Based on kernel version 4.2. Page generated on 2015-09-09 12:08 EST.
1 Block IO Controller 2 =================== 3 Overview 4 ======== 5 cgroup subsys "blkio" implements the block io controller. There seems to be 6 a need of various kinds of IO control policies (like proportional BW, max BW) 7 both at leaf nodes as well as at intermediate nodes in a storage hierarchy. 8 Plan is to use the same cgroup based management interface for blkio controller 9 and based on user options switch IO policies in the background. 10 11 Currently two IO control policies are implemented. First one is proportional 12 weight time based division of disk policy. It is implemented in CFQ. Hence 13 this policy takes effect only on leaf nodes when CFQ is being used. The second 14 one is throttling policy which can be used to specify upper IO rate limits 15 on devices. This policy is implemented in generic block layer and can be 16 used on leaf nodes as well as higher level logical devices like device mapper. 17 18 HOWTO 19 ===== 20 Proportional Weight division of bandwidth 21 ----------------------------------------- 22 You can do a very simple testing of running two dd threads in two different 23 cgroups. Here is what you can do. 24 25 - Enable Block IO controller 26 CONFIG_BLK_CGROUP=y 27 28 - Enable group scheduling in CFQ 29 CONFIG_CFQ_GROUP_IOSCHED=y 30 31 - Compile and boot into kernel and mount IO controller (blkio); see 32 cgroups.txt, Why are cgroups needed?. 33 34 mount -t tmpfs cgroup_root /sys/fs/cgroup 35 mkdir /sys/fs/cgroup/blkio 36 mount -t cgroup -o blkio none /sys/fs/cgroup/blkio 37 38 - Create two cgroups 39 mkdir -p /sys/fs/cgroup/blkio/test1/ /sys/fs/cgroup/blkio/test2 40 41 - Set weights of group test1 and test2 42 echo 1000 > /sys/fs/cgroup/blkio/test1/blkio.weight 43 echo 500 > /sys/fs/cgroup/blkio/test2/blkio.weight 44 45 - Create two same size files (say 512MB each) on same disk (file1, file2) and 46 launch two dd threads in different cgroup to read those files. 47 48 sync 49 echo 3 > /proc/sys/vm/drop_caches 50 51 dd if=/mnt/sdb/zerofile1 of=/dev/null & 52 echo $! > /sys/fs/cgroup/blkio/test1/tasks 53 cat /sys/fs/cgroup/blkio/test1/tasks 54 55 dd if=/mnt/sdb/zerofile2 of=/dev/null & 56 echo $! > /sys/fs/cgroup/blkio/test2/tasks 57 cat /sys/fs/cgroup/blkio/test2/tasks 58 59 - At macro level, first dd should finish first. To get more precise data, keep 60 on looking at (with the help of script), at blkio.disk_time and 61 blkio.disk_sectors files of both test1 and test2 groups. This will tell how 62 much disk time (in milli seconds), each group got and how many secotors each 63 group dispatched to the disk. We provide fairness in terms of disk time, so 64 ideally io.disk_time of cgroups should be in proportion to the weight. 65 66 Throttling/Upper Limit policy 67 ----------------------------- 68 - Enable Block IO controller 69 CONFIG_BLK_CGROUP=y 70 71 - Enable throttling in block layer 72 CONFIG_BLK_DEV_THROTTLING=y 73 74 - Mount blkio controller (see cgroups.txt, Why are cgroups needed?) 75 mount -t cgroup -o blkio none /sys/fs/cgroup/blkio 76 77 - Specify a bandwidth rate on particular device for root group. The format 78 for policy is "<major>:<minor> <bytes_per_second>". 79 80 echo "8:16 1048576" > /sys/fs/cgroup/blkio/blkio.throttle.read_bps_device 81 82 Above will put a limit of 1MB/second on reads happening for root group 83 on device having major/minor number 8:16. 84 85 - Run dd to read a file and see if rate is throttled to 1MB/s or not. 86 87 # dd if=/mnt/common/zerofile of=/dev/null bs=4K count=1024 88 # iflag=direct 89 1024+0 records in 90 1024+0 records out 91 4194304 bytes (4.2 MB) copied, 4.0001 s, 1.0 MB/s 92 93 Limits for writes can be put using blkio.throttle.write_bps_device file. 94 95 Hierarchical Cgroups 96 ==================== 97 98 Both CFQ and throttling implement hierarchy support; however, 99 throttling's hierarchy support is enabled iff "sane_behavior" is 100 enabled from cgroup side, which currently is a development option and 101 not publicly available. 102 103 If somebody created a hierarchy like as follows. 104 105 root 106 / \ 107 test1 test2 108 | 109 test3 110 111 CFQ by default and throttling with "sane_behavior" will handle the 112 hierarchy correctly. For details on CFQ hierarchy support, refer to 113 Documentation/block/cfq-iosched.txt. For throttling, all limits apply 114 to the whole subtree while all statistics are local to the IOs 115 directly generated by tasks in that cgroup. 116 117 Throttling without "sane_behavior" enabled from cgroup side will 118 practically treat all groups at same level as if it looks like the 119 following. 120 121 pivot 122 / / \ \ 123 root test1 test2 test3 124 125 Various user visible config options 126 =================================== 127 CONFIG_BLK_CGROUP 128 - Block IO controller. 129 130 CONFIG_DEBUG_BLK_CGROUP 131 - Debug help. Right now some additional stats file show up in cgroup 132 if this option is enabled. 133 134 CONFIG_CFQ_GROUP_IOSCHED 135 - Enables group scheduling in CFQ. Currently only 1 level of group 136 creation is allowed. 137 138 CONFIG_BLK_DEV_THROTTLING 139 - Enable block device throttling support in block layer. 140 141 Details of cgroup files 142 ======================= 143 Proportional weight policy files 144 -------------------------------- 145 - blkio.weight 146 - Specifies per cgroup weight. This is default weight of the group 147 on all the devices until and unless overridden by per device rule. 148 (See blkio.weight_device). 149 Currently allowed range of weights is from 10 to 1000. 150 151 - blkio.weight_device 152 - One can specify per cgroup per device rules using this interface. 153 These rules override the default value of group weight as specified 154 by blkio.weight. 155 156 Following is the format. 157 158 # echo dev_maj:dev_minor weight > blkio.weight_device 159 Configure weight=300 on /dev/sdb (8:16) in this cgroup 160 # echo 8:16 300 > blkio.weight_device 161 # cat blkio.weight_device 162 dev weight 163 8:16 300 164 165 Configure weight=500 on /dev/sda (8:0) in this cgroup 166 # echo 8:0 500 > blkio.weight_device 167 # cat blkio.weight_device 168 dev weight 169 8:0 500 170 8:16 300 171 172 Remove specific weight for /dev/sda in this cgroup 173 # echo 8:0 0 > blkio.weight_device 174 # cat blkio.weight_device 175 dev weight 176 8:16 300 177 178 - blkio.leaf_weight[_device] 179 - Equivalents of blkio.weight[_device] for the purpose of 180 deciding how much weight tasks in the given cgroup has while 181 competing with the cgroup's child cgroups. For details, 182 please refer to Documentation/block/cfq-iosched.txt. 183 184 - blkio.time 185 - disk time allocated to cgroup per device in milliseconds. First 186 two fields specify the major and minor number of the device and 187 third field specifies the disk time allocated to group in 188 milliseconds. 189 190 - blkio.sectors 191 - number of sectors transferred to/from disk by the group. First 192 two fields specify the major and minor number of the device and 193 third field specifies the number of sectors transferred by the 194 group to/from the device. 195 196 - blkio.io_service_bytes 197 - Number of bytes transferred to/from the disk by the group. These 198 are further divided by the type of operation - read or write, sync 199 or async. First two fields specify the major and minor number of the 200 device, third field specifies the operation type and the fourth field 201 specifies the number of bytes. 202 203 - blkio.io_serviced 204 - Number of IOs completed to/from the disk by the group. These 205 are further divided by the type of operation - read or write, sync 206 or async. First two fields specify the major and minor number of the 207 device, third field specifies the operation type and the fourth field 208 specifies the number of IOs. 209 210 - blkio.io_service_time 211 - Total amount of time between request dispatch and request completion 212 for the IOs done by this cgroup. This is in nanoseconds to make it 213 meaningful for flash devices too. For devices with queue depth of 1, 214 this time represents the actual service time. When queue_depth > 1, 215 that is no longer true as requests may be served out of order. This 216 may cause the service time for a given IO to include the service time 217 of multiple IOs when served out of order which may result in total 218 io_service_time > actual time elapsed. This time is further divided by 219 the type of operation - read or write, sync or async. First two fields 220 specify the major and minor number of the device, third field 221 specifies the operation type and the fourth field specifies the 222 io_service_time in ns. 223 224 - blkio.io_wait_time 225 - Total amount of time the IOs for this cgroup spent waiting in the 226 scheduler queues for service. This can be greater than the total time 227 elapsed since it is cumulative io_wait_time for all IOs. It is not a 228 measure of total time the cgroup spent waiting but rather a measure of 229 the wait_time for its individual IOs. For devices with queue_depth > 1 230 this metric does not include the time spent waiting for service once 231 the IO is dispatched to the device but till it actually gets serviced 232 (there might be a time lag here due to re-ordering of requests by the 233 device). This is in nanoseconds to make it meaningful for flash 234 devices too. This time is further divided by the type of operation - 235 read or write, sync or async. First two fields specify the major and 236 minor number of the device, third field specifies the operation type 237 and the fourth field specifies the io_wait_time in ns. 238 239 - blkio.io_merged 240 - Total number of bios/requests merged into requests belonging to this 241 cgroup. This is further divided by the type of operation - read or 242 write, sync or async. 243 244 - blkio.io_queued 245 - Total number of requests queued up at any given instant for this 246 cgroup. This is further divided by the type of operation - read or 247 write, sync or async. 248 249 - blkio.avg_queue_size 250 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. 251 The average queue size for this cgroup over the entire time of this 252 cgroup's existence. Queue size samples are taken each time one of the 253 queues of this cgroup gets a timeslice. 254 255 - blkio.group_wait_time 256 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. 257 This is the amount of time the cgroup had to wait since it became busy 258 (i.e., went from 0 to 1 request queued) to get a timeslice for one of 259 its queues. This is different from the io_wait_time which is the 260 cumulative total of the amount of time spent by each IO in that cgroup 261 waiting in the scheduler queue. This is in nanoseconds. If this is 262 read when the cgroup is in a waiting (for timeslice) state, the stat 263 will only report the group_wait_time accumulated till the last time it 264 got a timeslice and will not include the current delta. 265 266 - blkio.empty_time 267 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. 268 This is the amount of time a cgroup spends without any pending 269 requests when not being served, i.e., it does not include any time 270 spent idling for one of the queues of the cgroup. This is in 271 nanoseconds. If this is read when the cgroup is in an empty state, 272 the stat will only report the empty_time accumulated till the last 273 time it had a pending request and will not include the current delta. 274 275 - blkio.idle_time 276 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. 277 This is the amount of time spent by the IO scheduler idling for a 278 given cgroup in anticipation of a better request than the existing ones 279 from other queues/cgroups. This is in nanoseconds. If this is read 280 when the cgroup is in an idling state, the stat will only report the 281 idle_time accumulated till the last idle period and will not include 282 the current delta. 283 284 - blkio.dequeue 285 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. This 286 gives the statistics about how many a times a group was dequeued 287 from service tree of the device. First two fields specify the major 288 and minor number of the device and third field specifies the number 289 of times a group was dequeued from a particular device. 290 291 - blkio.*_recursive 292 - Recursive version of various stats. These files show the 293 same information as their non-recursive counterparts but 294 include stats from all the descendant cgroups. 295 296 Throttling/Upper limit policy files 297 ----------------------------------- 298 - blkio.throttle.read_bps_device 299 - Specifies upper limit on READ rate from the device. IO rate is 300 specified in bytes per second. Rules are per device. Following is 301 the format. 302 303 echo "<major>:<minor> <rate_bytes_per_second>" > /cgrp/blkio.throttle.read_bps_device 304 305 - blkio.throttle.write_bps_device 306 - Specifies upper limit on WRITE rate to the device. IO rate is 307 specified in bytes per second. Rules are per device. Following is 308 the format. 309 310 echo "<major>:<minor> <rate_bytes_per_second>" > /cgrp/blkio.throttle.write_bps_device 311 312 - blkio.throttle.read_iops_device 313 - Specifies upper limit on READ rate from the device. IO rate is 314 specified in IO per second. Rules are per device. Following is 315 the format. 316 317 echo "<major>:<minor> <rate_io_per_second>" > /cgrp/blkio.throttle.read_iops_device 318 319 - blkio.throttle.write_iops_device 320 - Specifies upper limit on WRITE rate to the device. IO rate is 321 specified in io per second. Rules are per device. Following is 322 the format. 323 324 echo "<major>:<minor> <rate_io_per_second>" > /cgrp/blkio.throttle.write_iops_device 325 326 Note: If both BW and IOPS rules are specified for a device, then IO is 327 subjected to both the constraints. 328 329 - blkio.throttle.io_serviced 330 - Number of IOs (bio) completed to/from the disk by the group (as 331 seen by throttling policy). These are further divided by the type 332 of operation - read or write, sync or async. First two fields specify 333 the major and minor number of the device, third field specifies the 334 operation type and the fourth field specifies the number of IOs. 335 336 blkio.io_serviced does accounting as seen by CFQ and counts are in 337 number of requests (struct request). On the other hand, 338 blkio.throttle.io_serviced counts number of IO in terms of number 339 of bios as seen by throttling policy. These bios can later be 340 merged by elevator and total number of requests completed can be 341 lesser. 342 343 - blkio.throttle.io_service_bytes 344 - Number of bytes transferred to/from the disk by the group. These 345 are further divided by the type of operation - read or write, sync 346 or async. First two fields specify the major and minor number of the 347 device, third field specifies the operation type and the fourth field 348 specifies the number of bytes. 349 350 These numbers should roughly be same as blkio.io_service_bytes as 351 updated by CFQ. The difference between two is that 352 blkio.io_service_bytes will not be updated if CFQ is not operating 353 on request queue. 354 355 Common files among various policies 356 ----------------------------------- 357 - blkio.reset_stats 358 - Writing an int to this file will result in resetting all the stats 359 for that cgroup. 360 361 CFQ sysfs tunable 362 ================= 363 /sys/block/<disk>/queue/iosched/slice_idle 364 ------------------------------------------ 365 On a faster hardware CFQ can be slow, especially with sequential workload. 366 This happens because CFQ idles on a single queue and single queue might not 367 drive deeper request queue depths to keep the storage busy. In such scenarios 368 one can try setting slice_idle=0 and that would switch CFQ to IOPS 369 (IO operations per second) mode on NCQ supporting hardware. 370 371 That means CFQ will not idle between cfq queues of a cfq group and hence be 372 able to driver higher queue depth and achieve better throughput. That also 373 means that cfq provides fairness among groups in terms of IOPS and not in 374 terms of disk time. 375 376 /sys/block/<disk>/queue/iosched/group_idle 377 ------------------------------------------ 378 If one disables idling on individual cfq queues and cfq service trees by 379 setting slice_idle=0, group_idle kicks in. That means CFQ will still idle 380 on the group in an attempt to provide fairness among groups. 381 382 By default group_idle is same as slice_idle and does not do anything if 383 slice_idle is enabled. 384 385 One can experience an overall throughput drop if you have created multiple 386 groups and put applications in that group which are not driving enough 387 IO to keep disk busy. In that case set group_idle=0, and CFQ will not idle 388 on individual groups and throughput should improve. 389 390 Writeback 391 ========= 392 393 Page cache is dirtied through buffered writes and shared mmaps and 394 written asynchronously to the backing filesystem by the writeback 395 mechanism. Writeback sits between the memory and IO domains and 396 regulates the proportion of dirty memory by balancing dirtying and 397 write IOs. 398 399 On traditional cgroup hierarchies, relationships between different 400 controllers cannot be established making it impossible for writeback 401 to operate accounting for cgroup resource restrictions and all 402 writeback IOs are attributed to the root cgroup. 403 404 If both the blkio and memory controllers are used on the v2 hierarchy 405 and the filesystem supports cgroup writeback, writeback operations 406 correctly follow the resource restrictions imposed by both memory and 407 blkio controllers. 408 409 Writeback examines both system-wide and per-cgroup dirty memory status 410 and enforces the more restrictive of the two. Also, writeback control 411 parameters which are absolute values - vm.dirty_bytes and 412 vm.dirty_background_bytes - are distributed across cgroups according 413 to their current writeback bandwidth. 414 415 There's a peculiarity stemming from the discrepancy in ownership 416 granularity between memory controller and writeback. While memory 417 controller tracks ownership per page, writeback operates on inode 418 basis. cgroup writeback bridges the gap by tracking ownership by 419 inode but migrating ownership if too many foreign pages, pages which 420 don't match the current inode ownership, have been encountered while 421 writing back the inode. 422 423 This is a conscious design choice as writeback operations are 424 inherently tied to inodes making strictly following page ownership 425 complicated and inefficient. The only use case which suffers from 426 this compromise is multiple cgroups concurrently dirtying disjoint 427 regions of the same inode, which is an unlikely use case and decided 428 to be unsupported. Note that as memory controller assigns page 429 ownership on the first use and doesn't update it until the page is 430 released, even if cgroup writeback strictly follows page ownership, 431 multiple cgroups dirtying overlapping areas wouldn't work as expected. 432 In general, write-sharing an inode across multiple cgroups is not well 433 supported. 434 435 Filesystem support for cgroup writeback 436 --------------------------------------- 437 438 A filesystem can make writeback IOs cgroup-aware by updating 439 address_space_operations->writepage[s]() to annotate bio's using the 440 following two functions. 441 442 * wbc_init_bio(@wbc, @bio) 443 444 Should be called for each bio carrying writeback data and associates 445 the bio with the inode's owner cgroup. Can be called anytime 446 between bio allocation and submission. 447 448 * wbc_account_io(@wbc, @page, @bytes) 449 450 Should be called for each data segment being written out. While 451 this function doesn't care exactly when it's called during the 452 writeback session, it's the easiest and most natural to call it as 453 data segments are added to a bio. 454 455 With writeback bio's annotated, cgroup support can be enabled per 456 super_block by setting MS_CGROUPWB in ->s_flags. This allows for 457 selective disabling of cgroup writeback support which is helpful when 458 certain filesystem features, e.g. journaled data mode, are 459 incompatible. 460 461 wbc_init_bio() binds the specified bio to its cgroup. Depending on 462 the configuration, the bio may be executed at a lower priority and if 463 the writeback session is holding shared resources, e.g. a journal 464 entry, may lead to priority inversion. There is no one easy solution 465 for the problem. Filesystems can try to work around specific problem 466 cases by skipping wbc_init_bio() or using bio_associate_blkcg() 467 directly.