sparc64: Eliminate PTE table memory wastage.
[~shefty/rdma-dev.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/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 tlb_batch *tb, unsigned long hash_shift,
50                             unsigned long tsb, unsigned long nentries)
51 {
52         unsigned long i;
53
54         for (i = 0; i < tb->tlb_nr; i++) {
55                 unsigned long v = tb->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 tlb_batch *tb)
69 {
70         struct mm_struct *mm = tb->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(tb, 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(tb, HPAGE_SHIFT, base, nentries);
88         }
89 #endif
90         spin_unlock_irqrestore(&mm->context.lock, flags);
91 }
92
93 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_8K
94 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_8K
95
96 #ifdef CONFIG_HUGETLB_PAGE
97 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_4MB
98 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_4MB
99 #endif
100
101 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
102 {
103         unsigned long tsb_reg, base, tsb_paddr;
104         unsigned long page_sz, tte;
105
106         mm->context.tsb_block[tsb_idx].tsb_nentries =
107                 tsb_bytes / sizeof(struct tsb);
108
109         base = TSBMAP_BASE;
110         tte = pgprot_val(PAGE_KERNEL_LOCKED);
111         tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
112         BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
113
114         /* Use the smallest page size that can map the whole TSB
115          * in one TLB entry.
116          */
117         switch (tsb_bytes) {
118         case 8192 << 0:
119                 tsb_reg = 0x0UL;
120 #ifdef DCACHE_ALIASING_POSSIBLE
121                 base += (tsb_paddr & 8192);
122 #endif
123                 page_sz = 8192;
124                 break;
125
126         case 8192 << 1:
127                 tsb_reg = 0x1UL;
128                 page_sz = 64 * 1024;
129                 break;
130
131         case 8192 << 2:
132                 tsb_reg = 0x2UL;
133                 page_sz = 64 * 1024;
134                 break;
135
136         case 8192 << 3:
137                 tsb_reg = 0x3UL;
138                 page_sz = 64 * 1024;
139                 break;
140
141         case 8192 << 4:
142                 tsb_reg = 0x4UL;
143                 page_sz = 512 * 1024;
144                 break;
145
146         case 8192 << 5:
147                 tsb_reg = 0x5UL;
148                 page_sz = 512 * 1024;
149                 break;
150
151         case 8192 << 6:
152                 tsb_reg = 0x6UL;
153                 page_sz = 512 * 1024;
154                 break;
155
156         case 8192 << 7:
157                 tsb_reg = 0x7UL;
158                 page_sz = 4 * 1024 * 1024;
159                 break;
160
161         default:
162                 printk(KERN_ERR "TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
163                        current->comm, current->pid, tsb_bytes);
164                 do_exit(SIGSEGV);
165         }
166         tte |= pte_sz_bits(page_sz);
167
168         if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
169                 /* Physical mapping, no locked TLB entry for TSB.  */
170                 tsb_reg |= tsb_paddr;
171
172                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
173                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
174                 mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
175         } else {
176                 tsb_reg |= base;
177                 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
178                 tte |= (tsb_paddr & ~(page_sz - 1UL));
179
180                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
181                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
182                 mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
183         }
184
185         /* Setup the Hypervisor TSB descriptor.  */
186         if (tlb_type == hypervisor) {
187                 struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
188
189                 switch (tsb_idx) {
190                 case MM_TSB_BASE:
191                         hp->pgsz_idx = HV_PGSZ_IDX_BASE;
192                         break;
193 #ifdef CONFIG_HUGETLB_PAGE
194                 case MM_TSB_HUGE:
195                         hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
196                         break;
197 #endif
198                 default:
199                         BUG();
200                 }
201                 hp->assoc = 1;
202                 hp->num_ttes = tsb_bytes / 16;
203                 hp->ctx_idx = 0;
204                 switch (tsb_idx) {
205                 case MM_TSB_BASE:
206                         hp->pgsz_mask = HV_PGSZ_MASK_BASE;
207                         break;
208 #ifdef CONFIG_HUGETLB_PAGE
209                 case MM_TSB_HUGE:
210                         hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
211                         break;
212 #endif
213                 default:
214                         BUG();
215                 }
216                 hp->tsb_base = tsb_paddr;
217                 hp->resv = 0;
218         }
219 }
220
221 struct kmem_cache *pgtable_cache __read_mostly;
222
223 static struct kmem_cache *tsb_caches[8] __read_mostly;
224
225 static const char *tsb_cache_names[8] = {
226         "tsb_8KB",
227         "tsb_16KB",
228         "tsb_32KB",
229         "tsb_64KB",
230         "tsb_128KB",
231         "tsb_256KB",
232         "tsb_512KB",
233         "tsb_1MB",
234 };
235
236 void __init pgtable_cache_init(void)
237 {
238         unsigned long i;
239
240         pgtable_cache = kmem_cache_create("pgtable_cache",
241                                           PAGE_SIZE, PAGE_SIZE,
242                                           0,
243                                           _clear_page);
244         if (!pgtable_cache) {
245                 prom_printf("pgtable_cache_init(): Could not create!\n");
246                 prom_halt();
247         }
248
249         for (i = 0; i < 8; i++) {
250                 unsigned long size = 8192 << i;
251                 const char *name = tsb_cache_names[i];
252
253                 tsb_caches[i] = kmem_cache_create(name,
254                                                   size, size,
255                                                   0, NULL);
256                 if (!tsb_caches[i]) {
257                         prom_printf("Could not create %s cache\n", name);
258                         prom_halt();
259                 }
260         }
261 }
262
263 int sysctl_tsb_ratio = -2;
264
265 static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
266 {
267         unsigned long num_ents = (new_size / sizeof(struct tsb));
268
269         if (sysctl_tsb_ratio < 0)
270                 return num_ents - (num_ents >> -sysctl_tsb_ratio);
271         else
272                 return num_ents + (num_ents >> sysctl_tsb_ratio);
273 }
274
275 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
276  * do_sparc64_fault() invokes this routine to try and grow it.
277  *
278  * When we reach the maximum TSB size supported, we stick ~0UL into
279  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
280  * will not trigger any longer.
281  *
282  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
283  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
284  * must be 512K aligned.  It also must be physically contiguous, so we
285  * cannot use vmalloc().
286  *
287  * The idea here is to grow the TSB when the RSS of the process approaches
288  * the number of entries that the current TSB can hold at once.  Currently,
289  * we trigger when the RSS hits 3/4 of the TSB capacity.
290  */
291 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
292 {
293         unsigned long max_tsb_size = 1 * 1024 * 1024;
294         unsigned long new_size, old_size, flags;
295         struct tsb *old_tsb, *new_tsb;
296         unsigned long new_cache_index, old_cache_index;
297         unsigned long new_rss_limit;
298         gfp_t gfp_flags;
299
300         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
301                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
302
303         new_cache_index = 0;
304         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
305                 new_rss_limit = tsb_size_to_rss_limit(new_size);
306                 if (new_rss_limit > rss)
307                         break;
308                 new_cache_index++;
309         }
310
311         if (new_size == max_tsb_size)
312                 new_rss_limit = ~0UL;
313
314 retry_tsb_alloc:
315         gfp_flags = GFP_KERNEL;
316         if (new_size > (PAGE_SIZE * 2))
317                 gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
318
319         new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
320                                         gfp_flags, numa_node_id());
321         if (unlikely(!new_tsb)) {
322                 /* Not being able to fork due to a high-order TSB
323                  * allocation failure is very bad behavior.  Just back
324                  * down to a 0-order allocation and force no TSB
325                  * growing for this address space.
326                  */
327                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
328                     new_cache_index > 0) {
329                         new_cache_index = 0;
330                         new_size = 8192;
331                         new_rss_limit = ~0UL;
332                         goto retry_tsb_alloc;
333                 }
334
335                 /* If we failed on a TSB grow, we are under serious
336                  * memory pressure so don't try to grow any more.
337                  */
338                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
339                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
340                 return;
341         }
342
343         /* Mark all tags as invalid.  */
344         tsb_init(new_tsb, new_size);
345
346         /* Ok, we are about to commit the changes.  If we are
347          * growing an existing TSB the locking is very tricky,
348          * so WATCH OUT!
349          *
350          * We have to hold mm->context.lock while committing to the
351          * new TSB, this synchronizes us with processors in
352          * flush_tsb_user() and switch_mm() for this address space.
353          *
354          * But even with that lock held, processors run asynchronously
355          * accessing the old TSB via TLB miss handling.  This is OK
356          * because those actions are just propagating state from the
357          * Linux page tables into the TSB, page table mappings are not
358          * being changed.  If a real fault occurs, the processor will
359          * synchronize with us when it hits flush_tsb_user(), this is
360          * also true for the case where vmscan is modifying the page
361          * tables.  The only thing we need to be careful with is to
362          * skip any locked TSB entries during copy_tsb().
363          *
364          * When we finish committing to the new TSB, we have to drop
365          * the lock and ask all other cpus running this address space
366          * to run tsb_context_switch() to see the new TSB table.
367          */
368         spin_lock_irqsave(&mm->context.lock, flags);
369
370         old_tsb = mm->context.tsb_block[tsb_index].tsb;
371         old_cache_index =
372                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
373         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
374                     sizeof(struct tsb));
375
376
377         /* Handle multiple threads trying to grow the TSB at the same time.
378          * One will get in here first, and bump the size and the RSS limit.
379          * The others will get in here next and hit this check.
380          */
381         if (unlikely(old_tsb &&
382                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
383                 spin_unlock_irqrestore(&mm->context.lock, flags);
384
385                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
386                 return;
387         }
388
389         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
390
391         if (old_tsb) {
392                 extern void copy_tsb(unsigned long old_tsb_base,
393                                      unsigned long old_tsb_size,
394                                      unsigned long new_tsb_base,
395                                      unsigned long new_tsb_size);
396                 unsigned long old_tsb_base = (unsigned long) old_tsb;
397                 unsigned long new_tsb_base = (unsigned long) new_tsb;
398
399                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
400                         old_tsb_base = __pa(old_tsb_base);
401                         new_tsb_base = __pa(new_tsb_base);
402                 }
403                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
404         }
405
406         mm->context.tsb_block[tsb_index].tsb = new_tsb;
407         setup_tsb_params(mm, tsb_index, new_size);
408
409         spin_unlock_irqrestore(&mm->context.lock, flags);
410
411         /* If old_tsb is NULL, we're being invoked for the first time
412          * from init_new_context().
413          */
414         if (old_tsb) {
415                 /* Reload it on the local cpu.  */
416                 tsb_context_switch(mm);
417
418                 /* Now force other processors to do the same.  */
419                 preempt_disable();
420                 smp_tsb_sync(mm);
421                 preempt_enable();
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         mm->context.pgtable_page = NULL;
449
450         /* copy_mm() copies over the parent's mm_struct before calling
451          * us, so we need to zero out the TSB pointer or else tsb_grow()
452          * will be confused and think there is an older TSB to free up.
453          */
454         for (i = 0; i < MM_NUM_TSBS; i++)
455                 mm->context.tsb_block[i].tsb = NULL;
456
457         /* If this is fork, inherit the parent's TSB size.  We would
458          * grow it to that size on the first page fault anyways.
459          */
460         tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
461
462 #ifdef CONFIG_HUGETLB_PAGE
463         if (unlikely(huge_pte_count))
464                 tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
465 #endif
466
467         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
468                 return -ENOMEM;
469
470         return 0;
471 }
472
473 static void tsb_destroy_one(struct tsb_config *tp)
474 {
475         unsigned long cache_index;
476
477         if (!tp->tsb)
478                 return;
479         cache_index = tp->tsb_reg_val & 0x7UL;
480         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
481         tp->tsb = NULL;
482         tp->tsb_reg_val = 0UL;
483 }
484
485 void destroy_context(struct mm_struct *mm)
486 {
487         unsigned long flags, i;
488         struct page *page;
489
490         for (i = 0; i < MM_NUM_TSBS; i++)
491                 tsb_destroy_one(&mm->context.tsb_block[i]);
492
493         page = mm->context.pgtable_page;
494         if (page && put_page_testzero(page)) {
495                 pgtable_page_dtor(page);
496                 free_hot_cold_page(page, 0);
497         }
498
499         spin_lock_irqsave(&ctx_alloc_lock, flags);
500
501         if (CTX_VALID(mm->context)) {
502                 unsigned long nr = CTX_NRBITS(mm->context);
503                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
504         }
505
506         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
507 }