Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition) |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
[ source navigation ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
|
Name | Size | Last modified (GMT) | Description | |
Parent directory | 2023-01-29 20:54:29 | |||
README | 15059 bytes | 2023-01-29 20:54:29 | ||
README.softupdates | 717 bytes | 2023-01-29 20:54:29 | ||
ffs_alloc.c | 46742 bytes | 2023-01-29 20:54:29 | ||
ffs_balloc.c | 10706 bytes | 2023-01-29 20:54:29 | ||
ffs_extern.h | 5815 bytes | 2023-01-29 20:54:29 | ||
ffs_inode.c | 15422 bytes | 2023-01-29 20:54:29 | ||
ffs_softdep.c | 134453 bytes | 2023-01-29 20:54:29 | ||
ffs_softdep_stub.c | 5621 bytes | 2023-01-29 20:54:29 | ||
ffs_subr.c | 6381 bytes | 2023-01-29 20:54:29 | ||
ffs_tables.c | 5873 bytes | 2023-01-29 20:54:29 | ||
ffs_vfsops.c | 34084 bytes | 2023-01-29 20:54:29 | ||
ffs_vnops.c | 7679 bytes | 2023-01-29 20:54:29 | ||
fs.h | 21986 bytes | 2023-01-29 20:54:29 | ||
softdep.h | 27527 bytes | 2023-01-29 20:54:29 |
1 Introduction 2 3 This package constitutes the alpha distribution of the soft update 4 code updates for the fast filesystem. 5 6 For More information on what Soft Updates is, see: 7 http://www.ece.cmu.edu/~ganger/papers/CSE-TR-254-95/ 8 9 Status 10 11 My `filesystem torture tests' (described below) run for days without 12 a hitch (no panic's, hangs, filesystem corruption, or memory leaks). 13 However, I have had several panic's reported to me by folks that 14 are field testing the code which I have not yet been able to 15 reproduce or fix. Although these panic's are rare and do not cause 16 filesystem corruption, the code should only be put into production 17 on systems where the system administrator is aware that it is being 18 run, and knows how to turn it off if problems arise. Thus, you may 19 hand out this code to others, but please ensure that this status 20 message is included with any distributions. Please also include 21 the file ffs_softdep.stub.c in any distributions so that folks that 22 cannot abide by the need to redistribute source will not be left 23 with a kernel that will not link. It will resolve all the calls 24 into the soft update code and simply ignores the request to enable 25 them. Thus you will be able to ensure that your other hooks have 26 not broken anything and that your kernel is softdep-ready for those 27 that wish to use them. Please report problems back to me with 28 kernel backtraces of panics if possible. This is massively complex 29 code, and people only have to have their filesystems hosed once or 30 twice to avoid future changes like the plague. I want to find and 31 fix as many bugs as soon as possible so as to get the code rock 32 solid before it gets widely released. Please report any bugs that 33 you uncover to mckusick@mckusick.com. 34 35 Performance 36 37 Running the Andrew Benchmarks yields the following raw data: 38 39 Phase Normal Softdep What it does 40 1 3s <1s Creating directories 41 2 8s 4s Copying files 42 3 6s 6s Recursive directory stats 43 4 8s 9s Scanning each file 44 5 25s 25s Compilation 45 46 Normal: 19.9u 29.2s 0:52.8 135+630io 47 Softdep: 20.3u 28.5s 0:47.8 103+363io 48 49 Another interesting datapoint are my `filesystem torture tests'. 50 They consist of 1000 runs of the andrew benchmarks, 1000 copy and 51 removes of /etc with randomly selected pauses of 0-60 seconds 52 between each copy and remove, and 500 find from / with randomly 53 selected pauses of 100 seconds between each run). The run of the 54 torture test compares as follows: 55 56 With soft updates: writes: 6 sync, 1,113,686 async; run time 19hr, 50min 57 Normal filesystem: writes: 1,459,147 sync, 487,031 async; run time 27hr, 15min 58 59 The upshot is 42% less I/O and 28% shorter running time. 60 61 Another interesting test point is a full MAKEDEV. Because it runs 62 as a shell script, it becomes mostly limited by the execution speed 63 of the machine on which it runs. Here are the numbers: 64 65 With soft updates: 66 67 labrat# time ./MAKEDEV std 68 2.2u 32.6s 0:34.82 100.0% 0+0k 11+36io 0pf+0w 69 70 labrat# ls | wc 71 522 522 3317 72 73 Without soft updates: 74 75 labrat# time ./MAKEDEV std 76 2.0u 40.5s 0:42.53 100.0% 0+0k 11+1221io 0pf+0w 77 78 labrat# ls | wc 79 522 522 3317 80 81 Of course, some of the system time is being pushed 82 to the syncer process, but that is a different story. 83 84 To show a benchmark designed to highlight the soft update code 85 consider a tar of zero-sized files and an rm -rf of a directory tree 86 that has at least 50 files or so at each level. Running a test with 87 a directory tree containing 28 directories holding 202 empty files 88 produces the following numbers: 89 90 With soft updates: 91 tar: 0.0u 0.5s 0:00.65 76.9% 0+0k 0+44io 0pf+0w (0 sync, 33 async writes) 92 rm: 0.0u 0.2s 0:00.20 100.0% 0+0k 0+37io 0pf+0w (0 sync, 72 async writes) 93 94 Normal filesystem: 95 tar: 0.0u 1.1s 0:07.27 16.5% 0+0k 60+586io 0pf+0w (523 sync, 0 async writes) 96 rm: 0.0u 0.5s 0:01.84 29.3% 0+0k 0+318io 0pf+0w (258 sync, 65 async writes) 97 98 The large reduction in writes is because inodes are clustered, so 99 most of a block gets allocated, then the whole block is written 100 out once rather than having the same block written once for each 101 inode allocated from it. Similarly each directory block is written 102 once rather than once for each new directory entry. Effectively 103 what the update code is doing is allocating a bunch of inodes 104 and directory entries without writing anything, then ensuring that 105 the block containing the inodes is written first followed by the 106 directory block that references them. If there were data in the 107 files it would further ensure that the data blocks were written 108 before their inodes claimed them. 109 110 Copyright Restrictions 111 112 Please familiarize yourself with the copyright restrictions 113 contained at the top of either the sys/ufs/ffs/softdep.h or 114 sys/ufs/ffs/ffs_softdep.c file. The key provision is similar 115 to the one used by the DB 2.0 package and goes as follows: 116 117 Redistributions in any form must be accompanied by information 118 on how to obtain complete source code for any accompanying 119 software that uses the this software. This source code must 120 either be included in the distribution or be available for 121 no more than the cost of distribution plus a nominal fee, 122 and must be freely redistributable under reasonable 123 conditions. For an executable file, complete source code 124 means the source code for all modules it contains. It does 125 not mean source code for modules or files that typically 126 accompany the operating system on which the executable file 127 runs, e.g., standard library modules or system header files. 128 129 The idea is to allow those of you freely redistributing your source 130 to use it while retaining for myself the right to peddle it for 131 money to the commercial UNIX vendors. Note that I have included a 132 stub file ffs_softdep.c.stub that is freely redistributable so that 133 you can put in all the necessary hooks to run the full soft updates 134 code, but still allow vendors that want to maintain proprietary 135 source to have a working system. I do plan to release the code with 136 a `Berkeley style' copyright once I have peddled it around to the 137 commercial vendors. If you have concerns about this copyright, 138 feel free to contact me with them and we can try to resolve any 139 difficulties. 140 141 Soft Dependency Operation 142 143 The soft update implementation does NOT require ANY changes 144 to the on-disk format of your filesystems. Furthermore it is 145 not used by default for any filesystems. It must be enabled on 146 a filesystem by filesystem basis by running tunefs to set a 147 bit in the superblock indicating that the filesystem should be 148 managed using soft updates. If you wish to stop using 149 soft updates due to performance or reliability reasons, 150 you can simply run tunefs on it again to turn off the bit and 151 revert to normal operation. The additional dynamic memory load 152 placed on the kernel malloc arena is approximately equal to 153 the amount of memory used by vnodes plus inodes (for a system 154 with 1000 vnodes, the additional peak memory load is about 300K). 155 156 Kernel Changes 157 158 There are two new changes to the kernel functionality that are not 159 contained in in the soft update files. The first is a `trickle 160 sync' facility running in the kernel as process 3. This trickle 161 sync process replaces the traditional `update' program (which should 162 be commented out of the /etc/rc startup script). When a vnode is 163 first written it is placed 30 seconds down on the trickle sync 164 queue. If it still exists and has dirty data when it reaches the 165 top of the queue, it is sync'ed. This approach evens out the load 166 on the underlying I/O system and avoids writing short-lived files. 167 The papers on trickle-sync tend to favor aging based on buffers 168 rather than files. However, I sync on file age rather than buffer 169 age because the data structures are much smaller as there are 170 typically far fewer files than buffers. Although this can make the 171 I/O spikey when a big file times out, it is still much better than 172 the wholesale sync's that were happening before. It also adapts 173 much better to the soft update code where I want to control 174 aging to improve performance (inodes age in 10 seconds, directories 175 in 15 seconds, files in 30 seconds). This ensures that most 176 dependencies are gone (e.g., inodes are written when directory 177 entries want to go to disk) reducing the amount of rollback that 178 is needed. 179 180 The other main kernel change is to split the vnode freelist into 181 two separate lists. One for vnodes that are still being used to 182 identify buffers and the other for those vnodes no longer identifying 183 any buffers. The latter list is used by getnewvnode in preference 184 to the former. 185 186 Packaging of Kernel Changes 187 188 The sys subdirectory contains the changes and additions to the 189 kernel. My goal in writing this code was to minimize the changes 190 that need to be made to the kernel. Thus, most of the new code 191 is contained in the two new files softdep.h and ffs_softdep.c. 192 The rest of the kernel changes are simply inserting hooks to 193 call into these two new files. Although there has been some 194 structural reorganization of the filesystem code to accommodate 195 gathering the information required by the soft update code, 196 the actual ordering of filesystem operations when soft updates 197 are disabled is unchanged. 198 199 The kernel changes are packaged as a set of diffs. As I am 200 doing my development in BSD/OS, the diffs are relative to the 201 BSD/OS versions of the files. Because BSD/OS recently had 202 4.4BSD-Lite2 merged into it, the Lite2 files are a good starting 203 point for figuring out the changes. There are 40 files that 204 require change plus the two new files. Most of these files have 205 only a few lines of changes in them. However, four files have 206 fairly extensive changes: kern/vfs_subr.c, ufs/ufs/ufs_lookup.c, 207 ufs/ufs/ufs_vnops.c, and ufs/ffs/ffs_alloc.c. For these four 208 files, I have provided the original Lite2 version, the Lite2 209 version with the diffs merged in, and the diffs between the 210 BSD/OS and merged version. Even so, I expect that there will 211 be some difficulty in doing the merge; I am certainly willing 212 to assist in helping get the code merged into your system. 213 214 Packaging of Utility Changes 215 216 The utilities subdirectory contains the changes and additions 217 to the utilities. There are diffs to three utilities enclosed: 218 219 tunefs - add a flag to enable and disable soft updates 220 221 mount - print out whether soft updates are enabled and 222 also statistics on number of sync and async writes 223 224 fsck - tighter checks on acceptable errors and a slightly 225 different policy for what to put in lost+found on 226 filesystems using soft updates 227 228 In addition you should recompile vmstat so as to get reports 229 on the 13 new memory types used by the soft update code. 230 It is not necessary to use the new version of fsck, however it 231 would aid in my debugging if you do. Also, because of the time 232 lag between deleting a directory entry and the inode it 233 references, you will find a lot more files showing up in your 234 lost+found if you do not use the new version. Note that the 235 new version checks for the soft update flag in the superblock 236 and only uses the new algorithms if it is set. So, it will run 237 unchanged on the filesystems that are not using soft updates. 238 239 Operation 240 241 Once you have booted a kernel that incorporates the soft update 242 code and installed the updated utilities, do the following: 243 244 1) Comment out the update program in /etc/rc. 245 246 2) Run `tunefs -n enable' on one or more test filesystems. 247 248 3) Mount these filesystems and then type `mount' to ensure that 249 they have been enabled for soft updates. 250 251 4) Copy the test directory to a softdep filesystem, chdir into 252 it and run `./doit'. You may want to check out each of the 253 three subtests individually first: doit1 - andrew benchmarks, 254 doit2 - copy and removal of /etc, doit3 - find from /. 255 256 ==== 257 Additional notes from Feb 13 258 259 hen removing huge directories of files, it is possible to get 260 the incore state arbitrarily far ahead of the disk. Maintaining 261 all the associated depedency information can exhaust the kernel 262 malloc arena. To avoid this senario, I have put some limits on 263 the soft update code so that it will not be allowed to rampage 264 through all of the kernel memory. I enclose below the relevant 265 patches to vnode.h and vfs_subr.c (which allow the soft update 266 code to speed up the filesystem syncer process). I have also 267 included the diffs for ffs_softdep.c. I hope to make a pass over 268 ffs_softdep.c to isolate the differences with my standard version 269 so that these diffs are less painful to incorporate. 270 271 Since I know you like to play with tuning, I have put the relevant 272 knobs on sysctl debug variables. The tuning knobs can be viewed 273 with `sysctl debug' and set with `sysctl -w debug.<name>=value'. 274 The knobs are as follows: 275 276 debug.max_softdeps - limit on any given resource 277 debug.tickdelay - ticks to delay before allocating 278 debug.max_limit_hit - number of times tickdelay imposed 279 debug.rush_requests - number of rush requests to filesystem syncer 280 281 The max_softdeps limit is derived from vnodesdesired which in 282 turn is sized based on the amount of memory on the machine. 283 When the limit is hit, a process requesting a resource first 284 tries to speed up the filesystem syncer process. Such a 285 request is recorded as a rush_request. After syncdelay / 2 286 unserviced rush requests (typically 15) are in the filesystem 287 syncers queue (i.e., it is more than 15 seconds behind in its 288 work), the process requesting the memory is put to sleep for 289 tickdelay seconds. Such a delay is recorded in max_limit_hit. 290 Following this delay it is granted its memory without further 291 delay. I have tried the following experiments in which I 292 delete an MH directory containing 16,703 files: 293 294 Run # 1 2 3 295 296 max_softdeps 4496 4496 4496 297 tickdelay 100 == 1 sec 20 == 0.2 sec 2 == 0.02 sec 298 max_limit_hit 16 == 16 sec 27 == 5.4 sec 203 == 4.1 sec 299 rush_requests 147 102 93 300 run time 57 sec 46 sec 45 sec 301 I/O's 781 859 936 302 303 When run with no limits, it completes in 40 seconds. So, the 304 time spent in delay is directly added to the bottom line. 305 Shortening the tick delay does cut down the total running time, 306 but at the expense of generating more total I/O operations 307 due to the rush orders being sent to the filesystem syncer. 308 Although the number of rush orders decreases with a shorter 309 tick delay, there are more requests in each order, hence the 310 increase in I/O count. Also, although the I/O count does rise 311 with a shorter delay, it is still at least an order of magnitude 312 less than without soft updates. Anyway, you may want to play 313 around with these value to see what works best and to see if 314 you can get an insight into how best to tune them. If you get 315 out of memory panic's, then you have max_softdeps set too high. 316 The max_limit_hit and rush_requests show be reset to zero 317 before each run. The minimum legal value for tickdelay is 2 318 (if you set it below that, the code will use 2). 319 320
[ source navigation ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.