OpenCores
URL https://opencores.org/ocsvn/test_project/test_project/trunk

Subversion Repositories test_project

[/] [test_project/] [trunk/] [linux_sd_driver/] [drivers/] [lguest/] [interrupts_and_traps.c] - Blame information for rev 65

Go to most recent revision | Details | Compare with Previous | View Log

Line No. Rev Author Line
1 62 marcus.erl
/*P:800 Interrupts (traps) are complicated enough to earn their own file.
2
 * There are three classes of interrupts:
3
 *
4
 * 1) Real hardware interrupts which occur while we're running the Guest,
5
 * 2) Interrupts for virtual devices attached to the Guest, and
6
 * 3) Traps and faults from the Guest.
7
 *
8
 * Real hardware interrupts must be delivered to the Host, not the Guest.
9
 * Virtual interrupts must be delivered to the Guest, but we make them look
10
 * just like real hardware would deliver them.  Traps from the Guest can be set
11
 * up to go directly back into the Guest, but sometimes the Host wants to see
12
 * them first, so we also have a way of "reflecting" them into the Guest as if
13
 * they had been delivered to it directly. :*/
14
#include <linux/uaccess.h>
15
#include <linux/interrupt.h>
16
#include <linux/module.h>
17
#include "lg.h"
18
 
19
/* Allow Guests to use a non-128 (ie. non-Linux) syscall trap. */
20
static unsigned int syscall_vector = SYSCALL_VECTOR;
21
module_param(syscall_vector, uint, 0444);
22
 
23
/* The address of the interrupt handler is split into two bits: */
24
static unsigned long idt_address(u32 lo, u32 hi)
25
{
26
        return (lo & 0x0000FFFF) | (hi & 0xFFFF0000);
27
}
28
 
29
/* The "type" of the interrupt handler is a 4 bit field: we only support a
30
 * couple of types. */
31
static int idt_type(u32 lo, u32 hi)
32
{
33
        return (hi >> 8) & 0xF;
34
}
35
 
36
/* An IDT entry can't be used unless the "present" bit is set. */
37
static int idt_present(u32 lo, u32 hi)
38
{
39
        return (hi & 0x8000);
40
}
41
 
42
/* We need a helper to "push" a value onto the Guest's stack, since that's a
43
 * big part of what delivering an interrupt does. */
44
static void push_guest_stack(struct lguest *lg, unsigned long *gstack, u32 val)
45
{
46
        /* Stack grows upwards: move stack then write value. */
47
        *gstack -= 4;
48
        lgwrite(lg, *gstack, u32, val);
49
}
50
 
51
/*H:210 The set_guest_interrupt() routine actually delivers the interrupt or
52
 * trap.  The mechanics of delivering traps and interrupts to the Guest are the
53
 * same, except some traps have an "error code" which gets pushed onto the
54
 * stack as well: the caller tells us if this is one.
55
 *
56
 * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this
57
 * interrupt or trap.  It's split into two parts for traditional reasons: gcc
58
 * on i386 used to be frightened by 64 bit numbers.
59
 *
60
 * We set up the stack just like the CPU does for a real interrupt, so it's
61
 * identical for the Guest (and the standard "iret" instruction will undo
62
 * it). */
63
static void set_guest_interrupt(struct lguest *lg, u32 lo, u32 hi, int has_err)
64
{
65
        unsigned long gstack, origstack;
66
        u32 eflags, ss, irq_enable;
67
        unsigned long virtstack;
68
 
69
        /* There are two cases for interrupts: one where the Guest is already
70
         * in the kernel, and a more complex one where the Guest is in
71
         * userspace.  We check the privilege level to find out. */
72
        if ((lg->regs->ss&0x3) != GUEST_PL) {
73
                /* The Guest told us their kernel stack with the SET_STACK
74
                 * hypercall: both the virtual address and the segment */
75
                virtstack = lg->esp1;
76
                ss = lg->ss1;
77
 
78
                origstack = gstack = guest_pa(lg, virtstack);
79
                /* We push the old stack segment and pointer onto the new
80
                 * stack: when the Guest does an "iret" back from the interrupt
81
                 * handler the CPU will notice they're dropping privilege
82
                 * levels and expect these here. */
83
                push_guest_stack(lg, &gstack, lg->regs->ss);
84
                push_guest_stack(lg, &gstack, lg->regs->esp);
85
        } else {
86
                /* We're staying on the same Guest (kernel) stack. */
87
                virtstack = lg->regs->esp;
88
                ss = lg->regs->ss;
89
 
90
                origstack = gstack = guest_pa(lg, virtstack);
91
        }
92
 
93
        /* Remember that we never let the Guest actually disable interrupts, so
94
         * the "Interrupt Flag" bit is always set.  We copy that bit from the
95
         * Guest's "irq_enabled" field into the eflags word: we saw the Guest
96
         * copy it back in "lguest_iret". */
97
        eflags = lg->regs->eflags;
98
        if (get_user(irq_enable, &lg->lguest_data->irq_enabled) == 0
99
            && !(irq_enable & X86_EFLAGS_IF))
100
                eflags &= ~X86_EFLAGS_IF;
101
 
102
        /* An interrupt is expected to push three things on the stack: the old
103
         * "eflags" word, the old code segment, and the old instruction
104
         * pointer. */
105
        push_guest_stack(lg, &gstack, eflags);
106
        push_guest_stack(lg, &gstack, lg->regs->cs);
107
        push_guest_stack(lg, &gstack, lg->regs->eip);
108
 
109
        /* For the six traps which supply an error code, we push that, too. */
110
        if (has_err)
111
                push_guest_stack(lg, &gstack, lg->regs->errcode);
112
 
113
        /* Now we've pushed all the old state, we change the stack, the code
114
         * segment and the address to execute. */
115
        lg->regs->ss = ss;
116
        lg->regs->esp = virtstack + (gstack - origstack);
117
        lg->regs->cs = (__KERNEL_CS|GUEST_PL);
118
        lg->regs->eip = idt_address(lo, hi);
119
 
120
        /* There are two kinds of interrupt handlers: 0xE is an "interrupt
121
         * gate" which expects interrupts to be disabled on entry. */
122
        if (idt_type(lo, hi) == 0xE)
123
                if (put_user(0, &lg->lguest_data->irq_enabled))
124
                        kill_guest(lg, "Disabling interrupts");
125
}
126
 
127
/*H:205
128
 * Virtual Interrupts.
129
 *
130
 * maybe_do_interrupt() gets called before every entry to the Guest, to see if
131
 * we should divert the Guest to running an interrupt handler. */
132
void maybe_do_interrupt(struct lguest *lg)
133
{
134
        unsigned int irq;
135
        DECLARE_BITMAP(blk, LGUEST_IRQS);
136
        struct desc_struct *idt;
137
 
138
        /* If the Guest hasn't even initialized yet, we can do nothing. */
139
        if (!lg->lguest_data)
140
                return;
141
 
142
        /* Take our "irqs_pending" array and remove any interrupts the Guest
143
         * wants blocked: the result ends up in "blk". */
144
        if (copy_from_user(&blk, lg->lguest_data->blocked_interrupts,
145
                           sizeof(blk)))
146
                return;
147
 
148
        bitmap_andnot(blk, lg->irqs_pending, blk, LGUEST_IRQS);
149
 
150
        /* Find the first interrupt. */
151
        irq = find_first_bit(blk, LGUEST_IRQS);
152
        /* None?  Nothing to do */
153
        if (irq >= LGUEST_IRQS)
154
                return;
155
 
156
        /* They may be in the middle of an iret, where they asked us never to
157
         * deliver interrupts. */
158
        if (lg->regs->eip >= lg->noirq_start && lg->regs->eip < lg->noirq_end)
159
                return;
160
 
161
        /* If they're halted, interrupts restart them. */
162
        if (lg->halted) {
163
                /* Re-enable interrupts. */
164
                if (put_user(X86_EFLAGS_IF, &lg->lguest_data->irq_enabled))
165
                        kill_guest(lg, "Re-enabling interrupts");
166
                lg->halted = 0;
167
        } else {
168
                /* Otherwise we check if they have interrupts disabled. */
169
                u32 irq_enabled;
170
                if (get_user(irq_enabled, &lg->lguest_data->irq_enabled))
171
                        irq_enabled = 0;
172
                if (!irq_enabled)
173
                        return;
174
        }
175
 
176
        /* Look at the IDT entry the Guest gave us for this interrupt.  The
177
         * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
178
         * over them. */
179
        idt = &lg->arch.idt[FIRST_EXTERNAL_VECTOR+irq];
180
        /* If they don't have a handler (yet?), we just ignore it */
181
        if (idt_present(idt->a, idt->b)) {
182
                /* OK, mark it no longer pending and deliver it. */
183
                clear_bit(irq, lg->irqs_pending);
184
                /* set_guest_interrupt() takes the interrupt descriptor and a
185
                 * flag to say whether this interrupt pushes an error code onto
186
                 * the stack as well: virtual interrupts never do. */
187
                set_guest_interrupt(lg, idt->a, idt->b, 0);
188
        }
189
 
190
        /* Every time we deliver an interrupt, we update the timestamp in the
191
         * Guest's lguest_data struct.  It would be better for the Guest if we
192
         * did this more often, but it can actually be quite slow: doing it
193
         * here is a compromise which means at least it gets updated every
194
         * timer interrupt. */
195
        write_timestamp(lg);
196
}
197
/*:*/
198
 
199
/* Linux uses trap 128 for system calls.  Plan9 uses 64, and Ron Minnich sent
200
 * me a patch, so we support that too.  It'd be a big step for lguest if half
201
 * the Plan 9 user base were to start using it.
202
 *
203
 * Actually now I think of it, it's possible that Ron *is* half the Plan 9
204
 * userbase.  Oh well. */
205
static bool could_be_syscall(unsigned int num)
206
{
207
        /* Normal Linux SYSCALL_VECTOR or reserved vector? */
208
        return num == SYSCALL_VECTOR || num == syscall_vector;
209
}
210
 
211
/* The syscall vector it wants must be unused by Host. */
212
bool check_syscall_vector(struct lguest *lg)
213
{
214
        u32 vector;
215
 
216
        if (get_user(vector, &lg->lguest_data->syscall_vec))
217
                return false;
218
 
219
        return could_be_syscall(vector);
220
}
221
 
222
int init_interrupts(void)
223
{
224
        /* If they want some strange system call vector, reserve it now */
225
        if (syscall_vector != SYSCALL_VECTOR
226
            && test_and_set_bit(syscall_vector, used_vectors)) {
227
                printk("lg: couldn't reserve syscall %u\n", syscall_vector);
228
                return -EBUSY;
229
        }
230
        return 0;
231
}
232
 
233
void free_interrupts(void)
234
{
235
        if (syscall_vector != SYSCALL_VECTOR)
236
                clear_bit(syscall_vector, used_vectors);
237
}
238
 
239
/*H:220 Now we've got the routines to deliver interrupts, delivering traps
240
 * like page fault is easy.  The only trick is that Intel decided that some
241
 * traps should have error codes: */
242
static int has_err(unsigned int trap)
243
{
244
        return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
245
}
246
 
247
/* deliver_trap() returns true if it could deliver the trap. */
248
int deliver_trap(struct lguest *lg, unsigned int num)
249
{
250
        /* Trap numbers are always 8 bit, but we set an impossible trap number
251
         * for traps inside the Switcher, so check that here. */
252
        if (num >= ARRAY_SIZE(lg->arch.idt))
253
                return 0;
254
 
255
        /* Early on the Guest hasn't set the IDT entries (or maybe it put a
256
         * bogus one in): if we fail here, the Guest will be killed. */
257
        if (!idt_present(lg->arch.idt[num].a, lg->arch.idt[num].b))
258
                return 0;
259
        set_guest_interrupt(lg, lg->arch.idt[num].a, lg->arch.idt[num].b,
260
                            has_err(num));
261
        return 1;
262
}
263
 
264
/*H:250 Here's the hard part: returning to the Host every time a trap happens
265
 * and then calling deliver_trap() and re-entering the Guest is slow.
266
 * Particularly because Guest userspace system calls are traps (usually trap
267
 * 128).
268
 *
269
 * So we'd like to set up the IDT to tell the CPU to deliver traps directly
270
 * into the Guest.  This is possible, but the complexities cause the size of
271
 * this file to double!  However, 150 lines of code is worth writing for taking
272
 * system calls down from 1750ns to 270ns.  Plus, if lguest didn't do it, all
273
 * the other hypervisors would beat it up at lunchtime.
274
 *
275
 * This routine indicates if a particular trap number could be delivered
276
 * directly. */
277
static int direct_trap(unsigned int num)
278
{
279
        /* Hardware interrupts don't go to the Guest at all (except system
280
         * call). */
281
        if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num))
282
                return 0;
283
 
284
        /* The Host needs to see page faults (for shadow paging and to save the
285
         * fault address), general protection faults (in/out emulation) and
286
         * device not available (TS handling), and of course, the hypercall
287
         * trap. */
288
        return num != 14 && num != 13 && num != 7 && num != LGUEST_TRAP_ENTRY;
289
}
290
/*:*/
291
 
292
/*M:005 The Guest has the ability to turn its interrupt gates into trap gates,
293
 * if it is careful.  The Host will let trap gates can go directly to the
294
 * Guest, but the Guest needs the interrupts atomically disabled for an
295
 * interrupt gate.  It can do this by pointing the trap gate at instructions
296
 * within noirq_start and noirq_end, where it can safely disable interrupts. */
297
 
298
/*M:006 The Guests do not use the sysenter (fast system call) instruction,
299
 * because it's hardcoded to enter privilege level 0 and so can't go direct.
300
 * It's about twice as fast as the older "int 0x80" system call, so it might
301
 * still be worthwhile to handle it in the Switcher and lcall down to the
302
 * Guest.  The sysenter semantics are hairy tho: search for that keyword in
303
 * entry.S :*/
304
 
305
/*H:260 When we make traps go directly into the Guest, we need to make sure
306
 * the kernel stack is valid (ie. mapped in the page tables).  Otherwise, the
307
 * CPU trying to deliver the trap will fault while trying to push the interrupt
308
 * words on the stack: this is called a double fault, and it forces us to kill
309
 * the Guest.
310
 *
311
 * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. */
312
void pin_stack_pages(struct lguest *lg)
313
{
314
        unsigned int i;
315
 
316
        /* Depending on the CONFIG_4KSTACKS option, the Guest can have one or
317
         * two pages of stack space. */
318
        for (i = 0; i < lg->stack_pages; i++)
319
                /* The stack grows *upwards*, so the address we're given is the
320
                 * start of the page after the kernel stack.  Subtract one to
321
                 * get back onto the first stack page, and keep subtracting to
322
                 * get to the rest of the stack pages. */
323
                pin_page(lg, lg->esp1 - 1 - i * PAGE_SIZE);
324
}
325
 
326
/* Direct traps also mean that we need to know whenever the Guest wants to use
327
 * a different kernel stack, so we can change the IDT entries to use that
328
 * stack.  The IDT entries expect a virtual address, so unlike most addresses
329
 * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
330
 * physical.
331
 *
332
 * In Linux each process has its own kernel stack, so this happens a lot: we
333
 * change stacks on each context switch. */
334
void guest_set_stack(struct lguest *lg, u32 seg, u32 esp, unsigned int pages)
335
{
336
        /* You are not allowed have a stack segment with privilege level 0: bad
337
         * Guest! */
338
        if ((seg & 0x3) != GUEST_PL)
339
                kill_guest(lg, "bad stack segment %i", seg);
340
        /* We only expect one or two stack pages. */
341
        if (pages > 2)
342
                kill_guest(lg, "bad stack pages %u", pages);
343
        /* Save where the stack is, and how many pages */
344
        lg->ss1 = seg;
345
        lg->esp1 = esp;
346
        lg->stack_pages = pages;
347
        /* Make sure the new stack pages are mapped */
348
        pin_stack_pages(lg);
349
}
350
 
351
/* All this reference to mapping stacks leads us neatly into the other complex
352
 * part of the Host: page table handling. */
353
 
354
/*H:235 This is the routine which actually checks the Guest's IDT entry and
355
 * transfers it into the entry in "struct lguest": */
356
static void set_trap(struct lguest *lg, struct desc_struct *trap,
357
                     unsigned int num, u32 lo, u32 hi)
358
{
359
        u8 type = idt_type(lo, hi);
360
 
361
        /* We zero-out a not-present entry */
362
        if (!idt_present(lo, hi)) {
363
                trap->a = trap->b = 0;
364
                return;
365
        }
366
 
367
        /* We only support interrupt and trap gates. */
368
        if (type != 0xE && type != 0xF)
369
                kill_guest(lg, "bad IDT type %i", type);
370
 
371
        /* We only copy the handler address, present bit, privilege level and
372
         * type.  The privilege level controls where the trap can be triggered
373
         * manually with an "int" instruction.  This is usually GUEST_PL,
374
         * except for system calls which userspace can use. */
375
        trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
376
        trap->b = (hi&0xFFFFEF00);
377
}
378
 
379
/*H:230 While we're here, dealing with delivering traps and interrupts to the
380
 * Guest, we might as well complete the picture: how the Guest tells us where
381
 * it wants them to go.  This would be simple, except making traps fast
382
 * requires some tricks.
383
 *
384
 * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
385
 * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. */
386
void load_guest_idt_entry(struct lguest *lg, unsigned int num, u32 lo, u32 hi)
387
{
388
        /* Guest never handles: NMI, doublefault, spurious interrupt or
389
         * hypercall.  We ignore when it tries to set them. */
390
        if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
391
                return;
392
 
393
        /* Mark the IDT as changed: next time the Guest runs we'll know we have
394
         * to copy this again. */
395
        lg->changed |= CHANGED_IDT;
396
 
397
        /* Check that the Guest doesn't try to step outside the bounds. */
398
        if (num >= ARRAY_SIZE(lg->arch.idt))
399
                kill_guest(lg, "Setting idt entry %u", num);
400
        else
401
                set_trap(lg, &lg->arch.idt[num], num, lo, hi);
402
}
403
 
404
/* The default entry for each interrupt points into the Switcher routines which
405
 * simply return to the Host.  The run_guest() loop will then call
406
 * deliver_trap() to bounce it back into the Guest. */
407
static void default_idt_entry(struct desc_struct *idt,
408
                              int trap,
409
                              const unsigned long handler)
410
{
411
        /* A present interrupt gate. */
412
        u32 flags = 0x8e00;
413
 
414
        /* Set the privilege level on the entry for the hypercall: this allows
415
         * the Guest to use the "int" instruction to trigger it. */
416
        if (trap == LGUEST_TRAP_ENTRY)
417
                flags |= (GUEST_PL << 13);
418
 
419
        /* Now pack it into the IDT entry in its weird format. */
420
        idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
421
        idt->b = (handler&0xFFFF0000) | flags;
422
}
423
 
424
/* When the Guest first starts, we put default entries into the IDT. */
425
void setup_default_idt_entries(struct lguest_ro_state *state,
426
                               const unsigned long *def)
427
{
428
        unsigned int i;
429
 
430
        for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
431
                default_idt_entry(&state->guest_idt[i], i, def[i]);
432
}
433
 
434
/*H:240 We don't use the IDT entries in the "struct lguest" directly, instead
435
 * we copy them into the IDT which we've set up for Guests on this CPU, just
436
 * before we run the Guest.  This routine does that copy. */
437
void copy_traps(const struct lguest *lg, struct desc_struct *idt,
438
                const unsigned long *def)
439
{
440
        unsigned int i;
441
 
442
        /* We can simply copy the direct traps, otherwise we use the default
443
         * ones in the Switcher: they will return to the Host. */
444
        for (i = 0; i < ARRAY_SIZE(lg->arch.idt); i++) {
445
                /* If no Guest can ever override this trap, leave it alone. */
446
                if (!direct_trap(i))
447
                        continue;
448
 
449
                /* Only trap gates (type 15) can go direct to the Guest.
450
                 * Interrupt gates (type 14) disable interrupts as they are
451
                 * entered, which we never let the Guest do.  Not present
452
                 * entries (type 0x0) also can't go direct, of course. */
453
                if (idt_type(lg->arch.idt[i].a, lg->arch.idt[i].b) == 0xF)
454
                        idt[i] = lg->arch.idt[i];
455
                else
456
                        /* Reset it to the default. */
457
                        default_idt_entry(&idt[i], i, def[i]);
458
        }
459
}
460
 
461
/*H:200
462
 * The Guest Clock.
463
 *
464
 * There are two sources of virtual interrupts.  We saw one in lguest_user.c:
465
 * the Launcher sending interrupts for virtual devices.  The other is the Guest
466
 * timer interrupt.
467
 *
468
 * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
469
 * the next timer interrupt (in nanoseconds).  We use the high-resolution timer
470
 * infrastructure to set a callback at that time.
471
 *
472
 * 0 means "turn off the clock". */
473
void guest_set_clockevent(struct lguest *lg, unsigned long delta)
474
{
475
        ktime_t expires;
476
 
477
        if (unlikely(delta == 0)) {
478
                /* Clock event device is shutting down. */
479
                hrtimer_cancel(&lg->hrt);
480
                return;
481
        }
482
 
483
        /* We use wallclock time here, so the Guest might not be running for
484
         * all the time between now and the timer interrupt it asked for.  This
485
         * is almost always the right thing to do. */
486
        expires = ktime_add_ns(ktime_get_real(), delta);
487
        hrtimer_start(&lg->hrt, expires, HRTIMER_MODE_ABS);
488
}
489
 
490
/* This is the function called when the Guest's timer expires. */
491
static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
492
{
493
        struct lguest *lg = container_of(timer, struct lguest, hrt);
494
 
495
        /* Remember the first interrupt is the timer interrupt. */
496
        set_bit(0, lg->irqs_pending);
497
        /* If the Guest is actually stopped, we need to wake it up. */
498
        if (lg->halted)
499
                wake_up_process(lg->tsk);
500
        return HRTIMER_NORESTART;
501
}
502
 
503
/* This sets up the timer for this Guest. */
504
void init_clockdev(struct lguest *lg)
505
{
506
        hrtimer_init(&lg->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
507
        lg->hrt.function = clockdev_fn;
508
}

powered by: WebSVN 2.1.0

© copyright 1999-2025 OpenCores.org, equivalent to Oliscience, all rights reserved. OpenCores®, registered trademark.