arch/powerpc/platforms/pseries/cmm.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Collaborative memory management interface.
   4  *
   5  * Copyright (C) 2008 IBM Corporation
   6  * Author(s): Brian King (brking@linux.vnet.ibm.com),
   7  */
   8
   9 #include <linux/ctype.h>
  10 #include <linux/delay.h>
  11 #include <linux/errno.h>
  12 #include <linux/fs.h>
  13 #include <linux/gfp.h>
  14 #include <linux/kthread.h>
  15 #include <linux/module.h>
  16 #include <linux/oom.h>
  17 #include <linux/reboot.h>
  18 #include <linux/sched.h>
  19 #include <linux/stringify.h>
  20 #include <linux/swap.h>
  21 #include <linux/device.h>
  22 #include <asm/firmware.h>
  23 #include <asm/hvcall.h>
  24 #include <asm/mmu.h>
  25 #include <asm/pgalloc.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/memory.h>
  28 #include <asm/plpar_wrappers.h>
  29
  30 #include "pseries.h"
  31
  32 #define CMM_DRIVER_VERSION      "1.0.0"
  33 #define CMM_DEFAULT_DELAY       1
  34 #define CMM_HOTPLUG_DELAY       5
  35 #define CMM_DEBUG                       0
  36 #define CMM_DISABLE             0
  37 #define CMM_OOM_KB              1024
  38 #define CMM_MIN_MEM_MB          256
  39 #define KB2PAGES(_p)            ((_p)>>(PAGE_SHIFT-10))
  40 #define PAGES2KB(_p)            ((_p)<<(PAGE_SHIFT-10))
  41
  42 #define CMM_MEM_HOTPLUG_PRI     1
  43
  44 static unsigned int delay = CMM_DEFAULT_DELAY;
  45 static unsigned int hotplug_delay = CMM_HOTPLUG_DELAY;
  46 static unsigned int oom_kb = CMM_OOM_KB;
  47 static unsigned int cmm_debug = CMM_DEBUG;
  48 static unsigned int cmm_disabled = CMM_DISABLE;
  49 static unsigned long min_mem_mb = CMM_MIN_MEM_MB;
  50 static struct device cmm_dev;
  51
  52 MODULE_AUTHOR("Brian King <brking@linux.vnet.ibm.com>");
  53 MODULE_DESCRIPTION("IBM System p Collaborative Memory Manager");
  54 MODULE_LICENSE("GPL");
  55 MODULE_VERSION(CMM_DRIVER_VERSION);
  56
  57 module_param_named(delay, delay, uint, 0644);
  58 MODULE_PARM_DESC(delay, "Delay (in seconds) between polls to query hypervisor paging requests. "
  59                  "[Default=" __stringify(CMM_DEFAULT_DELAY) "]");
  60 module_param_named(hotplug_delay, hotplug_delay, uint, 0644);
  61 MODULE_PARM_DESC(hotplug_delay, "Delay (in seconds) after memory hotplug remove "
  62                  "before loaning resumes. "
  63                  "[Default=" __stringify(CMM_HOTPLUG_DELAY) "]");
  64 module_param_named(oom_kb, oom_kb, uint, 0644);
  65 MODULE_PARM_DESC(oom_kb, "Amount of memory in kb to free on OOM. "
  66                  "[Default=" __stringify(CMM_OOM_KB) "]");
  67 module_param_named(min_mem_mb, min_mem_mb, ulong, 0644);
  68 MODULE_PARM_DESC(min_mem_mb, "Minimum amount of memory (in MB) to not balloon. "
  69                  "[Default=" __stringify(CMM_MIN_MEM_MB) "]");
  70 module_param_named(debug, cmm_debug, uint, 0644);
  71 MODULE_PARM_DESC(debug, "Enable module debugging logging. Set to 1 to enable. "
  72                  "[Default=" __stringify(CMM_DEBUG) "]");
  73
  74 #define cmm_dbg(...) if (cmm_debug) { printk(KERN_INFO "cmm: "__VA_ARGS__); }
  75
  76 static atomic_long_t loaned_pages;
  77 static unsigned long loaned_pages_target;
  78 static unsigned long oom_freed_pages;
  79
  80 static LIST_HEAD(cmm_page_list);
  81 static DEFINE_SPINLOCK(cmm_lock);
  82
  83 static DEFINE_MUTEX(hotplug_mutex);
  84 static int hotplug_occurred; /* protected by the hotplug mutex */
  85
  86 static struct task_struct *cmm_thread_ptr;
  87
  88 static long plpar_page_set_loaned(struct page *page)
  89 {
  90         const unsigned long vpa = page_to_phys(page);
  91         unsigned long cmo_page_sz = cmo_get_page_size();
  92         long rc = 0;
  93         int i;
  94
  95         for (i = 0; !rc && i < PAGE_SIZE; i += cmo_page_sz)
  96                 rc = plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_LOANED, vpa + i, 0);
  97
  98         for (i -= cmo_page_sz; rc && i != 0; i -= cmo_page_sz)
  99                 plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_ACTIVE,
 100                                    vpa + i - cmo_page_sz, 0);
 101
 102         return rc;
 103 }
 104
 105 static long plpar_page_set_active(struct page *page)
 106 {
 107         const unsigned long vpa = page_to_phys(page);
 108         unsigned long cmo_page_sz = cmo_get_page_size();
 109         long rc = 0;
 110         int i;
 111
 112         for (i = 0; !rc && i < PAGE_SIZE; i += cmo_page_sz)
 113                 rc = plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_ACTIVE, vpa + i, 0);
 114
 115         for (i -= cmo_page_sz; rc && i != 0; i -= cmo_page_sz)
 116                 plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_LOANED,
 117                                    vpa + i - cmo_page_sz, 0);
 118
 119         return rc;
 120 }
 121
 122 /**
 123  * cmm_alloc_pages - Allocate pages and mark them as loaned
 124  * @nr: number of pages to allocate
 125  *
 126  * Return value:
 127  *      number of pages requested to be allocated which were not
 128  **/
 129 static long cmm_alloc_pages(long nr)
 130 {
 131         struct page *page;
 132         long rc;
 133
 134         cmm_dbg("Begin request for %ld pages\n", nr);
 135
 136         while (nr) {
 137                 /* Exit if a hotplug operation is in progress or occurred */
 138                 if (mutex_trylock(&hotplug_mutex)) {
 139                         if (hotplug_occurred) {
 140                                 mutex_unlock(&hotplug_mutex);
 141                                 break;
 142                         }
 143                         mutex_unlock(&hotplug_mutex);
 144                 } else {
 145                         break;
 146                 }
 147
 148                 page = alloc_page(GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY |
 149                                   __GFP_NOMEMALLOC);
 150                 if (!page)
 151                         break;
 152                 spin_lock(&cmm_lock);
 153                 rc = plpar_page_set_loaned(page);
 154                 if (rc) {
 155                         pr_err("%s: Can not set page to loaned. rc=%ld\n", __func__, rc);
 156                         spin_unlock(&cmm_lock);
 157                         __free_page(page);
 158                         break;
 159                 }
 160
 161                 list_add(&page->lru, &cmm_page_list);
 162                 atomic_long_inc(&loaned_pages);
 163                 adjust_managed_page_count(page, -1);
 164                 spin_unlock(&cmm_lock);
 165                 nr--;
 166         }
 167
 168         cmm_dbg("End request with %ld pages unfulfilled\n", nr);
 169         return nr;
 170 }
 171
 172 /**
 173  * cmm_free_pages - Free pages and mark them as active
 174  * @nr: number of pages to free
 175  *
 176  * Return value:
 177  *      number of pages requested to be freed which were not
 178  **/
 179 static long cmm_free_pages(long nr)
 180 {
 181         struct page *page, *tmp;
 182
 183         cmm_dbg("Begin free of %ld pages.\n", nr);
 184         spin_lock(&cmm_lock);
 185         list_for_each_entry_safe(page, tmp, &cmm_page_list, lru) {
 186                 if (!nr)
 187                         break;
 188                 plpar_page_set_active(page);
 189                 list_del(&page->lru);
 190                 adjust_managed_page_count(page, 1);
 191                 __free_page(page);
 192                 atomic_long_dec(&loaned_pages);
 193                 nr--;
 194         }
 195         spin_unlock(&cmm_lock);
 196         cmm_dbg("End request with %ld pages unfulfilled\n", nr);
 197         return nr;
 198 }
 199
 200 /**
 201  * cmm_oom_notify - OOM notifier
 202  * @self:       notifier block struct
 203  * @dummy:      not used
 204  * @parm:       returned - number of pages freed
 205  *
 206  * Return value:
 207  *      NOTIFY_OK
 208  **/
 209 static int cmm_oom_notify(struct notifier_block *self,
 210                           unsigned long dummy, void *parm)
 211 {
 212         unsigned long *freed = parm;
 213         long nr = KB2PAGES(oom_kb);
 214
 215         cmm_dbg("OOM processing started\n");
 216         nr = cmm_free_pages(nr);
 217         loaned_pages_target = atomic_long_read(&loaned_pages);
 218         *freed += KB2PAGES(oom_kb) - nr;
 219         oom_freed_pages += KB2PAGES(oom_kb) - nr;
 220         cmm_dbg("OOM processing complete\n");
 221         return NOTIFY_OK;
 222 }
 223
 224 /**
 225  * cmm_get_mpp - Read memory performance parameters
 226  *
 227  * Makes hcall to query the current page loan request from the hypervisor.
 228  *
 229  * Return value:
 230  *      nothing
 231  **/
 232 static void cmm_get_mpp(void)
 233 {
 234         const long __loaned_pages = atomic_long_read(&loaned_pages);
 235         const long total_pages = totalram_pages() + __loaned_pages;
 236         int rc;
 237         struct hvcall_mpp_data mpp_data;
 238         signed long active_pages_target, page_loan_request, target;
 239         signed long min_mem_pages = (min_mem_mb * 1024 * 1024) / PAGE_SIZE;
 240
 241         rc = h_get_mpp(&mpp_data);
 242
 243         if (rc != H_SUCCESS)
 244                 return;
 245
 246         page_loan_request = div_s64((s64)mpp_data.loan_request, PAGE_SIZE);
 247         target = page_loan_request + __loaned_pages;
 248
 249         if (target < 0 || total_pages < min_mem_pages)
 250                 target = 0;
 251
 252         if (target > oom_freed_pages)
 253                 target -= oom_freed_pages;
 254         else
 255                 target = 0;
 256
 257         active_pages_target = total_pages - target;
 258
 259         if (min_mem_pages > active_pages_target)
 260                 target = total_pages - min_mem_pages;
 261
 262         if (target < 0)
 263                 target = 0;
 264
 265         loaned_pages_target = target;
 266
 267         cmm_dbg("delta = %ld, loaned = %lu, target = %lu, oom = %lu, totalram = %lu\n",
 268                 page_loan_request, __loaned_pages, loaned_pages_target,
 269                 oom_freed_pages, totalram_pages());
 270 }
 271
 272 static struct notifier_block cmm_oom_nb = {
 273         .notifier_call = cmm_oom_notify
 274 };
 275
 276 /**
 277  * cmm_thread - CMM task thread
 278  * @dummy:      not used
 279  *
 280  * Return value:
 281  *      0
 282  **/
 283 static int cmm_thread(void *dummy)
 284 {
 285         unsigned long timeleft;
 286         long __loaned_pages;
 287
 288         while (1) {
 289                 timeleft = msleep_interruptible(delay * 1000);
 290
 291                 if (kthread_should_stop() || timeleft)
 292                         break;
 293
 294                 if (mutex_trylock(&hotplug_mutex)) {
 295                         if (hotplug_occurred) {
 296                                 hotplug_occurred = 0;
 297                                 mutex_unlock(&hotplug_mutex);
 298                                 cmm_dbg("Hotplug operation has occurred, "
 299                                                 "loaning activity suspended "
 300                                                 "for %d seconds.\n",
 301                                                 hotplug_delay);
 302                                 timeleft = msleep_interruptible(hotplug_delay *
 303                                                 1000);
 304                                 if (kthread_should_stop() || timeleft)
 305                                         break;
 306                                 continue;
 307                         }
 308                         mutex_unlock(&hotplug_mutex);
 309                 } else {
 310                         cmm_dbg("Hotplug operation in progress, activity "
 311                                         "suspended\n");
 312                         continue;
 313                 }
 314
 315                 cmm_get_mpp();
 316
 317                 __loaned_pages = atomic_long_read(&loaned_pages);
 318                 if (loaned_pages_target > __loaned_pages) {
 319                         if (cmm_alloc_pages(loaned_pages_target - __loaned_pages))
 320                                 loaned_pages_target = __loaned_pages;
 321                 } else if (loaned_pages_target < __loaned_pages)
 322                         cmm_free_pages(__loaned_pages - loaned_pages_target);
 323         }
 324         return 0;
 325 }
 326
 327 #define CMM_SHOW(name, format, args...)                 \
 328         static ssize_t show_##name(struct device *dev,  \
 329                                    struct device_attribute *attr,       \
 330                                    char *buf)                   \
 331         {                                                       \
 332                 return sprintf(buf, format, ##args);            \
 333         }                                                       \
 334         static DEVICE_ATTR(name, 0444, show_##name, NULL)
 335
 336 CMM_SHOW(loaned_kb, "%lu\n", PAGES2KB(atomic_long_read(&loaned_pages)));
 337 CMM_SHOW(loaned_target_kb, "%lu\n", PAGES2KB(loaned_pages_target));
 338
 339 static ssize_t show_oom_pages(struct device *dev,
 340                               struct device_attribute *attr, char *buf)
 341 {
 342         return sprintf(buf, "%lu\n", PAGES2KB(oom_freed_pages));
 343 }
 344
 345 static ssize_t store_oom_pages(struct device *dev,
 346                                struct device_attribute *attr,
 347                                const char *buf, size_t count)
 348 {
 349         unsigned long val = simple_strtoul (buf, NULL, 10);
 350
 351         if (!capable(CAP_SYS_ADMIN))
 352                 return -EPERM;
 353         if (val != 0)
 354                 return -EBADMSG;
 355
 356         oom_freed_pages = 0;
 357         return count;
 358 }
 359
 360 static DEVICE_ATTR(oom_freed_kb, 0644,
 361                    show_oom_pages, store_oom_pages);
 362
 363 static struct device_attribute *cmm_attrs[] = {
 364         &dev_attr_loaned_kb,
 365         &dev_attr_loaned_target_kb,
 366         &dev_attr_oom_freed_kb,
 367 };
 368
 369 static struct bus_type cmm_subsys = {
 370         .name = "cmm",
 371         .dev_name = "cmm",
 372 };
 373
 374 static void cmm_release_device(struct device *dev)
 375 {
 376 }
 377
 378 /**
 379  * cmm_sysfs_register - Register with sysfs
 380  *
 381  * Return value:
 382  *      0 on success / other on failure
 383  **/
 384 static int cmm_sysfs_register(struct device *dev)
 385 {
 386         int i, rc;
 387
 388         if ((rc = subsys_system_register(&cmm_subsys, NULL)))
 389                 return rc;
 390
 391         dev->id = 0;
 392         dev->bus = &cmm_subsys;
 393         dev->release = cmm_release_device;
 394
 395         if ((rc = device_register(dev)))
 396                 goto subsys_unregister;
 397
 398         for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++) {
 399                 if ((rc = device_create_file(dev, cmm_attrs[i])))
 400                         goto fail;
 401         }
 402
 403         return 0;
 404
 405 fail:
 406         while (--i >= 0)
 407                 device_remove_file(dev, cmm_attrs[i]);
 408         device_unregister(dev);
 409 subsys_unregister:
 410         bus_unregister(&cmm_subsys);
 411         return rc;
 412 }
 413
 414 /**
 415  * cmm_unregister_sysfs - Unregister from sysfs
 416  *
 417  **/
 418 static void cmm_unregister_sysfs(struct device *dev)
 419 {
 420         int i;
 421
 422         for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++)
 423                 device_remove_file(dev, cmm_attrs[i]);
 424         device_unregister(dev);
 425         bus_unregister(&cmm_subsys);
 426 }
 427
 428 /**
 429  * cmm_reboot_notifier - Make sure pages are not still marked as "loaned"
 430  *
 431  **/
 432 static int cmm_reboot_notifier(struct notifier_block *nb,
 433                                unsigned long action, void *unused)
 434 {
 435         if (action == SYS_RESTART) {
 436                 if (cmm_thread_ptr)
 437                         kthread_stop(cmm_thread_ptr);
 438                 cmm_thread_ptr = NULL;
 439                 cmm_free_pages(atomic_long_read(&loaned_pages));
 440         }
 441         return NOTIFY_DONE;
 442 }
 443
 444 static struct notifier_block cmm_reboot_nb = {
 445         .notifier_call = cmm_reboot_notifier,
 446 };
 447
 448 /**
 449  * cmm_memory_cb - Handle memory hotplug notifier calls
 450  * @self:       notifier block struct
 451  * @action:     action to take
 452  * @arg:        struct memory_notify data for handler
 453  *
 454  * Return value:
 455  *      NOTIFY_OK or notifier error based on subfunction return value
 456  *
 457  **/
 458 static int cmm_memory_cb(struct notifier_block *self,
 459                         unsigned long action, void *arg)
 460 {
 461         int ret = 0;
 462
 463         switch (action) {
 464         case MEM_GOING_OFFLINE:
 465                 mutex_lock(&hotplug_mutex);
 466                 hotplug_occurred = 1;
 467                 break;
 468         case MEM_OFFLINE:
 469         case MEM_CANCEL_OFFLINE:
 470                 mutex_unlock(&hotplug_mutex);
 471                 cmm_dbg("Memory offline operation complete.\n");
 472                 break;
 473         case MEM_GOING_ONLINE:
 474         case MEM_ONLINE:
 475         case MEM_CANCEL_ONLINE:
 476                 break;
 477         }
 478
 479         return notifier_from_errno(ret);
 480 }
 481
 482 static struct notifier_block cmm_mem_nb = {
 483         .notifier_call = cmm_memory_cb,
 484         .priority = CMM_MEM_HOTPLUG_PRI
 485 };
 486
 487 /**
 488  * cmm_init - Module initialization
 489  *
 490  * Return value:
 491  *      0 on success / other on failure
 492  **/
 493 static int cmm_init(void)
 494 {
 495         int rc;
 496
 497         if (!firmware_has_feature(FW_FEATURE_CMO))
 498                 return -EOPNOTSUPP;
 499
 500         if ((rc = register_oom_notifier(&cmm_oom_nb)) < 0)
 501                 return rc;
 502
 503         if ((rc = register_reboot_notifier(&cmm_reboot_nb)))
 504                 goto out_oom_notifier;
 505
 506         if ((rc = cmm_sysfs_register(&cmm_dev)))
 507                 goto out_reboot_notifier;
 508
 509         rc = register_memory_notifier(&cmm_mem_nb);
 510         if (rc)
 511                 goto out_unregister_notifier;
 512
 513         if (cmm_disabled)
 514                 return 0;
 515
 516         cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
 517         if (IS_ERR(cmm_thread_ptr)) {
 518                 rc = PTR_ERR(cmm_thread_ptr);
 519                 goto out_unregister_notifier;
 520         }
 521
 522         return 0;
 523 out_unregister_notifier:
 524         unregister_memory_notifier(&cmm_mem_nb);
 525         cmm_unregister_sysfs(&cmm_dev);
 526 out_reboot_notifier:
 527         unregister_reboot_notifier(&cmm_reboot_nb);
 528 out_oom_notifier:
 529         unregister_oom_notifier(&cmm_oom_nb);
 530         return rc;
 531 }
 532
 533 /**
 534  * cmm_exit - Module exit
 535  *
 536  * Return value:
 537  *      nothing
 538  **/
 539 static void cmm_exit(void)
 540 {
 541         if (cmm_thread_ptr)
 542                 kthread_stop(cmm_thread_ptr);
 543         unregister_oom_notifier(&cmm_oom_nb);
 544         unregister_reboot_notifier(&cmm_reboot_nb);
 545         unregister_memory_notifier(&cmm_mem_nb);
 546         cmm_free_pages(atomic_long_read(&loaned_pages));
 547         cmm_unregister_sysfs(&cmm_dev);
 548 }
 549
 550 /**
 551  * cmm_set_disable - Disable/Enable CMM
 552  *
 553  * Return value:
 554  *      0 on success / other on failure
 555  **/
 556 static int cmm_set_disable(const char *val, const struct kernel_param *kp)
 557 {
 558         int disable = simple_strtoul(val, NULL, 10);
 559
 560         if (disable != 0 && disable != 1)
 561                 return -EINVAL;
 562
 563         if (disable && !cmm_disabled) {
 564                 if (cmm_thread_ptr)
 565                         kthread_stop(cmm_thread_ptr);
 566                 cmm_thread_ptr = NULL;
 567                 cmm_free_pages(atomic_long_read(&loaned_pages));
 568         } else if (!disable && cmm_disabled) {
 569                 cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
 570                 if (IS_ERR(cmm_thread_ptr))
 571                         return PTR_ERR(cmm_thread_ptr);
 572         }
 573
 574         cmm_disabled = disable;
 575         return 0;
 576 }
 577
 578 module_param_call(disable, cmm_set_disable, param_get_uint,
 579                   &cmm_disabled, 0644);
 580 MODULE_PARM_DESC(disable, "Disable CMM. Set to 1 to disable. "
 581                  "[Default=" __stringify(CMM_DISABLE) "]");
 582
 583 module_init(cmm_init);
 584 module_exit(cmm_exit);