Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[linux-2.6] / arch / sparc64 / mm / tsb.c
1 /* arch/sparc64/mm/tsb.c
2  *
3  * Copyright (C) 2006 David S. Miller <davem@davemloft.net>
4  */
5
6 #include <linux/kernel.h>
7 #include <asm/system.h>
8 #include <asm/page.h>
9 #include <asm/tlbflush.h>
10 #include <asm/tlb.h>
11 #include <asm/mmu_context.h>
12 #include <asm/pgtable.h>
13 #include <asm/tsb.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                         membar_storeload_storestore();
46                 }
47         }
48 }
49
50 static void __flush_tsb_one(struct mmu_gather *mp, unsigned long hash_shift, unsigned long tsb, unsigned long nentries)
51 {
52         unsigned long i;
53
54         for (i = 0; i < mp->tlb_nr; i++) {
55                 unsigned long v = mp->vaddrs[i];
56                 unsigned long tag, ent, hash;
57
58                 v &= ~0x1UL;
59
60                 hash = tsb_hash(v, hash_shift, nentries);
61                 ent = tsb + (hash * sizeof(struct tsb));
62                 tag = (v >> 22UL);
63
64                 tsb_flush(ent, tag);
65         }
66 }
67
68 void flush_tsb_user(struct mmu_gather *mp)
69 {
70         struct mm_struct *mm = mp->mm;
71         unsigned long nentries, base, flags;
72
73         spin_lock_irqsave(&mm->context.lock, flags);
74
75         base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
76         nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
77         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
78                 base = __pa(base);
79         __flush_tsb_one(mp, PAGE_SHIFT, base, nentries);
80
81 #ifdef CONFIG_HUGETLB_PAGE
82         if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
83                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
84                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
85                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
86                         base = __pa(base);
87                 __flush_tsb_one(mp, HPAGE_SHIFT, base, nentries);
88         }
89 #endif
90         spin_unlock_irqrestore(&mm->context.lock, flags);
91 }
92
93 #if defined(CONFIG_SPARC64_PAGE_SIZE_8KB)
94 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_8K
95 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_8K
96 #elif defined(CONFIG_SPARC64_PAGE_SIZE_64KB)
97 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_64K
98 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_64K
99 #elif defined(CONFIG_SPARC64_PAGE_SIZE_512KB)
100 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_512K
101 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_512K
102 #elif defined(CONFIG_SPARC64_PAGE_SIZE_4MB)
103 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_4MB
104 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_4MB
105 #else
106 #error Broken base page size setting...
107 #endif
108
109 #ifdef CONFIG_HUGETLB_PAGE
110 #if defined(CONFIG_HUGETLB_PAGE_SIZE_64K)
111 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_64K
112 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_64K
113 #elif defined(CONFIG_HUGETLB_PAGE_SIZE_512K)
114 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_512K
115 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_512K
116 #elif defined(CONFIG_HUGETLB_PAGE_SIZE_4MB)
117 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_4MB
118 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_4MB
119 #else
120 #error Broken huge page size setting...
121 #endif
122 #endif
123
124 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
125 {
126         unsigned long tsb_reg, base, tsb_paddr;
127         unsigned long page_sz, tte;
128
129         mm->context.tsb_block[tsb_idx].tsb_nentries =
130                 tsb_bytes / sizeof(struct tsb);
131
132         base = TSBMAP_BASE;
133         tte = pgprot_val(PAGE_KERNEL_LOCKED);
134         tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
135         BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
136
137         /* Use the smallest page size that can map the whole TSB
138          * in one TLB entry.
139          */
140         switch (tsb_bytes) {
141         case 8192 << 0:
142                 tsb_reg = 0x0UL;
143 #ifdef DCACHE_ALIASING_POSSIBLE
144                 base += (tsb_paddr & 8192);
145 #endif
146                 page_sz = 8192;
147                 break;
148
149         case 8192 << 1:
150                 tsb_reg = 0x1UL;
151                 page_sz = 64 * 1024;
152                 break;
153
154         case 8192 << 2:
155                 tsb_reg = 0x2UL;
156                 page_sz = 64 * 1024;
157                 break;
158
159         case 8192 << 3:
160                 tsb_reg = 0x3UL;
161                 page_sz = 64 * 1024;
162                 break;
163
164         case 8192 << 4:
165                 tsb_reg = 0x4UL;
166                 page_sz = 512 * 1024;
167                 break;
168
169         case 8192 << 5:
170                 tsb_reg = 0x5UL;
171                 page_sz = 512 * 1024;
172                 break;
173
174         case 8192 << 6:
175                 tsb_reg = 0x6UL;
176                 page_sz = 512 * 1024;
177                 break;
178
179         case 8192 << 7:
180                 tsb_reg = 0x7UL;
181                 page_sz = 4 * 1024 * 1024;
182                 break;
183
184         default:
185                 BUG();
186         };
187         tte |= pte_sz_bits(page_sz);
188
189         if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
190                 /* Physical mapping, no locked TLB entry for TSB.  */
191                 tsb_reg |= tsb_paddr;
192
193                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
194                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
195                 mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
196         } else {
197                 tsb_reg |= base;
198                 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
199                 tte |= (tsb_paddr & ~(page_sz - 1UL));
200
201                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
202                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
203                 mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
204         }
205
206         /* Setup the Hypervisor TSB descriptor.  */
207         if (tlb_type == hypervisor) {
208                 struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
209
210                 switch (tsb_idx) {
211                 case MM_TSB_BASE:
212                         hp->pgsz_idx = HV_PGSZ_IDX_BASE;
213                         break;
214 #ifdef CONFIG_HUGETLB_PAGE
215                 case MM_TSB_HUGE:
216                         hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
217                         break;
218 #endif
219                 default:
220                         BUG();
221                 };
222                 hp->assoc = 1;
223                 hp->num_ttes = tsb_bytes / 16;
224                 hp->ctx_idx = 0;
225                 switch (tsb_idx) {
226                 case MM_TSB_BASE:
227                         hp->pgsz_mask = HV_PGSZ_MASK_BASE;
228                         break;
229 #ifdef CONFIG_HUGETLB_PAGE
230                 case MM_TSB_HUGE:
231                         hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
232                         break;
233 #endif
234                 default:
235                         BUG();
236                 };
237                 hp->tsb_base = tsb_paddr;
238                 hp->resv = 0;
239         }
240 }
241
242 static struct kmem_cache *tsb_caches[8] __read_mostly;
243
244 static const char *tsb_cache_names[8] = {
245         "tsb_8KB",
246         "tsb_16KB",
247         "tsb_32KB",
248         "tsb_64KB",
249         "tsb_128KB",
250         "tsb_256KB",
251         "tsb_512KB",
252         "tsb_1MB",
253 };
254
255 void __init pgtable_cache_init(void)
256 {
257         unsigned long i;
258
259         for (i = 0; i < 8; i++) {
260                 unsigned long size = 8192 << i;
261                 const char *name = tsb_cache_names[i];
262
263                 tsb_caches[i] = kmem_cache_create(name,
264                                                   size, size,
265                                                   0, NULL);
266                 if (!tsb_caches[i]) {
267                         prom_printf("Could not create %s cache\n", name);
268                         prom_halt();
269                 }
270         }
271 }
272
273 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
274  * do_sparc64_fault() invokes this routine to try and grow it.
275  *
276  * When we reach the maximum TSB size supported, we stick ~0UL into
277  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
278  * will not trigger any longer.
279  *
280  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
281  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
282  * must be 512K aligned.  It also must be physically contiguous, so we
283  * cannot use vmalloc().
284  *
285  * The idea here is to grow the TSB when the RSS of the process approaches
286  * the number of entries that the current TSB can hold at once.  Currently,
287  * we trigger when the RSS hits 3/4 of the TSB capacity.
288  */
289 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
290 {
291         unsigned long max_tsb_size = 1 * 1024 * 1024;
292         unsigned long new_size, old_size, flags;
293         struct tsb *old_tsb, *new_tsb;
294         unsigned long new_cache_index, old_cache_index;
295         unsigned long new_rss_limit;
296         gfp_t gfp_flags;
297
298         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
299                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
300
301         new_cache_index = 0;
302         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
303                 unsigned long n_entries = new_size / sizeof(struct tsb);
304
305                 n_entries = (n_entries * 3) / 4;
306                 if (n_entries > rss)
307                         break;
308
309                 new_cache_index++;
310         }
311
312         if (new_size == max_tsb_size)
313                 new_rss_limit = ~0UL;
314         else
315                 new_rss_limit = ((new_size / sizeof(struct tsb)) * 3) / 4;
316
317 retry_tsb_alloc:
318         gfp_flags = GFP_KERNEL;
319         if (new_size > (PAGE_SIZE * 2))
320                 gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
321
322         new_tsb = kmem_cache_alloc(tsb_caches[new_cache_index], gfp_flags);
323         if (unlikely(!new_tsb)) {
324                 /* Not being able to fork due to a high-order TSB
325                  * allocation failure is very bad behavior.  Just back
326                  * down to a 0-order allocation and force no TSB
327                  * growing for this address space.
328                  */
329                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
330                     new_cache_index > 0) {
331                         new_cache_index = 0;
332                         new_size = 8192;
333                         new_rss_limit = ~0UL;
334                         goto retry_tsb_alloc;
335                 }
336
337                 /* If we failed on a TSB grow, we are under serious
338                  * memory pressure so don't try to grow any more.
339                  */
340                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
341                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
342                 return;
343         }
344
345         /* Mark all tags as invalid.  */
346         tsb_init(new_tsb, new_size);
347
348         /* Ok, we are about to commit the changes.  If we are
349          * growing an existing TSB the locking is very tricky,
350          * so WATCH OUT!
351          *
352          * We have to hold mm->context.lock while committing to the
353          * new TSB, this synchronizes us with processors in
354          * flush_tsb_user() and switch_mm() for this address space.
355          *
356          * But even with that lock held, processors run asynchronously
357          * accessing the old TSB via TLB miss handling.  This is OK
358          * because those actions are just propagating state from the
359          * Linux page tables into the TSB, page table mappings are not
360          * being changed.  If a real fault occurs, the processor will
361          * synchronize with us when it hits flush_tsb_user(), this is
362          * also true for the case where vmscan is modifying the page
363          * tables.  The only thing we need to be careful with is to
364          * skip any locked TSB entries during copy_tsb().
365          *
366          * When we finish committing to the new TSB, we have to drop
367          * the lock and ask all other cpus running this address space
368          * to run tsb_context_switch() to see the new TSB table.
369          */
370         spin_lock_irqsave(&mm->context.lock, flags);
371
372         old_tsb = mm->context.tsb_block[tsb_index].tsb;
373         old_cache_index =
374                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
375         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
376                     sizeof(struct tsb));
377
378
379         /* Handle multiple threads trying to grow the TSB at the same time.
380          * One will get in here first, and bump the size and the RSS limit.
381          * The others will get in here next and hit this check.
382          */
383         if (unlikely(old_tsb &&
384                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
385                 spin_unlock_irqrestore(&mm->context.lock, flags);
386
387                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
388                 return;
389         }
390
391         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
392
393         if (old_tsb) {
394                 extern void copy_tsb(unsigned long old_tsb_base,
395                                      unsigned long old_tsb_size,
396                                      unsigned long new_tsb_base,
397                                      unsigned long new_tsb_size);
398                 unsigned long old_tsb_base = (unsigned long) old_tsb;
399                 unsigned long new_tsb_base = (unsigned long) new_tsb;
400
401                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
402                         old_tsb_base = __pa(old_tsb_base);
403                         new_tsb_base = __pa(new_tsb_base);
404                 }
405                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
406         }
407
408         mm->context.tsb_block[tsb_index].tsb = new_tsb;
409         setup_tsb_params(mm, tsb_index, new_size);
410
411         spin_unlock_irqrestore(&mm->context.lock, flags);
412
413         /* If old_tsb is NULL, we're being invoked for the first time
414          * from init_new_context().
415          */
416         if (old_tsb) {
417                 /* Reload it on the local cpu.  */
418                 tsb_context_switch(mm);
419
420                 /* Now force other processors to do the same.  */
421                 smp_tsb_sync(mm);
422
423                 /* Now it is safe to free the old tsb.  */
424                 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
425         }
426 }
427
428 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
429 {
430 #ifdef CONFIG_HUGETLB_PAGE
431         unsigned long huge_pte_count;
432 #endif
433         unsigned int i;
434
435         spin_lock_init(&mm->context.lock);
436
437         mm->context.sparc64_ctx_val = 0UL;
438
439 #ifdef CONFIG_HUGETLB_PAGE
440         /* We reset it to zero because the fork() page copying
441          * will re-increment the counters as the parent PTEs are
442          * copied into the child address space.
443          */
444         huge_pte_count = mm->context.huge_pte_count;
445         mm->context.huge_pte_count = 0;
446 #endif
447
448         /* copy_mm() copies over the parent's mm_struct before calling
449          * us, so we need to zero out the TSB pointer or else tsb_grow()
450          * will be confused and think there is an older TSB to free up.
451          */
452         for (i = 0; i < MM_NUM_TSBS; i++)
453                 mm->context.tsb_block[i].tsb = NULL;
454
455         /* If this is fork, inherit the parent's TSB size.  We would
456          * grow it to that size on the first page fault anyways.
457          */
458         tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
459
460 #ifdef CONFIG_HUGETLB_PAGE
461         if (unlikely(huge_pte_count))
462                 tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
463 #endif
464
465         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
466                 return -ENOMEM;
467
468         return 0;
469 }
470
471 static void tsb_destroy_one(struct tsb_config *tp)
472 {
473         unsigned long cache_index;
474
475         if (!tp->tsb)
476                 return;
477         cache_index = tp->tsb_reg_val & 0x7UL;
478         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
479         tp->tsb = NULL;
480         tp->tsb_reg_val = 0UL;
481 }
482
483 void destroy_context(struct mm_struct *mm)
484 {
485         unsigned long flags, i;
486
487         for (i = 0; i < MM_NUM_TSBS; i++)
488                 tsb_destroy_one(&mm->context.tsb_block[i]);
489
490         spin_lock_irqsave(&ctx_alloc_lock, flags);
491
492         if (CTX_VALID(mm->context)) {
493                 unsigned long nr = CTX_NRBITS(mm->context);
494                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
495         }
496
497         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
498 }