Merge branch 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/benh/powerpc
[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 struct kmem_cache *pgtable_cache __read_mostly;
240
241 static struct kmem_cache *tsb_caches[8] __read_mostly;
242
243 static const char *tsb_cache_names[8] = {
244         "tsb_8KB",
245         "tsb_16KB",
246         "tsb_32KB",
247         "tsb_64KB",
248         "tsb_128KB",
249         "tsb_256KB",
250         "tsb_512KB",
251         "tsb_1MB",
252 };
253
254 void __init pgtable_cache_init(void)
255 {
256         unsigned long i;
257
258         pgtable_cache = kmem_cache_create("pgtable_cache",
259                                           PAGE_SIZE, PAGE_SIZE,
260                                           0,
261                                           _clear_page);
262         if (!pgtable_cache) {
263                 prom_printf("pgtable_cache_init(): Could not create!\n");
264                 prom_halt();
265         }
266
267         for (i = 0; i < 8; i++) {
268                 unsigned long size = 8192 << i;
269                 const char *name = tsb_cache_names[i];
270
271                 tsb_caches[i] = kmem_cache_create(name,
272                                                   size, size,
273                                                   0, NULL);
274                 if (!tsb_caches[i]) {
275                         prom_printf("Could not create %s cache\n", name);
276                         prom_halt();
277                 }
278         }
279 }
280
281 int sysctl_tsb_ratio = -2;
282
283 static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
284 {
285         unsigned long num_ents = (new_size / sizeof(struct tsb));
286
287         if (sysctl_tsb_ratio < 0)
288                 return num_ents - (num_ents >> -sysctl_tsb_ratio);
289         else
290                 return num_ents + (num_ents >> sysctl_tsb_ratio);
291 }
292
293 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
294  * do_sparc64_fault() invokes this routine to try and grow it.
295  *
296  * When we reach the maximum TSB size supported, we stick ~0UL into
297  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
298  * will not trigger any longer.
299  *
300  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
301  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
302  * must be 512K aligned.  It also must be physically contiguous, so we
303  * cannot use vmalloc().
304  *
305  * The idea here is to grow the TSB when the RSS of the process approaches
306  * the number of entries that the current TSB can hold at once.  Currently,
307  * we trigger when the RSS hits 3/4 of the TSB capacity.
308  */
309 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
310 {
311         unsigned long max_tsb_size = 1 * 1024 * 1024;
312         unsigned long new_size, old_size, flags;
313         struct tsb *old_tsb, *new_tsb;
314         unsigned long new_cache_index, old_cache_index;
315         unsigned long new_rss_limit;
316         gfp_t gfp_flags;
317
318         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
319                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
320
321         new_cache_index = 0;
322         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
323                 new_rss_limit = tsb_size_to_rss_limit(new_size);
324                 if (new_rss_limit > rss)
325                         break;
326                 new_cache_index++;
327         }
328
329         if (new_size == max_tsb_size)
330                 new_rss_limit = ~0UL;
331
332 retry_tsb_alloc:
333         gfp_flags = GFP_KERNEL;
334         if (new_size > (PAGE_SIZE * 2))
335                 gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
336
337         new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
338                                         gfp_flags, numa_node_id());
339         if (unlikely(!new_tsb)) {
340                 /* Not being able to fork due to a high-order TSB
341                  * allocation failure is very bad behavior.  Just back
342                  * down to a 0-order allocation and force no TSB
343                  * growing for this address space.
344                  */
345                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
346                     new_cache_index > 0) {
347                         new_cache_index = 0;
348                         new_size = 8192;
349                         new_rss_limit = ~0UL;
350                         goto retry_tsb_alloc;
351                 }
352
353                 /* If we failed on a TSB grow, we are under serious
354                  * memory pressure so don't try to grow any more.
355                  */
356                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
357                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
358                 return;
359         }
360
361         /* Mark all tags as invalid.  */
362         tsb_init(new_tsb, new_size);
363
364         /* Ok, we are about to commit the changes.  If we are
365          * growing an existing TSB the locking is very tricky,
366          * so WATCH OUT!
367          *
368          * We have to hold mm->context.lock while committing to the
369          * new TSB, this synchronizes us with processors in
370          * flush_tsb_user() and switch_mm() for this address space.
371          *
372          * But even with that lock held, processors run asynchronously
373          * accessing the old TSB via TLB miss handling.  This is OK
374          * because those actions are just propagating state from the
375          * Linux page tables into the TSB, page table mappings are not
376          * being changed.  If a real fault occurs, the processor will
377          * synchronize with us when it hits flush_tsb_user(), this is
378          * also true for the case where vmscan is modifying the page
379          * tables.  The only thing we need to be careful with is to
380          * skip any locked TSB entries during copy_tsb().
381          *
382          * When we finish committing to the new TSB, we have to drop
383          * the lock and ask all other cpus running this address space
384          * to run tsb_context_switch() to see the new TSB table.
385          */
386         spin_lock_irqsave(&mm->context.lock, flags);
387
388         old_tsb = mm->context.tsb_block[tsb_index].tsb;
389         old_cache_index =
390                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
391         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
392                     sizeof(struct tsb));
393
394
395         /* Handle multiple threads trying to grow the TSB at the same time.
396          * One will get in here first, and bump the size and the RSS limit.
397          * The others will get in here next and hit this check.
398          */
399         if (unlikely(old_tsb &&
400                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
401                 spin_unlock_irqrestore(&mm->context.lock, flags);
402
403                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
404                 return;
405         }
406
407         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
408
409         if (old_tsb) {
410                 extern void copy_tsb(unsigned long old_tsb_base,
411                                      unsigned long old_tsb_size,
412                                      unsigned long new_tsb_base,
413                                      unsigned long new_tsb_size);
414                 unsigned long old_tsb_base = (unsigned long) old_tsb;
415                 unsigned long new_tsb_base = (unsigned long) new_tsb;
416
417                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
418                         old_tsb_base = __pa(old_tsb_base);
419                         new_tsb_base = __pa(new_tsb_base);
420                 }
421                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
422         }
423
424         mm->context.tsb_block[tsb_index].tsb = new_tsb;
425         setup_tsb_params(mm, tsb_index, new_size);
426
427         spin_unlock_irqrestore(&mm->context.lock, flags);
428
429         /* If old_tsb is NULL, we're being invoked for the first time
430          * from init_new_context().
431          */
432         if (old_tsb) {
433                 /* Reload it on the local cpu.  */
434                 tsb_context_switch(mm);
435
436                 /* Now force other processors to do the same.  */
437                 preempt_disable();
438                 smp_tsb_sync(mm);
439                 preempt_enable();
440
441                 /* Now it is safe to free the old tsb.  */
442                 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
443         }
444 }
445
446 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
447 {
448 #ifdef CONFIG_HUGETLB_PAGE
449         unsigned long huge_pte_count;
450 #endif
451         unsigned int i;
452
453         spin_lock_init(&mm->context.lock);
454
455         mm->context.sparc64_ctx_val = 0UL;
456
457 #ifdef CONFIG_HUGETLB_PAGE
458         /* We reset it to zero because the fork() page copying
459          * will re-increment the counters as the parent PTEs are
460          * copied into the child address space.
461          */
462         huge_pte_count = mm->context.huge_pte_count;
463         mm->context.huge_pte_count = 0;
464 #endif
465
466         /* copy_mm() copies over the parent's mm_struct before calling
467          * us, so we need to zero out the TSB pointer or else tsb_grow()
468          * will be confused and think there is an older TSB to free up.
469          */
470         for (i = 0; i < MM_NUM_TSBS; i++)
471                 mm->context.tsb_block[i].tsb = NULL;
472
473         /* If this is fork, inherit the parent's TSB size.  We would
474          * grow it to that size on the first page fault anyways.
475          */
476         tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
477
478 #ifdef CONFIG_HUGETLB_PAGE
479         if (unlikely(huge_pte_count))
480                 tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
481 #endif
482
483         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
484                 return -ENOMEM;
485
486         return 0;
487 }
488
489 static void tsb_destroy_one(struct tsb_config *tp)
490 {
491         unsigned long cache_index;
492
493         if (!tp->tsb)
494                 return;
495         cache_index = tp->tsb_reg_val & 0x7UL;
496         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
497         tp->tsb = NULL;
498         tp->tsb_reg_val = 0UL;
499 }
500
501 void destroy_context(struct mm_struct *mm)
502 {
503         unsigned long flags, i;
504
505         for (i = 0; i < MM_NUM_TSBS; i++)
506                 tsb_destroy_one(&mm->context.tsb_block[i]);
507
508         spin_lock_irqsave(&ctx_alloc_lock, flags);
509
510         if (CTX_VALID(mm->context)) {
511                 unsigned long nr = CTX_NRBITS(mm->context);
512                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
513         }
514
515         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
516 }