tools/testing/selftests/vm/charge_reserved_hugetlb.sh

   1 #!/bin/sh
   2 # SPDX-License-Identifier: GPL-2.0
   3
   4 set -e
   5
   6 if [[ $(id -u) -ne 0 ]]; then
   7   echo "This test must be run as root. Skipping..."
   8   exit 0
   9 fi
  10
  11 fault_limit_file=limit_in_bytes
  12 reservation_limit_file=rsvd.limit_in_bytes
  13 fault_usage_file=usage_in_bytes
  14 reservation_usage_file=rsvd.usage_in_bytes
  15
  16 if [[ "$1" == "-cgroup-v2" ]]; then
  17   cgroup2=1
  18   fault_limit_file=max
  19   reservation_limit_file=rsvd.max
  20   fault_usage_file=current
  21   reservation_usage_file=rsvd.current
  22 fi
  23
  24 cgroup_path=/dev/cgroup/memory
  25 if [[ ! -e $cgroup_path ]]; then
  26   mkdir -p $cgroup_path
  27   if [[ $cgroup2 ]]; then
  28     mount -t cgroup2 none $cgroup_path
  29   else
  30     mount -t cgroup memory,hugetlb $cgroup_path
  31   fi
  32 fi
  33
  34 if [[ $cgroup2 ]]; then
  35   echo "+hugetlb" >/dev/cgroup/memory/cgroup.subtree_control
  36 fi
  37
  38 function cleanup() {
  39   if [[ $cgroup2 ]]; then
  40     echo $$ >$cgroup_path/cgroup.procs
  41   else
  42     echo $$ >$cgroup_path/tasks
  43   fi
  44
  45   if [[ -e /mnt/huge ]]; then
  46     rm -rf /mnt/huge/*
  47     umount /mnt/huge || echo error
  48     rmdir /mnt/huge
  49   fi
  50   if [[ -e $cgroup_path/hugetlb_cgroup_test ]]; then
  51     rmdir $cgroup_path/hugetlb_cgroup_test
  52   fi
  53   if [[ -e $cgroup_path/hugetlb_cgroup_test1 ]]; then
  54     rmdir $cgroup_path/hugetlb_cgroup_test1
  55   fi
  56   if [[ -e $cgroup_path/hugetlb_cgroup_test2 ]]; then
  57     rmdir $cgroup_path/hugetlb_cgroup_test2
  58   fi
  59   echo 0 >/proc/sys/vm/nr_hugepages
  60   echo CLEANUP DONE
  61 }
  62
  63 function expect_equal() {
  64   local expected="$1"
  65   local actual="$2"
  66   local error="$3"
  67
  68   if [[ "$expected" != "$actual" ]]; then
  69     echo "expected ($expected) != actual ($actual): $3"
  70     cleanup
  71     exit 1
  72   fi
  73 }
  74
  75 function get_machine_hugepage_size() {
  76   hpz=$(grep -i hugepagesize /proc/meminfo)
  77   kb=${hpz:14:-3}
  78   mb=$(($kb / 1024))
  79   echo $mb
  80 }
  81
  82 MB=$(get_machine_hugepage_size)
  83
  84 function setup_cgroup() {
  85   local name="$1"
  86   local cgroup_limit="$2"
  87   local reservation_limit="$3"
  88
  89   mkdir $cgroup_path/$name
  90
  91   echo writing cgroup limit: "$cgroup_limit"
  92   echo "$cgroup_limit" >$cgroup_path/$name/hugetlb.${MB}MB.$fault_limit_file
  93
  94   echo writing reseravation limit: "$reservation_limit"
  95   echo "$reservation_limit" > \
  96     $cgroup_path/$name/hugetlb.${MB}MB.$reservation_limit_file
  97
  98   if [ -e "$cgroup_path/$name/cpuset.cpus" ]; then
  99     echo 0 >$cgroup_path/$name/cpuset.cpus
 100   fi
 101   if [ -e "$cgroup_path/$name/cpuset.mems" ]; then
 102     echo 0 >$cgroup_path/$name/cpuset.mems
 103   fi
 104 }
 105
 106 function wait_for_hugetlb_memory_to_get_depleted() {
 107   local cgroup="$1"
 108   local path="/dev/cgroup/memory/$cgroup/hugetlb.${MB}MB.$reservation_usage_file"
 109   # Wait for hugetlbfs memory to get depleted.
 110   while [ $(cat $path) != 0 ]; do
 111     echo Waiting for hugetlb memory to get depleted.
 112     cat $path
 113     sleep 0.5
 114   done
 115 }
 116
 117 function wait_for_hugetlb_memory_to_get_reserved() {
 118   local cgroup="$1"
 119   local size="$2"
 120
 121   local path="/dev/cgroup/memory/$cgroup/hugetlb.${MB}MB.$reservation_usage_file"
 122   # Wait for hugetlbfs memory to get written.
 123   while [ $(cat $path) != $size ]; do
 124     echo Waiting for hugetlb memory reservation to reach size $size.
 125     cat $path
 126     sleep 0.5
 127   done
 128 }
 129
 130 function wait_for_hugetlb_memory_to_get_written() {
 131   local cgroup="$1"
 132   local size="$2"
 133
 134   local path="/dev/cgroup/memory/$cgroup/hugetlb.${MB}MB.$fault_usage_file"
 135   # Wait for hugetlbfs memory to get written.
 136   while [ $(cat $path) != $size ]; do
 137     echo Waiting for hugetlb memory to reach size $size.
 138     cat $path
 139     sleep 0.5
 140   done
 141 }
 142
 143 function write_hugetlbfs_and_get_usage() {
 144   local cgroup="$1"
 145   local size="$2"
 146   local populate="$3"
 147   local write="$4"
 148   local path="$5"
 149   local method="$6"
 150   local private="$7"
 151   local expect_failure="$8"
 152   local reserve="$9"
 153
 154   # Function return values.
 155   reservation_failed=0
 156   oom_killed=0
 157   hugetlb_difference=0
 158   reserved_difference=0
 159
 160   local hugetlb_usage=$cgroup_path/$cgroup/hugetlb.${MB}MB.$fault_usage_file
 161   local reserved_usage=$cgroup_path/$cgroup/hugetlb.${MB}MB.$reservation_usage_file
 162
 163   local hugetlb_before=$(cat $hugetlb_usage)
 164   local reserved_before=$(cat $reserved_usage)
 165
 166   echo
 167   echo Starting:
 168   echo hugetlb_usage="$hugetlb_before"
 169   echo reserved_usage="$reserved_before"
 170   echo expect_failure is "$expect_failure"
 171
 172   output=$(mktemp)
 173   set +e
 174   if [[ "$method" == "1" ]] || [[ "$method" == 2 ]] ||
 175     [[ "$private" == "-r" ]] && [[ "$expect_failure" != 1 ]]; then
 176
 177     bash write_hugetlb_memory.sh "$size" "$populate" "$write" \
 178       "$cgroup" "$path" "$method" "$private" "-l" "$reserve" 2>&1 | tee $output &
 179
 180     local write_result=$?
 181     local write_pid=$!
 182
 183     until grep -q -i "DONE" $output; do
 184       echo waiting for DONE signal.
 185       if ! ps $write_pid > /dev/null
 186       then
 187         echo "FAIL: The write died"
 188         cleanup
 189         exit 1
 190       fi
 191       sleep 0.5
 192     done
 193
 194     echo ================= write_hugetlb_memory.sh output is:
 195     cat $output
 196     echo ================= end output.
 197
 198     if [[ "$populate" == "-o" ]] || [[ "$write" == "-w" ]]; then
 199       wait_for_hugetlb_memory_to_get_written "$cgroup" "$size"
 200     elif [[ "$reserve" != "-n" ]]; then
 201       wait_for_hugetlb_memory_to_get_reserved "$cgroup" "$size"
 202     else
 203       # This case doesn't produce visible effects, but we still have
 204       # to wait for the async process to start and execute...
 205       sleep 0.5
 206     fi
 207
 208     echo write_result is $write_result
 209   else
 210     bash write_hugetlb_memory.sh "$size" "$populate" "$write" \
 211       "$cgroup" "$path" "$method" "$private" "$reserve"
 212     local write_result=$?
 213
 214     if [[ "$reserve" != "-n" ]]; then
 215       wait_for_hugetlb_memory_to_get_reserved "$cgroup" "$size"
 216     fi
 217   fi
 218   set -e
 219
 220   if [[ "$write_result" == 1 ]]; then
 221     reservation_failed=1
 222   fi
 223
 224   # On linus/master, the above process gets SIGBUS'd on oomkill, with
 225   # return code 135. On earlier kernels, it gets actual oomkill, with return
 226   # code 137, so just check for both conditions in case we're testing
 227   # against an earlier kernel.
 228   if [[ "$write_result" == 135 ]] || [[ "$write_result" == 137 ]]; then
 229     oom_killed=1
 230   fi
 231
 232   local hugetlb_after=$(cat $hugetlb_usage)
 233   local reserved_after=$(cat $reserved_usage)
 234
 235   echo After write:
 236   echo hugetlb_usage="$hugetlb_after"
 237   echo reserved_usage="$reserved_after"
 238
 239   hugetlb_difference=$(($hugetlb_after - $hugetlb_before))
 240   reserved_difference=$(($reserved_after - $reserved_before))
 241 }
 242
 243 function cleanup_hugetlb_memory() {
 244   set +e
 245   local cgroup="$1"
 246   if [[ "$(pgrep -f write_to_hugetlbfs)" != "" ]]; then
 247     echo killing write_to_hugetlbfs
 248     killall -2 write_to_hugetlbfs
 249     wait_for_hugetlb_memory_to_get_depleted $cgroup
 250   fi
 251   set -e
 252
 253   if [[ -e /mnt/huge ]]; then
 254     rm -rf /mnt/huge/*
 255     umount /mnt/huge
 256     rmdir /mnt/huge
 257   fi
 258 }
 259
 260 function run_test() {
 261   local size=$(($1 * ${MB} * 1024 * 1024))
 262   local populate="$2"
 263   local write="$3"
 264   local cgroup_limit=$(($4 * ${MB} * 1024 * 1024))
 265   local reservation_limit=$(($5 * ${MB} * 1024 * 1024))
 266   local nr_hugepages="$6"
 267   local method="$7"
 268   local private="$8"
 269   local expect_failure="$9"
 270   local reserve="${10}"
 271
 272   # Function return values.
 273   hugetlb_difference=0
 274   reserved_difference=0
 275   reservation_failed=0
 276   oom_killed=0
 277
 278   echo nr hugepages = "$nr_hugepages"
 279   echo "$nr_hugepages" >/proc/sys/vm/nr_hugepages
 280
 281   setup_cgroup "hugetlb_cgroup_test" "$cgroup_limit" "$reservation_limit"
 282
 283   mkdir -p /mnt/huge
 284   mount -t hugetlbfs -o pagesize=${MB}M,size=256M none /mnt/huge
 285
 286   write_hugetlbfs_and_get_usage "hugetlb_cgroup_test" "$size" "$populate" \
 287     "$write" "/mnt/huge/test" "$method" "$private" "$expect_failure" \
 288     "$reserve"
 289
 290   cleanup_hugetlb_memory "hugetlb_cgroup_test"
 291
 292   local final_hugetlb=$(cat $cgroup_path/hugetlb_cgroup_test/hugetlb.${MB}MB.$fault_usage_file)
 293   local final_reservation=$(cat $cgroup_path/hugetlb_cgroup_test/hugetlb.${MB}MB.$reservation_usage_file)
 294
 295   echo $hugetlb_difference
 296   echo $reserved_difference
 297   expect_equal "0" "$final_hugetlb" "final hugetlb is not zero"
 298   expect_equal "0" "$final_reservation" "final reservation is not zero"
 299 }
 300
 301 function run_multiple_cgroup_test() {
 302   local size1="$1"
 303   local populate1="$2"
 304   local write1="$3"
 305   local cgroup_limit1="$4"
 306   local reservation_limit1="$5"
 307
 308   local size2="$6"
 309   local populate2="$7"
 310   local write2="$8"
 311   local cgroup_limit2="$9"
 312   local reservation_limit2="${10}"
 313
 314   local nr_hugepages="${11}"
 315   local method="${12}"
 316   local private="${13}"
 317   local expect_failure="${14}"
 318   local reserve="${15}"
 319
 320   # Function return values.
 321   hugetlb_difference1=0
 322   reserved_difference1=0
 323   reservation_failed1=0
 324   oom_killed1=0
 325
 326   hugetlb_difference2=0
 327   reserved_difference2=0
 328   reservation_failed2=0
 329   oom_killed2=0
 330
 331   echo nr hugepages = "$nr_hugepages"
 332   echo "$nr_hugepages" >/proc/sys/vm/nr_hugepages
 333
 334   setup_cgroup "hugetlb_cgroup_test1" "$cgroup_limit1" "$reservation_limit1"
 335   setup_cgroup "hugetlb_cgroup_test2" "$cgroup_limit2" "$reservation_limit2"
 336
 337   mkdir -p /mnt/huge
 338   mount -t hugetlbfs -o pagesize=${MB}M,size=256M none /mnt/huge
 339
 340   write_hugetlbfs_and_get_usage "hugetlb_cgroup_test1" "$size1" \
 341     "$populate1" "$write1" "/mnt/huge/test1" "$method" "$private" \
 342     "$expect_failure" "$reserve"
 343
 344   hugetlb_difference1=$hugetlb_difference
 345   reserved_difference1=$reserved_difference
 346   reservation_failed1=$reservation_failed
 347   oom_killed1=$oom_killed
 348
 349   local cgroup1_hugetlb_usage=$cgroup_path/hugetlb_cgroup_test1/hugetlb.${MB}MB.$fault_usage_file
 350   local cgroup1_reservation_usage=$cgroup_path/hugetlb_cgroup_test1/hugetlb.${MB}MB.$reservation_usage_file
 351   local cgroup2_hugetlb_usage=$cgroup_path/hugetlb_cgroup_test2/hugetlb.${MB}MB.$fault_usage_file
 352   local cgroup2_reservation_usage=$cgroup_path/hugetlb_cgroup_test2/hugetlb.${MB}MB.$reservation_usage_file
 353
 354   local usage_before_second_write=$(cat $cgroup1_hugetlb_usage)
 355   local reservation_usage_before_second_write=$(cat $cgroup1_reservation_usage)
 356
 357   write_hugetlbfs_and_get_usage "hugetlb_cgroup_test2" "$size2" \
 358     "$populate2" "$write2" "/mnt/huge/test2" "$method" "$private" \
 359     "$expect_failure" "$reserve"
 360
 361   hugetlb_difference2=$hugetlb_difference
 362   reserved_difference2=$reserved_difference
 363   reservation_failed2=$reservation_failed
 364   oom_killed2=$oom_killed
 365
 366   expect_equal "$usage_before_second_write" \
 367     "$(cat $cgroup1_hugetlb_usage)" "Usage changed."
 368   expect_equal "$reservation_usage_before_second_write" \
 369     "$(cat $cgroup1_reservation_usage)" "Reservation usage changed."
 370
 371   cleanup_hugetlb_memory
 372
 373   local final_hugetlb=$(cat $cgroup1_hugetlb_usage)
 374   local final_reservation=$(cat $cgroup1_reservation_usage)
 375
 376   expect_equal "0" "$final_hugetlb" \
 377     "hugetlbt_cgroup_test1 final hugetlb is not zero"
 378   expect_equal "0" "$final_reservation" \
 379     "hugetlbt_cgroup_test1 final reservation is not zero"
 380
 381   local final_hugetlb=$(cat $cgroup2_hugetlb_usage)
 382   local final_reservation=$(cat $cgroup2_reservation_usage)
 383
 384   expect_equal "0" "$final_hugetlb" \
 385     "hugetlb_cgroup_test2 final hugetlb is not zero"
 386   expect_equal "0" "$final_reservation" \
 387     "hugetlb_cgroup_test2 final reservation is not zero"
 388 }
 389
 390 cleanup
 391
 392 for populate in "" "-o"; do
 393   for method in 0 1 2; do
 394     for private in "" "-r"; do
 395       for reserve in "" "-n"; do
 396
 397         # Skip mmap(MAP_HUGETLB | MAP_SHARED). Doesn't seem to be supported.
 398         if [[ "$method" == 1 ]] && [[ "$private" == "" ]]; then
 399           continue
 400         fi
 401
 402         # Skip populated shmem tests. Doesn't seem to be supported.
 403         if [[ "$method" == 2"" ]] && [[ "$populate" == "-o" ]]; then
 404           continue
 405         fi
 406
 407         if [[ "$method" == 2"" ]] && [[ "$reserve" == "-n" ]]; then
 408           continue
 409         fi
 410
 411         cleanup
 412         echo
 413         echo
 414         echo
 415         echo Test normal case.
 416         echo private=$private, populate=$populate, method=$method, reserve=$reserve
 417         run_test 5 "$populate" "" 10 10 10 "$method" "$private" "0" "$reserve"
 418
 419         echo Memory charged to hugtlb=$hugetlb_difference
 420         echo Memory charged to reservation=$reserved_difference
 421
 422         if [[ "$populate" == "-o" ]]; then
 423           expect_equal "$((5 * $MB * 1024 * 1024))" "$hugetlb_difference" \
 424             "Reserved memory charged to hugetlb cgroup."
 425         else
 426           expect_equal "0" "$hugetlb_difference" \
 427             "Reserved memory charged to hugetlb cgroup."
 428         fi
 429
 430         if [[ "$reserve" != "-n" ]] || [[ "$populate" == "-o" ]]; then
 431           expect_equal "$((5 * $MB * 1024 * 1024))" "$reserved_difference" \
 432             "Reserved memory not charged to reservation usage."
 433         else
 434           expect_equal "0" "$reserved_difference" \
 435             "Reserved memory not charged to reservation usage."
 436         fi
 437
 438         echo 'PASS'
 439
 440         cleanup
 441         echo
 442         echo
 443         echo
 444         echo Test normal case with write.
 445         echo private=$private, populate=$populate, method=$method, reserve=$reserve
 446         run_test 5 "$populate" '-w' 5 5 10 "$method" "$private" "0" "$reserve"
 447
 448         echo Memory charged to hugtlb=$hugetlb_difference
 449         echo Memory charged to reservation=$reserved_difference
 450
 451         expect_equal "$((5 * $MB * 1024 * 1024))" "$hugetlb_difference" \
 452           "Reserved memory charged to hugetlb cgroup."
 453
 454         expect_equal "$((5 * $MB * 1024 * 1024))" "$reserved_difference" \
 455           "Reserved memory not charged to reservation usage."
 456
 457         echo 'PASS'
 458
 459         cleanup
 460         continue
 461         echo
 462         echo
 463         echo
 464         echo Test more than reservation case.
 465         echo private=$private, populate=$populate, method=$method, reserve=$reserve
 466
 467         if [ "$reserve" != "-n" ]; then
 468           run_test "5" "$populate" '' "10" "2" "10" "$method" "$private" "1" \
 469             "$reserve"
 470
 471           expect_equal "1" "$reservation_failed" "Reservation succeeded."
 472         fi
 473
 474         echo 'PASS'
 475
 476         cleanup
 477
 478         echo
 479         echo
 480         echo
 481         echo Test more than cgroup limit case.
 482         echo private=$private, populate=$populate, method=$method, reserve=$reserve
 483
 484         # Not sure if shm memory can be cleaned up when the process gets sigbus'd.
 485         if [[ "$method" != 2 ]]; then
 486           run_test 5 "$populate" "-w" 2 10 10 "$method" "$private" "1" "$reserve"
 487
 488           expect_equal "1" "$oom_killed" "Not oom killed."
 489         fi
 490         echo 'PASS'
 491
 492         cleanup
 493
 494         echo
 495         echo
 496         echo
 497         echo Test normal case, multiple cgroups.
 498         echo private=$private, populate=$populate, method=$method, reserve=$reserve
 499         run_multiple_cgroup_test "3" "$populate" "" "10" "10" "5" \
 500           "$populate" "" "10" "10" "10" \
 501           "$method" "$private" "0" "$reserve"
 502
 503         echo Memory charged to hugtlb1=$hugetlb_difference1
 504         echo Memory charged to reservation1=$reserved_difference1
 505         echo Memory charged to hugtlb2=$hugetlb_difference2
 506         echo Memory charged to reservation2=$reserved_difference2
 507
 508         if [[ "$reserve" != "-n" ]] || [[ "$populate" == "-o" ]]; then
 509           expect_equal "3" "$reserved_difference1" \
 510             "Incorrect reservations charged to cgroup 1."
 511
 512           expect_equal "5" "$reserved_difference2" \
 513             "Incorrect reservation charged to cgroup 2."
 514
 515         else
 516           expect_equal "0" "$reserved_difference1" \
 517             "Incorrect reservations charged to cgroup 1."
 518
 519           expect_equal "0" "$reserved_difference2" \
 520             "Incorrect reservation charged to cgroup 2."
 521         fi
 522
 523         if [[ "$populate" == "-o" ]]; then
 524           expect_equal "3" "$hugetlb_difference1" \
 525             "Incorrect hugetlb charged to cgroup 1."
 526
 527           expect_equal "5" "$hugetlb_difference2" \
 528             "Incorrect hugetlb charged to cgroup 2."
 529
 530         else
 531           expect_equal "0" "$hugetlb_difference1" \
 532             "Incorrect hugetlb charged to cgroup 1."
 533
 534           expect_equal "0" "$hugetlb_difference2" \
 535             "Incorrect hugetlb charged to cgroup 2."
 536         fi
 537         echo 'PASS'
 538
 539         cleanup
 540         echo
 541         echo
 542         echo
 543         echo Test normal case with write, multiple cgroups.
 544         echo private=$private, populate=$populate, method=$method, reserve=$reserve
 545         run_multiple_cgroup_test "3" "$populate" "-w" "10" "10" "5" \
 546           "$populate" "-w" "10" "10" "10" \
 547           "$method" "$private" "0" "$reserve"
 548
 549         echo Memory charged to hugtlb1=$hugetlb_difference1
 550         echo Memory charged to reservation1=$reserved_difference1
 551         echo Memory charged to hugtlb2=$hugetlb_difference2
 552         echo Memory charged to reservation2=$reserved_difference2
 553
 554         expect_equal "3" "$hugetlb_difference1" \
 555           "Incorrect hugetlb charged to cgroup 1."
 556
 557         expect_equal "3" "$reserved_difference1" \
 558           "Incorrect reservation charged to cgroup 1."
 559
 560         expect_equal "5" "$hugetlb_difference2" \
 561           "Incorrect hugetlb charged to cgroup 2."
 562
 563         expect_equal "5" "$reserved_difference2" \
 564           "Incorrected reservation charged to cgroup 2."
 565         echo 'PASS'
 566
 567         cleanup
 568
 569       done # reserve
 570     done   # private
 571   done     # populate
 572 done       # method
 573
 574 umount $cgroup_path
 575 rmdir $cgroup_path