1================== 2HugeTLB Controller 3================== 4 5HugeTLB controller can be created by first mounting the cgroup filesystem. 6 7# mount -t cgroup -o hugetlb none /sys/fs/cgroup 8 9With the above step, the initial or the parent HugeTLB group becomes 10visible at /sys/fs/cgroup. At bootup, this group includes all the tasks in 11the system. /sys/fs/cgroup/tasks lists the tasks in this cgroup. 12 13New groups can be created under the parent group /sys/fs/cgroup:: 14 15 # cd /sys/fs/cgroup 16 # mkdir g1 17 # echo $$ > g1/tasks 18 19The above steps create a new group g1 and move the current shell 20process (bash) into it. 21 22Brief summary of control files:: 23 24 hugetlb.<hugepagesize>.rsvd.limit_in_bytes # set/show limit of "hugepagesize" hugetlb reservations 25 hugetlb.<hugepagesize>.rsvd.max_usage_in_bytes # show max "hugepagesize" hugetlb reservations and no-reserve faults 26 hugetlb.<hugepagesize>.rsvd.usage_in_bytes # show current reservations and no-reserve faults for "hugepagesize" hugetlb 27 hugetlb.<hugepagesize>.rsvd.failcnt # show the number of allocation failure due to HugeTLB reservation limit 28 hugetlb.<hugepagesize>.limit_in_bytes # set/show limit of "hugepagesize" hugetlb faults 29 hugetlb.<hugepagesize>.max_usage_in_bytes # show max "hugepagesize" hugetlb usage recorded 30 hugetlb.<hugepagesize>.usage_in_bytes # show current usage for "hugepagesize" hugetlb 31 hugetlb.<hugepagesize>.failcnt # show the number of allocation failure due to HugeTLB usage limit 32 hugetlb.<hugepagesize>.numa_stat # show the numa information of the hugetlb memory charged to this cgroup 33 34For a system supporting three hugepage sizes (64k, 32M and 1G), the control 35files include:: 36 37 hugetlb.1GB.limit_in_bytes 38 hugetlb.1GB.max_usage_in_bytes 39 hugetlb.1GB.numa_stat 40 hugetlb.1GB.usage_in_bytes 41 hugetlb.1GB.failcnt 42 hugetlb.1GB.rsvd.limit_in_bytes 43 hugetlb.1GB.rsvd.max_usage_in_bytes 44 hugetlb.1GB.rsvd.usage_in_bytes 45 hugetlb.1GB.rsvd.failcnt 46 hugetlb.64KB.limit_in_bytes 47 hugetlb.64KB.max_usage_in_bytes 48 hugetlb.64KB.numa_stat 49 hugetlb.64KB.usage_in_bytes 50 hugetlb.64KB.failcnt 51 hugetlb.64KB.rsvd.limit_in_bytes 52 hugetlb.64KB.rsvd.max_usage_in_bytes 53 hugetlb.64KB.rsvd.usage_in_bytes 54 hugetlb.64KB.rsvd.failcnt 55 hugetlb.32MB.limit_in_bytes 56 hugetlb.32MB.max_usage_in_bytes 57 hugetlb.32MB.numa_stat 58 hugetlb.32MB.usage_in_bytes 59 hugetlb.32MB.failcnt 60 hugetlb.32MB.rsvd.limit_in_bytes 61 hugetlb.32MB.rsvd.max_usage_in_bytes 62 hugetlb.32MB.rsvd.usage_in_bytes 63 hugetlb.32MB.rsvd.failcnt 64 65 661. Page fault accounting 67 68:: 69 70 hugetlb.<hugepagesize>.limit_in_bytes 71 hugetlb.<hugepagesize>.max_usage_in_bytes 72 hugetlb.<hugepagesize>.usage_in_bytes 73 hugetlb.<hugepagesize>.failcnt 74 75The HugeTLB controller allows users to limit the HugeTLB usage (page fault) per 76control group and enforces the limit during page fault. Since HugeTLB 77doesn't support page reclaim, enforcing the limit at page fault time implies 78that, the application will get SIGBUS signal if it tries to fault in HugeTLB 79pages beyond its limit. Therefore the application needs to know exactly how many 80HugeTLB pages it uses before hand, and the sysadmin needs to make sure that 81there are enough available on the machine for all the users to avoid processes 82getting SIGBUS. 83 84 852. Reservation accounting 86 87:: 88 89 hugetlb.<hugepagesize>.rsvd.limit_in_bytes 90 hugetlb.<hugepagesize>.rsvd.max_usage_in_bytes 91 hugetlb.<hugepagesize>.rsvd.usage_in_bytes 92 hugetlb.<hugepagesize>.rsvd.failcnt 93 94The HugeTLB controller allows to limit the HugeTLB reservations per control 95group and enforces the controller limit at reservation time and at the fault of 96HugeTLB memory for which no reservation exists. Since reservation limits are 97enforced at reservation time (on mmap or shget), reservation limits never causes 98the application to get SIGBUS signal if the memory was reserved before hand. For 99MAP_NORESERVE allocations, the reservation limit behaves the same as the fault 100limit, enforcing memory usage at fault time and causing the application to 101receive a SIGBUS if it's crossing its limit. 102 103Reservation limits are superior to page fault limits described above, since 104reservation limits are enforced at reservation time (on mmap or shget), and 105never causes the application to get SIGBUS signal if the memory was reserved 106before hand. This allows for easier fallback to alternatives such as 107non-HugeTLB memory for example. In the case of page fault accounting, it's very 108hard to avoid processes getting SIGBUS since the sysadmin needs precisely know 109the HugeTLB usage of all the tasks in the system and make sure there is enough 110pages to satisfy all requests. Avoiding tasks getting SIGBUS on overcommited 111systems is practically impossible with page fault accounting. 112 113 1143. Caveats with shared memory 115 116For shared HugeTLB memory, both HugeTLB reservation and page faults are charged 117to the first task that causes the memory to be reserved or faulted, and all 118subsequent uses of this reserved or faulted memory is done without charging. 119 120Shared HugeTLB memory is only uncharged when it is unreserved or deallocated. 121This is usually when the HugeTLB file is deleted, and not when the task that 122caused the reservation or fault has exited. 123 124 1254. Caveats with HugeTLB cgroup offline. 126 127When a HugeTLB cgroup goes offline with some reservations or faults still 128charged to it, the behavior is as follows: 129 130- The fault charges are charged to the parent HugeTLB cgroup (reparented), 131- the reservation charges remain on the offline HugeTLB cgroup. 132 133This means that if a HugeTLB cgroup gets offlined while there is still HugeTLB 134reservations charged to it, that cgroup persists as a zombie until all HugeTLB 135reservations are uncharged. HugeTLB reservations behave in this manner to match 136the memory controller whose cgroups also persist as zombie until all charged 137memory is uncharged. Also, the tracking of HugeTLB reservations is a bit more 138complex compared to the tracking of HugeTLB faults, so it is significantly 139harder to reparent reservations at offline time. 140