[IA64] Use bitmaps for efficient context allocation/free
[linux-2.6] / 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  */
14 #include <linux/config.h>
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,
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;
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         /* can't call flush_tlb_all() here because of race condition with O(1) scheduler [EF] */
76         {
77                 int cpu = get_cpu(); /* prevent preemption/migration */
78                 for_each_online_cpu(i) {
79                         if (i != cpu)
80                                 per_cpu(ia64_need_tlb_flush, i) = 1;
81                 }
82                 put_cpu();
83         }
84         local_flush_tlb_all();
85 }
86
87 void
88 ia64_global_tlb_purge (struct mm_struct *mm, unsigned long start, unsigned long end, unsigned long nbits)
89 {
90         static DEFINE_SPINLOCK(ptcg_lock);
91
92         if (mm != current->active_mm) {
93                 flush_tlb_all();
94                 return;
95         }
96
97         /* HW requires global serialization of ptc.ga.  */
98         spin_lock(&ptcg_lock);
99         {
100                 do {
101                         /*
102                          * Flush ALAT entries also.
103                          */
104                         ia64_ptcga(start, (nbits<<2));
105                         ia64_srlz_i();
106                         start += (1UL << nbits);
107                 } while (start < end);
108         }
109         spin_unlock(&ptcg_lock);
110 }
111
112 void
113 local_flush_tlb_all (void)
114 {
115         unsigned long i, j, flags, count0, count1, stride0, stride1, addr;
116
117         addr    = local_cpu_data->ptce_base;
118         count0  = local_cpu_data->ptce_count[0];
119         count1  = local_cpu_data->ptce_count[1];
120         stride0 = local_cpu_data->ptce_stride[0];
121         stride1 = local_cpu_data->ptce_stride[1];
122
123         local_irq_save(flags);
124         for (i = 0; i < count0; ++i) {
125                 for (j = 0; j < count1; ++j) {
126                         ia64_ptce(addr);
127                         addr += stride1;
128                 }
129                 addr += stride0;
130         }
131         local_irq_restore(flags);
132         ia64_srlz_i();                  /* srlz.i implies srlz.d */
133 }
134
135 void
136 flush_tlb_range (struct vm_area_struct *vma, unsigned long start, unsigned long end)
137 {
138         struct mm_struct *mm = vma->vm_mm;
139         unsigned long size = end - start;
140         unsigned long nbits;
141
142 #ifndef CONFIG_SMP
143         if (mm != current->active_mm) {
144                 mm->context = 0;
145                 return;
146         }
147 #endif
148
149         nbits = ia64_fls(size + 0xfff);
150         while (unlikely (((1UL << nbits) & purge.mask) == 0) && (nbits < purge.max_bits))
151                 ++nbits;
152         if (nbits > purge.max_bits)
153                 nbits = purge.max_bits;
154         start &= ~((1UL << nbits) - 1);
155
156 # ifdef CONFIG_SMP
157         platform_global_tlb_purge(mm, start, end, nbits);
158 # else
159         preempt_disable();
160         do {
161                 ia64_ptcl(start, (nbits<<2));
162                 start += (1UL << nbits);
163         } while (start < end);
164         preempt_enable();
165 # endif
166
167         ia64_srlz_i();                  /* srlz.i implies srlz.d */
168 }
169 EXPORT_SYMBOL(flush_tlb_range);
170
171 void __devinit
172 ia64_tlb_init (void)
173 {
174         ia64_ptce_info_t ptce_info;
175         unsigned long tr_pgbits;
176         long status;
177
178         if ((status = ia64_pal_vm_page_size(&tr_pgbits, &purge.mask)) != 0) {
179                 printk(KERN_ERR "PAL_VM_PAGE_SIZE failed with status=%ld;"
180                        "defaulting to architected purge page-sizes.\n", status);
181                 purge.mask = 0x115557000UL;
182         }
183         purge.max_bits = ia64_fls(purge.mask);
184
185         ia64_get_ptce(&ptce_info);
186         local_cpu_data->ptce_base = ptce_info.base;
187         local_cpu_data->ptce_count[0] = ptce_info.count[0];
188         local_cpu_data->ptce_count[1] = ptce_info.count[1];
189         local_cpu_data->ptce_stride[0] = ptce_info.stride[0];
190         local_cpu_data->ptce_stride[1] = ptce_info.stride[1];
191
192         local_flush_tlb_all();          /* nuke left overs from bootstrapping... */
193 }