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