[S390] Fix yet another two section mismatches.
[linux-2.6.git] / arch / sparc64 / mm / tsb.c
index 1af797a..8eb8a7c 100644 (file)
@@ -15,9 +15,9 @@
 
 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
 
-static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long nentries)
+static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long hash_shift, unsigned long nentries)
 {
-       vaddr >>= PAGE_SHIFT;
+       vaddr >>= hash_shift;
        return vaddr & (nentries - 1);
 }
 
@@ -36,7 +36,8 @@ void flush_tsb_kernel_range(unsigned long start, unsigned long end)
        unsigned long v;
 
        for (v = start; v < end; v += PAGE_SIZE) {
-               unsigned long hash = tsb_hash(v, KERNEL_TSB_NENTRIES);
+               unsigned long hash = tsb_hash(v, PAGE_SHIFT,
+                                             KERNEL_TSB_NENTRIES);
                struct tsb *ent = &swapper_tsb[hash];
 
                if (tag_compare(ent->tag, v)) {
@@ -46,49 +47,91 @@ void flush_tsb_kernel_range(unsigned long start, unsigned long end)
        }
 }
 
-void flush_tsb_user(struct mmu_gather *mp)
+static void __flush_tsb_one(struct mmu_gather *mp, unsigned long hash_shift, unsigned long tsb, unsigned long nentries)
 {
-       struct mm_struct *mm = mp->mm;
-       unsigned long nentries, base, flags;
-       struct tsb *tsb;
-       int i;
-
-       spin_lock_irqsave(&mm->context.lock, flags);
-
-       tsb = mm->context.tsb;
-       nentries = mm->context.tsb_nentries;
+       unsigned long i;
 
-       if (tlb_type == cheetah_plus || tlb_type == hypervisor)
-               base = __pa(tsb);
-       else
-               base = (unsigned long) tsb;
-       
        for (i = 0; i < mp->tlb_nr; i++) {
                unsigned long v = mp->vaddrs[i];
                unsigned long tag, ent, hash;
 
                v &= ~0x1UL;
 
-               hash = tsb_hash(v, nentries);
-               ent = base + (hash * sizeof(struct tsb));
+               hash = tsb_hash(v, hash_shift, nentries);
+               ent = tsb + (hash * sizeof(struct tsb));
                tag = (v >> 22UL);
 
                tsb_flush(ent, tag);
        }
+}
+
+void flush_tsb_user(struct mmu_gather *mp)
+{
+       struct mm_struct *mm = mp->mm;
+       unsigned long nentries, base, flags;
+
+       spin_lock_irqsave(&mm->context.lock, flags);
 
+       base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
+       nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
+       if (tlb_type == cheetah_plus || tlb_type == hypervisor)
+               base = __pa(base);
+       __flush_tsb_one(mp, PAGE_SHIFT, base, nentries);
+
+#ifdef CONFIG_HUGETLB_PAGE
+       if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
+               base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
+               nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
+               if (tlb_type == cheetah_plus || tlb_type == hypervisor)
+                       base = __pa(base);
+               __flush_tsb_one(mp, HPAGE_SHIFT, base, nentries);
+       }
+#endif
        spin_unlock_irqrestore(&mm->context.lock, flags);
 }
 
-static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_bytes)
+#if defined(CONFIG_SPARC64_PAGE_SIZE_8KB)
+#define HV_PGSZ_IDX_BASE       HV_PGSZ_IDX_8K
+#define HV_PGSZ_MASK_BASE      HV_PGSZ_MASK_8K
+#elif defined(CONFIG_SPARC64_PAGE_SIZE_64KB)
+#define HV_PGSZ_IDX_BASE       HV_PGSZ_IDX_64K
+#define HV_PGSZ_MASK_BASE      HV_PGSZ_MASK_64K
+#elif defined(CONFIG_SPARC64_PAGE_SIZE_512KB)
+#define HV_PGSZ_IDX_BASE       HV_PGSZ_IDX_512K
+#define HV_PGSZ_MASK_BASE      HV_PGSZ_MASK_512K
+#elif defined(CONFIG_SPARC64_PAGE_SIZE_4MB)
+#define HV_PGSZ_IDX_BASE       HV_PGSZ_IDX_4MB
+#define HV_PGSZ_MASK_BASE      HV_PGSZ_MASK_4MB
+#else
+#error Broken base page size setting...
+#endif
+
+#ifdef CONFIG_HUGETLB_PAGE
+#if defined(CONFIG_HUGETLB_PAGE_SIZE_64K)
+#define HV_PGSZ_IDX_HUGE       HV_PGSZ_IDX_64K
+#define HV_PGSZ_MASK_HUGE      HV_PGSZ_MASK_64K
+#elif defined(CONFIG_HUGETLB_PAGE_SIZE_512K)
+#define HV_PGSZ_IDX_HUGE       HV_PGSZ_IDX_512K
+#define HV_PGSZ_MASK_HUGE      HV_PGSZ_MASK_512K
+#elif defined(CONFIG_HUGETLB_PAGE_SIZE_4MB)
+#define HV_PGSZ_IDX_HUGE       HV_PGSZ_IDX_4MB
+#define HV_PGSZ_MASK_HUGE      HV_PGSZ_MASK_4MB
+#else
+#error Broken huge page size setting...
+#endif
+#endif
+
+static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
 {
        unsigned long tsb_reg, base, tsb_paddr;
        unsigned long page_sz, tte;
 
-       mm->context.tsb_nentries = tsb_bytes / sizeof(struct tsb);
+       mm->context.tsb_block[tsb_idx].tsb_nentries =
+               tsb_bytes / sizeof(struct tsb);
 
        base = TSBMAP_BASE;
        tte = pgprot_val(PAGE_KERNEL_LOCKED);
-       tsb_paddr = __pa(mm->context.tsb);
+       tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
        BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
 
        /* Use the smallest page size that can map the whole TSB
@@ -147,68 +190,56 @@ static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_bytes)
                /* Physical mapping, no locked TLB entry for TSB.  */
                tsb_reg |= tsb_paddr;
 
-               mm->context.tsb_reg_val = tsb_reg;
-               mm->context.tsb_map_vaddr = 0;
-               mm->context.tsb_map_pte = 0;
+               mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
+               mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
+               mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
        } else {
                tsb_reg |= base;
                tsb_reg |= (tsb_paddr & (page_sz - 1UL));
                tte |= (tsb_paddr & ~(page_sz - 1UL));
 
-               mm->context.tsb_reg_val = tsb_reg;
-               mm->context.tsb_map_vaddr = base;
-               mm->context.tsb_map_pte = tte;
+               mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
+               mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
+               mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
        }
 
        /* Setup the Hypervisor TSB descriptor.  */
        if (tlb_type == hypervisor) {
-               struct hv_tsb_descr *hp = &mm->context.tsb_descr;
+               struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
 
-               switch (PAGE_SIZE) {
-               case 8192:
-               default:
-                       hp->pgsz_idx = HV_PGSZ_IDX_8K;
+               switch (tsb_idx) {
+               case MM_TSB_BASE:
+                       hp->pgsz_idx = HV_PGSZ_IDX_BASE;
                        break;
-
-               case 64 * 1024:
-                       hp->pgsz_idx = HV_PGSZ_IDX_64K;
-                       break;
-
-               case 512 * 1024:
-                       hp->pgsz_idx = HV_PGSZ_IDX_512K;
-                       break;
-
-               case 4 * 1024 * 1024:
-                       hp->pgsz_idx = HV_PGSZ_IDX_4MB;
+#ifdef CONFIG_HUGETLB_PAGE
+               case MM_TSB_HUGE:
+                       hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
                        break;
+#endif
+               default:
+                       BUG();
                };
                hp->assoc = 1;
                hp->num_ttes = tsb_bytes / 16;
                hp->ctx_idx = 0;
-               switch (PAGE_SIZE) {
-               case 8192:
-               default:
-                       hp->pgsz_mask = HV_PGSZ_MASK_8K;
-                       break;
-
-               case 64 * 1024:
-                       hp->pgsz_mask = HV_PGSZ_MASK_64K;
-                       break;
-
-               case 512 * 1024:
-                       hp->pgsz_mask = HV_PGSZ_MASK_512K;
+               switch (tsb_idx) {
+               case MM_TSB_BASE:
+                       hp->pgsz_mask = HV_PGSZ_MASK_BASE;
                        break;
-
-               case 4 * 1024 * 1024:
-                       hp->pgsz_mask = HV_PGSZ_MASK_4MB;
+#ifdef CONFIG_HUGETLB_PAGE
+               case MM_TSB_HUGE:
+                       hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
                        break;
+#endif
+               default:
+                       BUG();
                };
                hp->tsb_base = tsb_paddr;
                hp->resv = 0;
        }
 }
 
-static kmem_cache_t *tsb_caches[8] __read_mostly;
+static struct kmem_cache *tsb_caches[8] __read_mostly;
 
 static const char *tsb_cache_names[8] = {
        "tsb_8KB",
@@ -221,7 +252,7 @@ static const char *tsb_cache_names[8] = {
        "tsb_1MB",
 };
 
-void __init tsb_cache_init(void)
+void __init pgtable_cache_init(void)
 {
        unsigned long i;
 
@@ -231,8 +262,7 @@ void __init tsb_cache_init(void)
 
                tsb_caches[i] = kmem_cache_create(name,
                                                  size, size,
-                                                 SLAB_HWCACHE_ALIGN |
-                                                 SLAB_MUST_HWCACHE_ALIGN,
+                                                 0,
                                                  NULL, NULL);
                if (!tsb_caches[i]) {
                        prom_printf("Could not create %s cache\n", name);
@@ -241,11 +271,11 @@ void __init tsb_cache_init(void)
        }
 }
 
-/* When the RSS of an address space exceeds mm->context.tsb_rss_limit,
- * do_sparc64_fault() invokes this routine to try and grow the TSB.
+/* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
+ * do_sparc64_fault() invokes this routine to try and grow it.
  *
  * When we reach the maximum TSB size supported, we stick ~0UL into
- * mm->context.tsb_rss_limit so the grow checks in update_mmu_cache()
+ * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
  * will not trigger any longer.
  *
  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
@@ -257,7 +287,7 @@ void __init tsb_cache_init(void)
  * the number of entries that the current TSB can hold at once.  Currently,
  * we trigger when the RSS hits 3/4 of the TSB capacity.
  */
-void tsb_grow(struct mm_struct *mm, unsigned long rss)
+void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
 {
        unsigned long max_tsb_size = 1 * 1024 * 1024;
        unsigned long new_size, old_size, flags;
@@ -297,7 +327,8 @@ retry_tsb_alloc:
                 * down to a 0-order allocation and force no TSB
                 * growing for this address space.
                 */
-               if (mm->context.tsb == NULL && new_cache_index > 0) {
+               if (mm->context.tsb_block[tsb_index].tsb == NULL &&
+                   new_cache_index > 0) {
                        new_cache_index = 0;
                        new_size = 8192;
                        new_rss_limit = ~0UL;
@@ -307,13 +338,13 @@ retry_tsb_alloc:
                /* If we failed on a TSB grow, we are under serious
                 * memory pressure so don't try to grow any more.
                 */
-               if (mm->context.tsb != NULL)
-                       mm->context.tsb_rss_limit = ~0UL;
+               if (mm->context.tsb_block[tsb_index].tsb != NULL)
+                       mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
                return;
        }
 
        /* Mark all tags as invalid.  */
-       memset(new_tsb, 0x40, new_size);
+       tsb_init(new_tsb, new_size);
 
        /* Ok, we are about to commit the changes.  If we are
         * growing an existing TSB the locking is very tricky,
@@ -339,23 +370,26 @@ retry_tsb_alloc:
         */
        spin_lock_irqsave(&mm->context.lock, flags);
 
-       old_tsb = mm->context.tsb;
-       old_cache_index = (mm->context.tsb_reg_val & 0x7UL);
-       old_size = mm->context.tsb_nentries * sizeof(struct tsb);
+       old_tsb = mm->context.tsb_block[tsb_index].tsb;
+       old_cache_index =
+               (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
+       old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
+                   sizeof(struct tsb));
 
 
        /* Handle multiple threads trying to grow the TSB at the same time.
         * One will get in here first, and bump the size and the RSS limit.
         * The others will get in here next and hit this check.
         */
-       if (unlikely(old_tsb && (rss < mm->context.tsb_rss_limit))) {
+       if (unlikely(old_tsb &&
+                    (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
                spin_unlock_irqrestore(&mm->context.lock, flags);
 
                kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
                return;
        }
 
-       mm->context.tsb_rss_limit = new_rss_limit;
+       mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
 
        if (old_tsb) {
                extern void copy_tsb(unsigned long old_tsb_base,
@@ -372,8 +406,8 @@ retry_tsb_alloc:
                copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
        }
 
-       mm->context.tsb = new_tsb;
-       setup_tsb_params(mm, new_size);
+       mm->context.tsb_block[tsb_index].tsb = new_tsb;
+       setup_tsb_params(mm, tsb_index, new_size);
 
        spin_unlock_irqrestore(&mm->context.lock, flags);
 
@@ -394,40 +428,65 @@ retry_tsb_alloc:
 
 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 {
+#ifdef CONFIG_HUGETLB_PAGE
+       unsigned long huge_pte_count;
+#endif
+       unsigned int i;
+
        spin_lock_init(&mm->context.lock);
 
        mm->context.sparc64_ctx_val = 0UL;
 
+#ifdef CONFIG_HUGETLB_PAGE
+       /* We reset it to zero because the fork() page copying
+        * will re-increment the counters as the parent PTEs are
+        * copied into the child address space.
+        */
+       huge_pte_count = mm->context.huge_pte_count;
+       mm->context.huge_pte_count = 0;
+#endif
+
        /* copy_mm() copies over the parent's mm_struct before calling
         * us, so we need to zero out the TSB pointer or else tsb_grow()
         * will be confused and think there is an older TSB to free up.
         */
-       mm->context.tsb = NULL;
+       for (i = 0; i < MM_NUM_TSBS; i++)
+               mm->context.tsb_block[i].tsb = NULL;
 
        /* If this is fork, inherit the parent's TSB size.  We would
         * grow it to that size on the first page fault anyways.
         */
-       tsb_grow(mm, get_mm_rss(mm));
+       tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
 
-       if (unlikely(!mm->context.tsb))
+#ifdef CONFIG_HUGETLB_PAGE
+       if (unlikely(huge_pte_count))
+               tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
+#endif
+
+       if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
                return -ENOMEM;
 
        return 0;
 }
 
-void destroy_context(struct mm_struct *mm)
+static void tsb_destroy_one(struct tsb_config *tp)
 {
-       unsigned long flags, cache_index;
+       unsigned long cache_index;
 
-       cache_index = (mm->context.tsb_reg_val & 0x7UL);
-       kmem_cache_free(tsb_caches[cache_index], mm->context.tsb);
+       if (!tp->tsb)
+               return;
+       cache_index = tp->tsb_reg_val & 0x7UL;
+       kmem_cache_free(tsb_caches[cache_index], tp->tsb);
+       tp->tsb = NULL;
+       tp->tsb_reg_val = 0UL;
+}
 
-       /* We can remove these later, but for now it's useful
-        * to catch any bogus post-destroy_context() references
-        * to the TSB.
-        */
-       mm->context.tsb = NULL;
-       mm->context.tsb_reg_val = 0UL;
+void destroy_context(struct mm_struct *mm)
+{
+       unsigned long flags, i;
+
+       for (i = 0; i < MM_NUM_TSBS; i++)
+               tsb_destroy_one(&mm->context.tsb_block[i]);
 
        spin_lock_irqsave(&ctx_alloc_lock, flags);