2 * A hypervisor allows multiple Operating Systems to run on a single machine.
3 * To quote David Wheeler: "Any problem in computer science can be solved with
4 * another layer of indirection."
6 * We keep things simple in two ways. First, we start with a normal Linux
7 * kernel and insert a module (lg.ko) which allows us to run other Linux
8 * kernels the same way we'd run processes. We call the first kernel the Host,
9 * and the others the Guests. The program which sets up and configures Guests
10 * (such as the example in Documentation/lguest/lguest.c) is called the
13 * Secondly, we only run specially modified Guests, not normal kernels: setting
14 * CONFIG_LGUEST_GUEST to "y" compiles this file into the kernel so it knows
15 * how to be a Guest at boot time. This means that you can use the same kernel
16 * you boot normally (ie. as a Host) as a Guest.
18 * These Guests know that they cannot do privileged operations, such as disable
19 * interrupts, and that they have to ask the Host to do such things explicitly.
20 * This file consists of all the replacements for such low-level native
21 * hardware operations: these special Guest versions call the Host.
23 * So how does the kernel know it's a Guest? We'll see that later, but let's
24 * just say that we end up here where we replace the native functions various
25 * "paravirt" structures with our Guest versions, then boot like normal.
29 * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation.
31 * This program is free software; you can redistribute it and/or modify
32 * it under the terms of the GNU General Public License as published by
33 * the Free Software Foundation; either version 2 of the License, or
34 * (at your option) any later version.
36 * This program is distributed in the hope that it will be useful, but
37 * WITHOUT ANY WARRANTY; without even the implied warranty of
38 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
39 * NON INFRINGEMENT. See the GNU General Public License for more
42 * You should have received a copy of the GNU General Public License
43 * along with this program; if not, write to the Free Software
44 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
46 #include <linux/kernel.h>
47 #include <linux/start_kernel.h>
48 #include <linux/string.h>
49 #include <linux/console.h>
50 #include <linux/screen_info.h>
51 #include <linux/irq.h>
52 #include <linux/interrupt.h>
53 #include <linux/clocksource.h>
54 #include <linux/clockchips.h>
55 #include <linux/lguest.h>
56 #include <linux/lguest_launcher.h>
57 #include <linux/virtio_console.h>
60 #include <asm/lguest.h>
61 #include <asm/paravirt.h>
62 #include <asm/param.h>
64 #include <asm/pgtable.h>
66 #include <asm/setup.h>
71 #include <asm/stackprotector.h>
72 #include <asm/reboot.h> /* for struct machine_ops */
74 /*G:010 Welcome to the Guest!
76 * The Guest in our tale is a simple creature: identical to the Host but
77 * behaving in simplified but equivalent ways. In particular, the Guest is the
78 * same kernel as the Host (or at least, built from the same source code).
81 struct lguest_data lguest_data = {
82 .hcall_status = { [0 ... LHCALL_RING_SIZE-1] = 0xFF },
83 .noirq_start = (u32)lguest_noirq_start,
84 .noirq_end = (u32)lguest_noirq_end,
85 .kernel_address = PAGE_OFFSET,
86 .blocked_interrupts = { 1 }, /* Block timer interrupts */
87 .syscall_vec = SYSCALL_VECTOR,
91 * async_hcall() is pretty simple: I'm quite proud of it really. We have a
92 * ring buffer of stored hypercalls which the Host will run though next time we
93 * do a normal hypercall. Each entry in the ring has 5 slots for the hypercall
94 * arguments, and a "hcall_status" word which is 0 if the call is ready to go,
95 * and 255 once the Host has finished with it.
97 * If we come around to a slot which hasn't been finished, then the table is
98 * full and we just make the hypercall directly. This has the nice side
99 * effect of causing the Host to run all the stored calls in the ring buffer
100 * which empties it for next time!
102 static void async_hcall(unsigned long call, unsigned long arg1,
103 unsigned long arg2, unsigned long arg3,
106 /* Note: This code assumes we're uniprocessor. */
107 static unsigned int next_call;
111 * Disable interrupts if not already disabled: we don't want an
112 * interrupt handler making a hypercall while we're already doing
115 local_irq_save(flags);
116 if (lguest_data.hcall_status[next_call] != 0xFF) {
117 /* Table full, so do normal hcall which will flush table. */
118 kvm_hypercall4(call, arg1, arg2, arg3, arg4);
120 lguest_data.hcalls[next_call].arg0 = call;
121 lguest_data.hcalls[next_call].arg1 = arg1;
122 lguest_data.hcalls[next_call].arg2 = arg2;
123 lguest_data.hcalls[next_call].arg3 = arg3;
124 lguest_data.hcalls[next_call].arg4 = arg4;
125 /* Arguments must all be written before we mark it to go */
127 lguest_data.hcall_status[next_call] = 0;
128 if (++next_call == LHCALL_RING_SIZE)
131 local_irq_restore(flags);
135 * Notice the lazy_hcall() above, rather than hcall(). This is our first real
136 * optimization trick!
138 * When lazy_mode is set, it means we're allowed to defer all hypercalls and do
139 * them as a batch when lazy_mode is eventually turned off. Because hypercalls
140 * are reasonably expensive, batching them up makes sense. For example, a
141 * large munmap might update dozens of page table entries: that code calls
142 * paravirt_enter_lazy_mmu(), does the dozen updates, then calls
143 * lguest_leave_lazy_mode().
145 * So, when we're in lazy mode, we call async_hcall() to store the call for
148 static void lazy_hcall1(unsigned long call,
151 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
152 kvm_hypercall1(call, arg1);
154 async_hcall(call, arg1, 0, 0, 0);
157 static void lazy_hcall2(unsigned long call,
161 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
162 kvm_hypercall2(call, arg1, arg2);
164 async_hcall(call, arg1, arg2, 0, 0);
167 static void lazy_hcall3(unsigned long call,
172 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
173 kvm_hypercall3(call, arg1, arg2, arg3);
175 async_hcall(call, arg1, arg2, arg3, 0);
178 #ifdef CONFIG_X86_PAE
179 static void lazy_hcall4(unsigned long call,
185 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
186 kvm_hypercall4(call, arg1, arg2, arg3, arg4);
188 async_hcall(call, arg1, arg2, arg3, arg4);
192 /* When lazy mode is turned off reset the per-cpu lazy mode variable and then
193 * issue the do-nothing hypercall to flush any stored calls. */
194 static void lguest_leave_lazy_mmu_mode(void)
196 kvm_hypercall0(LHCALL_FLUSH_ASYNC);
197 paravirt_leave_lazy_mmu();
200 static void lguest_end_context_switch(struct task_struct *next)
202 kvm_hypercall0(LHCALL_FLUSH_ASYNC);
203 paravirt_end_context_switch(next);
207 * After that diversion we return to our first native-instruction
208 * replacements: four functions for interrupt control.
210 * The simplest way of implementing these would be to have "turn interrupts
211 * off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow:
212 * these are by far the most commonly called functions of those we override.
214 * So instead we keep an "irq_enabled" field inside our "struct lguest_data",
215 * which the Guest can update with a single instruction. The Host knows to
216 * check there before it tries to deliver an interrupt.
220 * save_flags() is expected to return the processor state (ie. "flags"). The
221 * flags word contains all kind of stuff, but in practice Linux only cares
222 * about the interrupt flag. Our "save_flags()" just returns that.
224 static unsigned long save_fl(void)
226 return lguest_data.irq_enabled;
229 /* Interrupts go off... */
230 static void irq_disable(void)
232 lguest_data.irq_enabled = 0;
236 * Let's pause a moment. Remember how I said these are called so often?
237 * Jeremy Fitzhardinge optimized them so hard early in 2009 that he had to
238 * break some rules. In particular, these functions are assumed to save their
239 * own registers if they need to: normal C functions assume they can trash the
240 * eax register. To use normal C functions, we use
241 * PV_CALLEE_SAVE_REGS_THUNK(), which pushes %eax onto the stack, calls the
242 * C function, then restores it.
244 PV_CALLEE_SAVE_REGS_THUNK(save_fl);
245 PV_CALLEE_SAVE_REGS_THUNK(irq_disable);
248 /* These are in i386_head.S */
249 extern void lg_irq_enable(void);
250 extern void lg_restore_fl(unsigned long flags);
253 * Note that we don't check for outstanding interrupts when we re-enable them
254 * (or when we unmask an interrupt). This seems to work for the moment, since
255 * interrupts are rare and we'll just get the interrupt on the next timer tick,
256 * but now we can run with CONFIG_NO_HZ, we should revisit this. One way would
257 * be to put the "irq_enabled" field in a page by itself, and have the Host
258 * write-protect it when an interrupt comes in when irqs are disabled. There
259 * will then be a page fault as soon as interrupts are re-enabled.
261 * A better method is to implement soft interrupt disable generally for x86:
262 * instead of disabling interrupts, we set a flag. If an interrupt does come
263 * in, we then disable them for real. This is uncommon, so we could simply use
264 * a hypercall for interrupt control and not worry about efficiency.
268 * The Interrupt Descriptor Table (IDT).
270 * The IDT tells the processor what to do when an interrupt comes in. Each
271 * entry in the table is a 64-bit descriptor: this holds the privilege level,
272 * address of the handler, and... well, who cares? The Guest just asks the
273 * Host to make the change anyway, because the Host controls the real IDT.
275 static void lguest_write_idt_entry(gate_desc *dt,
276 int entrynum, const gate_desc *g)
279 * The gate_desc structure is 8 bytes long: we hand it to the Host in
280 * two 32-bit chunks. The whole 32-bit kernel used to hand descriptors
281 * around like this; typesafety wasn't a big concern in Linux's early
284 u32 *desc = (u32 *)g;
285 /* Keep the local copy up to date. */
286 native_write_idt_entry(dt, entrynum, g);
287 /* Tell Host about this new entry. */
288 kvm_hypercall3(LHCALL_LOAD_IDT_ENTRY, entrynum, desc[0], desc[1]);
292 * Changing to a different IDT is very rare: we keep the IDT up-to-date every
293 * time it is written, so we can simply loop through all entries and tell the
296 static void lguest_load_idt(const struct desc_ptr *desc)
299 struct desc_struct *idt = (void *)desc->address;
301 for (i = 0; i < (desc->size+1)/8; i++)
302 kvm_hypercall3(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b);
306 * The Global Descriptor Table.
308 * The Intel architecture defines another table, called the Global Descriptor
309 * Table (GDT). You tell the CPU where it is (and its size) using the "lgdt"
310 * instruction, and then several other instructions refer to entries in the
311 * table. There are three entries which the Switcher needs, so the Host simply
312 * controls the entire thing and the Guest asks it to make changes using the
313 * LOAD_GDT hypercall.
315 * This is the exactly like the IDT code.
317 static void lguest_load_gdt(const struct desc_ptr *desc)
320 struct desc_struct *gdt = (void *)desc->address;
322 for (i = 0; i < (desc->size+1)/8; i++)
323 kvm_hypercall3(LHCALL_LOAD_GDT_ENTRY, i, gdt[i].a, gdt[i].b);
327 * For a single GDT entry which changes, we do the lazy thing: alter our GDT,
328 * then tell the Host to reload the entire thing. This operation is so rare
329 * that this naive implementation is reasonable.
331 static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum,
332 const void *desc, int type)
334 native_write_gdt_entry(dt, entrynum, desc, type);
335 /* Tell Host about this new entry. */
336 kvm_hypercall3(LHCALL_LOAD_GDT_ENTRY, entrynum,
337 dt[entrynum].a, dt[entrynum].b);
341 * OK, I lied. There are three "thread local storage" GDT entries which change
342 * on every context switch (these three entries are how glibc implements
343 * __thread variables). So we have a hypercall specifically for this case.
345 static void lguest_load_tls(struct thread_struct *t, unsigned int cpu)
348 * There's one problem which normal hardware doesn't have: the Host
349 * can't handle us removing entries we're currently using. So we clear
350 * the GS register here: if it's needed it'll be reloaded anyway.
353 lazy_hcall2(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu);
357 * That's enough excitement for now, back to ploughing through each of the
358 * different pv_ops structures (we're about 1/3 of the way through).
360 * This is the Local Descriptor Table, another weird Intel thingy. Linux only
361 * uses this for some strange applications like Wine. We don't do anything
362 * here, so they'll get an informative and friendly Segmentation Fault.
364 static void lguest_set_ldt(const void *addr, unsigned entries)
369 * This loads a GDT entry into the "Task Register": that entry points to a
370 * structure called the Task State Segment. Some comments scattered though the
371 * kernel code indicate that this used for task switching in ages past, along
372 * with blood sacrifice and astrology.
374 * Now there's nothing interesting in here that we don't get told elsewhere.
375 * But the native version uses the "ltr" instruction, which makes the Host
376 * complain to the Guest about a Segmentation Fault and it'll oops. So we
377 * override the native version with a do-nothing version.
379 static void lguest_load_tr_desc(void)
384 * The "cpuid" instruction is a way of querying both the CPU identity
385 * (manufacturer, model, etc) and its features. It was introduced before the
386 * Pentium in 1993 and keeps getting extended by both Intel, AMD and others.
387 * As you might imagine, after a decade and a half this treatment, it is now a
388 * giant ball of hair. Its entry in the current Intel manual runs to 28 pages.
390 * This instruction even it has its own Wikipedia entry. The Wikipedia entry
391 * has been translated into 5 languages. I am not making this up!
393 * We could get funky here and identify ourselves as "GenuineLguest", but
394 * instead we just use the real "cpuid" instruction. Then I pretty much turned
395 * off feature bits until the Guest booted. (Don't say that: you'll damage
396 * lguest sales!) Shut up, inner voice! (Hey, just pointing out that this is
397 * hardly future proof.) Noone's listening! They don't like you anyway,
398 * parenthetic weirdo!
400 * Replacing the cpuid so we can turn features off is great for the kernel, but
401 * anyone (including userspace) can just use the raw "cpuid" instruction and
402 * the Host won't even notice since it isn't privileged. So we try not to get
403 * too worked up about it.
405 static void lguest_cpuid(unsigned int *ax, unsigned int *bx,
406 unsigned int *cx, unsigned int *dx)
410 native_cpuid(ax, bx, cx, dx);
413 * CPUID 0 gives the highest legal CPUID number (and the ID string).
414 * We futureproof our code a little by sticking to known CPUID values.
422 * CPUID 1 is a basic feature request.
424 * CX: we only allow kernel to see SSE3, CMPXCHG16B and SSSE3
425 * DX: SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU and PAE.
431 * The Host can do a nice optimization if it knows that the
432 * kernel mappings (addresses above 0xC0000000 or whatever
433 * PAGE_OFFSET is set to) haven't changed. But Linux calls
434 * flush_tlb_user() for both user and kernel mappings unless
435 * the Page Global Enable (PGE) feature bit is set.
439 * We also lie, and say we're family id 5. 6 or greater
440 * leads to a rdmsr in early_init_intel which we can't handle.
441 * Family ID is returned as bits 8-12 in ax.
447 * 0x80000000 returns the highest Extended Function, so we futureproof
448 * like we do above by limiting it to known fields.
451 if (*ax > 0x80000008)
456 * PAE systems can mark pages as non-executable. Linux calls this the
457 * NX bit. Intel calls it XD (eXecute Disable), AMD EVP (Enhanced
458 * Virus Protection). We just switch turn if off here, since we don't
468 * Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4.
469 * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother
470 * it. The Host needs to know when the Guest wants to change them, so we have
471 * a whole series of functions like read_cr0() and write_cr0().
473 * We start with cr0. cr0 allows you to turn on and off all kinds of basic
474 * features, but Linux only really cares about one: the horrifically-named Task
475 * Switched (TS) bit at bit 3 (ie. 8)
477 * What does the TS bit do? Well, it causes the CPU to trap (interrupt 7) if
478 * the floating point unit is used. Which allows us to restore FPU state
479 * lazily after a task switch, and Linux uses that gratefully, but wouldn't a
480 * name like "FPUTRAP bit" be a little less cryptic?
482 * We store cr0 locally because the Host never changes it. The Guest sometimes
483 * wants to read it and we'd prefer not to bother the Host unnecessarily.
485 static unsigned long current_cr0;
486 static void lguest_write_cr0(unsigned long val)
488 lazy_hcall1(LHCALL_TS, val & X86_CR0_TS);
492 static unsigned long lguest_read_cr0(void)
498 * Intel provided a special instruction to clear the TS bit for people too cool
499 * to use write_cr0() to do it. This "clts" instruction is faster, because all
500 * the vowels have been optimized out.
502 static void lguest_clts(void)
504 lazy_hcall1(LHCALL_TS, 0);
505 current_cr0 &= ~X86_CR0_TS;
509 * cr2 is the virtual address of the last page fault, which the Guest only ever
510 * reads. The Host kindly writes this into our "struct lguest_data", so we
511 * just read it out of there.
513 static unsigned long lguest_read_cr2(void)
515 return lguest_data.cr2;
518 /* See lguest_set_pte() below. */
519 static bool cr3_changed = false;
522 * cr3 is the current toplevel pagetable page: the principle is the same as
523 * cr0. Keep a local copy, and tell the Host when it changes. The only
524 * difference is that our local copy is in lguest_data because the Host needs
525 * to set it upon our initial hypercall.
527 static void lguest_write_cr3(unsigned long cr3)
529 lguest_data.pgdir = cr3;
530 lazy_hcall1(LHCALL_NEW_PGTABLE, cr3);
534 static unsigned long lguest_read_cr3(void)
536 return lguest_data.pgdir;
539 /* cr4 is used to enable and disable PGE, but we don't care. */
540 static unsigned long lguest_read_cr4(void)
545 static void lguest_write_cr4(unsigned long val)
550 * Page Table Handling.
552 * Now would be a good time to take a rest and grab a coffee or similarly
553 * relaxing stimulant. The easy parts are behind us, and the trek gradually
554 * winds uphill from here.
556 * Quick refresher: memory is divided into "pages" of 4096 bytes each. The CPU
557 * maps virtual addresses to physical addresses using "page tables". We could
558 * use one huge index of 1 million entries: each address is 4 bytes, so that's
559 * 1024 pages just to hold the page tables. But since most virtual addresses
560 * are unused, we use a two level index which saves space. The cr3 register
561 * contains the physical address of the top level "page directory" page, which
562 * contains physical addresses of up to 1024 second-level pages. Each of these
563 * second level pages contains up to 1024 physical addresses of actual pages,
564 * or Page Table Entries (PTEs).
566 * Here's a diagram, where arrows indicate physical addresses:
568 * cr3 ---> +---------+
569 * | --------->+---------+
571 * Top-level | | PADDR2 |
578 * So to convert a virtual address to a physical address, we look up the top
579 * level, which points us to the second level, which gives us the physical
580 * address of that page. If the top level entry was not present, or the second
581 * level entry was not present, then the virtual address is invalid (we
582 * say "the page was not mapped").
584 * Put another way, a 32-bit virtual address is divided up like so:
586 * 1 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
587 * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>|
588 * Index into top Index into second Offset within page
589 * page directory page pagetable page
591 * The kernel spends a lot of time changing both the top-level page directory
592 * and lower-level pagetable pages. The Guest doesn't know physical addresses,
593 * so while it maintains these page tables exactly like normal, it also needs
594 * to keep the Host informed whenever it makes a change: the Host will create
595 * the real page tables based on the Guests'.
599 * The Guest calls this to set a second-level entry (pte), ie. to map a page
600 * into a process' address space. We set the entry then tell the Host the
601 * toplevel and address this corresponds to. The Guest uses one pagetable per
602 * process, so we need to tell the Host which one we're changing (mm->pgd).
604 static void lguest_pte_update(struct mm_struct *mm, unsigned long addr,
607 #ifdef CONFIG_X86_PAE
608 lazy_hcall4(LHCALL_SET_PTE, __pa(mm->pgd), addr,
609 ptep->pte_low, ptep->pte_high);
611 lazy_hcall3(LHCALL_SET_PTE, __pa(mm->pgd), addr, ptep->pte_low);
615 static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr,
616 pte_t *ptep, pte_t pteval)
618 native_set_pte(ptep, pteval);
619 lguest_pte_update(mm, addr, ptep);
623 * The Guest calls lguest_set_pud to set a top-level entry and lguest_set_pmd
624 * to set a middle-level entry when PAE is activated.
626 * Again, we set the entry then tell the Host which page we changed,
627 * and the index of the entry we changed.
629 #ifdef CONFIG_X86_PAE
630 static void lguest_set_pud(pud_t *pudp, pud_t pudval)
632 native_set_pud(pudp, pudval);
634 /* 32 bytes aligned pdpt address and the index. */
635 lazy_hcall2(LHCALL_SET_PGD, __pa(pudp) & 0xFFFFFFE0,
636 (__pa(pudp) & 0x1F) / sizeof(pud_t));
639 static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
641 native_set_pmd(pmdp, pmdval);
642 lazy_hcall2(LHCALL_SET_PMD, __pa(pmdp) & PAGE_MASK,
643 (__pa(pmdp) & (PAGE_SIZE - 1)) / sizeof(pmd_t));
647 /* The Guest calls lguest_set_pmd to set a top-level entry when !PAE. */
648 static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
650 native_set_pmd(pmdp, pmdval);
651 lazy_hcall2(LHCALL_SET_PGD, __pa(pmdp) & PAGE_MASK,
652 (__pa(pmdp) & (PAGE_SIZE - 1)) / sizeof(pmd_t));
657 * There are a couple of legacy places where the kernel sets a PTE, but we
658 * don't know the top level any more. This is useless for us, since we don't
659 * know which pagetable is changing or what address, so we just tell the Host
660 * to forget all of them. Fortunately, this is very rare.
662 * ... except in early boot when the kernel sets up the initial pagetables,
663 * which makes booting astonishingly slow: 1.83 seconds! So we don't even tell
664 * the Host anything changed until we've done the first page table switch,
665 * which brings boot back to 0.25 seconds.
667 static void lguest_set_pte(pte_t *ptep, pte_t pteval)
669 native_set_pte(ptep, pteval);
671 lazy_hcall1(LHCALL_FLUSH_TLB, 1);
674 #ifdef CONFIG_X86_PAE
675 static void lguest_set_pte_atomic(pte_t *ptep, pte_t pte)
677 native_set_pte_atomic(ptep, pte);
679 lazy_hcall1(LHCALL_FLUSH_TLB, 1);
682 void lguest_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
684 native_pte_clear(mm, addr, ptep);
685 lguest_pte_update(mm, addr, ptep);
688 void lguest_pmd_clear(pmd_t *pmdp)
690 lguest_set_pmd(pmdp, __pmd(0));
695 * Unfortunately for Lguest, the pv_mmu_ops for page tables were based on
696 * native page table operations. On native hardware you can set a new page
697 * table entry whenever you want, but if you want to remove one you have to do
698 * a TLB flush (a TLB is a little cache of page table entries kept by the CPU).
700 * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only
701 * called when a valid entry is written, not when it's removed (ie. marked not
702 * present). Instead, this is where we come when the Guest wants to remove a
703 * page table entry: we tell the Host to set that entry to 0 (ie. the present
706 static void lguest_flush_tlb_single(unsigned long addr)
708 /* Simply set it to zero: if it was not, it will fault back in. */
709 lazy_hcall3(LHCALL_SET_PTE, lguest_data.pgdir, addr, 0);
713 * This is what happens after the Guest has removed a large number of entries.
714 * This tells the Host that any of the page table entries for userspace might
715 * have changed, ie. virtual addresses below PAGE_OFFSET.
717 static void lguest_flush_tlb_user(void)
719 lazy_hcall1(LHCALL_FLUSH_TLB, 0);
723 * This is called when the kernel page tables have changed. That's not very
724 * common (unless the Guest is using highmem, which makes the Guest extremely
725 * slow), so it's worth separating this from the user flushing above.
727 static void lguest_flush_tlb_kernel(void)
729 lazy_hcall1(LHCALL_FLUSH_TLB, 1);
733 * The Unadvanced Programmable Interrupt Controller.
735 * This is an attempt to implement the simplest possible interrupt controller.
736 * I spent some time looking though routines like set_irq_chip_and_handler,
737 * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and
738 * I *think* this is as simple as it gets.
740 * We can tell the Host what interrupts we want blocked ready for using the
741 * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as
742 * simple as setting a bit. We don't actually "ack" interrupts as such, we
743 * just mask and unmask them. I wonder if we should be cleverer?
745 static void disable_lguest_irq(unsigned int irq)
747 set_bit(irq, lguest_data.blocked_interrupts);
750 static void enable_lguest_irq(unsigned int irq)
752 clear_bit(irq, lguest_data.blocked_interrupts);
755 /* This structure describes the lguest IRQ controller. */
756 static struct irq_chip lguest_irq_controller = {
758 .mask = disable_lguest_irq,
759 .mask_ack = disable_lguest_irq,
760 .unmask = enable_lguest_irq,
764 * This sets up the Interrupt Descriptor Table (IDT) entry for each hardware
765 * interrupt (except 128, which is used for system calls), and then tells the
766 * Linux infrastructure that each interrupt is controlled by our level-based
767 * lguest interrupt controller.
769 static void __init lguest_init_IRQ(void)
773 for (i = FIRST_EXTERNAL_VECTOR; i < NR_VECTORS; i++) {
774 /* Some systems map "vectors" to interrupts weirdly. Not us! */
775 __get_cpu_var(vector_irq)[i] = i - FIRST_EXTERNAL_VECTOR;
776 if (i != SYSCALL_VECTOR)
777 set_intr_gate(i, interrupt[i - FIRST_EXTERNAL_VECTOR]);
781 * This call is required to set up for 4k stacks, where we have
782 * separate stacks for hard and soft interrupts.
784 irq_ctx_init(smp_processor_id());
787 void lguest_setup_irq(unsigned int irq)
789 irq_to_desc_alloc_node(irq, 0);
790 set_irq_chip_and_handler_name(irq, &lguest_irq_controller,
791 handle_level_irq, "level");
797 * It would be far better for everyone if the Guest had its own clock, but
798 * until then the Host gives us the time on every interrupt.
800 static unsigned long lguest_get_wallclock(void)
802 return lguest_data.time.tv_sec;
806 * The TSC is an Intel thing called the Time Stamp Counter. The Host tells us
807 * what speed it runs at, or 0 if it's unusable as a reliable clock source.
808 * This matches what we want here: if we return 0 from this function, the x86
809 * TSC clock will give up and not register itself.
811 static unsigned long lguest_tsc_khz(void)
813 return lguest_data.tsc_khz;
817 * If we can't use the TSC, the kernel falls back to our lower-priority
818 * "lguest_clock", where we read the time value given to us by the Host.
820 static cycle_t lguest_clock_read(struct clocksource *cs)
822 unsigned long sec, nsec;
825 * Since the time is in two parts (seconds and nanoseconds), we risk
826 * reading it just as it's changing from 99 & 0.999999999 to 100 and 0,
827 * and getting 99 and 0. As Linux tends to come apart under the stress
828 * of time travel, we must be careful:
831 /* First we read the seconds part. */
832 sec = lguest_data.time.tv_sec;
834 * This read memory barrier tells the compiler and the CPU that
835 * this can't be reordered: we have to complete the above
839 /* Now we read the nanoseconds part. */
840 nsec = lguest_data.time.tv_nsec;
841 /* Make sure we've done that. */
843 /* Now if the seconds part has changed, try again. */
844 } while (unlikely(lguest_data.time.tv_sec != sec));
846 /* Our lguest clock is in real nanoseconds. */
847 return sec*1000000000ULL + nsec;
850 /* This is the fallback clocksource: lower priority than the TSC clocksource. */
851 static struct clocksource lguest_clock = {
854 .read = lguest_clock_read,
855 .mask = CLOCKSOURCE_MASK(64),
858 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
862 * We also need a "struct clock_event_device": Linux asks us to set it to go
863 * off some time in the future. Actually, James Morris figured all this out, I
864 * just applied the patch.
866 static int lguest_clockevent_set_next_event(unsigned long delta,
867 struct clock_event_device *evt)
869 /* FIXME: I don't think this can ever happen, but James tells me he had
870 * to put this code in. Maybe we should remove it now. Anyone? */
871 if (delta < LG_CLOCK_MIN_DELTA) {
872 if (printk_ratelimit())
873 printk(KERN_DEBUG "%s: small delta %lu ns\n",
878 /* Please wake us this far in the future. */
879 kvm_hypercall1(LHCALL_SET_CLOCKEVENT, delta);
883 static void lguest_clockevent_set_mode(enum clock_event_mode mode,
884 struct clock_event_device *evt)
887 case CLOCK_EVT_MODE_UNUSED:
888 case CLOCK_EVT_MODE_SHUTDOWN:
889 /* A 0 argument shuts the clock down. */
890 kvm_hypercall0(LHCALL_SET_CLOCKEVENT);
892 case CLOCK_EVT_MODE_ONESHOT:
893 /* This is what we expect. */
895 case CLOCK_EVT_MODE_PERIODIC:
897 case CLOCK_EVT_MODE_RESUME:
902 /* This describes our primitive timer chip. */
903 static struct clock_event_device lguest_clockevent = {
905 .features = CLOCK_EVT_FEAT_ONESHOT,
906 .set_next_event = lguest_clockevent_set_next_event,
907 .set_mode = lguest_clockevent_set_mode,
911 .min_delta_ns = LG_CLOCK_MIN_DELTA,
912 .max_delta_ns = LG_CLOCK_MAX_DELTA,
916 * This is the Guest timer interrupt handler (hardware interrupt 0). We just
917 * call the clockevent infrastructure and it does whatever needs doing.
919 static void lguest_time_irq(unsigned int irq, struct irq_desc *desc)
923 /* Don't interrupt us while this is running. */
924 local_irq_save(flags);
925 lguest_clockevent.event_handler(&lguest_clockevent);
926 local_irq_restore(flags);
930 * At some point in the boot process, we get asked to set up our timing
931 * infrastructure. The kernel doesn't expect timer interrupts before this, but
932 * we cleverly initialized the "blocked_interrupts" field of "struct
933 * lguest_data" so that timer interrupts were blocked until now.
935 static void lguest_time_init(void)
937 /* Set up the timer interrupt (0) to go to our simple timer routine */
938 set_irq_handler(0, lguest_time_irq);
940 clocksource_register(&lguest_clock);
942 /* We can't set cpumask in the initializer: damn C limitations! Set it
943 * here and register our timer device. */
944 lguest_clockevent.cpumask = cpumask_of(0);
945 clockevents_register_device(&lguest_clockevent);
947 /* Finally, we unblock the timer interrupt. */
948 enable_lguest_irq(0);
952 * Miscellaneous bits and pieces.
954 * Here is an oddball collection of functions which the Guest needs for things
955 * to work. They're pretty simple.
959 * The Guest needs to tell the Host what stack it expects traps to use. For
960 * native hardware, this is part of the Task State Segment mentioned above in
961 * lguest_load_tr_desc(), but to help hypervisors there's this special call.
963 * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data
964 * segment), the privilege level (we're privilege level 1, the Host is 0 and
965 * will not tolerate us trying to use that), the stack pointer, and the number
966 * of pages in the stack.
968 static void lguest_load_sp0(struct tss_struct *tss,
969 struct thread_struct *thread)
971 lazy_hcall3(LHCALL_SET_STACK, __KERNEL_DS | 0x1, thread->sp0,
972 THREAD_SIZE / PAGE_SIZE);
975 /* Let's just say, I wouldn't do debugging under a Guest. */
976 static void lguest_set_debugreg(int regno, unsigned long value)
978 /* FIXME: Implement */
982 * There are times when the kernel wants to make sure that no memory writes are
983 * caught in the cache (that they've all reached real hardware devices). This
984 * doesn't matter for the Guest which has virtual hardware.
986 * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush
987 * (clflush) instruction is available and the kernel uses that. Otherwise, it
988 * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction.
989 * Unlike clflush, wbinvd can only be run at privilege level 0. So we can
990 * ignore clflush, but replace wbinvd.
992 static void lguest_wbinvd(void)
997 * If the Guest expects to have an Advanced Programmable Interrupt Controller,
998 * we play dumb by ignoring writes and returning 0 for reads. So it's no
999 * longer Programmable nor Controlling anything, and I don't think 8 lines of
1000 * code qualifies for Advanced. It will also never interrupt anything. It
1001 * does, however, allow us to get through the Linux boot code.
1003 #ifdef CONFIG_X86_LOCAL_APIC
1004 static void lguest_apic_write(u32 reg, u32 v)
1008 static u32 lguest_apic_read(u32 reg)
1013 static u64 lguest_apic_icr_read(void)
1018 static void lguest_apic_icr_write(u32 low, u32 id)
1020 /* Warn to see if there's any stray references */
1024 static void lguest_apic_wait_icr_idle(void)
1029 static u32 lguest_apic_safe_wait_icr_idle(void)
1034 static void set_lguest_basic_apic_ops(void)
1036 apic->read = lguest_apic_read;
1037 apic->write = lguest_apic_write;
1038 apic->icr_read = lguest_apic_icr_read;
1039 apic->icr_write = lguest_apic_icr_write;
1040 apic->wait_icr_idle = lguest_apic_wait_icr_idle;
1041 apic->safe_wait_icr_idle = lguest_apic_safe_wait_icr_idle;
1045 /* STOP! Until an interrupt comes in. */
1046 static void lguest_safe_halt(void)
1048 kvm_hypercall0(LHCALL_HALT);
1052 * The SHUTDOWN hypercall takes a string to describe what's happening, and
1053 * an argument which says whether this to restart (reboot) the Guest or not.
1055 * Note that the Host always prefers that the Guest speak in physical addresses
1056 * rather than virtual addresses, so we use __pa() here.
1058 static void lguest_power_off(void)
1060 kvm_hypercall2(LHCALL_SHUTDOWN, __pa("Power down"),
1061 LGUEST_SHUTDOWN_POWEROFF);
1067 * Don't. But if you did, this is what happens.
1069 static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p)
1071 kvm_hypercall2(LHCALL_SHUTDOWN, __pa(p), LGUEST_SHUTDOWN_POWEROFF);
1072 /* The hcall won't return, but to keep gcc happy, we're "done". */
1076 static struct notifier_block paniced = {
1077 .notifier_call = lguest_panic
1080 /* Setting up memory is fairly easy. */
1081 static __init char *lguest_memory_setup(void)
1083 /* We do this here and not earlier because lockcheck used to barf if we
1084 * did it before start_kernel(). I think we fixed that, so it'd be
1085 * nice to move it back to lguest_init. Patch welcome... */
1086 atomic_notifier_chain_register(&panic_notifier_list, &paniced);
1089 *The Linux bootloader header contains an "e820" memory map: the
1090 * Launcher populated the first entry with our memory limit.
1092 e820_add_region(boot_params.e820_map[0].addr,
1093 boot_params.e820_map[0].size,
1094 boot_params.e820_map[0].type);
1096 /* This string is for the boot messages. */
1101 * We will eventually use the virtio console device to produce console output,
1102 * but before that is set up we use LHCALL_NOTIFY on normal memory to produce
1105 static __init int early_put_chars(u32 vtermno, const char *buf, int count)
1108 unsigned int len = count;
1110 /* We use a nul-terminated string, so we make a copy. Icky, huh? */
1111 if (len > sizeof(scratch) - 1)
1112 len = sizeof(scratch) - 1;
1113 scratch[len] = '\0';
1114 memcpy(scratch, buf, len);
1115 kvm_hypercall1(LHCALL_NOTIFY, __pa(scratch));
1117 /* This routine returns the number of bytes actually written. */
1122 * Rebooting also tells the Host we're finished, but the RESTART flag tells the
1123 * Launcher to reboot us.
1125 static void lguest_restart(char *reason)
1127 kvm_hypercall2(LHCALL_SHUTDOWN, __pa(reason), LGUEST_SHUTDOWN_RESTART);
1131 * Patching (Powerfully Placating Performance Pedants)
1133 * We have already seen that pv_ops structures let us replace simple native
1134 * instructions with calls to the appropriate back end all throughout the
1135 * kernel. This allows the same kernel to run as a Guest and as a native
1136 * kernel, but it's slow because of all the indirect branches.
1138 * Remember that David Wheeler quote about "Any problem in computer science can
1139 * be solved with another layer of indirection"? The rest of that quote is
1140 * "... But that usually will create another problem." This is the first of
1143 * Our current solution is to allow the paravirt back end to optionally patch
1144 * over the indirect calls to replace them with something more efficient. We
1145 * patch two of the simplest of the most commonly called functions: disable
1146 * interrupts and save interrupts. We usually have 6 or 10 bytes to patch
1147 * into: the Guest versions of these operations are small enough that we can
1150 * First we need assembly templates of each of the patchable Guest operations,
1151 * and these are in i386_head.S.
1154 /*G:060 We construct a table from the assembler templates: */
1155 static const struct lguest_insns
1157 const char *start, *end;
1158 } lguest_insns[] = {
1159 [PARAVIRT_PATCH(pv_irq_ops.irq_disable)] = { lgstart_cli, lgend_cli },
1160 [PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf },
1164 * Now our patch routine is fairly simple (based on the native one in
1165 * paravirt.c). If we have a replacement, we copy it in and return how much of
1166 * the available space we used.
1168 static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf,
1169 unsigned long addr, unsigned len)
1171 unsigned int insn_len;
1173 /* Don't do anything special if we don't have a replacement */
1174 if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start)
1175 return paravirt_patch_default(type, clobber, ibuf, addr, len);
1177 insn_len = lguest_insns[type].end - lguest_insns[type].start;
1179 /* Similarly if it can't fit (doesn't happen, but let's be thorough). */
1181 return paravirt_patch_default(type, clobber, ibuf, addr, len);
1183 /* Copy in our instructions. */
1184 memcpy(ibuf, lguest_insns[type].start, insn_len);
1189 * Once we get to lguest_init(), we know we're a Guest. The various
1190 * pv_ops structures in the kernel provide points for (almost) every routine we
1191 * have to override to avoid privileged instructions.
1193 __init void lguest_init(void)
1195 /* We're under lguest. */
1196 pv_info.name = "lguest";
1197 /* Paravirt is enabled. */
1198 pv_info.paravirt_enabled = 1;
1199 /* We're running at privilege level 1, not 0 as normal. */
1200 pv_info.kernel_rpl = 1;
1201 /* Everyone except Xen runs with this set. */
1202 pv_info.shared_kernel_pmd = 1;
1205 * We set up all the lguest overrides for sensitive operations. These
1206 * are detailed with the operations themselves.
1209 /* Interrupt-related operations */
1210 pv_irq_ops.init_IRQ = lguest_init_IRQ;
1211 pv_irq_ops.save_fl = PV_CALLEE_SAVE(save_fl);
1212 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(lg_restore_fl);
1213 pv_irq_ops.irq_disable = PV_CALLEE_SAVE(irq_disable);
1214 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(lg_irq_enable);
1215 pv_irq_ops.safe_halt = lguest_safe_halt;
1217 /* Setup operations */
1218 pv_init_ops.memory_setup = lguest_memory_setup;
1219 pv_init_ops.patch = lguest_patch;
1221 /* Intercepts of various CPU instructions */
1222 pv_cpu_ops.load_gdt = lguest_load_gdt;
1223 pv_cpu_ops.cpuid = lguest_cpuid;
1224 pv_cpu_ops.load_idt = lguest_load_idt;
1225 pv_cpu_ops.iret = lguest_iret;
1226 pv_cpu_ops.load_sp0 = lguest_load_sp0;
1227 pv_cpu_ops.load_tr_desc = lguest_load_tr_desc;
1228 pv_cpu_ops.set_ldt = lguest_set_ldt;
1229 pv_cpu_ops.load_tls = lguest_load_tls;
1230 pv_cpu_ops.set_debugreg = lguest_set_debugreg;
1231 pv_cpu_ops.clts = lguest_clts;
1232 pv_cpu_ops.read_cr0 = lguest_read_cr0;
1233 pv_cpu_ops.write_cr0 = lguest_write_cr0;
1234 pv_cpu_ops.read_cr4 = lguest_read_cr4;
1235 pv_cpu_ops.write_cr4 = lguest_write_cr4;
1236 pv_cpu_ops.write_gdt_entry = lguest_write_gdt_entry;
1237 pv_cpu_ops.write_idt_entry = lguest_write_idt_entry;
1238 pv_cpu_ops.wbinvd = lguest_wbinvd;
1239 pv_cpu_ops.start_context_switch = paravirt_start_context_switch;
1240 pv_cpu_ops.end_context_switch = lguest_end_context_switch;
1242 /* Pagetable management */
1243 pv_mmu_ops.write_cr3 = lguest_write_cr3;
1244 pv_mmu_ops.flush_tlb_user = lguest_flush_tlb_user;
1245 pv_mmu_ops.flush_tlb_single = lguest_flush_tlb_single;
1246 pv_mmu_ops.flush_tlb_kernel = lguest_flush_tlb_kernel;
1247 pv_mmu_ops.set_pte = lguest_set_pte;
1248 pv_mmu_ops.set_pte_at = lguest_set_pte_at;
1249 pv_mmu_ops.set_pmd = lguest_set_pmd;
1250 #ifdef CONFIG_X86_PAE
1251 pv_mmu_ops.set_pte_atomic = lguest_set_pte_atomic;
1252 pv_mmu_ops.pte_clear = lguest_pte_clear;
1253 pv_mmu_ops.pmd_clear = lguest_pmd_clear;
1254 pv_mmu_ops.set_pud = lguest_set_pud;
1256 pv_mmu_ops.read_cr2 = lguest_read_cr2;
1257 pv_mmu_ops.read_cr3 = lguest_read_cr3;
1258 pv_mmu_ops.lazy_mode.enter = paravirt_enter_lazy_mmu;
1259 pv_mmu_ops.lazy_mode.leave = lguest_leave_lazy_mmu_mode;
1260 pv_mmu_ops.pte_update = lguest_pte_update;
1261 pv_mmu_ops.pte_update_defer = lguest_pte_update;
1263 #ifdef CONFIG_X86_LOCAL_APIC
1264 /* APIC read/write intercepts */
1265 set_lguest_basic_apic_ops();
1268 /* Time operations */
1269 pv_time_ops.get_wallclock = lguest_get_wallclock;
1270 pv_time_ops.time_init = lguest_time_init;
1271 pv_time_ops.get_tsc_khz = lguest_tsc_khz;
1274 * Now is a good time to look at the implementations of these functions
1275 * before returning to the rest of lguest_init().
1279 * Now we've seen all the paravirt_ops, we return to
1280 * lguest_init() where the rest of the fairly chaotic boot setup
1285 * The stack protector is a weird thing where gcc places a canary
1286 * value on the stack and then checks it on return. This file is
1287 * compiled with -fno-stack-protector it, so we got this far without
1288 * problems. The value of the canary is kept at offset 20 from the
1289 * %gs register, so we need to set that up before calling C functions
1292 setup_stack_canary_segment(0);
1295 * We could just call load_stack_canary_segment(), but we might as well
1296 * call switch_to_new_gdt() which loads the whole table and sets up the
1297 * per-cpu segment descriptor register %fs as well.
1299 switch_to_new_gdt(0);
1301 /* As described in head_32.S, we map the first 128M of memory. */
1302 max_pfn_mapped = (128*1024*1024) >> PAGE_SHIFT;
1305 * The Host<->Guest Switcher lives at the top of our address space, and
1306 * the Host told us how big it is when we made LGUEST_INIT hypercall:
1307 * it put the answer in lguest_data.reserve_mem
1309 reserve_top_address(lguest_data.reserve_mem);
1312 * If we don't initialize the lock dependency checker now, it crashes
1313 * paravirt_disable_iospace.
1318 * The IDE code spends about 3 seconds probing for disks: if we reserve
1319 * all the I/O ports up front it can't get them and so doesn't probe.
1320 * Other device drivers are similar (but less severe). This cuts the
1321 * kernel boot time on my machine from 4.1 seconds to 0.45 seconds.
1323 paravirt_disable_iospace();
1326 * This is messy CPU setup stuff which the native boot code does before
1327 * start_kernel, so we have to do, too:
1329 cpu_detect(&new_cpu_data);
1330 /* head.S usually sets up the first capability word, so do it here. */
1331 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1333 /* Math is always hard! */
1334 new_cpu_data.hard_math = 1;
1336 /* We don't have features. We have puppies! Puppies! */
1337 #ifdef CONFIG_X86_MCE
1346 * We set the preferred console to "hvc". This is the "hypervisor
1347 * virtual console" driver written by the PowerPC people, which we also
1348 * adapted for lguest's use.
1350 add_preferred_console("hvc", 0, NULL);
1352 /* Register our very early console. */
1353 virtio_cons_early_init(early_put_chars);
1356 * Last of all, we set the power management poweroff hook to point to
1357 * the Guest routine to power off, and the reboot hook to our restart
1360 pm_power_off = lguest_power_off;
1361 machine_ops.restart = lguest_restart;
1364 * Now we're set up, call i386_start_kernel() in head32.c and we proceed
1365 * to boot as normal. It never returns.
1367 i386_start_kernel();
1370 * This marks the end of stage II of our journey, The Guest.
1372 * It is now time for us to explore the layer of virtual drivers and complete
1373 * our understanding of the Guest in "make Drivers".