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         int idx;
  79
  80         for (idx = 0; idx < hugetlb_max_hstate; idx++) {
  81                 if (page_counter_read(
  82                                 hugetlb_cgroup_counter_from_cgroup(h_cg, idx)))
  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 -1 for offline nodes. */
 158                 int node_to_alloc =
 159                         node_state(node, N_NORMAL_MEMORY) ? node : -1;
 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         int idx;
 229
 230         do {
 231                 idx = 0;
 232                 for_each_hstate(h) {
 233                         spin_lock_irq(&hugetlb_lock);
 234                         list_for_each_entry(page, &h->hugepage_activelist, lru)
 235                                 hugetlb_cgroup_move_parent(idx, h_cg, page);
 236
 237                         spin_unlock_irq(&hugetlb_lock);
 238                         idx++;
 239                 }
 240                 cond_resched();
 241         } while (hugetlb_cgroup_have_usage(h_cg));
 242 }
 243
 244 static inline void hugetlb_event(struct hugetlb_cgroup *hugetlb, int idx,
 245                                  enum hugetlb_memory_event event)
 246 {
 247         atomic_long_inc(&hugetlb->events_local[idx][event]);
 248         cgroup_file_notify(&hugetlb->events_local_file[idx]);
 249
 250         do {
 251                 atomic_long_inc(&hugetlb->events[idx][event]);
 252                 cgroup_file_notify(&hugetlb->events_file[idx]);
 253         } while ((hugetlb = parent_hugetlb_cgroup(hugetlb)) &&
 254                  !hugetlb_cgroup_is_root(hugetlb));
 255 }
 256
 257 static int __hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
 258                                           struct hugetlb_cgroup **ptr,
 259                                           bool rsvd)
 260 {
 261         int ret = 0;
 262         struct page_counter *counter;
 263         struct hugetlb_cgroup *h_cg = NULL;
 264
 265         if (hugetlb_cgroup_disabled())
 266                 goto done;
 267         /*
 268          * We don't charge any cgroup if the compound page have less
 269          * than 3 pages.
 270          */
 271         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
 272                 goto done;
 273 again:
 274         rcu_read_lock();
 275         h_cg = hugetlb_cgroup_from_task(current);
 276         if (!css_tryget(&h_cg->css)) {
 277                 rcu_read_unlock();
 278                 goto again;
 279         }
 280         rcu_read_unlock();
 281
 282         if (!page_counter_try_charge(
 283                     __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
 284                     nr_pages, &counter)) {
 285                 ret = -ENOMEM;
 286                 hugetlb_event(h_cg, idx, HUGETLB_MAX);
 287                 css_put(&h_cg->css);
 288                 goto done;
 289         }
 290         /* Reservations take a reference to the css because they do not get
 291          * reparented.
 292          */
 293         if (!rsvd)
 294                 css_put(&h_cg->css);
 295 done:
 296         *ptr = h_cg;
 297         return ret;
 298 }
 299
 300 int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
 301                                  struct hugetlb_cgroup **ptr)
 302 {
 303         return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, false);
 304 }
 305
 306 int hugetlb_cgroup_charge_cgroup_rsvd(int idx, unsigned long nr_pages,
 307                                       struct hugetlb_cgroup **ptr)
 308 {
 309         return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, true);
 310 }
 311
 312 /* Should be called with hugetlb_lock held */
 313 static void __hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 314                                            struct hugetlb_cgroup *h_cg,
 315                                            struct page *page, bool rsvd)
 316 {
 317         if (hugetlb_cgroup_disabled() || !h_cg)
 318                 return;
 319
 320         __set_hugetlb_cgroup(page, h_cg, rsvd);
 321         if (!rsvd) {
 322                 unsigned long usage =
 323                         h_cg->nodeinfo[page_to_nid(page)]->usage[idx];
 324                 /*
 325                  * This write is not atomic due to fetching usage and writing
 326                  * to it, but that's fine because we call this with
 327                  * hugetlb_lock held anyway.
 328                  */
 329                 WRITE_ONCE(h_cg->nodeinfo[page_to_nid(page)]->usage[idx],
 330                            usage + nr_pages);
 331         }
 332 }
 333
 334 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 335                                   struct hugetlb_cgroup *h_cg,
 336                                   struct page *page)
 337 {
 338         __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, page, false);
 339 }
 340
 341 void hugetlb_cgroup_commit_charge_rsvd(int idx, unsigned long nr_pages,
 342                                        struct hugetlb_cgroup *h_cg,
 343                                        struct page *page)
 344 {
 345         __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, page, true);
 346 }
 347
 348 /*
 349  * Should be called with hugetlb_lock held
 350  */
 351 static void __hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
 352                                            struct page *page, bool rsvd)
 353 {
 354         struct hugetlb_cgroup *h_cg;
 355
 356         if (hugetlb_cgroup_disabled())
 357                 return;
 358         lockdep_assert_held(&hugetlb_lock);
 359         h_cg = __hugetlb_cgroup_from_page(page, rsvd);
 360         if (unlikely(!h_cg))
 361                 return;
 362         __set_hugetlb_cgroup(page, NULL, rsvd);
 363
 364         page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx,
 365                                                                    rsvd),
 366                               nr_pages);
 367
 368         if (rsvd)
 369                 css_put(&h_cg->css);
 370         else {
 371                 unsigned long usage =
 372                         h_cg->nodeinfo[page_to_nid(page)]->usage[idx];
 373                 /*
 374                  * This write is not atomic due to fetching usage and writing
 375                  * to it, but that's fine because we call this with
 376                  * hugetlb_lock held anyway.
 377                  */
 378                 WRITE_ONCE(h_cg->nodeinfo[page_to_nid(page)]->usage[idx],
 379                            usage - nr_pages);
 380         }
 381 }
 382
 383 void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
 384                                   struct page *page)
 385 {
 386         __hugetlb_cgroup_uncharge_page(idx, nr_pages, page, false);
 387 }
 388
 389 void hugetlb_cgroup_uncharge_page_rsvd(int idx, unsigned long nr_pages,
 390                                        struct page *page)
 391 {
 392         __hugetlb_cgroup_uncharge_page(idx, nr_pages, page, true);
 393 }
 394
 395 static void __hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 396                                              struct hugetlb_cgroup *h_cg,
 397                                              bool rsvd)
 398 {
 399         if (hugetlb_cgroup_disabled() || !h_cg)
 400                 return;
 401
 402         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
 403                 return;
 404
 405         page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx,
 406                                                                    rsvd),
 407                               nr_pages);
 408
 409         if (rsvd)
 410                 css_put(&h_cg->css);
 411 }
 412
 413 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 414                                     struct hugetlb_cgroup *h_cg)
 415 {
 416         __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, false);
 417 }
 418
 419 void hugetlb_cgroup_uncharge_cgroup_rsvd(int idx, unsigned long nr_pages,
 420                                          struct hugetlb_cgroup *h_cg)
 421 {
 422         __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, true);
 423 }
 424
 425 void hugetlb_cgroup_uncharge_counter(struct resv_map *resv, unsigned long start,
 426                                      unsigned long end)
 427 {
 428         if (hugetlb_cgroup_disabled() || !resv || !resv->reservation_counter ||
 429             !resv->css)
 430                 return;
 431
 432         page_counter_uncharge(resv->reservation_counter,
 433                               (end - start) * resv->pages_per_hpage);
 434         css_put(resv->css);
 435 }
 436
 437 void hugetlb_cgroup_uncharge_file_region(struct resv_map *resv,
 438                                          struct file_region *rg,
 439                                          unsigned long nr_pages,
 440                                          bool region_del)
 441 {
 442         if (hugetlb_cgroup_disabled() || !resv || !rg || !nr_pages)
 443                 return;
 444
 445         if (rg->reservation_counter && resv->pages_per_hpage && nr_pages > 0 &&
 446             !resv->reservation_counter) {
 447                 page_counter_uncharge(rg->reservation_counter,
 448                                       nr_pages * resv->pages_per_hpage);
 449                 /*
 450                  * Only do css_put(rg->css) when we delete the entire region
 451                  * because one file_region must hold exactly one css reference.
 452                  */
 453                 if (region_del)
 454                         css_put(rg->css);
 455         }
 456 }
 457
 458 enum {
 459         RES_USAGE,
 460         RES_RSVD_USAGE,
 461         RES_LIMIT,
 462         RES_RSVD_LIMIT,
 463         RES_MAX_USAGE,
 464         RES_RSVD_MAX_USAGE,
 465         RES_FAILCNT,
 466         RES_RSVD_FAILCNT,
 467 };
 468
 469 static int hugetlb_cgroup_read_numa_stat(struct seq_file *seq, void *dummy)
 470 {
 471         int nid;
 472         struct cftype *cft = seq_cft(seq);
 473         int idx = MEMFILE_IDX(cft->private);
 474         bool legacy = MEMFILE_ATTR(cft->private);
 475         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 476         struct cgroup_subsys_state *css;
 477         unsigned long usage;
 478
 479         if (legacy) {
 480                 /* Add up usage across all nodes for the non-hierarchical total. */
 481                 usage = 0;
 482                 for_each_node_state(nid, N_MEMORY)
 483                         usage += READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]);
 484                 seq_printf(seq, "total=%lu", usage * PAGE_SIZE);
 485
 486                 /* Simply print the per-node usage for the non-hierarchical total. */
 487                 for_each_node_state(nid, N_MEMORY)
 488                         seq_printf(seq, " N%d=%lu", nid,
 489                                    READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]) *
 490                                            PAGE_SIZE);
 491                 seq_putc(seq, '\n');
 492         }
 493
 494         /*
 495          * The hierarchical total is pretty much the value recorded by the
 496          * counter, so use that.
 497          */
 498         seq_printf(seq, "%stotal=%lu", legacy ? "hierarchical_" : "",
 499                    page_counter_read(&h_cg->hugepage[idx]) * PAGE_SIZE);
 500
 501         /*
 502          * For each node, transverse the css tree to obtain the hierarchical
 503          * node usage.
 504          */
 505         for_each_node_state(nid, N_MEMORY) {
 506                 usage = 0;
 507                 rcu_read_lock();
 508                 css_for_each_descendant_pre(css, &h_cg->css) {
 509                         usage += READ_ONCE(hugetlb_cgroup_from_css(css)
 510                                                    ->nodeinfo[nid]
 511                                                    ->usage[idx]);
 512                 }
 513                 rcu_read_unlock();
 514                 seq_printf(seq, " N%d=%lu", nid, usage * PAGE_SIZE);
 515         }
 516
 517         seq_putc(seq, '\n');
 518
 519         return 0;
 520 }
 521
 522 static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css,
 523                                    struct cftype *cft)
 524 {
 525         struct page_counter *counter;
 526         struct page_counter *rsvd_counter;
 527         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
 528
 529         counter = &h_cg->hugepage[MEMFILE_IDX(cft->private)];
 530         rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(cft->private)];
 531
 532         switch (MEMFILE_ATTR(cft->private)) {
 533         case RES_USAGE:
 534                 return (u64)page_counter_read(counter) * PAGE_SIZE;
 535         case RES_RSVD_USAGE:
 536                 return (u64)page_counter_read(rsvd_counter) * PAGE_SIZE;
 537         case RES_LIMIT:
 538                 return (u64)counter->max * PAGE_SIZE;
 539         case RES_RSVD_LIMIT:
 540                 return (u64)rsvd_counter->max * PAGE_SIZE;
 541         case RES_MAX_USAGE:
 542                 return (u64)counter->watermark * PAGE_SIZE;
 543         case RES_RSVD_MAX_USAGE:
 544                 return (u64)rsvd_counter->watermark * PAGE_SIZE;
 545         case RES_FAILCNT:
 546                 return counter->failcnt;
 547         case RES_RSVD_FAILCNT:
 548                 return rsvd_counter->failcnt;
 549         default:
 550                 BUG();
 551         }
 552 }
 553
 554 static int hugetlb_cgroup_read_u64_max(struct seq_file *seq, void *v)
 555 {
 556         int idx;
 557         u64 val;
 558         struct cftype *cft = seq_cft(seq);
 559         unsigned long limit;
 560         struct page_counter *counter;
 561         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 562
 563         idx = MEMFILE_IDX(cft->private);
 564         counter = &h_cg->hugepage[idx];
 565
 566         limit = round_down(PAGE_COUNTER_MAX,
 567                            pages_per_huge_page(&hstates[idx]));
 568
 569         switch (MEMFILE_ATTR(cft->private)) {
 570         case RES_RSVD_USAGE:
 571                 counter = &h_cg->rsvd_hugepage[idx];
 572                 fallthrough;
 573         case RES_USAGE:
 574                 val = (u64)page_counter_read(counter);
 575                 seq_printf(seq, "%llu\n", val * PAGE_SIZE);
 576                 break;
 577         case RES_RSVD_LIMIT:
 578                 counter = &h_cg->rsvd_hugepage[idx];
 579                 fallthrough;
 580         case RES_LIMIT:
 581                 val = (u64)counter->max;
 582                 if (val == limit)
 583                         seq_puts(seq, "max\n");
 584                 else
 585                         seq_printf(seq, "%llu\n", val * PAGE_SIZE);
 586                 break;
 587         default:
 588                 BUG();
 589         }
 590
 591         return 0;
 592 }
 593
 594 static DEFINE_MUTEX(hugetlb_limit_mutex);
 595
 596 static ssize_t hugetlb_cgroup_write(struct kernfs_open_file *of,
 597                                     char *buf, size_t nbytes, loff_t off,
 598                                     const char *max)
 599 {
 600         int ret, idx;
 601         unsigned long nr_pages;
 602         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
 603         bool rsvd = false;
 604
 605         if (hugetlb_cgroup_is_root(h_cg)) /* Can't set limit on root */
 606                 return -EINVAL;
 607
 608         buf = strstrip(buf);
 609         ret = page_counter_memparse(buf, max, &nr_pages);
 610         if (ret)
 611                 return ret;
 612
 613         idx = MEMFILE_IDX(of_cft(of)->private);
 614         nr_pages = round_down(nr_pages, pages_per_huge_page(&hstates[idx]));
 615
 616         switch (MEMFILE_ATTR(of_cft(of)->private)) {
 617         case RES_RSVD_LIMIT:
 618                 rsvd = true;
 619                 fallthrough;
 620         case RES_LIMIT:
 621                 mutex_lock(&hugetlb_limit_mutex);
 622                 ret = page_counter_set_max(
 623                         __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
 624                         nr_pages);
 625                 mutex_unlock(&hugetlb_limit_mutex);
 626                 break;
 627         default:
 628                 ret = -EINVAL;
 629                 break;
 630         }
 631         return ret ?: nbytes;
 632 }
 633
 634 static ssize_t hugetlb_cgroup_write_legacy(struct kernfs_open_file *of,
 635                                            char *buf, size_t nbytes, loff_t off)
 636 {
 637         return hugetlb_cgroup_write(of, buf, nbytes, off, "-1");
 638 }
 639
 640 static ssize_t hugetlb_cgroup_write_dfl(struct kernfs_open_file *of,
 641                                         char *buf, size_t nbytes, loff_t off)
 642 {
 643         return hugetlb_cgroup_write(of, buf, nbytes, off, "max");
 644 }
 645
 646 static ssize_t hugetlb_cgroup_reset(struct kernfs_open_file *of,
 647                                     char *buf, size_t nbytes, loff_t off)
 648 {
 649         int ret = 0;
 650         struct page_counter *counter, *rsvd_counter;
 651         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
 652
 653         counter = &h_cg->hugepage[MEMFILE_IDX(of_cft(of)->private)];
 654         rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(of_cft(of)->private)];
 655
 656         switch (MEMFILE_ATTR(of_cft(of)->private)) {
 657         case RES_MAX_USAGE:
 658                 page_counter_reset_watermark(counter);
 659                 break;
 660         case RES_RSVD_MAX_USAGE:
 661                 page_counter_reset_watermark(rsvd_counter);
 662                 break;
 663         case RES_FAILCNT:
 664                 counter->failcnt = 0;
 665                 break;
 666         case RES_RSVD_FAILCNT:
 667                 rsvd_counter->failcnt = 0;
 668                 break;
 669         default:
 670                 ret = -EINVAL;
 671                 break;
 672         }
 673         return ret ?: nbytes;
 674 }
 675
 676 static char *mem_fmt(char *buf, int size, unsigned long hsize)
 677 {
 678         if (hsize >= (1UL << 30))
 679                 snprintf(buf, size, "%luGB", hsize >> 30);
 680         else if (hsize >= (1UL << 20))
 681                 snprintf(buf, size, "%luMB", hsize >> 20);
 682         else
 683                 snprintf(buf, size, "%luKB", hsize >> 10);
 684         return buf;
 685 }
 686
 687 static int __hugetlb_events_show(struct seq_file *seq, bool local)
 688 {
 689         int idx;
 690         long max;
 691         struct cftype *cft = seq_cft(seq);
 692         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 693
 694         idx = MEMFILE_IDX(cft->private);
 695
 696         if (local)
 697                 max = atomic_long_read(&h_cg->events_local[idx][HUGETLB_MAX]);
 698         else
 699                 max = atomic_long_read(&h_cg->events[idx][HUGETLB_MAX]);
 700
 701         seq_printf(seq, "max %lu\n", max);
 702
 703         return 0;
 704 }
 705
 706 static int hugetlb_events_show(struct seq_file *seq, void *v)
 707 {
 708         return __hugetlb_events_show(seq, false);
 709 }
 710
 711 static int hugetlb_events_local_show(struct seq_file *seq, void *v)
 712 {
 713         return __hugetlb_events_show(seq, true);
 714 }
 715
 716 static void __init __hugetlb_cgroup_file_dfl_init(int idx)
 717 {
 718         char buf[32];
 719         struct cftype *cft;
 720         struct hstate *h = &hstates[idx];
 721
 722         /* format the size */
 723         mem_fmt(buf, sizeof(buf), huge_page_size(h));
 724
 725         /* Add the limit file */
 726         cft = &h->cgroup_files_dfl[0];
 727         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max", buf);
 728         cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
 729         cft->seq_show = hugetlb_cgroup_read_u64_max;
 730         cft->write = hugetlb_cgroup_write_dfl;
 731         cft->flags = CFTYPE_NOT_ON_ROOT;
 732
 733         /* Add the reservation limit file */
 734         cft = &h->cgroup_files_dfl[1];
 735         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max", buf);
 736         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT);
 737         cft->seq_show = hugetlb_cgroup_read_u64_max;
 738         cft->write = hugetlb_cgroup_write_dfl;
 739         cft->flags = CFTYPE_NOT_ON_ROOT;
 740
 741         /* Add the current usage file */
 742         cft = &h->cgroup_files_dfl[2];
 743         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.current", buf);
 744         cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
 745         cft->seq_show = hugetlb_cgroup_read_u64_max;
 746         cft->flags = CFTYPE_NOT_ON_ROOT;
 747
 748         /* Add the current reservation usage file */
 749         cft = &h->cgroup_files_dfl[3];
 750         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.current", buf);
 751         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
 752         cft->seq_show = hugetlb_cgroup_read_u64_max;
 753         cft->flags = CFTYPE_NOT_ON_ROOT;
 754
 755         /* Add the events file */
 756         cft = &h->cgroup_files_dfl[4];
 757         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events", buf);
 758         cft->private = MEMFILE_PRIVATE(idx, 0);
 759         cft->seq_show = hugetlb_events_show;
 760         cft->file_offset = offsetof(struct hugetlb_cgroup, events_file[idx]);
 761         cft->flags = CFTYPE_NOT_ON_ROOT;
 762
 763         /* Add the events.local file */
 764         cft = &h->cgroup_files_dfl[5];
 765         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events.local", buf);
 766         cft->private = MEMFILE_PRIVATE(idx, 0);
 767         cft->seq_show = hugetlb_events_local_show;
 768         cft->file_offset = offsetof(struct hugetlb_cgroup,
 769                                     events_local_file[idx]);
 770         cft->flags = CFTYPE_NOT_ON_ROOT;
 771
 772         /* Add the numa stat file */
 773         cft = &h->cgroup_files_dfl[6];
 774         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
 775         cft->seq_show = hugetlb_cgroup_read_numa_stat;
 776         cft->flags = CFTYPE_NOT_ON_ROOT;
 777
 778         /* NULL terminate the last cft */
 779         cft = &h->cgroup_files_dfl[7];
 780         memset(cft, 0, sizeof(*cft));
 781
 782         WARN_ON(cgroup_add_dfl_cftypes(&hugetlb_cgrp_subsys,
 783                                        h->cgroup_files_dfl));
 784 }
 785
 786 static void __init __hugetlb_cgroup_file_legacy_init(int idx)
 787 {
 788         char buf[32];
 789         struct cftype *cft;
 790         struct hstate *h = &hstates[idx];
 791
 792         /* format the size */
 793         mem_fmt(buf, sizeof(buf), huge_page_size(h));
 794
 795         /* Add the limit file */
 796         cft = &h->cgroup_files_legacy[0];
 797         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf);
 798         cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
 799         cft->read_u64 = hugetlb_cgroup_read_u64;
 800         cft->write = hugetlb_cgroup_write_legacy;
 801
 802         /* Add the reservation limit file */
 803         cft = &h->cgroup_files_legacy[1];
 804         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.limit_in_bytes", buf);
 805         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT);
 806         cft->read_u64 = hugetlb_cgroup_read_u64;
 807         cft->write = hugetlb_cgroup_write_legacy;
 808
 809         /* Add the usage file */
 810         cft = &h->cgroup_files_legacy[2];
 811         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf);
 812         cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
 813         cft->read_u64 = hugetlb_cgroup_read_u64;
 814
 815         /* Add the reservation usage file */
 816         cft = &h->cgroup_files_legacy[3];
 817         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.usage_in_bytes", buf);
 818         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
 819         cft->read_u64 = hugetlb_cgroup_read_u64;
 820
 821         /* Add the MAX usage file */
 822         cft = &h->cgroup_files_legacy[4];
 823         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf);
 824         cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE);
 825         cft->write = hugetlb_cgroup_reset;
 826         cft->read_u64 = hugetlb_cgroup_read_u64;
 827
 828         /* Add the MAX reservation usage file */
 829         cft = &h->cgroup_files_legacy[5];
 830         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max_usage_in_bytes", buf);
 831         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_MAX_USAGE);
 832         cft->write = hugetlb_cgroup_reset;
 833         cft->read_u64 = hugetlb_cgroup_read_u64;
 834
 835         /* Add the failcntfile */
 836         cft = &h->cgroup_files_legacy[6];
 837         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf);
 838         cft->private = MEMFILE_PRIVATE(idx, RES_FAILCNT);
 839         cft->write = hugetlb_cgroup_reset;
 840         cft->read_u64 = hugetlb_cgroup_read_u64;
 841
 842         /* Add the reservation failcntfile */
 843         cft = &h->cgroup_files_legacy[7];
 844         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.failcnt", buf);
 845         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_FAILCNT);
 846         cft->write = hugetlb_cgroup_reset;
 847         cft->read_u64 = hugetlb_cgroup_read_u64;
 848
 849         /* Add the numa stat file */
 850         cft = &h->cgroup_files_legacy[8];
 851         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
 852         cft->private = MEMFILE_PRIVATE(idx, 1);
 853         cft->seq_show = hugetlb_cgroup_read_numa_stat;
 854
 855         /* NULL terminate the last cft */
 856         cft = &h->cgroup_files_legacy[9];
 857         memset(cft, 0, sizeof(*cft));
 858
 859         WARN_ON(cgroup_add_legacy_cftypes(&hugetlb_cgrp_subsys,
 860                                           h->cgroup_files_legacy));
 861 }
 862
 863 static void __init __hugetlb_cgroup_file_init(int idx)
 864 {
 865         __hugetlb_cgroup_file_dfl_init(idx);
 866         __hugetlb_cgroup_file_legacy_init(idx);
 867 }
 868
 869 void __init hugetlb_cgroup_file_init(void)
 870 {
 871         struct hstate *h;
 872
 873         for_each_hstate(h) {
 874                 /*
 875                  * Add cgroup control files only if the huge page consists
 876                  * of more than two normal pages. This is because we use
 877                  * page[2].private for storing cgroup details.
 878                  */
 879                 if (huge_page_order(h) >= HUGETLB_CGROUP_MIN_ORDER)
 880                         __hugetlb_cgroup_file_init(hstate_index(h));
 881         }
 882 }
 883
 884 /*
 885  * hugetlb_lock will make sure a parallel cgroup rmdir won't happen
 886  * when we migrate hugepages
 887  */
 888 void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
 889 {
 890         struct hugetlb_cgroup *h_cg;
 891         struct hugetlb_cgroup *h_cg_rsvd;
 892         struct hstate *h = page_hstate(oldhpage);
 893
 894         if (hugetlb_cgroup_disabled())
 895                 return;
 896
 897         spin_lock_irq(&hugetlb_lock);
 898         h_cg = hugetlb_cgroup_from_page(oldhpage);
 899         h_cg_rsvd = hugetlb_cgroup_from_page_rsvd(oldhpage);
 900         set_hugetlb_cgroup(oldhpage, NULL);
 901         set_hugetlb_cgroup_rsvd(oldhpage, NULL);
 902
 903         /* move the h_cg details to new cgroup */
 904         set_hugetlb_cgroup(newhpage, h_cg);
 905         set_hugetlb_cgroup_rsvd(newhpage, h_cg_rsvd);
 906         list_move(&newhpage->lru, &h->hugepage_activelist);
 907         spin_unlock_irq(&hugetlb_lock);
 908         return;
 909 }
 910
 911 static struct cftype hugetlb_files[] = {
 912         {} /* terminate */
 913 };
 914
 915 struct cgroup_subsys hugetlb_cgrp_subsys = {
 916         .css_alloc      = hugetlb_cgroup_css_alloc,
 917         .css_offline    = hugetlb_cgroup_css_offline,
 918         .css_free       = hugetlb_cgroup_css_free,
 919         .dfl_cftypes    = hugetlb_files,
 920         .legacy_cftypes = hugetlb_files,
 921 };