mm/hugetlb_cgroup.c

   1 /*
   2  *
   3  * Copyright IBM Corporation, 2012
   4  * Author Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of version 2.1 of the GNU Lesser General Public License
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it would be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  13  *
  14  */
  15
  16 #include <linux/cgroup.h>
  17 #include <linux/slab.h>
  18 #include <linux/hugetlb.h>
  19 #include <linux/hugetlb_cgroup.h>
  20
  21 struct hugetlb_cgroup {
  22         struct cgroup_subsys_state css;
  23         /*
  24          * the counter to account for hugepages from hugetlb.
  25          */
  26         struct res_counter hugepage[HUGE_MAX_HSTATE];
  27 };
  28
  29 struct cgroup_subsys hugetlb_subsys __read_mostly;
  30 static struct hugetlb_cgroup *root_h_cgroup __read_mostly;
  31
  32 static inline
  33 struct hugetlb_cgroup *hugetlb_cgroup_from_css(struct cgroup_subsys_state *s)
  34 {
  35         return container_of(s, struct hugetlb_cgroup, css);
  36 }
  37
  38 static inline
  39 struct hugetlb_cgroup *hugetlb_cgroup_from_cgroup(struct cgroup *cgroup)
  40 {
  41         return hugetlb_cgroup_from_css(cgroup_subsys_state(cgroup,
  42                                                            hugetlb_subsys_id));
  43 }
  44
  45 static inline
  46 struct hugetlb_cgroup *hugetlb_cgroup_from_task(struct task_struct *task)
  47 {
  48         return hugetlb_cgroup_from_css(task_subsys_state(task,
  49                                                          hugetlb_subsys_id));
  50 }
  51
  52 static inline bool hugetlb_cgroup_is_root(struct hugetlb_cgroup *h_cg)
  53 {
  54         return (h_cg == root_h_cgroup);
  55 }
  56
  57 static inline struct hugetlb_cgroup *parent_hugetlb_cgroup(struct cgroup *cg)
  58 {
  59         if (!cg->parent)
  60                 return NULL;
  61         return hugetlb_cgroup_from_cgroup(cg->parent);
  62 }
  63
  64 static inline bool hugetlb_cgroup_have_usage(struct cgroup *cg)
  65 {
  66         int idx;
  67         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_cgroup(cg);
  68
  69         for (idx = 0; idx < hugetlb_max_hstate; idx++) {
  70                 if ((res_counter_read_u64(&h_cg->hugepage[idx], RES_USAGE)) > 0)
  71                         return true;
  72         }
  73         return false;
  74 }
  75
  76 static struct cgroup_subsys_state *hugetlb_cgroup_create(struct cgroup *cgroup)
  77 {
  78         int idx;
  79         struct cgroup *parent_cgroup;
  80         struct hugetlb_cgroup *h_cgroup, *parent_h_cgroup;
  81
  82         h_cgroup = kzalloc(sizeof(*h_cgroup), GFP_KERNEL);
  83         if (!h_cgroup)
  84                 return ERR_PTR(-ENOMEM);
  85
  86         parent_cgroup = cgroup->parent;
  87         if (parent_cgroup) {
  88                 parent_h_cgroup = hugetlb_cgroup_from_cgroup(parent_cgroup);
  89                 for (idx = 0; idx < HUGE_MAX_HSTATE; idx++)
  90                         res_counter_init(&h_cgroup->hugepage[idx],
  91                                          &parent_h_cgroup->hugepage[idx]);
  92         } else {
  93                 root_h_cgroup = h_cgroup;
  94                 for (idx = 0; idx < HUGE_MAX_HSTATE; idx++)
  95                         res_counter_init(&h_cgroup->hugepage[idx], NULL);
  96         }
  97         return &h_cgroup->css;
  98 }
  99
 100 static void hugetlb_cgroup_destroy(struct cgroup *cgroup)
 101 {
 102         struct hugetlb_cgroup *h_cgroup;
 103
 104         h_cgroup = hugetlb_cgroup_from_cgroup(cgroup);
 105         kfree(h_cgroup);
 106 }
 107
 108 static int hugetlb_cgroup_pre_destroy(struct cgroup *cgroup)
 109 {
 110         /* We will add the cgroup removal support in later patches */
 111            return -EBUSY;
 112 }
 113
 114 int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
 115                                  struct hugetlb_cgroup **ptr)
 116 {
 117         int ret = 0;
 118         struct res_counter *fail_res;
 119         struct hugetlb_cgroup *h_cg = NULL;
 120         unsigned long csize = nr_pages * PAGE_SIZE;
 121
 122         if (hugetlb_cgroup_disabled())
 123                 goto done;
 124         /*
 125          * We don't charge any cgroup if the compound page have less
 126          * than 3 pages.
 127          */
 128         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
 129                 goto done;
 130 again:
 131         rcu_read_lock();
 132         h_cg = hugetlb_cgroup_from_task(current);
 133         if (!css_tryget(&h_cg->css)) {
 134                 rcu_read_unlock();
 135                 goto again;
 136         }
 137         rcu_read_unlock();
 138
 139         ret = res_counter_charge(&h_cg->hugepage[idx], csize, &fail_res);
 140         css_put(&h_cg->css);
 141 done:
 142         *ptr = h_cg;
 143         return ret;
 144 }
 145
 146 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 147                                   struct hugetlb_cgroup *h_cg,
 148                                   struct page *page)
 149 {
 150         if (hugetlb_cgroup_disabled() || !h_cg)
 151                 return;
 152
 153         spin_lock(&hugetlb_lock);
 154         set_hugetlb_cgroup(page, h_cg);
 155         spin_unlock(&hugetlb_lock);
 156         return;
 157 }
 158
 159 /*
 160  * Should be called with hugetlb_lock held
 161  */
 162 void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
 163                                   struct page *page)
 164 {
 165         struct hugetlb_cgroup *h_cg;
 166         unsigned long csize = nr_pages * PAGE_SIZE;
 167
 168         if (hugetlb_cgroup_disabled())
 169                 return;
 170         VM_BUG_ON(!spin_is_locked(&hugetlb_lock));
 171         h_cg = hugetlb_cgroup_from_page(page);
 172         if (unlikely(!h_cg))
 173                 return;
 174         set_hugetlb_cgroup(page, NULL);
 175         res_counter_uncharge(&h_cg->hugepage[idx], csize);
 176         return;
 177 }
 178
 179 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 180                                     struct hugetlb_cgroup *h_cg)
 181 {
 182         unsigned long csize = nr_pages * PAGE_SIZE;
 183
 184         if (hugetlb_cgroup_disabled() || !h_cg)
 185                 return;
 186
 187         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
 188                 return;
 189
 190         res_counter_uncharge(&h_cg->hugepage[idx], csize);
 191         return;
 192 }
 193
 194 struct cgroup_subsys hugetlb_subsys = {
 195         .name = "hugetlb",
 196         .create     = hugetlb_cgroup_create,
 197         .pre_destroy = hugetlb_cgroup_pre_destroy,
 198         .destroy    = hugetlb_cgroup_destroy,
 199         .subsys_id  = hugetlb_subsys_id,
 200 };