arch/ia64/mm/tlb.c

   1 /*
   2  * TLB support routines.
   3  *
   4  * Copyright (C) 1998-2001, 2003 Hewlett-Packard Co
   5  *      David Mosberger-Tang <davidm@hpl.hp.com>
   6  *
   7  * 08/02/00 A. Mallick <asit.k.mallick@intel.com>
   8  *              Modified RID allocation for SMP
   9  *          Goutham Rao <goutham.rao@intel.com>
  10  *              IPI based ptc implementation and A-step IPI implementation.
  11  * Rohit Seth <rohit.seth@intel.com>
  12  * Ken Chen <kenneth.w.chen@intel.com>
  13  * Christophe de Dinechin <ddd@hp.com>: Avoid ptc.e on memory allocation
  14  */
  15 #include <linux/module.h>
  16 #include <linux/init.h>
  17 #include <linux/kernel.h>
  18 #include <linux/sched.h>
  19 #include <linux/smp.h>
  20 #include <linux/mm.h>
  21 #include <linux/bootmem.h>
  22
  23 #include <asm/delay.h>
  24 #include <asm/mmu_context.h>
  25 #include <asm/pgalloc.h>
  26 #include <asm/pal.h>
  27 #include <asm/tlbflush.h>
  28 #include <asm/dma.h>
  29
  30 static struct {
  31         unsigned long mask;     /* mask of supported purge page-sizes */
  32         unsigned long max_bits; /* log2 of largest supported purge page-size */
  33 } purge;
  34
  35 struct ia64_ctx ia64_ctx = {
  36         .lock = __SPIN_LOCK_UNLOCKED(ia64_ctx.lock),
  37         .next = 1,
  38         .max_ctx = ~0U
  39 };
  40
  41 DEFINE_PER_CPU(u8, ia64_need_tlb_flush);
  42
  43 /*
  44  * Initializes the ia64_ctx.bitmap array based on max_ctx+1.
  45  * Called after cpu_init() has setup ia64_ctx.max_ctx based on
  46  * maximum RID that is supported by boot CPU.
  47  */
  48 void __init
  49 mmu_context_init (void)
  50 {
  51         ia64_ctx.bitmap = alloc_bootmem((ia64_ctx.max_ctx+1)>>3);
  52         ia64_ctx.flushmap = alloc_bootmem((ia64_ctx.max_ctx+1)>>3);
  53 }
  54
  55 /*
  56  * Acquire the ia64_ctx.lock before calling this function!
  57  */
  58 void
  59 wrap_mmu_context (struct mm_struct *mm)
  60 {
  61         int i, cpu;
  62         unsigned long flush_bit;
  63
  64         for (i=0; i <= ia64_ctx.max_ctx / BITS_PER_LONG; i++) {
  65                 flush_bit = xchg(&ia64_ctx.flushmap[i], 0);
  66                 ia64_ctx.bitmap[i] ^= flush_bit;
  67         }
  68
  69         /* use offset at 300 to skip daemons */
  70         ia64_ctx.next = find_next_zero_bit(ia64_ctx.bitmap,
  71                                 ia64_ctx.max_ctx, 300);
  72         ia64_ctx.limit = find_next_bit(ia64_ctx.bitmap,
  73                                 ia64_ctx.max_ctx, ia64_ctx.next);
  74
  75         /*
  76          * can't call flush_tlb_all() here because of race condition
  77          * with O(1) scheduler [EF]
  78          */
  79         cpu = get_cpu(); /* prevent preemption/migration */
  80         for_each_online_cpu(i)
  81                 if (i != cpu)
  82                         per_cpu(ia64_need_tlb_flush, i) = 1;
  83         put_cpu();
  84         local_flush_tlb_all();
  85 }
  86
  87 void
  88 ia64_global_tlb_purge (struct mm_struct *mm, unsigned long start,
  89                        unsigned long end, unsigned long nbits)
  90 {
  91         static DEFINE_SPINLOCK(ptcg_lock);
  92
  93         struct mm_struct *active_mm = current->active_mm;
  94
  95         if (mm != active_mm) {
  96                 /* Restore region IDs for mm */
  97                 if (mm && active_mm) {
  98                         activate_context(mm);
  99                 } else {
 100                         flush_tlb_all();
 101                         return;
 102                 }
 103         }
 104
 105         /* HW requires global serialization of ptc.ga.  */
 106         spin_lock(&ptcg_lock);
 107         {
 108                 do {
 109                         /*
 110                          * Flush ALAT entries also.
 111                          */
 112                         ia64_ptcga(start, (nbits<<2));
 113                         ia64_srlz_i();
 114                         start += (1UL << nbits);
 115                 } while (start < end);
 116         }
 117         spin_unlock(&ptcg_lock);
 118
 119         if (mm != active_mm) {
 120                 activate_context(active_mm);
 121         }
 122 }
 123
 124 void
 125 local_flush_tlb_all (void)
 126 {
 127         unsigned long i, j, flags, count0, count1, stride0, stride1, addr;
 128
 129         addr    = local_cpu_data->ptce_base;
 130         count0  = local_cpu_data->ptce_count[0];
 131         count1  = local_cpu_data->ptce_count[1];
 132         stride0 = local_cpu_data->ptce_stride[0];
 133         stride1 = local_cpu_data->ptce_stride[1];
 134
 135         local_irq_save(flags);
 136         for (i = 0; i < count0; ++i) {
 137                 for (j = 0; j < count1; ++j) {
 138                         ia64_ptce(addr);
 139                         addr += stride1;
 140                 }
 141                 addr += stride0;
 142         }
 143         local_irq_restore(flags);
 144         ia64_srlz_i();                  /* srlz.i implies srlz.d */
 145 }
 146
 147 void
 148 flush_tlb_range (struct vm_area_struct *vma, unsigned long start,
 149                  unsigned long end)
 150 {
 151         struct mm_struct *mm = vma->vm_mm;
 152         unsigned long size = end - start;
 153         unsigned long nbits;
 154
 155 #ifndef CONFIG_SMP
 156         if (mm != current->active_mm) {
 157                 mm->context = 0;
 158                 return;
 159         }
 160 #endif
 161
 162         nbits = ia64_fls(size + 0xfff);
 163         while (unlikely (((1UL << nbits) & purge.mask) == 0) &&
 164                         (nbits < purge.max_bits))
 165                 ++nbits;
 166         if (nbits > purge.max_bits)
 167                 nbits = purge.max_bits;
 168         start &= ~((1UL << nbits) - 1);
 169
 170         preempt_disable();
 171 #ifdef CONFIG_SMP
 172         if (mm != current->active_mm || cpus_weight(mm->cpu_vm_mask) != 1) {
 173                 platform_global_tlb_purge(mm, start, end, nbits);
 174                 preempt_enable();
 175                 return;
 176         }
 177 #endif
 178         do {
 179                 ia64_ptcl(start, (nbits<<2));
 180                 start += (1UL << nbits);
 181         } while (start < end);
 182         preempt_enable();
 183         ia64_srlz_i();                  /* srlz.i implies srlz.d */
 184 }
 185 EXPORT_SYMBOL(flush_tlb_range);
 186
 187 void __devinit
 188 ia64_tlb_init (void)
 189 {
 190         ia64_ptce_info_t uninitialized_var(ptce_info); /* GCC be quiet */
 191         unsigned long tr_pgbits;
 192         long status;
 193
 194         if ((status = ia64_pal_vm_page_size(&tr_pgbits, &purge.mask)) != 0) {
 195                 printk(KERN_ERR "PAL_VM_PAGE_SIZE failed with status=%ld; "
 196                        "defaulting to architected purge page-sizes.\n", status);
 197                 purge.mask = 0x115557000UL;
 198         }
 199         purge.max_bits = ia64_fls(purge.mask);
 200
 201         ia64_get_ptce(&ptce_info);
 202         local_cpu_data->ptce_base = ptce_info.base;
 203         local_cpu_data->ptce_count[0] = ptce_info.count[0];
 204         local_cpu_data->ptce_count[1] = ptce_info.count[1];
 205         local_cpu_data->ptce_stride[0] = ptce_info.stride[0];
 206         local_cpu_data->ptce_stride[1] = ptce_info.stride[1];
 207
 208         local_flush_tlb_all();  /* nuke left overs from bootstrapping... */
 209 }