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. :*/
28 * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation.
30 * This program is free software; you can redistribute it and/or modify
31 * it under the terms of the GNU General Public License as published by
32 * the Free Software Foundation; either version 2 of the License, or
33 * (at your option) any later version.
35 * This program is distributed in the hope that it will be useful, but
36 * WITHOUT ANY WARRANTY; without even the implied warranty of
37 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
38 * NON INFRINGEMENT. See the GNU General Public License for more
41 * You should have received a copy of the GNU General Public License
42 * along with this program; if not, write to the Free Software
43 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
45 #include <linux/kernel.h>
46 #include <linux/start_kernel.h>
47 #include <linux/string.h>
48 #include <linux/console.h>
49 #include <linux/screen_info.h>
50 #include <linux/irq.h>
51 #include <linux/interrupt.h>
52 #include <linux/clocksource.h>
53 #include <linux/clockchips.h>
54 #include <linux/lguest.h>
55 #include <linux/lguest_launcher.h>
56 #include <linux/virtio_console.h>
58 #include <asm/lguest.h>
59 #include <asm/paravirt.h>
60 #include <asm/param.h>
62 #include <asm/pgtable.h>
64 #include <asm/setup.h>
69 #include <asm/reboot.h> /* for struct machine_ops */
71 /*G:010 Welcome to the Guest!
73 * The Guest in our tale is a simple creature: identical to the Host but
74 * behaving in simplified but equivalent ways. In particular, the Guest is the
75 * same kernel as the Host (or at least, built from the same source code). :*/
77 struct lguest_data lguest_data = {
78 .hcall_status = { [0 ... LHCALL_RING_SIZE-1] = 0xFF },
79 .noirq_start = (u32)lguest_noirq_start,
80 .noirq_end = (u32)lguest_noirq_end,
81 .kernel_address = PAGE_OFFSET,
82 .blocked_interrupts = { 1 }, /* Block timer interrupts */
83 .syscall_vec = SYSCALL_VECTOR,
86 /*G:037 async_hcall() is pretty simple: I'm quite proud of it really. We have a
87 * ring buffer of stored hypercalls which the Host will run though next time we
88 * do a normal hypercall. Each entry in the ring has 4 slots for the hypercall
89 * arguments, and a "hcall_status" word which is 0 if the call is ready to go,
90 * and 255 once the Host has finished with it.
92 * If we come around to a slot which hasn't been finished, then the table is
93 * full and we just make the hypercall directly. This has the nice side
94 * effect of causing the Host to run all the stored calls in the ring buffer
95 * which empties it for next time! */
96 static void async_hcall(unsigned long call, unsigned long arg1,
97 unsigned long arg2, unsigned long arg3)
99 /* Note: This code assumes we're uniprocessor. */
100 static unsigned int next_call;
103 /* Disable interrupts if not already disabled: we don't want an
104 * interrupt handler making a hypercall while we're already doing
106 local_irq_save(flags);
107 if (lguest_data.hcall_status[next_call] != 0xFF) {
108 /* Table full, so do normal hcall which will flush table. */
109 hcall(call, arg1, arg2, arg3);
111 lguest_data.hcalls[next_call].arg0 = call;
112 lguest_data.hcalls[next_call].arg1 = arg1;
113 lguest_data.hcalls[next_call].arg2 = arg2;
114 lguest_data.hcalls[next_call].arg3 = arg3;
115 /* Arguments must all be written before we mark it to go */
117 lguest_data.hcall_status[next_call] = 0;
118 if (++next_call == LHCALL_RING_SIZE)
121 local_irq_restore(flags);
124 /*G:035 Notice the lazy_hcall() above, rather than hcall(). This is our first
125 * real optimization trick!
127 * When lazy_mode is set, it means we're allowed to defer all hypercalls and do
128 * them as a batch when lazy_mode is eventually turned off. Because hypercalls
129 * are reasonably expensive, batching them up makes sense. For example, a
130 * large munmap might update dozens of page table entries: that code calls
131 * paravirt_enter_lazy_mmu(), does the dozen updates, then calls
132 * lguest_leave_lazy_mode().
134 * So, when we're in lazy mode, we call async_hcall() to store the call for
135 * future processing: */
136 static void lazy_hcall(unsigned long call,
141 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
142 hcall(call, arg1, arg2, arg3);
144 async_hcall(call, arg1, arg2, arg3);
147 /* When lazy mode is turned off reset the per-cpu lazy mode variable and then
148 * issue the do-nothing hypercall to flush any stored calls. */
149 static void lguest_leave_lazy_mode(void)
151 paravirt_leave_lazy(paravirt_get_lazy_mode());
152 hcall(LHCALL_FLUSH_ASYNC, 0, 0, 0);
156 * After that diversion we return to our first native-instruction
157 * replacements: four functions for interrupt control.
159 * The simplest way of implementing these would be to have "turn interrupts
160 * off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow:
161 * these are by far the most commonly called functions of those we override.
163 * So instead we keep an "irq_enabled" field inside our "struct lguest_data",
164 * which the Guest can update with a single instruction. The Host knows to
165 * check there before it tries to deliver an interrupt.
168 /* save_flags() is expected to return the processor state (ie. "flags"). The
169 * flags word contains all kind of stuff, but in practice Linux only cares
170 * about the interrupt flag. Our "save_flags()" just returns that. */
171 static unsigned long save_fl(void)
173 return lguest_data.irq_enabled;
176 /* restore_flags() just sets the flags back to the value given. */
177 static void restore_fl(unsigned long flags)
179 lguest_data.irq_enabled = flags;
182 /* Interrupts go off... */
183 static void irq_disable(void)
185 lguest_data.irq_enabled = 0;
188 /* Interrupts go on... */
189 static void irq_enable(void)
191 lguest_data.irq_enabled = X86_EFLAGS_IF;
194 /*M:003 Note that we don't check for outstanding interrupts when we re-enable
195 * them (or when we unmask an interrupt). This seems to work for the moment,
196 * since interrupts are rare and we'll just get the interrupt on the next timer
197 * tick, but now we can run with CONFIG_NO_HZ, we should revisit this. One way
198 * would be to put the "irq_enabled" field in a page by itself, and have the
199 * Host write-protect it when an interrupt comes in when irqs are disabled.
200 * There will then be a page fault as soon as interrupts are re-enabled.
202 * A better method is to implement soft interrupt disable generally for x86:
203 * instead of disabling interrupts, we set a flag. If an interrupt does come
204 * in, we then disable them for real. This is uncommon, so we could simply use
205 * a hypercall for interrupt control and not worry about efficiency. :*/
208 * The Interrupt Descriptor Table (IDT).
210 * The IDT tells the processor what to do when an interrupt comes in. Each
211 * entry in the table is a 64-bit descriptor: this holds the privilege level,
212 * address of the handler, and... well, who cares? The Guest just asks the
213 * Host to make the change anyway, because the Host controls the real IDT.
215 static void lguest_write_idt_entry(gate_desc *dt,
216 int entrynum, const gate_desc *g)
218 /* The gate_desc structure is 8 bytes long: we hand it to the Host in
219 * two 32-bit chunks. The whole 32-bit kernel used to hand descriptors
220 * around like this; typesafety wasn't a big concern in Linux's early
222 u32 *desc = (u32 *)g;
223 /* Keep the local copy up to date. */
224 native_write_idt_entry(dt, entrynum, g);
225 /* Tell Host about this new entry. */
226 hcall(LHCALL_LOAD_IDT_ENTRY, entrynum, desc[0], desc[1]);
229 /* Changing to a different IDT is very rare: we keep the IDT up-to-date every
230 * time it is written, so we can simply loop through all entries and tell the
231 * Host about them. */
232 static void lguest_load_idt(const struct desc_ptr *desc)
235 struct desc_struct *idt = (void *)desc->address;
237 for (i = 0; i < (desc->size+1)/8; i++)
238 hcall(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b);
242 * The Global Descriptor Table.
244 * The Intel architecture defines another table, called the Global Descriptor
245 * Table (GDT). You tell the CPU where it is (and its size) using the "lgdt"
246 * instruction, and then several other instructions refer to entries in the
247 * table. There are three entries which the Switcher needs, so the Host simply
248 * controls the entire thing and the Guest asks it to make changes using the
249 * LOAD_GDT hypercall.
251 * This is the opposite of the IDT code where we have a LOAD_IDT_ENTRY
252 * hypercall and use that repeatedly to load a new IDT. I don't think it
253 * really matters, but wouldn't it be nice if they were the same? Wouldn't
254 * it be even better if you were the one to send the patch to fix it?
256 static void lguest_load_gdt(const struct desc_ptr *desc)
258 BUG_ON((desc->size+1)/8 != GDT_ENTRIES);
259 hcall(LHCALL_LOAD_GDT, __pa(desc->address), GDT_ENTRIES, 0);
262 /* For a single GDT entry which changes, we do the lazy thing: alter our GDT,
263 * then tell the Host to reload the entire thing. This operation is so rare
264 * that this naive implementation is reasonable. */
265 static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum,
266 const void *desc, int type)
268 native_write_gdt_entry(dt, entrynum, desc, type);
269 hcall(LHCALL_LOAD_GDT, __pa(dt), GDT_ENTRIES, 0);
272 /* OK, I lied. There are three "thread local storage" GDT entries which change
273 * on every context switch (these three entries are how glibc implements
274 * __thread variables). So we have a hypercall specifically for this case. */
275 static void lguest_load_tls(struct thread_struct *t, unsigned int cpu)
277 /* There's one problem which normal hardware doesn't have: the Host
278 * can't handle us removing entries we're currently using. So we clear
279 * the GS register here: if it's needed it'll be reloaded anyway. */
281 lazy_hcall(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu, 0);
284 /*G:038 That's enough excitement for now, back to ploughing through each of
285 * the different pv_ops structures (we're about 1/3 of the way through).
287 * This is the Local Descriptor Table, another weird Intel thingy. Linux only
288 * uses this for some strange applications like Wine. We don't do anything
289 * here, so they'll get an informative and friendly Segmentation Fault. */
290 static void lguest_set_ldt(const void *addr, unsigned entries)
294 /* This loads a GDT entry into the "Task Register": that entry points to a
295 * structure called the Task State Segment. Some comments scattered though the
296 * kernel code indicate that this used for task switching in ages past, along
297 * with blood sacrifice and astrology.
299 * Now there's nothing interesting in here that we don't get told elsewhere.
300 * But the native version uses the "ltr" instruction, which makes the Host
301 * complain to the Guest about a Segmentation Fault and it'll oops. So we
302 * override the native version with a do-nothing version. */
303 static void lguest_load_tr_desc(void)
307 /* The "cpuid" instruction is a way of querying both the CPU identity
308 * (manufacturer, model, etc) and its features. It was introduced before the
309 * Pentium in 1993 and keeps getting extended by both Intel, AMD and others.
310 * As you might imagine, after a decade and a half this treatment, it is now a
311 * giant ball of hair. Its entry in the current Intel manual runs to 28 pages.
313 * This instruction even it has its own Wikipedia entry. The Wikipedia entry
314 * has been translated into 4 languages. I am not making this up!
316 * We could get funky here and identify ourselves as "GenuineLguest", but
317 * instead we just use the real "cpuid" instruction. Then I pretty much turned
318 * off feature bits until the Guest booted. (Don't say that: you'll damage
319 * lguest sales!) Shut up, inner voice! (Hey, just pointing out that this is
320 * hardly future proof.) Noone's listening! They don't like you anyway,
321 * parenthetic weirdo!
323 * Replacing the cpuid so we can turn features off is great for the kernel, but
324 * anyone (including userspace) can just use the raw "cpuid" instruction and
325 * the Host won't even notice since it isn't privileged. So we try not to get
326 * too worked up about it. */
327 static void lguest_cpuid(unsigned int *ax, unsigned int *bx,
328 unsigned int *cx, unsigned int *dx)
332 native_cpuid(ax, bx, cx, dx);
334 case 1: /* Basic feature request. */
335 /* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */
337 /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU. */
339 /* The Host can do a nice optimization if it knows that the
340 * kernel mappings (addresses above 0xC0000000 or whatever
341 * PAGE_OFFSET is set to) haven't changed. But Linux calls
342 * flush_tlb_user() for both user and kernel mappings unless
343 * the Page Global Enable (PGE) feature bit is set. */
347 /* Futureproof this a little: if they ask how much extended
348 * processor information there is, limit it to known fields. */
349 if (*ax > 0x80000008)
355 /* Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4.
356 * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother
357 * it. The Host needs to know when the Guest wants to change them, so we have
358 * a whole series of functions like read_cr0() and write_cr0().
360 * We start with cr0. cr0 allows you to turn on and off all kinds of basic
361 * features, but Linux only really cares about one: the horrifically-named Task
362 * Switched (TS) bit at bit 3 (ie. 8)
364 * What does the TS bit do? Well, it causes the CPU to trap (interrupt 7) if
365 * the floating point unit is used. Which allows us to restore FPU state
366 * lazily after a task switch, and Linux uses that gratefully, but wouldn't a
367 * name like "FPUTRAP bit" be a little less cryptic?
369 * We store cr0 (and cr3) locally, because the Host never changes it. The
370 * Guest sometimes wants to read it and we'd prefer not to bother the Host
372 static unsigned long current_cr0, current_cr3;
373 static void lguest_write_cr0(unsigned long val)
375 lazy_hcall(LHCALL_TS, val & X86_CR0_TS, 0, 0);
379 static unsigned long lguest_read_cr0(void)
384 /* Intel provided a special instruction to clear the TS bit for people too cool
385 * to use write_cr0() to do it. This "clts" instruction is faster, because all
386 * the vowels have been optimized out. */
387 static void lguest_clts(void)
389 lazy_hcall(LHCALL_TS, 0, 0, 0);
390 current_cr0 &= ~X86_CR0_TS;
393 /* cr2 is the virtual address of the last page fault, which the Guest only ever
394 * reads. The Host kindly writes this into our "struct lguest_data", so we
395 * just read it out of there. */
396 static unsigned long lguest_read_cr2(void)
398 return lguest_data.cr2;
401 /* cr3 is the current toplevel pagetable page: the principle is the same as
402 * cr0. Keep a local copy, and tell the Host when it changes. */
403 static void lguest_write_cr3(unsigned long cr3)
405 lazy_hcall(LHCALL_NEW_PGTABLE, cr3, 0, 0);
409 static unsigned long lguest_read_cr3(void)
414 /* cr4 is used to enable and disable PGE, but we don't care. */
415 static unsigned long lguest_read_cr4(void)
420 static void lguest_write_cr4(unsigned long val)
425 * Page Table Handling.
427 * Now would be a good time to take a rest and grab a coffee or similarly
428 * relaxing stimulant. The easy parts are behind us, and the trek gradually
429 * winds uphill from here.
431 * Quick refresher: memory is divided into "pages" of 4096 bytes each. The CPU
432 * maps virtual addresses to physical addresses using "page tables". We could
433 * use one huge index of 1 million entries: each address is 4 bytes, so that's
434 * 1024 pages just to hold the page tables. But since most virtual addresses
435 * are unused, we use a two level index which saves space. The cr3 register
436 * contains the physical address of the top level "page directory" page, which
437 * contains physical addresses of up to 1024 second-level pages. Each of these
438 * second level pages contains up to 1024 physical addresses of actual pages,
439 * or Page Table Entries (PTEs).
441 * Here's a diagram, where arrows indicate physical addresses:
443 * cr3 ---> +---------+
444 * | --------->+---------+
446 * Top-level | | PADDR2 |
453 * So to convert a virtual address to a physical address, we look up the top
454 * level, which points us to the second level, which gives us the physical
455 * address of that page. If the top level entry was not present, or the second
456 * level entry was not present, then the virtual address is invalid (we
457 * say "the page was not mapped").
459 * Put another way, a 32-bit virtual address is divided up like so:
461 * 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
462 * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>|
463 * Index into top Index into second Offset within page
464 * page directory page pagetable page
466 * The kernel spends a lot of time changing both the top-level page directory
467 * and lower-level pagetable pages. The Guest doesn't know physical addresses,
468 * so while it maintains these page tables exactly like normal, it also needs
469 * to keep the Host informed whenever it makes a change: the Host will create
470 * the real page tables based on the Guests'.
473 /* The Guest calls this to set a second-level entry (pte), ie. to map a page
474 * into a process' address space. We set the entry then tell the Host the
475 * toplevel and address this corresponds to. The Guest uses one pagetable per
476 * process, so we need to tell the Host which one we're changing (mm->pgd). */
477 static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr,
478 pte_t *ptep, pte_t pteval)
481 lazy_hcall(LHCALL_SET_PTE, __pa(mm->pgd), addr, pteval.pte_low);
484 /* The Guest calls this to set a top-level entry. Again, we set the entry then
485 * tell the Host which top-level page we changed, and the index of the entry we
487 static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
490 lazy_hcall(LHCALL_SET_PMD, __pa(pmdp)&PAGE_MASK,
491 (__pa(pmdp)&(PAGE_SIZE-1))/4, 0);
494 /* There are a couple of legacy places where the kernel sets a PTE, but we
495 * don't know the top level any more. This is useless for us, since we don't
496 * know which pagetable is changing or what address, so we just tell the Host
497 * to forget all of them. Fortunately, this is very rare.
499 * ... except in early boot when the kernel sets up the initial pagetables,
500 * which makes booting astonishingly slow. So we don't even tell the Host
501 * anything changed until we've done the first page table switch. */
502 static void lguest_set_pte(pte_t *ptep, pte_t pteval)
505 /* Don't bother with hypercall before initial setup. */
507 lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0);
510 /* Unfortunately for Lguest, the pv_mmu_ops for page tables were based on
511 * native page table operations. On native hardware you can set a new page
512 * table entry whenever you want, but if you want to remove one you have to do
513 * a TLB flush (a TLB is a little cache of page table entries kept by the CPU).
515 * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only
516 * called when a valid entry is written, not when it's removed (ie. marked not
517 * present). Instead, this is where we come when the Guest wants to remove a
518 * page table entry: we tell the Host to set that entry to 0 (ie. the present
520 static void lguest_flush_tlb_single(unsigned long addr)
522 /* Simply set it to zero: if it was not, it will fault back in. */
523 lazy_hcall(LHCALL_SET_PTE, current_cr3, addr, 0);
526 /* This is what happens after the Guest has removed a large number of entries.
527 * This tells the Host that any of the page table entries for userspace might
528 * have changed, ie. virtual addresses below PAGE_OFFSET. */
529 static void lguest_flush_tlb_user(void)
531 lazy_hcall(LHCALL_FLUSH_TLB, 0, 0, 0);
534 /* This is called when the kernel page tables have changed. That's not very
535 * common (unless the Guest is using highmem, which makes the Guest extremely
536 * slow), so it's worth separating this from the user flushing above. */
537 static void lguest_flush_tlb_kernel(void)
539 lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0);
543 * The Unadvanced Programmable Interrupt Controller.
545 * This is an attempt to implement the simplest possible interrupt controller.
546 * I spent some time looking though routines like set_irq_chip_and_handler,
547 * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and
548 * I *think* this is as simple as it gets.
550 * We can tell the Host what interrupts we want blocked ready for using the
551 * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as
552 * simple as setting a bit. We don't actually "ack" interrupts as such, we
553 * just mask and unmask them. I wonder if we should be cleverer?
555 static void disable_lguest_irq(unsigned int irq)
557 set_bit(irq, lguest_data.blocked_interrupts);
560 static void enable_lguest_irq(unsigned int irq)
562 clear_bit(irq, lguest_data.blocked_interrupts);
565 /* This structure describes the lguest IRQ controller. */
566 static struct irq_chip lguest_irq_controller = {
568 .mask = disable_lguest_irq,
569 .mask_ack = disable_lguest_irq,
570 .unmask = enable_lguest_irq,
573 /* This sets up the Interrupt Descriptor Table (IDT) entry for each hardware
574 * interrupt (except 128, which is used for system calls), and then tells the
575 * Linux infrastructure that each interrupt is controlled by our level-based
576 * lguest interrupt controller. */
577 static void __init lguest_init_IRQ(void)
581 for (i = 0; i < LGUEST_IRQS; i++) {
582 int vector = FIRST_EXTERNAL_VECTOR + i;
583 if (vector != SYSCALL_VECTOR) {
584 set_intr_gate(vector, interrupt[i]);
585 set_irq_chip_and_handler_name(i, &lguest_irq_controller,
590 /* This call is required to set up for 4k stacks, where we have
591 * separate stacks for hard and soft interrupts. */
592 irq_ctx_init(smp_processor_id());
598 * It would be far better for everyone if the Guest had its own clock, but
599 * until then the Host gives us the time on every interrupt.
601 static unsigned long lguest_get_wallclock(void)
603 return lguest_data.time.tv_sec;
606 /* The TSC is an Intel thing called the Time Stamp Counter. The Host tells us
607 * what speed it runs at, or 0 if it's unusable as a reliable clock source.
608 * This matches what we want here: if we return 0 from this function, the x86
609 * TSC clock will give up and not register itself. */
610 static unsigned long lguest_tsc_khz(void)
612 return lguest_data.tsc_khz;
615 /* If we can't use the TSC, the kernel falls back to our lower-priority
616 * "lguest_clock", where we read the time value given to us by the Host. */
617 static cycle_t lguest_clock_read(void)
619 unsigned long sec, nsec;
621 /* Since the time is in two parts (seconds and nanoseconds), we risk
622 * reading it just as it's changing from 99 & 0.999999999 to 100 and 0,
623 * and getting 99 and 0. As Linux tends to come apart under the stress
624 * of time travel, we must be careful: */
626 /* First we read the seconds part. */
627 sec = lguest_data.time.tv_sec;
628 /* This read memory barrier tells the compiler and the CPU that
629 * this can't be reordered: we have to complete the above
630 * before going on. */
632 /* Now we read the nanoseconds part. */
633 nsec = lguest_data.time.tv_nsec;
634 /* Make sure we've done that. */
636 /* Now if the seconds part has changed, try again. */
637 } while (unlikely(lguest_data.time.tv_sec != sec));
639 /* Our lguest clock is in real nanoseconds. */
640 return sec*1000000000ULL + nsec;
643 /* This is the fallback clocksource: lower priority than the TSC clocksource. */
644 static struct clocksource lguest_clock = {
647 .read = lguest_clock_read,
648 .mask = CLOCKSOURCE_MASK(64),
651 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
654 /* We also need a "struct clock_event_device": Linux asks us to set it to go
655 * off some time in the future. Actually, James Morris figured all this out, I
656 * just applied the patch. */
657 static int lguest_clockevent_set_next_event(unsigned long delta,
658 struct clock_event_device *evt)
660 /* FIXME: I don't think this can ever happen, but James tells me he had
661 * to put this code in. Maybe we should remove it now. Anyone? */
662 if (delta < LG_CLOCK_MIN_DELTA) {
663 if (printk_ratelimit())
664 printk(KERN_DEBUG "%s: small delta %lu ns\n",
669 /* Please wake us this far in the future. */
670 hcall(LHCALL_SET_CLOCKEVENT, delta, 0, 0);
674 static void lguest_clockevent_set_mode(enum clock_event_mode mode,
675 struct clock_event_device *evt)
678 case CLOCK_EVT_MODE_UNUSED:
679 case CLOCK_EVT_MODE_SHUTDOWN:
680 /* A 0 argument shuts the clock down. */
681 hcall(LHCALL_SET_CLOCKEVENT, 0, 0, 0);
683 case CLOCK_EVT_MODE_ONESHOT:
684 /* This is what we expect. */
686 case CLOCK_EVT_MODE_PERIODIC:
688 case CLOCK_EVT_MODE_RESUME:
693 /* This describes our primitive timer chip. */
694 static struct clock_event_device lguest_clockevent = {
696 .features = CLOCK_EVT_FEAT_ONESHOT,
697 .set_next_event = lguest_clockevent_set_next_event,
698 .set_mode = lguest_clockevent_set_mode,
702 .min_delta_ns = LG_CLOCK_MIN_DELTA,
703 .max_delta_ns = LG_CLOCK_MAX_DELTA,
706 /* This is the Guest timer interrupt handler (hardware interrupt 0). We just
707 * call the clockevent infrastructure and it does whatever needs doing. */
708 static void lguest_time_irq(unsigned int irq, struct irq_desc *desc)
712 /* Don't interrupt us while this is running. */
713 local_irq_save(flags);
714 lguest_clockevent.event_handler(&lguest_clockevent);
715 local_irq_restore(flags);
718 /* At some point in the boot process, we get asked to set up our timing
719 * infrastructure. The kernel doesn't expect timer interrupts before this, but
720 * we cleverly initialized the "blocked_interrupts" field of "struct
721 * lguest_data" so that timer interrupts were blocked until now. */
722 static void lguest_time_init(void)
724 /* Set up the timer interrupt (0) to go to our simple timer routine */
725 set_irq_handler(0, lguest_time_irq);
727 clocksource_register(&lguest_clock);
729 /* We can't set cpumask in the initializer: damn C limitations! Set it
730 * here and register our timer device. */
731 lguest_clockevent.cpumask = cpumask_of_cpu(0);
732 clockevents_register_device(&lguest_clockevent);
734 /* Finally, we unblock the timer interrupt. */
735 enable_lguest_irq(0);
739 * Miscellaneous bits and pieces.
741 * Here is an oddball collection of functions which the Guest needs for things
742 * to work. They're pretty simple.
745 /* The Guest needs to tell the Host what stack it expects traps to use. For
746 * native hardware, this is part of the Task State Segment mentioned above in
747 * lguest_load_tr_desc(), but to help hypervisors there's this special call.
749 * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data
750 * segment), the privilege level (we're privilege level 1, the Host is 0 and
751 * will not tolerate us trying to use that), the stack pointer, and the number
752 * of pages in the stack. */
753 static void lguest_load_sp0(struct tss_struct *tss,
754 struct thread_struct *thread)
756 lazy_hcall(LHCALL_SET_STACK, __KERNEL_DS|0x1, thread->sp0,
757 THREAD_SIZE/PAGE_SIZE);
760 /* Let's just say, I wouldn't do debugging under a Guest. */
761 static void lguest_set_debugreg(int regno, unsigned long value)
763 /* FIXME: Implement */
766 /* There are times when the kernel wants to make sure that no memory writes are
767 * caught in the cache (that they've all reached real hardware devices). This
768 * doesn't matter for the Guest which has virtual hardware.
770 * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush
771 * (clflush) instruction is available and the kernel uses that. Otherwise, it
772 * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction.
773 * Unlike clflush, wbinvd can only be run at privilege level 0. So we can
774 * ignore clflush, but replace wbinvd.
776 static void lguest_wbinvd(void)
780 /* If the Guest expects to have an Advanced Programmable Interrupt Controller,
781 * we play dumb by ignoring writes and returning 0 for reads. So it's no
782 * longer Programmable nor Controlling anything, and I don't think 8 lines of
783 * code qualifies for Advanced. It will also never interrupt anything. It
784 * does, however, allow us to get through the Linux boot code. */
785 #ifdef CONFIG_X86_LOCAL_APIC
786 static void lguest_apic_write(unsigned long reg, u32 v)
790 static u32 lguest_apic_read(unsigned long reg)
796 /* STOP! Until an interrupt comes in. */
797 static void lguest_safe_halt(void)
799 hcall(LHCALL_HALT, 0, 0, 0);
802 /* The SHUTDOWN hypercall takes a string to describe what's happening, and
803 * an argument which says whether this to restart (reboot) the Guest or not.
805 * Note that the Host always prefers that the Guest speak in physical addresses
806 * rather than virtual addresses, so we use __pa() here. */
807 static void lguest_power_off(void)
809 hcall(LHCALL_SHUTDOWN, __pa("Power down"), LGUEST_SHUTDOWN_POWEROFF, 0);
815 * Don't. But if you did, this is what happens.
817 static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p)
819 hcall(LHCALL_SHUTDOWN, __pa(p), LGUEST_SHUTDOWN_POWEROFF, 0);
820 /* The hcall won't return, but to keep gcc happy, we're "done". */
824 static struct notifier_block paniced = {
825 .notifier_call = lguest_panic
828 /* Setting up memory is fairly easy. */
829 static __init char *lguest_memory_setup(void)
831 /* We do this here and not earlier because lockcheck used to barf if we
832 * did it before start_kernel(). I think we fixed that, so it'd be
833 * nice to move it back to lguest_init. Patch welcome... */
834 atomic_notifier_chain_register(&panic_notifier_list, &paniced);
836 /* The Linux bootloader header contains an "e820" memory map: the
837 * Launcher populated the first entry with our memory limit. */
838 e820_add_region(boot_params.e820_map[0].addr,
839 boot_params.e820_map[0].size,
840 boot_params.e820_map[0].type);
842 /* This string is for the boot messages. */
846 /* We will eventually use the virtio console device to produce console output,
847 * but before that is set up we use LHCALL_NOTIFY on normal memory to produce
849 static __init int early_put_chars(u32 vtermno, const char *buf, int count)
852 unsigned int len = count;
854 /* We use a nul-terminated string, so we have to make a copy. Icky,
856 if (len > sizeof(scratch) - 1)
857 len = sizeof(scratch) - 1;
859 memcpy(scratch, buf, len);
860 hcall(LHCALL_NOTIFY, __pa(scratch), 0, 0);
862 /* This routine returns the number of bytes actually written. */
866 /* Rebooting also tells the Host we're finished, but the RESTART flag tells the
867 * Launcher to reboot us. */
868 static void lguest_restart(char *reason)
870 hcall(LHCALL_SHUTDOWN, __pa(reason), LGUEST_SHUTDOWN_RESTART, 0);
874 * Patching (Powerfully Placating Performance Pedants)
876 * We have already seen that pv_ops structures let us replace simple native
877 * instructions with calls to the appropriate back end all throughout the
878 * kernel. This allows the same kernel to run as a Guest and as a native
879 * kernel, but it's slow because of all the indirect branches.
881 * Remember that David Wheeler quote about "Any problem in computer science can
882 * be solved with another layer of indirection"? The rest of that quote is
883 * "... But that usually will create another problem." This is the first of
886 * Our current solution is to allow the paravirt back end to optionally patch
887 * over the indirect calls to replace them with something more efficient. We
888 * patch the four most commonly called functions: disable interrupts, enable
889 * interrupts, restore interrupts and save interrupts. We usually have 6 or 10
890 * bytes to patch into: the Guest versions of these operations are small enough
891 * that we can fit comfortably.
893 * First we need assembly templates of each of the patchable Guest operations,
894 * and these are in lguest_asm.S. */
896 /*G:060 We construct a table from the assembler templates: */
897 static const struct lguest_insns
899 const char *start, *end;
901 [PARAVIRT_PATCH(pv_irq_ops.irq_disable)] = { lgstart_cli, lgend_cli },
902 [PARAVIRT_PATCH(pv_irq_ops.irq_enable)] = { lgstart_sti, lgend_sti },
903 [PARAVIRT_PATCH(pv_irq_ops.restore_fl)] = { lgstart_popf, lgend_popf },
904 [PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf },
907 /* Now our patch routine is fairly simple (based on the native one in
908 * paravirt.c). If we have a replacement, we copy it in and return how much of
909 * the available space we used. */
910 static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf,
911 unsigned long addr, unsigned len)
913 unsigned int insn_len;
915 /* Don't do anything special if we don't have a replacement */
916 if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start)
917 return paravirt_patch_default(type, clobber, ibuf, addr, len);
919 insn_len = lguest_insns[type].end - lguest_insns[type].start;
921 /* Similarly if we can't fit replacement (shouldn't happen, but let's
924 return paravirt_patch_default(type, clobber, ibuf, addr, len);
926 /* Copy in our instructions. */
927 memcpy(ibuf, lguest_insns[type].start, insn_len);
931 /*G:030 Once we get to lguest_init(), we know we're a Guest. The various
932 * pv_ops structures in the kernel provide points for (almost) every routine we
933 * have to override to avoid privileged instructions. */
934 __init void lguest_init(void)
936 /* We're under lguest, paravirt is enabled, and we're running at
937 * privilege level 1, not 0 as normal. */
938 pv_info.name = "lguest";
939 pv_info.paravirt_enabled = 1;
940 pv_info.kernel_rpl = 1;
942 /* We set up all the lguest overrides for sensitive operations. These
943 * are detailed with the operations themselves. */
945 /* interrupt-related operations */
946 pv_irq_ops.init_IRQ = lguest_init_IRQ;
947 pv_irq_ops.save_fl = save_fl;
948 pv_irq_ops.restore_fl = restore_fl;
949 pv_irq_ops.irq_disable = irq_disable;
950 pv_irq_ops.irq_enable = irq_enable;
951 pv_irq_ops.safe_halt = lguest_safe_halt;
953 /* init-time operations */
954 pv_init_ops.memory_setup = lguest_memory_setup;
955 pv_init_ops.patch = lguest_patch;
957 /* Intercepts of various cpu instructions */
958 pv_cpu_ops.load_gdt = lguest_load_gdt;
959 pv_cpu_ops.cpuid = lguest_cpuid;
960 pv_cpu_ops.load_idt = lguest_load_idt;
961 pv_cpu_ops.iret = lguest_iret;
962 pv_cpu_ops.load_sp0 = lguest_load_sp0;
963 pv_cpu_ops.load_tr_desc = lguest_load_tr_desc;
964 pv_cpu_ops.set_ldt = lguest_set_ldt;
965 pv_cpu_ops.load_tls = lguest_load_tls;
966 pv_cpu_ops.set_debugreg = lguest_set_debugreg;
967 pv_cpu_ops.clts = lguest_clts;
968 pv_cpu_ops.read_cr0 = lguest_read_cr0;
969 pv_cpu_ops.write_cr0 = lguest_write_cr0;
970 pv_cpu_ops.read_cr4 = lguest_read_cr4;
971 pv_cpu_ops.write_cr4 = lguest_write_cr4;
972 pv_cpu_ops.write_gdt_entry = lguest_write_gdt_entry;
973 pv_cpu_ops.write_idt_entry = lguest_write_idt_entry;
974 pv_cpu_ops.wbinvd = lguest_wbinvd;
975 pv_cpu_ops.lazy_mode.enter = paravirt_enter_lazy_cpu;
976 pv_cpu_ops.lazy_mode.leave = lguest_leave_lazy_mode;
978 /* pagetable management */
979 pv_mmu_ops.write_cr3 = lguest_write_cr3;
980 pv_mmu_ops.flush_tlb_user = lguest_flush_tlb_user;
981 pv_mmu_ops.flush_tlb_single = lguest_flush_tlb_single;
982 pv_mmu_ops.flush_tlb_kernel = lguest_flush_tlb_kernel;
983 pv_mmu_ops.set_pte = lguest_set_pte;
984 pv_mmu_ops.set_pte_at = lguest_set_pte_at;
985 pv_mmu_ops.set_pmd = lguest_set_pmd;
986 pv_mmu_ops.read_cr2 = lguest_read_cr2;
987 pv_mmu_ops.read_cr3 = lguest_read_cr3;
988 pv_mmu_ops.lazy_mode.enter = paravirt_enter_lazy_mmu;
989 pv_mmu_ops.lazy_mode.leave = lguest_leave_lazy_mode;
991 #ifdef CONFIG_X86_LOCAL_APIC
992 /* apic read/write intercepts */
993 pv_apic_ops.apic_write = lguest_apic_write;
994 pv_apic_ops.apic_read = lguest_apic_read;
997 /* time operations */
998 pv_time_ops.get_wallclock = lguest_get_wallclock;
999 pv_time_ops.time_init = lguest_time_init;
1000 pv_time_ops.get_tsc_khz = lguest_tsc_khz;
1002 /* Now is a good time to look at the implementations of these functions
1003 * before returning to the rest of lguest_init(). */
1005 /*G:070 Now we've seen all the paravirt_ops, we return to
1006 * lguest_init() where the rest of the fairly chaotic boot setup
1009 /* The native boot code sets up initial page tables immediately after
1010 * the kernel itself, and sets init_pg_tables_end so they're not
1011 * clobbered. The Launcher places our initial pagetables somewhere at
1012 * the top of our physical memory, so we don't need extra space: set
1013 * init_pg_tables_end to the end of the kernel. */
1014 init_pg_tables_start = __pa(pg0);
1015 init_pg_tables_end = __pa(pg0);
1017 /* Load the %fs segment register (the per-cpu segment register) with
1018 * the normal data segment to get through booting. */
1019 asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS) : "memory");
1021 /* The Host<->Guest Switcher lives at the top of our address space, and
1022 * the Host told us how big it is when we made LGUEST_INIT hypercall:
1023 * it put the answer in lguest_data.reserve_mem */
1024 reserve_top_address(lguest_data.reserve_mem);
1026 /* If we don't initialize the lock dependency checker now, it crashes
1027 * paravirt_disable_iospace. */
1030 /* The IDE code spends about 3 seconds probing for disks: if we reserve
1031 * all the I/O ports up front it can't get them and so doesn't probe.
1032 * Other device drivers are similar (but less severe). This cuts the
1033 * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. */
1034 paravirt_disable_iospace();
1036 /* This is messy CPU setup stuff which the native boot code does before
1037 * start_kernel, so we have to do, too: */
1038 cpu_detect(&new_cpu_data);
1039 /* head.S usually sets up the first capability word, so do it here. */
1040 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1042 /* Math is always hard! */
1043 new_cpu_data.hard_math = 1;
1045 /* We don't have features. We have puppies! Puppies! */
1046 #ifdef CONFIG_X86_MCE
1054 /* We set the perferred console to "hvc". This is the "hypervisor
1055 * virtual console" driver written by the PowerPC people, which we also
1056 * adapted for lguest's use. */
1057 add_preferred_console("hvc", 0, NULL);
1059 /* Register our very early console. */
1060 virtio_cons_early_init(early_put_chars);
1062 /* Last of all, we set the power management poweroff hook to point to
1063 * the Guest routine to power off, and the reboot hook to our restart
1065 pm_power_off = lguest_power_off;
1066 machine_ops.restart = lguest_restart;
1068 /* Now we're set up, call i386_start_kernel() in head32.c and we proceed
1069 * to boot as normal. It never returns. */
1070 i386_start_kernel();
1073 * This marks the end of stage II of our journey, The Guest.
1075 * It is now time for us to explore the layer of virtual drivers and complete
1076 * our understanding of the Guest in "make Drivers".