1 /* arch/sparc64/mm/tsb.c
3 * Copyright (C) 2006, 2008 David S. Miller <davem@davemloft.net>
6 #include <linux/kernel.h>
7 #include <linux/preempt.h>
8 #include <linux/slab.h>
10 #include <asm/pgtable.h>
11 #include <asm/mmu_context.h>
12 #include <asm/setup.h>
15 #include <asm/oplib.h>
17 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
19 static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long hash_shift, unsigned long nentries)
22 return vaddr & (nentries - 1);
25 static inline int tag_compare(unsigned long tag, unsigned long vaddr)
27 return (tag == (vaddr >> 22));
30 static void flush_tsb_kernel_range_scan(unsigned long start, unsigned long end)
34 for (idx = 0; idx < KERNEL_TSB_NENTRIES; idx++) {
35 struct tsb *ent = &swapper_tsb[idx];
36 unsigned long match = idx << 13;
38 match |= (ent->tag << 22);
39 if (match >= start && match < end)
40 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
44 /* TSB flushes need only occur on the processor initiating the address
45 * space modification, not on each cpu the address space has run on.
46 * Only the TLB flush needs that treatment.
49 void flush_tsb_kernel_range(unsigned long start, unsigned long end)
53 if ((end - start) >> PAGE_SHIFT >= 2 * KERNEL_TSB_NENTRIES)
54 return flush_tsb_kernel_range_scan(start, end);
56 for (v = start; v < end; v += PAGE_SIZE) {
57 unsigned long hash = tsb_hash(v, PAGE_SHIFT,
59 struct tsb *ent = &swapper_tsb[hash];
61 if (tag_compare(ent->tag, v))
62 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
66 static void __flush_tsb_one_entry(unsigned long tsb, unsigned long v,
67 unsigned long hash_shift,
68 unsigned long nentries)
70 unsigned long tag, ent, hash;
73 hash = tsb_hash(v, hash_shift, nentries);
74 ent = tsb + (hash * sizeof(struct tsb));
80 static void __flush_tsb_one(struct tlb_batch *tb, unsigned long hash_shift,
81 unsigned long tsb, unsigned long nentries)
85 for (i = 0; i < tb->tlb_nr; i++)
86 __flush_tsb_one_entry(tsb, tb->vaddrs[i], hash_shift, nentries);
89 void flush_tsb_user(struct tlb_batch *tb)
91 struct mm_struct *mm = tb->mm;
92 unsigned long nentries, base, flags;
94 spin_lock_irqsave(&mm->context.lock, flags);
97 base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
98 nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
99 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
101 __flush_tsb_one(tb, PAGE_SHIFT, base, nentries);
103 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
104 if (tb->huge && mm->context.tsb_block[MM_TSB_HUGE].tsb) {
105 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
106 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
107 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
109 __flush_tsb_one(tb, REAL_HPAGE_SHIFT, base, nentries);
112 spin_unlock_irqrestore(&mm->context.lock, flags);
115 void flush_tsb_user_page(struct mm_struct *mm, unsigned long vaddr, bool huge)
117 unsigned long nentries, base, flags;
119 spin_lock_irqsave(&mm->context.lock, flags);
122 base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
123 nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
124 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
126 __flush_tsb_one_entry(base, vaddr, PAGE_SHIFT, nentries);
128 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
129 if (huge && mm->context.tsb_block[MM_TSB_HUGE].tsb) {
130 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
131 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
132 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
134 __flush_tsb_one_entry(base, vaddr, REAL_HPAGE_SHIFT, nentries);
137 spin_unlock_irqrestore(&mm->context.lock, flags);
140 #define HV_PGSZ_IDX_BASE HV_PGSZ_IDX_8K
141 #define HV_PGSZ_MASK_BASE HV_PGSZ_MASK_8K
143 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
144 #define HV_PGSZ_IDX_HUGE HV_PGSZ_IDX_4MB
145 #define HV_PGSZ_MASK_HUGE HV_PGSZ_MASK_4MB
148 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
150 unsigned long tsb_reg, base, tsb_paddr;
151 unsigned long page_sz, tte;
153 mm->context.tsb_block[tsb_idx].tsb_nentries =
154 tsb_bytes / sizeof(struct tsb);
158 base = TSBMAP_8K_BASE;
160 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
162 base = TSBMAP_4M_BASE;
169 tte = pgprot_val(PAGE_KERNEL_LOCKED);
170 tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
171 BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
173 /* Use the smallest page size that can map the whole TSB
179 #ifdef DCACHE_ALIASING_POSSIBLE
180 base += (tsb_paddr & 8192);
202 page_sz = 512 * 1024;
207 page_sz = 512 * 1024;
212 page_sz = 512 * 1024;
217 page_sz = 4 * 1024 * 1024;
221 printk(KERN_ERR "TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
222 current->comm, current->pid, tsb_bytes);
225 tte |= pte_sz_bits(page_sz);
227 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
228 /* Physical mapping, no locked TLB entry for TSB. */
229 tsb_reg |= tsb_paddr;
231 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
232 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
233 mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
236 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
237 tte |= (tsb_paddr & ~(page_sz - 1UL));
239 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
240 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
241 mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
244 /* Setup the Hypervisor TSB descriptor. */
245 if (tlb_type == hypervisor) {
246 struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
250 hp->pgsz_idx = HV_PGSZ_IDX_BASE;
252 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
254 hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
261 hp->num_ttes = tsb_bytes / 16;
265 hp->pgsz_mask = HV_PGSZ_MASK_BASE;
267 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
269 hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
275 hp->tsb_base = tsb_paddr;
280 struct kmem_cache *pgtable_cache __read_mostly;
282 static struct kmem_cache *tsb_caches[8] __read_mostly;
284 static const char *tsb_cache_names[8] = {
295 void __init pgtable_cache_init(void)
299 pgtable_cache = kmem_cache_create("pgtable_cache",
300 PAGE_SIZE, PAGE_SIZE,
303 if (!pgtable_cache) {
304 prom_printf("pgtable_cache_init(): Could not create!\n");
308 for (i = 0; i < ARRAY_SIZE(tsb_cache_names); i++) {
309 unsigned long size = 8192 << i;
310 const char *name = tsb_cache_names[i];
312 tsb_caches[i] = kmem_cache_create(name,
315 if (!tsb_caches[i]) {
316 prom_printf("Could not create %s cache\n", name);
322 int sysctl_tsb_ratio = -2;
324 static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
326 unsigned long num_ents = (new_size / sizeof(struct tsb));
328 if (sysctl_tsb_ratio < 0)
329 return num_ents - (num_ents >> -sysctl_tsb_ratio);
331 return num_ents + (num_ents >> sysctl_tsb_ratio);
334 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
335 * do_sparc64_fault() invokes this routine to try and grow it.
337 * When we reach the maximum TSB size supported, we stick ~0UL into
338 * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
339 * will not trigger any longer.
341 * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
342 * of two. The TSB must be aligned to it's size, so f.e. a 512K TSB
343 * must be 512K aligned. It also must be physically contiguous, so we
344 * cannot use vmalloc().
346 * The idea here is to grow the TSB when the RSS of the process approaches
347 * the number of entries that the current TSB can hold at once. Currently,
348 * we trigger when the RSS hits 3/4 of the TSB capacity.
350 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
352 unsigned long max_tsb_size = 1 * 1024 * 1024;
353 unsigned long new_size, old_size, flags;
354 struct tsb *old_tsb, *new_tsb;
355 unsigned long new_cache_index, old_cache_index;
356 unsigned long new_rss_limit;
359 if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
360 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
363 for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
364 new_rss_limit = tsb_size_to_rss_limit(new_size);
365 if (new_rss_limit > rss)
370 if (new_size == max_tsb_size)
371 new_rss_limit = ~0UL;
374 gfp_flags = GFP_KERNEL;
375 if (new_size > (PAGE_SIZE * 2))
376 gfp_flags |= __GFP_NOWARN | __GFP_NORETRY;
378 new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
379 gfp_flags, numa_node_id());
380 if (unlikely(!new_tsb)) {
381 /* Not being able to fork due to a high-order TSB
382 * allocation failure is very bad behavior. Just back
383 * down to a 0-order allocation and force no TSB
384 * growing for this address space.
386 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
387 new_cache_index > 0) {
390 new_rss_limit = ~0UL;
391 goto retry_tsb_alloc;
394 /* If we failed on a TSB grow, we are under serious
395 * memory pressure so don't try to grow any more.
397 if (mm->context.tsb_block[tsb_index].tsb != NULL)
398 mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
402 /* Mark all tags as invalid. */
403 tsb_init(new_tsb, new_size);
405 /* Ok, we are about to commit the changes. If we are
406 * growing an existing TSB the locking is very tricky,
409 * We have to hold mm->context.lock while committing to the
410 * new TSB, this synchronizes us with processors in
411 * flush_tsb_user() and switch_mm() for this address space.
413 * But even with that lock held, processors run asynchronously
414 * accessing the old TSB via TLB miss handling. This is OK
415 * because those actions are just propagating state from the
416 * Linux page tables into the TSB, page table mappings are not
417 * being changed. If a real fault occurs, the processor will
418 * synchronize with us when it hits flush_tsb_user(), this is
419 * also true for the case where vmscan is modifying the page
420 * tables. The only thing we need to be careful with is to
421 * skip any locked TSB entries during copy_tsb().
423 * When we finish committing to the new TSB, we have to drop
424 * the lock and ask all other cpus running this address space
425 * to run tsb_context_switch() to see the new TSB table.
427 spin_lock_irqsave(&mm->context.lock, flags);
429 old_tsb = mm->context.tsb_block[tsb_index].tsb;
431 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
432 old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
436 /* Handle multiple threads trying to grow the TSB at the same time.
437 * One will get in here first, and bump the size and the RSS limit.
438 * The others will get in here next and hit this check.
440 if (unlikely(old_tsb &&
441 (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
442 spin_unlock_irqrestore(&mm->context.lock, flags);
444 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
448 mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
451 extern void copy_tsb(unsigned long old_tsb_base,
452 unsigned long old_tsb_size,
453 unsigned long new_tsb_base,
454 unsigned long new_tsb_size);
455 unsigned long old_tsb_base = (unsigned long) old_tsb;
456 unsigned long new_tsb_base = (unsigned long) new_tsb;
458 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
459 old_tsb_base = __pa(old_tsb_base);
460 new_tsb_base = __pa(new_tsb_base);
462 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
465 mm->context.tsb_block[tsb_index].tsb = new_tsb;
466 setup_tsb_params(mm, tsb_index, new_size);
468 spin_unlock_irqrestore(&mm->context.lock, flags);
470 /* If old_tsb is NULL, we're being invoked for the first time
471 * from init_new_context().
474 /* Reload it on the local cpu. */
475 tsb_context_switch(mm);
477 /* Now force other processors to do the same. */
482 /* Now it is safe to free the old tsb. */
483 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
487 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
489 unsigned long mm_rss = get_mm_rss(mm);
490 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
491 unsigned long saved_hugetlb_pte_count;
492 unsigned long saved_thp_pte_count;
496 spin_lock_init(&mm->context.lock);
498 mm->context.sparc64_ctx_val = 0UL;
500 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
501 /* We reset them to zero because the fork() page copying
502 * will re-increment the counters as the parent PTEs are
503 * copied into the child address space.
505 saved_hugetlb_pte_count = mm->context.hugetlb_pte_count;
506 saved_thp_pte_count = mm->context.thp_pte_count;
507 mm->context.hugetlb_pte_count = 0;
508 mm->context.thp_pte_count = 0;
510 mm_rss -= saved_thp_pte_count * (HPAGE_SIZE / PAGE_SIZE);
513 /* copy_mm() copies over the parent's mm_struct before calling
514 * us, so we need to zero out the TSB pointer or else tsb_grow()
515 * will be confused and think there is an older TSB to free up.
517 for (i = 0; i < MM_NUM_TSBS; i++)
518 mm->context.tsb_block[i].tsb = NULL;
520 /* If this is fork, inherit the parent's TSB size. We would
521 * grow it to that size on the first page fault anyways.
523 tsb_grow(mm, MM_TSB_BASE, mm_rss);
525 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
526 if (unlikely(saved_hugetlb_pte_count + saved_thp_pte_count))
527 tsb_grow(mm, MM_TSB_HUGE,
528 (saved_hugetlb_pte_count + saved_thp_pte_count) *
529 REAL_HPAGE_PER_HPAGE);
532 if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
538 static void tsb_destroy_one(struct tsb_config *tp)
540 unsigned long cache_index;
544 cache_index = tp->tsb_reg_val & 0x7UL;
545 kmem_cache_free(tsb_caches[cache_index], tp->tsb);
547 tp->tsb_reg_val = 0UL;
550 void destroy_context(struct mm_struct *mm)
552 unsigned long flags, i;
554 for (i = 0; i < MM_NUM_TSBS; i++)
555 tsb_destroy_one(&mm->context.tsb_block[i]);
557 spin_lock_irqsave(&ctx_alloc_lock, flags);
559 if (CTX_VALID(mm->context)) {
560 unsigned long nr = CTX_NRBITS(mm->context);
561 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
564 spin_unlock_irqrestore(&ctx_alloc_lock, flags);