blob: e6023b86f31d344f5b3b63f27514526aee7dd653 [file] [log] [blame]
Rusty Russellf938d2c2007-07-26 10:41:02 -07001/*P:010
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."
Rusty Russell07ad1572007-07-19 01:49:22 -07005 *
Rusty Russellf938d2c2007-07-26 10:41:02 -07006 * 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
11 * Launcher.
12 *
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.
18 *
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.
23 *
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
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -070026 * calls here. We replace the native functions various "paravirt" structures
Rusty Russellf938d2c2007-07-26 10:41:02 -070027 * with our Guest versions, then boot like normal. :*/
28
29/*
Rusty Russell07ad1572007-07-19 01:49:22 -070030 * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation.
31 *
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.
36 *
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
41 * details.
42 *
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.
46 */
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>
Rusty Russelld7e28ff2007-07-19 01:49:23 -070054#include <linux/clocksource.h>
55#include <linux/clockchips.h>
Rusty Russell07ad1572007-07-19 01:49:22 -070056#include <linux/lguest.h>
57#include <linux/lguest_launcher.h>
Rusty Russell19f15372007-10-22 11:24:21 +100058#include <linux/virtio_console.h>
Jeff Garzik4cfe6c32007-10-25 14:15:09 +100059#include <linux/pm.h>
Rusty Russell07ad1572007-07-19 01:49:22 -070060#include <asm/paravirt.h>
61#include <asm/param.h>
62#include <asm/page.h>
63#include <asm/pgtable.h>
64#include <asm/desc.h>
65#include <asm/setup.h>
66#include <asm/e820.h>
67#include <asm/mce.h>
68#include <asm/io.h>
Jes Sorensen625efab2007-10-22 11:03:28 +100069#include <asm/i387.h>
Rusty Russell07ad1572007-07-19 01:49:22 -070070
Rusty Russellb2b47c22007-07-26 10:41:02 -070071/*G:010 Welcome to the Guest!
72 *
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). :*/
76
Rusty Russell07ad1572007-07-19 01:49:22 -070077/* Declarations for definitions in lguest_guest.S */
78extern char lguest_noirq_start[], lguest_noirq_end[];
79extern const char lgstart_cli[], lgend_cli[];
80extern const char lgstart_sti[], lgend_sti[];
81extern const char lgstart_popf[], lgend_popf[];
82extern const char lgstart_pushf[], lgend_pushf[];
83extern const char lgstart_iret[], lgend_iret[];
84extern void lguest_iret(void);
85
86struct lguest_data lguest_data = {
87 .hcall_status = { [0 ... LHCALL_RING_SIZE-1] = 0xFF },
88 .noirq_start = (u32)lguest_noirq_start,
89 .noirq_end = (u32)lguest_noirq_end,
Rusty Russell47436aa2007-10-22 11:03:36 +100090 .kernel_address = PAGE_OFFSET,
Rusty Russell07ad1572007-07-19 01:49:22 -070091 .blocked_interrupts = { 1 }, /* Block timer interrupts */
Rusty Russellc18acd72007-10-22 11:03:35 +100092 .syscall_vec = SYSCALL_VECTOR,
Rusty Russell07ad1572007-07-19 01:49:22 -070093};
Rusty Russell9d1ca6f2007-07-20 22:15:01 +100094static cycle_t clock_base;
Rusty Russell07ad1572007-07-19 01:49:22 -070095
Rusty Russellb2b47c22007-07-26 10:41:02 -070096/*G:035 Notice the lazy_hcall() above, rather than hcall(). This is our first
97 * real optimization trick!
98 *
99 * When lazy_mode is set, it means we're allowed to defer all hypercalls and do
100 * them as a batch when lazy_mode is eventually turned off. Because hypercalls
101 * are reasonably expensive, batching them up makes sense. For example, a
Rusty Russelle1e72962007-10-25 15:02:50 +1000102 * large munmap might update dozens of page table entries: that code calls
Jeremy Fitzhardinge8965c1c2007-10-16 11:51:29 -0700103 * paravirt_enter_lazy_mmu(), does the dozen updates, then calls
104 * lguest_leave_lazy_mode().
Rusty Russellb2b47c22007-07-26 10:41:02 -0700105 *
106 * So, when we're in lazy mode, we call async_hypercall() to store the call for
107 * future processing. When lazy mode is turned off we issue a hypercall to
108 * flush the stored calls.
Jeremy Fitzhardinge8965c1c2007-10-16 11:51:29 -0700109 */
110static void lguest_leave_lazy_mode(void)
Rusty Russell07ad1572007-07-19 01:49:22 -0700111{
Jeremy Fitzhardinge8965c1c2007-10-16 11:51:29 -0700112 paravirt_leave_lazy(paravirt_get_lazy_mode());
113 hcall(LHCALL_FLUSH_ASYNC, 0, 0, 0);
Rusty Russell07ad1572007-07-19 01:49:22 -0700114}
115
Rusty Russellb2b47c22007-07-26 10:41:02 -0700116/* async_hcall() is pretty simple: I'm quite proud of it really. We have a
117 * ring buffer of stored hypercalls which the Host will run though next time we
118 * do a normal hypercall. Each entry in the ring has 4 slots for the hypercall
119 * arguments, and a "hcall_status" word which is 0 if the call is ready to go,
120 * and 255 once the Host has finished with it.
121 *
122 * If we come around to a slot which hasn't been finished, then the table is
123 * full and we just make the hypercall directly. This has the nice side
124 * effect of causing the Host to run all the stored calls in the ring buffer
125 * which empties it for next time! */
Adrian Bunk9b56fdb2007-11-02 16:43:10 +0100126static void async_hcall(unsigned long call, unsigned long arg1,
127 unsigned long arg2, unsigned long arg3)
Rusty Russell07ad1572007-07-19 01:49:22 -0700128{
129 /* Note: This code assumes we're uniprocessor. */
130 static unsigned int next_call;
131 unsigned long flags;
132
Rusty Russellb2b47c22007-07-26 10:41:02 -0700133 /* Disable interrupts if not already disabled: we don't want an
134 * interrupt handler making a hypercall while we're already doing
135 * one! */
Rusty Russell07ad1572007-07-19 01:49:22 -0700136 local_irq_save(flags);
137 if (lguest_data.hcall_status[next_call] != 0xFF) {
138 /* Table full, so do normal hcall which will flush table. */
139 hcall(call, arg1, arg2, arg3);
140 } else {
Jes Sorensenb410e7b2007-10-22 11:03:31 +1000141 lguest_data.hcalls[next_call].arg0 = call;
142 lguest_data.hcalls[next_call].arg1 = arg1;
143 lguest_data.hcalls[next_call].arg2 = arg2;
144 lguest_data.hcalls[next_call].arg3 = arg3;
Rusty Russellb2b47c22007-07-26 10:41:02 -0700145 /* Arguments must all be written before we mark it to go */
Rusty Russell07ad1572007-07-19 01:49:22 -0700146 wmb();
147 lguest_data.hcall_status[next_call] = 0;
148 if (++next_call == LHCALL_RING_SIZE)
149 next_call = 0;
150 }
151 local_irq_restore(flags);
152}
Adrian Bunk9b56fdb2007-11-02 16:43:10 +0100153
154static void lazy_hcall(unsigned long call,
155 unsigned long arg1,
156 unsigned long arg2,
157 unsigned long arg3)
158{
159 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
160 hcall(call, arg1, arg2, arg3);
161 else
162 async_hcall(call, arg1, arg2, arg3);
163}
Rusty Russellb2b47c22007-07-26 10:41:02 -0700164/*:*/
Rusty Russell07ad1572007-07-19 01:49:22 -0700165
Rusty Russellb2b47c22007-07-26 10:41:02 -0700166/*G:033
Rusty Russelle1e72962007-10-25 15:02:50 +1000167 * After that diversion we return to our first native-instruction
168 * replacements: four functions for interrupt control.
Rusty Russellb2b47c22007-07-26 10:41:02 -0700169 *
170 * The simplest way of implementing these would be to have "turn interrupts
171 * off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow:
172 * these are by far the most commonly called functions of those we override.
173 *
174 * So instead we keep an "irq_enabled" field inside our "struct lguest_data",
175 * which the Guest can update with a single instruction. The Host knows to
176 * check there when it wants to deliver an interrupt.
177 */
178
179/* save_flags() is expected to return the processor state (ie. "eflags"). The
180 * eflags word contains all kind of stuff, but in practice Linux only cares
181 * about the interrupt flag. Our "save_flags()" just returns that. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700182static unsigned long save_fl(void)
183{
184 return lguest_data.irq_enabled;
185}
186
Rusty Russelle1e72962007-10-25 15:02:50 +1000187/* restore_flags() just sets the flags back to the value given. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700188static void restore_fl(unsigned long flags)
189{
Rusty Russell07ad1572007-07-19 01:49:22 -0700190 lguest_data.irq_enabled = flags;
191}
192
Rusty Russellb2b47c22007-07-26 10:41:02 -0700193/* Interrupts go off... */
Rusty Russell07ad1572007-07-19 01:49:22 -0700194static void irq_disable(void)
195{
196 lguest_data.irq_enabled = 0;
197}
198
Rusty Russellb2b47c22007-07-26 10:41:02 -0700199/* Interrupts go on... */
Rusty Russell07ad1572007-07-19 01:49:22 -0700200static void irq_enable(void)
201{
Rusty Russell07ad1572007-07-19 01:49:22 -0700202 lguest_data.irq_enabled = X86_EFLAGS_IF;
203}
Rusty Russellf56a3842007-07-26 10:41:05 -0700204/*:*/
205/*M:003 Note that we don't check for outstanding interrupts when we re-enable
206 * them (or when we unmask an interrupt). This seems to work for the moment,
207 * since interrupts are rare and we'll just get the interrupt on the next timer
208 * tick, but when we turn on CONFIG_NO_HZ, we should revisit this. One way
209 * would be to put the "irq_enabled" field in a page by itself, and have the
210 * Host write-protect it when an interrupt comes in when irqs are disabled.
211 * There will then be a page fault as soon as interrupts are re-enabled. :*/
Rusty Russell07ad1572007-07-19 01:49:22 -0700212
Rusty Russellb2b47c22007-07-26 10:41:02 -0700213/*G:034
214 * The Interrupt Descriptor Table (IDT).
215 *
216 * The IDT tells the processor what to do when an interrupt comes in. Each
217 * entry in the table is a 64-bit descriptor: this holds the privilege level,
218 * address of the handler, and... well, who cares? The Guest just asks the
219 * Host to make the change anyway, because the Host controls the real IDT.
220 */
Rusty Russell07ad1572007-07-19 01:49:22 -0700221static void lguest_write_idt_entry(struct desc_struct *dt,
222 int entrynum, u32 low, u32 high)
223{
Rusty Russellb2b47c22007-07-26 10:41:02 -0700224 /* Keep the local copy up to date. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700225 write_dt_entry(dt, entrynum, low, high);
Rusty Russellb2b47c22007-07-26 10:41:02 -0700226 /* Tell Host about this new entry. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700227 hcall(LHCALL_LOAD_IDT_ENTRY, entrynum, low, high);
228}
229
Rusty Russellb2b47c22007-07-26 10:41:02 -0700230/* Changing to a different IDT is very rare: we keep the IDT up-to-date every
231 * time it is written, so we can simply loop through all entries and tell the
232 * Host about them. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700233static void lguest_load_idt(const struct Xgt_desc_struct *desc)
234{
235 unsigned int i;
236 struct desc_struct *idt = (void *)desc->address;
237
238 for (i = 0; i < (desc->size+1)/8; i++)
239 hcall(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b);
240}
241
Rusty Russellb2b47c22007-07-26 10:41:02 -0700242/*
243 * The Global Descriptor Table.
244 *
245 * The Intel architecture defines another table, called the Global Descriptor
246 * Table (GDT). You tell the CPU where it is (and its size) using the "lgdt"
247 * instruction, and then several other instructions refer to entries in the
248 * table. There are three entries which the Switcher needs, so the Host simply
249 * controls the entire thing and the Guest asks it to make changes using the
250 * LOAD_GDT hypercall.
251 *
252 * This is the opposite of the IDT code where we have a LOAD_IDT_ENTRY
253 * hypercall and use that repeatedly to load a new IDT. I don't think it
254 * really matters, but wouldn't it be nice if they were the same?
255 */
Rusty Russell07ad1572007-07-19 01:49:22 -0700256static void lguest_load_gdt(const struct Xgt_desc_struct *desc)
257{
258 BUG_ON((desc->size+1)/8 != GDT_ENTRIES);
259 hcall(LHCALL_LOAD_GDT, __pa(desc->address), GDT_ENTRIES, 0);
260}
261
Rusty Russellb2b47c22007-07-26 10:41:02 -0700262/* 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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700265static void lguest_write_gdt_entry(struct desc_struct *dt,
266 int entrynum, u32 low, u32 high)
267{
268 write_dt_entry(dt, entrynum, low, high);
269 hcall(LHCALL_LOAD_GDT, __pa(dt), GDT_ENTRIES, 0);
270}
271
Rusty Russellb2b47c22007-07-26 10:41:02 -0700272/* 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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700275static void lguest_load_tls(struct thread_struct *t, unsigned int cpu)
276{
Rusty Russell0d027c02007-08-09 20:57:13 +1000277 /* 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. */
280 loadsegment(gs, 0);
Rusty Russell07ad1572007-07-19 01:49:22 -0700281 lazy_hcall(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu, 0);
282}
283
Rusty Russellb2b47c22007-07-26 10:41:02 -0700284/*G:038 That's enough excitement for now, back to ploughing through each of
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700285 * the different pv_ops structures (we're about 1/3 of the way through).
Rusty Russellb2b47c22007-07-26 10:41:02 -0700286 *
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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700290static void lguest_set_ldt(const void *addr, unsigned entries)
291{
292}
293
Rusty Russellb2b47c22007-07-26 10:41:02 -0700294/* 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.
298 *
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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700303static void lguest_load_tr_desc(void)
304{
305}
306
Rusty Russellb2b47c22007-07-26 10:41:02 -0700307/* 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 and AMD. As you
310 * might imagine, after a decade and a half this treatment, it is now a giant
311 * ball of hair. Its entry in the current Intel manual runs to 28 pages.
312 *
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!
315 *
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!
322 *
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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700327static void lguest_cpuid(unsigned int *eax, unsigned int *ebx,
328 unsigned int *ecx, unsigned int *edx)
329{
330 int function = *eax;
331
332 native_cpuid(eax, ebx, ecx, edx);
333 switch (function) {
334 case 1: /* Basic feature request. */
335 /* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */
336 *ecx &= 0x00002201;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700337 /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, FPU. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700338 *edx &= 0x07808101;
Rusty Russellb2b47c22007-07-26 10:41:02 -0700339 /* 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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700344 *edx |= 0x00002000;
345 break;
346 case 0x80000000:
347 /* Futureproof this a little: if they ask how much extended
Rusty Russellb2b47c22007-07-26 10:41:02 -0700348 * processor information there is, limit it to known fields. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700349 if (*eax > 0x80000008)
350 *eax = 0x80000008;
351 break;
352 }
353}
354
Rusty Russellb2b47c22007-07-26 10:41:02 -0700355/* 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().
359 *
Rusty Russelle1e72962007-10-25 15:02:50 +1000360 * We start with cr0. cr0 allows you to turn on and off all kinds of basic
Rusty Russellb2b47c22007-07-26 10:41:02 -0700361 * features, but Linux only really cares about one: the horrifically-named Task
362 * Switched (TS) bit at bit 3 (ie. 8)
363 *
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?
368 *
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
371 * unnecessarily. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700372static unsigned long current_cr0, current_cr3;
373static void lguest_write_cr0(unsigned long val)
374{
Rusty Russell25c47bb2007-10-25 14:09:53 +1000375 lazy_hcall(LHCALL_TS, val & X86_CR0_TS, 0, 0);
Rusty Russell07ad1572007-07-19 01:49:22 -0700376 current_cr0 = val;
377}
378
379static unsigned long lguest_read_cr0(void)
380{
381 return current_cr0;
382}
383
Rusty Russellb2b47c22007-07-26 10:41:02 -0700384/* 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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700387static void lguest_clts(void)
388{
389 lazy_hcall(LHCALL_TS, 0, 0, 0);
Rusty Russell25c47bb2007-10-25 14:09:53 +1000390 current_cr0 &= ~X86_CR0_TS;
Rusty Russell07ad1572007-07-19 01:49:22 -0700391}
392
Rusty Russelle1e72962007-10-25 15:02:50 +1000393/* cr2 is the virtual address of the last page fault, which the Guest only ever
Rusty Russellb2b47c22007-07-26 10:41:02 -0700394 * reads. The Host kindly writes this into our "struct lguest_data", so we
395 * just read it out of there. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700396static unsigned long lguest_read_cr2(void)
397{
398 return lguest_data.cr2;
399}
400
Rusty Russelle1e72962007-10-25 15:02:50 +1000401/* cr3 is the current toplevel pagetable page: the principle is the same as
Rusty Russellb2b47c22007-07-26 10:41:02 -0700402 * cr0. Keep a local copy, and tell the Host when it changes. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700403static void lguest_write_cr3(unsigned long cr3)
404{
405 lazy_hcall(LHCALL_NEW_PGTABLE, cr3, 0, 0);
406 current_cr3 = cr3;
407}
408
409static unsigned long lguest_read_cr3(void)
410{
411 return current_cr3;
412}
413
Rusty Russelle1e72962007-10-25 15:02:50 +1000414/* cr4 is used to enable and disable PGE, but we don't care. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700415static unsigned long lguest_read_cr4(void)
416{
417 return 0;
418}
419
420static void lguest_write_cr4(unsigned long val)
421{
422}
423
Rusty Russellb2b47c22007-07-26 10:41:02 -0700424/*
425 * Page Table Handling.
426 *
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.
430 *
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
Rusty Russelle1e72962007-10-25 15:02:50 +1000435 * are unused, we use a two level index which saves space. The cr3 register
Rusty Russellb2b47c22007-07-26 10:41:02 -0700436 * 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).
440 *
441 * Here's a diagram, where arrows indicate physical addresses:
442 *
Rusty Russelle1e72962007-10-25 15:02:50 +1000443 * cr3 ---> +---------+
Rusty Russellb2b47c22007-07-26 10:41:02 -0700444 * | --------->+---------+
445 * | | | PADDR1 |
446 * Top-level | | PADDR2 |
447 * (PMD) page | | |
448 * | | Lower-level |
449 * | | (PTE) page |
450 * | | | |
451 * .... ....
452 *
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").
458 *
459 * Put another way, a 32-bit virtual address is divided up like so:
460 *
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
465 *
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'.
471 */
472
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). */
Rusty Russell07ad1572007-07-19 01:49:22 -0700477static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr,
478 pte_t *ptep, pte_t pteval)
479{
480 *ptep = pteval;
481 lazy_hcall(LHCALL_SET_PTE, __pa(mm->pgd), addr, pteval.pte_low);
482}
483
Rusty Russellb2b47c22007-07-26 10:41:02 -0700484/* 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
486 * changed. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700487static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
488{
489 *pmdp = pmdval;
490 lazy_hcall(LHCALL_SET_PMD, __pa(pmdp)&PAGE_MASK,
491 (__pa(pmdp)&(PAGE_SIZE-1))/4, 0);
492}
493
Rusty Russellb2b47c22007-07-26 10:41:02 -0700494/* 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.
498 *
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
Rusty Russelle1e72962007-10-25 15:02:50 +1000501 * anything changed until we've done the first page table switch. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700502static void lguest_set_pte(pte_t *ptep, pte_t pteval)
503{
504 *ptep = pteval;
505 /* Don't bother with hypercall before initial setup. */
506 if (current_cr3)
507 lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0);
508}
509
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700510/* Unfortunately for Lguest, the pv_mmu_ops for page tables were based on
Rusty Russellb2b47c22007-07-26 10:41:02 -0700511 * 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).
514 *
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
519 * bit is zero). */
Rusty Russell07ad1572007-07-19 01:49:22 -0700520static void lguest_flush_tlb_single(unsigned long addr)
521{
Rusty Russellb2b47c22007-07-26 10:41:02 -0700522 /* Simply set it to zero: if it was not, it will fault back in. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700523 lazy_hcall(LHCALL_SET_PTE, current_cr3, addr, 0);
524}
525
Rusty Russellb2b47c22007-07-26 10:41:02 -0700526/* 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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700529static void lguest_flush_tlb_user(void)
530{
531 lazy_hcall(LHCALL_FLUSH_TLB, 0, 0, 0);
532}
533
Rusty Russellb2b47c22007-07-26 10:41:02 -0700534/* 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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700537static void lguest_flush_tlb_kernel(void)
538{
539 lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0);
540}
541
Rusty Russellb2b47c22007-07-26 10:41:02 -0700542/*
543 * The Unadvanced Programmable Interrupt Controller.
544 *
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.
549 *
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?
554 */
Rusty Russell07ad1572007-07-19 01:49:22 -0700555static void disable_lguest_irq(unsigned int irq)
556{
557 set_bit(irq, lguest_data.blocked_interrupts);
558}
559
560static void enable_lguest_irq(unsigned int irq)
561{
562 clear_bit(irq, lguest_data.blocked_interrupts);
Rusty Russell07ad1572007-07-19 01:49:22 -0700563}
564
Rusty Russellb2b47c22007-07-26 10:41:02 -0700565/* This structure describes the lguest IRQ controller. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700566static struct irq_chip lguest_irq_controller = {
567 .name = "lguest",
568 .mask = disable_lguest_irq,
569 .mask_ack = disable_lguest_irq,
570 .unmask = enable_lguest_irq,
571};
572
Rusty Russellb2b47c22007-07-26 10:41:02 -0700573/* 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. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700577static void __init lguest_init_IRQ(void)
578{
579 unsigned int i;
580
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(i, &lguest_irq_controller,
586 handle_level_irq);
587 }
588 }
Rusty Russellb2b47c22007-07-26 10:41:02 -0700589 /* This call is required to set up for 4k stacks, where we have
590 * separate stacks for hard and soft interrupts. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700591 irq_ctx_init(smp_processor_id());
592}
593
Rusty Russellb2b47c22007-07-26 10:41:02 -0700594/*
595 * Time.
596 *
597 * It would be far better for everyone if the Guest had its own clock, but
Rusty Russell6c8dca52007-07-27 13:42:52 +1000598 * until then the Host gives us the time on every interrupt.
Rusty Russellb2b47c22007-07-26 10:41:02 -0700599 */
Rusty Russell07ad1572007-07-19 01:49:22 -0700600static unsigned long lguest_get_wallclock(void)
601{
Rusty Russell6c8dca52007-07-27 13:42:52 +1000602 return lguest_data.time.tv_sec;
Rusty Russell07ad1572007-07-19 01:49:22 -0700603}
604
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700605static cycle_t lguest_clock_read(void)
Rusty Russell07ad1572007-07-19 01:49:22 -0700606{
Rusty Russell6c8dca52007-07-27 13:42:52 +1000607 unsigned long sec, nsec;
608
609 /* If the Host tells the TSC speed, we can trust that. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700610 if (lguest_data.tsc_khz)
611 return native_read_tsc();
Rusty Russell6c8dca52007-07-27 13:42:52 +1000612
613 /* If we can't use the TSC, we read the time value written by the Host.
614 * Since it's in two parts (seconds and nanoseconds), we risk reading
615 * it just as it's changing from 99 & 0.999999999 to 100 and 0, and
616 * getting 99 and 0. As Linux tends to come apart under the stress of
617 * time travel, we must be careful: */
618 do {
619 /* First we read the seconds part. */
620 sec = lguest_data.time.tv_sec;
621 /* This read memory barrier tells the compiler and the CPU that
622 * this can't be reordered: we have to complete the above
623 * before going on. */
624 rmb();
625 /* Now we read the nanoseconds part. */
626 nsec = lguest_data.time.tv_nsec;
627 /* Make sure we've done that. */
628 rmb();
629 /* Now if the seconds part has changed, try again. */
630 } while (unlikely(lguest_data.time.tv_sec != sec));
631
632 /* Our non-TSC clock is in real nanoseconds. */
633 return sec*1000000000ULL + nsec;
Rusty Russell07ad1572007-07-19 01:49:22 -0700634}
635
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700636/* This is what we tell the kernel is our clocksource. */
637static struct clocksource lguest_clock = {
638 .name = "lguest",
639 .rating = 400,
640 .read = lguest_clock_read,
Rusty Russell6c8dca52007-07-27 13:42:52 +1000641 .mask = CLOCKSOURCE_MASK(64),
Rusty Russell37250092007-08-09 20:52:35 +1000642 .mult = 1 << 22,
643 .shift = 22,
Tony Breeds05aa0262007-10-22 10:56:25 +1000644 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700645};
646
Rusty Russell6c8dca52007-07-27 13:42:52 +1000647/* The "scheduler clock" is just our real clock, adjusted to start at zero */
Rusty Russell9d1ca6f2007-07-20 22:15:01 +1000648static unsigned long long lguest_sched_clock(void)
649{
650 return cyc2ns(&lguest_clock, lguest_clock_read() - clock_base);
651}
652
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700653/* We also need a "struct clock_event_device": Linux asks us to set it to go
654 * off some time in the future. Actually, James Morris figured all this out, I
655 * just applied the patch. */
656static int lguest_clockevent_set_next_event(unsigned long delta,
657 struct clock_event_device *evt)
658{
659 if (delta < LG_CLOCK_MIN_DELTA) {
660 if (printk_ratelimit())
661 printk(KERN_DEBUG "%s: small delta %lu ns\n",
662 __FUNCTION__, delta);
663 return -ETIME;
664 }
665 hcall(LHCALL_SET_CLOCKEVENT, delta, 0, 0);
666 return 0;
667}
668
669static void lguest_clockevent_set_mode(enum clock_event_mode mode,
670 struct clock_event_device *evt)
671{
672 switch (mode) {
673 case CLOCK_EVT_MODE_UNUSED:
674 case CLOCK_EVT_MODE_SHUTDOWN:
675 /* A 0 argument shuts the clock down. */
676 hcall(LHCALL_SET_CLOCKEVENT, 0, 0, 0);
677 break;
678 case CLOCK_EVT_MODE_ONESHOT:
679 /* This is what we expect. */
680 break;
681 case CLOCK_EVT_MODE_PERIODIC:
682 BUG();
Thomas Gleixner18de5bc2007-07-21 04:37:34 -0700683 case CLOCK_EVT_MODE_RESUME:
684 break;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700685 }
686}
687
688/* This describes our primitive timer chip. */
689static struct clock_event_device lguest_clockevent = {
690 .name = "lguest",
691 .features = CLOCK_EVT_FEAT_ONESHOT,
692 .set_next_event = lguest_clockevent_set_next_event,
693 .set_mode = lguest_clockevent_set_mode,
694 .rating = INT_MAX,
695 .mult = 1,
696 .shift = 0,
697 .min_delta_ns = LG_CLOCK_MIN_DELTA,
698 .max_delta_ns = LG_CLOCK_MAX_DELTA,
699};
700
701/* This is the Guest timer interrupt handler (hardware interrupt 0). We just
702 * call the clockevent infrastructure and it does whatever needs doing. */
703static void lguest_time_irq(unsigned int irq, struct irq_desc *desc)
704{
705 unsigned long flags;
706
707 /* Don't interrupt us while this is running. */
708 local_irq_save(flags);
709 lguest_clockevent.event_handler(&lguest_clockevent);
710 local_irq_restore(flags);
711}
712
Rusty Russellb2b47c22007-07-26 10:41:02 -0700713/* At some point in the boot process, we get asked to set up our timing
714 * infrastructure. The kernel doesn't expect timer interrupts before this, but
715 * we cleverly initialized the "blocked_interrupts" field of "struct
716 * lguest_data" so that timer interrupts were blocked until now. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700717static void lguest_time_init(void)
718{
Rusty Russellb2b47c22007-07-26 10:41:02 -0700719 /* Set up the timer interrupt (0) to go to our simple timer routine */
Rusty Russell07ad1572007-07-19 01:49:22 -0700720 set_irq_handler(0, lguest_time_irq);
Rusty Russell07ad1572007-07-19 01:49:22 -0700721
Rusty Russelle1e72962007-10-25 15:02:50 +1000722 /* Our clock structure looks like arch/x86/kernel/tsc_32.c if we can
723 * use the TSC, otherwise it's a dumb nanosecond-resolution clock.
724 * Either way, the "rating" is set so high that it's always chosen over
725 * any other clocksource. */
Tony Breeds05aa0262007-10-22 10:56:25 +1000726 if (lguest_data.tsc_khz)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700727 lguest_clock.mult = clocksource_khz2mult(lguest_data.tsc_khz,
728 lguest_clock.shift);
Rusty Russell9d1ca6f2007-07-20 22:15:01 +1000729 clock_base = lguest_clock_read();
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700730 clocksource_register(&lguest_clock);
731
Rusty Russell6c8dca52007-07-27 13:42:52 +1000732 /* Now we've set up our clock, we can use it as the scheduler clock */
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700733 pv_time_ops.sched_clock = lguest_sched_clock;
Rusty Russell6c8dca52007-07-27 13:42:52 +1000734
Rusty Russellb2b47c22007-07-26 10:41:02 -0700735 /* We can't set cpumask in the initializer: damn C limitations! Set it
736 * here and register our timer device. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700737 lguest_clockevent.cpumask = cpumask_of_cpu(0);
738 clockevents_register_device(&lguest_clockevent);
739
Rusty Russellb2b47c22007-07-26 10:41:02 -0700740 /* Finally, we unblock the timer interrupt. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700741 enable_lguest_irq(0);
Rusty Russell07ad1572007-07-19 01:49:22 -0700742}
743
Rusty Russellb2b47c22007-07-26 10:41:02 -0700744/*
745 * Miscellaneous bits and pieces.
746 *
747 * Here is an oddball collection of functions which the Guest needs for things
748 * to work. They're pretty simple.
749 */
750
Rusty Russelle1e72962007-10-25 15:02:50 +1000751/* The Guest needs to tell the Host what stack it expects traps to use. For
Rusty Russellb2b47c22007-07-26 10:41:02 -0700752 * native hardware, this is part of the Task State Segment mentioned above in
753 * lguest_load_tr_desc(), but to help hypervisors there's this special call.
754 *
755 * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data
756 * segment), the privilege level (we're privilege level 1, the Host is 0 and
757 * will not tolerate us trying to use that), the stack pointer, and the number
758 * of pages in the stack. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700759static void lguest_load_esp0(struct tss_struct *tss,
760 struct thread_struct *thread)
761{
762 lazy_hcall(LHCALL_SET_STACK, __KERNEL_DS|0x1, thread->esp0,
763 THREAD_SIZE/PAGE_SIZE);
764}
765
Rusty Russellb2b47c22007-07-26 10:41:02 -0700766/* Let's just say, I wouldn't do debugging under a Guest. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700767static void lguest_set_debugreg(int regno, unsigned long value)
768{
769 /* FIXME: Implement */
770}
771
Rusty Russellb2b47c22007-07-26 10:41:02 -0700772/* There are times when the kernel wants to make sure that no memory writes are
773 * caught in the cache (that they've all reached real hardware devices). This
774 * doesn't matter for the Guest which has virtual hardware.
775 *
776 * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush
777 * (clflush) instruction is available and the kernel uses that. Otherwise, it
778 * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction.
779 * Unlike clflush, wbinvd can only be run at privilege level 0. So we can
780 * ignore clflush, but replace wbinvd.
781 */
Rusty Russell07ad1572007-07-19 01:49:22 -0700782static void lguest_wbinvd(void)
783{
784}
785
Rusty Russellb2b47c22007-07-26 10:41:02 -0700786/* If the Guest expects to have an Advanced Programmable Interrupt Controller,
787 * we play dumb by ignoring writes and returning 0 for reads. So it's no
788 * longer Programmable nor Controlling anything, and I don't think 8 lines of
789 * code qualifies for Advanced. It will also never interrupt anything. It
790 * does, however, allow us to get through the Linux boot code. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700791#ifdef CONFIG_X86_LOCAL_APIC
792static void lguest_apic_write(unsigned long reg, unsigned long v)
793{
794}
795
796static unsigned long lguest_apic_read(unsigned long reg)
797{
798 return 0;
799}
800#endif
801
Rusty Russellb2b47c22007-07-26 10:41:02 -0700802/* STOP! Until an interrupt comes in. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700803static void lguest_safe_halt(void)
804{
805 hcall(LHCALL_HALT, 0, 0, 0);
806}
807
Rusty Russellb2b47c22007-07-26 10:41:02 -0700808/* Perhaps CRASH isn't the best name for this hypercall, but we use it to get a
809 * message out when we're crashing as well as elegant termination like powering
810 * off.
811 *
812 * Note that the Host always prefers that the Guest speak in physical addresses
813 * rather than virtual addresses, so we use __pa() here. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700814static void lguest_power_off(void)
815{
816 hcall(LHCALL_CRASH, __pa("Power down"), 0, 0);
817}
818
Rusty Russellb2b47c22007-07-26 10:41:02 -0700819/*
820 * Panicing.
821 *
822 * Don't. But if you did, this is what happens.
823 */
Rusty Russell07ad1572007-07-19 01:49:22 -0700824static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p)
825{
826 hcall(LHCALL_CRASH, __pa(p), 0, 0);
Rusty Russellb2b47c22007-07-26 10:41:02 -0700827 /* The hcall won't return, but to keep gcc happy, we're "done". */
Rusty Russell07ad1572007-07-19 01:49:22 -0700828 return NOTIFY_DONE;
829}
830
831static struct notifier_block paniced = {
832 .notifier_call = lguest_panic
833};
834
Rusty Russellb2b47c22007-07-26 10:41:02 -0700835/* Setting up memory is fairly easy. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700836static __init char *lguest_memory_setup(void)
837{
Rusty Russellb2b47c22007-07-26 10:41:02 -0700838 /* We do this here and not earlier because lockcheck barfs if we do it
839 * before start_kernel() */
Rusty Russell07ad1572007-07-19 01:49:22 -0700840 atomic_notifier_chain_register(&panic_notifier_list, &paniced);
841
Rusty Russellb2b47c22007-07-26 10:41:02 -0700842 /* The Linux bootloader header contains an "e820" memory map: the
843 * Launcher populated the first entry with our memory limit. */
H. Peter Anvin30c82642007-10-15 17:13:22 -0700844 add_memory_region(boot_params.e820_map[0].addr,
845 boot_params.e820_map[0].size,
846 boot_params.e820_map[0].type);
Rusty Russellb2b47c22007-07-26 10:41:02 -0700847
848 /* This string is for the boot messages. */
Rusty Russell07ad1572007-07-19 01:49:22 -0700849 return "LGUEST";
850}
851
Rusty Russelle1e72962007-10-25 15:02:50 +1000852/* We will eventually use the virtio console device to produce console output,
853 * but before that is set up we use LHCALL_NOTIFY on normal memory to produce
854 * console output. */
Rusty Russell19f15372007-10-22 11:24:21 +1000855static __init int early_put_chars(u32 vtermno, const char *buf, int count)
856{
857 char scratch[17];
858 unsigned int len = count;
859
Rusty Russelle1e72962007-10-25 15:02:50 +1000860 /* We use a nul-terminated string, so we have to make a copy. Icky,
861 * huh? */
Rusty Russell19f15372007-10-22 11:24:21 +1000862 if (len > sizeof(scratch) - 1)
863 len = sizeof(scratch) - 1;
864 scratch[len] = '\0';
865 memcpy(scratch, buf, len);
866 hcall(LHCALL_NOTIFY, __pa(scratch), 0, 0);
867
868 /* This routine returns the number of bytes actually written. */
869 return len;
870}
871
Rusty Russellb2b47c22007-07-26 10:41:02 -0700872/*G:050
873 * Patching (Powerfully Placating Performance Pedants)
874 *
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700875 * We have already seen that pv_ops structures let us replace simple
Rusty Russellb2b47c22007-07-26 10:41:02 -0700876 * native instructions with calls to the appropriate back end all throughout
877 * the kernel. This allows the same kernel to run as a Guest and as a native
878 * kernel, but it's slow because of all the indirect branches.
879 *
880 * Remember that David Wheeler quote about "Any problem in computer science can
881 * be solved with another layer of indirection"? The rest of that quote is
882 * "... But that usually will create another problem." This is the first of
883 * those problems.
884 *
885 * Our current solution is to allow the paravirt back end to optionally patch
886 * over the indirect calls to replace them with something more efficient. We
887 * patch the four most commonly called functions: disable interrupts, enable
Rusty Russelle1e72962007-10-25 15:02:50 +1000888 * interrupts, restore interrupts and save interrupts. We usually have 6 or 10
Rusty Russellb2b47c22007-07-26 10:41:02 -0700889 * bytes to patch into: the Guest versions of these operations are small enough
890 * that we can fit comfortably.
891 *
892 * First we need assembly templates of each of the patchable Guest operations,
893 * and these are in lguest_asm.S. */
894
895/*G:060 We construct a table from the assembler templates: */
Rusty Russell07ad1572007-07-19 01:49:22 -0700896static const struct lguest_insns
897{
898 const char *start, *end;
899} lguest_insns[] = {
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700900 [PARAVIRT_PATCH(pv_irq_ops.irq_disable)] = { lgstart_cli, lgend_cli },
901 [PARAVIRT_PATCH(pv_irq_ops.irq_enable)] = { lgstart_sti, lgend_sti },
902 [PARAVIRT_PATCH(pv_irq_ops.restore_fl)] = { lgstart_popf, lgend_popf },
903 [PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf },
Rusty Russell07ad1572007-07-19 01:49:22 -0700904};
Rusty Russellb2b47c22007-07-26 10:41:02 -0700905
906/* Now our patch routine is fairly simple (based on the native one in
907 * paravirt.c). If we have a replacement, we copy it in and return how much of
908 * the available space we used. */
Andi Kleenab144f52007-08-10 22:31:03 +0200909static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf,
910 unsigned long addr, unsigned len)
Rusty Russell07ad1572007-07-19 01:49:22 -0700911{
912 unsigned int insn_len;
913
Rusty Russellb2b47c22007-07-26 10:41:02 -0700914 /* Don't do anything special if we don't have a replacement */
Rusty Russell07ad1572007-07-19 01:49:22 -0700915 if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start)
Andi Kleenab144f52007-08-10 22:31:03 +0200916 return paravirt_patch_default(type, clobber, ibuf, addr, len);
Rusty Russell07ad1572007-07-19 01:49:22 -0700917
918 insn_len = lguest_insns[type].end - lguest_insns[type].start;
919
Rusty Russellb2b47c22007-07-26 10:41:02 -0700920 /* Similarly if we can't fit replacement (shouldn't happen, but let's
921 * be thorough). */
Rusty Russell07ad1572007-07-19 01:49:22 -0700922 if (len < insn_len)
Andi Kleenab144f52007-08-10 22:31:03 +0200923 return paravirt_patch_default(type, clobber, ibuf, addr, len);
Rusty Russell07ad1572007-07-19 01:49:22 -0700924
Rusty Russellb2b47c22007-07-26 10:41:02 -0700925 /* Copy in our instructions. */
Andi Kleenab144f52007-08-10 22:31:03 +0200926 memcpy(ibuf, lguest_insns[type].start, insn_len);
Rusty Russell07ad1572007-07-19 01:49:22 -0700927 return insn_len;
928}
929
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700930/*G:030 Once we get to lguest_init(), we know we're a Guest. The pv_ops
931 * structures in the kernel provide points for (almost) every routine we have
932 * to override to avoid privileged instructions. */
Rusty Russell814a0e52007-10-22 11:29:44 +1000933__init void lguest_init(void)
Rusty Russell07ad1572007-07-19 01:49:22 -0700934{
Rusty Russellb2b47c22007-07-26 10:41:02 -0700935 /* We're under lguest, paravirt is enabled, and we're running at
936 * privilege level 1, not 0 as normal. */
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700937 pv_info.name = "lguest";
938 pv_info.paravirt_enabled = 1;
939 pv_info.kernel_rpl = 1;
Rusty Russell07ad1572007-07-19 01:49:22 -0700940
Rusty Russellb2b47c22007-07-26 10:41:02 -0700941 /* We set up all the lguest overrides for sensitive operations. These
942 * are detailed with the operations themselves. */
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700943
944 /* interrupt-related operations */
945 pv_irq_ops.init_IRQ = lguest_init_IRQ;
946 pv_irq_ops.save_fl = save_fl;
947 pv_irq_ops.restore_fl = restore_fl;
948 pv_irq_ops.irq_disable = irq_disable;
949 pv_irq_ops.irq_enable = irq_enable;
950 pv_irq_ops.safe_halt = lguest_safe_halt;
951
952 /* init-time operations */
953 pv_init_ops.memory_setup = lguest_memory_setup;
954 pv_init_ops.patch = lguest_patch;
955
956 /* Intercepts of various cpu instructions */
957 pv_cpu_ops.load_gdt = lguest_load_gdt;
958 pv_cpu_ops.cpuid = lguest_cpuid;
959 pv_cpu_ops.load_idt = lguest_load_idt;
960 pv_cpu_ops.iret = lguest_iret;
961 pv_cpu_ops.load_esp0 = lguest_load_esp0;
962 pv_cpu_ops.load_tr_desc = lguest_load_tr_desc;
963 pv_cpu_ops.set_ldt = lguest_set_ldt;
964 pv_cpu_ops.load_tls = lguest_load_tls;
965 pv_cpu_ops.set_debugreg = lguest_set_debugreg;
966 pv_cpu_ops.clts = lguest_clts;
967 pv_cpu_ops.read_cr0 = lguest_read_cr0;
968 pv_cpu_ops.write_cr0 = lguest_write_cr0;
969 pv_cpu_ops.read_cr4 = lguest_read_cr4;
970 pv_cpu_ops.write_cr4 = lguest_write_cr4;
971 pv_cpu_ops.write_gdt_entry = lguest_write_gdt_entry;
972 pv_cpu_ops.write_idt_entry = lguest_write_idt_entry;
973 pv_cpu_ops.wbinvd = lguest_wbinvd;
Jeremy Fitzhardinge8965c1c2007-10-16 11:51:29 -0700974 pv_cpu_ops.lazy_mode.enter = paravirt_enter_lazy_cpu;
975 pv_cpu_ops.lazy_mode.leave = lguest_leave_lazy_mode;
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700976
977 /* pagetable management */
978 pv_mmu_ops.write_cr3 = lguest_write_cr3;
979 pv_mmu_ops.flush_tlb_user = lguest_flush_tlb_user;
980 pv_mmu_ops.flush_tlb_single = lguest_flush_tlb_single;
981 pv_mmu_ops.flush_tlb_kernel = lguest_flush_tlb_kernel;
982 pv_mmu_ops.set_pte = lguest_set_pte;
983 pv_mmu_ops.set_pte_at = lguest_set_pte_at;
984 pv_mmu_ops.set_pmd = lguest_set_pmd;
985 pv_mmu_ops.read_cr2 = lguest_read_cr2;
986 pv_mmu_ops.read_cr3 = lguest_read_cr3;
Jeremy Fitzhardinge8965c1c2007-10-16 11:51:29 -0700987 pv_mmu_ops.lazy_mode.enter = paravirt_enter_lazy_mmu;
988 pv_mmu_ops.lazy_mode.leave = lguest_leave_lazy_mode;
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700989
Rusty Russell07ad1572007-07-19 01:49:22 -0700990#ifdef CONFIG_X86_LOCAL_APIC
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700991 /* apic read/write intercepts */
992 pv_apic_ops.apic_write = lguest_apic_write;
993 pv_apic_ops.apic_write_atomic = lguest_apic_write;
994 pv_apic_ops.apic_read = lguest_apic_read;
Rusty Russell07ad1572007-07-19 01:49:22 -0700995#endif
Jeremy Fitzhardinge93b1eab2007-10-16 11:51:29 -0700996
997 /* time operations */
998 pv_time_ops.get_wallclock = lguest_get_wallclock;
999 pv_time_ops.time_init = lguest_time_init;
1000
Rusty Russellb2b47c22007-07-26 10:41:02 -07001001 /* Now is a good time to look at the implementations of these functions
1002 * before returning to the rest of lguest_init(). */
Rusty Russell07ad1572007-07-19 01:49:22 -07001003
Rusty Russellb2b47c22007-07-26 10:41:02 -07001004 /*G:070 Now we've seen all the paravirt_ops, we return to
1005 * lguest_init() where the rest of the fairly chaotic boot setup
Rusty Russell47436aa2007-10-22 11:03:36 +10001006 * occurs. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001007
Rusty Russellb2b47c22007-07-26 10:41:02 -07001008 /* The native boot code sets up initial page tables immediately after
1009 * the kernel itself, and sets init_pg_tables_end so they're not
1010 * clobbered. The Launcher places our initial pagetables somewhere at
1011 * the top of our physical memory, so we don't need extra space: set
1012 * init_pg_tables_end to the end of the kernel. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001013 init_pg_tables_end = __pa(pg0);
1014
Rusty Russellb2b47c22007-07-26 10:41:02 -07001015 /* Load the %fs segment register (the per-cpu segment register) with
1016 * the normal data segment to get through booting. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001017 asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS) : "memory");
1018
Rusty Russellb2b47c22007-07-26 10:41:02 -07001019 /* The Host uses the top of the Guest's virtual address space for the
Rusty Russelle1e72962007-10-25 15:02:50 +10001020 * Host<->Guest Switcher, and it tells us how big that is in
Rusty Russellb2b47c22007-07-26 10:41:02 -07001021 * lguest_data.reserve_mem, set up on the LGUEST_INIT hypercall. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001022 reserve_top_address(lguest_data.reserve_mem);
1023
Rusty Russellb2b47c22007-07-26 10:41:02 -07001024 /* If we don't initialize the lock dependency checker now, it crashes
1025 * paravirt_disable_iospace. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001026 lockdep_init();
1027
Rusty Russellb2b47c22007-07-26 10:41:02 -07001028 /* The IDE code spends about 3 seconds probing for disks: if we reserve
1029 * all the I/O ports up front it can't get them and so doesn't probe.
1030 * Other device drivers are similar (but less severe). This cuts the
1031 * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001032 paravirt_disable_iospace();
1033
Rusty Russellb2b47c22007-07-26 10:41:02 -07001034 /* This is messy CPU setup stuff which the native boot code does before
1035 * start_kernel, so we have to do, too: */
Rusty Russell07ad1572007-07-19 01:49:22 -07001036 cpu_detect(&new_cpu_data);
1037 /* head.S usually sets up the first capability word, so do it here. */
1038 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1039
1040 /* Math is always hard! */
1041 new_cpu_data.hard_math = 1;
1042
1043#ifdef CONFIG_X86_MCE
1044 mce_disabled = 1;
1045#endif
Rusty Russell07ad1572007-07-19 01:49:22 -07001046#ifdef CONFIG_ACPI
1047 acpi_disabled = 1;
1048 acpi_ht = 0;
1049#endif
1050
Rusty Russellb2b47c22007-07-26 10:41:02 -07001051 /* We set the perferred console to "hvc". This is the "hypervisor
1052 * virtual console" driver written by the PowerPC people, which we also
1053 * adapted for lguest's use. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001054 add_preferred_console("hvc", 0, NULL);
1055
Rusty Russell19f15372007-10-22 11:24:21 +10001056 /* Register our very early console. */
1057 virtio_cons_early_init(early_put_chars);
1058
Rusty Russellb2b47c22007-07-26 10:41:02 -07001059 /* Last of all, we set the power management poweroff hook to point to
1060 * the Guest routine to power off. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001061 pm_power_off = lguest_power_off;
Rusty Russellb2b47c22007-07-26 10:41:02 -07001062
1063 /* Now we're set up, call start_kernel() in init/main.c and we proceed
1064 * to boot as normal. It never returns. */
Rusty Russell07ad1572007-07-19 01:49:22 -07001065 start_kernel();
1066}
Rusty Russellb2b47c22007-07-26 10:41:02 -07001067/*
1068 * This marks the end of stage II of our journey, The Guest.
1069 *
Rusty Russelle1e72962007-10-25 15:02:50 +10001070 * It is now time for us to explore the layer of virtual drivers and complete
1071 * our understanding of the Guest in "make Drivers".
Rusty Russellb2b47c22007-07-26 10:41:02 -07001072 */