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         struct folio *folio = page_folio(page);
 195
 196         page_hcg = hugetlb_cgroup_from_folio(folio);
 197         /*
 198          * We can have pages in active list without any cgroup
 199          * ie, hugepage with less than 3 pages. We can safely
 200          * ignore those pages.
 201          */
 202         if (!page_hcg || page_hcg != h_cg)
 203                 goto out;
 204
 205         nr_pages = compound_nr(page);
 206         if (!parent) {
 207                 parent = root_h_cgroup;
 208                 /* root has no limit */
 209                 page_counter_charge(&parent->hugepage[idx], nr_pages);
 210         }
 211         counter = &h_cg->hugepage[idx];
 212         /* Take the pages off the local counter */
 213         page_counter_cancel(counter, nr_pages);
 214
 215         set_hugetlb_cgroup(folio, parent);
 216 out:
 217         return;
 218 }
 219
 220 /*
 221  * Force the hugetlb cgroup to empty the hugetlb resources by moving them to
 222  * the parent cgroup.
 223  */
 224 static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css)
 225 {
 226         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
 227         struct hstate *h;
 228         struct page *page;
 229
 230         do {
 231                 for_each_hstate(h) {
 232                         spin_lock_irq(&hugetlb_lock);
 233                         list_for_each_entry(page, &h->hugepage_activelist, lru)
 234                                 hugetlb_cgroup_move_parent(hstate_index(h), h_cg, page);
 235
 236                         spin_unlock_irq(&hugetlb_lock);
 237                 }
 238                 cond_resched();
 239         } while (hugetlb_cgroup_have_usage(h_cg));
 240 }
 241
 242 static inline void hugetlb_event(struct hugetlb_cgroup *hugetlb, int idx,
 243                                  enum hugetlb_memory_event event)
 244 {
 245         atomic_long_inc(&hugetlb->events_local[idx][event]);
 246         cgroup_file_notify(&hugetlb->events_local_file[idx]);
 247
 248         do {
 249                 atomic_long_inc(&hugetlb->events[idx][event]);
 250                 cgroup_file_notify(&hugetlb->events_file[idx]);
 251         } while ((hugetlb = parent_hugetlb_cgroup(hugetlb)) &&
 252                  !hugetlb_cgroup_is_root(hugetlb));
 253 }
 254
 255 static int __hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
 256                                           struct hugetlb_cgroup **ptr,
 257                                           bool rsvd)
 258 {
 259         int ret = 0;
 260         struct page_counter *counter;
 261         struct hugetlb_cgroup *h_cg = NULL;
 262
 263         if (hugetlb_cgroup_disabled())
 264                 goto done;
 265 again:
 266         rcu_read_lock();
 267         h_cg = hugetlb_cgroup_from_task(current);
 268         if (!css_tryget(&h_cg->css)) {
 269                 rcu_read_unlock();
 270                 goto again;
 271         }
 272         rcu_read_unlock();
 273
 274         if (!page_counter_try_charge(
 275                     __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
 276                     nr_pages, &counter)) {
 277                 ret = -ENOMEM;
 278                 hugetlb_event(h_cg, idx, HUGETLB_MAX);
 279                 css_put(&h_cg->css);
 280                 goto done;
 281         }
 282         /* Reservations take a reference to the css because they do not get
 283          * reparented.
 284          */
 285         if (!rsvd)
 286                 css_put(&h_cg->css);
 287 done:
 288         *ptr = h_cg;
 289         return ret;
 290 }
 291
 292 int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
 293                                  struct hugetlb_cgroup **ptr)
 294 {
 295         return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, false);
 296 }
 297
 298 int hugetlb_cgroup_charge_cgroup_rsvd(int idx, unsigned long nr_pages,
 299                                       struct hugetlb_cgroup **ptr)
 300 {
 301         return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, true);
 302 }
 303
 304 /* Should be called with hugetlb_lock held */
 305 static void __hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 306                                            struct hugetlb_cgroup *h_cg,
 307                                            struct folio *folio, bool rsvd)
 308 {
 309         if (hugetlb_cgroup_disabled() || !h_cg)
 310                 return;
 311
 312         __set_hugetlb_cgroup(folio, h_cg, rsvd);
 313         if (!rsvd) {
 314                 unsigned long usage =
 315                         h_cg->nodeinfo[folio_nid(folio)]->usage[idx];
 316                 /*
 317                  * This write is not atomic due to fetching usage and writing
 318                  * to it, but that's fine because we call this with
 319                  * hugetlb_lock held anyway.
 320                  */
 321                 WRITE_ONCE(h_cg->nodeinfo[folio_nid(folio)]->usage[idx],
 322                            usage + nr_pages);
 323         }
 324 }
 325
 326 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 327                                   struct hugetlb_cgroup *h_cg,
 328                                   struct folio *folio)
 329 {
 330         __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, folio, false);
 331 }
 332
 333 void hugetlb_cgroup_commit_charge_rsvd(int idx, unsigned long nr_pages,
 334                                        struct hugetlb_cgroup *h_cg,
 335                                        struct folio *folio)
 336 {
 337         __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, folio, true);
 338 }
 339
 340 /*
 341  * Should be called with hugetlb_lock held
 342  */
 343 static void __hugetlb_cgroup_uncharge_folio(int idx, unsigned long nr_pages,
 344                                            struct folio *folio, bool rsvd)
 345 {
 346         struct hugetlb_cgroup *h_cg;
 347
 348         if (hugetlb_cgroup_disabled())
 349                 return;
 350         lockdep_assert_held(&hugetlb_lock);
 351         h_cg = __hugetlb_cgroup_from_folio(folio, rsvd);
 352         if (unlikely(!h_cg))
 353                 return;
 354         __set_hugetlb_cgroup(folio, NULL, rsvd);
 355
 356         page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx,
 357                                                                    rsvd),
 358                               nr_pages);
 359
 360         if (rsvd)
 361                 css_put(&h_cg->css);
 362         else {
 363                 unsigned long usage =
 364                         h_cg->nodeinfo[folio_nid(folio)]->usage[idx];
 365                 /*
 366                  * This write is not atomic due to fetching usage and writing
 367                  * to it, but that's fine because we call this with
 368                  * hugetlb_lock held anyway.
 369                  */
 370                 WRITE_ONCE(h_cg->nodeinfo[folio_nid(folio)]->usage[idx],
 371                            usage - nr_pages);
 372         }
 373 }
 374
 375 void hugetlb_cgroup_uncharge_folio(int idx, unsigned long nr_pages,
 376                                   struct folio *folio)
 377 {
 378         __hugetlb_cgroup_uncharge_folio(idx, nr_pages, folio, false);
 379 }
 380
 381 void hugetlb_cgroup_uncharge_folio_rsvd(int idx, unsigned long nr_pages,
 382                                        struct folio *folio)
 383 {
 384         __hugetlb_cgroup_uncharge_folio(idx, nr_pages, folio, true);
 385 }
 386
 387 static void __hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 388                                              struct hugetlb_cgroup *h_cg,
 389                                              bool rsvd)
 390 {
 391         if (hugetlb_cgroup_disabled() || !h_cg)
 392                 return;
 393
 394         page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx,
 395                                                                    rsvd),
 396                               nr_pages);
 397
 398         if (rsvd)
 399                 css_put(&h_cg->css);
 400 }
 401
 402 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 403                                     struct hugetlb_cgroup *h_cg)
 404 {
 405         __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, false);
 406 }
 407
 408 void hugetlb_cgroup_uncharge_cgroup_rsvd(int idx, unsigned long nr_pages,
 409                                          struct hugetlb_cgroup *h_cg)
 410 {
 411         __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, true);
 412 }
 413
 414 void hugetlb_cgroup_uncharge_counter(struct resv_map *resv, unsigned long start,
 415                                      unsigned long end)
 416 {
 417         if (hugetlb_cgroup_disabled() || !resv || !resv->reservation_counter ||
 418             !resv->css)
 419                 return;
 420
 421         page_counter_uncharge(resv->reservation_counter,
 422                               (end - start) * resv->pages_per_hpage);
 423         css_put(resv->css);
 424 }
 425
 426 void hugetlb_cgroup_uncharge_file_region(struct resv_map *resv,
 427                                          struct file_region *rg,
 428                                          unsigned long nr_pages,
 429                                          bool region_del)
 430 {
 431         if (hugetlb_cgroup_disabled() || !resv || !rg || !nr_pages)
 432                 return;
 433
 434         if (rg->reservation_counter && resv->pages_per_hpage &&
 435             !resv->reservation_counter) {
 436                 page_counter_uncharge(rg->reservation_counter,
 437                                       nr_pages * resv->pages_per_hpage);
 438                 /*
 439                  * Only do css_put(rg->css) when we delete the entire region
 440                  * because one file_region must hold exactly one css reference.
 441                  */
 442                 if (region_del)
 443                         css_put(rg->css);
 444         }
 445 }
 446
 447 enum {
 448         RES_USAGE,
 449         RES_RSVD_USAGE,
 450         RES_LIMIT,
 451         RES_RSVD_LIMIT,
 452         RES_MAX_USAGE,
 453         RES_RSVD_MAX_USAGE,
 454         RES_FAILCNT,
 455         RES_RSVD_FAILCNT,
 456 };
 457
 458 static int hugetlb_cgroup_read_numa_stat(struct seq_file *seq, void *dummy)
 459 {
 460         int nid;
 461         struct cftype *cft = seq_cft(seq);
 462         int idx = MEMFILE_IDX(cft->private);
 463         bool legacy = MEMFILE_ATTR(cft->private);
 464         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 465         struct cgroup_subsys_state *css;
 466         unsigned long usage;
 467
 468         if (legacy) {
 469                 /* Add up usage across all nodes for the non-hierarchical total. */
 470                 usage = 0;
 471                 for_each_node_state(nid, N_MEMORY)
 472                         usage += READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]);
 473                 seq_printf(seq, "total=%lu", usage * PAGE_SIZE);
 474
 475                 /* Simply print the per-node usage for the non-hierarchical total. */
 476                 for_each_node_state(nid, N_MEMORY)
 477                         seq_printf(seq, " N%d=%lu", nid,
 478                                    READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]) *
 479                                            PAGE_SIZE);
 480                 seq_putc(seq, '\n');
 481         }
 482
 483         /*
 484          * The hierarchical total is pretty much the value recorded by the
 485          * counter, so use that.
 486          */
 487         seq_printf(seq, "%stotal=%lu", legacy ? "hierarchical_" : "",
 488                    page_counter_read(&h_cg->hugepage[idx]) * PAGE_SIZE);
 489
 490         /*
 491          * For each node, transverse the css tree to obtain the hierarchical
 492          * node usage.
 493          */
 494         for_each_node_state(nid, N_MEMORY) {
 495                 usage = 0;
 496                 rcu_read_lock();
 497                 css_for_each_descendant_pre(css, &h_cg->css) {
 498                         usage += READ_ONCE(hugetlb_cgroup_from_css(css)
 499                                                    ->nodeinfo[nid]
 500                                                    ->usage[idx]);
 501                 }
 502                 rcu_read_unlock();
 503                 seq_printf(seq, " N%d=%lu", nid, usage * PAGE_SIZE);
 504         }
 505
 506         seq_putc(seq, '\n');
 507
 508         return 0;
 509 }
 510
 511 static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css,
 512                                    struct cftype *cft)
 513 {
 514         struct page_counter *counter;
 515         struct page_counter *rsvd_counter;
 516         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
 517
 518         counter = &h_cg->hugepage[MEMFILE_IDX(cft->private)];
 519         rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(cft->private)];
 520
 521         switch (MEMFILE_ATTR(cft->private)) {
 522         case RES_USAGE:
 523                 return (u64)page_counter_read(counter) * PAGE_SIZE;
 524         case RES_RSVD_USAGE:
 525                 return (u64)page_counter_read(rsvd_counter) * PAGE_SIZE;
 526         case RES_LIMIT:
 527                 return (u64)counter->max * PAGE_SIZE;
 528         case RES_RSVD_LIMIT:
 529                 return (u64)rsvd_counter->max * PAGE_SIZE;
 530         case RES_MAX_USAGE:
 531                 return (u64)counter->watermark * PAGE_SIZE;
 532         case RES_RSVD_MAX_USAGE:
 533                 return (u64)rsvd_counter->watermark * PAGE_SIZE;
 534         case RES_FAILCNT:
 535                 return counter->failcnt;
 536         case RES_RSVD_FAILCNT:
 537                 return rsvd_counter->failcnt;
 538         default:
 539                 BUG();
 540         }
 541 }
 542
 543 static int hugetlb_cgroup_read_u64_max(struct seq_file *seq, void *v)
 544 {
 545         int idx;
 546         u64 val;
 547         struct cftype *cft = seq_cft(seq);
 548         unsigned long limit;
 549         struct page_counter *counter;
 550         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 551
 552         idx = MEMFILE_IDX(cft->private);
 553         counter = &h_cg->hugepage[idx];
 554
 555         limit = round_down(PAGE_COUNTER_MAX,
 556                            pages_per_huge_page(&hstates[idx]));
 557
 558         switch (MEMFILE_ATTR(cft->private)) {
 559         case RES_RSVD_USAGE:
 560                 counter = &h_cg->rsvd_hugepage[idx];
 561                 fallthrough;
 562         case RES_USAGE:
 563                 val = (u64)page_counter_read(counter);
 564                 seq_printf(seq, "%llu\n", val * PAGE_SIZE);
 565                 break;
 566         case RES_RSVD_LIMIT:
 567                 counter = &h_cg->rsvd_hugepage[idx];
 568                 fallthrough;
 569         case RES_LIMIT:
 570                 val = (u64)counter->max;
 571                 if (val == limit)
 572                         seq_puts(seq, "max\n");
 573                 else
 574                         seq_printf(seq, "%llu\n", val * PAGE_SIZE);
 575                 break;
 576         default:
 577                 BUG();
 578         }
 579
 580         return 0;
 581 }
 582
 583 static DEFINE_MUTEX(hugetlb_limit_mutex);
 584
 585 static ssize_t hugetlb_cgroup_write(struct kernfs_open_file *of,
 586                                     char *buf, size_t nbytes, loff_t off,
 587                                     const char *max)
 588 {
 589         int ret, idx;
 590         unsigned long nr_pages;
 591         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
 592         bool rsvd = false;
 593
 594         if (hugetlb_cgroup_is_root(h_cg)) /* Can't set limit on root */
 595                 return -EINVAL;
 596
 597         buf = strstrip(buf);
 598         ret = page_counter_memparse(buf, max, &nr_pages);
 599         if (ret)
 600                 return ret;
 601
 602         idx = MEMFILE_IDX(of_cft(of)->private);
 603         nr_pages = round_down(nr_pages, pages_per_huge_page(&hstates[idx]));
 604
 605         switch (MEMFILE_ATTR(of_cft(of)->private)) {
 606         case RES_RSVD_LIMIT:
 607                 rsvd = true;
 608                 fallthrough;
 609         case RES_LIMIT:
 610                 mutex_lock(&hugetlb_limit_mutex);
 611                 ret = page_counter_set_max(
 612                         __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
 613                         nr_pages);
 614                 mutex_unlock(&hugetlb_limit_mutex);
 615                 break;
 616         default:
 617                 ret = -EINVAL;
 618                 break;
 619         }
 620         return ret ?: nbytes;
 621 }
 622
 623 static ssize_t hugetlb_cgroup_write_legacy(struct kernfs_open_file *of,
 624                                            char *buf, size_t nbytes, loff_t off)
 625 {
 626         return hugetlb_cgroup_write(of, buf, nbytes, off, "-1");
 627 }
 628
 629 static ssize_t hugetlb_cgroup_write_dfl(struct kernfs_open_file *of,
 630                                         char *buf, size_t nbytes, loff_t off)
 631 {
 632         return hugetlb_cgroup_write(of, buf, nbytes, off, "max");
 633 }
 634
 635 static ssize_t hugetlb_cgroup_reset(struct kernfs_open_file *of,
 636                                     char *buf, size_t nbytes, loff_t off)
 637 {
 638         int ret = 0;
 639         struct page_counter *counter, *rsvd_counter;
 640         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
 641
 642         counter = &h_cg->hugepage[MEMFILE_IDX(of_cft(of)->private)];
 643         rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(of_cft(of)->private)];
 644
 645         switch (MEMFILE_ATTR(of_cft(of)->private)) {
 646         case RES_MAX_USAGE:
 647                 page_counter_reset_watermark(counter);
 648                 break;
 649         case RES_RSVD_MAX_USAGE:
 650                 page_counter_reset_watermark(rsvd_counter);
 651                 break;
 652         case RES_FAILCNT:
 653                 counter->failcnt = 0;
 654                 break;
 655         case RES_RSVD_FAILCNT:
 656                 rsvd_counter->failcnt = 0;
 657                 break;
 658         default:
 659                 ret = -EINVAL;
 660                 break;
 661         }
 662         return ret ?: nbytes;
 663 }
 664
 665 static char *mem_fmt(char *buf, int size, unsigned long hsize)
 666 {
 667         if (hsize >= SZ_1G)
 668                 snprintf(buf, size, "%luGB", hsize / SZ_1G);
 669         else if (hsize >= SZ_1M)
 670                 snprintf(buf, size, "%luMB", hsize / SZ_1M);
 671         else
 672                 snprintf(buf, size, "%luKB", hsize / SZ_1K);
 673         return buf;
 674 }
 675
 676 static int __hugetlb_events_show(struct seq_file *seq, bool local)
 677 {
 678         int idx;
 679         long max;
 680         struct cftype *cft = seq_cft(seq);
 681         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 682
 683         idx = MEMFILE_IDX(cft->private);
 684
 685         if (local)
 686                 max = atomic_long_read(&h_cg->events_local[idx][HUGETLB_MAX]);
 687         else
 688                 max = atomic_long_read(&h_cg->events[idx][HUGETLB_MAX]);
 689
 690         seq_printf(seq, "max %lu\n", max);
 691
 692         return 0;
 693 }
 694
 695 static int hugetlb_events_show(struct seq_file *seq, void *v)
 696 {
 697         return __hugetlb_events_show(seq, false);
 698 }
 699
 700 static int hugetlb_events_local_show(struct seq_file *seq, void *v)
 701 {
 702         return __hugetlb_events_show(seq, true);
 703 }
 704
 705 static void __init __hugetlb_cgroup_file_dfl_init(int idx)
 706 {
 707         char buf[32];
 708         struct cftype *cft;
 709         struct hstate *h = &hstates[idx];
 710
 711         /* format the size */
 712         mem_fmt(buf, sizeof(buf), huge_page_size(h));
 713
 714         /* Add the limit file */
 715         cft = &h->cgroup_files_dfl[0];
 716         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max", buf);
 717         cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
 718         cft->seq_show = hugetlb_cgroup_read_u64_max;
 719         cft->write = hugetlb_cgroup_write_dfl;
 720         cft->flags = CFTYPE_NOT_ON_ROOT;
 721
 722         /* Add the reservation limit file */
 723         cft = &h->cgroup_files_dfl[1];
 724         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max", buf);
 725         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_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 current usage file */
 731         cft = &h->cgroup_files_dfl[2];
 732         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.current", buf);
 733         cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
 734         cft->seq_show = hugetlb_cgroup_read_u64_max;
 735         cft->flags = CFTYPE_NOT_ON_ROOT;
 736
 737         /* Add the current reservation usage file */
 738         cft = &h->cgroup_files_dfl[3];
 739         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.current", buf);
 740         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
 741         cft->seq_show = hugetlb_cgroup_read_u64_max;
 742         cft->flags = CFTYPE_NOT_ON_ROOT;
 743
 744         /* Add the events file */
 745         cft = &h->cgroup_files_dfl[4];
 746         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events", buf);
 747         cft->private = MEMFILE_PRIVATE(idx, 0);
 748         cft->seq_show = hugetlb_events_show;
 749         cft->file_offset = offsetof(struct hugetlb_cgroup, events_file[idx]);
 750         cft->flags = CFTYPE_NOT_ON_ROOT;
 751
 752         /* Add the events.local file */
 753         cft = &h->cgroup_files_dfl[5];
 754         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events.local", buf);
 755         cft->private = MEMFILE_PRIVATE(idx, 0);
 756         cft->seq_show = hugetlb_events_local_show;
 757         cft->file_offset = offsetof(struct hugetlb_cgroup,
 758                                     events_local_file[idx]);
 759         cft->flags = CFTYPE_NOT_ON_ROOT;
 760
 761         /* Add the numa stat file */
 762         cft = &h->cgroup_files_dfl[6];
 763         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
 764         cft->private = MEMFILE_PRIVATE(idx, 0);
 765         cft->seq_show = hugetlb_cgroup_read_numa_stat;
 766         cft->flags = CFTYPE_NOT_ON_ROOT;
 767
 768         /* NULL terminate the last cft */
 769         cft = &h->cgroup_files_dfl[7];
 770         memset(cft, 0, sizeof(*cft));
 771
 772         WARN_ON(cgroup_add_dfl_cftypes(&hugetlb_cgrp_subsys,
 773                                        h->cgroup_files_dfl));
 774 }
 775
 776 static void __init __hugetlb_cgroup_file_legacy_init(int idx)
 777 {
 778         char buf[32];
 779         struct cftype *cft;
 780         struct hstate *h = &hstates[idx];
 781
 782         /* format the size */
 783         mem_fmt(buf, sizeof(buf), huge_page_size(h));
 784
 785         /* Add the limit file */
 786         cft = &h->cgroup_files_legacy[0];
 787         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf);
 788         cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
 789         cft->read_u64 = hugetlb_cgroup_read_u64;
 790         cft->write = hugetlb_cgroup_write_legacy;
 791
 792         /* Add the reservation limit file */
 793         cft = &h->cgroup_files_legacy[1];
 794         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.limit_in_bytes", buf);
 795         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT);
 796         cft->read_u64 = hugetlb_cgroup_read_u64;
 797         cft->write = hugetlb_cgroup_write_legacy;
 798
 799         /* Add the usage file */
 800         cft = &h->cgroup_files_legacy[2];
 801         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf);
 802         cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
 803         cft->read_u64 = hugetlb_cgroup_read_u64;
 804
 805         /* Add the reservation usage file */
 806         cft = &h->cgroup_files_legacy[3];
 807         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.usage_in_bytes", buf);
 808         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
 809         cft->read_u64 = hugetlb_cgroup_read_u64;
 810
 811         /* Add the MAX usage file */
 812         cft = &h->cgroup_files_legacy[4];
 813         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf);
 814         cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE);
 815         cft->write = hugetlb_cgroup_reset;
 816         cft->read_u64 = hugetlb_cgroup_read_u64;
 817
 818         /* Add the MAX reservation usage file */
 819         cft = &h->cgroup_files_legacy[5];
 820         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max_usage_in_bytes", buf);
 821         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_MAX_USAGE);
 822         cft->write = hugetlb_cgroup_reset;
 823         cft->read_u64 = hugetlb_cgroup_read_u64;
 824
 825         /* Add the failcntfile */
 826         cft = &h->cgroup_files_legacy[6];
 827         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf);
 828         cft->private = MEMFILE_PRIVATE(idx, RES_FAILCNT);
 829         cft->write = hugetlb_cgroup_reset;
 830         cft->read_u64 = hugetlb_cgroup_read_u64;
 831
 832         /* Add the reservation failcntfile */
 833         cft = &h->cgroup_files_legacy[7];
 834         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.failcnt", buf);
 835         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_FAILCNT);
 836         cft->write = hugetlb_cgroup_reset;
 837         cft->read_u64 = hugetlb_cgroup_read_u64;
 838
 839         /* Add the numa stat file */
 840         cft = &h->cgroup_files_legacy[8];
 841         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
 842         cft->private = MEMFILE_PRIVATE(idx, 1);
 843         cft->seq_show = hugetlb_cgroup_read_numa_stat;
 844
 845         /* NULL terminate the last cft */
 846         cft = &h->cgroup_files_legacy[9];
 847         memset(cft, 0, sizeof(*cft));
 848
 849         WARN_ON(cgroup_add_legacy_cftypes(&hugetlb_cgrp_subsys,
 850                                           h->cgroup_files_legacy));
 851 }
 852
 853 static void __init __hugetlb_cgroup_file_init(int idx)
 854 {
 855         __hugetlb_cgroup_file_dfl_init(idx);
 856         __hugetlb_cgroup_file_legacy_init(idx);
 857 }
 858
 859 void __init hugetlb_cgroup_file_init(void)
 860 {
 861         struct hstate *h;
 862
 863         for_each_hstate(h)
 864                 __hugetlb_cgroup_file_init(hstate_index(h));
 865 }
 866
 867 /*
 868  * hugetlb_lock will make sure a parallel cgroup rmdir won't happen
 869  * when we migrate hugepages
 870  */
 871 void hugetlb_cgroup_migrate(struct folio *old_folio, struct folio *new_folio)
 872 {
 873         struct hugetlb_cgroup *h_cg;
 874         struct hugetlb_cgroup *h_cg_rsvd;
 875         struct hstate *h = folio_hstate(old_folio);
 876
 877         if (hugetlb_cgroup_disabled())
 878                 return;
 879
 880         spin_lock_irq(&hugetlb_lock);
 881         h_cg = hugetlb_cgroup_from_folio(old_folio);
 882         h_cg_rsvd = hugetlb_cgroup_from_folio_rsvd(old_folio);
 883         set_hugetlb_cgroup(old_folio, NULL);
 884         set_hugetlb_cgroup_rsvd(old_folio, NULL);
 885
 886         /* move the h_cg details to new cgroup */
 887         set_hugetlb_cgroup(new_folio, h_cg);
 888         set_hugetlb_cgroup_rsvd(new_folio, h_cg_rsvd);
 889         list_move(&new_folio->lru, &h->hugepage_activelist);
 890         spin_unlock_irq(&hugetlb_lock);
 891         return;
 892 }
 893
 894 static struct cftype hugetlb_files[] = {
 895         {} /* terminate */
 896 };
 897
 898 struct cgroup_subsys hugetlb_cgrp_subsys = {
 899         .css_alloc      = hugetlb_cgroup_css_alloc,
 900         .css_offline    = hugetlb_cgroup_css_offline,
 901         .css_free       = hugetlb_cgroup_css_free,
 902         .dfl_cftypes    = hugetlb_files,
 903         .legacy_cftypes = hugetlb_files,
 904 };