sparc,sparc64: unify mm/
[linux-2.6] / 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 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
269  * do_sparc64_fault() invokes this routine to try and grow it.
270  *
271  * When we reach the maximum TSB size supported, we stick ~0UL into
272  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
273  * will not trigger any longer.
274  *
275  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
276  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
277  * must be 512K aligned.  It also must be physically contiguous, so we
278  * cannot use vmalloc().
279  *
280  * The idea here is to grow the TSB when the RSS of the process approaches
281  * the number of entries that the current TSB can hold at once.  Currently,
282  * we trigger when the RSS hits 3/4 of the TSB capacity.
283  */
284 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
285 {
286         unsigned long max_tsb_size = 1 * 1024 * 1024;
287         unsigned long new_size, old_size, flags;
288         struct tsb *old_tsb, *new_tsb;
289         unsigned long new_cache_index, old_cache_index;
290         unsigned long new_rss_limit;
291         gfp_t gfp_flags;
292
293         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
294                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
295
296         new_cache_index = 0;
297         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
298                 unsigned long n_entries = new_size / sizeof(struct tsb);
299
300                 n_entries = (n_entries * 3) / 4;
301                 if (n_entries > rss)
302                         break;
303
304                 new_cache_index++;
305         }
306
307         if (new_size == max_tsb_size)
308                 new_rss_limit = ~0UL;
309         else
310                 new_rss_limit = ((new_size / sizeof(struct tsb)) * 3) / 4;
311
312 retry_tsb_alloc:
313         gfp_flags = GFP_KERNEL;
314         if (new_size > (PAGE_SIZE * 2))
315                 gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
316
317         new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
318                                         gfp_flags, numa_node_id());
319         if (unlikely(!new_tsb)) {
320                 /* Not being able to fork due to a high-order TSB
321                  * allocation failure is very bad behavior.  Just back
322                  * down to a 0-order allocation and force no TSB
323                  * growing for this address space.
324                  */
325                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
326                     new_cache_index > 0) {
327                         new_cache_index = 0;
328                         new_size = 8192;
329                         new_rss_limit = ~0UL;
330                         goto retry_tsb_alloc;
331                 }
332
333                 /* If we failed on a TSB grow, we are under serious
334                  * memory pressure so don't try to grow any more.
335                  */
336                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
337                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
338                 return;
339         }
340
341         /* Mark all tags as invalid.  */
342         tsb_init(new_tsb, new_size);
343
344         /* Ok, we are about to commit the changes.  If we are
345          * growing an existing TSB the locking is very tricky,
346          * so WATCH OUT!
347          *
348          * We have to hold mm->context.lock while committing to the
349          * new TSB, this synchronizes us with processors in
350          * flush_tsb_user() and switch_mm() for this address space.
351          *
352          * But even with that lock held, processors run asynchronously
353          * accessing the old TSB via TLB miss handling.  This is OK
354          * because those actions are just propagating state from the
355          * Linux page tables into the TSB, page table mappings are not
356          * being changed.  If a real fault occurs, the processor will
357          * synchronize with us when it hits flush_tsb_user(), this is
358          * also true for the case where vmscan is modifying the page
359          * tables.  The only thing we need to be careful with is to
360          * skip any locked TSB entries during copy_tsb().
361          *
362          * When we finish committing to the new TSB, we have to drop
363          * the lock and ask all other cpus running this address space
364          * to run tsb_context_switch() to see the new TSB table.
365          */
366         spin_lock_irqsave(&mm->context.lock, flags);
367
368         old_tsb = mm->context.tsb_block[tsb_index].tsb;
369         old_cache_index =
370                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
371         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
372                     sizeof(struct tsb));
373
374
375         /* Handle multiple threads trying to grow the TSB at the same time.
376          * One will get in here first, and bump the size and the RSS limit.
377          * The others will get in here next and hit this check.
378          */
379         if (unlikely(old_tsb &&
380                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
381                 spin_unlock_irqrestore(&mm->context.lock, flags);
382
383                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
384                 return;
385         }
386
387         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
388
389         if (old_tsb) {
390                 extern void copy_tsb(unsigned long old_tsb_base,
391                                      unsigned long old_tsb_size,
392                                      unsigned long new_tsb_base,
393                                      unsigned long new_tsb_size);
394                 unsigned long old_tsb_base = (unsigned long) old_tsb;
395                 unsigned long new_tsb_base = (unsigned long) new_tsb;
396
397                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
398                         old_tsb_base = __pa(old_tsb_base);
399                         new_tsb_base = __pa(new_tsb_base);
400                 }
401                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
402         }
403
404         mm->context.tsb_block[tsb_index].tsb = new_tsb;
405         setup_tsb_params(mm, tsb_index, new_size);
406
407         spin_unlock_irqrestore(&mm->context.lock, flags);
408
409         /* If old_tsb is NULL, we're being invoked for the first time
410          * from init_new_context().
411          */
412         if (old_tsb) {
413                 /* Reload it on the local cpu.  */
414                 tsb_context_switch(mm);
415
416                 /* Now force other processors to do the same.  */
417                 preempt_disable();
418                 smp_tsb_sync(mm);
419                 preempt_enable();
420
421                 /* Now it is safe to free the old tsb.  */
422                 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
423         }
424 }
425
426 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
427 {
428 #ifdef CONFIG_HUGETLB_PAGE
429         unsigned long huge_pte_count;
430 #endif
431         unsigned int i;
432
433         spin_lock_init(&mm->context.lock);
434
435         mm->context.sparc64_ctx_val = 0UL;
436
437 #ifdef CONFIG_HUGETLB_PAGE
438         /* We reset it to zero because the fork() page copying
439          * will re-increment the counters as the parent PTEs are
440          * copied into the child address space.
441          */
442         huge_pte_count = mm->context.huge_pte_count;
443         mm->context.huge_pte_count = 0;
444 #endif
445
446         /* copy_mm() copies over the parent's mm_struct before calling
447          * us, so we need to zero out the TSB pointer or else tsb_grow()
448          * will be confused and think there is an older TSB to free up.
449          */
450         for (i = 0; i < MM_NUM_TSBS; i++)
451                 mm->context.tsb_block[i].tsb = NULL;
452
453         /* If this is fork, inherit the parent's TSB size.  We would
454          * grow it to that size on the first page fault anyways.
455          */
456         tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
457
458 #ifdef CONFIG_HUGETLB_PAGE
459         if (unlikely(huge_pte_count))
460                 tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
461 #endif
462
463         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
464                 return -ENOMEM;
465
466         return 0;
467 }
468
469 static void tsb_destroy_one(struct tsb_config *tp)
470 {
471         unsigned long cache_index;
472
473         if (!tp->tsb)
474                 return;
475         cache_index = tp->tsb_reg_val & 0x7UL;
476         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
477         tp->tsb = NULL;
478         tp->tsb_reg_val = 0UL;
479 }
480
481 void destroy_context(struct mm_struct *mm)
482 {
483         unsigned long flags, i;
484
485         for (i = 0; i < MM_NUM_TSBS; i++)
486                 tsb_destroy_one(&mm->context.tsb_block[i]);
487
488         spin_lock_irqsave(&ctx_alloc_lock, flags);
489
490         if (CTX_VALID(mm->context)) {
491                 unsigned long nr = CTX_NRBITS(mm->context);
492                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
493         }
494
495         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
496 }