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. When
14 * you set CONFIG_LGUEST to 'y' or 'm', this automatically sets
15 * CONFIG_LGUEST_GUEST=y, which compiles this file into the kernel so it knows
16 * how to be a Guest. This means that you can use the same kernel you boot
17 * normally (ie. as a Host) as a Guest.
19 * These Guests know that they cannot do privileged operations, such as disable
20 * interrupts, and that they have to ask the Host to do such things explicitly.
21 * This file consists of all the replacements for such low-level native
22 * hardware operations: these special Guest versions call the Host.
24 * So how does the kernel know it's a Guest? The Guest starts at a special
25 * entry point marked with a magic string, which sets up a few things then
26 * calls here. We replace the native functions various "paravirt" structures
27 * with our Guest versions, then boot like normal. :*/
30 * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation.
32 * This program is free software; you can redistribute it and/or modify
33 * it under the terms of the GNU General Public License as published by
34 * the Free Software Foundation; either version 2 of the License, or
35 * (at your option) any later version.
37 * This program is distributed in the hope that it will be useful, but
38 * WITHOUT ANY WARRANTY; without even the implied warranty of
39 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
40 * NON INFRINGEMENT. See the GNU General Public License for more
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
47 #include <linux/kernel.h>
48 #include <linux/start_kernel.h>
49 #include <linux/string.h>
50 #include <linux/console.h>
51 #include <linux/screen_info.h>
52 #include <linux/irq.h>
53 #include <linux/interrupt.h>
54 #include <linux/clocksource.h>
55 #include <linux/clockchips.h>
56 #include <linux/lguest.h>
57 #include <linux/lguest_launcher.h>
58 #include <linux/lguest_bus.h>
59 #include <asm/paravirt.h>
60 #include <asm/param.h>
62 #include <asm/pgtable.h>
64 #include <asm/setup.h>
69 /*G:010 Welcome to the Guest!
71 * The Guest in our tale is a simple creature: identical to the Host but
72 * behaving in simplified but equivalent ways. In particular, the Guest is the
73 * same kernel as the Host (or at least, built from the same source code). :*/
75 /* Declarations for definitions in lguest_guest.S */
76 extern char lguest_noirq_start
[], lguest_noirq_end
[];
77 extern const char lgstart_cli
[], lgend_cli
[];
78 extern const char lgstart_sti
[], lgend_sti
[];
79 extern const char lgstart_popf
[], lgend_popf
[];
80 extern const char lgstart_pushf
[], lgend_pushf
[];
81 extern const char lgstart_iret
[], lgend_iret
[];
82 extern void lguest_iret(void);
84 struct lguest_data lguest_data
= {
85 .hcall_status
= { [0 ... LHCALL_RING_SIZE
-1] = 0xFF },
86 .noirq_start
= (u32
)lguest_noirq_start
,
87 .noirq_end
= (u32
)lguest_noirq_end
,
88 .blocked_interrupts
= { 1 }, /* Block timer interrupts */
90 static cycle_t clock_base
;
92 /*G:035 Notice the lazy_hcall() above, rather than hcall(). This is our first
93 * real optimization trick!
95 * When lazy_mode is set, it means we're allowed to defer all hypercalls and do
96 * them as a batch when lazy_mode is eventually turned off. Because hypercalls
97 * are reasonably expensive, batching them up makes sense. For example, a
98 * large mmap might update dozens of page table entries: that code calls
99 * paravirt_enter_lazy_mmu(), does the dozen updates, then calls
100 * lguest_leave_lazy_mode().
102 * So, when we're in lazy mode, we call async_hypercall() to store the call for
103 * future processing. When lazy mode is turned off we issue a hypercall to
104 * flush the stored calls.
106 static void lguest_leave_lazy_mode(void)
108 paravirt_leave_lazy(paravirt_get_lazy_mode());
109 hcall(LHCALL_FLUSH_ASYNC
, 0, 0, 0);
112 static void lazy_hcall(unsigned long call
,
117 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE
)
118 hcall(call
, arg1
, arg2
, arg3
);
120 async_hcall(call
, arg1
, arg2
, arg3
);
123 /* async_hcall() is pretty simple: I'm quite proud of it really. We have a
124 * ring buffer of stored hypercalls which the Host will run though next time we
125 * do a normal hypercall. Each entry in the ring has 4 slots for the hypercall
126 * arguments, and a "hcall_status" word which is 0 if the call is ready to go,
127 * and 255 once the Host has finished with it.
129 * If we come around to a slot which hasn't been finished, then the table is
130 * full and we just make the hypercall directly. This has the nice side
131 * effect of causing the Host to run all the stored calls in the ring buffer
132 * which empties it for next time! */
133 void async_hcall(unsigned long call
,
134 unsigned long arg1
, unsigned long arg2
, unsigned long arg3
)
136 /* Note: This code assumes we're uniprocessor. */
137 static unsigned int next_call
;
140 /* Disable interrupts if not already disabled: we don't want an
141 * interrupt handler making a hypercall while we're already doing
143 local_irq_save(flags
);
144 if (lguest_data
.hcall_status
[next_call
] != 0xFF) {
145 /* Table full, so do normal hcall which will flush table. */
146 hcall(call
, arg1
, arg2
, arg3
);
148 lguest_data
.hcalls
[next_call
].eax
= call
;
149 lguest_data
.hcalls
[next_call
].edx
= arg1
;
150 lguest_data
.hcalls
[next_call
].ebx
= arg2
;
151 lguest_data
.hcalls
[next_call
].ecx
= arg3
;
152 /* Arguments must all be written before we mark it to go */
154 lguest_data
.hcall_status
[next_call
] = 0;
155 if (++next_call
== LHCALL_RING_SIZE
)
158 local_irq_restore(flags
);
162 /* Wrappers for the SEND_DMA and BIND_DMA hypercalls. This is mainly because
163 * Jeff Garzik complained that __pa() should never appear in drivers, and this
164 * helps remove most of them. But also, it wraps some ugliness. */
165 void lguest_send_dma(unsigned long key
, struct lguest_dma
*dma
)
167 /* The hcall might not write this if something goes wrong */
169 hcall(LHCALL_SEND_DMA
, key
, __pa(dma
), 0);
172 int lguest_bind_dma(unsigned long key
, struct lguest_dma
*dmas
,
173 unsigned int num
, u8 irq
)
175 /* This is the only hypercall which actually wants 5 arguments, and we
176 * only support 4. Fortunately the interrupt number is always less
177 * than 256, so we can pack it with the number of dmas in the final
179 if (!hcall(LHCALL_BIND_DMA
, key
, __pa(dmas
), (num
<< 8) | irq
))
184 /* Unbinding is the same hypercall as binding, but with 0 num & irq. */
185 void lguest_unbind_dma(unsigned long key
, struct lguest_dma
*dmas
)
187 hcall(LHCALL_BIND_DMA
, key
, __pa(dmas
), 0);
190 /* For guests, device memory can be used as normal memory, so we cast away the
191 * __iomem to quieten sparse. */
192 void *lguest_map(unsigned long phys_addr
, unsigned long pages
)
194 return (__force
void *)ioremap(phys_addr
, PAGE_SIZE
*pages
);
197 void lguest_unmap(void *addr
)
199 iounmap((__force
void __iomem
*)addr
);
203 * Here are our first native-instruction replacements: four functions for
206 * The simplest way of implementing these would be to have "turn interrupts
207 * off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow:
208 * these are by far the most commonly called functions of those we override.
210 * So instead we keep an "irq_enabled" field inside our "struct lguest_data",
211 * which the Guest can update with a single instruction. The Host knows to
212 * check there when it wants to deliver an interrupt.
215 /* save_flags() is expected to return the processor state (ie. "eflags"). The
216 * eflags word contains all kind of stuff, but in practice Linux only cares
217 * about the interrupt flag. Our "save_flags()" just returns that. */
218 static unsigned long save_fl(void)
220 return lguest_data
.irq_enabled
;
223 /* "restore_flags" just sets the flags back to the value given. */
224 static void restore_fl(unsigned long flags
)
226 lguest_data
.irq_enabled
= flags
;
229 /* Interrupts go off... */
230 static void irq_disable(void)
232 lguest_data
.irq_enabled
= 0;
235 /* Interrupts go on... */
236 static void irq_enable(void)
238 lguest_data
.irq_enabled
= X86_EFLAGS_IF
;
241 /*M:003 Note that we don't check for outstanding interrupts when we re-enable
242 * them (or when we unmask an interrupt). This seems to work for the moment,
243 * since interrupts are rare and we'll just get the interrupt on the next timer
244 * tick, but when we turn on CONFIG_NO_HZ, we should revisit this. One way
245 * would be to put the "irq_enabled" field in a page by itself, and have the
246 * Host write-protect it when an interrupt comes in when irqs are disabled.
247 * There will then be a page fault as soon as interrupts are re-enabled. :*/
250 * The Interrupt Descriptor Table (IDT).
252 * The IDT tells the processor what to do when an interrupt comes in. Each
253 * entry in the table is a 64-bit descriptor: this holds the privilege level,
254 * address of the handler, and... well, who cares? The Guest just asks the
255 * Host to make the change anyway, because the Host controls the real IDT.
257 static void lguest_write_idt_entry(struct desc_struct
*dt
,
258 int entrynum
, u32 low
, u32 high
)
260 /* Keep the local copy up to date. */
261 write_dt_entry(dt
, entrynum
, low
, high
);
262 /* Tell Host about this new entry. */
263 hcall(LHCALL_LOAD_IDT_ENTRY
, entrynum
, low
, high
);
266 /* Changing to a different IDT is very rare: we keep the IDT up-to-date every
267 * time it is written, so we can simply loop through all entries and tell the
268 * Host about them. */
269 static void lguest_load_idt(const struct Xgt_desc_struct
*desc
)
272 struct desc_struct
*idt
= (void *)desc
->address
;
274 for (i
= 0; i
< (desc
->size
+1)/8; i
++)
275 hcall(LHCALL_LOAD_IDT_ENTRY
, i
, idt
[i
].a
, idt
[i
].b
);
279 * The Global Descriptor Table.
281 * The Intel architecture defines another table, called the Global Descriptor
282 * Table (GDT). You tell the CPU where it is (and its size) using the "lgdt"
283 * instruction, and then several other instructions refer to entries in the
284 * table. There are three entries which the Switcher needs, so the Host simply
285 * controls the entire thing and the Guest asks it to make changes using the
286 * LOAD_GDT hypercall.
288 * This is the opposite of the IDT code where we have a LOAD_IDT_ENTRY
289 * hypercall and use that repeatedly to load a new IDT. I don't think it
290 * really matters, but wouldn't it be nice if they were the same?
292 static void lguest_load_gdt(const struct Xgt_desc_struct
*desc
)
294 BUG_ON((desc
->size
+1)/8 != GDT_ENTRIES
);
295 hcall(LHCALL_LOAD_GDT
, __pa(desc
->address
), GDT_ENTRIES
, 0);
298 /* For a single GDT entry which changes, we do the lazy thing: alter our GDT,
299 * then tell the Host to reload the entire thing. This operation is so rare
300 * that this naive implementation is reasonable. */
301 static void lguest_write_gdt_entry(struct desc_struct
*dt
,
302 int entrynum
, u32 low
, u32 high
)
304 write_dt_entry(dt
, entrynum
, low
, high
);
305 hcall(LHCALL_LOAD_GDT
, __pa(dt
), GDT_ENTRIES
, 0);
308 /* OK, I lied. There are three "thread local storage" GDT entries which change
309 * on every context switch (these three entries are how glibc implements
310 * __thread variables). So we have a hypercall specifically for this case. */
311 static void lguest_load_tls(struct thread_struct
*t
, unsigned int cpu
)
313 /* There's one problem which normal hardware doesn't have: the Host
314 * can't handle us removing entries we're currently using. So we clear
315 * the GS register here: if it's needed it'll be reloaded anyway. */
317 lazy_hcall(LHCALL_LOAD_TLS
, __pa(&t
->tls_array
), cpu
, 0);
320 /*G:038 That's enough excitement for now, back to ploughing through each of
321 * the different pv_ops structures (we're about 1/3 of the way through).
323 * This is the Local Descriptor Table, another weird Intel thingy. Linux only
324 * uses this for some strange applications like Wine. We don't do anything
325 * here, so they'll get an informative and friendly Segmentation Fault. */
326 static void lguest_set_ldt(const void *addr
, unsigned entries
)
330 /* This loads a GDT entry into the "Task Register": that entry points to a
331 * structure called the Task State Segment. Some comments scattered though the
332 * kernel code indicate that this used for task switching in ages past, along
333 * with blood sacrifice and astrology.
335 * Now there's nothing interesting in here that we don't get told elsewhere.
336 * But the native version uses the "ltr" instruction, which makes the Host
337 * complain to the Guest about a Segmentation Fault and it'll oops. So we
338 * override the native version with a do-nothing version. */
339 static void lguest_load_tr_desc(void)
343 /* The "cpuid" instruction is a way of querying both the CPU identity
344 * (manufacturer, model, etc) and its features. It was introduced before the
345 * Pentium in 1993 and keeps getting extended by both Intel and AMD. As you
346 * might imagine, after a decade and a half this treatment, it is now a giant
347 * ball of hair. Its entry in the current Intel manual runs to 28 pages.
349 * This instruction even it has its own Wikipedia entry. The Wikipedia entry
350 * has been translated into 4 languages. I am not making this up!
352 * We could get funky here and identify ourselves as "GenuineLguest", but
353 * instead we just use the real "cpuid" instruction. Then I pretty much turned
354 * off feature bits until the Guest booted. (Don't say that: you'll damage
355 * lguest sales!) Shut up, inner voice! (Hey, just pointing out that this is
356 * hardly future proof.) Noone's listening! They don't like you anyway,
357 * parenthetic weirdo!
359 * Replacing the cpuid so we can turn features off is great for the kernel, but
360 * anyone (including userspace) can just use the raw "cpuid" instruction and
361 * the Host won't even notice since it isn't privileged. So we try not to get
362 * too worked up about it. */
363 static void lguest_cpuid(unsigned int *eax
, unsigned int *ebx
,
364 unsigned int *ecx
, unsigned int *edx
)
368 native_cpuid(eax
, ebx
, ecx
, edx
);
370 case 1: /* Basic feature request. */
371 /* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */
373 /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, FPU. */
375 /* The Host can do a nice optimization if it knows that the
376 * kernel mappings (addresses above 0xC0000000 or whatever
377 * PAGE_OFFSET is set to) haven't changed. But Linux calls
378 * flush_tlb_user() for both user and kernel mappings unless
379 * the Page Global Enable (PGE) feature bit is set. */
383 /* Futureproof this a little: if they ask how much extended
384 * processor information there is, limit it to known fields. */
385 if (*eax
> 0x80000008)
391 /* Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4.
392 * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother
393 * it. The Host needs to know when the Guest wants to change them, so we have
394 * a whole series of functions like read_cr0() and write_cr0().
396 * We start with CR0. CR0 allows you to turn on and off all kinds of basic
397 * features, but Linux only really cares about one: the horrifically-named Task
398 * Switched (TS) bit at bit 3 (ie. 8)
400 * What does the TS bit do? Well, it causes the CPU to trap (interrupt 7) if
401 * the floating point unit is used. Which allows us to restore FPU state
402 * lazily after a task switch, and Linux uses that gratefully, but wouldn't a
403 * name like "FPUTRAP bit" be a little less cryptic?
405 * We store cr0 (and cr3) locally, because the Host never changes it. The
406 * Guest sometimes wants to read it and we'd prefer not to bother the Host
408 static unsigned long current_cr0
, current_cr3
;
409 static void lguest_write_cr0(unsigned long val
)
412 lazy_hcall(LHCALL_TS
, val
& 8, 0, 0);
416 static unsigned long lguest_read_cr0(void)
421 /* Intel provided a special instruction to clear the TS bit for people too cool
422 * to use write_cr0() to do it. This "clts" instruction is faster, because all
423 * the vowels have been optimized out. */
424 static void lguest_clts(void)
426 lazy_hcall(LHCALL_TS
, 0, 0, 0);
430 /* CR2 is the virtual address of the last page fault, which the Guest only ever
431 * reads. The Host kindly writes this into our "struct lguest_data", so we
432 * just read it out of there. */
433 static unsigned long lguest_read_cr2(void)
435 return lguest_data
.cr2
;
438 /* CR3 is the current toplevel pagetable page: the principle is the same as
439 * cr0. Keep a local copy, and tell the Host when it changes. */
440 static void lguest_write_cr3(unsigned long cr3
)
442 lazy_hcall(LHCALL_NEW_PGTABLE
, cr3
, 0, 0);
446 static unsigned long lguest_read_cr3(void)
451 /* CR4 is used to enable and disable PGE, but we don't care. */
452 static unsigned long lguest_read_cr4(void)
457 static void lguest_write_cr4(unsigned long val
)
462 * Page Table Handling.
464 * Now would be a good time to take a rest and grab a coffee or similarly
465 * relaxing stimulant. The easy parts are behind us, and the trek gradually
466 * winds uphill from here.
468 * Quick refresher: memory is divided into "pages" of 4096 bytes each. The CPU
469 * maps virtual addresses to physical addresses using "page tables". We could
470 * use one huge index of 1 million entries: each address is 4 bytes, so that's
471 * 1024 pages just to hold the page tables. But since most virtual addresses
472 * are unused, we use a two level index which saves space. The CR3 register
473 * contains the physical address of the top level "page directory" page, which
474 * contains physical addresses of up to 1024 second-level pages. Each of these
475 * second level pages contains up to 1024 physical addresses of actual pages,
476 * or Page Table Entries (PTEs).
478 * Here's a diagram, where arrows indicate physical addresses:
480 * CR3 ---> +---------+
481 * | --------->+---------+
483 * Top-level | | PADDR2 |
490 * So to convert a virtual address to a physical address, we look up the top
491 * level, which points us to the second level, which gives us the physical
492 * address of that page. If the top level entry was not present, or the second
493 * level entry was not present, then the virtual address is invalid (we
494 * say "the page was not mapped").
496 * Put another way, a 32-bit virtual address is divided up like so:
498 * 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
499 * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>|
500 * Index into top Index into second Offset within page
501 * page directory page pagetable page
503 * The kernel spends a lot of time changing both the top-level page directory
504 * and lower-level pagetable pages. The Guest doesn't know physical addresses,
505 * so while it maintains these page tables exactly like normal, it also needs
506 * to keep the Host informed whenever it makes a change: the Host will create
507 * the real page tables based on the Guests'.
510 /* The Guest calls this to set a second-level entry (pte), ie. to map a page
511 * into a process' address space. We set the entry then tell the Host the
512 * toplevel and address this corresponds to. The Guest uses one pagetable per
513 * process, so we need to tell the Host which one we're changing (mm->pgd). */
514 static void lguest_set_pte_at(struct mm_struct
*mm
, unsigned long addr
,
515 pte_t
*ptep
, pte_t pteval
)
518 lazy_hcall(LHCALL_SET_PTE
, __pa(mm
->pgd
), addr
, pteval
.pte_low
);
521 /* The Guest calls this to set a top-level entry. Again, we set the entry then
522 * tell the Host which top-level page we changed, and the index of the entry we
524 static void lguest_set_pmd(pmd_t
*pmdp
, pmd_t pmdval
)
527 lazy_hcall(LHCALL_SET_PMD
, __pa(pmdp
)&PAGE_MASK
,
528 (__pa(pmdp
)&(PAGE_SIZE
-1))/4, 0);
531 /* There are a couple of legacy places where the kernel sets a PTE, but we
532 * don't know the top level any more. This is useless for us, since we don't
533 * know which pagetable is changing or what address, so we just tell the Host
534 * to forget all of them. Fortunately, this is very rare.
536 * ... except in early boot when the kernel sets up the initial pagetables,
537 * which makes booting astonishingly slow. So we don't even tell the Host
538 * anything changed until we've done the first page table switch.
540 static void lguest_set_pte(pte_t
*ptep
, pte_t pteval
)
543 /* Don't bother with hypercall before initial setup. */
545 lazy_hcall(LHCALL_FLUSH_TLB
, 1, 0, 0);
548 /* Unfortunately for Lguest, the pv_mmu_ops for page tables were based on
549 * native page table operations. On native hardware you can set a new page
550 * table entry whenever you want, but if you want to remove one you have to do
551 * a TLB flush (a TLB is a little cache of page table entries kept by the CPU).
553 * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only
554 * called when a valid entry is written, not when it's removed (ie. marked not
555 * present). Instead, this is where we come when the Guest wants to remove a
556 * page table entry: we tell the Host to set that entry to 0 (ie. the present
558 static void lguest_flush_tlb_single(unsigned long addr
)
560 /* Simply set it to zero: if it was not, it will fault back in. */
561 lazy_hcall(LHCALL_SET_PTE
, current_cr3
, addr
, 0);
564 /* This is what happens after the Guest has removed a large number of entries.
565 * This tells the Host that any of the page table entries for userspace might
566 * have changed, ie. virtual addresses below PAGE_OFFSET. */
567 static void lguest_flush_tlb_user(void)
569 lazy_hcall(LHCALL_FLUSH_TLB
, 0, 0, 0);
572 /* This is called when the kernel page tables have changed. That's not very
573 * common (unless the Guest is using highmem, which makes the Guest extremely
574 * slow), so it's worth separating this from the user flushing above. */
575 static void lguest_flush_tlb_kernel(void)
577 lazy_hcall(LHCALL_FLUSH_TLB
, 1, 0, 0);
581 * The Unadvanced Programmable Interrupt Controller.
583 * This is an attempt to implement the simplest possible interrupt controller.
584 * I spent some time looking though routines like set_irq_chip_and_handler,
585 * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and
586 * I *think* this is as simple as it gets.
588 * We can tell the Host what interrupts we want blocked ready for using the
589 * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as
590 * simple as setting a bit. We don't actually "ack" interrupts as such, we
591 * just mask and unmask them. I wonder if we should be cleverer?
593 static void disable_lguest_irq(unsigned int irq
)
595 set_bit(irq
, lguest_data
.blocked_interrupts
);
598 static void enable_lguest_irq(unsigned int irq
)
600 clear_bit(irq
, lguest_data
.blocked_interrupts
);
603 /* This structure describes the lguest IRQ controller. */
604 static struct irq_chip lguest_irq_controller
= {
606 .mask
= disable_lguest_irq
,
607 .mask_ack
= disable_lguest_irq
,
608 .unmask
= enable_lguest_irq
,
611 /* This sets up the Interrupt Descriptor Table (IDT) entry for each hardware
612 * interrupt (except 128, which is used for system calls), and then tells the
613 * Linux infrastructure that each interrupt is controlled by our level-based
614 * lguest interrupt controller. */
615 static void __init
lguest_init_IRQ(void)
619 for (i
= 0; i
< LGUEST_IRQS
; i
++) {
620 int vector
= FIRST_EXTERNAL_VECTOR
+ i
;
621 if (vector
!= SYSCALL_VECTOR
) {
622 set_intr_gate(vector
, interrupt
[i
]);
623 set_irq_chip_and_handler(i
, &lguest_irq_controller
,
627 /* This call is required to set up for 4k stacks, where we have
628 * separate stacks for hard and soft interrupts. */
629 irq_ctx_init(smp_processor_id());
635 * It would be far better for everyone if the Guest had its own clock, but
636 * until then the Host gives us the time on every interrupt.
638 static unsigned long lguest_get_wallclock(void)
640 return lguest_data
.time
.tv_sec
;
643 static cycle_t
lguest_clock_read(void)
645 unsigned long sec
, nsec
;
647 /* If the Host tells the TSC speed, we can trust that. */
648 if (lguest_data
.tsc_khz
)
649 return native_read_tsc();
651 /* If we can't use the TSC, we read the time value written by the Host.
652 * Since it's in two parts (seconds and nanoseconds), we risk reading
653 * it just as it's changing from 99 & 0.999999999 to 100 and 0, and
654 * getting 99 and 0. As Linux tends to come apart under the stress of
655 * time travel, we must be careful: */
657 /* First we read the seconds part. */
658 sec
= lguest_data
.time
.tv_sec
;
659 /* This read memory barrier tells the compiler and the CPU that
660 * this can't be reordered: we have to complete the above
661 * before going on. */
663 /* Now we read the nanoseconds part. */
664 nsec
= lguest_data
.time
.tv_nsec
;
665 /* Make sure we've done that. */
667 /* Now if the seconds part has changed, try again. */
668 } while (unlikely(lguest_data
.time
.tv_sec
!= sec
));
670 /* Our non-TSC clock is in real nanoseconds. */
671 return sec
*1000000000ULL + nsec
;
674 /* This is what we tell the kernel is our clocksource. */
675 static struct clocksource lguest_clock
= {
678 .read
= lguest_clock_read
,
679 .mask
= CLOCKSOURCE_MASK(64),
682 .flags
= CLOCK_SOURCE_IS_CONTINUOUS
,
685 /* The "scheduler clock" is just our real clock, adjusted to start at zero */
686 static unsigned long long lguest_sched_clock(void)
688 return cyc2ns(&lguest_clock
, lguest_clock_read() - clock_base
);
691 /* We also need a "struct clock_event_device": Linux asks us to set it to go
692 * off some time in the future. Actually, James Morris figured all this out, I
693 * just applied the patch. */
694 static int lguest_clockevent_set_next_event(unsigned long delta
,
695 struct clock_event_device
*evt
)
697 if (delta
< LG_CLOCK_MIN_DELTA
) {
698 if (printk_ratelimit())
699 printk(KERN_DEBUG
"%s: small delta %lu ns\n",
700 __FUNCTION__
, delta
);
703 hcall(LHCALL_SET_CLOCKEVENT
, delta
, 0, 0);
707 static void lguest_clockevent_set_mode(enum clock_event_mode mode
,
708 struct clock_event_device
*evt
)
711 case CLOCK_EVT_MODE_UNUSED
:
712 case CLOCK_EVT_MODE_SHUTDOWN
:
713 /* A 0 argument shuts the clock down. */
714 hcall(LHCALL_SET_CLOCKEVENT
, 0, 0, 0);
716 case CLOCK_EVT_MODE_ONESHOT
:
717 /* This is what we expect. */
719 case CLOCK_EVT_MODE_PERIODIC
:
721 case CLOCK_EVT_MODE_RESUME
:
726 /* This describes our primitive timer chip. */
727 static struct clock_event_device lguest_clockevent
= {
729 .features
= CLOCK_EVT_FEAT_ONESHOT
,
730 .set_next_event
= lguest_clockevent_set_next_event
,
731 .set_mode
= lguest_clockevent_set_mode
,
735 .min_delta_ns
= LG_CLOCK_MIN_DELTA
,
736 .max_delta_ns
= LG_CLOCK_MAX_DELTA
,
739 /* This is the Guest timer interrupt handler (hardware interrupt 0). We just
740 * call the clockevent infrastructure and it does whatever needs doing. */
741 static void lguest_time_irq(unsigned int irq
, struct irq_desc
*desc
)
745 /* Don't interrupt us while this is running. */
746 local_irq_save(flags
);
747 lguest_clockevent
.event_handler(&lguest_clockevent
);
748 local_irq_restore(flags
);
751 /* At some point in the boot process, we get asked to set up our timing
752 * infrastructure. The kernel doesn't expect timer interrupts before this, but
753 * we cleverly initialized the "blocked_interrupts" field of "struct
754 * lguest_data" so that timer interrupts were blocked until now. */
755 static void lguest_time_init(void)
757 /* Set up the timer interrupt (0) to go to our simple timer routine */
758 set_irq_handler(0, lguest_time_irq
);
760 /* Our clock structure look like arch/i386/kernel/tsc.c if we can use
761 * the TSC, otherwise it's a dumb nanosecond-resolution clock. Either
762 * way, the "rating" is initialized so high that it's always chosen
763 * over any other clocksource. */
764 if (lguest_data
.tsc_khz
)
765 lguest_clock
.mult
= clocksource_khz2mult(lguest_data
.tsc_khz
,
767 clock_base
= lguest_clock_read();
768 clocksource_register(&lguest_clock
);
770 /* Now we've set up our clock, we can use it as the scheduler clock */
771 pv_time_ops
.sched_clock
= lguest_sched_clock
;
773 /* We can't set cpumask in the initializer: damn C limitations! Set it
774 * here and register our timer device. */
775 lguest_clockevent
.cpumask
= cpumask_of_cpu(0);
776 clockevents_register_device(&lguest_clockevent
);
778 /* Finally, we unblock the timer interrupt. */
779 enable_lguest_irq(0);
783 * Miscellaneous bits and pieces.
785 * Here is an oddball collection of functions which the Guest needs for things
786 * to work. They're pretty simple.
789 /* The Guest needs to tell the host what stack it expects traps to use. For
790 * native hardware, this is part of the Task State Segment mentioned above in
791 * lguest_load_tr_desc(), but to help hypervisors there's this special call.
793 * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data
794 * segment), the privilege level (we're privilege level 1, the Host is 0 and
795 * will not tolerate us trying to use that), the stack pointer, and the number
796 * of pages in the stack. */
797 static void lguest_load_esp0(struct tss_struct
*tss
,
798 struct thread_struct
*thread
)
800 lazy_hcall(LHCALL_SET_STACK
, __KERNEL_DS
|0x1, thread
->esp0
,
801 THREAD_SIZE
/PAGE_SIZE
);
804 /* Let's just say, I wouldn't do debugging under a Guest. */
805 static void lguest_set_debugreg(int regno
, unsigned long value
)
807 /* FIXME: Implement */
810 /* There are times when the kernel wants to make sure that no memory writes are
811 * caught in the cache (that they've all reached real hardware devices). This
812 * doesn't matter for the Guest which has virtual hardware.
814 * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush
815 * (clflush) instruction is available and the kernel uses that. Otherwise, it
816 * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction.
817 * Unlike clflush, wbinvd can only be run at privilege level 0. So we can
818 * ignore clflush, but replace wbinvd.
820 static void lguest_wbinvd(void)
824 /* If the Guest expects to have an Advanced Programmable Interrupt Controller,
825 * we play dumb by ignoring writes and returning 0 for reads. So it's no
826 * longer Programmable nor Controlling anything, and I don't think 8 lines of
827 * code qualifies for Advanced. It will also never interrupt anything. It
828 * does, however, allow us to get through the Linux boot code. */
829 #ifdef CONFIG_X86_LOCAL_APIC
830 static void lguest_apic_write(unsigned long reg
, unsigned long v
)
834 static unsigned long lguest_apic_read(unsigned long reg
)
840 /* STOP! Until an interrupt comes in. */
841 static void lguest_safe_halt(void)
843 hcall(LHCALL_HALT
, 0, 0, 0);
846 /* Perhaps CRASH isn't the best name for this hypercall, but we use it to get a
847 * message out when we're crashing as well as elegant termination like powering
850 * Note that the Host always prefers that the Guest speak in physical addresses
851 * rather than virtual addresses, so we use __pa() here. */
852 static void lguest_power_off(void)
854 hcall(LHCALL_CRASH
, __pa("Power down"), 0, 0);
860 * Don't. But if you did, this is what happens.
862 static int lguest_panic(struct notifier_block
*nb
, unsigned long l
, void *p
)
864 hcall(LHCALL_CRASH
, __pa(p
), 0, 0);
865 /* The hcall won't return, but to keep gcc happy, we're "done". */
869 static struct notifier_block paniced
= {
870 .notifier_call
= lguest_panic
873 /* Setting up memory is fairly easy. */
874 static __init
char *lguest_memory_setup(void)
876 /* We do this here and not earlier because lockcheck barfs if we do it
877 * before start_kernel() */
878 atomic_notifier_chain_register(&panic_notifier_list
, &paniced
);
880 /* The Linux bootloader header contains an "e820" memory map: the
881 * Launcher populated the first entry with our memory limit. */
882 add_memory_region(boot_params
.e820_map
[0].addr
,
883 boot_params
.e820_map
[0].size
,
884 boot_params
.e820_map
[0].type
);
886 /* This string is for the boot messages. */
891 * Patching (Powerfully Placating Performance Pedants)
893 * We have already seen that pv_ops structures let us replace simple
894 * native instructions with calls to the appropriate back end all throughout
895 * the kernel. This allows the same kernel to run as a Guest and as a native
896 * kernel, but it's slow because of all the indirect branches.
898 * Remember that David Wheeler quote about "Any problem in computer science can
899 * be solved with another layer of indirection"? The rest of that quote is
900 * "... But that usually will create another problem." This is the first of
903 * Our current solution is to allow the paravirt back end to optionally patch
904 * over the indirect calls to replace them with something more efficient. We
905 * patch the four most commonly called functions: disable interrupts, enable
906 * interrupts, restore interrupts and save interrupts. We usually have 10
907 * bytes to patch into: the Guest versions of these operations are small enough
908 * that we can fit comfortably.
910 * First we need assembly templates of each of the patchable Guest operations,
911 * and these are in lguest_asm.S. */
913 /*G:060 We construct a table from the assembler templates: */
914 static const struct lguest_insns
916 const char *start
, *end
;
918 [PARAVIRT_PATCH(pv_irq_ops
.irq_disable
)] = { lgstart_cli
, lgend_cli
},
919 [PARAVIRT_PATCH(pv_irq_ops
.irq_enable
)] = { lgstart_sti
, lgend_sti
},
920 [PARAVIRT_PATCH(pv_irq_ops
.restore_fl
)] = { lgstart_popf
, lgend_popf
},
921 [PARAVIRT_PATCH(pv_irq_ops
.save_fl
)] = { lgstart_pushf
, lgend_pushf
},
924 /* Now our patch routine is fairly simple (based on the native one in
925 * paravirt.c). If we have a replacement, we copy it in and return how much of
926 * the available space we used. */
927 static unsigned lguest_patch(u8 type
, u16 clobber
, void *ibuf
,
928 unsigned long addr
, unsigned len
)
930 unsigned int insn_len
;
932 /* Don't do anything special if we don't have a replacement */
933 if (type
>= ARRAY_SIZE(lguest_insns
) || !lguest_insns
[type
].start
)
934 return paravirt_patch_default(type
, clobber
, ibuf
, addr
, len
);
936 insn_len
= lguest_insns
[type
].end
- lguest_insns
[type
].start
;
938 /* Similarly if we can't fit replacement (shouldn't happen, but let's
941 return paravirt_patch_default(type
, clobber
, ibuf
, addr
, len
);
943 /* Copy in our instructions. */
944 memcpy(ibuf
, lguest_insns
[type
].start
, insn_len
);
948 /*G:030 Once we get to lguest_init(), we know we're a Guest. The pv_ops
949 * structures in the kernel provide points for (almost) every routine we have
950 * to override to avoid privileged instructions. */
951 __init
void lguest_init(void *boot
)
953 /* Copy boot parameters first: the Launcher put the physical location
954 * in %esi, and head.S converted that to a virtual address and handed
955 * it to us. We use "__memcpy" because "memcpy" sometimes tries to do
956 * tricky things to go faster, and we're not ready for that. */
957 __memcpy(&boot_params
, boot
, PARAM_SIZE
);
958 /* The boot parameters also tell us where the command-line is: save
960 __memcpy(boot_command_line
, __va(boot_params
.hdr
.cmd_line_ptr
),
963 /* We're under lguest, paravirt is enabled, and we're running at
964 * privilege level 1, not 0 as normal. */
965 pv_info
.name
= "lguest";
966 pv_info
.paravirt_enabled
= 1;
967 pv_info
.kernel_rpl
= 1;
969 /* We set up all the lguest overrides for sensitive operations. These
970 * are detailed with the operations themselves. */
972 /* interrupt-related operations */
973 pv_irq_ops
.init_IRQ
= lguest_init_IRQ
;
974 pv_irq_ops
.save_fl
= save_fl
;
975 pv_irq_ops
.restore_fl
= restore_fl
;
976 pv_irq_ops
.irq_disable
= irq_disable
;
977 pv_irq_ops
.irq_enable
= irq_enable
;
978 pv_irq_ops
.safe_halt
= lguest_safe_halt
;
980 /* init-time operations */
981 pv_init_ops
.memory_setup
= lguest_memory_setup
;
982 pv_init_ops
.patch
= lguest_patch
;
984 /* Intercepts of various cpu instructions */
985 pv_cpu_ops
.load_gdt
= lguest_load_gdt
;
986 pv_cpu_ops
.cpuid
= lguest_cpuid
;
987 pv_cpu_ops
.load_idt
= lguest_load_idt
;
988 pv_cpu_ops
.iret
= lguest_iret
;
989 pv_cpu_ops
.load_esp0
= lguest_load_esp0
;
990 pv_cpu_ops
.load_tr_desc
= lguest_load_tr_desc
;
991 pv_cpu_ops
.set_ldt
= lguest_set_ldt
;
992 pv_cpu_ops
.load_tls
= lguest_load_tls
;
993 pv_cpu_ops
.set_debugreg
= lguest_set_debugreg
;
994 pv_cpu_ops
.clts
= lguest_clts
;
995 pv_cpu_ops
.read_cr0
= lguest_read_cr0
;
996 pv_cpu_ops
.write_cr0
= lguest_write_cr0
;
997 pv_cpu_ops
.read_cr4
= lguest_read_cr4
;
998 pv_cpu_ops
.write_cr4
= lguest_write_cr4
;
999 pv_cpu_ops
.write_gdt_entry
= lguest_write_gdt_entry
;
1000 pv_cpu_ops
.write_idt_entry
= lguest_write_idt_entry
;
1001 pv_cpu_ops
.wbinvd
= lguest_wbinvd
;
1002 pv_cpu_ops
.lazy_mode
.enter
= paravirt_enter_lazy_cpu
;
1003 pv_cpu_ops
.lazy_mode
.leave
= lguest_leave_lazy_mode
;
1005 /* pagetable management */
1006 pv_mmu_ops
.write_cr3
= lguest_write_cr3
;
1007 pv_mmu_ops
.flush_tlb_user
= lguest_flush_tlb_user
;
1008 pv_mmu_ops
.flush_tlb_single
= lguest_flush_tlb_single
;
1009 pv_mmu_ops
.flush_tlb_kernel
= lguest_flush_tlb_kernel
;
1010 pv_mmu_ops
.set_pte
= lguest_set_pte
;
1011 pv_mmu_ops
.set_pte_at
= lguest_set_pte_at
;
1012 pv_mmu_ops
.set_pmd
= lguest_set_pmd
;
1013 pv_mmu_ops
.read_cr2
= lguest_read_cr2
;
1014 pv_mmu_ops
.read_cr3
= lguest_read_cr3
;
1015 pv_mmu_ops
.lazy_mode
.enter
= paravirt_enter_lazy_mmu
;
1016 pv_mmu_ops
.lazy_mode
.leave
= lguest_leave_lazy_mode
;
1018 #ifdef CONFIG_X86_LOCAL_APIC
1019 /* apic read/write intercepts */
1020 pv_apic_ops
.apic_write
= lguest_apic_write
;
1021 pv_apic_ops
.apic_write_atomic
= lguest_apic_write
;
1022 pv_apic_ops
.apic_read
= lguest_apic_read
;
1025 /* time operations */
1026 pv_time_ops
.get_wallclock
= lguest_get_wallclock
;
1027 pv_time_ops
.time_init
= lguest_time_init
;
1029 /* Now is a good time to look at the implementations of these functions
1030 * before returning to the rest of lguest_init(). */
1032 /*G:070 Now we've seen all the paravirt_ops, we return to
1033 * lguest_init() where the rest of the fairly chaotic boot setup
1036 * The Host expects our first hypercall to tell it where our "struct
1037 * lguest_data" is, so we do that first. */
1038 hcall(LHCALL_LGUEST_INIT
, __pa(&lguest_data
), 0, 0);
1040 /* The native boot code sets up initial page tables immediately after
1041 * the kernel itself, and sets init_pg_tables_end so they're not
1042 * clobbered. The Launcher places our initial pagetables somewhere at
1043 * the top of our physical memory, so we don't need extra space: set
1044 * init_pg_tables_end to the end of the kernel. */
1045 init_pg_tables_end
= __pa(pg0
);
1047 /* Load the %fs segment register (the per-cpu segment register) with
1048 * the normal data segment to get through booting. */
1049 asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS
) : "memory");
1051 /* Clear the part of the kernel data which is expected to be zero.
1052 * Normally it will be anyway, but if we're loading from a bzImage with
1053 * CONFIG_RELOCATALE=y, the relocations will be sitting here. */
1054 memset(__bss_start
, 0, __bss_stop
- __bss_start
);
1056 /* The Host uses the top of the Guest's virtual address space for the
1057 * Host<->Guest Switcher, and it tells us how much it needs in
1058 * lguest_data.reserve_mem, set up on the LGUEST_INIT hypercall. */
1059 reserve_top_address(lguest_data
.reserve_mem
);
1061 /* If we don't initialize the lock dependency checker now, it crashes
1062 * paravirt_disable_iospace. */
1065 /* The IDE code spends about 3 seconds probing for disks: if we reserve
1066 * all the I/O ports up front it can't get them and so doesn't probe.
1067 * Other device drivers are similar (but less severe). This cuts the
1068 * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. */
1069 paravirt_disable_iospace();
1071 /* This is messy CPU setup stuff which the native boot code does before
1072 * start_kernel, so we have to do, too: */
1073 cpu_detect(&new_cpu_data
);
1074 /* head.S usually sets up the first capability word, so do it here. */
1075 new_cpu_data
.x86_capability
[0] = cpuid_edx(1);
1077 /* Math is always hard! */
1078 new_cpu_data
.hard_math
= 1;
1080 #ifdef CONFIG_X86_MCE
1088 /* We set the perferred console to "hvc". This is the "hypervisor
1089 * virtual console" driver written by the PowerPC people, which we also
1090 * adapted for lguest's use. */
1091 add_preferred_console("hvc", 0, NULL
);
1093 /* Last of all, we set the power management poweroff hook to point to
1094 * the Guest routine to power off. */
1095 pm_power_off
= lguest_power_off
;
1097 /* Now we're set up, call start_kernel() in init/main.c and we proceed
1098 * to boot as normal. It never returns. */
1102 * This marks the end of stage II of our journey, The Guest.
1104 * It is now time for us to explore the nooks and crannies of the three Guest
1105 * devices and complete our understanding of the Guest in "make Drivers".