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> | |
625efab1 JS |
32 | #include <asm/paravirt.h> |
33 | #include <asm/param.h> | |
34 | #include <asm/page.h> | |
35 | #include <asm/pgtable.h> | |
36 | #include <asm/desc.h> | |
37 | #include <asm/setup.h> | |
38 | #include <asm/lguest.h> | |
39 | #include <asm/uaccess.h> | |
40 | #include <asm/i387.h> | |
41 | #include "../lg.h" | |
42 | ||
43 | static int cpu_had_pge; | |
44 | ||
45 | static struct { | |
46 | unsigned long offset; | |
47 | unsigned short segment; | |
48 | } lguest_entry; | |
49 | ||
50 | /* Offset from where switcher.S was compiled to where we've copied it */ | |
51 | static unsigned long switcher_offset(void) | |
52 | { | |
53 | return SWITCHER_ADDR - (unsigned long)start_switcher_text; | |
54 | } | |
55 | ||
56 | /* This cpu's struct lguest_pages. */ | |
57 | static struct lguest_pages *lguest_pages(unsigned int cpu) | |
58 | { | |
59 | return &(((struct lguest_pages *) | |
60 | (SWITCHER_ADDR + SHARED_SWITCHER_PAGES*PAGE_SIZE))[cpu]); | |
61 | } | |
62 | ||
63 | static DEFINE_PER_CPU(struct lguest *, last_guest); | |
64 | ||
65 | /*S:010 | |
e1e72965 | 66 | * We approach the Switcher. |
625efab1 JS |
67 | * |
68 | * Remember that each CPU has two pages which are visible to the Guest when it | |
69 | * runs on that CPU. This has to contain the state for that Guest: we copy the | |
70 | * state in just before we run the Guest. | |
71 | * | |
72 | * Each Guest has "changed" flags which indicate what has changed in the Guest | |
73 | * since it last ran. We saw this set in interrupts_and_traps.c and | |
74 | * segments.c. | |
75 | */ | |
d0953d42 | 76 | static void copy_in_guest_info(struct lg_cpu *cpu, struct lguest_pages *pages) |
625efab1 | 77 | { |
d0953d42 | 78 | struct lguest *lg = cpu->lg; |
625efab1 JS |
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). */ | |
0c78441c | 94 | map_switcher_in_guest(cpu, pages); |
625efab1 JS |
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) | |
fc708b3e | 103 | copy_traps(cpu, pages->state.guest_idt, default_idt_entries); |
625efab1 JS |
104 | |
105 | /* Copy all GDT entries which the Guest can change. */ | |
106 | if (lg->changed & CHANGED_GDT) | |
fc708b3e | 107 | copy_gdt(cpu, pages->state.guest_gdt); |
625efab1 JS |
108 | /* If only the TLS entries have changed, copy them. */ |
109 | else if (lg->changed & CHANGED_GDT_TLS) | |
fc708b3e | 110 | copy_gdt_tls(cpu, pages->state.guest_gdt); |
625efab1 JS |
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. */ | |
d0953d42 | 117 | static void run_guest_once(struct lg_cpu *cpu, struct lguest_pages *pages) |
625efab1 JS |
118 | { |
119 | /* This is a dummy value we need for GCC's sake. */ | |
120 | unsigned int clobber; | |
d0953d42 | 121 | struct lguest *lg = cpu->lg; |
625efab1 JS |
122 | |
123 | /* Copy the guest-specific information into this CPU's "struct | |
124 | * lguest_pages". */ | |
d0953d42 | 125 | copy_in_guest_info(cpu, pages); |
625efab1 JS |
126 | |
127 | /* Set the trap number to 256 (impossible value). If we fault while | |
128 | * switching to the Guest (bad segment registers or bug), this will | |
129 | * cause us to abort the Guest. */ | |
a53a35a8 | 130 | cpu->regs->trapnum = 256; |
625efab1 JS |
131 | |
132 | /* Now: we push the "eflags" register on the stack, then do an "lcall". | |
133 | * This is how we change from using the kernel code segment to using | |
134 | * the dedicated lguest code segment, as well as jumping into the | |
135 | * Switcher. | |
136 | * | |
137 | * The lcall also pushes the old code segment (KERNEL_CS) onto the | |
138 | * stack, then the address of this call. This stack layout happens to | |
e1e72965 | 139 | * exactly match the stack layout created by an interrupt... */ |
625efab1 JS |
140 | asm volatile("pushf; lcall *lguest_entry" |
141 | /* This is how we tell GCC that %eax ("a") and %ebx ("b") | |
142 | * are changed by this routine. The "=" means output. */ | |
143 | : "=a"(clobber), "=b"(clobber) | |
144 | /* %eax contains the pages pointer. ("0" refers to the | |
145 | * 0-th argument above, ie "a"). %ebx contains the | |
146 | * physical address of the Guest's top-level page | |
147 | * directory. */ | |
148 | : "0"(pages), "1"(__pa(lg->pgdirs[lg->pgdidx].pgdir)) | |
149 | /* We tell gcc that all these registers could change, | |
150 | * which means we don't have to save and restore them in | |
151 | * the Switcher. */ | |
152 | : "memory", "%edx", "%ecx", "%edi", "%esi"); | |
153 | } | |
154 | /*:*/ | |
155 | ||
e1e72965 RR |
156 | /*M:002 There are hooks in the scheduler which we can register to tell when we |
157 | * get kicked off the CPU (preempt_notifier_register()). This would allow us | |
158 | * to lazily disable SYSENTER which would regain some performance, and should | |
159 | * also simplify copy_in_guest_info(). Note that we'd still need to restore | |
160 | * things when we exit to Launcher userspace, but that's fairly easy. | |
161 | * | |
162 | * The hooks were designed for KVM, but we can also put them to good use. :*/ | |
163 | ||
625efab1 JS |
164 | /*H:040 This is the i386-specific code to setup and run the Guest. Interrupts |
165 | * are disabled: we own the CPU. */ | |
d0953d42 | 166 | void lguest_arch_run_guest(struct lg_cpu *cpu) |
625efab1 | 167 | { |
d0953d42 GOC |
168 | struct lguest *lg = cpu->lg; |
169 | ||
e1e72965 RR |
170 | /* Remember the awfully-named TS bit? If the Guest has asked to set it |
171 | * we set it now, so we can trap and pass that trap to the Guest if it | |
172 | * uses the FPU. */ | |
625efab1 JS |
173 | if (lg->ts) |
174 | lguest_set_ts(); | |
175 | ||
e1e72965 RR |
176 | /* SYSENTER is an optimized way of doing system calls. We can't allow |
177 | * it because it always jumps to privilege level 0. A normal Guest | |
178 | * won't try it because we don't advertise it in CPUID, but a malicious | |
179 | * Guest (or malicious Guest userspace program) could, so we tell the | |
180 | * CPU to disable it before running the Guest. */ | |
625efab1 JS |
181 | if (boot_cpu_has(X86_FEATURE_SEP)) |
182 | wrmsr(MSR_IA32_SYSENTER_CS, 0, 0); | |
183 | ||
e1e72965 RR |
184 | /* Now we actually run the Guest. It will return when something |
185 | * interesting happens, and we can examine its registers to see what it | |
186 | * was doing. */ | |
d0953d42 | 187 | run_guest_once(cpu, lguest_pages(raw_smp_processor_id())); |
625efab1 | 188 | |
e1e72965 RR |
189 | /* Note that the "regs" pointer contains two extra entries which are |
190 | * not really registers: a trap number which says what interrupt or | |
191 | * trap made the switcher code come back, and an error code which some | |
192 | * traps set. */ | |
625efab1 | 193 | |
e1e72965 RR |
194 | /* If the Guest page faulted, then the cr2 register will tell us the |
195 | * bad virtual address. We have to grab this now, because once we | |
196 | * re-enable interrupts an interrupt could fault and thus overwrite | |
197 | * cr2, or we could even move off to a different CPU. */ | |
a53a35a8 | 198 | if (cpu->regs->trapnum == 14) |
fc708b3e | 199 | cpu->arch.last_pagefault = read_cr2(); |
625efab1 JS |
200 | /* Similarly, if we took a trap because the Guest used the FPU, |
201 | * we have to restore the FPU it expects to see. */ | |
a53a35a8 | 202 | else if (cpu->regs->trapnum == 7) |
625efab1 JS |
203 | math_state_restore(); |
204 | ||
205 | /* Restore SYSENTER if it's supposed to be on. */ | |
206 | if (boot_cpu_has(X86_FEATURE_SEP)) | |
207 | wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0); | |
208 | } | |
209 | ||
e1e72965 RR |
210 | /*H:130 Now we've examined the hypercall code; our Guest can make requests. |
211 | * Our Guest is usually so well behaved; it never tries to do things it isn't | |
212 | * allowed to, and uses hypercalls instead. Unfortunately, Linux's paravirtual | |
213 | * infrastructure isn't quite complete, because it doesn't contain replacements | |
214 | * for the Intel I/O instructions. As a result, the Guest sometimes fumbles | |
215 | * across one during the boot process as it probes for various things which are | |
216 | * usually attached to a PC. | |
625efab1 | 217 | * |
e1e72965 | 218 | * When the Guest uses one of these instructions, we get a trap (General |
625efab1 JS |
219 | * Protection Fault) and come here. We see if it's one of those troublesome |
220 | * instructions and skip over it. We return true if we did. */ | |
a3863f68 | 221 | static int emulate_insn(struct lg_cpu *cpu) |
625efab1 | 222 | { |
a3863f68 | 223 | struct lguest *lg = cpu->lg; |
625efab1 JS |
224 | u8 insn; |
225 | unsigned int insnlen = 0, in = 0, shift = 0; | |
226 | /* The eip contains the *virtual* address of the Guest's instruction: | |
227 | * guest_pa just subtracts the Guest's page_offset. */ | |
a53a35a8 | 228 | unsigned long physaddr = guest_pa(lg, cpu->regs->eip); |
625efab1 | 229 | |
47436aa4 RR |
230 | /* This must be the Guest kernel trying to do something, not userspace! |
231 | * The bottom two bits of the CS segment register are the privilege | |
232 | * level. */ | |
a53a35a8 | 233 | if ((cpu->regs->cs & 3) != GUEST_PL) |
625efab1 JS |
234 | return 0; |
235 | ||
236 | /* Decoding x86 instructions is icky. */ | |
2d37f94a | 237 | insn = lgread(lg, physaddr, u8); |
625efab1 JS |
238 | |
239 | /* 0x66 is an "operand prefix". It means it's using the upper 16 bits | |
240 | of the eax register. */ | |
241 | if (insn == 0x66) { | |
242 | shift = 16; | |
243 | /* The instruction is 1 byte so far, read the next byte. */ | |
244 | insnlen = 1; | |
2d37f94a | 245 | insn = lgread(lg, physaddr + insnlen, u8); |
625efab1 JS |
246 | } |
247 | ||
248 | /* We can ignore the lower bit for the moment and decode the 4 opcodes | |
249 | * we need to emulate. */ | |
250 | switch (insn & 0xFE) { | |
251 | case 0xE4: /* in <next byte>,%al */ | |
252 | insnlen += 2; | |
253 | in = 1; | |
254 | break; | |
255 | case 0xEC: /* in (%dx),%al */ | |
256 | insnlen += 1; | |
257 | in = 1; | |
258 | break; | |
259 | case 0xE6: /* out %al,<next byte> */ | |
260 | insnlen += 2; | |
261 | break; | |
262 | case 0xEE: /* out %al,(%dx) */ | |
263 | insnlen += 1; | |
264 | break; | |
265 | default: | |
266 | /* OK, we don't know what this is, can't emulate. */ | |
267 | return 0; | |
268 | } | |
269 | ||
270 | /* If it was an "IN" instruction, they expect the result to be read | |
271 | * into %eax, so we change %eax. We always return all-ones, which | |
272 | * traditionally means "there's nothing there". */ | |
273 | if (in) { | |
274 | /* Lower bit tells is whether it's a 16 or 32 bit access */ | |
275 | if (insn & 0x1) | |
a53a35a8 | 276 | cpu->regs->eax = 0xFFFFFFFF; |
625efab1 | 277 | else |
a53a35a8 | 278 | cpu->regs->eax |= (0xFFFF << shift); |
625efab1 JS |
279 | } |
280 | /* Finally, we've "done" the instruction, so move past it. */ | |
a53a35a8 | 281 | cpu->regs->eip += insnlen; |
625efab1 JS |
282 | /* Success! */ |
283 | return 1; | |
284 | } | |
285 | ||
286 | /*H:050 Once we've re-enabled interrupts, we look at why the Guest exited. */ | |
73044f05 | 287 | void lguest_arch_handle_trap(struct lg_cpu *cpu) |
625efab1 | 288 | { |
73044f05 | 289 | struct lguest *lg = cpu->lg; |
a53a35a8 | 290 | switch (cpu->regs->trapnum) { |
e1e72965 RR |
291 | case 13: /* We've intercepted a General Protection Fault. */ |
292 | /* Check if this was one of those annoying IN or OUT | |
293 | * instructions which we need to emulate. If so, we just go | |
294 | * back into the Guest after we've done it. */ | |
a53a35a8 | 295 | if (cpu->regs->errcode == 0) { |
a3863f68 | 296 | if (emulate_insn(cpu)) |
625efab1 JS |
297 | return; |
298 | } | |
299 | break; | |
e1e72965 RR |
300 | case 14: /* We've intercepted a Page Fault. */ |
301 | /* The Guest accessed a virtual address that wasn't mapped. | |
302 | * This happens a lot: we don't actually set up most of the | |
303 | * page tables for the Guest at all when we start: as it runs | |
304 | * it asks for more and more, and we set them up as | |
305 | * required. In this case, we don't even tell the Guest that | |
306 | * the fault happened. | |
307 | * | |
308 | * The errcode tells whether this was a read or a write, and | |
309 | * whether kernel or userspace code. */ | |
fc708b3e | 310 | if (demand_page(lg,cpu->arch.last_pagefault,cpu->regs->errcode)) |
625efab1 JS |
311 | return; |
312 | ||
e1e72965 RR |
313 | /* OK, it's really not there (or not OK): the Guest needs to |
314 | * know. We write out the cr2 value so it knows where the | |
315 | * fault occurred. | |
316 | * | |
317 | * Note that if the Guest were really messed up, this could | |
318 | * happen before it's done the LHCALL_LGUEST_INIT hypercall, so | |
319 | * lg->lguest_data could be NULL */ | |
625efab1 | 320 | if (lg->lguest_data && |
fc708b3e | 321 | put_user(cpu->arch.last_pagefault, &lg->lguest_data->cr2)) |
625efab1 JS |
322 | kill_guest(lg, "Writing cr2"); |
323 | break; | |
324 | case 7: /* We've intercepted a Device Not Available fault. */ | |
e1e72965 RR |
325 | /* If the Guest doesn't want to know, we already restored the |
326 | * Floating Point Unit, so we just continue without telling | |
327 | * it. */ | |
625efab1 JS |
328 | if (!lg->ts) |
329 | return; | |
330 | break; | |
331 | case 32 ... 255: | |
cc6d4fbc RR |
332 | /* These values mean a real interrupt occurred, in which case |
333 | * the Host handler has already been run. We just do a | |
334 | * friendly check if another process should now be run, then | |
335 | * return to run the Guest again */ | |
625efab1 | 336 | cond_resched(); |
cc6d4fbc RR |
337 | return; |
338 | case LGUEST_TRAP_ENTRY: | |
b410e7b1 JS |
339 | /* Our 'struct hcall_args' maps directly over our regs: we set |
340 | * up the pointer now to indicate a hypercall is pending. */ | |
a53a35a8 | 341 | cpu->hcall = (struct hcall_args *)cpu->regs; |
625efab1 JS |
342 | return; |
343 | } | |
344 | ||
345 | /* We didn't handle the trap, so it needs to go to the Guest. */ | |
a53a35a8 | 346 | if (!deliver_trap(cpu, cpu->regs->trapnum)) |
625efab1 JS |
347 | /* If the Guest doesn't have a handler (either it hasn't |
348 | * registered any yet, or it's one of the faults we don't let | |
349 | * it handle), it dies with a cryptic error message. */ | |
350 | kill_guest(lg, "unhandled trap %li at %#lx (%#lx)", | |
a53a35a8 | 351 | cpu->regs->trapnum, cpu->regs->eip, |
fc708b3e | 352 | cpu->regs->trapnum == 14 ? cpu->arch.last_pagefault |
a53a35a8 | 353 | : cpu->regs->errcode); |
625efab1 JS |
354 | } |
355 | ||
356 | /* Now we can look at each of the routines this calls, in increasing order of | |
357 | * complexity: do_hypercalls(), emulate_insn(), maybe_do_interrupt(), | |
358 | * deliver_trap() and demand_page(). After all those, we'll be ready to | |
359 | * examine the Switcher, and our philosophical understanding of the Host/Guest | |
360 | * duality will be complete. :*/ | |
361 | static void adjust_pge(void *on) | |
362 | { | |
363 | if (on) | |
364 | write_cr4(read_cr4() | X86_CR4_PGE); | |
365 | else | |
366 | write_cr4(read_cr4() & ~X86_CR4_PGE); | |
367 | } | |
368 | ||
369 | /*H:020 Now the Switcher is mapped and every thing else is ready, we need to do | |
370 | * some more i386-specific initialization. */ | |
371 | void __init lguest_arch_host_init(void) | |
372 | { | |
373 | int i; | |
374 | ||
375 | /* Most of the i386/switcher.S doesn't care that it's been moved; on | |
376 | * Intel, jumps are relative, and it doesn't access any references to | |
377 | * external code or data. | |
378 | * | |
379 | * The only exception is the interrupt handlers in switcher.S: their | |
380 | * addresses are placed in a table (default_idt_entries), so we need to | |
381 | * update the table with the new addresses. switcher_offset() is a | |
382 | * convenience function which returns the distance between the builtin | |
383 | * switcher code and the high-mapped copy we just made. */ | |
384 | for (i = 0; i < IDT_ENTRIES; i++) | |
385 | default_idt_entries[i] += switcher_offset(); | |
386 | ||
387 | /* | |
388 | * Set up the Switcher's per-cpu areas. | |
389 | * | |
390 | * Each CPU gets two pages of its own within the high-mapped region | |
391 | * (aka. "struct lguest_pages"). Much of this can be initialized now, | |
392 | * but some depends on what Guest we are running (which is set up in | |
393 | * copy_in_guest_info()). | |
394 | */ | |
395 | for_each_possible_cpu(i) { | |
396 | /* lguest_pages() returns this CPU's two pages. */ | |
397 | struct lguest_pages *pages = lguest_pages(i); | |
398 | /* This is a convenience pointer to make the code fit one | |
399 | * statement to a line. */ | |
400 | struct lguest_ro_state *state = &pages->state; | |
401 | ||
402 | /* The Global Descriptor Table: the Host has a different one | |
403 | * for each CPU. We keep a descriptor for the GDT which says | |
404 | * where it is and how big it is (the size is actually the last | |
405 | * byte, not the size, hence the "-1"). */ | |
406 | state->host_gdt_desc.size = GDT_SIZE-1; | |
407 | state->host_gdt_desc.address = (long)get_cpu_gdt_table(i); | |
408 | ||
409 | /* All CPUs on the Host use the same Interrupt Descriptor | |
410 | * Table, so we just use store_idt(), which gets this CPU's IDT | |
411 | * descriptor. */ | |
412 | store_idt(&state->host_idt_desc); | |
413 | ||
414 | /* The descriptors for the Guest's GDT and IDT can be filled | |
415 | * out now, too. We copy the GDT & IDT into ->guest_gdt and | |
416 | * ->guest_idt before actually running the Guest. */ | |
417 | state->guest_idt_desc.size = sizeof(state->guest_idt)-1; | |
418 | state->guest_idt_desc.address = (long)&state->guest_idt; | |
419 | state->guest_gdt_desc.size = sizeof(state->guest_gdt)-1; | |
420 | state->guest_gdt_desc.address = (long)&state->guest_gdt; | |
421 | ||
422 | /* We know where we want the stack to be when the Guest enters | |
423 | * the switcher: in pages->regs. The stack grows upwards, so | |
424 | * we start it at the end of that structure. */ | |
425 | state->guest_tss.esp0 = (long)(&pages->regs + 1); | |
426 | /* And this is the GDT entry to use for the stack: we keep a | |
427 | * couple of special LGUEST entries. */ | |
428 | state->guest_tss.ss0 = LGUEST_DS; | |
429 | ||
430 | /* x86 can have a finegrained bitmap which indicates what I/O | |
431 | * ports the process can use. We set it to the end of our | |
432 | * structure, meaning "none". */ | |
433 | state->guest_tss.io_bitmap_base = sizeof(state->guest_tss); | |
434 | ||
435 | /* Some GDT entries are the same across all Guests, so we can | |
436 | * set them up now. */ | |
437 | setup_default_gdt_entries(state); | |
438 | /* Most IDT entries are the same for all Guests, too.*/ | |
439 | setup_default_idt_entries(state, default_idt_entries); | |
440 | ||
441 | /* The Host needs to be able to use the LGUEST segments on this | |
442 | * CPU, too, so put them in the Host GDT. */ | |
443 | get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT; | |
444 | get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT; | |
445 | } | |
446 | ||
447 | /* In the Switcher, we want the %cs segment register to use the | |
448 | * LGUEST_CS GDT entry: we've put that in the Host and Guest GDTs, so | |
449 | * it will be undisturbed when we switch. To change %cs and jump we | |
450 | * need this structure to feed to Intel's "lcall" instruction. */ | |
451 | lguest_entry.offset = (long)switch_to_guest + switcher_offset(); | |
452 | lguest_entry.segment = LGUEST_CS; | |
453 | ||
454 | /* Finally, we need to turn off "Page Global Enable". PGE is an | |
455 | * optimization where page table entries are specially marked to show | |
456 | * they never change. The Host kernel marks all the kernel pages this | |
457 | * way because it's always present, even when userspace is running. | |
458 | * | |
459 | * Lguest breaks this: unbeknownst to the rest of the Host kernel, we | |
460 | * switch to the Guest kernel. If you don't disable this on all CPUs, | |
461 | * you'll get really weird bugs that you'll chase for two days. | |
462 | * | |
463 | * I used to turn PGE off every time we switched to the Guest and back | |
464 | * on when we return, but that slowed the Switcher down noticibly. */ | |
465 | ||
466 | /* We don't need the complexity of CPUs coming and going while we're | |
467 | * doing this. */ | |
86ef5c9a | 468 | get_online_cpus(); |
625efab1 JS |
469 | if (cpu_has_pge) { /* We have a broader idea of "global". */ |
470 | /* Remember that this was originally set (for cleanup). */ | |
471 | cpu_had_pge = 1; | |
472 | /* adjust_pge is a helper function which sets or unsets the PGE | |
473 | * bit on its CPU, depending on the argument (0 == unset). */ | |
474 | on_each_cpu(adjust_pge, (void *)0, 0, 1); | |
475 | /* Turn off the feature in the global feature set. */ | |
476 | clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); | |
477 | } | |
86ef5c9a | 478 | put_online_cpus(); |
625efab1 JS |
479 | }; |
480 | /*:*/ | |
481 | ||
482 | void __exit lguest_arch_host_fini(void) | |
483 | { | |
484 | /* If we had PGE before we started, turn it back on now. */ | |
86ef5c9a | 485 | get_online_cpus(); |
625efab1 JS |
486 | if (cpu_had_pge) { |
487 | set_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); | |
488 | /* adjust_pge's argument "1" means set PGE. */ | |
489 | on_each_cpu(adjust_pge, (void *)1, 0, 1); | |
490 | } | |
86ef5c9a | 491 | put_online_cpus(); |
625efab1 | 492 | } |
b410e7b1 JS |
493 | |
494 | ||
495 | /*H:122 The i386-specific hypercalls simply farm out to the right functions. */ | |
73044f05 | 496 | int lguest_arch_do_hcall(struct lg_cpu *cpu, struct hcall_args *args) |
b410e7b1 JS |
497 | { |
498 | switch (args->arg0) { | |
499 | case LHCALL_LOAD_GDT: | |
fc708b3e | 500 | load_guest_gdt(cpu, args->arg1, args->arg2); |
b410e7b1 JS |
501 | break; |
502 | case LHCALL_LOAD_IDT_ENTRY: | |
fc708b3e | 503 | load_guest_idt_entry(cpu, args->arg1, args->arg2, args->arg3); |
b410e7b1 JS |
504 | break; |
505 | case LHCALL_LOAD_TLS: | |
fc708b3e | 506 | guest_load_tls(cpu, args->arg1); |
b410e7b1 JS |
507 | break; |
508 | default: | |
509 | /* Bad Guest. Bad! */ | |
510 | return -EIO; | |
511 | } | |
512 | return 0; | |
513 | } | |
514 | ||
515 | /*H:126 i386-specific hypercall initialization: */ | |
73044f05 | 516 | int lguest_arch_init_hypercalls(struct lg_cpu *cpu) |
b410e7b1 JS |
517 | { |
518 | u32 tsc_speed; | |
73044f05 | 519 | struct lguest *lg = cpu->lg; |
b410e7b1 JS |
520 | |
521 | /* The pointer to the Guest's "struct lguest_data" is the only | |
522 | * argument. We check that address now. */ | |
73044f05 | 523 | if (!lguest_address_ok(lg, cpu->hcall->arg1, sizeof(*lg->lguest_data))) |
b410e7b1 JS |
524 | return -EFAULT; |
525 | ||
526 | /* Having checked it, we simply set lg->lguest_data to point straight | |
527 | * into the Launcher's memory at the right place and then use | |
528 | * copy_to_user/from_user from now on, instead of lgread/write. I put | |
529 | * this in to show that I'm not immune to writing stupid | |
530 | * optimizations. */ | |
73044f05 | 531 | lg->lguest_data = lg->mem_base + cpu->hcall->arg1; |
b410e7b1 JS |
532 | |
533 | /* We insist that the Time Stamp Counter exist and doesn't change with | |
534 | * cpu frequency. Some devious chip manufacturers decided that TSC | |
535 | * changes could be handled in software. I decided that time going | |
536 | * backwards might be good for benchmarks, but it's bad for users. | |
537 | * | |
538 | * We also insist that the TSC be stable: the kernel detects unreliable | |
539 | * TSCs for its own purposes, and we use that here. */ | |
540 | if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) && !check_tsc_unstable()) | |
541 | tsc_speed = tsc_khz; | |
542 | else | |
543 | tsc_speed = 0; | |
544 | if (put_user(tsc_speed, &lg->lguest_data->tsc_khz)) | |
545 | return -EFAULT; | |
546 | ||
c18acd73 RR |
547 | /* The interrupt code might not like the system call vector. */ |
548 | if (!check_syscall_vector(lg)) | |
549 | kill_guest(lg, "bad syscall vector"); | |
550 | ||
b410e7b1 JS |
551 | return 0; |
552 | } | |
d612cde0 JS |
553 | |
554 | /*L:030 lguest_arch_setup_regs() | |
555 | * | |
556 | * Most of the Guest's registers are left alone: we used get_zeroed_page() to | |
557 | * allocate the structure, so they will be 0. */ | |
a53a35a8 | 558 | void lguest_arch_setup_regs(struct lg_cpu *cpu, unsigned long start) |
d612cde0 | 559 | { |
a53a35a8 | 560 | struct lguest_regs *regs = cpu->regs; |
d612cde0 JS |
561 | |
562 | /* There are four "segment" registers which the Guest needs to boot: | |
563 | * The "code segment" register (cs) refers to the kernel code segment | |
564 | * __KERNEL_CS, and the "data", "extra" and "stack" segment registers | |
565 | * refer to the kernel data segment __KERNEL_DS. | |
566 | * | |
567 | * The privilege level is packed into the lower bits. The Guest runs | |
568 | * at privilege level 1 (GUEST_PL).*/ | |
569 | regs->ds = regs->es = regs->ss = __KERNEL_DS|GUEST_PL; | |
570 | regs->cs = __KERNEL_CS|GUEST_PL; | |
571 | ||
572 | /* The "eflags" register contains miscellaneous flags. Bit 1 (0x002) | |
573 | * is supposed to always be "1". Bit 9 (0x200) controls whether | |
574 | * interrupts are enabled. We always leave interrupts enabled while | |
575 | * running the Guest. */ | |
25c47bb3 | 576 | regs->eflags = X86_EFLAGS_IF | 0x2; |
d612cde0 JS |
577 | |
578 | /* The "Extended Instruction Pointer" register says where the Guest is | |
579 | * running. */ | |
580 | regs->eip = start; | |
581 | ||
582 | /* %esi points to our boot information, at physical address 0, so don't | |
583 | * touch it. */ | |
e1e72965 | 584 | |
d612cde0 JS |
585 | /* There are a couple of GDT entries the Guest expects when first |
586 | * booting. */ | |
fc708b3e | 587 | setup_guest_gdt(cpu); |
d612cde0 | 588 | } |