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 <asm/system.h>
   9 #include <asm/page.h>
  10 #include <asm/tlbflush.h>
  11 #include <asm/tlb.h>
  12 #include <asm/mmu_context.h>
  13 #include <asm/pgtable.h>
  14 #include <asm/tsb.h>
  15 #include <asm/oplib.h>
  16
  17 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
  18
  19 static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long hash_shift, unsigned long nentries)
  20 {
  21         vaddr >>= hash_shift;
  22         return vaddr & (nentries - 1);
  23 }
  24
  25 static inline int tag_compare(unsigned long tag, unsigned long vaddr)
  26 {
  27         return (tag == (vaddr >> 22));
  28 }
  29
  30 /* TSB flushes need only occur on the processor initiating the address
  31  * space modification, not on each cpu the address space has run on.
  32  * Only the TLB flush needs that treatment.
  33  */
  34
  35 void flush_tsb_kernel_range(unsigned long start, unsigned long end)
  36 {
  37         unsigned long v;
  38
  39         for (v = start; v < end; v += PAGE_SIZE) {
  40                 unsigned long hash = tsb_hash(v, PAGE_SHIFT,
  41                                               KERNEL_TSB_NENTRIES);
  42                 struct tsb *ent = &swapper_tsb[hash];
  43
  44                 if (tag_compare(ent->tag, v))
  45                         ent->tag = (1UL << TSB_TAG_INVALID_BIT);
  46         }
  47 }
  48
  49 static void __flush_tsb_one(struct mmu_gather *mp, unsigned long hash_shift, unsigned long tsb, unsigned long nentries)
  50 {
  51         unsigned long i;
  52
  53         for (i = 0; i < mp->tlb_nr; i++) {
  54                 unsigned long v = mp->vaddrs[i];
  55                 unsigned long tag, ent, hash;
  56
  57                 v &= ~0x1UL;
  58
  59                 hash = tsb_hash(v, hash_shift, nentries);
  60                 ent = tsb + (hash * sizeof(struct tsb));
  61                 tag = (v >> 22UL);
  62
  63                 tsb_flush(ent, tag);
  64         }
  65 }
  66
  67 void flush_tsb_user(struct mmu_gather *mp)
  68 {
  69         struct mm_struct *mm = mp->mm;
  70         unsigned long nentries, base, flags;
  71
  72         spin_lock_irqsave(&mm->context.lock, flags);
  73
  74         base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
  75         nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
  76         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
  77                 base = __pa(base);
  78         __flush_tsb_one(mp, PAGE_SHIFT, base, nentries);
  79
  80 #ifdef CONFIG_HUGETLB_PAGE
  81         if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
  82                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
  83                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
  84                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
  85                         base = __pa(base);
  86                 __flush_tsb_one(mp, HPAGE_SHIFT, base, nentries);
  87         }
  88 #endif
  89         spin_unlock_irqrestore(&mm->context.lock, flags);
  90 }
  91
  92 #if defined(CONFIG_SPARC64_PAGE_SIZE_8KB)
  93 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_8K
  94 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_8K
  95 #elif defined(CONFIG_SPARC64_PAGE_SIZE_64KB)
  96 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_64K
  97 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_64K
  98 #else
  99 #error Broken base page size setting...
 100 #endif
 101
 102 #ifdef CONFIG_HUGETLB_PAGE
 103 #if defined(CONFIG_HUGETLB_PAGE_SIZE_64K)
 104 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_64K
 105 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_64K
 106 #elif defined(CONFIG_HUGETLB_PAGE_SIZE_512K)
 107 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_512K
 108 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_512K
 109 #elif defined(CONFIG_HUGETLB_PAGE_SIZE_4MB)
 110 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_4MB
 111 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_4MB
 112 #else
 113 #error Broken huge page size setting...
 114 #endif
 115 #endif
 116
 117 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
 118 {
 119         unsigned long tsb_reg, base, tsb_paddr;
 120         unsigned long page_sz, tte;
 121
 122         mm->context.tsb_block[tsb_idx].tsb_nentries =
 123                 tsb_bytes / sizeof(struct tsb);
 124
 125         base = TSBMAP_BASE;
 126         tte = pgprot_val(PAGE_KERNEL_LOCKED);
 127         tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
 128         BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
 129
 130         /* Use the smallest page size that can map the whole TSB
 131          * in one TLB entry.
 132          */
 133         switch (tsb_bytes) {
 134         case 8192 << 0:
 135                 tsb_reg = 0x0UL;
 136 #ifdef DCACHE_ALIASING_POSSIBLE
 137                 base += (tsb_paddr & 8192);
 138 #endif
 139                 page_sz = 8192;
 140                 break;
 141
 142         case 8192 << 1:
 143                 tsb_reg = 0x1UL;
 144                 page_sz = 64 * 1024;
 145                 break;
 146
 147         case 8192 << 2:
 148                 tsb_reg = 0x2UL;
 149                 page_sz = 64 * 1024;
 150                 break;
 151
 152         case 8192 << 3:
 153                 tsb_reg = 0x3UL;
 154                 page_sz = 64 * 1024;
 155                 break;
 156
 157         case 8192 << 4:
 158                 tsb_reg = 0x4UL;
 159                 page_sz = 512 * 1024;
 160                 break;
 161
 162         case 8192 << 5:
 163                 tsb_reg = 0x5UL;
 164                 page_sz = 512 * 1024;
 165                 break;
 166
 167         case 8192 << 6:
 168                 tsb_reg = 0x6UL;
 169                 page_sz = 512 * 1024;
 170                 break;
 171
 172         case 8192 << 7:
 173                 tsb_reg = 0x7UL;
 174                 page_sz = 4 * 1024 * 1024;
 175                 break;
 176
 177         default:
 178                 printk(KERN_ERR "TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
 179                        current->comm, current->pid, tsb_bytes);
 180                 do_exit(SIGSEGV);
 181         };
 182         tte |= pte_sz_bits(page_sz);
 183
 184         if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
 185                 /* Physical mapping, no locked TLB entry for TSB.  */
 186                 tsb_reg |= tsb_paddr;
 187
 188                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
 189                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
 190                 mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
 191         } else {
 192                 tsb_reg |= base;
 193                 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
 194                 tte |= (tsb_paddr & ~(page_sz - 1UL));
 195
 196                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
 197                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
 198                 mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
 199         }
 200
 201         /* Setup the Hypervisor TSB descriptor.  */
 202         if (tlb_type == hypervisor) {
 203                 struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
 204
 205                 switch (tsb_idx) {
 206                 case MM_TSB_BASE:
 207                         hp->pgsz_idx = HV_PGSZ_IDX_BASE;
 208                         break;
 209 #ifdef CONFIG_HUGETLB_PAGE
 210                 case MM_TSB_HUGE:
 211                         hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
 212                         break;
 213 #endif
 214                 default:
 215                         BUG();
 216                 };
 217                 hp->assoc = 1;
 218                 hp->num_ttes = tsb_bytes / 16;
 219                 hp->ctx_idx = 0;
 220                 switch (tsb_idx) {
 221                 case MM_TSB_BASE:
 222                         hp->pgsz_mask = HV_PGSZ_MASK_BASE;
 223                         break;
 224 #ifdef CONFIG_HUGETLB_PAGE
 225                 case MM_TSB_HUGE:
 226                         hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
 227                         break;
 228 #endif
 229                 default:
 230                         BUG();
 231                 };
 232                 hp->tsb_base = tsb_paddr;
 233                 hp->resv = 0;
 234         }
 235 }
 236
 237 static struct kmem_cache *tsb_caches[8] __read_mostly;
 238
 239 static const char *tsb_cache_names[8] = {
 240         "tsb_8KB",
 241         "tsb_16KB",
 242         "tsb_32KB",
 243         "tsb_64KB",
 244         "tsb_128KB",
 245         "tsb_256KB",
 246         "tsb_512KB",
 247         "tsb_1MB",
 248 };
 249
 250 void __init pgtable_cache_init(void)
 251 {
 252         unsigned long i;
 253
 254         for (i = 0; i < 8; i++) {
 255                 unsigned long size = 8192 << i;
 256                 const char *name = tsb_cache_names[i];
 257
 258                 tsb_caches[i] = kmem_cache_create(name,
 259                                                   size, size,
 260                                                   0, NULL);
 261                 if (!tsb_caches[i]) {
 262                         prom_printf("Could not create %s cache\n", name);
 263                         prom_halt();
 264                 }
 265         }
 266 }
 267
 268 int sysctl_tsb_ratio = -2;
 269
 270 static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
 271 {
 272         unsigned long num_ents = (new_size / sizeof(struct tsb));
 273
 274         if (sysctl_tsb_ratio < 0)
 275                 return num_ents - (num_ents >> -sysctl_tsb_ratio);
 276         else
 277                 return num_ents + (num_ents >> sysctl_tsb_ratio);
 278 }
 279
 280 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
 281  * do_sparc64_fault() invokes this routine to try and grow it.
 282  *
 283  * When we reach the maximum TSB size supported, we stick ~0UL into
 284  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
 285  * will not trigger any longer.
 286  *
 287  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
 288  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
 289  * must be 512K aligned.  It also must be physically contiguous, so we
 290  * cannot use vmalloc().
 291  *
 292  * The idea here is to grow the TSB when the RSS of the process approaches
 293  * the number of entries that the current TSB can hold at once.  Currently,
 294  * we trigger when the RSS hits 3/4 of the TSB capacity.
 295  */
 296 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
 297 {
 298         unsigned long max_tsb_size = 1 * 1024 * 1024;
 299         unsigned long new_size, old_size, flags;
 300         struct tsb *old_tsb, *new_tsb;
 301         unsigned long new_cache_index, old_cache_index;
 302         unsigned long new_rss_limit;
 303         gfp_t gfp_flags;
 304
 305         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
 306                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
 307
 308         new_cache_index = 0;
 309         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
 310                 new_rss_limit = tsb_size_to_rss_limit(new_size);
 311                 if (new_rss_limit > rss)
 312                         break;
 313                 new_cache_index++;
 314         }
 315
 316         if (new_size == max_tsb_size)
 317                 new_rss_limit = ~0UL;
 318
 319 retry_tsb_alloc:
 320         gfp_flags = GFP_KERNEL;
 321         if (new_size > (PAGE_SIZE * 2))
 322                 gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
 323
 324         new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
 325                                         gfp_flags, numa_node_id());
 326         if (unlikely(!new_tsb)) {
 327                 /* Not being able to fork due to a high-order TSB
 328                  * allocation failure is very bad behavior.  Just back
 329                  * down to a 0-order allocation and force no TSB
 330                  * growing for this address space.
 331                  */
 332                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
 333                     new_cache_index > 0) {
 334                         new_cache_index = 0;
 335                         new_size = 8192;
 336                         new_rss_limit = ~0UL;
 337                         goto retry_tsb_alloc;
 338                 }
 339
 340                 /* If we failed on a TSB grow, we are under serious
 341                  * memory pressure so don't try to grow any more.
 342                  */
 343                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
 344                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
 345                 return;
 346         }
 347
 348         /* Mark all tags as invalid.  */
 349         tsb_init(new_tsb, new_size);
 350
 351         /* Ok, we are about to commit the changes.  If we are
 352          * growing an existing TSB the locking is very tricky,
 353          * so WATCH OUT!
 354          *
 355          * We have to hold mm->context.lock while committing to the
 356          * new TSB, this synchronizes us with processors in
 357          * flush_tsb_user() and switch_mm() for this address space.
 358          *
 359          * But even with that lock held, processors run asynchronously
 360          * accessing the old TSB via TLB miss handling.  This is OK
 361          * because those actions are just propagating state from the
 362          * Linux page tables into the TSB, page table mappings are not
 363          * being changed.  If a real fault occurs, the processor will
 364          * synchronize with us when it hits flush_tsb_user(), this is
 365          * also true for the case where vmscan is modifying the page
 366          * tables.  The only thing we need to be careful with is to
 367          * skip any locked TSB entries during copy_tsb().
 368          *
 369          * When we finish committing to the new TSB, we have to drop
 370          * the lock and ask all other cpus running this address space
 371          * to run tsb_context_switch() to see the new TSB table.
 372          */
 373         spin_lock_irqsave(&mm->context.lock, flags);
 374
 375         old_tsb = mm->context.tsb_block[tsb_index].tsb;
 376         old_cache_index =
 377                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
 378         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
 379                     sizeof(struct tsb));
 380
 381
 382         /* Handle multiple threads trying to grow the TSB at the same time.
 383          * One will get in here first, and bump the size and the RSS limit.
 384          * The others will get in here next and hit this check.
 385          */
 386         if (unlikely(old_tsb &&
 387                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
 388                 spin_unlock_irqrestore(&mm->context.lock, flags);
 389
 390                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
 391                 return;
 392         }
 393
 394         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
 395
 396         if (old_tsb) {
 397                 extern void copy_tsb(unsigned long old_tsb_base,
 398                                      unsigned long old_tsb_size,
 399                                      unsigned long new_tsb_base,
 400                                      unsigned long new_tsb_size);
 401                 unsigned long old_tsb_base = (unsigned long) old_tsb;
 402                 unsigned long new_tsb_base = (unsigned long) new_tsb;
 403
 404                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
 405                         old_tsb_base = __pa(old_tsb_base);
 406                         new_tsb_base = __pa(new_tsb_base);
 407                 }
 408                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
 409         }
 410
 411         mm->context.tsb_block[tsb_index].tsb = new_tsb;
 412         setup_tsb_params(mm, tsb_index, new_size);
 413
 414         spin_unlock_irqrestore(&mm->context.lock, flags);
 415
 416         /* If old_tsb is NULL, we're being invoked for the first time
 417          * from init_new_context().
 418          */
 419         if (old_tsb) {
 420                 /* Reload it on the local cpu.  */
 421                 tsb_context_switch(mm);
 422
 423                 /* Now force other processors to do the same.  */
 424                 preempt_disable();
 425                 smp_tsb_sync(mm);
 426                 preempt_enable();
 427
 428                 /* Now it is safe to free the old tsb.  */
 429                 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
 430         }
 431 }
 432
 433 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 434 {
 435 #ifdef CONFIG_HUGETLB_PAGE
 436         unsigned long huge_pte_count;
 437 #endif
 438         unsigned int i;
 439
 440         spin_lock_init(&mm->context.lock);
 441
 442         mm->context.sparc64_ctx_val = 0UL;
 443
 444 #ifdef CONFIG_HUGETLB_PAGE
 445         /* We reset it to zero because the fork() page copying
 446          * will re-increment the counters as the parent PTEs are
 447          * copied into the child address space.
 448          */
 449         huge_pte_count = mm->context.huge_pte_count;
 450         mm->context.huge_pte_count = 0;
 451 #endif
 452
 453         /* copy_mm() copies over the parent's mm_struct before calling
 454          * us, so we need to zero out the TSB pointer or else tsb_grow()
 455          * will be confused and think there is an older TSB to free up.
 456          */
 457         for (i = 0; i < MM_NUM_TSBS; i++)
 458                 mm->context.tsb_block[i].tsb = NULL;
 459
 460         /* If this is fork, inherit the parent's TSB size.  We would
 461          * grow it to that size on the first page fault anyways.
 462          */
 463         tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
 464
 465 #ifdef CONFIG_HUGETLB_PAGE
 466         if (unlikely(huge_pte_count))
 467                 tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
 468 #endif
 469
 470         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
 471                 return -ENOMEM;
 472
 473         return 0;
 474 }
 475
 476 static void tsb_destroy_one(struct tsb_config *tp)
 477 {
 478         unsigned long cache_index;
 479
 480         if (!tp->tsb)
 481                 return;
 482         cache_index = tp->tsb_reg_val & 0x7UL;
 483         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
 484         tp->tsb = NULL;
 485         tp->tsb_reg_val = 0UL;
 486 }
 487
 488 void destroy_context(struct mm_struct *mm)
 489 {
 490         unsigned long flags, i;
 491
 492         for (i = 0; i < MM_NUM_TSBS; i++)
 493                 tsb_destroy_one(&mm->context.tsb_block[i]);
 494
 495         spin_lock_irqsave(&ctx_alloc_lock, flags);
 496
 497         if (CTX_VALID(mm->context)) {
 498                 unsigned long nr = CTX_NRBITS(mm->context);
 499                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
 500         }
 501
 502         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
 503 }