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  * Cgroup v2
   7  * Copyright (C) 2019 Red Hat, Inc.
   8  * Author: Giuseppe Scrivano <gscrivan@redhat.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify it
  11  * under the terms of version 2.1 of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it would be useful, but
  15  * WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  17  *
  18  */
  19
  20 #include <linux/cgroup.h>
  21 #include <linux/page_counter.h>
  22 #include <linux/slab.h>
  23 #include <linux/hugetlb.h>
  24 #include <linux/hugetlb_cgroup.h>
  25
  26 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
  27 #define MEMFILE_IDX(val)        (((val) >> 16) & 0xffff)
  28 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
  29
  30 static struct hugetlb_cgroup *root_h_cgroup __read_mostly;
  31
  32 static inline struct page_counter *
  33 __hugetlb_cgroup_counter_from_cgroup(struct hugetlb_cgroup *h_cg, int idx,
  34                                      bool rsvd)
  35 {
  36         if (rsvd)
  37                 return &h_cg->rsvd_hugepage[idx];
  38         return &h_cg->hugepage[idx];
  39 }
  40
  41 static inline struct page_counter *
  42 hugetlb_cgroup_counter_from_cgroup(struct hugetlb_cgroup *h_cg, int idx)
  43 {
  44         return __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, false);
  45 }
  46
  47 static inline struct page_counter *
  48 hugetlb_cgroup_counter_from_cgroup_rsvd(struct hugetlb_cgroup *h_cg, int idx)
  49 {
  50         return __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, true);
  51 }
  52
  53 static inline
  54 struct hugetlb_cgroup *hugetlb_cgroup_from_css(struct cgroup_subsys_state *s)
  55 {
  56         return s ? container_of(s, struct hugetlb_cgroup, css) : NULL;
  57 }
  58
  59 static inline
  60 struct hugetlb_cgroup *hugetlb_cgroup_from_task(struct task_struct *task)
  61 {
  62         return hugetlb_cgroup_from_css(task_css(task, hugetlb_cgrp_id));
  63 }
  64
  65 static inline bool hugetlb_cgroup_is_root(struct hugetlb_cgroup *h_cg)
  66 {
  67         return (h_cg == root_h_cgroup);
  68 }
  69
  70 static inline struct hugetlb_cgroup *
  71 parent_hugetlb_cgroup(struct hugetlb_cgroup *h_cg)
  72 {
  73         return hugetlb_cgroup_from_css(h_cg->css.parent);
  74 }
  75
  76 static inline bool hugetlb_cgroup_have_usage(struct hugetlb_cgroup *h_cg)
  77 {
  78         struct hstate *h;
  79
  80         for_each_hstate(h) {
  81                 if (page_counter_read(
  82                     hugetlb_cgroup_counter_from_cgroup(h_cg, hstate_index(h))))
  83                         return true;
  84         }
  85         return false;
  86 }
  87
  88 static void hugetlb_cgroup_init(struct hugetlb_cgroup *h_cgroup,
  89                                 struct hugetlb_cgroup *parent_h_cgroup)
  90 {
  91         int idx;
  92
  93         for (idx = 0; idx < HUGE_MAX_HSTATE; idx++) {
  94                 struct page_counter *fault_parent = NULL;
  95                 struct page_counter *rsvd_parent = NULL;
  96                 unsigned long limit;
  97                 int ret;
  98
  99                 if (parent_h_cgroup) {
 100                         fault_parent = hugetlb_cgroup_counter_from_cgroup(
 101                                 parent_h_cgroup, idx);
 102                         rsvd_parent = hugetlb_cgroup_counter_from_cgroup_rsvd(
 103                                 parent_h_cgroup, idx);
 104                 }
 105                 page_counter_init(hugetlb_cgroup_counter_from_cgroup(h_cgroup,
 106                                                                      idx),
 107                                   fault_parent);
 108                 page_counter_init(
 109                         hugetlb_cgroup_counter_from_cgroup_rsvd(h_cgroup, idx),
 110                         rsvd_parent);
 111
 112                 limit = round_down(PAGE_COUNTER_MAX,
 113                                    pages_per_huge_page(&hstates[idx]));
 114
 115                 ret = page_counter_set_max(
 116                         hugetlb_cgroup_counter_from_cgroup(h_cgroup, idx),
 117                         limit);
 118                 VM_BUG_ON(ret);
 119                 ret = page_counter_set_max(
 120                         hugetlb_cgroup_counter_from_cgroup_rsvd(h_cgroup, idx),
 121                         limit);
 122                 VM_BUG_ON(ret);
 123         }
 124 }
 125
 126 static void hugetlb_cgroup_free(struct hugetlb_cgroup *h_cgroup)
 127 {
 128         int node;
 129
 130         for_each_node(node)
 131                 kfree(h_cgroup->nodeinfo[node]);
 132         kfree(h_cgroup);
 133 }
 134
 135 static struct cgroup_subsys_state *
 136 hugetlb_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
 137 {
 138         struct hugetlb_cgroup *parent_h_cgroup = hugetlb_cgroup_from_css(parent_css);
 139         struct hugetlb_cgroup *h_cgroup;
 140         int node;
 141
 142         h_cgroup = kzalloc(struct_size(h_cgroup, nodeinfo, nr_node_ids),
 143                            GFP_KERNEL);
 144
 145         if (!h_cgroup)
 146                 return ERR_PTR(-ENOMEM);
 147
 148         if (!parent_h_cgroup)
 149                 root_h_cgroup = h_cgroup;
 150
 151         /*
 152          * TODO: this routine can waste much memory for nodes which will
 153          * never be onlined. It's better to use memory hotplug callback
 154          * function.
 155          */
 156         for_each_node(node) {
 157                 /* Set node_to_alloc to NUMA_NO_NODE for offline nodes. */
 158                 int node_to_alloc =
 159                         node_state(node, N_NORMAL_MEMORY) ? node : NUMA_NO_NODE;
 160                 h_cgroup->nodeinfo[node] =
 161                         kzalloc_node(sizeof(struct hugetlb_cgroup_per_node),
 162                                      GFP_KERNEL, node_to_alloc);
 163                 if (!h_cgroup->nodeinfo[node])
 164                         goto fail_alloc_nodeinfo;
 165         }
 166
 167         hugetlb_cgroup_init(h_cgroup, parent_h_cgroup);
 168         return &h_cgroup->css;
 169
 170 fail_alloc_nodeinfo:
 171         hugetlb_cgroup_free(h_cgroup);
 172         return ERR_PTR(-ENOMEM);
 173 }
 174
 175 static void hugetlb_cgroup_css_free(struct cgroup_subsys_state *css)
 176 {
 177         hugetlb_cgroup_free(hugetlb_cgroup_from_css(css));
 178 }
 179
 180 /*
 181  * Should be called with hugetlb_lock held.
 182  * Since we are holding hugetlb_lock, pages cannot get moved from
 183  * active list or uncharged from the cgroup, So no need to get
 184  * page reference and test for page active here. This function
 185  * cannot fail.
 186  */
 187 static void hugetlb_cgroup_move_parent(int idx, struct hugetlb_cgroup *h_cg,
 188                                        struct page *page)
 189 {
 190         unsigned int nr_pages;
 191         struct page_counter *counter;
 192         struct hugetlb_cgroup *page_hcg;
 193         struct hugetlb_cgroup *parent = parent_hugetlb_cgroup(h_cg);
 194
 195         page_hcg = hugetlb_cgroup_from_page(page);
 196         /*
 197          * We can have pages in active list without any cgroup
 198          * ie, hugepage with less than 3 pages. We can safely
 199          * ignore those pages.
 200          */
 201         if (!page_hcg || page_hcg != h_cg)
 202                 goto out;
 203
 204         nr_pages = compound_nr(page);
 205         if (!parent) {
 206                 parent = root_h_cgroup;
 207                 /* root has no limit */
 208                 page_counter_charge(&parent->hugepage[idx], nr_pages);
 209         }
 210         counter = &h_cg->hugepage[idx];
 211         /* Take the pages off the local counter */
 212         page_counter_cancel(counter, nr_pages);
 213
 214         set_hugetlb_cgroup(page, parent);
 215 out:
 216         return;
 217 }
 218
 219 /*
 220  * Force the hugetlb cgroup to empty the hugetlb resources by moving them to
 221  * the parent cgroup.
 222  */
 223 static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css)
 224 {
 225         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
 226         struct hstate *h;
 227         struct page *page;
 228
 229         do {
 230                 for_each_hstate(h) {
 231                         spin_lock_irq(&hugetlb_lock);
 232                         list_for_each_entry(page, &h->hugepage_activelist, lru)
 233                                 hugetlb_cgroup_move_parent(hstate_index(h), h_cg, page);
 234
 235                         spin_unlock_irq(&hugetlb_lock);
 236                 }
 237                 cond_resched();
 238         } while (hugetlb_cgroup_have_usage(h_cg));
 239 }
 240
 241 static inline void hugetlb_event(struct hugetlb_cgroup *hugetlb, int idx,
 242                                  enum hugetlb_memory_event event)
 243 {
 244         atomic_long_inc(&hugetlb->events_local[idx][event]);
 245         cgroup_file_notify(&hugetlb->events_local_file[idx]);
 246
 247         do {
 248                 atomic_long_inc(&hugetlb->events[idx][event]);
 249                 cgroup_file_notify(&hugetlb->events_file[idx]);
 250         } while ((hugetlb = parent_hugetlb_cgroup(hugetlb)) &&
 251                  !hugetlb_cgroup_is_root(hugetlb));
 252 }
 253
 254 static int __hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
 255                                           struct hugetlb_cgroup **ptr,
 256                                           bool rsvd)
 257 {
 258         int ret = 0;
 259         struct page_counter *counter;
 260         struct hugetlb_cgroup *h_cg = NULL;
 261
 262         if (hugetlb_cgroup_disabled())
 263                 goto done;
 264         /*
 265          * We don't charge any cgroup if the compound page have less
 266          * than 3 pages.
 267          */
 268         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
 269                 goto done;
 270 again:
 271         rcu_read_lock();
 272         h_cg = hugetlb_cgroup_from_task(current);
 273         if (!css_tryget(&h_cg->css)) {
 274                 rcu_read_unlock();
 275                 goto again;
 276         }
 277         rcu_read_unlock();
 278
 279         if (!page_counter_try_charge(
 280                     __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
 281                     nr_pages, &counter)) {
 282                 ret = -ENOMEM;
 283                 hugetlb_event(h_cg, idx, HUGETLB_MAX);
 284                 css_put(&h_cg->css);
 285                 goto done;
 286         }
 287         /* Reservations take a reference to the css because they do not get
 288          * reparented.
 289          */
 290         if (!rsvd)
 291                 css_put(&h_cg->css);
 292 done:
 293         *ptr = h_cg;
 294         return ret;
 295 }
 296
 297 int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
 298                                  struct hugetlb_cgroup **ptr)
 299 {
 300         return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, false);
 301 }
 302
 303 int hugetlb_cgroup_charge_cgroup_rsvd(int idx, unsigned long nr_pages,
 304                                       struct hugetlb_cgroup **ptr)
 305 {
 306         return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, true);
 307 }
 308
 309 /* Should be called with hugetlb_lock held */
 310 static void __hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 311                                            struct hugetlb_cgroup *h_cg,
 312                                            struct page *page, bool rsvd)
 313 {
 314         if (hugetlb_cgroup_disabled() || !h_cg)
 315                 return;
 316
 317         __set_hugetlb_cgroup(page, h_cg, rsvd);
 318         if (!rsvd) {
 319                 unsigned long usage =
 320                         h_cg->nodeinfo[page_to_nid(page)]->usage[idx];
 321                 /*
 322                  * This write is not atomic due to fetching usage and writing
 323                  * to it, but that's fine because we call this with
 324                  * hugetlb_lock held anyway.
 325                  */
 326                 WRITE_ONCE(h_cg->nodeinfo[page_to_nid(page)]->usage[idx],
 327                            usage + nr_pages);
 328         }
 329 }
 330
 331 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 332                                   struct hugetlb_cgroup *h_cg,
 333                                   struct page *page)
 334 {
 335         __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, page, false);
 336 }
 337
 338 void hugetlb_cgroup_commit_charge_rsvd(int idx, unsigned long nr_pages,
 339                                        struct hugetlb_cgroup *h_cg,
 340                                        struct page *page)
 341 {
 342         __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, page, true);
 343 }
 344
 345 /*
 346  * Should be called with hugetlb_lock held
 347  */
 348 static void __hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
 349                                            struct page *page, bool rsvd)
 350 {
 351         struct hugetlb_cgroup *h_cg;
 352
 353         if (hugetlb_cgroup_disabled())
 354                 return;
 355         lockdep_assert_held(&hugetlb_lock);
 356         h_cg = __hugetlb_cgroup_from_page(page, rsvd);
 357         if (unlikely(!h_cg))
 358                 return;
 359         __set_hugetlb_cgroup(page, NULL, rsvd);
 360
 361         page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx,
 362                                                                    rsvd),
 363                               nr_pages);
 364
 365         if (rsvd)
 366                 css_put(&h_cg->css);
 367         else {
 368                 unsigned long usage =
 369                         h_cg->nodeinfo[page_to_nid(page)]->usage[idx];
 370                 /*
 371                  * This write is not atomic due to fetching usage and writing
 372                  * to it, but that's fine because we call this with
 373                  * hugetlb_lock held anyway.
 374                  */
 375                 WRITE_ONCE(h_cg->nodeinfo[page_to_nid(page)]->usage[idx],
 376                            usage - nr_pages);
 377         }
 378 }
 379
 380 void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
 381                                   struct page *page)
 382 {
 383         __hugetlb_cgroup_uncharge_page(idx, nr_pages, page, false);
 384 }
 385
 386 void hugetlb_cgroup_uncharge_page_rsvd(int idx, unsigned long nr_pages,
 387                                        struct page *page)
 388 {
 389         __hugetlb_cgroup_uncharge_page(idx, nr_pages, page, true);
 390 }
 391
 392 static void __hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 393                                              struct hugetlb_cgroup *h_cg,
 394                                              bool rsvd)
 395 {
 396         if (hugetlb_cgroup_disabled() || !h_cg)
 397                 return;
 398
 399         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
 400                 return;
 401
 402         page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx,
 403                                                                    rsvd),
 404                               nr_pages);
 405
 406         if (rsvd)
 407                 css_put(&h_cg->css);
 408 }
 409
 410 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 411                                     struct hugetlb_cgroup *h_cg)
 412 {
 413         __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, false);
 414 }
 415
 416 void hugetlb_cgroup_uncharge_cgroup_rsvd(int idx, unsigned long nr_pages,
 417                                          struct hugetlb_cgroup *h_cg)
 418 {
 419         __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, true);
 420 }
 421
 422 void hugetlb_cgroup_uncharge_counter(struct resv_map *resv, unsigned long start,
 423                                      unsigned long end)
 424 {
 425         if (hugetlb_cgroup_disabled() || !resv || !resv->reservation_counter ||
 426             !resv->css)
 427                 return;
 428
 429         page_counter_uncharge(resv->reservation_counter,
 430                               (end - start) * resv->pages_per_hpage);
 431         css_put(resv->css);
 432 }
 433
 434 void hugetlb_cgroup_uncharge_file_region(struct resv_map *resv,
 435                                          struct file_region *rg,
 436                                          unsigned long nr_pages,
 437                                          bool region_del)
 438 {
 439         if (hugetlb_cgroup_disabled() || !resv || !rg || !nr_pages)
 440                 return;
 441
 442         if (rg->reservation_counter && resv->pages_per_hpage &&
 443             !resv->reservation_counter) {
 444                 page_counter_uncharge(rg->reservation_counter,
 445                                       nr_pages * resv->pages_per_hpage);
 446                 /*
 447                  * Only do css_put(rg->css) when we delete the entire region
 448                  * because one file_region must hold exactly one css reference.
 449                  */
 450                 if (region_del)
 451                         css_put(rg->css);
 452         }
 453 }
 454
 455 enum {
 456         RES_USAGE,
 457         RES_RSVD_USAGE,
 458         RES_LIMIT,
 459         RES_RSVD_LIMIT,
 460         RES_MAX_USAGE,
 461         RES_RSVD_MAX_USAGE,
 462         RES_FAILCNT,
 463         RES_RSVD_FAILCNT,
 464 };
 465
 466 static int hugetlb_cgroup_read_numa_stat(struct seq_file *seq, void *dummy)
 467 {
 468         int nid;
 469         struct cftype *cft = seq_cft(seq);
 470         int idx = MEMFILE_IDX(cft->private);
 471         bool legacy = MEMFILE_ATTR(cft->private);
 472         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 473         struct cgroup_subsys_state *css;
 474         unsigned long usage;
 475
 476         if (legacy) {
 477                 /* Add up usage across all nodes for the non-hierarchical total. */
 478                 usage = 0;
 479                 for_each_node_state(nid, N_MEMORY)
 480                         usage += READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]);
 481                 seq_printf(seq, "total=%lu", usage * PAGE_SIZE);
 482
 483                 /* Simply print the per-node usage for the non-hierarchical total. */
 484                 for_each_node_state(nid, N_MEMORY)
 485                         seq_printf(seq, " N%d=%lu", nid,
 486                                    READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]) *
 487                                            PAGE_SIZE);
 488                 seq_putc(seq, '\n');
 489         }
 490
 491         /*
 492          * The hierarchical total is pretty much the value recorded by the
 493          * counter, so use that.
 494          */
 495         seq_printf(seq, "%stotal=%lu", legacy ? "hierarchical_" : "",
 496                    page_counter_read(&h_cg->hugepage[idx]) * PAGE_SIZE);
 497
 498         /*
 499          * For each node, transverse the css tree to obtain the hierarchical
 500          * node usage.
 501          */
 502         for_each_node_state(nid, N_MEMORY) {
 503                 usage = 0;
 504                 rcu_read_lock();
 505                 css_for_each_descendant_pre(css, &h_cg->css) {
 506                         usage += READ_ONCE(hugetlb_cgroup_from_css(css)
 507                                                    ->nodeinfo[nid]
 508                                                    ->usage[idx]);
 509                 }
 510                 rcu_read_unlock();
 511                 seq_printf(seq, " N%d=%lu", nid, usage * PAGE_SIZE);
 512         }
 513
 514         seq_putc(seq, '\n');
 515
 516         return 0;
 517 }
 518
 519 static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css,
 520                                    struct cftype *cft)
 521 {
 522         struct page_counter *counter;
 523         struct page_counter *rsvd_counter;
 524         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
 525
 526         counter = &h_cg->hugepage[MEMFILE_IDX(cft->private)];
 527         rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(cft->private)];
 528
 529         switch (MEMFILE_ATTR(cft->private)) {
 530         case RES_USAGE:
 531                 return (u64)page_counter_read(counter) * PAGE_SIZE;
 532         case RES_RSVD_USAGE:
 533                 return (u64)page_counter_read(rsvd_counter) * PAGE_SIZE;
 534         case RES_LIMIT:
 535                 return (u64)counter->max * PAGE_SIZE;
 536         case RES_RSVD_LIMIT:
 537                 return (u64)rsvd_counter->max * PAGE_SIZE;
 538         case RES_MAX_USAGE:
 539                 return (u64)counter->watermark * PAGE_SIZE;
 540         case RES_RSVD_MAX_USAGE:
 541                 return (u64)rsvd_counter->watermark * PAGE_SIZE;
 542         case RES_FAILCNT:
 543                 return counter->failcnt;
 544         case RES_RSVD_FAILCNT:
 545                 return rsvd_counter->failcnt;
 546         default:
 547                 BUG();
 548         }
 549 }
 550
 551 static int hugetlb_cgroup_read_u64_max(struct seq_file *seq, void *v)
 552 {
 553         int idx;
 554         u64 val;
 555         struct cftype *cft = seq_cft(seq);
 556         unsigned long limit;
 557         struct page_counter *counter;
 558         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 559
 560         idx = MEMFILE_IDX(cft->private);
 561         counter = &h_cg->hugepage[idx];
 562
 563         limit = round_down(PAGE_COUNTER_MAX,
 564                            pages_per_huge_page(&hstates[idx]));
 565
 566         switch (MEMFILE_ATTR(cft->private)) {
 567         case RES_RSVD_USAGE:
 568                 counter = &h_cg->rsvd_hugepage[idx];
 569                 fallthrough;
 570         case RES_USAGE:
 571                 val = (u64)page_counter_read(counter);
 572                 seq_printf(seq, "%llu\n", val * PAGE_SIZE);
 573                 break;
 574         case RES_RSVD_LIMIT:
 575                 counter = &h_cg->rsvd_hugepage[idx];
 576                 fallthrough;
 577         case RES_LIMIT:
 578                 val = (u64)counter->max;
 579                 if (val == limit)
 580                         seq_puts(seq, "max\n");
 581                 else
 582                         seq_printf(seq, "%llu\n", val * PAGE_SIZE);
 583                 break;
 584         default:
 585                 BUG();
 586         }
 587
 588         return 0;
 589 }
 590
 591 static DEFINE_MUTEX(hugetlb_limit_mutex);
 592
 593 static ssize_t hugetlb_cgroup_write(struct kernfs_open_file *of,
 594                                     char *buf, size_t nbytes, loff_t off,
 595                                     const char *max)
 596 {
 597         int ret, idx;
 598         unsigned long nr_pages;
 599         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
 600         bool rsvd = false;
 601
 602         if (hugetlb_cgroup_is_root(h_cg)) /* Can't set limit on root */
 603                 return -EINVAL;
 604
 605         buf = strstrip(buf);
 606         ret = page_counter_memparse(buf, max, &nr_pages);
 607         if (ret)
 608                 return ret;
 609
 610         idx = MEMFILE_IDX(of_cft(of)->private);
 611         nr_pages = round_down(nr_pages, pages_per_huge_page(&hstates[idx]));
 612
 613         switch (MEMFILE_ATTR(of_cft(of)->private)) {
 614         case RES_RSVD_LIMIT:
 615                 rsvd = true;
 616                 fallthrough;
 617         case RES_LIMIT:
 618                 mutex_lock(&hugetlb_limit_mutex);
 619                 ret = page_counter_set_max(
 620                         __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
 621                         nr_pages);
 622                 mutex_unlock(&hugetlb_limit_mutex);
 623                 break;
 624         default:
 625                 ret = -EINVAL;
 626                 break;
 627         }
 628         return ret ?: nbytes;
 629 }
 630
 631 static ssize_t hugetlb_cgroup_write_legacy(struct kernfs_open_file *of,
 632                                            char *buf, size_t nbytes, loff_t off)
 633 {
 634         return hugetlb_cgroup_write(of, buf, nbytes, off, "-1");
 635 }
 636
 637 static ssize_t hugetlb_cgroup_write_dfl(struct kernfs_open_file *of,
 638                                         char *buf, size_t nbytes, loff_t off)
 639 {
 640         return hugetlb_cgroup_write(of, buf, nbytes, off, "max");
 641 }
 642
 643 static ssize_t hugetlb_cgroup_reset(struct kernfs_open_file *of,
 644                                     char *buf, size_t nbytes, loff_t off)
 645 {
 646         int ret = 0;
 647         struct page_counter *counter, *rsvd_counter;
 648         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
 649
 650         counter = &h_cg->hugepage[MEMFILE_IDX(of_cft(of)->private)];
 651         rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(of_cft(of)->private)];
 652
 653         switch (MEMFILE_ATTR(of_cft(of)->private)) {
 654         case RES_MAX_USAGE:
 655                 page_counter_reset_watermark(counter);
 656                 break;
 657         case RES_RSVD_MAX_USAGE:
 658                 page_counter_reset_watermark(rsvd_counter);
 659                 break;
 660         case RES_FAILCNT:
 661                 counter->failcnt = 0;
 662                 break;
 663         case RES_RSVD_FAILCNT:
 664                 rsvd_counter->failcnt = 0;
 665                 break;
 666         default:
 667                 ret = -EINVAL;
 668                 break;
 669         }
 670         return ret ?: nbytes;
 671 }
 672
 673 static char *mem_fmt(char *buf, int size, unsigned long hsize)
 674 {
 675         if (hsize >= SZ_1G)
 676                 snprintf(buf, size, "%luGB", hsize / SZ_1G);
 677         else if (hsize >= SZ_1M)
 678                 snprintf(buf, size, "%luMB", hsize / SZ_1M);
 679         else
 680                 snprintf(buf, size, "%luKB", hsize / SZ_1K);
 681         return buf;
 682 }
 683
 684 static int __hugetlb_events_show(struct seq_file *seq, bool local)
 685 {
 686         int idx;
 687         long max;
 688         struct cftype *cft = seq_cft(seq);
 689         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 690
 691         idx = MEMFILE_IDX(cft->private);
 692
 693         if (local)
 694                 max = atomic_long_read(&h_cg->events_local[idx][HUGETLB_MAX]);
 695         else
 696                 max = atomic_long_read(&h_cg->events[idx][HUGETLB_MAX]);
 697
 698         seq_printf(seq, "max %lu\n", max);
 699
 700         return 0;
 701 }
 702
 703 static int hugetlb_events_show(struct seq_file *seq, void *v)
 704 {
 705         return __hugetlb_events_show(seq, false);
 706 }
 707
 708 static int hugetlb_events_local_show(struct seq_file *seq, void *v)
 709 {
 710         return __hugetlb_events_show(seq, true);
 711 }
 712
 713 static void __init __hugetlb_cgroup_file_dfl_init(int idx)
 714 {
 715         char buf[32];
 716         struct cftype *cft;
 717         struct hstate *h = &hstates[idx];
 718
 719         /* format the size */
 720         mem_fmt(buf, sizeof(buf), huge_page_size(h));
 721
 722         /* Add the limit file */
 723         cft = &h->cgroup_files_dfl[0];
 724         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max", buf);
 725         cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
 726         cft->seq_show = hugetlb_cgroup_read_u64_max;
 727         cft->write = hugetlb_cgroup_write_dfl;
 728         cft->flags = CFTYPE_NOT_ON_ROOT;
 729
 730         /* Add the reservation limit file */
 731         cft = &h->cgroup_files_dfl[1];
 732         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max", buf);
 733         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT);
 734         cft->seq_show = hugetlb_cgroup_read_u64_max;
 735         cft->write = hugetlb_cgroup_write_dfl;
 736         cft->flags = CFTYPE_NOT_ON_ROOT;
 737
 738         /* Add the current usage file */
 739         cft = &h->cgroup_files_dfl[2];
 740         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.current", buf);
 741         cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
 742         cft->seq_show = hugetlb_cgroup_read_u64_max;
 743         cft->flags = CFTYPE_NOT_ON_ROOT;
 744
 745         /* Add the current reservation usage file */
 746         cft = &h->cgroup_files_dfl[3];
 747         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.current", buf);
 748         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
 749         cft->seq_show = hugetlb_cgroup_read_u64_max;
 750         cft->flags = CFTYPE_NOT_ON_ROOT;
 751
 752         /* Add the events file */
 753         cft = &h->cgroup_files_dfl[4];
 754         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events", buf);
 755         cft->private = MEMFILE_PRIVATE(idx, 0);
 756         cft->seq_show = hugetlb_events_show;
 757         cft->file_offset = offsetof(struct hugetlb_cgroup, events_file[idx]);
 758         cft->flags = CFTYPE_NOT_ON_ROOT;
 759
 760         /* Add the events.local file */
 761         cft = &h->cgroup_files_dfl[5];
 762         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events.local", buf);
 763         cft->private = MEMFILE_PRIVATE(idx, 0);
 764         cft->seq_show = hugetlb_events_local_show;
 765         cft->file_offset = offsetof(struct hugetlb_cgroup,
 766                                     events_local_file[idx]);
 767         cft->flags = CFTYPE_NOT_ON_ROOT;
 768
 769         /* Add the numa stat file */
 770         cft = &h->cgroup_files_dfl[6];
 771         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
 772         cft->private = MEMFILE_PRIVATE(idx, 0);
 773         cft->seq_show = hugetlb_cgroup_read_numa_stat;
 774         cft->flags = CFTYPE_NOT_ON_ROOT;
 775
 776         /* NULL terminate the last cft */
 777         cft = &h->cgroup_files_dfl[7];
 778         memset(cft, 0, sizeof(*cft));
 779
 780         WARN_ON(cgroup_add_dfl_cftypes(&hugetlb_cgrp_subsys,
 781                                        h->cgroup_files_dfl));
 782 }
 783
 784 static void __init __hugetlb_cgroup_file_legacy_init(int idx)
 785 {
 786         char buf[32];
 787         struct cftype *cft;
 788         struct hstate *h = &hstates[idx];
 789
 790         /* format the size */
 791         mem_fmt(buf, sizeof(buf), huge_page_size(h));
 792
 793         /* Add the limit file */
 794         cft = &h->cgroup_files_legacy[0];
 795         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf);
 796         cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
 797         cft->read_u64 = hugetlb_cgroup_read_u64;
 798         cft->write = hugetlb_cgroup_write_legacy;
 799
 800         /* Add the reservation limit file */
 801         cft = &h->cgroup_files_legacy[1];
 802         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.limit_in_bytes", buf);
 803         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT);
 804         cft->read_u64 = hugetlb_cgroup_read_u64;
 805         cft->write = hugetlb_cgroup_write_legacy;
 806
 807         /* Add the usage file */
 808         cft = &h->cgroup_files_legacy[2];
 809         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf);
 810         cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
 811         cft->read_u64 = hugetlb_cgroup_read_u64;
 812
 813         /* Add the reservation usage file */
 814         cft = &h->cgroup_files_legacy[3];
 815         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.usage_in_bytes", buf);
 816         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
 817         cft->read_u64 = hugetlb_cgroup_read_u64;
 818
 819         /* Add the MAX usage file */
 820         cft = &h->cgroup_files_legacy[4];
 821         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf);
 822         cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE);
 823         cft->write = hugetlb_cgroup_reset;
 824         cft->read_u64 = hugetlb_cgroup_read_u64;
 825
 826         /* Add the MAX reservation usage file */
 827         cft = &h->cgroup_files_legacy[5];
 828         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max_usage_in_bytes", buf);
 829         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_MAX_USAGE);
 830         cft->write = hugetlb_cgroup_reset;
 831         cft->read_u64 = hugetlb_cgroup_read_u64;
 832
 833         /* Add the failcntfile */
 834         cft = &h->cgroup_files_legacy[6];
 835         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf);
 836         cft->private = MEMFILE_PRIVATE(idx, RES_FAILCNT);
 837         cft->write = hugetlb_cgroup_reset;
 838         cft->read_u64 = hugetlb_cgroup_read_u64;
 839
 840         /* Add the reservation failcntfile */
 841         cft = &h->cgroup_files_legacy[7];
 842         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.failcnt", buf);
 843         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_FAILCNT);
 844         cft->write = hugetlb_cgroup_reset;
 845         cft->read_u64 = hugetlb_cgroup_read_u64;
 846
 847         /* Add the numa stat file */
 848         cft = &h->cgroup_files_legacy[8];
 849         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
 850         cft->private = MEMFILE_PRIVATE(idx, 1);
 851         cft->seq_show = hugetlb_cgroup_read_numa_stat;
 852
 853         /* NULL terminate the last cft */
 854         cft = &h->cgroup_files_legacy[9];
 855         memset(cft, 0, sizeof(*cft));
 856
 857         WARN_ON(cgroup_add_legacy_cftypes(&hugetlb_cgrp_subsys,
 858                                           h->cgroup_files_legacy));
 859 }
 860
 861 static void __init __hugetlb_cgroup_file_init(int idx)
 862 {
 863         __hugetlb_cgroup_file_dfl_init(idx);
 864         __hugetlb_cgroup_file_legacy_init(idx);
 865 }
 866
 867 void __init hugetlb_cgroup_file_init(void)
 868 {
 869         struct hstate *h;
 870
 871         for_each_hstate(h) {
 872                 /*
 873                  * Add cgroup control files only if the huge page consists
 874                  * of more than two normal pages. This is because we use
 875                  * page[2].private for storing cgroup details.
 876                  */
 877                 if (huge_page_order(h) >= HUGETLB_CGROUP_MIN_ORDER)
 878                         __hugetlb_cgroup_file_init(hstate_index(h));
 879         }
 880 }
 881
 882 /*
 883  * hugetlb_lock will make sure a parallel cgroup rmdir won't happen
 884  * when we migrate hugepages
 885  */
 886 void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
 887 {
 888         struct hugetlb_cgroup *h_cg;
 889         struct hugetlb_cgroup *h_cg_rsvd;
 890         struct hstate *h = page_hstate(oldhpage);
 891
 892         if (hugetlb_cgroup_disabled())
 893                 return;
 894
 895         spin_lock_irq(&hugetlb_lock);
 896         h_cg = hugetlb_cgroup_from_page(oldhpage);
 897         h_cg_rsvd = hugetlb_cgroup_from_page_rsvd(oldhpage);
 898         set_hugetlb_cgroup(oldhpage, NULL);
 899         set_hugetlb_cgroup_rsvd(oldhpage, NULL);
 900
 901         /* move the h_cg details to new cgroup */
 902         set_hugetlb_cgroup(newhpage, h_cg);
 903         set_hugetlb_cgroup_rsvd(newhpage, h_cg_rsvd);
 904         list_move(&newhpage->lru, &h->hugepage_activelist);
 905         spin_unlock_irq(&hugetlb_lock);
 906         return;
 907 }
 908
 909 static struct cftype hugetlb_files[] = {
 910         {} /* terminate */
 911 };
 912
 913 struct cgroup_subsys hugetlb_cgrp_subsys = {
 914         .css_alloc      = hugetlb_cgroup_css_alloc,
 915         .css_offline    = hugetlb_cgroup_css_offline,
 916         .css_free       = hugetlb_cgroup_css_free,
 917         .dfl_cftypes    = hugetlb_files,
 918         .legacy_cftypes = hugetlb_files,
 919 };