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(page, 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         /*
 266          * We don't charge any cgroup if the compound page have less
 267          * than 3 pages.
 268          */
 269         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
 270                 goto done;
 271 again:
 272         rcu_read_lock();
 273         h_cg = hugetlb_cgroup_from_task(current);
 274         if (!css_tryget(&h_cg->css)) {
 275                 rcu_read_unlock();
 276                 goto again;
 277         }
 278         rcu_read_unlock();
 279
 280         if (!page_counter_try_charge(
 281                     __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
 282                     nr_pages, &counter)) {
 283                 ret = -ENOMEM;
 284                 hugetlb_event(h_cg, idx, HUGETLB_MAX);
 285                 css_put(&h_cg->css);
 286                 goto done;
 287         }
 288         /* Reservations take a reference to the css because they do not get
 289          * reparented.
 290          */
 291         if (!rsvd)
 292                 css_put(&h_cg->css);
 293 done:
 294         *ptr = h_cg;
 295         return ret;
 296 }
 297
 298 int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
 299                                  struct hugetlb_cgroup **ptr)
 300 {
 301         return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, false);
 302 }
 303
 304 int hugetlb_cgroup_charge_cgroup_rsvd(int idx, unsigned long nr_pages,
 305                                       struct hugetlb_cgroup **ptr)
 306 {
 307         return __hugetlb_cgroup_charge_cgroup(idx, nr_pages, ptr, true);
 308 }
 309
 310 /* Should be called with hugetlb_lock held */
 311 static void __hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 312                                            struct hugetlb_cgroup *h_cg,
 313                                            struct page *page, bool rsvd)
 314 {
 315         if (hugetlb_cgroup_disabled() || !h_cg)
 316                 return;
 317
 318         __set_hugetlb_cgroup(page_folio(page), h_cg, rsvd);
 319         if (!rsvd) {
 320                 unsigned long usage =
 321                         h_cg->nodeinfo[page_to_nid(page)]->usage[idx];
 322                 /*
 323                  * This write is not atomic due to fetching usage and writing
 324                  * to it, but that's fine because we call this with
 325                  * hugetlb_lock held anyway.
 326                  */
 327                 WRITE_ONCE(h_cg->nodeinfo[page_to_nid(page)]->usage[idx],
 328                            usage + nr_pages);
 329         }
 330 }
 331
 332 void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
 333                                   struct hugetlb_cgroup *h_cg,
 334                                   struct page *page)
 335 {
 336         __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, page, false);
 337 }
 338
 339 void hugetlb_cgroup_commit_charge_rsvd(int idx, unsigned long nr_pages,
 340                                        struct hugetlb_cgroup *h_cg,
 341                                        struct page *page)
 342 {
 343         __hugetlb_cgroup_commit_charge(idx, nr_pages, h_cg, page, true);
 344 }
 345
 346 /*
 347  * Should be called with hugetlb_lock held
 348  */
 349 static void __hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
 350                                            struct page *page, bool rsvd)
 351 {
 352         struct hugetlb_cgroup *h_cg;
 353         struct folio *folio = page_folio(page);
 354
 355         if (hugetlb_cgroup_disabled())
 356                 return;
 357         lockdep_assert_held(&hugetlb_lock);
 358         h_cg = __hugetlb_cgroup_from_folio(folio, rsvd);
 359         if (unlikely(!h_cg))
 360                 return;
 361         __set_hugetlb_cgroup(folio, NULL, rsvd);
 362
 363         page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx,
 364                                                                    rsvd),
 365                               nr_pages);
 366
 367         if (rsvd)
 368                 css_put(&h_cg->css);
 369         else {
 370                 unsigned long usage =
 371                         h_cg->nodeinfo[page_to_nid(page)]->usage[idx];
 372                 /*
 373                  * This write is not atomic due to fetching usage and writing
 374                  * to it, but that's fine because we call this with
 375                  * hugetlb_lock held anyway.
 376                  */
 377                 WRITE_ONCE(h_cg->nodeinfo[page_to_nid(page)]->usage[idx],
 378                            usage - nr_pages);
 379         }
 380 }
 381
 382 void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
 383                                   struct page *page)
 384 {
 385         __hugetlb_cgroup_uncharge_page(idx, nr_pages, page, false);
 386 }
 387
 388 void hugetlb_cgroup_uncharge_page_rsvd(int idx, unsigned long nr_pages,
 389                                        struct page *page)
 390 {
 391         __hugetlb_cgroup_uncharge_page(idx, nr_pages, page, true);
 392 }
 393
 394 static void __hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 395                                              struct hugetlb_cgroup *h_cg,
 396                                              bool rsvd)
 397 {
 398         if (hugetlb_cgroup_disabled() || !h_cg)
 399                 return;
 400
 401         if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
 402                 return;
 403
 404         page_counter_uncharge(__hugetlb_cgroup_counter_from_cgroup(h_cg, idx,
 405                                                                    rsvd),
 406                               nr_pages);
 407
 408         if (rsvd)
 409                 css_put(&h_cg->css);
 410 }
 411
 412 void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
 413                                     struct hugetlb_cgroup *h_cg)
 414 {
 415         __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, false);
 416 }
 417
 418 void hugetlb_cgroup_uncharge_cgroup_rsvd(int idx, unsigned long nr_pages,
 419                                          struct hugetlb_cgroup *h_cg)
 420 {
 421         __hugetlb_cgroup_uncharge_cgroup(idx, nr_pages, h_cg, true);
 422 }
 423
 424 void hugetlb_cgroup_uncharge_counter(struct resv_map *resv, unsigned long start,
 425                                      unsigned long end)
 426 {
 427         if (hugetlb_cgroup_disabled() || !resv || !resv->reservation_counter ||
 428             !resv->css)
 429                 return;
 430
 431         page_counter_uncharge(resv->reservation_counter,
 432                               (end - start) * resv->pages_per_hpage);
 433         css_put(resv->css);
 434 }
 435
 436 void hugetlb_cgroup_uncharge_file_region(struct resv_map *resv,
 437                                          struct file_region *rg,
 438                                          unsigned long nr_pages,
 439                                          bool region_del)
 440 {
 441         if (hugetlb_cgroup_disabled() || !resv || !rg || !nr_pages)
 442                 return;
 443
 444         if (rg->reservation_counter && resv->pages_per_hpage &&
 445             !resv->reservation_counter) {
 446                 page_counter_uncharge(rg->reservation_counter,
 447                                       nr_pages * resv->pages_per_hpage);
 448                 /*
 449                  * Only do css_put(rg->css) when we delete the entire region
 450                  * because one file_region must hold exactly one css reference.
 451                  */
 452                 if (region_del)
 453                         css_put(rg->css);
 454         }
 455 }
 456
 457 enum {
 458         RES_USAGE,
 459         RES_RSVD_USAGE,
 460         RES_LIMIT,
 461         RES_RSVD_LIMIT,
 462         RES_MAX_USAGE,
 463         RES_RSVD_MAX_USAGE,
 464         RES_FAILCNT,
 465         RES_RSVD_FAILCNT,
 466 };
 467
 468 static int hugetlb_cgroup_read_numa_stat(struct seq_file *seq, void *dummy)
 469 {
 470         int nid;
 471         struct cftype *cft = seq_cft(seq);
 472         int idx = MEMFILE_IDX(cft->private);
 473         bool legacy = MEMFILE_ATTR(cft->private);
 474         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 475         struct cgroup_subsys_state *css;
 476         unsigned long usage;
 477
 478         if (legacy) {
 479                 /* Add up usage across all nodes for the non-hierarchical total. */
 480                 usage = 0;
 481                 for_each_node_state(nid, N_MEMORY)
 482                         usage += READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]);
 483                 seq_printf(seq, "total=%lu", usage * PAGE_SIZE);
 484
 485                 /* Simply print the per-node usage for the non-hierarchical total. */
 486                 for_each_node_state(nid, N_MEMORY)
 487                         seq_printf(seq, " N%d=%lu", nid,
 488                                    READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]) *
 489                                            PAGE_SIZE);
 490                 seq_putc(seq, '\n');
 491         }
 492
 493         /*
 494          * The hierarchical total is pretty much the value recorded by the
 495          * counter, so use that.
 496          */
 497         seq_printf(seq, "%stotal=%lu", legacy ? "hierarchical_" : "",
 498                    page_counter_read(&h_cg->hugepage[idx]) * PAGE_SIZE);
 499
 500         /*
 501          * For each node, transverse the css tree to obtain the hierarchical
 502          * node usage.
 503          */
 504         for_each_node_state(nid, N_MEMORY) {
 505                 usage = 0;
 506                 rcu_read_lock();
 507                 css_for_each_descendant_pre(css, &h_cg->css) {
 508                         usage += READ_ONCE(hugetlb_cgroup_from_css(css)
 509                                                    ->nodeinfo[nid]
 510                                                    ->usage[idx]);
 511                 }
 512                 rcu_read_unlock();
 513                 seq_printf(seq, " N%d=%lu", nid, usage * PAGE_SIZE);
 514         }
 515
 516         seq_putc(seq, '\n');
 517
 518         return 0;
 519 }
 520
 521 static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css,
 522                                    struct cftype *cft)
 523 {
 524         struct page_counter *counter;
 525         struct page_counter *rsvd_counter;
 526         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
 527
 528         counter = &h_cg->hugepage[MEMFILE_IDX(cft->private)];
 529         rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(cft->private)];
 530
 531         switch (MEMFILE_ATTR(cft->private)) {
 532         case RES_USAGE:
 533                 return (u64)page_counter_read(counter) * PAGE_SIZE;
 534         case RES_RSVD_USAGE:
 535                 return (u64)page_counter_read(rsvd_counter) * PAGE_SIZE;
 536         case RES_LIMIT:
 537                 return (u64)counter->max * PAGE_SIZE;
 538         case RES_RSVD_LIMIT:
 539                 return (u64)rsvd_counter->max * PAGE_SIZE;
 540         case RES_MAX_USAGE:
 541                 return (u64)counter->watermark * PAGE_SIZE;
 542         case RES_RSVD_MAX_USAGE:
 543                 return (u64)rsvd_counter->watermark * PAGE_SIZE;
 544         case RES_FAILCNT:
 545                 return counter->failcnt;
 546         case RES_RSVD_FAILCNT:
 547                 return rsvd_counter->failcnt;
 548         default:
 549                 BUG();
 550         }
 551 }
 552
 553 static int hugetlb_cgroup_read_u64_max(struct seq_file *seq, void *v)
 554 {
 555         int idx;
 556         u64 val;
 557         struct cftype *cft = seq_cft(seq);
 558         unsigned long limit;
 559         struct page_counter *counter;
 560         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 561
 562         idx = MEMFILE_IDX(cft->private);
 563         counter = &h_cg->hugepage[idx];
 564
 565         limit = round_down(PAGE_COUNTER_MAX,
 566                            pages_per_huge_page(&hstates[idx]));
 567
 568         switch (MEMFILE_ATTR(cft->private)) {
 569         case RES_RSVD_USAGE:
 570                 counter = &h_cg->rsvd_hugepage[idx];
 571                 fallthrough;
 572         case RES_USAGE:
 573                 val = (u64)page_counter_read(counter);
 574                 seq_printf(seq, "%llu\n", val * PAGE_SIZE);
 575                 break;
 576         case RES_RSVD_LIMIT:
 577                 counter = &h_cg->rsvd_hugepage[idx];
 578                 fallthrough;
 579         case RES_LIMIT:
 580                 val = (u64)counter->max;
 581                 if (val == limit)
 582                         seq_puts(seq, "max\n");
 583                 else
 584                         seq_printf(seq, "%llu\n", val * PAGE_SIZE);
 585                 break;
 586         default:
 587                 BUG();
 588         }
 589
 590         return 0;
 591 }
 592
 593 static DEFINE_MUTEX(hugetlb_limit_mutex);
 594
 595 static ssize_t hugetlb_cgroup_write(struct kernfs_open_file *of,
 596                                     char *buf, size_t nbytes, loff_t off,
 597                                     const char *max)
 598 {
 599         int ret, idx;
 600         unsigned long nr_pages;
 601         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
 602         bool rsvd = false;
 603
 604         if (hugetlb_cgroup_is_root(h_cg)) /* Can't set limit on root */
 605                 return -EINVAL;
 606
 607         buf = strstrip(buf);
 608         ret = page_counter_memparse(buf, max, &nr_pages);
 609         if (ret)
 610                 return ret;
 611
 612         idx = MEMFILE_IDX(of_cft(of)->private);
 613         nr_pages = round_down(nr_pages, pages_per_huge_page(&hstates[idx]));
 614
 615         switch (MEMFILE_ATTR(of_cft(of)->private)) {
 616         case RES_RSVD_LIMIT:
 617                 rsvd = true;
 618                 fallthrough;
 619         case RES_LIMIT:
 620                 mutex_lock(&hugetlb_limit_mutex);
 621                 ret = page_counter_set_max(
 622                         __hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
 623                         nr_pages);
 624                 mutex_unlock(&hugetlb_limit_mutex);
 625                 break;
 626         default:
 627                 ret = -EINVAL;
 628                 break;
 629         }
 630         return ret ?: nbytes;
 631 }
 632
 633 static ssize_t hugetlb_cgroup_write_legacy(struct kernfs_open_file *of,
 634                                            char *buf, size_t nbytes, loff_t off)
 635 {
 636         return hugetlb_cgroup_write(of, buf, nbytes, off, "-1");
 637 }
 638
 639 static ssize_t hugetlb_cgroup_write_dfl(struct kernfs_open_file *of,
 640                                         char *buf, size_t nbytes, loff_t off)
 641 {
 642         return hugetlb_cgroup_write(of, buf, nbytes, off, "max");
 643 }
 644
 645 static ssize_t hugetlb_cgroup_reset(struct kernfs_open_file *of,
 646                                     char *buf, size_t nbytes, loff_t off)
 647 {
 648         int ret = 0;
 649         struct page_counter *counter, *rsvd_counter;
 650         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
 651
 652         counter = &h_cg->hugepage[MEMFILE_IDX(of_cft(of)->private)];
 653         rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(of_cft(of)->private)];
 654
 655         switch (MEMFILE_ATTR(of_cft(of)->private)) {
 656         case RES_MAX_USAGE:
 657                 page_counter_reset_watermark(counter);
 658                 break;
 659         case RES_RSVD_MAX_USAGE:
 660                 page_counter_reset_watermark(rsvd_counter);
 661                 break;
 662         case RES_FAILCNT:
 663                 counter->failcnt = 0;
 664                 break;
 665         case RES_RSVD_FAILCNT:
 666                 rsvd_counter->failcnt = 0;
 667                 break;
 668         default:
 669                 ret = -EINVAL;
 670                 break;
 671         }
 672         return ret ?: nbytes;
 673 }
 674
 675 static char *mem_fmt(char *buf, int size, unsigned long hsize)
 676 {
 677         if (hsize >= SZ_1G)
 678                 snprintf(buf, size, "%luGB", hsize / SZ_1G);
 679         else if (hsize >= SZ_1M)
 680                 snprintf(buf, size, "%luMB", hsize / SZ_1M);
 681         else
 682                 snprintf(buf, size, "%luKB", hsize / SZ_1K);
 683         return buf;
 684 }
 685
 686 static int __hugetlb_events_show(struct seq_file *seq, bool local)
 687 {
 688         int idx;
 689         long max;
 690         struct cftype *cft = seq_cft(seq);
 691         struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
 692
 693         idx = MEMFILE_IDX(cft->private);
 694
 695         if (local)
 696                 max = atomic_long_read(&h_cg->events_local[idx][HUGETLB_MAX]);
 697         else
 698                 max = atomic_long_read(&h_cg->events[idx][HUGETLB_MAX]);
 699
 700         seq_printf(seq, "max %lu\n", max);
 701
 702         return 0;
 703 }
 704
 705 static int hugetlb_events_show(struct seq_file *seq, void *v)
 706 {
 707         return __hugetlb_events_show(seq, false);
 708 }
 709
 710 static int hugetlb_events_local_show(struct seq_file *seq, void *v)
 711 {
 712         return __hugetlb_events_show(seq, true);
 713 }
 714
 715 static void __init __hugetlb_cgroup_file_dfl_init(int idx)
 716 {
 717         char buf[32];
 718         struct cftype *cft;
 719         struct hstate *h = &hstates[idx];
 720
 721         /* format the size */
 722         mem_fmt(buf, sizeof(buf), huge_page_size(h));
 723
 724         /* Add the limit file */
 725         cft = &h->cgroup_files_dfl[0];
 726         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max", buf);
 727         cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
 728         cft->seq_show = hugetlb_cgroup_read_u64_max;
 729         cft->write = hugetlb_cgroup_write_dfl;
 730         cft->flags = CFTYPE_NOT_ON_ROOT;
 731
 732         /* Add the reservation limit file */
 733         cft = &h->cgroup_files_dfl[1];
 734         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max", buf);
 735         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT);
 736         cft->seq_show = hugetlb_cgroup_read_u64_max;
 737         cft->write = hugetlb_cgroup_write_dfl;
 738         cft->flags = CFTYPE_NOT_ON_ROOT;
 739
 740         /* Add the current usage file */
 741         cft = &h->cgroup_files_dfl[2];
 742         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.current", buf);
 743         cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
 744         cft->seq_show = hugetlb_cgroup_read_u64_max;
 745         cft->flags = CFTYPE_NOT_ON_ROOT;
 746
 747         /* Add the current reservation usage file */
 748         cft = &h->cgroup_files_dfl[3];
 749         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.current", buf);
 750         cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
 751         cft->seq_show = hugetlb_cgroup_read_u64_max;
 752         cft->flags = CFTYPE_NOT_ON_ROOT;
 753
 754         /* Add the events file */
 755         cft = &h->cgroup_files_dfl[4];
 756         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events", buf);
 757         cft->private = MEMFILE_PRIVATE(idx, 0);
 758         cft->seq_show = hugetlb_events_show;
 759         cft->file_offset = offsetof(struct hugetlb_cgroup, events_file[idx]);
 760         cft->flags = CFTYPE_NOT_ON_ROOT;
 761
 762         /* Add the events.local file */
 763         cft = &h->cgroup_files_dfl[5];
 764         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events.local", buf);
 765         cft->private = MEMFILE_PRIVATE(idx, 0);
 766         cft->seq_show = hugetlb_events_local_show;
 767         cft->file_offset = offsetof(struct hugetlb_cgroup,
 768                                     events_local_file[idx]);
 769         cft->flags = CFTYPE_NOT_ON_ROOT;
 770
 771         /* Add the numa stat file */
 772         cft = &h->cgroup_files_dfl[6];
 773         snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
 774         cft->private = MEMFILE_PRIVATE(idx, 0);
 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         struct folio *old_folio = page_folio(oldhpage);
 894
 895         if (hugetlb_cgroup_disabled())
 896                 return;
 897
 898         spin_lock_irq(&hugetlb_lock);
 899         h_cg = hugetlb_cgroup_from_folio(old_folio);
 900         h_cg_rsvd = hugetlb_cgroup_from_folio_rsvd(old_folio);
 901         set_hugetlb_cgroup(oldhpage, NULL);
 902         set_hugetlb_cgroup_rsvd(oldhpage, NULL);
 903
 904         /* move the h_cg details to new cgroup */
 905         set_hugetlb_cgroup(newhpage, h_cg);
 906         set_hugetlb_cgroup_rsvd(newhpage, h_cg_rsvd);
 907         list_move(&newhpage->lru, &h->hugepage_activelist);
 908         spin_unlock_irq(&hugetlb_lock);
 909         return;
 910 }
 911
 912 static struct cftype hugetlb_files[] = {
 913         {} /* terminate */
 914 };
 915
 916 struct cgroup_subsys hugetlb_cgrp_subsys = {
 917         .css_alloc      = hugetlb_cgroup_css_alloc,
 918         .css_offline    = hugetlb_cgroup_css_offline,
 919         .css_free       = hugetlb_cgroup_css_free,
 920         .dfl_cftypes    = hugetlb_files,
 921         .legacy_cftypes = hugetlb_files,
 922 };