arch/sparc/mm/tsb.c

   1 /* arch/sparc64/mm/tsb.c
   2  *
   3  * Copyright (C) 2006, 2008 David S. Miller <davem@davemloft.net>
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/preempt.h>
   8 #include <linux/slab.h>
   9 #include <asm/page.h>
  10 #include <asm/pgtable.h>
  11 #include <asm/mmu_context.h>
  12 #include <asm/tsb.h>
  13 #include <asm/tlb.h>
  14 #include <asm/oplib.h>
  15
  16 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
  17
  18 static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long hash_shift, unsigned long nentries)
  19 {
  20         vaddr >>= hash_shift;
  21         return vaddr & (nentries - 1);
  22 }
  23
  24 static inline int tag_compare(unsigned long tag, unsigned long vaddr)
  25 {
  26         return (tag == (vaddr >> 22));
  27 }
  28
  29 /* TSB flushes need only occur on the processor initiating the address
  30  * space modification, not on each cpu the address space has run on.
  31  * Only the TLB flush needs that treatment.
  32  */
  33
  34 void flush_tsb_kernel_range(unsigned long start, unsigned long end)
  35 {
  36         unsigned long v;
  37
  38         for (v = start; v < end; v += PAGE_SIZE) {
  39                 unsigned long hash = tsb_hash(v, PAGE_SHIFT,
  40                                               KERNEL_TSB_NENTRIES);
  41                 struct tsb *ent = &swapper_tsb[hash];
  42
  43                 if (tag_compare(ent->tag, v))
  44                         ent->tag = (1UL << TSB_TAG_INVALID_BIT);
  45         }
  46 }
  47
  48 static void __flush_tsb_one_entry(unsigned long tsb, unsigned long v,
  49                                   unsigned long hash_shift,
  50                                   unsigned long nentries)
  51 {
  52         unsigned long tag, ent, hash;
  53
  54         v &= ~0x1UL;
  55         hash = tsb_hash(v, hash_shift, nentries);
  56         ent = tsb + (hash * sizeof(struct tsb));
  57         tag = (v >> 22UL);
  58
  59         tsb_flush(ent, tag);
  60 }
  61
  62 static void __flush_tsb_one(struct tlb_batch *tb, unsigned long hash_shift,
  63                             unsigned long tsb, unsigned long nentries)
  64 {
  65         unsigned long i;
  66
  67         for (i = 0; i < tb->tlb_nr; i++)
  68                 __flush_tsb_one_entry(tsb, tb->vaddrs[i], hash_shift, nentries);
  69 }
  70
  71 void flush_tsb_user(struct tlb_batch *tb)
  72 {
  73         struct mm_struct *mm = tb->mm;
  74         unsigned long nentries, base, flags;
  75
  76         spin_lock_irqsave(&mm->context.lock, flags);
  77
  78         base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
  79         nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
  80         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
  81                 base = __pa(base);
  82         __flush_tsb_one(tb, PAGE_SHIFT, base, nentries);
  83
  84 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
  85         if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
  86                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
  87                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
  88                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
  89                         base = __pa(base);
  90                 __flush_tsb_one(tb, REAL_HPAGE_SHIFT, base, nentries);
  91         }
  92 #endif
  93         spin_unlock_irqrestore(&mm->context.lock, flags);
  94 }
  95
  96 void flush_tsb_user_page(struct mm_struct *mm, unsigned long vaddr)
  97 {
  98         unsigned long nentries, base, flags;
  99
 100         spin_lock_irqsave(&mm->context.lock, flags);
 101
 102         base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
 103         nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
 104         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 105                 base = __pa(base);
 106         __flush_tsb_one_entry(base, vaddr, PAGE_SHIFT, nentries);
 107
 108 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 109         if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
 110                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
 111                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
 112                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 113                         base = __pa(base);
 114                 __flush_tsb_one_entry(base, vaddr, REAL_HPAGE_SHIFT, nentries);
 115         }
 116 #endif
 117         spin_unlock_irqrestore(&mm->context.lock, flags);
 118 }
 119
 120 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_8K
 121 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_8K
 122
 123 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 124 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_4MB
 125 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_4MB
 126 #endif
 127
 128 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
 129 {
 130         unsigned long tsb_reg, base, tsb_paddr;
 131         unsigned long page_sz, tte;
 132
 133         mm->context.tsb_block[tsb_idx].tsb_nentries =
 134                 tsb_bytes / sizeof(struct tsb);
 135
 136         switch (tsb_idx) {
 137         case MM_TSB_BASE:
 138                 base = TSBMAP_8K_BASE;
 139                 break;
 140 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 141         case MM_TSB_HUGE:
 142                 base = TSBMAP_4M_BASE;
 143                 break;
 144 #endif
 145         default:
 146                 BUG();
 147         }
 148
 149         tte = pgprot_val(PAGE_KERNEL_LOCKED);
 150         tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
 151         BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
 152
 153         /* Use the smallest page size that can map the whole TSB
 154          * in one TLB entry.
 155          */
 156         switch (tsb_bytes) {
 157         case 8192 << 0:
 158                 tsb_reg = 0x0UL;
 159 #ifdef DCACHE_ALIASING_POSSIBLE
 160                 base += (tsb_paddr & 8192);
 161 #endif
 162                 page_sz = 8192;
 163                 break;
 164
 165         case 8192 << 1:
 166                 tsb_reg = 0x1UL;
 167                 page_sz = 64 * 1024;
 168                 break;
 169
 170         case 8192 << 2:
 171                 tsb_reg = 0x2UL;
 172                 page_sz = 64 * 1024;
 173                 break;
 174
 175         case 8192 << 3:
 176                 tsb_reg = 0x3UL;
 177                 page_sz = 64 * 1024;
 178                 break;
 179
 180         case 8192 << 4:
 181                 tsb_reg = 0x4UL;
 182                 page_sz = 512 * 1024;
 183                 break;
 184
 185         case 8192 << 5:
 186                 tsb_reg = 0x5UL;
 187                 page_sz = 512 * 1024;
 188                 break;
 189
 190         case 8192 << 6:
 191                 tsb_reg = 0x6UL;
 192                 page_sz = 512 * 1024;
 193                 break;
 194
 195         case 8192 << 7:
 196                 tsb_reg = 0x7UL;
 197                 page_sz = 4 * 1024 * 1024;
 198                 break;
 199
 200         default:
 201                 printk(KERN_ERR "TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
 202                        current->comm, current->pid, tsb_bytes);
 203                 do_exit(SIGSEGV);
 204         }
 205         tte |= pte_sz_bits(page_sz);
 206
 207         if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
 208                 /* Physical mapping, no locked TLB entry for TSB.  */
 209                 tsb_reg |= tsb_paddr;
 210
 211                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
 212                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
 213                 mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
 214         } else {
 215                 tsb_reg |= base;
 216                 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
 217                 tte |= (tsb_paddr & ~(page_sz - 1UL));
 218
 219                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
 220                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
 221                 mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
 222         }
 223
 224         /* Setup the Hypervisor TSB descriptor.  */
 225         if (tlb_type == hypervisor) {
 226                 struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
 227
 228                 switch (tsb_idx) {
 229                 case MM_TSB_BASE:
 230                         hp->pgsz_idx = HV_PGSZ_IDX_BASE;
 231                         break;
 232 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 233                 case MM_TSB_HUGE:
 234                         hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
 235                         break;
 236 #endif
 237                 default:
 238                         BUG();
 239                 }
 240                 hp->assoc = 1;
 241                 hp->num_ttes = tsb_bytes / 16;
 242                 hp->ctx_idx = 0;
 243                 switch (tsb_idx) {
 244                 case MM_TSB_BASE:
 245                         hp->pgsz_mask = HV_PGSZ_MASK_BASE;
 246                         break;
 247 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 248                 case MM_TSB_HUGE:
 249                         hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
 250                         break;
 251 #endif
 252                 default:
 253                         BUG();
 254                 }
 255                 hp->tsb_base = tsb_paddr;
 256                 hp->resv = 0;
 257         }
 258 }
 259
 260 struct kmem_cache *pgtable_cache __read_mostly;
 261
 262 static struct kmem_cache *tsb_caches[8] __read_mostly;
 263
 264 static const char *tsb_cache_names[8] = {
 265         "tsb_8KB",
 266         "tsb_16KB",
 267         "tsb_32KB",
 268         "tsb_64KB",
 269         "tsb_128KB",
 270         "tsb_256KB",
 271         "tsb_512KB",
 272         "tsb_1MB",
 273 };
 274
 275 void __init pgtable_cache_init(void)
 276 {
 277         unsigned long i;
 278
 279         pgtable_cache = kmem_cache_create("pgtable_cache",
 280                                           PAGE_SIZE, PAGE_SIZE,
 281                                           0,
 282                                           _clear_page);
 283         if (!pgtable_cache) {
 284                 prom_printf("pgtable_cache_init(): Could not create!\n");
 285                 prom_halt();
 286         }
 287
 288         for (i = 0; i < 8; i++) {
 289                 unsigned long size = 8192 << i;
 290                 const char *name = tsb_cache_names[i];
 291
 292                 tsb_caches[i] = kmem_cache_create(name,
 293                                                   size, size,
 294                                                   0, NULL);
 295                 if (!tsb_caches[i]) {
 296                         prom_printf("Could not create %s cache\n", name);
 297                         prom_halt();
 298                 }
 299         }
 300 }
 301
 302 int sysctl_tsb_ratio = -2;
 303
 304 static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
 305 {
 306         unsigned long num_ents = (new_size / sizeof(struct tsb));
 307
 308         if (sysctl_tsb_ratio < 0)
 309                 return num_ents - (num_ents >> -sysctl_tsb_ratio);
 310         else
 311                 return num_ents + (num_ents >> sysctl_tsb_ratio);
 312 }
 313
 314 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
 315  * do_sparc64_fault() invokes this routine to try and grow it.
 316  *
 317  * When we reach the maximum TSB size supported, we stick ~0UL into
 318  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
 319  * will not trigger any longer.
 320  *
 321  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
 322  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
 323  * must be 512K aligned.  It also must be physically contiguous, so we
 324  * cannot use vmalloc().
 325  *
 326  * The idea here is to grow the TSB when the RSS of the process approaches
 327  * the number of entries that the current TSB can hold at once.  Currently,
 328  * we trigger when the RSS hits 3/4 of the TSB capacity.
 329  */
 330 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
 331 {
 332         unsigned long max_tsb_size = 1 * 1024 * 1024;
 333         unsigned long new_size, old_size, flags;
 334         struct tsb *old_tsb, *new_tsb;
 335         unsigned long new_cache_index, old_cache_index;
 336         unsigned long new_rss_limit;
 337         gfp_t gfp_flags;
 338
 339         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
 340                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
 341
 342         new_cache_index = 0;
 343         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
 344                 new_rss_limit = tsb_size_to_rss_limit(new_size);
 345                 if (new_rss_limit > rss)
 346                         break;
 347                 new_cache_index++;
 348         }
 349
 350         if (new_size == max_tsb_size)
 351                 new_rss_limit = ~0UL;
 352
 353 retry_tsb_alloc:
 354         gfp_flags = GFP_KERNEL;
 355         if (new_size > (PAGE_SIZE * 2))
 356                 gfp_flags |= __GFP_NOWARN | __GFP_NORETRY;
 357
 358         new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
 359                                         gfp_flags, numa_node_id());
 360         if (unlikely(!new_tsb)) {
 361                 /* Not being able to fork due to a high-order TSB
 362                  * allocation failure is very bad behavior.  Just back
 363                  * down to a 0-order allocation and force no TSB
 364                  * growing for this address space.
 365                  */
 366                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
 367                     new_cache_index > 0) {
 368                         new_cache_index = 0;
 369                         new_size = 8192;
 370                         new_rss_limit = ~0UL;
 371                         goto retry_tsb_alloc;
 372                 }
 373
 374                 /* If we failed on a TSB grow, we are under serious
 375                  * memory pressure so don't try to grow any more.
 376                  */
 377                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
 378                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
 379                 return;
 380         }
 381
 382         /* Mark all tags as invalid.  */
 383         tsb_init(new_tsb, new_size);
 384
 385         /* Ok, we are about to commit the changes.  If we are
 386          * growing an existing TSB the locking is very tricky,
 387          * so WATCH OUT!
 388          *
 389          * We have to hold mm->context.lock while committing to the
 390          * new TSB, this synchronizes us with processors in
 391          * flush_tsb_user() and switch_mm() for this address space.
 392          *
 393          * But even with that lock held, processors run asynchronously
 394          * accessing the old TSB via TLB miss handling.  This is OK
 395          * because those actions are just propagating state from the
 396          * Linux page tables into the TSB, page table mappings are not
 397          * being changed.  If a real fault occurs, the processor will
 398          * synchronize with us when it hits flush_tsb_user(), this is
 399          * also true for the case where vmscan is modifying the page
 400          * tables.  The only thing we need to be careful with is to
 401          * skip any locked TSB entries during copy_tsb().
 402          *
 403          * When we finish committing to the new TSB, we have to drop
 404          * the lock and ask all other cpus running this address space
 405          * to run tsb_context_switch() to see the new TSB table.
 406          */
 407         spin_lock_irqsave(&mm->context.lock, flags);
 408
 409         old_tsb = mm->context.tsb_block[tsb_index].tsb;
 410         old_cache_index =
 411                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
 412         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
 413                     sizeof(struct tsb));
 414
 415
 416         /* Handle multiple threads trying to grow the TSB at the same time.
 417          * One will get in here first, and bump the size and the RSS limit.
 418          * The others will get in here next and hit this check.
 419          */
 420         if (unlikely(old_tsb &&
 421                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
 422                 spin_unlock_irqrestore(&mm->context.lock, flags);
 423
 424                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
 425                 return;
 426         }
 427
 428         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
 429
 430         if (old_tsb) {
 431                 extern void copy_tsb(unsigned long old_tsb_base,
 432                                      unsigned long old_tsb_size,
 433                                      unsigned long new_tsb_base,
 434                                      unsigned long new_tsb_size);
 435                 unsigned long old_tsb_base = (unsigned long) old_tsb;
 436                 unsigned long new_tsb_base = (unsigned long) new_tsb;
 437
 438                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
 439                         old_tsb_base = __pa(old_tsb_base);
 440                         new_tsb_base = __pa(new_tsb_base);
 441                 }
 442                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
 443         }
 444
 445         mm->context.tsb_block[tsb_index].tsb = new_tsb;
 446         setup_tsb_params(mm, tsb_index, new_size);
 447
 448         spin_unlock_irqrestore(&mm->context.lock, flags);
 449
 450         /* If old_tsb is NULL, we're being invoked for the first time
 451          * from init_new_context().
 452          */
 453         if (old_tsb) {
 454                 /* Reload it on the local cpu.  */
 455                 tsb_context_switch(mm);
 456
 457                 /* Now force other processors to do the same.  */
 458                 preempt_disable();
 459                 smp_tsb_sync(mm);
 460                 preempt_enable();
 461
 462                 /* Now it is safe to free the old tsb.  */
 463                 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
 464         }
 465 }
 466
 467 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 468 {
 469 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 470         unsigned long huge_pte_count;
 471 #endif
 472         unsigned int i;
 473
 474         spin_lock_init(&mm->context.lock);
 475
 476         mm->context.sparc64_ctx_val = 0UL;
 477
 478 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 479         /* We reset it to zero because the fork() page copying
 480          * will re-increment the counters as the parent PTEs are
 481          * copied into the child address space.
 482          */
 483         huge_pte_count = mm->context.huge_pte_count;
 484         mm->context.huge_pte_count = 0;
 485 #endif
 486
 487         /* copy_mm() copies over the parent's mm_struct before calling
 488          * us, so we need to zero out the TSB pointer or else tsb_grow()
 489          * will be confused and think there is an older TSB to free up.
 490          */
 491         for (i = 0; i < MM_NUM_TSBS; i++)
 492                 mm->context.tsb_block[i].tsb = NULL;
 493
 494         /* If this is fork, inherit the parent's TSB size.  We would
 495          * grow it to that size on the first page fault anyways.
 496          */
 497         tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
 498
 499 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 500         if (unlikely(huge_pte_count))
 501                 tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
 502 #endif
 503
 504         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
 505                 return -ENOMEM;
 506
 507         return 0;
 508 }
 509
 510 static void tsb_destroy_one(struct tsb_config *tp)
 511 {
 512         unsigned long cache_index;
 513
 514         if (!tp->tsb)
 515                 return;
 516         cache_index = tp->tsb_reg_val & 0x7UL;
 517         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
 518         tp->tsb = NULL;
 519         tp->tsb_reg_val = 0UL;
 520 }
 521
 522 void destroy_context(struct mm_struct *mm)
 523 {
 524         unsigned long flags, i;
 525
 526         for (i = 0; i < MM_NUM_TSBS; i++)
 527                 tsb_destroy_one(&mm->context.tsb_block[i]);
 528
 529         spin_lock_irqsave(&ctx_alloc_lock, flags);
 530
 531         if (CTX_VALID(mm->context)) {
 532                 unsigned long nr = CTX_NRBITS(mm->context);
 533                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
 534         }
 535
 536         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
 537 }