry and before inserting | 
 | 110 | it in the new hash bucket. So what is done is to insert the dentry into the | 
 | 111 | new list immediately. | 
 | 112 |  | 
 | 113 | However, when the dentry's list pointers are updated to point to objects in the | 
 | 114 | new list before waiting for a grace period, this can result in a concurrent RCU | 
 | 115 | lookup of the old list veering off into the new (incorrect) list and missing | 
 | 116 | the remaining dentries on the list. | 
 | 117 |  | 
 | 118 | There is no fundamental problem with walking down the wrong list, because the | 
 | 119 | dentry comparisons will never match. However it is fatal to miss a matching | 
 | 120 | dentry. So a seqlock is used to detect when a rename has occurred, and so the | 
 | 121 | lookup can be retried. | 
 | 122 |  | 
 | 123 |          1      2      3 | 
 | 124 |         +---+  +---+  +---+ | 
 | 125 | hlist-->| N-+->| N-+->| N-+-> | 
 | 126 | head <--+-P |<-+-P |<-+-P | | 
 | 127 |         +---+  +---+  +---+ | 
 | 128 |  | 
 | 129 | Rename of dentry 2 may require it deleted from the above list, and inserted | 
 | 130 | into a new list. Deleting 2 gives the following list. | 
 | 131 |  | 
 | 132 |          1             3 | 
 | 133 |         +---+         +---+     (don't worry, the longer pointers do not | 
 | 134 | hlist-->| N-+-------->| N-+->    impose a measurable performance overhead | 
 | 135 | head <--+-P |<--------+-P |      on modern CPUs) | 
 | 136 |         +---+         +---+ | 
 | 137 |           ^      2      ^ | 
 | 138 |           |    +---+    | | 
 | 139 |           |    | N-+----+ | 
 | 140 |           +----+-P | | 
 | 141 |                +---+ | 
 | 142 |  | 
 | 143 | This is a standard RCU-list deletion, which leaves the deleted object's | 
 | 144 | pointers intact, so a concurrent list walker that is currently looking at | 
 | 145 | object 2 will correctly continue to object 3 when it is time to traverse the | 
 | 146 | next object. | 
 | 147 |  | 
 | 148 | However, when inserting object 2 onto a new list, we end up with this: | 
 | 149 |  | 
 | 150 |          1             3 | 
 | 151 |         +---+         +---+ | 
 | 152 | hlist-->| N-+-------->| N-+-> | 
 | 153 | head <--+-P |<--------+-P | | 
 | 154 |         +---+         +---+ | 
 | 155 |                  2 | 
 | 156 |                +---+ | 
 | 157 |                | N-+----> | 
 | 158 |           <----+-P | | 
 | 159 |                +---+ | 
 | 160 |  | 
 | 161 | Because we didn't wait for a grace period, there may be a concurrent lookup | 
 | 162 | still at 2. Now when it follows 2's 'next' pointer, it will walk off into | 
 | 163 | another list without ever having checked object 3. | 
 | 164 |  | 
 | 165 | A related, but distinctly different, issue is that of rename atomicity versus | 
 | 166 | lookup operations. If a file is renamed from 'A' to 'B', a lookup must only | 
 | 167 | find either 'A' or 'B'. So if a lookup of 'A' returns NULL, a subsequent lookup | 
 | 168 | of 'B' must succeed (note the reverse is not true). | 
 | 169 |  | 
 | 170 | Between deleting the dentry from the old hash list, and inserting it on the new | 
 | 171 | hash list, a lookup may find neither 'A' nor 'B' matching the dentry. The same | 
 | 172 | rename seqlock is also used to cover this race in much the same way, by | 
 | 173 | retrying a negative lookup result if a rename was in progress. | 
 | 174 |  | 
 | 175 | Seqcount based lookups | 
 | 176 | ---------------------- | 
 | 177 | In refcount based dcache lookups, d_lock is used to serialise access to | 
 | 178 | the dentry, stabilising it while comparing its name and parent and then | 
 | 179 | taking a reference count (the reference count then gives a stable place to | 
 | 180 | start the next part of the path walk from). | 
 | 181 |  | 
 | 182 | As explained above, we would like to do path walking without taking locks or | 
 | 183 | reference counts on intermediate dentries along the path. To do this, a per | 
 | 184 | dentry seqlock (d_seq) is used to take a "coherent snapshot" of what the dentry | 
 | 185 | looks like (its name, parent, and inode). That snapshot is then used to start | 
 | 186 | the next part of the path walk. When loading the coherent snapshot under d_seq, | 
 | 187 | care must be taken to load the members up-front, and use those pointers rather | 
 | 188 | than reloading from the dentry later on (otherwise we'd have interesting things | 
 | 189 | like d_inode going NULL underneath us, if the name was unlinked). | 
 | 190 |  | 
 | 191 | Also important is to avoid performing any destructive operations (pretty much: | 
 | 192 | no non-atomic stores to shared data), and to recheck the seqcount when we are | 
 | 193 | "done" with the operation. Retry or abort if the seqcount does not match. | 
 | 194 | Avoiding destructive or changing operations means we can easily unwind from | 
 | 195 | failure. | 
 | 196 |  | 
 | 197 | What this means is that a caller, provided they are holding RCU lock to | 
 | 198 | protect the dentry object from disappearing, can perform a seqcount based | 
 | 199 | lookup which does not increment the refcount on the dentry or write to | 
 | 200 | it in any way. This returned dentry can be used for subsequent operations, | 
 | 201 | provided that d_seq is rechecked after that operation is complete. | 
 | 202 |  | 
 | 203 | Inodes are also rcu freed, so the seqcount lookup dentry's inode may also be | 
 | 204 | queried for permissions. | 
 | 205 |  | 
 | 206 | With this two parts of the puzzle, we can do path lookups without taking | 
 | 207 | locks or refcounts on dentry elements. | 
 | 208 |  | 
 | 209 | RCU-walk path walking design | 
 | 210 | ============================ | 
 | 211 |  | 
 | 212 | Path walking code now has two distinct modes, ref-walk and rcu-walk. ref-walk | 
 | 213 | is the traditional[*] way of performing dcache lookups using d_lock to | 
 | 214 | serialise concurrent modifications to the dentry and take a reference count on | 
 | 215 | it. ref-walk is simple and obvious, and may sleep, take locks, etc while path | 
 | 216 | walking is operating on each dentry. rcu-walk uses seqcount based dentry | 
 | 217 | lookups, and can perform lookup of intermediate elements without any stores to | 
 | 218 | shared data in the dentry or inode. rcu-walk can not be applied to all cases, | 
 | 219 | eg. if the filesystem must sleep or perform non trivial operations, rcu-walk | 
 | 220 | must be switched to ref-walk mode. | 
 | 221 |  | 
 | 222 | [*] RCU is still used for the dentry hash lookup in ref-walk, but not the full | 
 | 223 |     path walk. | 
 | 224 |  | 
 | 225 | Where ref-walk uses a stable, refcounted ``parent'' to walk the remaining | 
 | 226 | path string, rcu-walk uses a d_seq protected snapshot. When looking up a | 
 | 227 | child of this parent snapshot, we open d_seq critical section on the child | 
 | 228 | before closing d_seq critical section on the parent. This gives an interlocking | 
 | 229 | ladder of snapshots to walk down. | 
 | 230 |  | 
 | 231 |  | 
 | 232 |      proc 101 | 
 | 233 |       /----------------\ | 
 | 234 |      / comm:    "vi"    \ | 
 | 235 |     /  fs.root: dentry0  \ | 
 | 236 |     \  fs.cwd:  dentry2  / | 
 | 237 |      \                  / | 
 | 238 |       \----------------/ | 
 | 239 |  | 
 | 240 | So when vi wants to open("/home/npiggin/test.c", O_RDWR), then it will | 
 | 241 | start from current->fs->root, which is a pinned dentry. Alternatively, | 
 | 242 | "./test.c" would start from cwd; both names refer to the same path in | 
 | 243 | the context of proc101. | 
 | 244 |  | 
 | 245 |      dentry 0 | 
 | 246 |     +---------------------+   rcu-walk begins here, we note d_seq, check the | 
 | 247 |     | name:    "/"        |   inode's permission, and then look up the next | 
 | 248 |     | inode:   10         |   path element which is "home"... | 
 | 249 |     | children:"home", ...| | 
 | 250 |     +---------------------+ | 
 | 251 |               | | 
 | 252 |      dentry 1 V | 
 | 253 |     +---------------------+   ... which brings us here. We find dentry1 via | 
 | 254 |     | name:    "home"     |   hash lookup, then note d_seq and compare name | 
 | 255 |     | inode:   678        |   string and parent pointer. When we have a match, | 
 | 256 |     | children:"npiggin"  |   we now recheck the d_seq of dentry0. Then we | 
 | 257 |     +---------------------+   check inode and look up the next element. | 
 | 258 |               | | 
 | 259 |      dentry2  V | 
 | 260 |     +---------------------+   Note: if dentry0 is now modified, lookup is | 
 | 261 |     | name:    "npiggin"  |   not necessarily invalid, so we need only keep a | 
 | 262 |     | inode:   543        |   parent for d_seq verification, and grandparents | 
 | 263 |     | children:"a.c", ... |   can be forgotten. | 
 | 264 |     +---------------------+ | 
 | 265 |               | | 
 | 266 |      dentry3  V | 
 | 267 |     +---------------------+   At this point we have our destination dentry. | 
 | 268 |     | name:    "a.c"      |   We now take its d_lock, verify d_seq of this | 
 | 269 |     | inode:   14221      |   dentry. If that checks out, we can increment | 
 | 270 |     | children:NULL       |   its refcount because we're holding d_lock. | 
 | 271 |     +---------------------+ | 
 | 272 |  | 
 | 273 | Taking a refcount on a dentry from rcu-walk mode, by taking its d_lock, | 
 | 274 | re-checking its d_seq, and then incrementing its refcount is called | 
 | 275 | "dropping rcu" or dropping from rcu-walk into ref-walk mode. | 
 | 276 |  | 
 | 277 | It is, in some sense, a bit of a house of cards. If the seqcount check of the | 
 | 278 | parent snapshot fails, the house comes down, because we had closed the d_seq | 
 | 279 | section on the grandparent, so we have nothing left to stand on. In that case, | 
 | 280 | the path walk must be fully restarted (which we do in ref-walk mode, to avoid | 
 | 281 | live locks). It is costly to have a full restart, but fortunately they are | 
 | 282 | quite rare. | 
 | 283 |  | 
 | 284 | When we reach a point where sleeping is required, or a filesystem callout | 
 | 285 | requires ref-walk, then instead of restarting the walk, we attempt to drop rcu | 
 | 286 | at the last known good dentry we have. Avoiding a full restart in ref-walk in | 
 | 287 | these cases is fundamental for performance and scalability because blocking | 
 | 288 | operations such as creates and unlinks are not uncommon. | 
 | 289 |  | 
 | 290 | The detailed design for rcu-walk is like this: | 
 | 291 | * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. | 
 | 292 | * Take the RCU lock for the entire path walk, starting with the acquiring | 
 | 293 |   of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are | 
 | 294 |   not required for dentry persistence. | 
 | 295 | * synchronize_rcu is called when unregistering a filesystem, so we can | 
 | 296 |   access d_ops and i_ops during rcu-walk. | 
 | 297 | * Similarly take the vfsmount lock for the entire path walk. So now mnt | 
 | 298 |   refcounts are not required for persistence. Also we are free to perform mount | 
 | 299 |   lookups, and to assume dentry mount points and mount roots are stable up and | 
 | 300 |   down the path. | 
 | 301 | * Have a per-dentry seqlock to protect the dentry name, parent, and inode, | 
 | 302 |   so we can load this tuple atomically, and also check whether any of its | 
 | 303 |   members have changed. | 
 | 304 | * Dentry lookups (based on parent, candidate string tuple) recheck the parent | 
 | 305 |   sequence after the child is found in case anything changed in the parent | 
 | 306 |   during the path walk. | 
 | 307 | * inode is also RCU protected so we can load d_inode and use the inode for | 
 | 308 |   limited things. | 
 | 309 | * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. | 
 | 310 | * i_op can be loaded. | 
 | 311 | * When the destination dentry is reached, drop rcu there (ie. take d_lock, | 
 | 312 |   verify d_seq, increment refcount). | 
 | 313 | * If seqlock verification fails anywhere along the path, do a full restart | 
 | 314 |   of the path lookup in ref-walk mode. -ECHILD tends to be used (for want of | 
 | 315 |   a better errno) to signal an rcu-walk failure. | 
 | 316 |  | 
 | 317 | The cases where rcu-walk cannot continue are: | 
 | 318 | * NULL dentry (ie. any uncached path element) | 
| Nick Piggin | 31e6b01 | 2011-01-07 17:49:52 +1100 | [diff] [blame] | 319 | * Following links | 
 | 320 |  | 
| Nick Piggin | b74c79e | 2011-01-07 17:49:58 +1100 | [diff] [blame] | 321 | It may be possible eventually to make following links rcu-walk aware. | 
| Nick Piggin | 31e6b01 | 2011-01-07 17:49:52 +1100 | [diff] [blame] | 322 |  | 
 | 323 | Uncached path elements will always require dropping to ref-walk mode, at the | 
 | 324 | very least because i_mutex needs to be grabbed, and objects allocated. | 
 | 325 |  | 
 | 326 | Final note: | 
 | 327 | "store-free" path walking is not strictly store free. We take vfsmount lock | 
 | 328 | and refcounts (both of which can be made per-cpu), and we also store to the | 
 | 329 | stack (which is essentially CPU-local), and we also have to take locks and | 
 | 330 | refcount on final dentry. | 
 | 331 |  | 
 | 332 | The point is that shared data, where practically possible, is not locked | 
 | 333 | or stored into. The result is massive improvements in performance and | 
 | 334 | scalability of path resolution. | 
 | 335 |  | 
 | 336 |  | 
| Nick Piggin | b74c79e | 2011-01-07 17:49:58 +1100 | [diff] [blame] | 337 | Interesting statistics | 
 | 338 | ====================== | 
 | 339 |  | 
 | 340 | The following table gives rcu lookup statistics for a few simple workloads | 
 | 341 | (2s12c24t Westmere, debian non-graphical system). Ungraceful are attempts to | 
 | 342 | drop rcu that fail due to d_seq failure and requiring the entire path lookup | 
 | 343 | again. Other cases are successful rcu-drops that are required before the final | 
 | 344 | element, nodentry for missing dentry, revalidate for filesystem revalidate | 
 | 345 | routine requiring rcu drop, permission for permission check requiring drop, | 
 | 346 | and link for symlink traversal requiring drop. | 
 | 347 |  | 
 | 348 |      rcu-lookups     restart  nodentry          link  revalidate  permission | 
 | 349 | bootup     47121           0      4624          1010       10283        7852 | 
 | 350 | dbench  25386793           0   6778659(26.7%)     55         549        1156 | 
 | 351 | kbuild   2696672          10     64442(2.3%)  108764(4.0%)     1        1590 | 
 | 352 | git diff   39605           0        28             2           0         106 | 
 | 353 | vfstest 24185492        4945    708725(2.9%) 1076136(4.4%)     0        2651 | 
 | 354 |  | 
 | 355 | What this shows is that failed rcu-walk lookups, ie. ones that are restarted | 
 | 356 | entirely with ref-walk, are quite rare. Even the "vfstest" case which | 
| Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 357 | specifically has concurrent renames/mkdir/rmdir/ creat/unlink/etc to exercise | 
| Nick Piggin | b74c79e | 2011-01-07 17:49:58 +1100 | [diff] [blame] | 358 | such races is not showing a huge amount of restarts. | 
 | 359 |  | 
 | 360 | Dropping from rcu-walk to ref-walk mean that we have encountered a dentry where | 
 | 361 | the reference count needs to be taken for some reason. This is either because | 
 | 362 | we have reached the target of the path walk, or because we have encountered a | 
 | 363 | condition that can't be resolved in rcu-walk mode.  Ideally, we drop rcu-walk | 
 | 364 | only when we have reached the target dentry, so the other statistics show where | 
 | 365 | this does not happen. | 
 | 366 |  | 
 | 367 | Note that a graceful drop from rcu-walk mode due to something such as the | 
 | 368 | dentry not existing (which can be common) is not necessarily a failure of | 
 | 369 | rcu-walk scheme, because some elements of the path may have been walked in | 
 | 370 | rcu-walk mode. The further we get from common path elements (such as cwd or | 
 | 371 | root), the less contended the dentry is likely to be. The closer we are to | 
 | 372 | common path elements, the more likely they will exist in dentry cache. | 
 | 373 |  | 
 | 374 |  | 
| Nick Piggin | 31e6b01 | 2011-01-07 17:49:52 +1100 | [diff] [blame] | 375 | Papers and other documentation on dcache locking | 
 | 376 | ================================================ | 
 | 377 |  | 
 | 378 | 1. Scaling dcache with RCU (http://linuxjournal.com/article.php?sid=7124). | 
 | 379 |  | 
 | 380 | 2. http://lse.sourceforge.net/locking/dcache/dcache.html | 
| Nick Piggin | b74c79e | 2011-01-07 17:49:58 +1100 | [diff] [blame] | 381 |  | 
 | 382 |  |