Commit | Line | Data |
---|---|---|
f938d2c8 RR |
1 | /*P:400 This contains run_guest() which actually calls into the Host<->Guest |
2 | * Switcher and analyzes the return, such as determining if the Guest wants the | |
3 | * Host to do something. This file also contains useful helper routines, and a | |
4 | * couple of non-obvious setup and teardown pieces which were implemented after | |
5 | * days of debugging pain. :*/ | |
d7e28ffe RR |
6 | #include <linux/module.h> |
7 | #include <linux/stringify.h> | |
8 | #include <linux/stddef.h> | |
9 | #include <linux/io.h> | |
10 | #include <linux/mm.h> | |
11 | #include <linux/vmalloc.h> | |
12 | #include <linux/cpu.h> | |
13 | #include <linux/freezer.h> | |
14 | #include <asm/paravirt.h> | |
15 | #include <asm/desc.h> | |
16 | #include <asm/pgtable.h> | |
17 | #include <asm/uaccess.h> | |
18 | #include <asm/poll.h> | |
19 | #include <asm/highmem.h> | |
20 | #include <asm/asm-offsets.h> | |
21 | #include <asm/i387.h> | |
22 | #include "lg.h" | |
23 | ||
24 | /* Found in switcher.S */ | |
25 | extern char start_switcher_text[], end_switcher_text[], switch_to_guest[]; | |
26 | extern unsigned long default_idt_entries[]; | |
27 | ||
28 | /* Every guest maps the core switcher code. */ | |
29 | #define SHARED_SWITCHER_PAGES \ | |
30 | DIV_ROUND_UP(end_switcher_text - start_switcher_text, PAGE_SIZE) | |
31 | /* Pages for switcher itself, then two pages per cpu */ | |
32 | #define TOTAL_SWITCHER_PAGES (SHARED_SWITCHER_PAGES + 2 * NR_CPUS) | |
33 | ||
34 | /* We map at -4M for ease of mapping into the guest (one PTE page). */ | |
35 | #define SWITCHER_ADDR 0xFFC00000 | |
36 | ||
37 | static struct vm_struct *switcher_vma; | |
38 | static struct page **switcher_page; | |
39 | ||
40 | static int cpu_had_pge; | |
41 | static struct { | |
42 | unsigned long offset; | |
43 | unsigned short segment; | |
44 | } lguest_entry; | |
45 | ||
46 | /* This One Big lock protects all inter-guest data structures. */ | |
47 | DEFINE_MUTEX(lguest_lock); | |
48 | static DEFINE_PER_CPU(struct lguest *, last_guest); | |
49 | ||
50 | /* FIXME: Make dynamic. */ | |
51 | #define MAX_LGUEST_GUESTS 16 | |
52 | struct lguest lguests[MAX_LGUEST_GUESTS]; | |
53 | ||
54 | /* Offset from where switcher.S was compiled to where we've copied it */ | |
55 | static unsigned long switcher_offset(void) | |
56 | { | |
57 | return SWITCHER_ADDR - (unsigned long)start_switcher_text; | |
58 | } | |
59 | ||
60 | /* This cpu's struct lguest_pages. */ | |
61 | static struct lguest_pages *lguest_pages(unsigned int cpu) | |
62 | { | |
63 | return &(((struct lguest_pages *) | |
64 | (SWITCHER_ADDR + SHARED_SWITCHER_PAGES*PAGE_SIZE))[cpu]); | |
65 | } | |
66 | ||
bff672e6 RR |
67 | /*H:010 We need to set up the Switcher at a high virtual address. Remember the |
68 | * Switcher is a few hundred bytes of assembler code which actually changes the | |
69 | * CPU to run the Guest, and then changes back to the Host when a trap or | |
70 | * interrupt happens. | |
71 | * | |
72 | * The Switcher code must be at the same virtual address in the Guest as the | |
73 | * Host since it will be running as the switchover occurs. | |
74 | * | |
75 | * Trying to map memory at a particular address is an unusual thing to do, so | |
76 | * it's not a simple one-liner. We also set up the per-cpu parts of the | |
77 | * Switcher here. | |
78 | */ | |
d7e28ffe RR |
79 | static __init int map_switcher(void) |
80 | { | |
81 | int i, err; | |
82 | struct page **pagep; | |
83 | ||
bff672e6 RR |
84 | /* |
85 | * Map the Switcher in to high memory. | |
86 | * | |
87 | * It turns out that if we choose the address 0xFFC00000 (4MB under the | |
88 | * top virtual address), it makes setting up the page tables really | |
89 | * easy. | |
90 | */ | |
91 | ||
92 | /* We allocate an array of "struct page"s. map_vm_area() wants the | |
93 | * pages in this form, rather than just an array of pointers. */ | |
d7e28ffe RR |
94 | switcher_page = kmalloc(sizeof(switcher_page[0])*TOTAL_SWITCHER_PAGES, |
95 | GFP_KERNEL); | |
96 | if (!switcher_page) { | |
97 | err = -ENOMEM; | |
98 | goto out; | |
99 | } | |
100 | ||
bff672e6 RR |
101 | /* Now we actually allocate the pages. The Guest will see these pages, |
102 | * so we make sure they're zeroed. */ | |
d7e28ffe RR |
103 | for (i = 0; i < TOTAL_SWITCHER_PAGES; i++) { |
104 | unsigned long addr = get_zeroed_page(GFP_KERNEL); | |
105 | if (!addr) { | |
106 | err = -ENOMEM; | |
107 | goto free_some_pages; | |
108 | } | |
109 | switcher_page[i] = virt_to_page(addr); | |
110 | } | |
111 | ||
bff672e6 RR |
112 | /* Now we reserve the "virtual memory area" we want: 0xFFC00000 |
113 | * (SWITCHER_ADDR). We might not get it in theory, but in practice | |
114 | * it's worked so far. */ | |
d7e28ffe RR |
115 | switcher_vma = __get_vm_area(TOTAL_SWITCHER_PAGES * PAGE_SIZE, |
116 | VM_ALLOC, SWITCHER_ADDR, VMALLOC_END); | |
117 | if (!switcher_vma) { | |
118 | err = -ENOMEM; | |
119 | printk("lguest: could not map switcher pages high\n"); | |
120 | goto free_pages; | |
121 | } | |
122 | ||
bff672e6 RR |
123 | /* This code actually sets up the pages we've allocated to appear at |
124 | * SWITCHER_ADDR. map_vm_area() takes the vma we allocated above, the | |
125 | * kind of pages we're mapping (kernel pages), and a pointer to our | |
126 | * array of struct pages. It increments that pointer, but we don't | |
127 | * care. */ | |
d7e28ffe RR |
128 | pagep = switcher_page; |
129 | err = map_vm_area(switcher_vma, PAGE_KERNEL, &pagep); | |
130 | if (err) { | |
131 | printk("lguest: map_vm_area failed: %i\n", err); | |
132 | goto free_vma; | |
133 | } | |
bff672e6 RR |
134 | |
135 | /* Now the switcher is mapped at the right address, we can't fail! | |
136 | * Copy in the compiled-in Switcher code (from switcher.S). */ | |
d7e28ffe RR |
137 | memcpy(switcher_vma->addr, start_switcher_text, |
138 | end_switcher_text - start_switcher_text); | |
139 | ||
bff672e6 RR |
140 | /* Most of the switcher.S doesn't care that it's been moved; on Intel, |
141 | * jumps are relative, and it doesn't access any references to external | |
142 | * code or data. | |
143 | * | |
144 | * The only exception is the interrupt handlers in switcher.S: their | |
145 | * addresses are placed in a table (default_idt_entries), so we need to | |
146 | * update the table with the new addresses. switcher_offset() is a | |
147 | * convenience function which returns the distance between the builtin | |
148 | * switcher code and the high-mapped copy we just made. */ | |
d7e28ffe RR |
149 | for (i = 0; i < IDT_ENTRIES; i++) |
150 | default_idt_entries[i] += switcher_offset(); | |
151 | ||
bff672e6 RR |
152 | /* |
153 | * Set up the Switcher's per-cpu areas. | |
154 | * | |
155 | * Each CPU gets two pages of its own within the high-mapped region | |
156 | * (aka. "struct lguest_pages"). Much of this can be initialized now, | |
157 | * but some depends on what Guest we are running (which is set up in | |
158 | * copy_in_guest_info()). | |
159 | */ | |
d7e28ffe | 160 | for_each_possible_cpu(i) { |
bff672e6 | 161 | /* lguest_pages() returns this CPU's two pages. */ |
d7e28ffe | 162 | struct lguest_pages *pages = lguest_pages(i); |
bff672e6 RR |
163 | /* This is a convenience pointer to make the code fit one |
164 | * statement to a line. */ | |
d7e28ffe RR |
165 | struct lguest_ro_state *state = &pages->state; |
166 | ||
bff672e6 RR |
167 | /* The Global Descriptor Table: the Host has a different one |
168 | * for each CPU. We keep a descriptor for the GDT which says | |
169 | * where it is and how big it is (the size is actually the last | |
170 | * byte, not the size, hence the "-1"). */ | |
d7e28ffe RR |
171 | state->host_gdt_desc.size = GDT_SIZE-1; |
172 | state->host_gdt_desc.address = (long)get_cpu_gdt_table(i); | |
bff672e6 RR |
173 | |
174 | /* All CPUs on the Host use the same Interrupt Descriptor | |
175 | * Table, so we just use store_idt(), which gets this CPU's IDT | |
176 | * descriptor. */ | |
d7e28ffe | 177 | store_idt(&state->host_idt_desc); |
bff672e6 RR |
178 | |
179 | /* The descriptors for the Guest's GDT and IDT can be filled | |
180 | * out now, too. We copy the GDT & IDT into ->guest_gdt and | |
181 | * ->guest_idt before actually running the Guest. */ | |
d7e28ffe RR |
182 | state->guest_idt_desc.size = sizeof(state->guest_idt)-1; |
183 | state->guest_idt_desc.address = (long)&state->guest_idt; | |
184 | state->guest_gdt_desc.size = sizeof(state->guest_gdt)-1; | |
185 | state->guest_gdt_desc.address = (long)&state->guest_gdt; | |
bff672e6 RR |
186 | |
187 | /* We know where we want the stack to be when the Guest enters | |
188 | * the switcher: in pages->regs. The stack grows upwards, so | |
189 | * we start it at the end of that structure. */ | |
d7e28ffe | 190 | state->guest_tss.esp0 = (long)(&pages->regs + 1); |
bff672e6 RR |
191 | /* And this is the GDT entry to use for the stack: we keep a |
192 | * couple of special LGUEST entries. */ | |
d7e28ffe | 193 | state->guest_tss.ss0 = LGUEST_DS; |
bff672e6 RR |
194 | |
195 | /* x86 can have a finegrained bitmap which indicates what I/O | |
196 | * ports the process can use. We set it to the end of our | |
197 | * structure, meaning "none". */ | |
d7e28ffe | 198 | state->guest_tss.io_bitmap_base = sizeof(state->guest_tss); |
bff672e6 RR |
199 | |
200 | /* Some GDT entries are the same across all Guests, so we can | |
201 | * set them up now. */ | |
d7e28ffe | 202 | setup_default_gdt_entries(state); |
bff672e6 | 203 | /* Most IDT entries are the same for all Guests, too.*/ |
d7e28ffe RR |
204 | setup_default_idt_entries(state, default_idt_entries); |
205 | ||
bff672e6 RR |
206 | /* The Host needs to be able to use the LGUEST segments on this |
207 | * CPU, too, so put them in the Host GDT. */ | |
d7e28ffe RR |
208 | get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT; |
209 | get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT; | |
210 | } | |
211 | ||
bff672e6 RR |
212 | /* In the Switcher, we want the %cs segment register to use the |
213 | * LGUEST_CS GDT entry: we've put that in the Host and Guest GDTs, so | |
214 | * it will be undisturbed when we switch. To change %cs and jump we | |
215 | * need this structure to feed to Intel's "lcall" instruction. */ | |
d7e28ffe RR |
216 | lguest_entry.offset = (long)switch_to_guest + switcher_offset(); |
217 | lguest_entry.segment = LGUEST_CS; | |
218 | ||
219 | printk(KERN_INFO "lguest: mapped switcher at %p\n", | |
220 | switcher_vma->addr); | |
bff672e6 | 221 | /* And we succeeded... */ |
d7e28ffe RR |
222 | return 0; |
223 | ||
224 | free_vma: | |
225 | vunmap(switcher_vma->addr); | |
226 | free_pages: | |
227 | i = TOTAL_SWITCHER_PAGES; | |
228 | free_some_pages: | |
229 | for (--i; i >= 0; i--) | |
230 | __free_pages(switcher_page[i], 0); | |
231 | kfree(switcher_page); | |
232 | out: | |
233 | return err; | |
234 | } | |
bff672e6 | 235 | /*:*/ |
d7e28ffe | 236 | |
bff672e6 RR |
237 | /* Cleaning up the mapping when the module is unloaded is almost... |
238 | * too easy. */ | |
d7e28ffe RR |
239 | static void unmap_switcher(void) |
240 | { | |
241 | unsigned int i; | |
242 | ||
bff672e6 | 243 | /* vunmap() undoes *both* map_vm_area() and __get_vm_area(). */ |
d7e28ffe | 244 | vunmap(switcher_vma->addr); |
bff672e6 | 245 | /* Now we just need to free the pages we copied the switcher into */ |
d7e28ffe RR |
246 | for (i = 0; i < TOTAL_SWITCHER_PAGES; i++) |
247 | __free_pages(switcher_page[i], 0); | |
248 | } | |
249 | ||
bff672e6 | 250 | /*H:130 Our Guest is usually so well behaved; it never tries to do things it |
93b1eab3 JF |
251 | * isn't allowed to. Unfortunately, Linux's paravirtual infrastructure isn't |
252 | * quite complete, because it doesn't contain replacements for the Intel I/O | |
bff672e6 RR |
253 | * instructions. As a result, the Guest sometimes fumbles across one during |
254 | * the boot process as it probes for various things which are usually attached | |
255 | * to a PC. | |
256 | * | |
257 | * When the Guest uses one of these instructions, we get trap #13 (General | |
258 | * Protection Fault) and come here. We see if it's one of those troublesome | |
259 | * instructions and skip over it. We return true if we did. */ | |
d7e28ffe RR |
260 | static int emulate_insn(struct lguest *lg) |
261 | { | |
262 | u8 insn; | |
263 | unsigned int insnlen = 0, in = 0, shift = 0; | |
bff672e6 RR |
264 | /* The eip contains the *virtual* address of the Guest's instruction: |
265 | * guest_pa just subtracts the Guest's page_offset. */ | |
d7e28ffe RR |
266 | unsigned long physaddr = guest_pa(lg, lg->regs->eip); |
267 | ||
bff672e6 RR |
268 | /* The guest_pa() function only works for Guest kernel addresses, but |
269 | * that's all we're trying to do anyway. */ | |
d7e28ffe RR |
270 | if (lg->regs->eip < lg->page_offset) |
271 | return 0; | |
bff672e6 RR |
272 | |
273 | /* Decoding x86 instructions is icky. */ | |
d7e28ffe RR |
274 | lgread(lg, &insn, physaddr, 1); |
275 | ||
bff672e6 RR |
276 | /* 0x66 is an "operand prefix". It means it's using the upper 16 bits |
277 | of the eax register. */ | |
d7e28ffe RR |
278 | if (insn == 0x66) { |
279 | shift = 16; | |
bff672e6 | 280 | /* The instruction is 1 byte so far, read the next byte. */ |
d7e28ffe RR |
281 | insnlen = 1; |
282 | lgread(lg, &insn, physaddr + insnlen, 1); | |
283 | } | |
284 | ||
bff672e6 RR |
285 | /* We can ignore the lower bit for the moment and decode the 4 opcodes |
286 | * we need to emulate. */ | |
d7e28ffe RR |
287 | switch (insn & 0xFE) { |
288 | case 0xE4: /* in <next byte>,%al */ | |
289 | insnlen += 2; | |
290 | in = 1; | |
291 | break; | |
292 | case 0xEC: /* in (%dx),%al */ | |
293 | insnlen += 1; | |
294 | in = 1; | |
295 | break; | |
296 | case 0xE6: /* out %al,<next byte> */ | |
297 | insnlen += 2; | |
298 | break; | |
299 | case 0xEE: /* out %al,(%dx) */ | |
300 | insnlen += 1; | |
301 | break; | |
302 | default: | |
bff672e6 | 303 | /* OK, we don't know what this is, can't emulate. */ |
d7e28ffe RR |
304 | return 0; |
305 | } | |
306 | ||
bff672e6 RR |
307 | /* If it was an "IN" instruction, they expect the result to be read |
308 | * into %eax, so we change %eax. We always return all-ones, which | |
309 | * traditionally means "there's nothing there". */ | |
d7e28ffe RR |
310 | if (in) { |
311 | /* Lower bit tells is whether it's a 16 or 32 bit access */ | |
312 | if (insn & 0x1) | |
313 | lg->regs->eax = 0xFFFFFFFF; | |
314 | else | |
315 | lg->regs->eax |= (0xFFFF << shift); | |
316 | } | |
bff672e6 | 317 | /* Finally, we've "done" the instruction, so move past it. */ |
d7e28ffe | 318 | lg->regs->eip += insnlen; |
bff672e6 | 319 | /* Success! */ |
d7e28ffe RR |
320 | return 1; |
321 | } | |
bff672e6 | 322 | /*:*/ |
d7e28ffe | 323 | |
dde79789 RR |
324 | /*L:305 |
325 | * Dealing With Guest Memory. | |
326 | * | |
327 | * When the Guest gives us (what it thinks is) a physical address, we can use | |
328 | * the normal copy_from_user() & copy_to_user() on that address: remember, | |
329 | * Guest physical == Launcher virtual. | |
330 | * | |
331 | * But we can't trust the Guest: it might be trying to access the Launcher | |
332 | * code. We have to check that the range is below the pfn_limit the Launcher | |
333 | * gave us. We have to make sure that addr + len doesn't give us a false | |
334 | * positive by overflowing, too. */ | |
d7e28ffe RR |
335 | int lguest_address_ok(const struct lguest *lg, |
336 | unsigned long addr, unsigned long len) | |
337 | { | |
338 | return (addr+len) / PAGE_SIZE < lg->pfn_limit && (addr+len >= addr); | |
339 | } | |
340 | ||
dde79789 RR |
341 | /* This is a convenient routine to get a 32-bit value from the Guest (a very |
342 | * common operation). Here we can see how useful the kill_lguest() routine we | |
343 | * met in the Launcher can be: we return a random value (0) instead of needing | |
344 | * to return an error. */ | |
d7e28ffe RR |
345 | u32 lgread_u32(struct lguest *lg, unsigned long addr) |
346 | { | |
347 | u32 val = 0; | |
348 | ||
dde79789 | 349 | /* Don't let them access lguest binary. */ |
d7e28ffe RR |
350 | if (!lguest_address_ok(lg, addr, sizeof(val)) |
351 | || get_user(val, (u32 __user *)addr) != 0) | |
352 | kill_guest(lg, "bad read address %#lx", addr); | |
353 | return val; | |
354 | } | |
355 | ||
dde79789 | 356 | /* Same thing for writing a value. */ |
d7e28ffe RR |
357 | void lgwrite_u32(struct lguest *lg, unsigned long addr, u32 val) |
358 | { | |
359 | if (!lguest_address_ok(lg, addr, sizeof(val)) | |
360 | || put_user(val, (u32 __user *)addr) != 0) | |
361 | kill_guest(lg, "bad write address %#lx", addr); | |
362 | } | |
363 | ||
dde79789 RR |
364 | /* This routine is more generic, and copies a range of Guest bytes into a |
365 | * buffer. If the copy_from_user() fails, we fill the buffer with zeroes, so | |
366 | * the caller doesn't end up using uninitialized kernel memory. */ | |
d7e28ffe RR |
367 | void lgread(struct lguest *lg, void *b, unsigned long addr, unsigned bytes) |
368 | { | |
369 | if (!lguest_address_ok(lg, addr, bytes) | |
370 | || copy_from_user(b, (void __user *)addr, bytes) != 0) { | |
371 | /* copy_from_user should do this, but as we rely on it... */ | |
372 | memset(b, 0, bytes); | |
373 | kill_guest(lg, "bad read address %#lx len %u", addr, bytes); | |
374 | } | |
375 | } | |
376 | ||
dde79789 | 377 | /* Similarly, our generic routine to copy into a range of Guest bytes. */ |
d7e28ffe RR |
378 | void lgwrite(struct lguest *lg, unsigned long addr, const void *b, |
379 | unsigned bytes) | |
380 | { | |
381 | if (!lguest_address_ok(lg, addr, bytes) | |
382 | || copy_to_user((void __user *)addr, b, bytes) != 0) | |
383 | kill_guest(lg, "bad write address %#lx len %u", addr, bytes); | |
384 | } | |
dde79789 | 385 | /* (end of memory access helper routines) :*/ |
d7e28ffe RR |
386 | |
387 | static void set_ts(void) | |
388 | { | |
389 | u32 cr0; | |
390 | ||
391 | cr0 = read_cr0(); | |
392 | if (!(cr0 & 8)) | |
393 | write_cr0(cr0|8); | |
394 | } | |
395 | ||
f8f0fdcd RR |
396 | /*S:010 |
397 | * We are getting close to the Switcher. | |
398 | * | |
399 | * Remember that each CPU has two pages which are visible to the Guest when it | |
400 | * runs on that CPU. This has to contain the state for that Guest: we copy the | |
401 | * state in just before we run the Guest. | |
402 | * | |
403 | * Each Guest has "changed" flags which indicate what has changed in the Guest | |
404 | * since it last ran. We saw this set in interrupts_and_traps.c and | |
405 | * segments.c. | |
406 | */ | |
d7e28ffe RR |
407 | static void copy_in_guest_info(struct lguest *lg, struct lguest_pages *pages) |
408 | { | |
f8f0fdcd RR |
409 | /* Copying all this data can be quite expensive. We usually run the |
410 | * same Guest we ran last time (and that Guest hasn't run anywhere else | |
411 | * meanwhile). If that's not the case, we pretend everything in the | |
412 | * Guest has changed. */ | |
d7e28ffe RR |
413 | if (__get_cpu_var(last_guest) != lg || lg->last_pages != pages) { |
414 | __get_cpu_var(last_guest) = lg; | |
415 | lg->last_pages = pages; | |
416 | lg->changed = CHANGED_ALL; | |
417 | } | |
418 | ||
f8f0fdcd RR |
419 | /* These copies are pretty cheap, so we do them unconditionally: */ |
420 | /* Save the current Host top-level page directory. */ | |
d7e28ffe | 421 | pages->state.host_cr3 = __pa(current->mm->pgd); |
f8f0fdcd RR |
422 | /* Set up the Guest's page tables to see this CPU's pages (and no |
423 | * other CPU's pages). */ | |
d7e28ffe | 424 | map_switcher_in_guest(lg, pages); |
f8f0fdcd RR |
425 | /* Set up the two "TSS" members which tell the CPU what stack to use |
426 | * for traps which do directly into the Guest (ie. traps at privilege | |
427 | * level 1). */ | |
d7e28ffe RR |
428 | pages->state.guest_tss.esp1 = lg->esp1; |
429 | pages->state.guest_tss.ss1 = lg->ss1; | |
430 | ||
f8f0fdcd | 431 | /* Copy direct-to-Guest trap entries. */ |
d7e28ffe RR |
432 | if (lg->changed & CHANGED_IDT) |
433 | copy_traps(lg, pages->state.guest_idt, default_idt_entries); | |
434 | ||
f8f0fdcd | 435 | /* Copy all GDT entries which the Guest can change. */ |
d7e28ffe RR |
436 | if (lg->changed & CHANGED_GDT) |
437 | copy_gdt(lg, pages->state.guest_gdt); | |
438 | /* If only the TLS entries have changed, copy them. */ | |
439 | else if (lg->changed & CHANGED_GDT_TLS) | |
440 | copy_gdt_tls(lg, pages->state.guest_gdt); | |
441 | ||
f8f0fdcd | 442 | /* Mark the Guest as unchanged for next time. */ |
d7e28ffe RR |
443 | lg->changed = 0; |
444 | } | |
445 | ||
f8f0fdcd | 446 | /* Finally: the code to actually call into the Switcher to run the Guest. */ |
d7e28ffe RR |
447 | static void run_guest_once(struct lguest *lg, struct lguest_pages *pages) |
448 | { | |
f8f0fdcd | 449 | /* This is a dummy value we need for GCC's sake. */ |
d7e28ffe RR |
450 | unsigned int clobber; |
451 | ||
f8f0fdcd RR |
452 | /* Copy the guest-specific information into this CPU's "struct |
453 | * lguest_pages". */ | |
d7e28ffe RR |
454 | copy_in_guest_info(lg, pages); |
455 | ||
0d027c01 RR |
456 | /* Set the trap number to 256 (impossible value). If we fault while |
457 | * switching to the Guest (bad segment registers or bug), this will | |
458 | * cause us to abort the Guest. */ | |
459 | lg->regs->trapnum = 256; | |
460 | ||
f8f0fdcd RR |
461 | /* Now: we push the "eflags" register on the stack, then do an "lcall". |
462 | * This is how we change from using the kernel code segment to using | |
463 | * the dedicated lguest code segment, as well as jumping into the | |
464 | * Switcher. | |
465 | * | |
466 | * The lcall also pushes the old code segment (KERNEL_CS) onto the | |
467 | * stack, then the address of this call. This stack layout happens to | |
468 | * exactly match the stack of an interrupt... */ | |
d7e28ffe | 469 | asm volatile("pushf; lcall *lguest_entry" |
f8f0fdcd RR |
470 | /* This is how we tell GCC that %eax ("a") and %ebx ("b") |
471 | * are changed by this routine. The "=" means output. */ | |
d7e28ffe | 472 | : "=a"(clobber), "=b"(clobber) |
f8f0fdcd RR |
473 | /* %eax contains the pages pointer. ("0" refers to the |
474 | * 0-th argument above, ie "a"). %ebx contains the | |
475 | * physical address of the Guest's top-level page | |
476 | * directory. */ | |
d7e28ffe | 477 | : "0"(pages), "1"(__pa(lg->pgdirs[lg->pgdidx].pgdir)) |
f8f0fdcd RR |
478 | /* We tell gcc that all these registers could change, |
479 | * which means we don't have to save and restore them in | |
480 | * the Switcher. */ | |
d7e28ffe RR |
481 | : "memory", "%edx", "%ecx", "%edi", "%esi"); |
482 | } | |
f8f0fdcd | 483 | /*:*/ |
d7e28ffe | 484 | |
bff672e6 RR |
485 | /*H:030 Let's jump straight to the the main loop which runs the Guest. |
486 | * Remember, this is called by the Launcher reading /dev/lguest, and we keep | |
487 | * going around and around until something interesting happens. */ | |
d7e28ffe RR |
488 | int run_guest(struct lguest *lg, unsigned long __user *user) |
489 | { | |
bff672e6 | 490 | /* We stop running once the Guest is dead. */ |
d7e28ffe | 491 | while (!lg->dead) { |
bff672e6 RR |
492 | /* We need to initialize this, otherwise gcc complains. It's |
493 | * not (yet) clever enough to see that it's initialized when we | |
494 | * need it. */ | |
d7e28ffe RR |
495 | unsigned int cr2 = 0; /* Damn gcc */ |
496 | ||
bff672e6 RR |
497 | /* First we run any hypercalls the Guest wants done: either in |
498 | * the hypercall ring in "struct lguest_data", or directly by | |
499 | * using int 31 (LGUEST_TRAP_ENTRY). */ | |
d7e28ffe | 500 | do_hypercalls(lg); |
bff672e6 RR |
501 | /* It's possible the Guest did a SEND_DMA hypercall to the |
502 | * Launcher, in which case we return from the read() now. */ | |
d7e28ffe RR |
503 | if (lg->dma_is_pending) { |
504 | if (put_user(lg->pending_dma, user) || | |
505 | put_user(lg->pending_key, user+1)) | |
506 | return -EFAULT; | |
507 | return sizeof(unsigned long)*2; | |
508 | } | |
509 | ||
bff672e6 | 510 | /* Check for signals */ |
d7e28ffe RR |
511 | if (signal_pending(current)) |
512 | return -ERESTARTSYS; | |
513 | ||
514 | /* If Waker set break_out, return to Launcher. */ | |
515 | if (lg->break_out) | |
516 | return -EAGAIN; | |
517 | ||
bff672e6 RR |
518 | /* Check if there are any interrupts which can be delivered |
519 | * now: if so, this sets up the hander to be executed when we | |
520 | * next run the Guest. */ | |
d7e28ffe RR |
521 | maybe_do_interrupt(lg); |
522 | ||
bff672e6 RR |
523 | /* All long-lived kernel loops need to check with this horrible |
524 | * thing called the freezer. If the Host is trying to suspend, | |
525 | * it stops us. */ | |
d7e28ffe RR |
526 | try_to_freeze(); |
527 | ||
bff672e6 RR |
528 | /* Just make absolutely sure the Guest is still alive. One of |
529 | * those hypercalls could have been fatal, for example. */ | |
d7e28ffe RR |
530 | if (lg->dead) |
531 | break; | |
532 | ||
bff672e6 RR |
533 | /* If the Guest asked to be stopped, we sleep. The Guest's |
534 | * clock timer or LHCALL_BREAK from the Waker will wake us. */ | |
d7e28ffe RR |
535 | if (lg->halted) { |
536 | set_current_state(TASK_INTERRUPTIBLE); | |
537 | schedule(); | |
538 | continue; | |
539 | } | |
540 | ||
bff672e6 RR |
541 | /* OK, now we're ready to jump into the Guest. First we put up |
542 | * the "Do Not Disturb" sign: */ | |
d7e28ffe RR |
543 | local_irq_disable(); |
544 | ||
bff672e6 RR |
545 | /* Remember the awfully-named TS bit? If the Guest has asked |
546 | * to set it we set it now, so we can trap and pass that trap | |
547 | * to the Guest if it uses the FPU. */ | |
d7e28ffe RR |
548 | if (lg->ts) |
549 | set_ts(); | |
550 | ||
bff672e6 RR |
551 | /* SYSENTER is an optimized way of doing system calls. We |
552 | * can't allow it because it always jumps to privilege level 0. | |
553 | * A normal Guest won't try it because we don't advertise it in | |
554 | * CPUID, but a malicious Guest (or malicious Guest userspace | |
555 | * program) could, so we tell the CPU to disable it before | |
556 | * running the Guest. */ | |
d7e28ffe RR |
557 | if (boot_cpu_has(X86_FEATURE_SEP)) |
558 | wrmsr(MSR_IA32_SYSENTER_CS, 0, 0); | |
559 | ||
bff672e6 RR |
560 | /* Now we actually run the Guest. It will pop back out when |
561 | * something interesting happens, and we can examine its | |
562 | * registers to see what it was doing. */ | |
d7e28ffe RR |
563 | run_guest_once(lg, lguest_pages(raw_smp_processor_id())); |
564 | ||
bff672e6 RR |
565 | /* The "regs" pointer contains two extra entries which are not |
566 | * really registers: a trap number which says what interrupt or | |
567 | * trap made the switcher code come back, and an error code | |
568 | * which some traps set. */ | |
569 | ||
570 | /* If the Guest page faulted, then the cr2 register will tell | |
571 | * us the bad virtual address. We have to grab this now, | |
572 | * because once we re-enable interrupts an interrupt could | |
573 | * fault and thus overwrite cr2, or we could even move off to a | |
574 | * different CPU. */ | |
d7e28ffe RR |
575 | if (lg->regs->trapnum == 14) |
576 | cr2 = read_cr2(); | |
bff672e6 RR |
577 | /* Similarly, if we took a trap because the Guest used the FPU, |
578 | * we have to restore the FPU it expects to see. */ | |
d7e28ffe RR |
579 | else if (lg->regs->trapnum == 7) |
580 | math_state_restore(); | |
581 | ||
bff672e6 | 582 | /* Restore SYSENTER if it's supposed to be on. */ |
d7e28ffe RR |
583 | if (boot_cpu_has(X86_FEATURE_SEP)) |
584 | wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0); | |
bff672e6 RR |
585 | |
586 | /* Now we're ready to be interrupted or moved to other CPUs */ | |
d7e28ffe RR |
587 | local_irq_enable(); |
588 | ||
bff672e6 | 589 | /* OK, so what happened? */ |
d7e28ffe RR |
590 | switch (lg->regs->trapnum) { |
591 | case 13: /* We've intercepted a GPF. */ | |
bff672e6 RR |
592 | /* Check if this was one of those annoying IN or OUT |
593 | * instructions which we need to emulate. If so, we | |
594 | * just go back into the Guest after we've done it. */ | |
d7e28ffe RR |
595 | if (lg->regs->errcode == 0) { |
596 | if (emulate_insn(lg)) | |
597 | continue; | |
598 | } | |
599 | break; | |
600 | case 14: /* We've intercepted a page fault. */ | |
bff672e6 RR |
601 | /* The Guest accessed a virtual address that wasn't |
602 | * mapped. This happens a lot: we don't actually set | |
603 | * up most of the page tables for the Guest at all when | |
604 | * we start: as it runs it asks for more and more, and | |
605 | * we set them up as required. In this case, we don't | |
606 | * even tell the Guest that the fault happened. | |
607 | * | |
608 | * The errcode tells whether this was a read or a | |
609 | * write, and whether kernel or userspace code. */ | |
d7e28ffe RR |
610 | if (demand_page(lg, cr2, lg->regs->errcode)) |
611 | continue; | |
612 | ||
bff672e6 RR |
613 | /* OK, it's really not there (or not OK): the Guest |
614 | * needs to know. We write out the cr2 value so it | |
615 | * knows where the fault occurred. | |
616 | * | |
617 | * Note that if the Guest were really messed up, this | |
618 | * could happen before it's done the INITIALIZE | |
619 | * hypercall, so lg->lguest_data will be NULL, so | |
620 | * &lg->lguest_data->cr2 will be address 8. Writing | |
621 | * into that address won't hurt the Host at all, | |
622 | * though. */ | |
d7e28ffe RR |
623 | if (put_user(cr2, &lg->lguest_data->cr2)) |
624 | kill_guest(lg, "Writing cr2"); | |
625 | break; | |
626 | case 7: /* We've intercepted a Device Not Available fault. */ | |
bff672e6 RR |
627 | /* If the Guest doesn't want to know, we already |
628 | * restored the Floating Point Unit, so we just | |
629 | * continue without telling it. */ | |
d7e28ffe RR |
630 | if (!lg->ts) |
631 | continue; | |
632 | break; | |
bff672e6 RR |
633 | case 32 ... 255: |
634 | /* These values mean a real interrupt occurred, in | |
635 | * which case the Host handler has already been run. | |
636 | * We just do a friendly check if another process | |
637 | * should now be run, then fall through to loop | |
638 | * around: */ | |
d7e28ffe RR |
639 | cond_resched(); |
640 | case LGUEST_TRAP_ENTRY: /* Handled at top of loop */ | |
641 | continue; | |
642 | } | |
643 | ||
bff672e6 RR |
644 | /* If we get here, it's a trap the Guest wants to know |
645 | * about. */ | |
d7e28ffe RR |
646 | if (deliver_trap(lg, lg->regs->trapnum)) |
647 | continue; | |
648 | ||
bff672e6 RR |
649 | /* If the Guest doesn't have a handler (either it hasn't |
650 | * registered any yet, or it's one of the faults we don't let | |
651 | * it handle), it dies with a cryptic error message. */ | |
d7e28ffe RR |
652 | kill_guest(lg, "unhandled trap %li at %#lx (%#lx)", |
653 | lg->regs->trapnum, lg->regs->eip, | |
654 | lg->regs->trapnum == 14 ? cr2 : lg->regs->errcode); | |
655 | } | |
bff672e6 | 656 | /* The Guest is dead => "No such file or directory" */ |
d7e28ffe RR |
657 | return -ENOENT; |
658 | } | |
659 | ||
bff672e6 RR |
660 | /* Now we can look at each of the routines this calls, in increasing order of |
661 | * complexity: do_hypercalls(), emulate_insn(), maybe_do_interrupt(), | |
662 | * deliver_trap() and demand_page(). After all those, we'll be ready to | |
663 | * examine the Switcher, and our philosophical understanding of the Host/Guest | |
664 | * duality will be complete. :*/ | |
665 | ||
d7e28ffe RR |
666 | int find_free_guest(void) |
667 | { | |
668 | unsigned int i; | |
669 | for (i = 0; i < MAX_LGUEST_GUESTS; i++) | |
670 | if (!lguests[i].tsk) | |
671 | return i; | |
672 | return -1; | |
673 | } | |
674 | ||
675 | static void adjust_pge(void *on) | |
676 | { | |
677 | if (on) | |
678 | write_cr4(read_cr4() | X86_CR4_PGE); | |
679 | else | |
680 | write_cr4(read_cr4() & ~X86_CR4_PGE); | |
681 | } | |
682 | ||
bff672e6 RR |
683 | /*H:000 |
684 | * Welcome to the Host! | |
685 | * | |
686 | * By this point your brain has been tickled by the Guest code and numbed by | |
687 | * the Launcher code; prepare for it to be stretched by the Host code. This is | |
688 | * the heart. Let's begin at the initialization routine for the Host's lg | |
689 | * module. | |
690 | */ | |
d7e28ffe RR |
691 | static int __init init(void) |
692 | { | |
693 | int err; | |
694 | ||
bff672e6 | 695 | /* Lguest can't run under Xen, VMI or itself. It does Tricky Stuff. */ |
d7e28ffe | 696 | if (paravirt_enabled()) { |
93b1eab3 | 697 | printk("lguest is afraid of %s\n", pv_info.name); |
d7e28ffe RR |
698 | return -EPERM; |
699 | } | |
700 | ||
bff672e6 | 701 | /* First we put the Switcher up in very high virtual memory. */ |
d7e28ffe RR |
702 | err = map_switcher(); |
703 | if (err) | |
704 | return err; | |
705 | ||
bff672e6 | 706 | /* Now we set up the pagetable implementation for the Guests. */ |
d7e28ffe RR |
707 | err = init_pagetables(switcher_page, SHARED_SWITCHER_PAGES); |
708 | if (err) { | |
709 | unmap_switcher(); | |
710 | return err; | |
711 | } | |
bff672e6 RR |
712 | |
713 | /* The I/O subsystem needs some things initialized. */ | |
d7e28ffe RR |
714 | lguest_io_init(); |
715 | ||
bff672e6 | 716 | /* /dev/lguest needs to be registered. */ |
d7e28ffe RR |
717 | err = lguest_device_init(); |
718 | if (err) { | |
719 | free_pagetables(); | |
720 | unmap_switcher(); | |
721 | return err; | |
722 | } | |
bff672e6 RR |
723 | |
724 | /* Finally, we need to turn off "Page Global Enable". PGE is an | |
725 | * optimization where page table entries are specially marked to show | |
726 | * they never change. The Host kernel marks all the kernel pages this | |
727 | * way because it's always present, even when userspace is running. | |
728 | * | |
729 | * Lguest breaks this: unbeknownst to the rest of the Host kernel, we | |
730 | * switch to the Guest kernel. If you don't disable this on all CPUs, | |
731 | * you'll get really weird bugs that you'll chase for two days. | |
732 | * | |
733 | * I used to turn PGE off every time we switched to the Guest and back | |
734 | * on when we return, but that slowed the Switcher down noticibly. */ | |
735 | ||
736 | /* We don't need the complexity of CPUs coming and going while we're | |
737 | * doing this. */ | |
d7e28ffe RR |
738 | lock_cpu_hotplug(); |
739 | if (cpu_has_pge) { /* We have a broader idea of "global". */ | |
bff672e6 | 740 | /* Remember that this was originally set (for cleanup). */ |
d7e28ffe | 741 | cpu_had_pge = 1; |
bff672e6 RR |
742 | /* adjust_pge is a helper function which sets or unsets the PGE |
743 | * bit on its CPU, depending on the argument (0 == unset). */ | |
d7e28ffe | 744 | on_each_cpu(adjust_pge, (void *)0, 0, 1); |
bff672e6 | 745 | /* Turn off the feature in the global feature set. */ |
d7e28ffe RR |
746 | clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); |
747 | } | |
748 | unlock_cpu_hotplug(); | |
bff672e6 RR |
749 | |
750 | /* All good! */ | |
d7e28ffe RR |
751 | return 0; |
752 | } | |
753 | ||
bff672e6 | 754 | /* Cleaning up is just the same code, backwards. With a little French. */ |
d7e28ffe RR |
755 | static void __exit fini(void) |
756 | { | |
757 | lguest_device_remove(); | |
758 | free_pagetables(); | |
759 | unmap_switcher(); | |
bff672e6 RR |
760 | |
761 | /* If we had PGE before we started, turn it back on now. */ | |
d7e28ffe RR |
762 | lock_cpu_hotplug(); |
763 | if (cpu_had_pge) { | |
764 | set_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); | |
bff672e6 | 765 | /* adjust_pge's argument "1" means set PGE. */ |
d7e28ffe RR |
766 | on_each_cpu(adjust_pge, (void *)1, 0, 1); |
767 | } | |
768 | unlock_cpu_hotplug(); | |
769 | } | |
770 | ||
bff672e6 RR |
771 | /* The Host side of lguest can be a module. This is a nice way for people to |
772 | * play with it. */ | |
d7e28ffe RR |
773 | module_init(init); |
774 | module_exit(fini); | |
775 | MODULE_LICENSE("GPL"); | |
776 | MODULE_AUTHOR("Rusty Russell <rusty@rustcorp.com.au>"); |