2 * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation.
3 * Copyright (C) 2007, Jes Sorensen <jes@sgi.com> SGI.
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.
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
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.
20 #include <linux/kernel.h>
21 #include <linux/start_kernel.h>
22 #include <linux/string.h>
23 #include <linux/console.h>
24 #include <linux/screen_info.h>
25 #include <linux/irq.h>
26 #include <linux/interrupt.h>
27 #include <linux/clocksource.h>
28 #include <linux/clockchips.h>
29 #include <linux/cpu.h>
30 #include <linux/lguest.h>
31 #include <linux/lguest_launcher.h>
32 #include <asm/paravirt.h>
33 #include <asm/param.h>
35 #include <asm/pgtable.h>
37 #include <asm/setup.h>
38 #include <asm/lguest.h>
39 #include <asm/uaccess.h>
43 static int cpu_had_pge;
47 unsigned short segment;
50 /* Offset from where switcher.S was compiled to where we've copied it */
51 static unsigned long switcher_offset(void)
53 return SWITCHER_ADDR - (unsigned long)start_switcher_text;
56 /* This cpu's struct lguest_pages. */
57 static struct lguest_pages *lguest_pages(unsigned int cpu)
59 return &(((struct lguest_pages *)
60 (SWITCHER_ADDR + SHARED_SWITCHER_PAGES*PAGE_SIZE))[cpu]);
63 static DEFINE_PER_CPU(struct lguest *, last_guest);
66 * We are getting close to the Switcher.
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.
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
76 static void copy_in_guest_info(struct lguest *lg, struct lguest_pages *pages)
78 /* Copying all this data can be quite expensive. We usually run the
79 * same Guest we ran last time (and that Guest hasn't run anywhere else
80 * meanwhile). If that's not the case, we pretend everything in the
81 * Guest has changed. */
82 if (__get_cpu_var(last_guest) != lg || lg->last_pages != pages) {
83 __get_cpu_var(last_guest) = lg;
84 lg->last_pages = pages;
85 lg->changed = CHANGED_ALL;
88 /* These copies are pretty cheap, so we do them unconditionally: */
89 /* Save the current Host top-level page directory. */
90 pages->state.host_cr3 = __pa(current->mm->pgd);
91 /* Set up the Guest's page tables to see this CPU's pages (and no
92 * other CPU's pages). */
93 map_switcher_in_guest(lg, pages);
94 /* Set up the two "TSS" members which tell the CPU what stack to use
95 * for traps which do directly into the Guest (ie. traps at privilege
97 pages->state.guest_tss.esp1 = lg->esp1;
98 pages->state.guest_tss.ss1 = lg->ss1;
100 /* Copy direct-to-Guest trap entries. */
101 if (lg->changed & CHANGED_IDT)
102 copy_traps(lg, pages->state.guest_idt, default_idt_entries);
104 /* Copy all GDT entries which the Guest can change. */
105 if (lg->changed & CHANGED_GDT)
106 copy_gdt(lg, pages->state.guest_gdt);
107 /* If only the TLS entries have changed, copy them. */
108 else if (lg->changed & CHANGED_GDT_TLS)
109 copy_gdt_tls(lg, pages->state.guest_gdt);
111 /* Mark the Guest as unchanged for next time. */
115 /* Finally: the code to actually call into the Switcher to run the Guest. */
116 static void run_guest_once(struct lguest *lg, struct lguest_pages *pages)
118 /* This is a dummy value we need for GCC's sake. */
119 unsigned int clobber;
121 /* Copy the guest-specific information into this CPU's "struct
123 copy_in_guest_info(lg, pages);
125 /* Set the trap number to 256 (impossible value). If we fault while
126 * switching to the Guest (bad segment registers or bug), this will
127 * cause us to abort the Guest. */
128 lg->regs->trapnum = 256;
130 /* Now: we push the "eflags" register on the stack, then do an "lcall".
131 * This is how we change from using the kernel code segment to using
132 * the dedicated lguest code segment, as well as jumping into the
135 * The lcall also pushes the old code segment (KERNEL_CS) onto the
136 * stack, then the address of this call. This stack layout happens to
137 * exactly match the stack of an interrupt... */
138 asm volatile("pushf; lcall *lguest_entry"
139 /* This is how we tell GCC that %eax ("a") and %ebx ("b")
140 * are changed by this routine. The "=" means output. */
141 : "=a"(clobber), "=b"(clobber)
142 /* %eax contains the pages pointer. ("0" refers to the
143 * 0-th argument above, ie "a"). %ebx contains the
144 * physical address of the Guest's top-level page
146 : "0"(pages), "1"(__pa(lg->pgdirs[lg->pgdidx].pgdir))
147 /* We tell gcc that all these registers could change,
148 * which means we don't have to save and restore them in
150 : "memory", "%edx", "%ecx", "%edi", "%esi");
154 /*H:040 This is the i386-specific code to setup and run the Guest. Interrupts
155 * are disabled: we own the CPU. */
156 void lguest_arch_run_guest(struct lguest *lg)
158 /* Remember the awfully-named TS bit? If the Guest has asked
159 * to set it we set it now, so we can trap and pass that trap
160 * to the Guest if it uses the FPU. */
164 /* SYSENTER is an optimized way of doing system calls. We
165 * can't allow it because it always jumps to privilege level 0.
166 * A normal Guest won't try it because we don't advertise it in
167 * CPUID, but a malicious Guest (or malicious Guest userspace
168 * program) could, so we tell the CPU to disable it before
169 * running the Guest. */
170 if (boot_cpu_has(X86_FEATURE_SEP))
171 wrmsr(MSR_IA32_SYSENTER_CS, 0, 0);
173 /* Now we actually run the Guest. It will pop back out when
174 * something interesting happens, and we can examine its
175 * registers to see what it was doing. */
176 run_guest_once(lg, lguest_pages(raw_smp_processor_id()));
178 /* The "regs" pointer contains two extra entries which are not
179 * really registers: a trap number which says what interrupt or
180 * trap made the switcher code come back, and an error code
181 * which some traps set. */
183 /* If the Guest page faulted, then the cr2 register will tell
184 * us the bad virtual address. We have to grab this now,
185 * because once we re-enable interrupts an interrupt could
186 * fault and thus overwrite cr2, or we could even move off to a
188 if (lg->regs->trapnum == 14)
189 lg->arch.last_pagefault = read_cr2();
190 /* Similarly, if we took a trap because the Guest used the FPU,
191 * we have to restore the FPU it expects to see. */
192 else if (lg->regs->trapnum == 7)
193 math_state_restore();
195 /* Restore SYSENTER if it's supposed to be on. */
196 if (boot_cpu_has(X86_FEATURE_SEP))
197 wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0);
200 /*H:130 Our Guest is usually so well behaved; it never tries to do things it
201 * isn't allowed to. Unfortunately, Linux's paravirtual infrastructure isn't
202 * quite complete, because it doesn't contain replacements for the Intel I/O
203 * instructions. As a result, the Guest sometimes fumbles across one during
204 * the boot process as it probes for various things which are usually attached
207 * When the Guest uses one of these instructions, we get trap #13 (General
208 * Protection Fault) and come here. We see if it's one of those troublesome
209 * instructions and skip over it. We return true if we did. */
210 static int emulate_insn(struct lguest *lg)
213 unsigned int insnlen = 0, in = 0, shift = 0;
214 /* The eip contains the *virtual* address of the Guest's instruction:
215 * guest_pa just subtracts the Guest's page_offset. */
216 unsigned long physaddr = guest_pa(lg, lg->regs->eip);
218 /* This must be the Guest kernel trying to do something, not userspace!
219 * The bottom two bits of the CS segment register are the privilege
221 if ((lg->regs->cs & 3) != GUEST_PL)
224 /* Decoding x86 instructions is icky. */
225 insn = lgread(lg, physaddr, u8);
227 /* 0x66 is an "operand prefix". It means it's using the upper 16 bits
228 of the eax register. */
231 /* The instruction is 1 byte so far, read the next byte. */
233 insn = lgread(lg, physaddr + insnlen, u8);
236 /* We can ignore the lower bit for the moment and decode the 4 opcodes
237 * we need to emulate. */
238 switch (insn & 0xFE) {
239 case 0xE4: /* in <next byte>,%al */
243 case 0xEC: /* in (%dx),%al */
247 case 0xE6: /* out %al,<next byte> */
250 case 0xEE: /* out %al,(%dx) */
254 /* OK, we don't know what this is, can't emulate. */
258 /* If it was an "IN" instruction, they expect the result to be read
259 * into %eax, so we change %eax. We always return all-ones, which
260 * traditionally means "there's nothing there". */
262 /* Lower bit tells is whether it's a 16 or 32 bit access */
264 lg->regs->eax = 0xFFFFFFFF;
266 lg->regs->eax |= (0xFFFF << shift);
268 /* Finally, we've "done" the instruction, so move past it. */
269 lg->regs->eip += insnlen;
274 /*H:050 Once we've re-enabled interrupts, we look at why the Guest exited. */
275 void lguest_arch_handle_trap(struct lguest *lg)
277 switch (lg->regs->trapnum) {
278 case 13: /* We've intercepted a GPF. */
279 /* Check if this was one of those annoying IN or OUT
280 * instructions which we need to emulate. If so, we
281 * just go back into the Guest after we've done it. */
282 if (lg->regs->errcode == 0) {
283 if (emulate_insn(lg))
287 case 14: /* We've intercepted a page fault. */
288 /* The Guest accessed a virtual address that wasn't
289 * mapped. This happens a lot: we don't actually set
290 * up most of the page tables for the Guest at all when
291 * we start: as it runs it asks for more and more, and
292 * we set them up as required. In this case, we don't
293 * even tell the Guest that the fault happened.
295 * The errcode tells whether this was a read or a
296 * write, and whether kernel or userspace code. */
297 if (demand_page(lg, lg->arch.last_pagefault, lg->regs->errcode))
300 /* OK, it's really not there (or not OK): the Guest
301 * needs to know. We write out the cr2 value so it
302 * knows where the fault occurred.
304 * Note that if the Guest were really messed up, this
305 * could happen before it's done the INITIALIZE
306 * hypercall, so lg->lguest_data will be NULL */
307 if (lg->lguest_data &&
308 put_user(lg->arch.last_pagefault, &lg->lguest_data->cr2))
309 kill_guest(lg, "Writing cr2");
311 case 7: /* We've intercepted a Device Not Available fault. */
312 /* If the Guest doesn't want to know, we already
313 * restored the Floating Point Unit, so we just
314 * continue without telling it. */
319 /* These values mean a real interrupt occurred, in which case
320 * the Host handler has already been run. We just do a
321 * friendly check if another process should now be run, then
322 * return to run the Guest again */
325 case LGUEST_TRAP_ENTRY:
326 /* Our 'struct hcall_args' maps directly over our regs: we set
327 * up the pointer now to indicate a hypercall is pending. */
328 lg->hcall = (struct hcall_args *)lg->regs;
332 /* We didn't handle the trap, so it needs to go to the Guest. */
333 if (!deliver_trap(lg, lg->regs->trapnum))
334 /* If the Guest doesn't have a handler (either it hasn't
335 * registered any yet, or it's one of the faults we don't let
336 * it handle), it dies with a cryptic error message. */
337 kill_guest(lg, "unhandled trap %li at %#lx (%#lx)",
338 lg->regs->trapnum, lg->regs->eip,
339 lg->regs->trapnum == 14 ? lg->arch.last_pagefault
340 : lg->regs->errcode);
343 /* Now we can look at each of the routines this calls, in increasing order of
344 * complexity: do_hypercalls(), emulate_insn(), maybe_do_interrupt(),
345 * deliver_trap() and demand_page(). After all those, we'll be ready to
346 * examine the Switcher, and our philosophical understanding of the Host/Guest
347 * duality will be complete. :*/
348 static void adjust_pge(void *on)
351 write_cr4(read_cr4() | X86_CR4_PGE);
353 write_cr4(read_cr4() & ~X86_CR4_PGE);
356 /*H:020 Now the Switcher is mapped and every thing else is ready, we need to do
357 * some more i386-specific initialization. */
358 void __init lguest_arch_host_init(void)
362 /* Most of the i386/switcher.S doesn't care that it's been moved; on
363 * Intel, jumps are relative, and it doesn't access any references to
364 * external code or data.
366 * The only exception is the interrupt handlers in switcher.S: their
367 * addresses are placed in a table (default_idt_entries), so we need to
368 * update the table with the new addresses. switcher_offset() is a
369 * convenience function which returns the distance between the builtin
370 * switcher code and the high-mapped copy we just made. */
371 for (i = 0; i < IDT_ENTRIES; i++)
372 default_idt_entries[i] += switcher_offset();
375 * Set up the Switcher's per-cpu areas.
377 * Each CPU gets two pages of its own within the high-mapped region
378 * (aka. "struct lguest_pages"). Much of this can be initialized now,
379 * but some depends on what Guest we are running (which is set up in
380 * copy_in_guest_info()).
382 for_each_possible_cpu(i) {
383 /* lguest_pages() returns this CPU's two pages. */
384 struct lguest_pages *pages = lguest_pages(i);
385 /* This is a convenience pointer to make the code fit one
386 * statement to a line. */
387 struct lguest_ro_state *state = &pages->state;
389 /* The Global Descriptor Table: the Host has a different one
390 * for each CPU. We keep a descriptor for the GDT which says
391 * where it is and how big it is (the size is actually the last
392 * byte, not the size, hence the "-1"). */
393 state->host_gdt_desc.size = GDT_SIZE-1;
394 state->host_gdt_desc.address = (long)get_cpu_gdt_table(i);
396 /* All CPUs on the Host use the same Interrupt Descriptor
397 * Table, so we just use store_idt(), which gets this CPU's IDT
399 store_idt(&state->host_idt_desc);
401 /* The descriptors for the Guest's GDT and IDT can be filled
402 * out now, too. We copy the GDT & IDT into ->guest_gdt and
403 * ->guest_idt before actually running the Guest. */
404 state->guest_idt_desc.size = sizeof(state->guest_idt)-1;
405 state->guest_idt_desc.address = (long)&state->guest_idt;
406 state->guest_gdt_desc.size = sizeof(state->guest_gdt)-1;
407 state->guest_gdt_desc.address = (long)&state->guest_gdt;
409 /* We know where we want the stack to be when the Guest enters
410 * the switcher: in pages->regs. The stack grows upwards, so
411 * we start it at the end of that structure. */
412 state->guest_tss.esp0 = (long)(&pages->regs + 1);
413 /* And this is the GDT entry to use for the stack: we keep a
414 * couple of special LGUEST entries. */
415 state->guest_tss.ss0 = LGUEST_DS;
417 /* x86 can have a finegrained bitmap which indicates what I/O
418 * ports the process can use. We set it to the end of our
419 * structure, meaning "none". */
420 state->guest_tss.io_bitmap_base = sizeof(state->guest_tss);
422 /* Some GDT entries are the same across all Guests, so we can
423 * set them up now. */
424 setup_default_gdt_entries(state);
425 /* Most IDT entries are the same for all Guests, too.*/
426 setup_default_idt_entries(state, default_idt_entries);
428 /* The Host needs to be able to use the LGUEST segments on this
429 * CPU, too, so put them in the Host GDT. */
430 get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT;
431 get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT;
434 /* In the Switcher, we want the %cs segment register to use the
435 * LGUEST_CS GDT entry: we've put that in the Host and Guest GDTs, so
436 * it will be undisturbed when we switch. To change %cs and jump we
437 * need this structure to feed to Intel's "lcall" instruction. */
438 lguest_entry.offset = (long)switch_to_guest + switcher_offset();
439 lguest_entry.segment = LGUEST_CS;
441 /* Finally, we need to turn off "Page Global Enable". PGE is an
442 * optimization where page table entries are specially marked to show
443 * they never change. The Host kernel marks all the kernel pages this
444 * way because it's always present, even when userspace is running.
446 * Lguest breaks this: unbeknownst to the rest of the Host kernel, we
447 * switch to the Guest kernel. If you don't disable this on all CPUs,
448 * you'll get really weird bugs that you'll chase for two days.
450 * I used to turn PGE off every time we switched to the Guest and back
451 * on when we return, but that slowed the Switcher down noticibly. */
453 /* We don't need the complexity of CPUs coming and going while we're
456 if (cpu_has_pge) { /* We have a broader idea of "global". */
457 /* Remember that this was originally set (for cleanup). */
459 /* adjust_pge is a helper function which sets or unsets the PGE
460 * bit on its CPU, depending on the argument (0 == unset). */
461 on_each_cpu(adjust_pge, (void *)0, 0, 1);
462 /* Turn off the feature in the global feature set. */
463 clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability);
465 unlock_cpu_hotplug();
469 void __exit lguest_arch_host_fini(void)
471 /* If we had PGE before we started, turn it back on now. */
474 set_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability);
475 /* adjust_pge's argument "1" means set PGE. */
476 on_each_cpu(adjust_pge, (void *)1, 0, 1);
478 unlock_cpu_hotplug();
482 /*H:122 The i386-specific hypercalls simply farm out to the right functions. */
483 int lguest_arch_do_hcall(struct lguest *lg, struct hcall_args *args)
485 switch (args->arg0) {
486 case LHCALL_LOAD_GDT:
487 load_guest_gdt(lg, args->arg1, args->arg2);
489 case LHCALL_LOAD_IDT_ENTRY:
490 load_guest_idt_entry(lg, args->arg1, args->arg2, args->arg3);
492 case LHCALL_LOAD_TLS:
493 guest_load_tls(lg, args->arg1);
496 /* Bad Guest. Bad! */
502 /*H:126 i386-specific hypercall initialization: */
503 int lguest_arch_init_hypercalls(struct lguest *lg)
507 /* The pointer to the Guest's "struct lguest_data" is the only
508 * argument. We check that address now. */
509 if (!lguest_address_ok(lg, lg->hcall->arg1, sizeof(*lg->lguest_data)))
512 /* Having checked it, we simply set lg->lguest_data to point straight
513 * into the Launcher's memory at the right place and then use
514 * copy_to_user/from_user from now on, instead of lgread/write. I put
515 * this in to show that I'm not immune to writing stupid
517 lg->lguest_data = lg->mem_base + lg->hcall->arg1;
519 /* We insist that the Time Stamp Counter exist and doesn't change with
520 * cpu frequency. Some devious chip manufacturers decided that TSC
521 * changes could be handled in software. I decided that time going
522 * backwards might be good for benchmarks, but it's bad for users.
524 * We also insist that the TSC be stable: the kernel detects unreliable
525 * TSCs for its own purposes, and we use that here. */
526 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) && !check_tsc_unstable())
530 if (put_user(tsc_speed, &lg->lguest_data->tsc_khz))
533 /* The interrupt code might not like the system call vector. */
534 if (!check_syscall_vector(lg))
535 kill_guest(lg, "bad syscall vector");
539 /* Now we've examined the hypercall code; our Guest can make requests. There
540 * is one other way we can do things for the Guest, as we see in
541 * emulate_insn(). :*/
543 /*L:030 lguest_arch_setup_regs()
545 * Most of the Guest's registers are left alone: we used get_zeroed_page() to
546 * allocate the structure, so they will be 0. */
547 void lguest_arch_setup_regs(struct lguest *lg, unsigned long start)
549 struct lguest_regs *regs = lg->regs;
551 /* There are four "segment" registers which the Guest needs to boot:
552 * The "code segment" register (cs) refers to the kernel code segment
553 * __KERNEL_CS, and the "data", "extra" and "stack" segment registers
554 * refer to the kernel data segment __KERNEL_DS.
556 * The privilege level is packed into the lower bits. The Guest runs
557 * at privilege level 1 (GUEST_PL).*/
558 regs->ds = regs->es = regs->ss = __KERNEL_DS|GUEST_PL;
559 regs->cs = __KERNEL_CS|GUEST_PL;
561 /* The "eflags" register contains miscellaneous flags. Bit 1 (0x002)
562 * is supposed to always be "1". Bit 9 (0x200) controls whether
563 * interrupts are enabled. We always leave interrupts enabled while
564 * running the Guest. */
565 regs->eflags = 0x202;
567 /* The "Extended Instruction Pointer" register says where the Guest is
571 /* %esi points to our boot information, at physical address 0, so don't
573 /* There are a couple of GDT entries the Guest expects when first
projects / linux-2.6 / blob