Linux Audio

Check our new training course

Loading...
v3.1
 
  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
 18extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
 19
 20static 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
 26static 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
 36void 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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 50static 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
 69void 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
119static 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
239struct kmem_cache *pgtable_cache __read_mostly;
240
241static struct kmem_cache *tsb_caches[8] __read_mostly;
242
243static 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
254void __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
281int sysctl_tsb_ratio = -2;
282
283static 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 */
309void 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
332retry_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
446int 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
489static 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
501void 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}
v6.2
  1// SPDX-License-Identifier: GPL-2.0
  2/* arch/sparc64/mm/tsb.c
  3 *
  4 * Copyright (C) 2006, 2008 David S. Miller <davem@davemloft.net>
  5 */
  6
  7#include <linux/kernel.h>
  8#include <linux/preempt.h>
  9#include <linux/slab.h>
 10#include <linux/mm_types.h>
 11#include <linux/pgtable.h>
 12
 13#include <asm/page.h>
 
 
 14#include <asm/mmu_context.h>
 15#include <asm/setup.h>
 16#include <asm/tsb.h>
 17#include <asm/tlb.h>
 18#include <asm/oplib.h>
 19
 20extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
 21
 22static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long hash_shift, unsigned long nentries)
 23{
 24	vaddr >>= hash_shift;
 25	return vaddr & (nentries - 1);
 26}
 27
 28static inline int tag_compare(unsigned long tag, unsigned long vaddr)
 29{
 30	return (tag == (vaddr >> 22));
 31}
 32
 33static void flush_tsb_kernel_range_scan(unsigned long start, unsigned long end)
 34{
 35	unsigned long idx;
 36
 37	for (idx = 0; idx < KERNEL_TSB_NENTRIES; idx++) {
 38		struct tsb *ent = &swapper_tsb[idx];
 39		unsigned long match = idx << 13;
 40
 41		match |= (ent->tag << 22);
 42		if (match >= start && match < end)
 43			ent->tag = (1UL << TSB_TAG_INVALID_BIT);
 44	}
 45}
 46
 47/* TSB flushes need only occur on the processor initiating the address
 48 * space modification, not on each cpu the address space has run on.
 49 * Only the TLB flush needs that treatment.
 50 */
 51
 52void flush_tsb_kernel_range(unsigned long start, unsigned long end)
 53{
 54	unsigned long v;
 55
 56	if ((end - start) >> PAGE_SHIFT >= 2 * KERNEL_TSB_NENTRIES)
 57		return flush_tsb_kernel_range_scan(start, end);
 58
 59	for (v = start; v < end; v += PAGE_SIZE) {
 60		unsigned long hash = tsb_hash(v, PAGE_SHIFT,
 61					      KERNEL_TSB_NENTRIES);
 62		struct tsb *ent = &swapper_tsb[hash];
 63
 64		if (tag_compare(ent->tag, v))
 65			ent->tag = (1UL << TSB_TAG_INVALID_BIT);
 66	}
 67}
 68
 69static void __flush_tsb_one_entry(unsigned long tsb, unsigned long v,
 70				  unsigned long hash_shift,
 71				  unsigned long nentries)
 72{
 73	unsigned long tag, ent, hash;
 74
 75	v &= ~0x1UL;
 76	hash = tsb_hash(v, hash_shift, nentries);
 77	ent = tsb + (hash * sizeof(struct tsb));
 78	tag = (v >> 22UL);
 79
 80	tsb_flush(ent, tag);
 81}
 82
 83static void __flush_tsb_one(struct tlb_batch *tb, unsigned long hash_shift,
 84			    unsigned long tsb, unsigned long nentries)
 85{
 86	unsigned long i;
 87
 88	for (i = 0; i < tb->tlb_nr; i++)
 89		__flush_tsb_one_entry(tsb, tb->vaddrs[i], hash_shift, nentries);
 90}
 91
 92#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 93static void __flush_huge_tsb_one_entry(unsigned long tsb, unsigned long v,
 94				       unsigned long hash_shift,
 95				       unsigned long nentries,
 96				       unsigned int hugepage_shift)
 97{
 98	unsigned int hpage_entries;
 99	unsigned int i;
100
101	hpage_entries = 1 << (hugepage_shift - hash_shift);
102	for (i = 0; i < hpage_entries; i++)
103		__flush_tsb_one_entry(tsb, v + (i << hash_shift), hash_shift,
104				      nentries);
105}
106
107static void __flush_huge_tsb_one(struct tlb_batch *tb, unsigned long hash_shift,
108				 unsigned long tsb, unsigned long nentries,
109				 unsigned int hugepage_shift)
110{
111	unsigned long i;
112
113	for (i = 0; i < tb->tlb_nr; i++)
114		__flush_huge_tsb_one_entry(tsb, tb->vaddrs[i], hash_shift,
115					   nentries, hugepage_shift);
116}
117#endif
118
119void flush_tsb_user(struct tlb_batch *tb)
120{
121	struct mm_struct *mm = tb->mm;
122	unsigned long nentries, base, flags;
123
124	spin_lock_irqsave(&mm->context.lock, flags);
125
126	if (tb->hugepage_shift < REAL_HPAGE_SHIFT) {
127		base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
128		nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
129		if (tlb_type == cheetah_plus || tlb_type == hypervisor)
130			base = __pa(base);
131		if (tb->hugepage_shift == PAGE_SHIFT)
132			__flush_tsb_one(tb, PAGE_SHIFT, base, nentries);
133#if defined(CONFIG_HUGETLB_PAGE)
134		else
135			__flush_huge_tsb_one(tb, PAGE_SHIFT, base, nentries,
136					     tb->hugepage_shift);
137#endif
138	}
139#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
140	else if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
141		base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
142		nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
143		if (tlb_type == cheetah_plus || tlb_type == hypervisor)
144			base = __pa(base);
145		__flush_huge_tsb_one(tb, REAL_HPAGE_SHIFT, base, nentries,
146				     tb->hugepage_shift);
147	}
148#endif
149	spin_unlock_irqrestore(&mm->context.lock, flags);
150}
151
152void flush_tsb_user_page(struct mm_struct *mm, unsigned long vaddr,
153			 unsigned int hugepage_shift)
154{
155	unsigned long nentries, base, flags;
156
157	spin_lock_irqsave(&mm->context.lock, flags);
158
159	if (hugepage_shift < REAL_HPAGE_SHIFT) {
160		base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
161		nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
162		if (tlb_type == cheetah_plus || tlb_type == hypervisor)
163			base = __pa(base);
164		if (hugepage_shift == PAGE_SHIFT)
165			__flush_tsb_one_entry(base, vaddr, PAGE_SHIFT,
166					      nentries);
167#if defined(CONFIG_HUGETLB_PAGE)
168		else
169			__flush_huge_tsb_one_entry(base, vaddr, PAGE_SHIFT,
170						   nentries, hugepage_shift);
171#endif
172	}
173#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
174	else if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
175		base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
176		nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
177		if (tlb_type == cheetah_plus || tlb_type == hypervisor)
178			base = __pa(base);
179		__flush_huge_tsb_one_entry(base, vaddr, REAL_HPAGE_SHIFT,
180					   nentries, hugepage_shift);
181	}
182#endif
183	spin_unlock_irqrestore(&mm->context.lock, flags);
184}
185
 
186#define HV_PGSZ_IDX_BASE	HV_PGSZ_IDX_8K
187#define HV_PGSZ_MASK_BASE	HV_PGSZ_MASK_8K
 
 
 
 
 
 
188
189#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 
 
 
 
 
 
 
190#define HV_PGSZ_IDX_HUGE	HV_PGSZ_IDX_4MB
191#define HV_PGSZ_MASK_HUGE	HV_PGSZ_MASK_4MB
 
 
 
192#endif
193
194static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
195{
196	unsigned long tsb_reg, base, tsb_paddr;
197	unsigned long page_sz, tte;
198
199	mm->context.tsb_block[tsb_idx].tsb_nentries =
200		tsb_bytes / sizeof(struct tsb);
201
202	switch (tsb_idx) {
203	case MM_TSB_BASE:
204		base = TSBMAP_8K_BASE;
205		break;
206#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
207	case MM_TSB_HUGE:
208		base = TSBMAP_4M_BASE;
209		break;
210#endif
211	default:
212		BUG();
213	}
214
215	tte = pgprot_val(PAGE_KERNEL_LOCKED);
216	tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
217	BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
218
219	/* Use the smallest page size that can map the whole TSB
220	 * in one TLB entry.
221	 */
222	switch (tsb_bytes) {
223	case 8192 << 0:
224		tsb_reg = 0x0UL;
225#ifdef DCACHE_ALIASING_POSSIBLE
226		base += (tsb_paddr & 8192);
227#endif
228		page_sz = 8192;
229		break;
230
231	case 8192 << 1:
232		tsb_reg = 0x1UL;
233		page_sz = 64 * 1024;
234		break;
235
236	case 8192 << 2:
237		tsb_reg = 0x2UL;
238		page_sz = 64 * 1024;
239		break;
240
241	case 8192 << 3:
242		tsb_reg = 0x3UL;
243		page_sz = 64 * 1024;
244		break;
245
246	case 8192 << 4:
247		tsb_reg = 0x4UL;
248		page_sz = 512 * 1024;
249		break;
250
251	case 8192 << 5:
252		tsb_reg = 0x5UL;
253		page_sz = 512 * 1024;
254		break;
255
256	case 8192 << 6:
257		tsb_reg = 0x6UL;
258		page_sz = 512 * 1024;
259		break;
260
261	case 8192 << 7:
262		tsb_reg = 0x7UL;
263		page_sz = 4 * 1024 * 1024;
264		break;
265
266	default:
267		printk(KERN_ERR "TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
268		       current->comm, current->pid, tsb_bytes);
269		BUG();
270	}
271	tte |= pte_sz_bits(page_sz);
272
273	if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
274		/* Physical mapping, no locked TLB entry for TSB.  */
275		tsb_reg |= tsb_paddr;
276
277		mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
278		mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
279		mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
280	} else {
281		tsb_reg |= base;
282		tsb_reg |= (tsb_paddr & (page_sz - 1UL));
283		tte |= (tsb_paddr & ~(page_sz - 1UL));
284
285		mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
286		mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
287		mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
288	}
289
290	/* Setup the Hypervisor TSB descriptor.  */
291	if (tlb_type == hypervisor) {
292		struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
293
294		switch (tsb_idx) {
295		case MM_TSB_BASE:
296			hp->pgsz_idx = HV_PGSZ_IDX_BASE;
297			break;
298#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
299		case MM_TSB_HUGE:
300			hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
301			break;
302#endif
303		default:
304			BUG();
305		}
306		hp->assoc = 1;
307		hp->num_ttes = tsb_bytes / 16;
308		hp->ctx_idx = 0;
309		switch (tsb_idx) {
310		case MM_TSB_BASE:
311			hp->pgsz_mask = HV_PGSZ_MASK_BASE;
312			break;
313#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
314		case MM_TSB_HUGE:
315			hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
316			break;
317#endif
318		default:
319			BUG();
320		}
321		hp->tsb_base = tsb_paddr;
322		hp->resv = 0;
323	}
324}
325
326struct kmem_cache *pgtable_cache __read_mostly;
327
328static struct kmem_cache *tsb_caches[8] __read_mostly;
329
330static const char *tsb_cache_names[8] = {
331	"tsb_8KB",
332	"tsb_16KB",
333	"tsb_32KB",
334	"tsb_64KB",
335	"tsb_128KB",
336	"tsb_256KB",
337	"tsb_512KB",
338	"tsb_1MB",
339};
340
341void __init pgtable_cache_init(void)
342{
343	unsigned long i;
344
345	pgtable_cache = kmem_cache_create("pgtable_cache",
346					  PAGE_SIZE, PAGE_SIZE,
347					  0,
348					  _clear_page);
349	if (!pgtable_cache) {
350		prom_printf("pgtable_cache_init(): Could not create!\n");
351		prom_halt();
352	}
353
354	for (i = 0; i < ARRAY_SIZE(tsb_cache_names); i++) {
355		unsigned long size = 8192 << i;
356		const char *name = tsb_cache_names[i];
357
358		tsb_caches[i] = kmem_cache_create(name,
359						  size, size,
360						  0, NULL);
361		if (!tsb_caches[i]) {
362			prom_printf("Could not create %s cache\n", name);
363			prom_halt();
364		}
365	}
366}
367
368int sysctl_tsb_ratio = -2;
369
370static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
371{
372	unsigned long num_ents = (new_size / sizeof(struct tsb));
373
374	if (sysctl_tsb_ratio < 0)
375		return num_ents - (num_ents >> -sysctl_tsb_ratio);
376	else
377		return num_ents + (num_ents >> sysctl_tsb_ratio);
378}
379
380/* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
381 * do_sparc64_fault() invokes this routine to try and grow it.
382 *
383 * When we reach the maximum TSB size supported, we stick ~0UL into
384 * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
385 * will not trigger any longer.
386 *
387 * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
388 * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
389 * must be 512K aligned.  It also must be physically contiguous, so we
390 * cannot use vmalloc().
391 *
392 * The idea here is to grow the TSB when the RSS of the process approaches
393 * the number of entries that the current TSB can hold at once.  Currently,
394 * we trigger when the RSS hits 3/4 of the TSB capacity.
395 */
396void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
397{
398	unsigned long max_tsb_size = 1 * 1024 * 1024;
399	unsigned long new_size, old_size, flags;
400	struct tsb *old_tsb, *new_tsb;
401	unsigned long new_cache_index, old_cache_index;
402	unsigned long new_rss_limit;
403	gfp_t gfp_flags;
404
405	if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
406		max_tsb_size = (PAGE_SIZE << MAX_ORDER);
407
408	new_cache_index = 0;
409	for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
410		new_rss_limit = tsb_size_to_rss_limit(new_size);
411		if (new_rss_limit > rss)
412			break;
413		new_cache_index++;
414	}
415
416	if (new_size == max_tsb_size)
417		new_rss_limit = ~0UL;
418
419retry_tsb_alloc:
420	gfp_flags = GFP_KERNEL;
421	if (new_size > (PAGE_SIZE * 2))
422		gfp_flags |= __GFP_NOWARN | __GFP_NORETRY;
423
424	new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
425					gfp_flags, numa_node_id());
426	if (unlikely(!new_tsb)) {
427		/* Not being able to fork due to a high-order TSB
428		 * allocation failure is very bad behavior.  Just back
429		 * down to a 0-order allocation and force no TSB
430		 * growing for this address space.
431		 */
432		if (mm->context.tsb_block[tsb_index].tsb == NULL &&
433		    new_cache_index > 0) {
434			new_cache_index = 0;
435			new_size = 8192;
436			new_rss_limit = ~0UL;
437			goto retry_tsb_alloc;
438		}
439
440		/* If we failed on a TSB grow, we are under serious
441		 * memory pressure so don't try to grow any more.
442		 */
443		if (mm->context.tsb_block[tsb_index].tsb != NULL)
444			mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
445		return;
446	}
447
448	/* Mark all tags as invalid.  */
449	tsb_init(new_tsb, new_size);
450
451	/* Ok, we are about to commit the changes.  If we are
452	 * growing an existing TSB the locking is very tricky,
453	 * so WATCH OUT!
454	 *
455	 * We have to hold mm->context.lock while committing to the
456	 * new TSB, this synchronizes us with processors in
457	 * flush_tsb_user() and switch_mm() for this address space.
458	 *
459	 * But even with that lock held, processors run asynchronously
460	 * accessing the old TSB via TLB miss handling.  This is OK
461	 * because those actions are just propagating state from the
462	 * Linux page tables into the TSB, page table mappings are not
463	 * being changed.  If a real fault occurs, the processor will
464	 * synchronize with us when it hits flush_tsb_user(), this is
465	 * also true for the case where vmscan is modifying the page
466	 * tables.  The only thing we need to be careful with is to
467	 * skip any locked TSB entries during copy_tsb().
468	 *
469	 * When we finish committing to the new TSB, we have to drop
470	 * the lock and ask all other cpus running this address space
471	 * to run tsb_context_switch() to see the new TSB table.
472	 */
473	spin_lock_irqsave(&mm->context.lock, flags);
474
475	old_tsb = mm->context.tsb_block[tsb_index].tsb;
476	old_cache_index =
477		(mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
478	old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
479		    sizeof(struct tsb));
480
481
482	/* Handle multiple threads trying to grow the TSB at the same time.
483	 * One will get in here first, and bump the size and the RSS limit.
484	 * The others will get in here next and hit this check.
485	 */
486	if (unlikely(old_tsb &&
487		     (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
488		spin_unlock_irqrestore(&mm->context.lock, flags);
489
490		kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
491		return;
492	}
493
494	mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
495
496	if (old_tsb) {
497		extern void copy_tsb(unsigned long old_tsb_base,
498				     unsigned long old_tsb_size,
499				     unsigned long new_tsb_base,
500				     unsigned long new_tsb_size,
501				     unsigned long page_size_shift);
502		unsigned long old_tsb_base = (unsigned long) old_tsb;
503		unsigned long new_tsb_base = (unsigned long) new_tsb;
504
505		if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
506			old_tsb_base = __pa(old_tsb_base);
507			new_tsb_base = __pa(new_tsb_base);
508		}
509		copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size,
510			tsb_index == MM_TSB_BASE ?
511			PAGE_SHIFT : REAL_HPAGE_SHIFT);
512	}
513
514	mm->context.tsb_block[tsb_index].tsb = new_tsb;
515	setup_tsb_params(mm, tsb_index, new_size);
516
517	spin_unlock_irqrestore(&mm->context.lock, flags);
518
519	/* If old_tsb is NULL, we're being invoked for the first time
520	 * from init_new_context().
521	 */
522	if (old_tsb) {
523		/* Reload it on the local cpu.  */
524		tsb_context_switch(mm);
525
526		/* Now force other processors to do the same.  */
527		preempt_disable();
528		smp_tsb_sync(mm);
529		preempt_enable();
530
531		/* Now it is safe to free the old tsb.  */
532		kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
533	}
534}
535
536int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
537{
538	unsigned long mm_rss = get_mm_rss(mm);
539#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
540	unsigned long saved_hugetlb_pte_count;
541	unsigned long saved_thp_pte_count;
542#endif
543	unsigned int i;
544
545	spin_lock_init(&mm->context.lock);
546
547	mm->context.sparc64_ctx_val = 0UL;
548
549	mm->context.tag_store = NULL;
550	spin_lock_init(&mm->context.tag_lock);
551
552#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
553	/* We reset them to zero because the fork() page copying
554	 * will re-increment the counters as the parent PTEs are
555	 * copied into the child address space.
556	 */
557	saved_hugetlb_pte_count = mm->context.hugetlb_pte_count;
558	saved_thp_pte_count = mm->context.thp_pte_count;
559	mm->context.hugetlb_pte_count = 0;
560	mm->context.thp_pte_count = 0;
561
562	mm_rss -= saved_thp_pte_count * (HPAGE_SIZE / PAGE_SIZE);
563#endif
564
565	/* copy_mm() copies over the parent's mm_struct before calling
566	 * us, so we need to zero out the TSB pointer or else tsb_grow()
567	 * will be confused and think there is an older TSB to free up.
568	 */
569	for (i = 0; i < MM_NUM_TSBS; i++)
570		mm->context.tsb_block[i].tsb = NULL;
571
572	/* If this is fork, inherit the parent's TSB size.  We would
573	 * grow it to that size on the first page fault anyways.
574	 */
575	tsb_grow(mm, MM_TSB_BASE, mm_rss);
576
577#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
578	if (unlikely(saved_hugetlb_pte_count + saved_thp_pte_count))
579		tsb_grow(mm, MM_TSB_HUGE,
580			 (saved_hugetlb_pte_count + saved_thp_pte_count) *
581			 REAL_HPAGE_PER_HPAGE);
582#endif
583
584	if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
585		return -ENOMEM;
586
587	return 0;
588}
589
590static void tsb_destroy_one(struct tsb_config *tp)
591{
592	unsigned long cache_index;
593
594	if (!tp->tsb)
595		return;
596	cache_index = tp->tsb_reg_val & 0x7UL;
597	kmem_cache_free(tsb_caches[cache_index], tp->tsb);
598	tp->tsb = NULL;
599	tp->tsb_reg_val = 0UL;
600}
601
602void destroy_context(struct mm_struct *mm)
603{
604	unsigned long flags, i;
605
606	for (i = 0; i < MM_NUM_TSBS; i++)
607		tsb_destroy_one(&mm->context.tsb_block[i]);
608
609	spin_lock_irqsave(&ctx_alloc_lock, flags);
610
611	if (CTX_VALID(mm->context)) {
612		unsigned long nr = CTX_NRBITS(mm->context);
613		mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
614	}
615
616	spin_unlock_irqrestore(&ctx_alloc_lock, flags);
617
618	/* If ADI tag storage was allocated for this task, free it */
619	if (mm->context.tag_store) {
620		tag_storage_desc_t *tag_desc;
621		unsigned long max_desc;
622		unsigned char *tags;
623
624		tag_desc = mm->context.tag_store;
625		max_desc = PAGE_SIZE/sizeof(tag_storage_desc_t);
626		for (i = 0; i < max_desc; i++) {
627			tags = tag_desc->tags;
628			tag_desc->tags = NULL;
629			kfree(tags);
630			tag_desc++;
631		}
632		kfree(mm->context.tag_store);
633		mm->context.tag_store = NULL;
634	}
635}