drivers/staging/zcache/ramster/ramster.c

   1 /*
   2  * ramster.c
   3  *
   4  * Copyright (c) 2010-2012, Dan Magenheimer, Oracle Corp.
   5  *
   6  * RAMster implements peer-to-peer transcendent memory, allowing a "cluster" of
   7  * kernels to dynamically pool their RAM so that a RAM-hungry workload on one
   8  * machine can temporarily and transparently utilize RAM on another machine
   9  * which is presumably idle or running a non-RAM-hungry workload.
  10  *
  11  * RAMster combines a clustering and messaging foundation based on the ocfs2
  12  * cluster layer with the in-kernel compression implementation of zcache, and
  13  * adds code to glue them together.  When a page is "put" to RAMster, it is
  14  * compressed and stored locally.  Periodically, a thread will "remotify" these
  15  * pages by sending them via messages to a remote machine.  When the page is
  16  * later needed as indicated by a page fault, a "get" is issued.  If the data
  17  * is local, it is uncompressed and the fault is resolved.  If the data is
  18  * remote, a message is sent to fetch the data and the faulting thread sleeps;
  19  * when the data arrives, the thread awakens, the data is decompressed and
  20  * the fault is resolved.
  21
  22  * As of V5, clusters up to eight nodes are supported; each node can remotify
  23  * pages to one specified node, so clusters can be configured as clients to
  24  * a "memory server".  Some simple policy is in place that will need to be
  25  * refined over time.  Larger clusters and fault-resistant protocols can also
  26  * be added over time.
  27  */
  28
  29 #include <linux/module.h>
  30 #include <linux/cpu.h>
  31 #include <linux/highmem.h>
  32 #include <linux/list.h>
  33 #include <linux/lzo.h>
  34 #include <linux/slab.h>
  35 #include <linux/spinlock.h>
  36 #include <linux/types.h>
  37 #include <linux/atomic.h>
  38 #include <linux/frontswap.h>
  39 #include "../tmem.h"
  40 #include "../zcache.h"
  41 #include "../zbud.h"
  42 #include "ramster.h"
  43 #include "ramster_nodemanager.h"
  44 #include "tcp.h"
  45 #include "debug.h"
  46
  47 #define RAMSTER_TESTING
  48
  49 #ifndef CONFIG_SYSFS
  50 #error "ramster needs sysfs to define cluster nodes to use"
  51 #endif
  52
  53 static bool use_cleancache __read_mostly;
  54 static bool use_frontswap __read_mostly;
  55 static bool use_frontswap_exclusive_gets __read_mostly;
  56
  57 /* These must be sysfs not debugfs as they are checked/used by userland!! */
  58 static unsigned long ramster_interface_revision __read_mostly =
  59         R2NM_API_VERSION; /* interface revision must match userspace! */
  60 static unsigned long ramster_pers_remotify_enable __read_mostly;
  61 static unsigned long ramster_eph_remotify_enable __read_mostly;
  62 static atomic_t ramster_remote_pers_pages = ATOMIC_INIT(0);
  63 #define MANUAL_NODES 8
  64 static bool ramster_nodes_manual_up[MANUAL_NODES] __read_mostly;
  65 static int ramster_remote_target_nodenum __read_mostly = -1;
  66
  67 /* Used by this code. */
  68 long ramster_flnodes;
  69 ssize_t ramster_foreign_eph_pages;
  70 ssize_t ramster_foreign_pers_pages;
  71 /* FIXME frontswap selfshrinking knobs in debugfs? */
  72
  73 static LIST_HEAD(ramster_rem_op_list);
  74 static DEFINE_SPINLOCK(ramster_rem_op_list_lock);
  75 static DEFINE_PER_CPU(struct ramster_preload, ramster_preloads);
  76
  77 static DEFINE_PER_CPU(unsigned char *, ramster_remoteputmem1);
  78 static DEFINE_PER_CPU(unsigned char *, ramster_remoteputmem2);
  79
  80 static struct kmem_cache *ramster_flnode_cache __read_mostly;
  81
  82 static struct flushlist_node *ramster_flnode_alloc(struct tmem_pool *pool)
  83 {
  84         struct flushlist_node *flnode = NULL;
  85         struct ramster_preload *kp;
  86
  87         kp = &__get_cpu_var(ramster_preloads);
  88         flnode = kp->flnode;
  89         BUG_ON(flnode == NULL);
  90         kp->flnode = NULL;
  91         inc_ramster_flnodes();
  92         return flnode;
  93 }
  94
  95 /* the "flush list" asynchronously collects pages to remotely flush */
  96 #define FLUSH_ENTIRE_OBJECT ((uint32_t)-1)
  97 static void ramster_flnode_free(struct flushlist_node *flnode,
  98                                 struct tmem_pool *pool)
  99 {
 100         dec_ramster_flnodes();
 101         BUG_ON(ramster_flnodes < 0);
 102         kmem_cache_free(ramster_flnode_cache, flnode);
 103 }
 104
 105 int ramster_do_preload_flnode(struct tmem_pool *pool)
 106 {
 107         struct ramster_preload *kp;
 108         struct flushlist_node *flnode;
 109         int ret = -ENOMEM;
 110
 111         BUG_ON(!irqs_disabled());
 112         if (unlikely(ramster_flnode_cache == NULL))
 113                 BUG();
 114         kp = &__get_cpu_var(ramster_preloads);
 115         flnode = kmem_cache_alloc(ramster_flnode_cache, GFP_ATOMIC);
 116         if (unlikely(flnode == NULL) && kp->flnode == NULL)
 117                 BUG();  /* FIXME handle more gracefully, but how??? */
 118         else if (kp->flnode == NULL)
 119                 kp->flnode = flnode;
 120         else
 121                 kmem_cache_free(ramster_flnode_cache, flnode);
 122         return ret;
 123 }
 124
 125 /*
 126  * Called by the message handler after a (still compressed) page has been
 127  * fetched from the remote machine in response to an "is_remote" tmem_get
 128  * or persistent tmem_localify.  For a tmem_get, "extra" is the address of
 129  * the page that is to be filled to successfully resolve the tmem_get; for
 130  * a (persistent) tmem_localify, "extra" is NULL (as the data is placed only
 131  * in the local zcache).  "data" points to "size" bytes of (compressed) data
 132  * passed in the message.  In the case of a persistent remote get, if
 133  * pre-allocation was successful (see ramster_repatriate_preload), the page
 134  * is placed into both local zcache and at "extra".
 135  */
 136 int ramster_localify(int pool_id, struct tmem_oid *oidp, uint32_t index,
 137                         char *data, unsigned int size, void *extra)
 138 {
 139         int ret = -ENOENT;
 140         unsigned long flags;
 141         struct tmem_pool *pool;
 142         bool eph, delete = false;
 143         void *pampd, *saved_hb;
 144         struct tmem_obj *obj;
 145
 146         pool = zcache_get_pool_by_id(LOCAL_CLIENT, pool_id);
 147         if (unlikely(pool == NULL))
 148                 /* pool doesn't exist anymore */
 149                 goto out;
 150         eph = is_ephemeral(pool);
 151         local_irq_save(flags);  /* FIXME: maybe only disable softirqs? */
 152         pampd = tmem_localify_get_pampd(pool, oidp, index, &obj, &saved_hb);
 153         if (pampd == NULL) {
 154                 /* hmmm... must have been a flush while waiting */
 155 #ifdef RAMSTER_TESTING
 156                 pr_err("UNTESTED pampd==NULL in ramster_localify\n");
 157 #endif
 158                 if (eph)
 159                         inc_ramster_remote_eph_pages_unsucc_get();
 160                 else
 161                         inc_ramster_remote_pers_pages_unsucc_get();
 162                 obj = NULL;
 163                 goto finish;
 164         } else if (unlikely(!pampd_is_remote(pampd))) {
 165                 /* hmmm... must have been a dup put while waiting */
 166 #ifdef RAMSTER_TESTING
 167                 pr_err("UNTESTED dup while waiting in ramster_localify\n");
 168 #endif
 169                 if (eph)
 170                         inc_ramster_remote_eph_pages_unsucc_get();
 171                 else
 172                         inc_ramster_remote_pers_pages_unsucc_get();
 173                 obj = NULL;
 174                 pampd = NULL;
 175                 ret = -EEXIST;
 176                 goto finish;
 177         } else if (size == 0) {
 178                 /* no remote data, delete the local is_remote pampd */
 179                 pampd = NULL;
 180                 if (eph)
 181                         inc_ramster_remote_eph_pages_unsucc_get();
 182                 else
 183                         BUG();
 184                 delete = true;
 185                 goto finish;
 186         }
 187         if (pampd_is_intransit(pampd)) {
 188                 /*
 189                  *  a pampd is marked intransit if it is remote and space has
 190                  *  been allocated for it locally (note, only happens for
 191                  *  persistent pages, in which case the remote copy is freed)
 192                  */
 193                 BUG_ON(eph);
 194                 pampd = pampd_mask_intransit_and_remote(pampd);
 195                 zbud_copy_to_zbud(pampd, data, size);
 196         } else {
 197                 /*
 198                  * setting pampd to NULL tells tmem_localify_finish to leave
 199                  * pampd alone... meaning it is left pointing to the
 200                  * remote copy
 201                  */
 202                 pampd = NULL;
 203                 obj = NULL;
 204         }
 205         /*
 206          * but in all cases, we decompress direct-to-memory to complete
 207          * the remotify and return success
 208          */
 209         BUG_ON(extra == NULL);
 210         zcache_decompress_to_page(data, size, (struct page *)extra);
 211         if (eph)
 212                 inc_ramster_remote_eph_pages_succ_get();
 213         else
 214                 inc_ramster_remote_pers_pages_succ_get();
 215         ret = 0;
 216 finish:
 217         tmem_localify_finish(obj, index, pampd, saved_hb, delete);
 218         zcache_put_pool(pool);
 219         local_irq_restore(flags);
 220 out:
 221         return ret;
 222 }
 223
 224 void ramster_pampd_new_obj(struct tmem_obj *obj)
 225 {
 226         obj->extra = NULL;
 227 }
 228
 229 void ramster_pampd_free_obj(struct tmem_pool *pool, struct tmem_obj *obj,
 230                                 bool pool_destroy)
 231 {
 232         struct flushlist_node *flnode;
 233
 234         BUG_ON(preemptible());
 235         if (obj->extra == NULL)
 236                 return;
 237         if (pool_destroy && is_ephemeral(pool))
 238                 /* FIXME don't bother with remote eph data for now */
 239                 return;
 240         BUG_ON(!pampd_is_remote(obj->extra));
 241         flnode = ramster_flnode_alloc(pool);
 242         flnode->xh.client_id = pampd_remote_node(obj->extra);
 243         flnode->xh.pool_id = pool->pool_id;
 244         flnode->xh.oid = obj->oid;
 245         flnode->xh.index = FLUSH_ENTIRE_OBJECT;
 246         flnode->rem_op.op = RAMSTER_REMOTIFY_FLUSH_OBJ;
 247         spin_lock(&ramster_rem_op_list_lock);
 248         list_add(&flnode->rem_op.list, &ramster_rem_op_list);
 249         spin_unlock(&ramster_rem_op_list_lock);
 250 }
 251
 252 /*
 253  * Called on a remote persistent tmem_get to attempt to preallocate
 254  * local storage for the data contained in the remote persistent page.
 255  * If successfully preallocated, returns the pampd, marked as remote and
 256  * in_transit.  Else returns NULL.  Note that the appropriate tmem data
 257  * structure must be locked.
 258  */
 259 void *ramster_pampd_repatriate_preload(void *pampd, struct tmem_pool *pool,
 260                                         struct tmem_oid *oidp, uint32_t index,
 261                                         bool *intransit)
 262 {
 263         int clen = pampd_remote_size(pampd), c;
 264         void *ret_pampd = NULL;
 265         unsigned long flags;
 266         struct tmem_handle th;
 267
 268         BUG_ON(!pampd_is_remote(pampd));
 269         BUG_ON(is_ephemeral(pool));
 270         if (use_frontswap_exclusive_gets)
 271                 /* don't need local storage */
 272                 goto out;
 273         if (pampd_is_intransit(pampd)) {
 274                 /*
 275                  * to avoid multiple allocations (and maybe a memory leak)
 276                  * don't preallocate if already in the process of being
 277                  * repatriated
 278                  */
 279                 *intransit = true;
 280                 goto out;
 281         }
 282         *intransit = false;
 283         local_irq_save(flags);
 284         th.client_id = pampd_remote_node(pampd);
 285         th.pool_id = pool->pool_id;
 286         th.oid = *oidp;
 287         th.index = index;
 288         ret_pampd = zcache_pampd_create(NULL, clen, true, false, &th);
 289         if (ret_pampd != NULL) {
 290                 /*
 291                  *  a pampd is marked intransit if it is remote and space has
 292                  *  been allocated for it locally (note, only happens for
 293                  *  persistent pages, in which case the remote copy is freed)
 294                  */
 295                 ret_pampd = pampd_mark_intransit(ret_pampd);
 296                 c = atomic_dec_return(&ramster_remote_pers_pages);
 297                 WARN_ON_ONCE(c < 0);
 298         } else {
 299                 inc_ramster_pers_pages_remote_nomem();
 300         }
 301         local_irq_restore(flags);
 302 out:
 303         return ret_pampd;
 304 }
 305
 306 /*
 307  * Called on a remote tmem_get to invoke a message to fetch the page.
 308  * Might sleep so no tmem locks can be held.  "extra" is passed
 309  * all the way through the round-trip messaging to ramster_localify.
 310  */
 311 int ramster_pampd_repatriate(void *fake_pampd, void *real_pampd,
 312                                 struct tmem_pool *pool,
 313                                 struct tmem_oid *oid, uint32_t index,
 314                                 bool free, void *extra)
 315 {
 316         struct tmem_xhandle xh;
 317         int ret;
 318
 319         if (pampd_is_intransit(real_pampd))
 320                 /* have local space pre-reserved, so free remote copy */
 321                 free = true;
 322         xh = tmem_xhandle_fill(LOCAL_CLIENT, pool, oid, index);
 323         /* unreliable request/response for now */
 324         ret = r2net_remote_async_get(&xh, free,
 325                                         pampd_remote_node(fake_pampd),
 326                                         pampd_remote_size(fake_pampd),
 327                                         pampd_remote_cksum(fake_pampd),
 328                                         extra);
 329         return ret;
 330 }
 331
 332 bool ramster_pampd_is_remote(void *pampd)
 333 {
 334         return pampd_is_remote(pampd);
 335 }
 336
 337 int ramster_pampd_replace_in_obj(void *new_pampd, struct tmem_obj *obj)
 338 {
 339         int ret = -1;
 340
 341         if (new_pampd != NULL) {
 342                 if (obj->extra == NULL)
 343                         obj->extra = new_pampd;
 344                 /* enforce that all remote pages in an object reside
 345                  * in the same node! */
 346                 else if (pampd_remote_node(new_pampd) !=
 347                                 pampd_remote_node((void *)(obj->extra)))
 348                         BUG();
 349                 ret = 0;
 350         }
 351         return ret;
 352 }
 353
 354 void *ramster_pampd_free(void *pampd, struct tmem_pool *pool,
 355                               struct tmem_oid *oid, uint32_t index, bool acct)
 356 {
 357         bool eph = is_ephemeral(pool);
 358         void *local_pampd = NULL;
 359         int c;
 360
 361         BUG_ON(preemptible());
 362         BUG_ON(!pampd_is_remote(pampd));
 363         WARN_ON(acct == false);
 364         if (oid == NULL) {
 365                 /*
 366                  * a NULL oid means to ignore this pampd free
 367                  * as the remote freeing will be handled elsewhere
 368                  */
 369         } else if (eph) {
 370                 /* FIXME remote flush optional but probably good idea */
 371         } else if (pampd_is_intransit(pampd)) {
 372                 /* did a pers remote get_and_free, so just free local */
 373                 local_pampd = pampd_mask_intransit_and_remote(pampd);
 374         } else {
 375                 struct flushlist_node *flnode =
 376                         ramster_flnode_alloc(pool);
 377
 378                 flnode->xh.client_id = pampd_remote_node(pampd);
 379                 flnode->xh.pool_id = pool->pool_id;
 380                 flnode->xh.oid = *oid;
 381                 flnode->xh.index = index;
 382                 flnode->rem_op.op = RAMSTER_REMOTIFY_FLUSH_PAGE;
 383                 spin_lock(&ramster_rem_op_list_lock);
 384                 list_add(&flnode->rem_op.list, &ramster_rem_op_list);
 385                 spin_unlock(&ramster_rem_op_list_lock);
 386                 c = atomic_dec_return(&ramster_remote_pers_pages);
 387                 WARN_ON_ONCE(c < 0);
 388         }
 389         return local_pampd;
 390 }
 391
 392 void ramster_count_foreign_pages(bool eph, int count)
 393 {
 394         BUG_ON(count != 1 && count != -1);
 395         if (eph) {
 396                 if (count > 0) {
 397                         inc_ramster_foreign_eph_pages();
 398                 } else {
 399                         dec_ramster_foreign_eph_pages();
 400                         WARN_ON_ONCE(ramster_foreign_eph_pages < 0);
 401                 }
 402         } else {
 403                 if (count > 0) {
 404                         inc_ramster_foreign_pers_pages();
 405                 } else {
 406                         dec_ramster_foreign_pers_pages();
 407                         WARN_ON_ONCE(ramster_foreign_pers_pages < 0);
 408                 }
 409         }
 410 }
 411
 412 /*
 413  * For now, just push over a few pages every few seconds to
 414  * ensure that it basically works
 415  */
 416 static struct workqueue_struct *ramster_remotify_workqueue;
 417 static void ramster_remotify_process(struct work_struct *work);
 418 static DECLARE_DELAYED_WORK(ramster_remotify_worker,
 419                 ramster_remotify_process);
 420
 421 static void ramster_remotify_queue_delayed_work(unsigned long delay)
 422 {
 423         if (!queue_delayed_work(ramster_remotify_workqueue,
 424                                 &ramster_remotify_worker, delay))
 425                 pr_err("ramster_remotify: bad workqueue\n");
 426 }
 427
 428 static void ramster_remote_flush_page(struct flushlist_node *flnode)
 429 {
 430         struct tmem_xhandle *xh;
 431         int remotenode, ret;
 432
 433         preempt_disable();
 434         xh = &flnode->xh;
 435         remotenode = flnode->xh.client_id;
 436         ret = r2net_remote_flush(xh, remotenode);
 437         if (ret >= 0)
 438                 inc_ramster_remote_pages_flushed();
 439         else
 440                 inc_ramster_remote_page_flushes_failed();
 441         preempt_enable_no_resched();
 442         ramster_flnode_free(flnode, NULL);
 443 }
 444
 445 static void ramster_remote_flush_object(struct flushlist_node *flnode)
 446 {
 447         struct tmem_xhandle *xh;
 448         int remotenode, ret;
 449
 450         preempt_disable();
 451         xh = &flnode->xh;
 452         remotenode = flnode->xh.client_id;
 453         ret = r2net_remote_flush_object(xh, remotenode);
 454         if (ret >= 0)
 455                 inc_ramster_remote_objects_flushed();
 456         else
 457                 inc_ramster_remote_object_flushes_failed();
 458         preempt_enable_no_resched();
 459         ramster_flnode_free(flnode, NULL);
 460 }
 461
 462 int ramster_remotify_pageframe(bool eph)
 463 {
 464         struct tmem_xhandle xh;
 465         unsigned int size;
 466         int remotenode, ret, zbuds;
 467         struct tmem_pool *pool;
 468         unsigned long flags;
 469         unsigned char cksum;
 470         char *p;
 471         int i, j;
 472         unsigned char *tmpmem[2];
 473         struct tmem_handle th[2];
 474         unsigned int zsize[2];
 475
 476         tmpmem[0] = __get_cpu_var(ramster_remoteputmem1);
 477         tmpmem[1] = __get_cpu_var(ramster_remoteputmem2);
 478         local_bh_disable();
 479         zbuds = zbud_make_zombie_lru(&th[0], &tmpmem[0], &zsize[0], eph);
 480         /* now OK to release lock set in caller */
 481         local_bh_enable();
 482         if (zbuds == 0)
 483                 goto out;
 484         BUG_ON(zbuds > 2);
 485         for (i = 0; i < zbuds; i++) {
 486                 xh.client_id = th[i].client_id;
 487                 xh.pool_id = th[i].pool_id;
 488                 xh.oid = th[i].oid;
 489                 xh.index = th[i].index;
 490                 size = zsize[i];
 491                 BUG_ON(size == 0 || size > zbud_max_buddy_size());
 492                 for (p = tmpmem[i], cksum = 0, j = 0; j < size; j++)
 493                         cksum += *p++;
 494                 ret = r2net_remote_put(&xh, tmpmem[i], size, eph, &remotenode);
 495                 if (ret != 0) {
 496                 /*
 497                  * This is some form of a memory leak... if the remote put
 498                  * fails, there will never be another attempt to remotify
 499                  * this page.  But since we've dropped the zv pointer,
 500                  * the page may have been freed or the data replaced
 501                  * so we can't just "put it back" in the remote op list.
 502                  * Even if we could, not sure where to put it in the list
 503                  * because there may be flushes that must be strictly
 504                  * ordered vs the put.  So leave this as a FIXME for now.
 505                  * But count them so we know if it becomes a problem.
 506                  */
 507                         if (eph)
 508                                 inc_ramster_eph_pages_remote_failed();
 509                         else
 510                                 inc_ramster_pers_pages_remote_failed();
 511                         break;
 512                 } else {
 513                         if (!eph)
 514                                 atomic_inc(&ramster_remote_pers_pages);
 515                 }
 516                 if (eph)
 517                         inc_ramster_eph_pages_remoted();
 518                 else
 519                         inc_ramster_pers_pages_remoted();
 520                 /*
 521                  * data was successfully remoted so change the local version to
 522                  * point to the remote node where it landed
 523                  */
 524                 local_bh_disable();
 525                 pool = zcache_get_pool_by_id(LOCAL_CLIENT, xh.pool_id);
 526                 local_irq_save(flags);
 527                 (void)tmem_replace(pool, &xh.oid, xh.index,
 528                                 pampd_make_remote(remotenode, size, cksum));
 529                 local_irq_restore(flags);
 530                 zcache_put_pool(pool);
 531                 local_bh_enable();
 532         }
 533 out:
 534         return zbuds;
 535 }
 536
 537 static void zcache_do_remotify_flushes(void)
 538 {
 539         struct ramster_remotify_hdr *rem_op;
 540         union remotify_list_node *u;
 541
 542         while (1) {
 543                 spin_lock(&ramster_rem_op_list_lock);
 544                 if (list_empty(&ramster_rem_op_list)) {
 545                         spin_unlock(&ramster_rem_op_list_lock);
 546                         goto out;
 547                 }
 548                 rem_op = list_first_entry(&ramster_rem_op_list,
 549                                 struct ramster_remotify_hdr, list);
 550                 list_del_init(&rem_op->list);
 551                 spin_unlock(&ramster_rem_op_list_lock);
 552                 u = (union remotify_list_node *)rem_op;
 553                 switch (rem_op->op) {
 554                 case RAMSTER_REMOTIFY_FLUSH_PAGE:
 555                         ramster_remote_flush_page((struct flushlist_node *)u);
 556                         break;
 557                 case RAMSTER_REMOTIFY_FLUSH_OBJ:
 558                         ramster_remote_flush_object((struct flushlist_node *)u);
 559                         break;
 560                 default:
 561                         BUG();
 562                 }
 563         }
 564 out:
 565         return;
 566 }
 567
 568 static void ramster_remotify_process(struct work_struct *work)
 569 {
 570         static bool remotify_in_progress;
 571         int i;
 572
 573         BUG_ON(irqs_disabled());
 574         if (remotify_in_progress)
 575                 goto requeue;
 576         if (ramster_remote_target_nodenum == -1)
 577                 goto requeue;
 578         remotify_in_progress = true;
 579         if (use_cleancache && ramster_eph_remotify_enable) {
 580                 for (i = 0; i < 100; i++) {
 581                         zcache_do_remotify_flushes();
 582                         (void)ramster_remotify_pageframe(true);
 583                 }
 584         }
 585         if (use_frontswap && ramster_pers_remotify_enable) {
 586                 for (i = 0; i < 100; i++) {
 587                         zcache_do_remotify_flushes();
 588                         (void)ramster_remotify_pageframe(false);
 589                 }
 590         }
 591         remotify_in_progress = false;
 592 requeue:
 593         ramster_remotify_queue_delayed_work(HZ);
 594 }
 595
 596 void __init ramster_remotify_init(void)
 597 {
 598         unsigned long n = 60UL;
 599         ramster_remotify_workqueue =
 600                 create_singlethread_workqueue("ramster_remotify");
 601         ramster_remotify_queue_delayed_work(n * HZ);
 602 }
 603
 604 static ssize_t ramster_manual_node_up_show(struct kobject *kobj,
 605                                 struct kobj_attribute *attr, char *buf)
 606 {
 607         int i;
 608         char *p = buf;
 609         for (i = 0; i < MANUAL_NODES; i++)
 610                 if (ramster_nodes_manual_up[i])
 611                         p += sprintf(p, "%d ", i);
 612         p += sprintf(p, "\n");
 613         return p - buf;
 614 }
 615
 616 static ssize_t ramster_manual_node_up_store(struct kobject *kobj,
 617                 struct kobj_attribute *attr, const char *buf, size_t count)
 618 {
 619         int err;
 620         unsigned long node_num;
 621
 622         err = kstrtoul(buf, 10, &node_num);
 623         if (err) {
 624                 pr_err("ramster: bad strtoul?\n");
 625                 return -EINVAL;
 626         }
 627         if (node_num >= MANUAL_NODES) {
 628                 pr_err("ramster: bad node_num=%lu?\n", node_num);
 629                 return -EINVAL;
 630         }
 631         if (ramster_nodes_manual_up[node_num]) {
 632                 pr_err("ramster: node %d already up, ignoring\n",
 633                                                         (int)node_num);
 634         } else {
 635                 ramster_nodes_manual_up[node_num] = true;
 636                 r2net_hb_node_up_manual((int)node_num);
 637         }
 638         return count;
 639 }
 640
 641 static struct kobj_attribute ramster_manual_node_up_attr = {
 642         .attr = { .name = "manual_node_up", .mode = 0644 },
 643         .show = ramster_manual_node_up_show,
 644         .store = ramster_manual_node_up_store,
 645 };
 646
 647 static ssize_t ramster_remote_target_nodenum_show(struct kobject *kobj,
 648                                 struct kobj_attribute *attr, char *buf)
 649 {
 650         if (ramster_remote_target_nodenum == -1UL)
 651                 return sprintf(buf, "unset\n");
 652         else
 653                 return sprintf(buf, "%d\n", ramster_remote_target_nodenum);
 654 }
 655
 656 static ssize_t ramster_remote_target_nodenum_store(struct kobject *kobj,
 657                 struct kobj_attribute *attr, const char *buf, size_t count)
 658 {
 659         int err;
 660         unsigned long node_num;
 661
 662         err = kstrtoul(buf, 10, &node_num);
 663         if (err) {
 664                 pr_err("ramster: bad strtoul?\n");
 665                 return -EINVAL;
 666         } else if (node_num == -1UL) {
 667                 pr_err("ramster: disabling all remotification, "
 668                         "data may still reside on remote nodes however\n");
 669                 return -EINVAL;
 670         } else if (node_num >= MANUAL_NODES) {
 671                 pr_err("ramster: bad node_num=%lu?\n", node_num);
 672                 return -EINVAL;
 673         } else if (!ramster_nodes_manual_up[node_num]) {
 674                 pr_err("ramster: node %d not up, ignoring setting "
 675                         "of remotification target\n", (int)node_num);
 676         } else if (r2net_remote_target_node_set((int)node_num) >= 0) {
 677                 pr_info("ramster: node %d set as remotification target\n",
 678                                 (int)node_num);
 679                 ramster_remote_target_nodenum = (int)node_num;
 680         } else {
 681                 pr_err("ramster: bad num to node node_num=%d?\n",
 682                                 (int)node_num);
 683                 return -EINVAL;
 684         }
 685         return count;
 686 }
 687
 688 static struct kobj_attribute ramster_remote_target_nodenum_attr = {
 689         .attr = { .name = "remote_target_nodenum", .mode = 0644 },
 690         .show = ramster_remote_target_nodenum_show,
 691         .store = ramster_remote_target_nodenum_store,
 692 };
 693
 694 #define RAMSTER_SYSFS_RO(_name) \
 695         static ssize_t ramster_##_name##_show(struct kobject *kobj, \
 696                                 struct kobj_attribute *attr, char *buf) \
 697         { \
 698                 return sprintf(buf, "%lu\n", ramster_##_name); \
 699         } \
 700         static struct kobj_attribute ramster_##_name##_attr = { \
 701                 .attr = { .name = __stringify(_name), .mode = 0444 }, \
 702                 .show = ramster_##_name##_show, \
 703         }
 704
 705 #define RAMSTER_SYSFS_RW(_name) \
 706         static ssize_t ramster_##_name##_show(struct kobject *kobj, \
 707                                 struct kobj_attribute *attr, char *buf) \
 708         { \
 709                 return sprintf(buf, "%lu\n", ramster_##_name); \
 710         } \
 711         static ssize_t ramster_##_name##_store(struct kobject *kobj, \
 712                 struct kobj_attribute *attr, const char *buf, size_t count) \
 713         { \
 714                 int err; \
 715                 unsigned long enable; \
 716                 err = kstrtoul(buf, 10, &enable); \
 717                 if (err) \
 718                         return -EINVAL; \
 719                 ramster_##_name = enable; \
 720                 return count; \
 721         } \
 722         static struct kobj_attribute ramster_##_name##_attr = { \
 723                 .attr = { .name = __stringify(_name), .mode = 0644 }, \
 724                 .show = ramster_##_name##_show, \
 725                 .store = ramster_##_name##_store, \
 726         }
 727
 728 #define RAMSTER_SYSFS_RO_ATOMIC(_name) \
 729         static ssize_t ramster_##_name##_show(struct kobject *kobj, \
 730                                 struct kobj_attribute *attr, char *buf) \
 731         { \
 732             return sprintf(buf, "%d\n", atomic_read(&ramster_##_name)); \
 733         } \
 734         static struct kobj_attribute ramster_##_name##_attr = { \
 735                 .attr = { .name = __stringify(_name), .mode = 0444 }, \
 736                 .show = ramster_##_name##_show, \
 737         }
 738
 739 RAMSTER_SYSFS_RO(interface_revision);
 740 RAMSTER_SYSFS_RO_ATOMIC(remote_pers_pages);
 741 RAMSTER_SYSFS_RW(pers_remotify_enable);
 742 RAMSTER_SYSFS_RW(eph_remotify_enable);
 743
 744 static struct attribute *ramster_attrs[] = {
 745         &ramster_interface_revision_attr.attr,
 746         &ramster_remote_pers_pages_attr.attr,
 747         &ramster_manual_node_up_attr.attr,
 748         &ramster_remote_target_nodenum_attr.attr,
 749         &ramster_pers_remotify_enable_attr.attr,
 750         &ramster_eph_remotify_enable_attr.attr,
 751         NULL,
 752 };
 753
 754 static struct attribute_group ramster_attr_group = {
 755         .attrs = ramster_attrs,
 756         .name = "ramster",
 757 };
 758
 759 /*
 760  * frontswap selfshrinking
 761  */
 762
 763 /* In HZ, controls frequency of worker invocation. */
 764 static unsigned int selfshrink_interval __read_mostly = 5;
 765 /* Enable/disable with sysfs. */
 766 static bool frontswap_selfshrinking __read_mostly;
 767
 768 static void selfshrink_process(struct work_struct *work);
 769 static DECLARE_DELAYED_WORK(selfshrink_worker, selfshrink_process);
 770
 771 /* Enable/disable with kernel boot option. */
 772 static bool use_frontswap_selfshrink __initdata = true;
 773
 774 /*
 775  * The default values for the following parameters were deemed reasonable
 776  * by experimentation, may be workload-dependent, and can all be
 777  * adjusted via sysfs.
 778  */
 779
 780 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
 781 static unsigned int frontswap_hysteresis __read_mostly = 20;
 782
 783 /*
 784  * Number of selfshrink worker invocations to wait before observing that
 785  * frontswap selfshrinking should commence. Note that selfshrinking does
 786  * not use a separate worker thread.
 787  */
 788 static unsigned int frontswap_inertia __read_mostly = 3;
 789
 790 /* Countdown to next invocation of frontswap_shrink() */
 791 static unsigned long frontswap_inertia_counter;
 792
 793 /*
 794  * Invoked by the selfshrink worker thread, uses current number of pages
 795  * in frontswap (frontswap_curr_pages()), previous status, and control
 796  * values (hysteresis and inertia) to determine if frontswap should be
 797  * shrunk and what the new frontswap size should be.  Note that
 798  * frontswap_shrink is essentially a partial swapoff that immediately
 799  * transfers pages from the "swap device" (frontswap) back into kernel
 800  * RAM; despite the name, frontswap "shrinking" is very different from
 801  * the "shrinker" interface used by the kernel MM subsystem to reclaim
 802  * memory.
 803  */
 804 static void frontswap_selfshrink(void)
 805 {
 806         static unsigned long cur_frontswap_pages;
 807         static unsigned long last_frontswap_pages;
 808         static unsigned long tgt_frontswap_pages;
 809
 810         last_frontswap_pages = cur_frontswap_pages;
 811         cur_frontswap_pages = frontswap_curr_pages();
 812         if (!cur_frontswap_pages ||
 813                         (cur_frontswap_pages > last_frontswap_pages)) {
 814                 frontswap_inertia_counter = frontswap_inertia;
 815                 return;
 816         }
 817         if (frontswap_inertia_counter && --frontswap_inertia_counter)
 818                 return;
 819         if (cur_frontswap_pages <= frontswap_hysteresis)
 820                 tgt_frontswap_pages = 0;
 821         else
 822                 tgt_frontswap_pages = cur_frontswap_pages -
 823                         (cur_frontswap_pages / frontswap_hysteresis);
 824         frontswap_shrink(tgt_frontswap_pages);
 825 }
 826
 827 static int __init ramster_nofrontswap_selfshrink_setup(char *s)
 828 {
 829         use_frontswap_selfshrink = false;
 830         return 1;
 831 }
 832
 833 __setup("noselfshrink", ramster_nofrontswap_selfshrink_setup);
 834
 835 static void selfshrink_process(struct work_struct *work)
 836 {
 837         if (frontswap_selfshrinking && frontswap_enabled) {
 838                 frontswap_selfshrink();
 839                 schedule_delayed_work(&selfshrink_worker,
 840                         selfshrink_interval * HZ);
 841         }
 842 }
 843
 844 void ramster_cpu_up(int cpu)
 845 {
 846         unsigned char *p1 = kzalloc(PAGE_SIZE, GFP_KERNEL | __GFP_REPEAT);
 847         unsigned char *p2 = kzalloc(PAGE_SIZE, GFP_KERNEL | __GFP_REPEAT);
 848         BUG_ON(!p1 || !p2);
 849         per_cpu(ramster_remoteputmem1, cpu) = p1;
 850         per_cpu(ramster_remoteputmem2, cpu) = p2;
 851 }
 852
 853 void ramster_cpu_down(int cpu)
 854 {
 855         struct ramster_preload *kp;
 856
 857         kfree(per_cpu(ramster_remoteputmem1, cpu));
 858         per_cpu(ramster_remoteputmem1, cpu) = NULL;
 859         kfree(per_cpu(ramster_remoteputmem2, cpu));
 860         per_cpu(ramster_remoteputmem2, cpu) = NULL;
 861         kp = &per_cpu(ramster_preloads, cpu);
 862         if (kp->flnode) {
 863                 kmem_cache_free(ramster_flnode_cache, kp->flnode);
 864                 kp->flnode = NULL;
 865         }
 866 }
 867
 868 void ramster_register_pamops(struct tmem_pamops *pamops)
 869 {
 870         pamops->free_obj = ramster_pampd_free_obj;
 871         pamops->new_obj = ramster_pampd_new_obj;
 872         pamops->replace_in_obj = ramster_pampd_replace_in_obj;
 873         pamops->is_remote = ramster_pampd_is_remote;
 874         pamops->repatriate = ramster_pampd_repatriate;
 875         pamops->repatriate_preload = ramster_pampd_repatriate_preload;
 876 }
 877
 878 void __init ramster_init(bool cleancache, bool frontswap,
 879                                 bool frontswap_exclusive_gets)
 880 {
 881         int ret = 0;
 882
 883         if (cleancache)
 884                 use_cleancache = true;
 885         if (frontswap)
 886                 use_frontswap = true;
 887         if (frontswap_exclusive_gets)
 888                 use_frontswap_exclusive_gets = true;
 889         ramster_debugfs_init();
 890         ret = sysfs_create_group(mm_kobj, &ramster_attr_group);
 891         if (ret)
 892                 pr_err("ramster: can't create sysfs for ramster\n");
 893         (void)r2net_register_handlers();
 894         INIT_LIST_HEAD(&ramster_rem_op_list);
 895         ramster_flnode_cache = kmem_cache_create("ramster_flnode",
 896                                 sizeof(struct flushlist_node), 0, 0, NULL);
 897         frontswap_selfshrinking = use_frontswap_selfshrink;
 898         if (frontswap_selfshrinking) {
 899                 pr_info("ramster: Initializing frontswap selfshrink driver.\n");
 900                 schedule_delayed_work(&selfshrink_worker,
 901                                         selfshrink_interval * HZ);
 902         }
 903         ramster_remotify_init();
 904 }