blob: a6b717644be0d50efc2f257e52e9e543cc545833 [file] [log] [blame]
Jes Sorensen625efab2007-10-22 11:03:28 +10001/*
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 */
Rusty Russella6bd8e12008-03-28 11:05:53 -050020/*P:450 This file contains the x86-specific lguest code. It used to be all
21 * mixed in with drivers/lguest/core.c but several foolhardy code slashers
22 * wrestled most of the dependencies out to here in preparation for porting
23 * lguest to other architectures (see what I mean by foolhardy?).
24 *
25 * This also contains a couple of non-obvious setup and teardown pieces which
26 * were implemented after days of debugging pain. :*/
Jes Sorensen625efab2007-10-22 11:03:28 +100027#include <linux/kernel.h>
28#include <linux/start_kernel.h>
29#include <linux/string.h>
30#include <linux/console.h>
31#include <linux/screen_info.h>
32#include <linux/irq.h>
33#include <linux/interrupt.h>
34#include <linux/clocksource.h>
35#include <linux/clockchips.h>
36#include <linux/cpu.h>
37#include <linux/lguest.h>
38#include <linux/lguest_launcher.h>
Jes Sorensen625efab2007-10-22 11:03:28 +100039#include <asm/paravirt.h>
40#include <asm/param.h>
41#include <asm/page.h>
42#include <asm/pgtable.h>
43#include <asm/desc.h>
44#include <asm/setup.h>
45#include <asm/lguest.h>
46#include <asm/uaccess.h>
47#include <asm/i387.h>
48#include "../lg.h"
49
50static int cpu_had_pge;
51
52static struct {
53 unsigned long offset;
54 unsigned short segment;
55} lguest_entry;
56
57/* Offset from where switcher.S was compiled to where we've copied it */
58static unsigned long switcher_offset(void)
59{
60 return SWITCHER_ADDR - (unsigned long)start_switcher_text;
61}
62
63/* This cpu's struct lguest_pages. */
64static struct lguest_pages *lguest_pages(unsigned int cpu)
65{
66 return &(((struct lguest_pages *)
67 (SWITCHER_ADDR + SHARED_SWITCHER_PAGES*PAGE_SIZE))[cpu]);
68}
69
Glauber de Oliveira Costac40a9f42008-01-17 19:11:20 -020070static DEFINE_PER_CPU(struct lg_cpu *, last_cpu);
Jes Sorensen625efab2007-10-22 11:03:28 +100071
72/*S:010
Rusty Russelle1e72962007-10-25 15:02:50 +100073 * We approach the Switcher.
Jes Sorensen625efab2007-10-22 11:03:28 +100074 *
75 * Remember that each CPU has two pages which are visible to the Guest when it
76 * runs on that CPU. This has to contain the state for that Guest: we copy the
77 * state in just before we run the Guest.
78 *
79 * Each Guest has "changed" flags which indicate what has changed in the Guest
80 * since it last ran. We saw this set in interrupts_and_traps.c and
81 * segments.c.
82 */
Glauber de Oliveira Costad0953d42008-01-07 11:05:25 -020083static void copy_in_guest_info(struct lg_cpu *cpu, struct lguest_pages *pages)
Jes Sorensen625efab2007-10-22 11:03:28 +100084{
85 /* Copying all this data can be quite expensive. We usually run the
86 * same Guest we ran last time (and that Guest hasn't run anywhere else
87 * meanwhile). If that's not the case, we pretend everything in the
88 * Guest has changed. */
Glauber de Oliveira Costaf34f8c52008-01-17 19:13:26 -020089 if (__get_cpu_var(last_cpu) != cpu || cpu->last_pages != pages) {
Glauber de Oliveira Costac40a9f42008-01-17 19:11:20 -020090 __get_cpu_var(last_cpu) = cpu;
Glauber de Oliveira Costaf34f8c52008-01-17 19:13:26 -020091 cpu->last_pages = pages;
Glauber de Oliveira Costaae3749d2008-01-17 19:14:46 -020092 cpu->changed = CHANGED_ALL;
Jes Sorensen625efab2007-10-22 11:03:28 +100093 }
94
95 /* These copies are pretty cheap, so we do them unconditionally: */
96 /* Save the current Host top-level page directory. */
97 pages->state.host_cr3 = __pa(current->mm->pgd);
98 /* Set up the Guest's page tables to see this CPU's pages (and no
99 * other CPU's pages). */
Glauber de Oliveira Costa0c784412008-01-07 11:05:30 -0200100 map_switcher_in_guest(cpu, pages);
Jes Sorensen625efab2007-10-22 11:03:28 +1000101 /* Set up the two "TSS" members which tell the CPU what stack to use
102 * for traps which do directly into the Guest (ie. traps at privilege
103 * level 1). */
Rusty Russelle95035c2008-01-31 18:00:47 +1100104 pages->state.guest_tss.sp1 = cpu->esp1;
Glauber de Oliveira Costa4665ac82008-01-07 11:05:35 -0200105 pages->state.guest_tss.ss1 = cpu->ss1;
Jes Sorensen625efab2007-10-22 11:03:28 +1000106
107 /* Copy direct-to-Guest trap entries. */
Glauber de Oliveira Costaae3749d2008-01-17 19:14:46 -0200108 if (cpu->changed & CHANGED_IDT)
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200109 copy_traps(cpu, pages->state.guest_idt, default_idt_entries);
Jes Sorensen625efab2007-10-22 11:03:28 +1000110
111 /* Copy all GDT entries which the Guest can change. */
Glauber de Oliveira Costaae3749d2008-01-17 19:14:46 -0200112 if (cpu->changed & CHANGED_GDT)
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200113 copy_gdt(cpu, pages->state.guest_gdt);
Jes Sorensen625efab2007-10-22 11:03:28 +1000114 /* If only the TLS entries have changed, copy them. */
Glauber de Oliveira Costaae3749d2008-01-17 19:14:46 -0200115 else if (cpu->changed & CHANGED_GDT_TLS)
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200116 copy_gdt_tls(cpu, pages->state.guest_gdt);
Jes Sorensen625efab2007-10-22 11:03:28 +1000117
118 /* Mark the Guest as unchanged for next time. */
Glauber de Oliveira Costaae3749d2008-01-17 19:14:46 -0200119 cpu->changed = 0;
Jes Sorensen625efab2007-10-22 11:03:28 +1000120}
121
122/* Finally: the code to actually call into the Switcher to run the Guest. */
Glauber de Oliveira Costad0953d42008-01-07 11:05:25 -0200123static void run_guest_once(struct lg_cpu *cpu, struct lguest_pages *pages)
Jes Sorensen625efab2007-10-22 11:03:28 +1000124{
125 /* This is a dummy value we need for GCC's sake. */
126 unsigned int clobber;
127
128 /* Copy the guest-specific information into this CPU's "struct
129 * lguest_pages". */
Glauber de Oliveira Costad0953d42008-01-07 11:05:25 -0200130 copy_in_guest_info(cpu, pages);
Jes Sorensen625efab2007-10-22 11:03:28 +1000131
132 /* Set the trap number to 256 (impossible value). If we fault while
133 * switching to the Guest (bad segment registers or bug), this will
134 * cause us to abort the Guest. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200135 cpu->regs->trapnum = 256;
Jes Sorensen625efab2007-10-22 11:03:28 +1000136
137 /* Now: we push the "eflags" register on the stack, then do an "lcall".
138 * This is how we change from using the kernel code segment to using
139 * the dedicated lguest code segment, as well as jumping into the
140 * Switcher.
141 *
142 * The lcall also pushes the old code segment (KERNEL_CS) onto the
143 * stack, then the address of this call. This stack layout happens to
Rusty Russelle1e72962007-10-25 15:02:50 +1000144 * exactly match the stack layout created by an interrupt... */
Jes Sorensen625efab2007-10-22 11:03:28 +1000145 asm volatile("pushf; lcall *lguest_entry"
146 /* This is how we tell GCC that %eax ("a") and %ebx ("b")
147 * are changed by this routine. The "=" means output. */
148 : "=a"(clobber), "=b"(clobber)
149 /* %eax contains the pages pointer. ("0" refers to the
150 * 0-th argument above, ie "a"). %ebx contains the
151 * physical address of the Guest's top-level page
152 * directory. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200153 : "0"(pages), "1"(__pa(cpu->lg->pgdirs[cpu->cpu_pgd].pgdir))
Jes Sorensen625efab2007-10-22 11:03:28 +1000154 /* We tell gcc that all these registers could change,
155 * which means we don't have to save and restore them in
156 * the Switcher. */
157 : "memory", "%edx", "%ecx", "%edi", "%esi");
158}
159/*:*/
160
Rusty Russelle1e72962007-10-25 15:02:50 +1000161/*M:002 There are hooks in the scheduler which we can register to tell when we
162 * get kicked off the CPU (preempt_notifier_register()). This would allow us
163 * to lazily disable SYSENTER which would regain some performance, and should
164 * also simplify copy_in_guest_info(). Note that we'd still need to restore
165 * things when we exit to Launcher userspace, but that's fairly easy.
166 *
Rusty Russella6bd8e12008-03-28 11:05:53 -0500167 * We could also try using this hooks for PGE, but that might be too expensive.
168 *
Rusty Russelle1e72962007-10-25 15:02:50 +1000169 * The hooks were designed for KVM, but we can also put them to good use. :*/
170
Jes Sorensen625efab2007-10-22 11:03:28 +1000171/*H:040 This is the i386-specific code to setup and run the Guest. Interrupts
172 * are disabled: we own the CPU. */
Glauber de Oliveira Costad0953d42008-01-07 11:05:25 -0200173void lguest_arch_run_guest(struct lg_cpu *cpu)
Jes Sorensen625efab2007-10-22 11:03:28 +1000174{
Rusty Russelle1e72962007-10-25 15:02:50 +1000175 /* Remember the awfully-named TS bit? If the Guest has asked to set it
176 * we set it now, so we can trap and pass that trap to the Guest if it
177 * uses the FPU. */
Glauber de Oliveira Costa4665ac82008-01-07 11:05:35 -0200178 if (cpu->ts)
Suresh Siddha54481cf82008-06-19 09:41:22 -0700179 unlazy_fpu(current);
Jes Sorensen625efab2007-10-22 11:03:28 +1000180
Rusty Russelle1e72962007-10-25 15:02:50 +1000181 /* SYSENTER is an optimized way of doing system calls. We can't allow
182 * it because it always jumps to privilege level 0. A normal Guest
183 * won't try it because we don't advertise it in CPUID, but a malicious
184 * Guest (or malicious Guest userspace program) could, so we tell the
185 * CPU to disable it before running the Guest. */
Jes Sorensen625efab2007-10-22 11:03:28 +1000186 if (boot_cpu_has(X86_FEATURE_SEP))
187 wrmsr(MSR_IA32_SYSENTER_CS, 0, 0);
188
Rusty Russelle1e72962007-10-25 15:02:50 +1000189 /* Now we actually run the Guest. It will return when something
190 * interesting happens, and we can examine its registers to see what it
191 * was doing. */
Glauber de Oliveira Costad0953d42008-01-07 11:05:25 -0200192 run_guest_once(cpu, lguest_pages(raw_smp_processor_id()));
Jes Sorensen625efab2007-10-22 11:03:28 +1000193
Rusty Russella6bd8e12008-03-28 11:05:53 -0500194 /* Note that the "regs" structure contains two extra entries which are
Rusty Russelle1e72962007-10-25 15:02:50 +1000195 * not really registers: a trap number which says what interrupt or
196 * trap made the switcher code come back, and an error code which some
197 * traps set. */
Jes Sorensen625efab2007-10-22 11:03:28 +1000198
Suresh Siddha54481cf82008-06-19 09:41:22 -0700199 /* Restore SYSENTER if it's supposed to be on. */
200 if (boot_cpu_has(X86_FEATURE_SEP))
201 wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0);
202
Rusty Russelle1e72962007-10-25 15:02:50 +1000203 /* If the Guest page faulted, then the cr2 register will tell us the
204 * bad virtual address. We have to grab this now, because once we
205 * re-enable interrupts an interrupt could fault and thus overwrite
206 * cr2, or we could even move off to a different CPU. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200207 if (cpu->regs->trapnum == 14)
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200208 cpu->arch.last_pagefault = read_cr2();
Jes Sorensen625efab2007-10-22 11:03:28 +1000209 /* Similarly, if we took a trap because the Guest used the FPU,
Suresh Siddha54481cf82008-06-19 09:41:22 -0700210 * we have to restore the FPU it expects to see.
211 * math_state_restore() may sleep and we may even move off to
212 * a different CPU. So all the critical stuff should be done
213 * before this. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200214 else if (cpu->regs->trapnum == 7)
Jes Sorensen625efab2007-10-22 11:03:28 +1000215 math_state_restore();
Jes Sorensen625efab2007-10-22 11:03:28 +1000216}
217
Rusty Russelle1e72962007-10-25 15:02:50 +1000218/*H:130 Now we've examined the hypercall code; our Guest can make requests.
219 * Our Guest is usually so well behaved; it never tries to do things it isn't
220 * allowed to, and uses hypercalls instead. Unfortunately, Linux's paravirtual
221 * infrastructure isn't quite complete, because it doesn't contain replacements
222 * for the Intel I/O instructions. As a result, the Guest sometimes fumbles
223 * across one during the boot process as it probes for various things which are
224 * usually attached to a PC.
Jes Sorensen625efab2007-10-22 11:03:28 +1000225 *
Rusty Russelle1e72962007-10-25 15:02:50 +1000226 * When the Guest uses one of these instructions, we get a trap (General
Jes Sorensen625efab2007-10-22 11:03:28 +1000227 * Protection Fault) and come here. We see if it's one of those troublesome
228 * instructions and skip over it. We return true if we did. */
Glauber de Oliveira Costaa3863f62008-01-07 11:05:31 -0200229static int emulate_insn(struct lg_cpu *cpu)
Jes Sorensen625efab2007-10-22 11:03:28 +1000230{
231 u8 insn;
232 unsigned int insnlen = 0, in = 0, shift = 0;
233 /* The eip contains the *virtual* address of the Guest's instruction:
234 * guest_pa just subtracts the Guest's page_offset. */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200235 unsigned long physaddr = guest_pa(cpu, cpu->regs->eip);
Jes Sorensen625efab2007-10-22 11:03:28 +1000236
Rusty Russell47436aa2007-10-22 11:03:36 +1000237 /* This must be the Guest kernel trying to do something, not userspace!
238 * The bottom two bits of the CS segment register are the privilege
239 * level. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200240 if ((cpu->regs->cs & 3) != GUEST_PL)
Jes Sorensen625efab2007-10-22 11:03:28 +1000241 return 0;
242
243 /* Decoding x86 instructions is icky. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200244 insn = lgread(cpu, physaddr, u8);
Jes Sorensen625efab2007-10-22 11:03:28 +1000245
246 /* 0x66 is an "operand prefix". It means it's using the upper 16 bits
247 of the eax register. */
248 if (insn == 0x66) {
249 shift = 16;
250 /* The instruction is 1 byte so far, read the next byte. */
251 insnlen = 1;
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200252 insn = lgread(cpu, physaddr + insnlen, u8);
Jes Sorensen625efab2007-10-22 11:03:28 +1000253 }
254
255 /* We can ignore the lower bit for the moment and decode the 4 opcodes
256 * we need to emulate. */
257 switch (insn & 0xFE) {
258 case 0xE4: /* in <next byte>,%al */
259 insnlen += 2;
260 in = 1;
261 break;
262 case 0xEC: /* in (%dx),%al */
263 insnlen += 1;
264 in = 1;
265 break;
266 case 0xE6: /* out %al,<next byte> */
267 insnlen += 2;
268 break;
269 case 0xEE: /* out %al,(%dx) */
270 insnlen += 1;
271 break;
272 default:
273 /* OK, we don't know what this is, can't emulate. */
274 return 0;
275 }
276
277 /* If it was an "IN" instruction, they expect the result to be read
278 * into %eax, so we change %eax. We always return all-ones, which
279 * traditionally means "there's nothing there". */
280 if (in) {
281 /* Lower bit tells is whether it's a 16 or 32 bit access */
282 if (insn & 0x1)
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200283 cpu->regs->eax = 0xFFFFFFFF;
Jes Sorensen625efab2007-10-22 11:03:28 +1000284 else
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200285 cpu->regs->eax |= (0xFFFF << shift);
Jes Sorensen625efab2007-10-22 11:03:28 +1000286 }
287 /* Finally, we've "done" the instruction, so move past it. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200288 cpu->regs->eip += insnlen;
Jes Sorensen625efab2007-10-22 11:03:28 +1000289 /* Success! */
290 return 1;
291}
292
Matias Zabaljauregui4cd8b5e2009-03-14 13:37:52 -0200293/* Our hypercalls mechanism used to be based on direct software interrupts.
294 * After Anthony's "Refactor hypercall infrastructure" kvm patch, we decided to
295 * change over to using kvm hypercalls.
296 *
297 * KVM_HYPERCALL is actually a "vmcall" instruction, which generates an invalid
298 * opcode fault (fault 6) on non-VT cpus, so the easiest solution seemed to be
299 * an *emulation approach*: if the fault was really produced by an hypercall
300 * (is_hypercall() does exactly this check), we can just call the corresponding
301 * hypercall host implementation function.
302 *
303 * But these invalid opcode faults are notably slower than software interrupts.
304 * So we implemented the *patching (or rewriting) approach*: every time we hit
305 * the KVM_HYPERCALL opcode in Guest code, we patch it to the old "int 0x1f"
306 * opcode, so next time the Guest calls this hypercall it will use the
307 * faster trap mechanism.
308 *
309 * Matias even benchmarked it to convince you: this shows the average cycle
310 * cost of a hypercall. For each alternative solution mentioned above we've
311 * made 5 runs of the benchmark:
312 *
313 * 1) direct software interrupt: 2915, 2789, 2764, 2721, 2898
314 * 2) emulation technique: 3410, 3681, 3466, 3392, 3780
315 * 3) patching (rewrite) technique: 2977, 2975, 2891, 2637, 2884
316 *
317 * One two-line function is worth a 20% hypercall speed boost!
318 */
319static void rewrite_hypercall(struct lg_cpu *cpu)
320{
321 /* This are the opcodes we use to patch the Guest. The opcode for "int
322 * $0x1f" is "0xcd 0x1f" but vmcall instruction is 3 bytes long, so we
323 * complete the sequence with a NOP (0x90). */
324 u8 insn[3] = {0xcd, 0x1f, 0x90};
325
326 __lgwrite(cpu, guest_pa(cpu, cpu->regs->eip), insn, sizeof(insn));
327}
328
329static bool is_hypercall(struct lg_cpu *cpu)
330{
331 u8 insn[3];
332
333 /* This must be the Guest kernel trying to do something.
334 * The bottom two bits of the CS segment register are the privilege
335 * level. */
336 if ((cpu->regs->cs & 3) != GUEST_PL)
337 return false;
338
339 /* Is it a vmcall? */
340 __lgread(cpu, insn, guest_pa(cpu, cpu->regs->eip), sizeof(insn));
341 return insn[0] == 0x0f && insn[1] == 0x01 && insn[2] == 0xc1;
342}
343
Jes Sorensen625efab2007-10-22 11:03:28 +1000344/*H:050 Once we've re-enabled interrupts, we look at why the Guest exited. */
Glauber de Oliveira Costa73044f02008-01-07 11:05:27 -0200345void lguest_arch_handle_trap(struct lg_cpu *cpu)
Jes Sorensen625efab2007-10-22 11:03:28 +1000346{
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200347 switch (cpu->regs->trapnum) {
Rusty Russelle1e72962007-10-25 15:02:50 +1000348 case 13: /* We've intercepted a General Protection Fault. */
349 /* Check if this was one of those annoying IN or OUT
350 * instructions which we need to emulate. If so, we just go
351 * back into the Guest after we've done it. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200352 if (cpu->regs->errcode == 0) {
Glauber de Oliveira Costaa3863f62008-01-07 11:05:31 -0200353 if (emulate_insn(cpu))
Jes Sorensen625efab2007-10-22 11:03:28 +1000354 return;
355 }
356 break;
Rusty Russelle1e72962007-10-25 15:02:50 +1000357 case 14: /* We've intercepted a Page Fault. */
358 /* The Guest accessed a virtual address that wasn't mapped.
Rusty Russella6bd8e12008-03-28 11:05:53 -0500359 * This happens a lot: we don't actually set up most of the page
360 * tables for the Guest at all when we start: as it runs it asks
361 * for more and more, and we set them up as required. In this
362 * case, we don't even tell the Guest that the fault happened.
Rusty Russelle1e72962007-10-25 15:02:50 +1000363 *
364 * The errcode tells whether this was a read or a write, and
365 * whether kernel or userspace code. */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200366 if (demand_page(cpu, cpu->arch.last_pagefault,
367 cpu->regs->errcode))
Jes Sorensen625efab2007-10-22 11:03:28 +1000368 return;
369
Rusty Russelle1e72962007-10-25 15:02:50 +1000370 /* OK, it's really not there (or not OK): the Guest needs to
371 * know. We write out the cr2 value so it knows where the
372 * fault occurred.
373 *
374 * Note that if the Guest were really messed up, this could
375 * happen before it's done the LHCALL_LGUEST_INIT hypercall, so
376 * lg->lguest_data could be NULL */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200377 if (cpu->lg->lguest_data &&
378 put_user(cpu->arch.last_pagefault,
379 &cpu->lg->lguest_data->cr2))
380 kill_guest(cpu, "Writing cr2");
Jes Sorensen625efab2007-10-22 11:03:28 +1000381 break;
382 case 7: /* We've intercepted a Device Not Available fault. */
Rusty Russelle1e72962007-10-25 15:02:50 +1000383 /* If the Guest doesn't want to know, we already restored the
384 * Floating Point Unit, so we just continue without telling
385 * it. */
Glauber de Oliveira Costa4665ac82008-01-07 11:05:35 -0200386 if (!cpu->ts)
Jes Sorensen625efab2007-10-22 11:03:28 +1000387 return;
388 break;
389 case 32 ... 255:
Rusty Russellcc6d4fb2007-10-22 11:03:30 +1000390 /* These values mean a real interrupt occurred, in which case
Matias Zabaljauregui4cd8b5e2009-03-14 13:37:52 -0200391 * the Host handler has already been run. We just do a
Rusty Russellcc6d4fb2007-10-22 11:03:30 +1000392 * friendly check if another process should now be run, then
393 * return to run the Guest again */
Jes Sorensen625efab2007-10-22 11:03:28 +1000394 cond_resched();
Rusty Russellcc6d4fb2007-10-22 11:03:30 +1000395 return;
396 case LGUEST_TRAP_ENTRY:
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000397 /* Our 'struct hcall_args' maps directly over our regs: we set
398 * up the pointer now to indicate a hypercall is pending. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200399 cpu->hcall = (struct hcall_args *)cpu->regs;
Jes Sorensen625efab2007-10-22 11:03:28 +1000400 return;
Matias Zabaljauregui4cd8b5e2009-03-14 13:37:52 -0200401 case 6:
402 /* kvm hypercalls trigger an invalid opcode fault (6).
403 * We need to check if ring == GUEST_PL and
404 * faulting instruction == vmcall. */
405 if (is_hypercall(cpu)) {
406 rewrite_hypercall(cpu);
407 return;
408 }
409 break;
Jes Sorensen625efab2007-10-22 11:03:28 +1000410 }
411
412 /* We didn't handle the trap, so it needs to go to the Guest. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200413 if (!deliver_trap(cpu, cpu->regs->trapnum))
Jes Sorensen625efab2007-10-22 11:03:28 +1000414 /* If the Guest doesn't have a handler (either it hasn't
415 * registered any yet, or it's one of the faults we don't let
Rusty Russella6bd8e12008-03-28 11:05:53 -0500416 * it handle), it dies with this cryptic error message. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200417 kill_guest(cpu, "unhandled trap %li at %#lx (%#lx)",
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200418 cpu->regs->trapnum, cpu->regs->eip,
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200419 cpu->regs->trapnum == 14 ? cpu->arch.last_pagefault
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200420 : cpu->regs->errcode);
Jes Sorensen625efab2007-10-22 11:03:28 +1000421}
422
423/* Now we can look at each of the routines this calls, in increasing order of
424 * complexity: do_hypercalls(), emulate_insn(), maybe_do_interrupt(),
425 * deliver_trap() and demand_page(). After all those, we'll be ready to
426 * examine the Switcher, and our philosophical understanding of the Host/Guest
427 * duality will be complete. :*/
428static void adjust_pge(void *on)
429{
430 if (on)
431 write_cr4(read_cr4() | X86_CR4_PGE);
432 else
433 write_cr4(read_cr4() & ~X86_CR4_PGE);
434}
435
436/*H:020 Now the Switcher is mapped and every thing else is ready, we need to do
437 * some more i386-specific initialization. */
438void __init lguest_arch_host_init(void)
439{
440 int i;
441
442 /* Most of the i386/switcher.S doesn't care that it's been moved; on
443 * Intel, jumps are relative, and it doesn't access any references to
444 * external code or data.
445 *
446 * The only exception is the interrupt handlers in switcher.S: their
447 * addresses are placed in a table (default_idt_entries), so we need to
448 * update the table with the new addresses. switcher_offset() is a
Rusty Russella6bd8e12008-03-28 11:05:53 -0500449 * convenience function which returns the distance between the
450 * compiled-in switcher code and the high-mapped copy we just made. */
Jes Sorensen625efab2007-10-22 11:03:28 +1000451 for (i = 0; i < IDT_ENTRIES; i++)
452 default_idt_entries[i] += switcher_offset();
453
454 /*
455 * Set up the Switcher's per-cpu areas.
456 *
457 * Each CPU gets two pages of its own within the high-mapped region
458 * (aka. "struct lguest_pages"). Much of this can be initialized now,
459 * but some depends on what Guest we are running (which is set up in
460 * copy_in_guest_info()).
461 */
462 for_each_possible_cpu(i) {
463 /* lguest_pages() returns this CPU's two pages. */
464 struct lguest_pages *pages = lguest_pages(i);
465 /* This is a convenience pointer to make the code fit one
466 * statement to a line. */
467 struct lguest_ro_state *state = &pages->state;
468
469 /* The Global Descriptor Table: the Host has a different one
470 * for each CPU. We keep a descriptor for the GDT which says
471 * where it is and how big it is (the size is actually the last
472 * byte, not the size, hence the "-1"). */
473 state->host_gdt_desc.size = GDT_SIZE-1;
474 state->host_gdt_desc.address = (long)get_cpu_gdt_table(i);
475
476 /* All CPUs on the Host use the same Interrupt Descriptor
477 * Table, so we just use store_idt(), which gets this CPU's IDT
478 * descriptor. */
479 store_idt(&state->host_idt_desc);
480
481 /* The descriptors for the Guest's GDT and IDT can be filled
482 * out now, too. We copy the GDT & IDT into ->guest_gdt and
483 * ->guest_idt before actually running the Guest. */
484 state->guest_idt_desc.size = sizeof(state->guest_idt)-1;
485 state->guest_idt_desc.address = (long)&state->guest_idt;
486 state->guest_gdt_desc.size = sizeof(state->guest_gdt)-1;
487 state->guest_gdt_desc.address = (long)&state->guest_gdt;
488
489 /* We know where we want the stack to be when the Guest enters
Rusty Russella6bd8e12008-03-28 11:05:53 -0500490 * the Switcher: in pages->regs. The stack grows upwards, so
Jes Sorensen625efab2007-10-22 11:03:28 +1000491 * we start it at the end of that structure. */
H. Peter Anvinfaca6222008-01-30 13:31:02 +0100492 state->guest_tss.sp0 = (long)(&pages->regs + 1);
Jes Sorensen625efab2007-10-22 11:03:28 +1000493 /* And this is the GDT entry to use for the stack: we keep a
494 * couple of special LGUEST entries. */
495 state->guest_tss.ss0 = LGUEST_DS;
496
497 /* x86 can have a finegrained bitmap which indicates what I/O
498 * ports the process can use. We set it to the end of our
499 * structure, meaning "none". */
500 state->guest_tss.io_bitmap_base = sizeof(state->guest_tss);
501
502 /* Some GDT entries are the same across all Guests, so we can
503 * set them up now. */
504 setup_default_gdt_entries(state);
505 /* Most IDT entries are the same for all Guests, too.*/
506 setup_default_idt_entries(state, default_idt_entries);
507
508 /* The Host needs to be able to use the LGUEST segments on this
509 * CPU, too, so put them in the Host GDT. */
510 get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT;
511 get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT;
512 }
513
514 /* In the Switcher, we want the %cs segment register to use the
515 * LGUEST_CS GDT entry: we've put that in the Host and Guest GDTs, so
516 * it will be undisturbed when we switch. To change %cs and jump we
517 * need this structure to feed to Intel's "lcall" instruction. */
518 lguest_entry.offset = (long)switch_to_guest + switcher_offset();
519 lguest_entry.segment = LGUEST_CS;
520
521 /* Finally, we need to turn off "Page Global Enable". PGE is an
522 * optimization where page table entries are specially marked to show
523 * they never change. The Host kernel marks all the kernel pages this
524 * way because it's always present, even when userspace is running.
525 *
526 * Lguest breaks this: unbeknownst to the rest of the Host kernel, we
527 * switch to the Guest kernel. If you don't disable this on all CPUs,
528 * you'll get really weird bugs that you'll chase for two days.
529 *
530 * I used to turn PGE off every time we switched to the Guest and back
531 * on when we return, but that slowed the Switcher down noticibly. */
532
533 /* We don't need the complexity of CPUs coming and going while we're
534 * doing this. */
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +0100535 get_online_cpus();
Jes Sorensen625efab2007-10-22 11:03:28 +1000536 if (cpu_has_pge) { /* We have a broader idea of "global". */
537 /* Remember that this was originally set (for cleanup). */
538 cpu_had_pge = 1;
539 /* adjust_pge is a helper function which sets or unsets the PGE
540 * bit on its CPU, depending on the argument (0 == unset). */
Jens Axboe15c8b6c2008-05-09 09:39:44 +0200541 on_each_cpu(adjust_pge, (void *)0, 1);
Jes Sorensen625efab2007-10-22 11:03:28 +1000542 /* Turn off the feature in the global feature set. */
Andrew Mortoncf485e52008-06-09 16:22:48 -0700543 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_PGE);
Jes Sorensen625efab2007-10-22 11:03:28 +1000544 }
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +0100545 put_online_cpus();
Jes Sorensen625efab2007-10-22 11:03:28 +1000546};
547/*:*/
548
549void __exit lguest_arch_host_fini(void)
550{
551 /* If we had PGE before we started, turn it back on now. */
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +0100552 get_online_cpus();
Jes Sorensen625efab2007-10-22 11:03:28 +1000553 if (cpu_had_pge) {
Andrew Mortoncf485e52008-06-09 16:22:48 -0700554 set_cpu_cap(&boot_cpu_data, X86_FEATURE_PGE);
Jes Sorensen625efab2007-10-22 11:03:28 +1000555 /* adjust_pge's argument "1" means set PGE. */
Jens Axboe15c8b6c2008-05-09 09:39:44 +0200556 on_each_cpu(adjust_pge, (void *)1, 1);
Jes Sorensen625efab2007-10-22 11:03:28 +1000557 }
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +0100558 put_online_cpus();
Jes Sorensen625efab2007-10-22 11:03:28 +1000559}
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000560
561
562/*H:122 The i386-specific hypercalls simply farm out to the right functions. */
Glauber de Oliveira Costa73044f02008-01-07 11:05:27 -0200563int lguest_arch_do_hcall(struct lg_cpu *cpu, struct hcall_args *args)
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000564{
565 switch (args->arg0) {
566 case LHCALL_LOAD_GDT:
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200567 load_guest_gdt(cpu, args->arg1, args->arg2);
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000568 break;
569 case LHCALL_LOAD_IDT_ENTRY:
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200570 load_guest_idt_entry(cpu, args->arg1, args->arg2, args->arg3);
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000571 break;
572 case LHCALL_LOAD_TLS:
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200573 guest_load_tls(cpu, args->arg1);
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000574 break;
575 default:
576 /* Bad Guest. Bad! */
577 return -EIO;
578 }
579 return 0;
580}
581
582/*H:126 i386-specific hypercall initialization: */
Glauber de Oliveira Costa73044f02008-01-07 11:05:27 -0200583int lguest_arch_init_hypercalls(struct lg_cpu *cpu)
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000584{
585 u32 tsc_speed;
586
Rusty Russella6bd8e12008-03-28 11:05:53 -0500587 /* The pointer to the Guest's "struct lguest_data" is the only argument.
588 * We check that address now. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200589 if (!lguest_address_ok(cpu->lg, cpu->hcall->arg1,
590 sizeof(*cpu->lg->lguest_data)))
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000591 return -EFAULT;
592
593 /* Having checked it, we simply set lg->lguest_data to point straight
594 * into the Launcher's memory at the right place and then use
595 * copy_to_user/from_user from now on, instead of lgread/write. I put
596 * this in to show that I'm not immune to writing stupid
597 * optimizations. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200598 cpu->lg->lguest_data = cpu->lg->mem_base + cpu->hcall->arg1;
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000599
600 /* We insist that the Time Stamp Counter exist and doesn't change with
601 * cpu frequency. Some devious chip manufacturers decided that TSC
602 * changes could be handled in software. I decided that time going
603 * backwards might be good for benchmarks, but it's bad for users.
604 *
605 * We also insist that the TSC be stable: the kernel detects unreliable
606 * TSCs for its own purposes, and we use that here. */
607 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) && !check_tsc_unstable())
608 tsc_speed = tsc_khz;
609 else
610 tsc_speed = 0;
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200611 if (put_user(tsc_speed, &cpu->lg->lguest_data->tsc_khz))
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000612 return -EFAULT;
613
Rusty Russellc18acd72007-10-22 11:03:35 +1000614 /* The interrupt code might not like the system call vector. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200615 if (!check_syscall_vector(cpu->lg))
616 kill_guest(cpu, "bad syscall vector");
Rusty Russellc18acd72007-10-22 11:03:35 +1000617
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000618 return 0;
619}
Rusty Russella6bd8e12008-03-28 11:05:53 -0500620/*:*/
Jes Sorensend612cde2007-10-22 11:03:32 +1000621
622/*L:030 lguest_arch_setup_regs()
623 *
624 * Most of the Guest's registers are left alone: we used get_zeroed_page() to
625 * allocate the structure, so they will be 0. */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200626void lguest_arch_setup_regs(struct lg_cpu *cpu, unsigned long start)
Jes Sorensend612cde2007-10-22 11:03:32 +1000627{
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -0200628 struct lguest_regs *regs = cpu->regs;
Jes Sorensend612cde2007-10-22 11:03:32 +1000629
630 /* There are four "segment" registers which the Guest needs to boot:
631 * The "code segment" register (cs) refers to the kernel code segment
632 * __KERNEL_CS, and the "data", "extra" and "stack" segment registers
633 * refer to the kernel data segment __KERNEL_DS.
634 *
635 * The privilege level is packed into the lower bits. The Guest runs
636 * at privilege level 1 (GUEST_PL).*/
637 regs->ds = regs->es = regs->ss = __KERNEL_DS|GUEST_PL;
638 regs->cs = __KERNEL_CS|GUEST_PL;
639
640 /* The "eflags" register contains miscellaneous flags. Bit 1 (0x002)
641 * is supposed to always be "1". Bit 9 (0x200) controls whether
642 * interrupts are enabled. We always leave interrupts enabled while
643 * running the Guest. */
Rusty Russell25c47bb2007-10-25 14:09:53 +1000644 regs->eflags = X86_EFLAGS_IF | 0x2;
Jes Sorensend612cde2007-10-22 11:03:32 +1000645
646 /* The "Extended Instruction Pointer" register says where the Guest is
647 * running. */
648 regs->eip = start;
649
650 /* %esi points to our boot information, at physical address 0, so don't
651 * touch it. */
Rusty Russelle1e72962007-10-25 15:02:50 +1000652
Jes Sorensend612cde2007-10-22 11:03:32 +1000653 /* There are a couple of GDT entries the Guest expects when first
654 * booting. */
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200655 setup_guest_gdt(cpu);
Jes Sorensend612cde2007-10-22 11:03:32 +1000656}