IRQ: Maintain regs pointer globally rather than passing to IRQ handlers
[linux-2.6.git] / arch / i386 / kernel / irq.c
1 /*
2  *      linux/arch/i386/kernel/irq.c
3  *
4  *      Copyright (C) 1992, 1998 Linus Torvalds, Ingo Molnar
5  *
6  * This file contains the lowest level x86-specific interrupt
7  * entry, irq-stacks and irq statistics code. All the remaining
8  * irq logic is done by the generic kernel/irq/ code and
9  * by the x86-specific irq controller code. (e.g. i8259.c and
10  * io_apic.c.)
11  */
12
13 #include <asm/uaccess.h>
14 #include <linux/module.h>
15 #include <linux/seq_file.h>
16 #include <linux/interrupt.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/notifier.h>
19 #include <linux/cpu.h>
20 #include <linux/delay.h>
21
22 DEFINE_PER_CPU(irq_cpustat_t, irq_stat) ____cacheline_internodealigned_in_smp;
23 EXPORT_PER_CPU_SYMBOL(irq_stat);
24
25 #ifndef CONFIG_X86_LOCAL_APIC
26 /*
27  * 'what should we do if we get a hw irq event on an illegal vector'.
28  * each architecture has to answer this themselves.
29  */
30 void ack_bad_irq(unsigned int irq)
31 {
32         printk("unexpected IRQ trap at vector %02x\n", irq);
33 }
34 #endif
35
36 #ifdef CONFIG_4KSTACKS
37 /*
38  * per-CPU IRQ handling contexts (thread information and stack)
39  */
40 union irq_ctx {
41         struct thread_info      tinfo;
42         u32                     stack[THREAD_SIZE/sizeof(u32)];
43 };
44
45 static union irq_ctx *hardirq_ctx[NR_CPUS] __read_mostly;
46 static union irq_ctx *softirq_ctx[NR_CPUS] __read_mostly;
47 #endif
48
49 /*
50  * do_IRQ handles all normal device IRQ's (the special
51  * SMP cross-CPU interrupts have their own specific
52  * handlers).
53  */
54 fastcall unsigned int do_IRQ(struct pt_regs *regs)
55 {       
56         struct pt_regs *old_regs;
57         /* high bit used in ret_from_ code */
58         int irq = ~regs->orig_eax;
59         struct irq_desc *desc = irq_desc + irq;
60 #ifdef CONFIG_4KSTACKS
61         union irq_ctx *curctx, *irqctx;
62         u32 *isp;
63 #endif
64
65         if (unlikely((unsigned)irq >= NR_IRQS)) {
66                 printk(KERN_EMERG "%s: cannot handle IRQ %d\n",
67                                         __FUNCTION__, irq);
68                 BUG();
69         }
70
71         old_regs = set_irq_regs(regs);
72         irq_enter();
73 #ifdef CONFIG_DEBUG_STACKOVERFLOW
74         /* Debugging check for stack overflow: is there less than 1KB free? */
75         {
76                 long esp;
77
78                 __asm__ __volatile__("andl %%esp,%0" :
79                                         "=r" (esp) : "0" (THREAD_SIZE - 1));
80                 if (unlikely(esp < (sizeof(struct thread_info) + STACK_WARN))) {
81                         printk("do_IRQ: stack overflow: %ld\n",
82                                 esp - sizeof(struct thread_info));
83                         dump_stack();
84                 }
85         }
86 #endif
87
88 #ifdef CONFIG_4KSTACKS
89
90         curctx = (union irq_ctx *) current_thread_info();
91         irqctx = hardirq_ctx[smp_processor_id()];
92
93         /*
94          * this is where we switch to the IRQ stack. However, if we are
95          * already using the IRQ stack (because we interrupted a hardirq
96          * handler) we can't do that and just have to keep using the
97          * current stack (which is the irq stack already after all)
98          */
99         if (curctx != irqctx) {
100                 int arg1, arg2, ebx;
101
102                 /* build the stack frame on the IRQ stack */
103                 isp = (u32*) ((char*)irqctx + sizeof(*irqctx));
104                 irqctx->tinfo.task = curctx->tinfo.task;
105                 irqctx->tinfo.previous_esp = current_stack_pointer;
106
107                 /*
108                  * Copy the softirq bits in preempt_count so that the
109                  * softirq checks work in the hardirq context.
110                  */
111                 irqctx->tinfo.preempt_count =
112                         (irqctx->tinfo.preempt_count & ~SOFTIRQ_MASK) |
113                         (curctx->tinfo.preempt_count & SOFTIRQ_MASK);
114
115                 asm volatile(
116                         "       xchgl  %%ebx,%%esp      \n"
117                         "       call   *%%edi           \n"
118                         "       movl   %%ebx,%%esp      \n"
119                         : "=a" (arg1), "=d" (arg2), "=b" (ebx)
120                         :  "0" (irq),   "1" (desc),  "2" (isp),
121                            "D" (desc->handle_irq)
122                         : "memory", "cc"
123                 );
124         } else
125 #endif
126                 desc->handle_irq(irq, desc);
127
128         irq_exit();
129         set_irq_regs(old_regs);
130         return 1;
131 }
132
133 #ifdef CONFIG_4KSTACKS
134
135 /*
136  * These should really be __section__(".bss.page_aligned") as well, but
137  * gcc's 3.0 and earlier don't handle that correctly.
138  */
139 static char softirq_stack[NR_CPUS * THREAD_SIZE]
140                 __attribute__((__aligned__(THREAD_SIZE)));
141
142 static char hardirq_stack[NR_CPUS * THREAD_SIZE]
143                 __attribute__((__aligned__(THREAD_SIZE)));
144
145 /*
146  * allocate per-cpu stacks for hardirq and for softirq processing
147  */
148 void irq_ctx_init(int cpu)
149 {
150         union irq_ctx *irqctx;
151
152         if (hardirq_ctx[cpu])
153                 return;
154
155         irqctx = (union irq_ctx*) &hardirq_stack[cpu*THREAD_SIZE];
156         irqctx->tinfo.task              = NULL;
157         irqctx->tinfo.exec_domain       = NULL;
158         irqctx->tinfo.cpu               = cpu;
159         irqctx->tinfo.preempt_count     = HARDIRQ_OFFSET;
160         irqctx->tinfo.addr_limit        = MAKE_MM_SEG(0);
161
162         hardirq_ctx[cpu] = irqctx;
163
164         irqctx = (union irq_ctx*) &softirq_stack[cpu*THREAD_SIZE];
165         irqctx->tinfo.task              = NULL;
166         irqctx->tinfo.exec_domain       = NULL;
167         irqctx->tinfo.cpu               = cpu;
168         irqctx->tinfo.preempt_count     = 0;
169         irqctx->tinfo.addr_limit        = MAKE_MM_SEG(0);
170
171         softirq_ctx[cpu] = irqctx;
172
173         printk("CPU %u irqstacks, hard=%p soft=%p\n",
174                 cpu,hardirq_ctx[cpu],softirq_ctx[cpu]);
175 }
176
177 void irq_ctx_exit(int cpu)
178 {
179         hardirq_ctx[cpu] = NULL;
180 }
181
182 extern asmlinkage void __do_softirq(void);
183
184 asmlinkage void do_softirq(void)
185 {
186         unsigned long flags;
187         struct thread_info *curctx;
188         union irq_ctx *irqctx;
189         u32 *isp;
190
191         if (in_interrupt())
192                 return;
193
194         local_irq_save(flags);
195
196         if (local_softirq_pending()) {
197                 curctx = current_thread_info();
198                 irqctx = softirq_ctx[smp_processor_id()];
199                 irqctx->tinfo.task = curctx->task;
200                 irqctx->tinfo.previous_esp = current_stack_pointer;
201
202                 /* build the stack frame on the softirq stack */
203                 isp = (u32*) ((char*)irqctx + sizeof(*irqctx));
204
205                 asm volatile(
206                         "       xchgl   %%ebx,%%esp     \n"
207                         "       call    __do_softirq    \n"
208                         "       movl    %%ebx,%%esp     \n"
209                         : "=b"(isp)
210                         : "0"(isp)
211                         : "memory", "cc", "edx", "ecx", "eax"
212                 );
213                 /*
214                  * Shouldnt happen, we returned above if in_interrupt():
215                  */
216                 WARN_ON_ONCE(softirq_count());
217         }
218
219         local_irq_restore(flags);
220 }
221
222 EXPORT_SYMBOL(do_softirq);
223 #endif
224
225 /*
226  * Interrupt statistics:
227  */
228
229 atomic_t irq_err_count;
230
231 /*
232  * /proc/interrupts printing:
233  */
234
235 int show_interrupts(struct seq_file *p, void *v)
236 {
237         int i = *(loff_t *) v, j;
238         struct irqaction * action;
239         unsigned long flags;
240
241         if (i == 0) {
242                 seq_printf(p, "           ");
243                 for_each_online_cpu(j)
244                         seq_printf(p, "CPU%-8d",j);
245                 seq_putc(p, '\n');
246         }
247
248         if (i < NR_IRQS) {
249                 spin_lock_irqsave(&irq_desc[i].lock, flags);
250                 action = irq_desc[i].action;
251                 if (!action)
252                         goto skip;
253                 seq_printf(p, "%3d: ",i);
254 #ifndef CONFIG_SMP
255                 seq_printf(p, "%10u ", kstat_irqs(i));
256 #else
257                 for_each_online_cpu(j)
258                         seq_printf(p, "%10u ", kstat_cpu(j).irqs[i]);
259 #endif
260                 seq_printf(p, " %8s", irq_desc[i].chip->name);
261                 seq_printf(p, "-%s", handle_irq_name(irq_desc[i].handle_irq));
262                 seq_printf(p, "  %s", action->name);
263
264                 for (action=action->next; action; action = action->next)
265                         seq_printf(p, ", %s", action->name);
266
267                 seq_putc(p, '\n');
268 skip:
269                 spin_unlock_irqrestore(&irq_desc[i].lock, flags);
270         } else if (i == NR_IRQS) {
271                 seq_printf(p, "NMI: ");
272                 for_each_online_cpu(j)
273                         seq_printf(p, "%10u ", nmi_count(j));
274                 seq_putc(p, '\n');
275 #ifdef CONFIG_X86_LOCAL_APIC
276                 seq_printf(p, "LOC: ");
277                 for_each_online_cpu(j)
278                         seq_printf(p, "%10u ",
279                                 per_cpu(irq_stat,j).apic_timer_irqs);
280                 seq_putc(p, '\n');
281 #endif
282                 seq_printf(p, "ERR: %10u\n", atomic_read(&irq_err_count));
283 #if defined(CONFIG_X86_IO_APIC)
284                 seq_printf(p, "MIS: %10u\n", atomic_read(&irq_mis_count));
285 #endif
286         }
287         return 0;
288 }
289
290 #ifdef CONFIG_HOTPLUG_CPU
291 #include <mach_apic.h>
292
293 void fixup_irqs(cpumask_t map)
294 {
295         unsigned int irq;
296         static int warned;
297
298         for (irq = 0; irq < NR_IRQS; irq++) {
299                 cpumask_t mask;
300                 if (irq == 2)
301                         continue;
302
303                 cpus_and(mask, irq_desc[irq].affinity, map);
304                 if (any_online_cpu(mask) == NR_CPUS) {
305                         printk("Breaking affinity for irq %i\n", irq);
306                         mask = map;
307                 }
308                 if (irq_desc[irq].chip->set_affinity)
309                         irq_desc[irq].chip->set_affinity(irq, mask);
310                 else if (irq_desc[irq].action && !(warned++))
311                         printk("Cannot set affinity for irq %i\n", irq);
312         }
313
314 #if 0
315         barrier();
316         /* Ingo Molnar says: "after the IO-APIC masks have been redirected
317            [note the nop - the interrupt-enable boundary on x86 is two
318            instructions from sti] - to flush out pending hardirqs and
319            IPIs. After this point nothing is supposed to reach this CPU." */
320         __asm__ __volatile__("sti; nop; cli");
321         barrier();
322 #else
323         /* That doesn't seem sufficient.  Give it 1ms. */
324         local_irq_enable();
325         mdelay(1);
326         local_irq_disable();
327 #endif
328 }
329 #endif
330