Commit | Line | Data |
---|---|---|
625efab1 JS |
1 | /* |
2 | * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation. | |
3 | * Copyright (C) 2007, Jes Sorensen <jes@sgi.com> SGI. | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or modify | |
6 | * it under the terms of the GNU General Public License as published by | |
7 | * the Free Software Foundation; either version 2 of the License, or | |
8 | * (at your option) any later version. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, but | |
11 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or | |
13 | * NON INFRINGEMENT. See the GNU General Public License for more | |
14 | * details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * along with this program; if not, write to the Free Software | |
18 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
19 | */ | |
20 | #include <linux/kernel.h> | |
21 | #include <linux/start_kernel.h> | |
22 | #include <linux/string.h> | |
23 | #include <linux/console.h> | |
24 | #include <linux/screen_info.h> | |
25 | #include <linux/irq.h> | |
26 | #include <linux/interrupt.h> | |
27 | #include <linux/clocksource.h> | |
28 | #include <linux/clockchips.h> | |
29 | #include <linux/cpu.h> | |
30 | #include <linux/lguest.h> | |
31 | #include <linux/lguest_launcher.h> | |
32 | #include <linux/lguest_bus.h> | |
33 | #include <asm/paravirt.h> | |
34 | #include <asm/param.h> | |
35 | #include <asm/page.h> | |
36 | #include <asm/pgtable.h> | |
37 | #include <asm/desc.h> | |
38 | #include <asm/setup.h> | |
39 | #include <asm/lguest.h> | |
40 | #include <asm/uaccess.h> | |
41 | #include <asm/i387.h> | |
42 | #include "../lg.h" | |
43 | ||
44 | static int cpu_had_pge; | |
45 | ||
46 | static struct { | |
47 | unsigned long offset; | |
48 | unsigned short segment; | |
49 | } lguest_entry; | |
50 | ||
51 | /* Offset from where switcher.S was compiled to where we've copied it */ | |
52 | static unsigned long switcher_offset(void) | |
53 | { | |
54 | return SWITCHER_ADDR - (unsigned long)start_switcher_text; | |
55 | } | |
56 | ||
57 | /* This cpu's struct lguest_pages. */ | |
58 | static struct lguest_pages *lguest_pages(unsigned int cpu) | |
59 | { | |
60 | return &(((struct lguest_pages *) | |
61 | (SWITCHER_ADDR + SHARED_SWITCHER_PAGES*PAGE_SIZE))[cpu]); | |
62 | } | |
63 | ||
64 | static DEFINE_PER_CPU(struct lguest *, last_guest); | |
65 | ||
66 | /*S:010 | |
67 | * We are getting close to the Switcher. | |
68 | * | |
69 | * Remember that each CPU has two pages which are visible to the Guest when it | |
70 | * runs on that CPU. This has to contain the state for that Guest: we copy the | |
71 | * state in just before we run the Guest. | |
72 | * | |
73 | * Each Guest has "changed" flags which indicate what has changed in the Guest | |
74 | * since it last ran. We saw this set in interrupts_and_traps.c and | |
75 | * segments.c. | |
76 | */ | |
77 | static void copy_in_guest_info(struct lguest *lg, struct lguest_pages *pages) | |
78 | { | |
79 | /* Copying all this data can be quite expensive. We usually run the | |
80 | * same Guest we ran last time (and that Guest hasn't run anywhere else | |
81 | * meanwhile). If that's not the case, we pretend everything in the | |
82 | * Guest has changed. */ | |
83 | if (__get_cpu_var(last_guest) != lg || lg->last_pages != pages) { | |
84 | __get_cpu_var(last_guest) = lg; | |
85 | lg->last_pages = pages; | |
86 | lg->changed = CHANGED_ALL; | |
87 | } | |
88 | ||
89 | /* These copies are pretty cheap, so we do them unconditionally: */ | |
90 | /* Save the current Host top-level page directory. */ | |
91 | pages->state.host_cr3 = __pa(current->mm->pgd); | |
92 | /* Set up the Guest's page tables to see this CPU's pages (and no | |
93 | * other CPU's pages). */ | |
94 | map_switcher_in_guest(lg, pages); | |
95 | /* Set up the two "TSS" members which tell the CPU what stack to use | |
96 | * for traps which do directly into the Guest (ie. traps at privilege | |
97 | * level 1). */ | |
98 | pages->state.guest_tss.esp1 = lg->esp1; | |
99 | pages->state.guest_tss.ss1 = lg->ss1; | |
100 | ||
101 | /* Copy direct-to-Guest trap entries. */ | |
102 | if (lg->changed & CHANGED_IDT) | |
103 | copy_traps(lg, pages->state.guest_idt, default_idt_entries); | |
104 | ||
105 | /* Copy all GDT entries which the Guest can change. */ | |
106 | if (lg->changed & CHANGED_GDT) | |
107 | copy_gdt(lg, pages->state.guest_gdt); | |
108 | /* If only the TLS entries have changed, copy them. */ | |
109 | else if (lg->changed & CHANGED_GDT_TLS) | |
110 | copy_gdt_tls(lg, pages->state.guest_gdt); | |
111 | ||
112 | /* Mark the Guest as unchanged for next time. */ | |
113 | lg->changed = 0; | |
114 | } | |
115 | ||
116 | /* Finally: the code to actually call into the Switcher to run the Guest. */ | |
117 | static void run_guest_once(struct lguest *lg, struct lguest_pages *pages) | |
118 | { | |
119 | /* This is a dummy value we need for GCC's sake. */ | |
120 | unsigned int clobber; | |
121 | ||
122 | /* Copy the guest-specific information into this CPU's "struct | |
123 | * lguest_pages". */ | |
124 | copy_in_guest_info(lg, pages); | |
125 | ||
126 | /* Set the trap number to 256 (impossible value). If we fault while | |
127 | * switching to the Guest (bad segment registers or bug), this will | |
128 | * cause us to abort the Guest. */ | |
129 | lg->regs->trapnum = 256; | |
130 | ||
131 | /* Now: we push the "eflags" register on the stack, then do an "lcall". | |
132 | * This is how we change from using the kernel code segment to using | |
133 | * the dedicated lguest code segment, as well as jumping into the | |
134 | * Switcher. | |
135 | * | |
136 | * The lcall also pushes the old code segment (KERNEL_CS) onto the | |
137 | * stack, then the address of this call. This stack layout happens to | |
138 | * exactly match the stack of an interrupt... */ | |
139 | asm volatile("pushf; lcall *lguest_entry" | |
140 | /* This is how we tell GCC that %eax ("a") and %ebx ("b") | |
141 | * are changed by this routine. The "=" means output. */ | |
142 | : "=a"(clobber), "=b"(clobber) | |
143 | /* %eax contains the pages pointer. ("0" refers to the | |
144 | * 0-th argument above, ie "a"). %ebx contains the | |
145 | * physical address of the Guest's top-level page | |
146 | * directory. */ | |
147 | : "0"(pages), "1"(__pa(lg->pgdirs[lg->pgdidx].pgdir)) | |
148 | /* We tell gcc that all these registers could change, | |
149 | * which means we don't have to save and restore them in | |
150 | * the Switcher. */ | |
151 | : "memory", "%edx", "%ecx", "%edi", "%esi"); | |
152 | } | |
153 | /*:*/ | |
154 | ||
155 | /*H:040 This is the i386-specific code to setup and run the Guest. Interrupts | |
156 | * are disabled: we own the CPU. */ | |
157 | void lguest_arch_run_guest(struct lguest *lg) | |
158 | { | |
159 | /* Remember the awfully-named TS bit? If the Guest has asked | |
160 | * to set it we set it now, so we can trap and pass that trap | |
161 | * to the Guest if it uses the FPU. */ | |
162 | if (lg->ts) | |
163 | lguest_set_ts(); | |
164 | ||
165 | /* SYSENTER is an optimized way of doing system calls. We | |
166 | * can't allow it because it always jumps to privilege level 0. | |
167 | * A normal Guest won't try it because we don't advertise it in | |
168 | * CPUID, but a malicious Guest (or malicious Guest userspace | |
169 | * program) could, so we tell the CPU to disable it before | |
170 | * running the Guest. */ | |
171 | if (boot_cpu_has(X86_FEATURE_SEP)) | |
172 | wrmsr(MSR_IA32_SYSENTER_CS, 0, 0); | |
173 | ||
174 | /* Now we actually run the Guest. It will pop back out when | |
175 | * something interesting happens, and we can examine its | |
176 | * registers to see what it was doing. */ | |
177 | run_guest_once(lg, lguest_pages(raw_smp_processor_id())); | |
178 | ||
179 | /* The "regs" pointer contains two extra entries which are not | |
180 | * really registers: a trap number which says what interrupt or | |
181 | * trap made the switcher code come back, and an error code | |
182 | * which some traps set. */ | |
183 | ||
184 | /* If the Guest page faulted, then the cr2 register will tell | |
185 | * us the bad virtual address. We have to grab this now, | |
186 | * because once we re-enable interrupts an interrupt could | |
187 | * fault and thus overwrite cr2, or we could even move off to a | |
188 | * different CPU. */ | |
189 | if (lg->regs->trapnum == 14) | |
190 | lg->arch.last_pagefault = read_cr2(); | |
191 | /* Similarly, if we took a trap because the Guest used the FPU, | |
192 | * we have to restore the FPU it expects to see. */ | |
193 | else if (lg->regs->trapnum == 7) | |
194 | math_state_restore(); | |
195 | ||
196 | /* Restore SYSENTER if it's supposed to be on. */ | |
197 | if (boot_cpu_has(X86_FEATURE_SEP)) | |
198 | wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0); | |
199 | } | |
200 | ||
201 | /*H:130 Our Guest is usually so well behaved; it never tries to do things it | |
202 | * isn't allowed to. Unfortunately, Linux's paravirtual infrastructure isn't | |
203 | * quite complete, because it doesn't contain replacements for the Intel I/O | |
204 | * instructions. As a result, the Guest sometimes fumbles across one during | |
205 | * the boot process as it probes for various things which are usually attached | |
206 | * to a PC. | |
207 | * | |
208 | * When the Guest uses one of these instructions, we get trap #13 (General | |
209 | * Protection Fault) and come here. We see if it's one of those troublesome | |
210 | * instructions and skip over it. We return true if we did. */ | |
211 | static int emulate_insn(struct lguest *lg) | |
212 | { | |
213 | u8 insn; | |
214 | unsigned int insnlen = 0, in = 0, shift = 0; | |
215 | /* The eip contains the *virtual* address of the Guest's instruction: | |
216 | * guest_pa just subtracts the Guest's page_offset. */ | |
217 | unsigned long physaddr = guest_pa(lg, lg->regs->eip); | |
218 | ||
219 | /* The guest_pa() function only works for Guest kernel addresses, but | |
220 | * that's all we're trying to do anyway. */ | |
221 | if (lg->regs->eip < lg->page_offset) | |
222 | return 0; | |
223 | ||
224 | /* Decoding x86 instructions is icky. */ | |
225 | lgread(lg, &insn, physaddr, 1); | |
226 | ||
227 | /* 0x66 is an "operand prefix". It means it's using the upper 16 bits | |
228 | of the eax register. */ | |
229 | if (insn == 0x66) { | |
230 | shift = 16; | |
231 | /* The instruction is 1 byte so far, read the next byte. */ | |
232 | insnlen = 1; | |
233 | lgread(lg, &insn, physaddr + insnlen, 1); | |
234 | } | |
235 | ||
236 | /* We can ignore the lower bit for the moment and decode the 4 opcodes | |
237 | * we need to emulate. */ | |
238 | switch (insn & 0xFE) { | |
239 | case 0xE4: /* in <next byte>,%al */ | |
240 | insnlen += 2; | |
241 | in = 1; | |
242 | break; | |
243 | case 0xEC: /* in (%dx),%al */ | |
244 | insnlen += 1; | |
245 | in = 1; | |
246 | break; | |
247 | case 0xE6: /* out %al,<next byte> */ | |
248 | insnlen += 2; | |
249 | break; | |
250 | case 0xEE: /* out %al,(%dx) */ | |
251 | insnlen += 1; | |
252 | break; | |
253 | default: | |
254 | /* OK, we don't know what this is, can't emulate. */ | |
255 | return 0; | |
256 | } | |
257 | ||
258 | /* If it was an "IN" instruction, they expect the result to be read | |
259 | * into %eax, so we change %eax. We always return all-ones, which | |
260 | * traditionally means "there's nothing there". */ | |
261 | if (in) { | |
262 | /* Lower bit tells is whether it's a 16 or 32 bit access */ | |
263 | if (insn & 0x1) | |
264 | lg->regs->eax = 0xFFFFFFFF; | |
265 | else | |
266 | lg->regs->eax |= (0xFFFF << shift); | |
267 | } | |
268 | /* Finally, we've "done" the instruction, so move past it. */ | |
269 | lg->regs->eip += insnlen; | |
270 | /* Success! */ | |
271 | return 1; | |
272 | } | |
273 | ||
274 | /*H:050 Once we've re-enabled interrupts, we look at why the Guest exited. */ | |
275 | void lguest_arch_handle_trap(struct lguest *lg) | |
276 | { | |
277 | switch (lg->regs->trapnum) { | |
278 | case 13: /* We've intercepted a GPF. */ | |
279 | /* Check if this was one of those annoying IN or OUT | |
280 | * instructions which we need to emulate. If so, we | |
281 | * just go back into the Guest after we've done it. */ | |
282 | if (lg->regs->errcode == 0) { | |
283 | if (emulate_insn(lg)) | |
284 | return; | |
285 | } | |
286 | break; | |
287 | case 14: /* We've intercepted a page fault. */ | |
288 | /* The Guest accessed a virtual address that wasn't | |
289 | * mapped. This happens a lot: we don't actually set | |
290 | * up most of the page tables for the Guest at all when | |
291 | * we start: as it runs it asks for more and more, and | |
292 | * we set them up as required. In this case, we don't | |
293 | * even tell the Guest that the fault happened. | |
294 | * | |
295 | * The errcode tells whether this was a read or a | |
296 | * write, and whether kernel or userspace code. */ | |
297 | if (demand_page(lg, lg->arch.last_pagefault, lg->regs->errcode)) | |
298 | return; | |
299 | ||
300 | /* OK, it's really not there (or not OK): the Guest | |
301 | * needs to know. We write out the cr2 value so it | |
302 | * knows where the fault occurred. | |
303 | * | |
304 | * Note that if the Guest were really messed up, this | |
305 | * could happen before it's done the INITIALIZE | |
306 | * hypercall, so lg->lguest_data will be NULL */ | |
307 | if (lg->lguest_data && | |
308 | put_user(lg->arch.last_pagefault, &lg->lguest_data->cr2)) | |
309 | kill_guest(lg, "Writing cr2"); | |
310 | break; | |
311 | case 7: /* We've intercepted a Device Not Available fault. */ | |
312 | /* If the Guest doesn't want to know, we already | |
313 | * restored the Floating Point Unit, so we just | |
314 | * continue without telling it. */ | |
315 | if (!lg->ts) | |
316 | return; | |
317 | break; | |
318 | case 32 ... 255: | |
319 | /* These values mean a real interrupt occurred, in | |
320 | * which case the Host handler has already been run. | |
321 | * We just do a friendly check if another process | |
322 | * should now be run, then fall through to loop | |
323 | * around: */ | |
324 | cond_resched(); | |
325 | case LGUEST_TRAP_ENTRY: /* Handled before re-entering Guest */ | |
326 | return; | |
327 | } | |
328 | ||
329 | /* We didn't handle the trap, so it needs to go to the Guest. */ | |
330 | if (!deliver_trap(lg, lg->regs->trapnum)) | |
331 | /* If the Guest doesn't have a handler (either it hasn't | |
332 | * registered any yet, or it's one of the faults we don't let | |
333 | * it handle), it dies with a cryptic error message. */ | |
334 | kill_guest(lg, "unhandled trap %li at %#lx (%#lx)", | |
335 | lg->regs->trapnum, lg->regs->eip, | |
336 | lg->regs->trapnum == 14 ? lg->arch.last_pagefault | |
337 | : lg->regs->errcode); | |
338 | } | |
339 | ||
340 | /* Now we can look at each of the routines this calls, in increasing order of | |
341 | * complexity: do_hypercalls(), emulate_insn(), maybe_do_interrupt(), | |
342 | * deliver_trap() and demand_page(). After all those, we'll be ready to | |
343 | * examine the Switcher, and our philosophical understanding of the Host/Guest | |
344 | * duality will be complete. :*/ | |
345 | static void adjust_pge(void *on) | |
346 | { | |
347 | if (on) | |
348 | write_cr4(read_cr4() | X86_CR4_PGE); | |
349 | else | |
350 | write_cr4(read_cr4() & ~X86_CR4_PGE); | |
351 | } | |
352 | ||
353 | /*H:020 Now the Switcher is mapped and every thing else is ready, we need to do | |
354 | * some more i386-specific initialization. */ | |
355 | void __init lguest_arch_host_init(void) | |
356 | { | |
357 | int i; | |
358 | ||
359 | /* Most of the i386/switcher.S doesn't care that it's been moved; on | |
360 | * Intel, jumps are relative, and it doesn't access any references to | |
361 | * external code or data. | |
362 | * | |
363 | * The only exception is the interrupt handlers in switcher.S: their | |
364 | * addresses are placed in a table (default_idt_entries), so we need to | |
365 | * update the table with the new addresses. switcher_offset() is a | |
366 | * convenience function which returns the distance between the builtin | |
367 | * switcher code and the high-mapped copy we just made. */ | |
368 | for (i = 0; i < IDT_ENTRIES; i++) | |
369 | default_idt_entries[i] += switcher_offset(); | |
370 | ||
371 | /* | |
372 | * Set up the Switcher's per-cpu areas. | |
373 | * | |
374 | * Each CPU gets two pages of its own within the high-mapped region | |
375 | * (aka. "struct lguest_pages"). Much of this can be initialized now, | |
376 | * but some depends on what Guest we are running (which is set up in | |
377 | * copy_in_guest_info()). | |
378 | */ | |
379 | for_each_possible_cpu(i) { | |
380 | /* lguest_pages() returns this CPU's two pages. */ | |
381 | struct lguest_pages *pages = lguest_pages(i); | |
382 | /* This is a convenience pointer to make the code fit one | |
383 | * statement to a line. */ | |
384 | struct lguest_ro_state *state = &pages->state; | |
385 | ||
386 | /* The Global Descriptor Table: the Host has a different one | |
387 | * for each CPU. We keep a descriptor for the GDT which says | |
388 | * where it is and how big it is (the size is actually the last | |
389 | * byte, not the size, hence the "-1"). */ | |
390 | state->host_gdt_desc.size = GDT_SIZE-1; | |
391 | state->host_gdt_desc.address = (long)get_cpu_gdt_table(i); | |
392 | ||
393 | /* All CPUs on the Host use the same Interrupt Descriptor | |
394 | * Table, so we just use store_idt(), which gets this CPU's IDT | |
395 | * descriptor. */ | |
396 | store_idt(&state->host_idt_desc); | |
397 | ||
398 | /* The descriptors for the Guest's GDT and IDT can be filled | |
399 | * out now, too. We copy the GDT & IDT into ->guest_gdt and | |
400 | * ->guest_idt before actually running the Guest. */ | |
401 | state->guest_idt_desc.size = sizeof(state->guest_idt)-1; | |
402 | state->guest_idt_desc.address = (long)&state->guest_idt; | |
403 | state->guest_gdt_desc.size = sizeof(state->guest_gdt)-1; | |
404 | state->guest_gdt_desc.address = (long)&state->guest_gdt; | |
405 | ||
406 | /* We know where we want the stack to be when the Guest enters | |
407 | * the switcher: in pages->regs. The stack grows upwards, so | |
408 | * we start it at the end of that structure. */ | |
409 | state->guest_tss.esp0 = (long)(&pages->regs + 1); | |
410 | /* And this is the GDT entry to use for the stack: we keep a | |
411 | * couple of special LGUEST entries. */ | |
412 | state->guest_tss.ss0 = LGUEST_DS; | |
413 | ||
414 | /* x86 can have a finegrained bitmap which indicates what I/O | |
415 | * ports the process can use. We set it to the end of our | |
416 | * structure, meaning "none". */ | |
417 | state->guest_tss.io_bitmap_base = sizeof(state->guest_tss); | |
418 | ||
419 | /* Some GDT entries are the same across all Guests, so we can | |
420 | * set them up now. */ | |
421 | setup_default_gdt_entries(state); | |
422 | /* Most IDT entries are the same for all Guests, too.*/ | |
423 | setup_default_idt_entries(state, default_idt_entries); | |
424 | ||
425 | /* The Host needs to be able to use the LGUEST segments on this | |
426 | * CPU, too, so put them in the Host GDT. */ | |
427 | get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT; | |
428 | get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT; | |
429 | } | |
430 | ||
431 | /* In the Switcher, we want the %cs segment register to use the | |
432 | * LGUEST_CS GDT entry: we've put that in the Host and Guest GDTs, so | |
433 | * it will be undisturbed when we switch. To change %cs and jump we | |
434 | * need this structure to feed to Intel's "lcall" instruction. */ | |
435 | lguest_entry.offset = (long)switch_to_guest + switcher_offset(); | |
436 | lguest_entry.segment = LGUEST_CS; | |
437 | ||
438 | /* Finally, we need to turn off "Page Global Enable". PGE is an | |
439 | * optimization where page table entries are specially marked to show | |
440 | * they never change. The Host kernel marks all the kernel pages this | |
441 | * way because it's always present, even when userspace is running. | |
442 | * | |
443 | * Lguest breaks this: unbeknownst to the rest of the Host kernel, we | |
444 | * switch to the Guest kernel. If you don't disable this on all CPUs, | |
445 | * you'll get really weird bugs that you'll chase for two days. | |
446 | * | |
447 | * I used to turn PGE off every time we switched to the Guest and back | |
448 | * on when we return, but that slowed the Switcher down noticibly. */ | |
449 | ||
450 | /* We don't need the complexity of CPUs coming and going while we're | |
451 | * doing this. */ | |
452 | lock_cpu_hotplug(); | |
453 | if (cpu_has_pge) { /* We have a broader idea of "global". */ | |
454 | /* Remember that this was originally set (for cleanup). */ | |
455 | cpu_had_pge = 1; | |
456 | /* adjust_pge is a helper function which sets or unsets the PGE | |
457 | * bit on its CPU, depending on the argument (0 == unset). */ | |
458 | on_each_cpu(adjust_pge, (void *)0, 0, 1); | |
459 | /* Turn off the feature in the global feature set. */ | |
460 | clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); | |
461 | } | |
462 | unlock_cpu_hotplug(); | |
463 | }; | |
464 | /*:*/ | |
465 | ||
466 | void __exit lguest_arch_host_fini(void) | |
467 | { | |
468 | /* If we had PGE before we started, turn it back on now. */ | |
469 | lock_cpu_hotplug(); | |
470 | if (cpu_had_pge) { | |
471 | set_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); | |
472 | /* adjust_pge's argument "1" means set PGE. */ | |
473 | on_each_cpu(adjust_pge, (void *)1, 0, 1); | |
474 | } | |
475 | unlock_cpu_hotplug(); | |
476 | } |