1 /* arch/sparc64/mm/tsb.c
3 * Copyright (C) 2006 David S. Miller <davem@davemloft.net>
6 #include <linux/kernel.h>
7 #include <asm/system.h>
9 #include <asm/tlbflush.h>
11 #include <asm/mmu_context.h>
12 #include <asm/pgtable.h>
15 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
17 static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long nentries)
20 return vaddr & (nentries - 1);
23 static inline int tag_compare(unsigned long tag, unsigned long vaddr, unsigned long context)
25 return (tag == ((vaddr >> 22) | (context << 48)));
28 /* TSB flushes need only occur on the processor initiating the address
29 * space modification, not on each cpu the address space has run on.
30 * Only the TLB flush needs that treatment.
33 void flush_tsb_kernel_range(unsigned long start, unsigned long end)
37 for (v = start; v < end; v += PAGE_SIZE) {
38 unsigned long hash = tsb_hash(v, KERNEL_TSB_NENTRIES);
39 struct tsb *ent = &swapper_tsb[hash];
41 if (tag_compare(ent->tag, v, 0)) {
43 membar_storeload_storestore();
48 void flush_tsb_user(struct mmu_gather *mp)
50 struct mm_struct *mm = mp->mm;
51 struct tsb *tsb = mm->context.tsb;
52 unsigned long nentries = mm->context.tsb_nentries;
53 unsigned long ctx, base;
56 if (unlikely(!CTX_VALID(mm->context)))
59 ctx = CTX_HWBITS(mm->context);
61 if (tlb_type == cheetah_plus)
64 base = (unsigned long) tsb;
66 for (i = 0; i < mp->tlb_nr; i++) {
67 unsigned long v = mp->vaddrs[i];
68 unsigned long tag, ent, hash;
72 hash = tsb_hash(v, nentries);
73 ent = base + (hash * sizeof(struct tsb));
74 tag = (v >> 22UL) | (ctx << 48UL);
80 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_bytes)
82 unsigned long tsb_reg, base, tsb_paddr;
83 unsigned long page_sz, tte;
85 mm->context.tsb_nentries = tsb_bytes / sizeof(struct tsb);
88 tte = (_PAGE_VALID | _PAGE_L | _PAGE_CP |
89 _PAGE_CV | _PAGE_P | _PAGE_W);
90 tsb_paddr = __pa(mm->context.tsb);
91 BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
93 /* Use the smallest page size that can map the whole TSB
99 #ifdef DCACHE_ALIASING_POSSIBLE
100 base += (tsb_paddr & 8192);
127 page_sz = 512 * 1024;
133 page_sz = 512 * 1024;
139 page_sz = 512 * 1024;
145 page_sz = 4 * 1024 * 1024;
152 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
153 /* Physical mapping, no locked TLB entry for TSB. */
154 tsb_reg |= tsb_paddr;
156 mm->context.tsb_reg_val = tsb_reg;
157 mm->context.tsb_map_vaddr = 0;
158 mm->context.tsb_map_pte = 0;
161 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
162 tte |= (tsb_paddr & ~(page_sz - 1UL));
164 mm->context.tsb_reg_val = tsb_reg;
165 mm->context.tsb_map_vaddr = base;
166 mm->context.tsb_map_pte = tte;
169 /* Setup the Hypervisor TSB descriptor. */
170 if (tlb_type == hypervisor) {
171 struct hv_tsb_descr *hp = &mm->context.tsb_descr;
176 hp->pgsz_idx = HV_PGSZ_IDX_8K;
180 hp->pgsz_idx = HV_PGSZ_IDX_64K;
184 hp->pgsz_idx = HV_PGSZ_IDX_512K;
187 case 4 * 1024 * 1024:
188 hp->pgsz_idx = HV_PGSZ_IDX_4MB;
192 hp->num_ttes = tsb_bytes / 16;
197 hp->pgsz_mask = HV_PGSZ_MASK_8K;
201 hp->pgsz_mask = HV_PGSZ_MASK_64K;
205 hp->pgsz_mask = HV_PGSZ_MASK_512K;
208 case 4 * 1024 * 1024:
209 hp->pgsz_mask = HV_PGSZ_MASK_4MB;
212 hp->tsb_base = tsb_paddr;
217 /* The page tables are locked against modifications while this
220 * XXX do some prefetching...
222 static void copy_tsb(struct tsb *old_tsb, unsigned long old_size,
223 struct tsb *new_tsb, unsigned long new_size)
225 unsigned long old_nentries = old_size / sizeof(struct tsb);
226 unsigned long new_nentries = new_size / sizeof(struct tsb);
229 for (i = 0; i < old_nentries; i++) {
230 register unsigned long tag asm("o4");
231 register unsigned long pte asm("o5");
232 unsigned long v, hash;
234 if (tlb_type == cheetah_plus) {
235 __asm__ __volatile__(
237 : "=r" (tag), "=r" (pte)
238 : "r" (__pa(&old_tsb[i])),
239 "i" (ASI_QUAD_LDD_PHYS));
241 __asm__ __volatile__(
243 : "=r" (tag), "=r" (pte)
245 "i" (ASI_NUCLEUS_QUAD_LDD));
248 if (!tag || (tag & (1UL << TSB_TAG_LOCK_BIT)))
251 /* We only put base page size PTEs into the TSB,
252 * but that might change in the future. This code
253 * would need to be changed if we start putting larger
254 * page size PTEs into there.
256 WARN_ON((pte & _PAGE_ALL_SZ_BITS) != _PAGE_SZBITS);
258 /* The tag holds bits 22 to 63 of the virtual address
259 * and the context. Clear out the context, and shift
260 * up to make a virtual address.
262 v = (tag & ((1UL << 42UL) - 1UL)) << 22UL;
264 /* The implied bits of the tag (bits 13 to 21) are
265 * determined by the TSB entry index, so fill that in.
267 v |= (i & (512UL - 1UL)) << 13UL;
269 hash = tsb_hash(v, new_nentries);
270 if (tlb_type == cheetah_plus) {
271 __asm__ __volatile__(
272 "stxa %0, [%1] %2\n\t"
276 "r" (__pa(&new_tsb[hash].tag)),
277 "i" (ASI_PHYS_USE_EC),
279 "r" (__pa(&new_tsb[hash].pte)));
281 new_tsb[hash].tag = tag;
282 new_tsb[hash].pte = pte;
287 /* When the RSS of an address space exceeds mm->context.tsb_rss_limit,
288 * update_mmu_cache() invokes this routine to try and grow the TSB.
289 * When we reach the maximum TSB size supported, we stick ~0UL into
290 * mm->context.tsb_rss_limit so the grow checks in update_mmu_cache()
291 * will not trigger any longer.
293 * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
294 * of two. The TSB must be aligned to it's size, so f.e. a 512K TSB
295 * must be 512K aligned.
297 * The idea here is to grow the TSB when the RSS of the process approaches
298 * the number of entries that the current TSB can hold at once. Currently,
299 * we trigger when the RSS hits 3/4 of the TSB capacity.
301 void tsb_grow(struct mm_struct *mm, unsigned long rss, gfp_t gfp_flags)
303 unsigned long max_tsb_size = 1 * 1024 * 1024;
304 unsigned long size, old_size;
308 if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
309 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
311 for (size = PAGE_SIZE; size < max_tsb_size; size <<= 1UL) {
312 unsigned long n_entries = size / sizeof(struct tsb);
314 n_entries = (n_entries * 3) / 4;
319 page = alloc_pages(gfp_flags | __GFP_ZERO, get_order(size));
323 if (size == max_tsb_size)
324 mm->context.tsb_rss_limit = ~0UL;
326 mm->context.tsb_rss_limit =
327 ((size / sizeof(struct tsb)) * 3) / 4;
329 old_tsb = mm->context.tsb;
330 old_size = mm->context.tsb_nentries * sizeof(struct tsb);
333 copy_tsb(old_tsb, old_size, page_address(page), size);
335 mm->context.tsb = page_address(page);
336 setup_tsb_params(mm, size);
338 /* If old_tsb is NULL, we're being invoked for the first time
339 * from init_new_context().
342 /* Now force all other processors to reload the new
347 /* Finally reload it on the local cpu. No further
348 * references will remain to the old TSB and we can
351 tsb_context_switch(mm);
353 free_pages((unsigned long) old_tsb, get_order(old_size));
357 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
360 mm->context.sparc64_ctx_val = 0UL;
362 /* copy_mm() copies over the parent's mm_struct before calling
363 * us, so we need to zero out the TSB pointer or else tsb_grow()
364 * will be confused and think there is an older TSB to free up.
366 mm->context.tsb = NULL;
367 tsb_grow(mm, 0, GFP_KERNEL);
369 if (unlikely(!mm->context.tsb))
375 void destroy_context(struct mm_struct *mm)
377 unsigned long size = mm->context.tsb_nentries * sizeof(struct tsb);
379 free_pages((unsigned long) mm->context.tsb, get_order(size));
381 /* We can remove these later, but for now it's useful
382 * to catch any bogus post-destroy_context() references
385 mm->context.tsb = NULL;
386 mm->context.tsb_reg_val = 0UL;
388 spin_lock(&ctx_alloc_lock);
390 if (CTX_VALID(mm->context)) {
391 unsigned long nr = CTX_NRBITS(mm->context);
392 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
395 spin_unlock(&ctx_alloc_lock);