94846151349991387435c0a61a925b93b09debb2
[linux-2.6.git] / 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/system.h>
10 #include <asm/page.h>
11 #include <asm/tlbflush.h>
12 #include <asm/tlb.h>
13 #include <asm/mmu_context.h>
14 #include <asm/pgtable.h>
15 #include <asm/tsb.h>
16 #include <asm/oplib.h>
17
18 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
19
20 static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long hash_shift, unsigned long nentries)
21 {
22         vaddr >>= hash_shift;
23         return vaddr & (nentries - 1);
24 }
25
26 static inline int tag_compare(unsigned long tag, unsigned long vaddr)
27 {
28         return (tag == (vaddr >> 22));
29 }
30
31 /* TSB flushes need only occur on the processor initiating the address
32  * space modification, not on each cpu the address space has run on.
33  * Only the TLB flush needs that treatment.
34  */
35
36 void flush_tsb_kernel_range(unsigned long start, unsigned long end)
37 {
38         unsigned long v;
39
40         for (v = start; v < end; v += PAGE_SIZE) {
41                 unsigned long hash = tsb_hash(v, PAGE_SHIFT,
42                                               KERNEL_TSB_NENTRIES);
43                 struct tsb *ent = &swapper_tsb[hash];
44
45                 if (tag_compare(ent->tag, v))
46                         ent->tag = (1UL << TSB_TAG_INVALID_BIT);
47         }
48 }
49
50 static void __flush_tsb_one(struct tlb_batch *tb, unsigned long hash_shift,
51                             unsigned long tsb, unsigned long nentries)
52 {
53         unsigned long i;
54
55         for (i = 0; i < tb->tlb_nr; i++) {
56                 unsigned long v = tb->vaddrs[i];
57                 unsigned long tag, ent, hash;
58
59                 v &= ~0x1UL;
60
61                 hash = tsb_hash(v, hash_shift, nentries);
62                 ent = tsb + (hash * sizeof(struct tsb));
63                 tag = (v >> 22UL);
64
65                 tsb_flush(ent, tag);
66         }
67 }
68
69 void flush_tsb_user(struct tlb_batch *tb)
70 {
71         struct mm_struct *mm = tb->mm;
72         unsigned long nentries, base, flags;
73
74         spin_lock_irqsave(&mm->context.lock, flags);
75
76         base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
77         nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
78         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
79                 base = __pa(base);
80         __flush_tsb_one(tb, PAGE_SHIFT, base, nentries);
81
82 #ifdef CONFIG_HUGETLB_PAGE
83         if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
84                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
85                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
86                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
87                         base = __pa(base);
88                 __flush_tsb_one(tb, HPAGE_SHIFT, base, nentries);
89         }
90 #endif
91         spin_unlock_irqrestore(&mm->context.lock, flags);
92 }
93
94 #if defined(CONFIG_SPARC64_PAGE_SIZE_8KB)
95 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_8K
96 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_8K
97 #elif defined(CONFIG_SPARC64_PAGE_SIZE_64KB)
98 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_64K
99 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_64K
100 #else
101 #error Broken base page size setting...
102 #endif
103
104 #ifdef CONFIG_HUGETLB_PAGE
105 #if defined(CONFIG_HUGETLB_PAGE_SIZE_64K)
106 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_64K
107 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_64K
108 #elif defined(CONFIG_HUGETLB_PAGE_SIZE_512K)
109 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_512K
110 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_512K
111 #elif defined(CONFIG_HUGETLB_PAGE_SIZE_4MB)
112 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_4MB
113 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_4MB
114 #else
115 #error Broken huge page size setting...
116 #endif
117 #endif
118
119 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
120 {
121         unsigned long tsb_reg, base, tsb_paddr;
122         unsigned long page_sz, tte;
123
124         mm->context.tsb_block[tsb_idx].tsb_nentries =
125                 tsb_bytes / sizeof(struct tsb);
126
127         base = TSBMAP_BASE;
128         tte = pgprot_val(PAGE_KERNEL_LOCKED);
129         tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
130         BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
131
132         /* Use the smallest page size that can map the whole TSB
133          * in one TLB entry.
134          */
135         switch (tsb_bytes) {
136         case 8192 << 0:
137                 tsb_reg = 0x0UL;
138 #ifdef DCACHE_ALIASING_POSSIBLE
139                 base += (tsb_paddr & 8192);
140 #endif
141                 page_sz = 8192;
142                 break;
143
144         case 8192 << 1:
145                 tsb_reg = 0x1UL;
146                 page_sz = 64 * 1024;
147                 break;
148
149         case 8192 << 2:
150                 tsb_reg = 0x2UL;
151                 page_sz = 64 * 1024;
152                 break;
153
154         case 8192 << 3:
155                 tsb_reg = 0x3UL;
156                 page_sz = 64 * 1024;
157                 break;
158
159         case 8192 << 4:
160                 tsb_reg = 0x4UL;
161                 page_sz = 512 * 1024;
162                 break;
163
164         case 8192 << 5:
165                 tsb_reg = 0x5UL;
166                 page_sz = 512 * 1024;
167                 break;
168
169         case 8192 << 6:
170                 tsb_reg = 0x6UL;
171                 page_sz = 512 * 1024;
172                 break;
173
174         case 8192 << 7:
175                 tsb_reg = 0x7UL;
176                 page_sz = 4 * 1024 * 1024;
177                 break;
178
179         default:
180                 printk(KERN_ERR "TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
181                        current->comm, current->pid, tsb_bytes);
182                 do_exit(SIGSEGV);
183         };
184         tte |= pte_sz_bits(page_sz);
185
186         if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
187                 /* Physical mapping, no locked TLB entry for TSB.  */
188                 tsb_reg |= tsb_paddr;
189
190                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
191                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
192                 mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
193         } else {
194                 tsb_reg |= base;
195                 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
196                 tte |= (tsb_paddr & ~(page_sz - 1UL));
197
198                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
199                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
200                 mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
201         }
202
203         /* Setup the Hypervisor TSB descriptor.  */
204         if (tlb_type == hypervisor) {
205                 struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
206
207                 switch (tsb_idx) {
208                 case MM_TSB_BASE:
209                         hp->pgsz_idx = HV_PGSZ_IDX_BASE;
210                         break;
211 #ifdef CONFIG_HUGETLB_PAGE
212                 case MM_TSB_HUGE:
213                         hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
214                         break;
215 #endif
216                 default:
217                         BUG();
218                 };
219                 hp->assoc = 1;
220                 hp->num_ttes = tsb_bytes / 16;
221                 hp->ctx_idx = 0;
222                 switch (tsb_idx) {
223                 case MM_TSB_BASE:
224                         hp->pgsz_mask = HV_PGSZ_MASK_BASE;
225                         break;
226 #ifdef CONFIG_HUGETLB_PAGE
227                 case MM_TSB_HUGE:
228                         hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
229                         break;
230 #endif
231                 default:
232                         BUG();
233                 };
234                 hp->tsb_base = tsb_paddr;
235                 hp->resv = 0;
236         }
237 }
238
239 static struct kmem_cache *tsb_caches[8] __read_mostly;
240
241 static const char *tsb_cache_names[8] = {
242         "tsb_8KB",
243         "tsb_16KB",
244         "tsb_32KB",
245         "tsb_64KB",
246         "tsb_128KB",
247         "tsb_256KB",
248         "tsb_512KB",
249         "tsb_1MB",
250 };
251
252 void __init pgtable_cache_init(void)
253 {
254         unsigned long i;
255
256         for (i = 0; i < 8; i++) {
257                 unsigned long size = 8192 << i;
258                 const char *name = tsb_cache_names[i];
259
260                 tsb_caches[i] = kmem_cache_create(name,
261                                                   size, size,
262                                                   0, NULL);
263                 if (!tsb_caches[i]) {
264                         prom_printf("Could not create %s cache\n", name);
265                         prom_halt();
266                 }
267         }
268 }
269
270 int sysctl_tsb_ratio = -2;
271
272 static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
273 {
274         unsigned long num_ents = (new_size / sizeof(struct tsb));
275
276         if (sysctl_tsb_ratio < 0)
277                 return num_ents - (num_ents >> -sysctl_tsb_ratio);
278         else
279                 return num_ents + (num_ents >> sysctl_tsb_ratio);
280 }
281
282 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
283  * do_sparc64_fault() invokes this routine to try and grow it.
284  *
285  * When we reach the maximum TSB size supported, we stick ~0UL into
286  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
287  * will not trigger any longer.
288  *
289  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
290  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
291  * must be 512K aligned.  It also must be physically contiguous, so we
292  * cannot use vmalloc().
293  *
294  * The idea here is to grow the TSB when the RSS of the process approaches
295  * the number of entries that the current TSB can hold at once.  Currently,
296  * we trigger when the RSS hits 3/4 of the TSB capacity.
297  */
298 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
299 {
300         unsigned long max_tsb_size = 1 * 1024 * 1024;
301         unsigned long new_size, old_size, flags;
302         struct tsb *old_tsb, *new_tsb;
303         unsigned long new_cache_index, old_cache_index;
304         unsigned long new_rss_limit;
305         gfp_t gfp_flags;
306
307         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
308                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
309
310         new_cache_index = 0;
311         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
312                 new_rss_limit = tsb_size_to_rss_limit(new_size);
313                 if (new_rss_limit > rss)
314                         break;
315                 new_cache_index++;
316         }
317
318         if (new_size == max_tsb_size)
319                 new_rss_limit = ~0UL;
320
321 retry_tsb_alloc:
322         gfp_flags = GFP_KERNEL;
323         if (new_size > (PAGE_SIZE * 2))
324                 gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
325
326         new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
327                                         gfp_flags, numa_node_id());
328         if (unlikely(!new_tsb)) {
329                 /* Not being able to fork due to a high-order TSB
330                  * allocation failure is very bad behavior.  Just back
331                  * down to a 0-order allocation and force no TSB
332                  * growing for this address space.
333                  */
334                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
335                     new_cache_index > 0) {
336                         new_cache_index = 0;
337                         new_size = 8192;
338                         new_rss_limit = ~0UL;
339                         goto retry_tsb_alloc;
340                 }
341
342                 /* If we failed on a TSB grow, we are under serious
343                  * memory pressure so don't try to grow any more.
344                  */
345                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
346                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
347                 return;
348         }
349
350         /* Mark all tags as invalid.  */
351         tsb_init(new_tsb, new_size);
352
353         /* Ok, we are about to commit the changes.  If we are
354          * growing an existing TSB the locking is very tricky,
355          * so WATCH OUT!
356          *
357          * We have to hold mm->context.lock while committing to the
358          * new TSB, this synchronizes us with processors in
359          * flush_tsb_user() and switch_mm() for this address space.
360          *
361          * But even with that lock held, processors run asynchronously
362          * accessing the old TSB via TLB miss handling.  This is OK
363          * because those actions are just propagating state from the
364          * Linux page tables into the TSB, page table mappings are not
365          * being changed.  If a real fault occurs, the processor will
366          * synchronize with us when it hits flush_tsb_user(), this is
367          * also true for the case where vmscan is modifying the page
368          * tables.  The only thing we need to be careful with is to
369          * skip any locked TSB entries during copy_tsb().
370          *
371          * When we finish committing to the new TSB, we have to drop
372          * the lock and ask all other cpus running this address space
373          * to run tsb_context_switch() to see the new TSB table.
374          */
375         spin_lock_irqsave(&mm->context.lock, flags);
376
377         old_tsb = mm->context.tsb_block[tsb_index].tsb;
378         old_cache_index =
379                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
380         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
381                     sizeof(struct tsb));
382
383
384         /* Handle multiple threads trying to grow the TSB at the same time.
385          * One will get in here first, and bump the size and the RSS limit.
386          * The others will get in here next and hit this check.
387          */
388         if (unlikely(old_tsb &&
389                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
390                 spin_unlock_irqrestore(&mm->context.lock, flags);
391
392                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
393                 return;
394         }
395
396         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
397
398         if (old_tsb) {
399                 extern void copy_tsb(unsigned long old_tsb_base,
400                                      unsigned long old_tsb_size,
401                                      unsigned long new_tsb_base,
402                                      unsigned long new_tsb_size);
403                 unsigned long old_tsb_base = (unsigned long) old_tsb;
404                 unsigned long new_tsb_base = (unsigned long) new_tsb;
405
406                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
407                         old_tsb_base = __pa(old_tsb_base);
408                         new_tsb_base = __pa(new_tsb_base);
409                 }
410                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
411         }
412
413         mm->context.tsb_block[tsb_index].tsb = new_tsb;
414         setup_tsb_params(mm, tsb_index, new_size);
415
416         spin_unlock_irqrestore(&mm->context.lock, flags);
417
418         /* If old_tsb is NULL, we're being invoked for the first time
419          * from init_new_context().
420          */
421         if (old_tsb) {
422                 /* Reload it on the local cpu.  */
423                 tsb_context_switch(mm);
424
425                 /* Now force other processors to do the same.  */
426                 preempt_disable();
427                 smp_tsb_sync(mm);
428                 preempt_enable();
429
430                 /* Now it is safe to free the old tsb.  */
431                 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
432         }
433 }
434
435 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
436 {
437 #ifdef CONFIG_HUGETLB_PAGE
438         unsigned long huge_pte_count;
439 #endif
440         unsigned int i;
441
442         spin_lock_init(&mm->context.lock);
443
444         mm->context.sparc64_ctx_val = 0UL;
445
446 #ifdef CONFIG_HUGETLB_PAGE
447         /* We reset it to zero because the fork() page copying
448          * will re-increment the counters as the parent PTEs are
449          * copied into the child address space.
450          */
451         huge_pte_count = mm->context.huge_pte_count;
452         mm->context.huge_pte_count = 0;
453 #endif
454
455         /* copy_mm() copies over the parent's mm_struct before calling
456          * us, so we need to zero out the TSB pointer or else tsb_grow()
457          * will be confused and think there is an older TSB to free up.
458          */
459         for (i = 0; i < MM_NUM_TSBS; i++)
460                 mm->context.tsb_block[i].tsb = NULL;
461
462         /* If this is fork, inherit the parent's TSB size.  We would
463          * grow it to that size on the first page fault anyways.
464          */
465         tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
466
467 #ifdef CONFIG_HUGETLB_PAGE
468         if (unlikely(huge_pte_count))
469                 tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
470 #endif
471
472         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
473                 return -ENOMEM;
474
475         return 0;
476 }
477
478 static void tsb_destroy_one(struct tsb_config *tp)
479 {
480         unsigned long cache_index;
481
482         if (!tp->tsb)
483                 return;
484         cache_index = tp->tsb_reg_val & 0x7UL;
485         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
486         tp->tsb = NULL;
487         tp->tsb_reg_val = 0UL;
488 }
489
490 void destroy_context(struct mm_struct *mm)
491 {
492         unsigned long flags, i;
493
494         for (i = 0; i < MM_NUM_TSBS; i++)
495                 tsb_destroy_one(&mm->context.tsb_block[i]);
496
497         spin_lock_irqsave(&ctx_alloc_lock, flags);
498
499         if (CTX_VALID(mm->context)) {
500                 unsigned long nr = CTX_NRBITS(mm->context);
501                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
502         }
503
504         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
505 }