aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/lguest/lguest.c
diff options
context:
space:
mode:
authorRusty Russell <rusty@rustcorp.com.au>2007-07-26 13:41:02 -0400
committerLinus Torvalds <torvalds@woody.linux-foundation.org>2007-07-26 14:35:17 -0400
commitb2b47c214f4e85ce3968120d42e8b18eccb4f4e3 (patch)
treef77d6898a769b8e0fcb552207e87f273bdc19f09 /drivers/lguest/lguest.c
parentf938d2c892db0d80d144253d4a7b7083efdbedeb (diff)
lguest: documentation II: Guest
Documentation: The Guest Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'drivers/lguest/lguest.c')
-rw-r--r--drivers/lguest/lguest.c450
1 files changed, 429 insertions, 21 deletions
diff --git a/drivers/lguest/lguest.c b/drivers/lguest/lguest.c
index e7d128312b23..7e7e9fb3aefd 100644
--- a/drivers/lguest/lguest.c
+++ b/drivers/lguest/lguest.c
@@ -66,6 +66,12 @@
66#include <asm/mce.h> 66#include <asm/mce.h>
67#include <asm/io.h> 67#include <asm/io.h>
68 68
69/*G:010 Welcome to the Guest!
70 *
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). :*/
74
69/* Declarations for definitions in lguest_guest.S */ 75/* Declarations for definitions in lguest_guest.S */
70extern char lguest_noirq_start[], lguest_noirq_end[]; 76extern char lguest_noirq_start[], lguest_noirq_end[];
71extern const char lgstart_cli[], lgend_cli[]; 77extern const char lgstart_cli[], lgend_cli[];
@@ -84,7 +90,26 @@ struct lguest_data lguest_data = {
84struct lguest_device_desc *lguest_devices; 90struct lguest_device_desc *lguest_devices;
85static cycle_t clock_base; 91static cycle_t clock_base;
86 92
87static enum paravirt_lazy_mode lazy_mode; 93/*G:035 Notice the lazy_hcall() above, rather than hcall(). This is our first
94 * real optimization trick!
95 *
96 * When lazy_mode is set, it means we're allowed to defer all hypercalls and do
97 * them as a batch when lazy_mode is eventually turned off. Because hypercalls
98 * are reasonably expensive, batching them up makes sense. For example, a
99 * large mmap might update dozens of page table entries: that code calls
100 * lguest_lazy_mode(PARAVIRT_LAZY_MMU), does the dozen updates, then calls
101 * lguest_lazy_mode(PARAVIRT_LAZY_NONE).
102 *
103 * So, when we're in lazy mode, we call async_hypercall() to store the call for
104 * future processing. When lazy mode is turned off we issue a hypercall to
105 * flush the stored calls.
106 *
107 * There's also a hack where "mode" is set to "PARAVIRT_LAZY_FLUSH" which
108 * indicates we're to flush any outstanding calls immediately. This is used
109 * when an interrupt handler does a kmap_atomic(): the page table changes must
110 * happen immediately even if we're in the middle of a batch. Usually we're
111 * not, though, so there's nothing to do. */
112static enum paravirt_lazy_mode lazy_mode; /* Note: not SMP-safe! */
88static void lguest_lazy_mode(enum paravirt_lazy_mode mode) 113static void lguest_lazy_mode(enum paravirt_lazy_mode mode)
89{ 114{
90 if (mode == PARAVIRT_LAZY_FLUSH) { 115 if (mode == PARAVIRT_LAZY_FLUSH) {
@@ -108,6 +133,16 @@ static void lazy_hcall(unsigned long call,
108 async_hcall(call, arg1, arg2, arg3); 133 async_hcall(call, arg1, arg2, arg3);
109} 134}
110 135
136/* async_hcall() is pretty simple: I'm quite proud of it really. We have a
137 * ring buffer of stored hypercalls which the Host will run though next time we
138 * do a normal hypercall. Each entry in the ring has 4 slots for the hypercall
139 * arguments, and a "hcall_status" word which is 0 if the call is ready to go,
140 * and 255 once the Host has finished with it.
141 *
142 * If we come around to a slot which hasn't been finished, then the table is
143 * full and we just make the hypercall directly. This has the nice side
144 * effect of causing the Host to run all the stored calls in the ring buffer
145 * which empties it for next time! */
111void async_hcall(unsigned long call, 146void async_hcall(unsigned long call,
112 unsigned long arg1, unsigned long arg2, unsigned long arg3) 147 unsigned long arg1, unsigned long arg2, unsigned long arg3)
113{ 148{
@@ -115,6 +150,9 @@ void async_hcall(unsigned long call,
115 static unsigned int next_call; 150 static unsigned int next_call;
116 unsigned long flags; 151 unsigned long flags;
117 152
153 /* Disable interrupts if not already disabled: we don't want an
154 * interrupt handler making a hypercall while we're already doing
155 * one! */
118 local_irq_save(flags); 156 local_irq_save(flags);
119 if (lguest_data.hcall_status[next_call] != 0xFF) { 157 if (lguest_data.hcall_status[next_call] != 0xFF) {
120 /* Table full, so do normal hcall which will flush table. */ 158 /* Table full, so do normal hcall which will flush table. */
@@ -124,7 +162,7 @@ void async_hcall(unsigned long call,
124 lguest_data.hcalls[next_call].edx = arg1; 162 lguest_data.hcalls[next_call].edx = arg1;
125 lguest_data.hcalls[next_call].ebx = arg2; 163 lguest_data.hcalls[next_call].ebx = arg2;
126 lguest_data.hcalls[next_call].ecx = arg3; 164 lguest_data.hcalls[next_call].ecx = arg3;
127 /* Make sure host sees arguments before "valid" flag. */ 165 /* Arguments must all be written before we mark it to go */
128 wmb(); 166 wmb();
129 lguest_data.hcall_status[next_call] = 0; 167 lguest_data.hcall_status[next_call] = 0;
130 if (++next_call == LHCALL_RING_SIZE) 168 if (++next_call == LHCALL_RING_SIZE)
@@ -132,9 +170,14 @@ void async_hcall(unsigned long call,
132 } 170 }
133 local_irq_restore(flags); 171 local_irq_restore(flags);
134} 172}
173/*:*/
135 174
175/* Wrappers for the SEND_DMA and BIND_DMA hypercalls. This is mainly because
176 * Jeff Garzik complained that __pa() should never appear in drivers, and this
177 * helps remove most of them. But also, it wraps some ugliness. */
136void lguest_send_dma(unsigned long key, struct lguest_dma *dma) 178void lguest_send_dma(unsigned long key, struct lguest_dma *dma)
137{ 179{
180 /* The hcall might not write this if something goes wrong */
138 dma->used_len = 0; 181 dma->used_len = 0;
139 hcall(LHCALL_SEND_DMA, key, __pa(dma), 0); 182 hcall(LHCALL_SEND_DMA, key, __pa(dma), 0);
140} 183}
@@ -142,11 +185,16 @@ void lguest_send_dma(unsigned long key, struct lguest_dma *dma)
142int lguest_bind_dma(unsigned long key, struct lguest_dma *dmas, 185int lguest_bind_dma(unsigned long key, struct lguest_dma *dmas,
143 unsigned int num, u8 irq) 186 unsigned int num, u8 irq)
144{ 187{
188 /* This is the only hypercall which actually wants 5 arguments, and we
189 * only support 4. Fortunately the interrupt number is always less
190 * than 256, so we can pack it with the number of dmas in the final
191 * argument. */
145 if (!hcall(LHCALL_BIND_DMA, key, __pa(dmas), (num << 8) | irq)) 192 if (!hcall(LHCALL_BIND_DMA, key, __pa(dmas), (num << 8) | irq))
146 return -ENOMEM; 193 return -ENOMEM;
147 return 0; 194 return 0;
148} 195}
149 196
197/* Unbinding is the same hypercall as binding, but with 0 num & irq. */
150void lguest_unbind_dma(unsigned long key, struct lguest_dma *dmas) 198void lguest_unbind_dma(unsigned long key, struct lguest_dma *dmas)
151{ 199{
152 hcall(LHCALL_BIND_DMA, key, __pa(dmas), 0); 200 hcall(LHCALL_BIND_DMA, key, __pa(dmas), 0);
@@ -164,35 +212,65 @@ void lguest_unmap(void *addr)
164 iounmap((__force void __iomem *)addr); 212 iounmap((__force void __iomem *)addr);
165} 213}
166 214
215/*G:033
216 * Here are our first native-instruction replacements: four functions for
217 * interrupt control.
218 *
219 * The simplest way of implementing these would be to have "turn interrupts
220 * off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow:
221 * these are by far the most commonly called functions of those we override.
222 *
223 * So instead we keep an "irq_enabled" field inside our "struct lguest_data",
224 * which the Guest can update with a single instruction. The Host knows to
225 * check there when it wants to deliver an interrupt.
226 */
227
228/* save_flags() is expected to return the processor state (ie. "eflags"). The
229 * eflags word contains all kind of stuff, but in practice Linux only cares
230 * about the interrupt flag. Our "save_flags()" just returns that. */
167static unsigned long save_fl(void) 231static unsigned long save_fl(void)
168{ 232{
169 return lguest_data.irq_enabled; 233 return lguest_data.irq_enabled;
170} 234}
171 235
236/* "restore_flags" just sets the flags back to the value given. */
172static void restore_fl(unsigned long flags) 237static void restore_fl(unsigned long flags)
173{ 238{
174 /* FIXME: Check if interrupt pending... */
175 lguest_data.irq_enabled = flags; 239 lguest_data.irq_enabled = flags;
176} 240}
177 241
242/* Interrupts go off... */
178static void irq_disable(void) 243static void irq_disable(void)
179{ 244{
180 lguest_data.irq_enabled = 0; 245 lguest_data.irq_enabled = 0;
181} 246}
182 247
248/* Interrupts go on... */
183static void irq_enable(void) 249static void irq_enable(void)
184{ 250{
185 /* FIXME: Check if interrupt pending... */
186 lguest_data.irq_enabled = X86_EFLAGS_IF; 251 lguest_data.irq_enabled = X86_EFLAGS_IF;
187} 252}
188 253
254/*G:034
255 * The Interrupt Descriptor Table (IDT).
256 *
257 * The IDT tells the processor what to do when an interrupt comes in. Each
258 * entry in the table is a 64-bit descriptor: this holds the privilege level,
259 * address of the handler, and... well, who cares? The Guest just asks the
260 * Host to make the change anyway, because the Host controls the real IDT.
261 */
189static void lguest_write_idt_entry(struct desc_struct *dt, 262static void lguest_write_idt_entry(struct desc_struct *dt,
190 int entrynum, u32 low, u32 high) 263 int entrynum, u32 low, u32 high)
191{ 264{
265 /* Keep the local copy up to date. */
192 write_dt_entry(dt, entrynum, low, high); 266 write_dt_entry(dt, entrynum, low, high);
267 /* Tell Host about this new entry. */
193 hcall(LHCALL_LOAD_IDT_ENTRY, entrynum, low, high); 268 hcall(LHCALL_LOAD_IDT_ENTRY, entrynum, low, high);
194} 269}
195 270
271/* Changing to a different IDT is very rare: we keep the IDT up-to-date every
272 * time it is written, so we can simply loop through all entries and tell the
273 * Host about them. */
196static void lguest_load_idt(const struct Xgt_desc_struct *desc) 274static void lguest_load_idt(const struct Xgt_desc_struct *desc)
197{ 275{
198 unsigned int i; 276 unsigned int i;
@@ -202,12 +280,29 @@ static void lguest_load_idt(const struct Xgt_desc_struct *desc)
202 hcall(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b); 280 hcall(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b);
203} 281}
204 282
283/*
284 * The Global Descriptor Table.
285 *
286 * The Intel architecture defines another table, called the Global Descriptor
287 * Table (GDT). You tell the CPU where it is (and its size) using the "lgdt"
288 * instruction, and then several other instructions refer to entries in the
289 * table. There are three entries which the Switcher needs, so the Host simply
290 * controls the entire thing and the Guest asks it to make changes using the
291 * LOAD_GDT hypercall.
292 *
293 * This is the opposite of the IDT code where we have a LOAD_IDT_ENTRY
294 * hypercall and use that repeatedly to load a new IDT. I don't think it
295 * really matters, but wouldn't it be nice if they were the same?
296 */
205static void lguest_load_gdt(const struct Xgt_desc_struct *desc) 297static void lguest_load_gdt(const struct Xgt_desc_struct *desc)
206{ 298{
207 BUG_ON((desc->size+1)/8 != GDT_ENTRIES); 299 BUG_ON((desc->size+1)/8 != GDT_ENTRIES);
208 hcall(LHCALL_LOAD_GDT, __pa(desc->address), GDT_ENTRIES, 0); 300 hcall(LHCALL_LOAD_GDT, __pa(desc->address), GDT_ENTRIES, 0);
209} 301}
210 302
303/* For a single GDT entry which changes, we do the lazy thing: alter our GDT,
304 * then tell the Host to reload the entire thing. This operation is so rare
305 * that this naive implementation is reasonable. */
211static void lguest_write_gdt_entry(struct desc_struct *dt, 306static void lguest_write_gdt_entry(struct desc_struct *dt,
212 int entrynum, u32 low, u32 high) 307 int entrynum, u32 low, u32 high)
213{ 308{
@@ -215,19 +310,58 @@ static void lguest_write_gdt_entry(struct desc_struct *dt,
215 hcall(LHCALL_LOAD_GDT, __pa(dt), GDT_ENTRIES, 0); 310 hcall(LHCALL_LOAD_GDT, __pa(dt), GDT_ENTRIES, 0);
216} 311}
217 312
313/* OK, I lied. There are three "thread local storage" GDT entries which change
314 * on every context switch (these three entries are how glibc implements
315 * __thread variables). So we have a hypercall specifically for this case. */
218static void lguest_load_tls(struct thread_struct *t, unsigned int cpu) 316static void lguest_load_tls(struct thread_struct *t, unsigned int cpu)
219{ 317{
220 lazy_hcall(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu, 0); 318 lazy_hcall(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu, 0);
221} 319}
320/*:*/
222 321
322/*G:038 That's enough excitement for now, back to ploughing through each of
323 * the paravirt_ops (we're about 1/3 of the way through).
324 *
325 * This is the Local Descriptor Table, another weird Intel thingy. Linux only
326 * uses this for some strange applications like Wine. We don't do anything
327 * here, so they'll get an informative and friendly Segmentation Fault. */
223static void lguest_set_ldt(const void *addr, unsigned entries) 328static void lguest_set_ldt(const void *addr, unsigned entries)
224{ 329{
225} 330}
226 331
332/* This loads a GDT entry into the "Task Register": that entry points to a
333 * structure called the Task State Segment. Some comments scattered though the
334 * kernel code indicate that this used for task switching in ages past, along
335 * with blood sacrifice and astrology.
336 *
337 * Now there's nothing interesting in here that we don't get told elsewhere.
338 * But the native version uses the "ltr" instruction, which makes the Host
339 * complain to the Guest about a Segmentation Fault and it'll oops. So we
340 * override the native version with a do-nothing version. */
227static void lguest_load_tr_desc(void) 341static void lguest_load_tr_desc(void)
228{ 342{
229} 343}
230 344
345/* The "cpuid" instruction is a way of querying both the CPU identity
346 * (manufacturer, model, etc) and its features. It was introduced before the
347 * Pentium in 1993 and keeps getting extended by both Intel and AMD. As you
348 * might imagine, after a decade and a half this treatment, it is now a giant
349 * ball of hair. Its entry in the current Intel manual runs to 28 pages.
350 *
351 * This instruction even it has its own Wikipedia entry. The Wikipedia entry
352 * has been translated into 4 languages. I am not making this up!
353 *
354 * We could get funky here and identify ourselves as "GenuineLguest", but
355 * instead we just use the real "cpuid" instruction. Then I pretty much turned
356 * off feature bits until the Guest booted. (Don't say that: you'll damage
357 * lguest sales!) Shut up, inner voice! (Hey, just pointing out that this is
358 * hardly future proof.) Noone's listening! They don't like you anyway,
359 * parenthetic weirdo!
360 *
361 * Replacing the cpuid so we can turn features off is great for the kernel, but
362 * anyone (including userspace) can just use the raw "cpuid" instruction and
363 * the Host won't even notice since it isn't privileged. So we try not to get
364 * too worked up about it. */
231static void lguest_cpuid(unsigned int *eax, unsigned int *ebx, 365static void lguest_cpuid(unsigned int *eax, unsigned int *ebx,
232 unsigned int *ecx, unsigned int *edx) 366 unsigned int *ecx, unsigned int *edx)
233{ 367{
@@ -240,21 +374,43 @@ static void lguest_cpuid(unsigned int *eax, unsigned int *ebx,
240 *ecx &= 0x00002201; 374 *ecx &= 0x00002201;
241 /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, FPU. */ 375 /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, FPU. */
242 *edx &= 0x07808101; 376 *edx &= 0x07808101;
243 /* Host wants to know when we flush kernel pages: set PGE. */ 377 /* The Host can do a nice optimization if it knows that the
378 * kernel mappings (addresses above 0xC0000000 or whatever
379 * PAGE_OFFSET is set to) haven't changed. But Linux calls
380 * flush_tlb_user() for both user and kernel mappings unless
381 * the Page Global Enable (PGE) feature bit is set. */
244 *edx |= 0x00002000; 382 *edx |= 0x00002000;
245 break; 383 break;
246 case 0x80000000: 384 case 0x80000000:
247 /* Futureproof this a little: if they ask how much extended 385 /* Futureproof this a little: if they ask how much extended
248 * processor information, limit it to known fields. */ 386 * processor information there is, limit it to known fields. */
249 if (*eax > 0x80000008) 387 if (*eax > 0x80000008)
250 *eax = 0x80000008; 388 *eax = 0x80000008;
251 break; 389 break;
252 } 390 }
253} 391}
254 392
393/* Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4.
394 * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother
395 * it. The Host needs to know when the Guest wants to change them, so we have
396 * a whole series of functions like read_cr0() and write_cr0().
397 *
398 * We start with CR0. CR0 allows you to turn on and off all kinds of basic
399 * features, but Linux only really cares about one: the horrifically-named Task
400 * Switched (TS) bit at bit 3 (ie. 8)
401 *
402 * What does the TS bit do? Well, it causes the CPU to trap (interrupt 7) if
403 * the floating point unit is used. Which allows us to restore FPU state
404 * lazily after a task switch, and Linux uses that gratefully, but wouldn't a
405 * name like "FPUTRAP bit" be a little less cryptic?
406 *
407 * We store cr0 (and cr3) locally, because the Host never changes it. The
408 * Guest sometimes wants to read it and we'd prefer not to bother the Host
409 * unnecessarily. */
255static unsigned long current_cr0, current_cr3; 410static unsigned long current_cr0, current_cr3;
256static void lguest_write_cr0(unsigned long val) 411static void lguest_write_cr0(unsigned long val)
257{ 412{
413 /* 8 == TS bit. */
258 lazy_hcall(LHCALL_TS, val & 8, 0, 0); 414 lazy_hcall(LHCALL_TS, val & 8, 0, 0);
259 current_cr0 = val; 415 current_cr0 = val;
260} 416}
@@ -264,17 +420,25 @@ static unsigned long lguest_read_cr0(void)
264 return current_cr0; 420 return current_cr0;
265} 421}
266 422
423/* Intel provided a special instruction to clear the TS bit for people too cool
424 * to use write_cr0() to do it. This "clts" instruction is faster, because all
425 * the vowels have been optimized out. */
267static void lguest_clts(void) 426static void lguest_clts(void)
268{ 427{
269 lazy_hcall(LHCALL_TS, 0, 0, 0); 428 lazy_hcall(LHCALL_TS, 0, 0, 0);
270 current_cr0 &= ~8U; 429 current_cr0 &= ~8U;
271} 430}
272 431
432/* CR2 is the virtual address of the last page fault, which the Guest only ever
433 * reads. The Host kindly writes this into our "struct lguest_data", so we
434 * just read it out of there. */
273static unsigned long lguest_read_cr2(void) 435static unsigned long lguest_read_cr2(void)
274{ 436{
275 return lguest_data.cr2; 437 return lguest_data.cr2;
276} 438}
277 439
440/* CR3 is the current toplevel pagetable page: the principle is the same as
441 * cr0. Keep a local copy, and tell the Host when it changes. */
278static void lguest_write_cr3(unsigned long cr3) 442static void lguest_write_cr3(unsigned long cr3)
279{ 443{
280 lazy_hcall(LHCALL_NEW_PGTABLE, cr3, 0, 0); 444 lazy_hcall(LHCALL_NEW_PGTABLE, cr3, 0, 0);
@@ -286,7 +450,7 @@ static unsigned long lguest_read_cr3(void)
286 return current_cr3; 450 return current_cr3;
287} 451}
288 452
289/* Used to enable/disable PGE, but we don't care. */ 453/* CR4 is used to enable and disable PGE, but we don't care. */
290static unsigned long lguest_read_cr4(void) 454static unsigned long lguest_read_cr4(void)
291{ 455{
292 return 0; 456 return 0;
@@ -296,6 +460,59 @@ static void lguest_write_cr4(unsigned long val)
296{ 460{
297} 461}
298 462
463/*
464 * Page Table Handling.
465 *
466 * Now would be a good time to take a rest and grab a coffee or similarly
467 * relaxing stimulant. The easy parts are behind us, and the trek gradually
468 * winds uphill from here.
469 *
470 * Quick refresher: memory is divided into "pages" of 4096 bytes each. The CPU
471 * maps virtual addresses to physical addresses using "page tables". We could
472 * use one huge index of 1 million entries: each address is 4 bytes, so that's
473 * 1024 pages just to hold the page tables. But since most virtual addresses
474 * are unused, we use a two level index which saves space. The CR3 register
475 * contains the physical address of the top level "page directory" page, which
476 * contains physical addresses of up to 1024 second-level pages. Each of these
477 * second level pages contains up to 1024 physical addresses of actual pages,
478 * or Page Table Entries (PTEs).
479 *
480 * Here's a diagram, where arrows indicate physical addresses:
481 *
482 * CR3 ---> +---------+
483 * | --------->+---------+
484 * | | | PADDR1 |
485 * Top-level | | PADDR2 |
486 * (PMD) page | | |
487 * | | Lower-level |
488 * | | (PTE) page |
489 * | | | |
490 * .... ....
491 *
492 * So to convert a virtual address to a physical address, we look up the top
493 * level, which points us to the second level, which gives us the physical
494 * address of that page. If the top level entry was not present, or the second
495 * level entry was not present, then the virtual address is invalid (we
496 * say "the page was not mapped").
497 *
498 * Put another way, a 32-bit virtual address is divided up like so:
499 *
500 * 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
501 * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>|
502 * Index into top Index into second Offset within page
503 * page directory page pagetable page
504 *
505 * The kernel spends a lot of time changing both the top-level page directory
506 * and lower-level pagetable pages. The Guest doesn't know physical addresses,
507 * so while it maintains these page tables exactly like normal, it also needs
508 * to keep the Host informed whenever it makes a change: the Host will create
509 * the real page tables based on the Guests'.
510 */
511
512/* The Guest calls this to set a second-level entry (pte), ie. to map a page
513 * into a process' address space. We set the entry then tell the Host the
514 * toplevel and address this corresponds to. The Guest uses one pagetable per
515 * process, so we need to tell the Host which one we're changing (mm->pgd). */
299static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr, 516static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr,
300 pte_t *ptep, pte_t pteval) 517 pte_t *ptep, pte_t pteval)
301{ 518{
@@ -303,7 +520,9 @@ static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr,
303 lazy_hcall(LHCALL_SET_PTE, __pa(mm->pgd), addr, pteval.pte_low); 520 lazy_hcall(LHCALL_SET_PTE, __pa(mm->pgd), addr, pteval.pte_low);
304} 521}
305 522
306/* We only support two-level pagetables at the moment. */ 523/* The Guest calls this to set a top-level entry. Again, we set the entry then
524 * tell the Host which top-level page we changed, and the index of the entry we
525 * changed. */
307static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) 526static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
308{ 527{
309 *pmdp = pmdval; 528 *pmdp = pmdval;
@@ -311,7 +530,15 @@ static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
311 (__pa(pmdp)&(PAGE_SIZE-1))/4, 0); 530 (__pa(pmdp)&(PAGE_SIZE-1))/4, 0);
312} 531}
313 532
314/* FIXME: Eliminate all callers of this. */ 533/* There are a couple of legacy places where the kernel sets a PTE, but we
534 * don't know the top level any more. This is useless for us, since we don't
535 * know which pagetable is changing or what address, so we just tell the Host
536 * to forget all of them. Fortunately, this is very rare.
537 *
538 * ... except in early boot when the kernel sets up the initial pagetables,
539 * which makes booting astonishingly slow. So we don't even tell the Host
540 * anything changed until we've done the first page table switch.
541 */
315static void lguest_set_pte(pte_t *ptep, pte_t pteval) 542static void lguest_set_pte(pte_t *ptep, pte_t pteval)
316{ 543{
317 *ptep = pteval; 544 *ptep = pteval;
@@ -320,22 +547,51 @@ static void lguest_set_pte(pte_t *ptep, pte_t pteval)
320 lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0); 547 lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0);
321} 548}
322 549
550/* Unfortunately for Lguest, the paravirt_ops for page tables were based on
551 * native page table operations. On native hardware you can set a new page
552 * table entry whenever you want, but if you want to remove one you have to do
553 * a TLB flush (a TLB is a little cache of page table entries kept by the CPU).
554 *
555 * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only
556 * called when a valid entry is written, not when it's removed (ie. marked not
557 * present). Instead, this is where we come when the Guest wants to remove a
558 * page table entry: we tell the Host to set that entry to 0 (ie. the present
559 * bit is zero). */
323static void lguest_flush_tlb_single(unsigned long addr) 560static void lguest_flush_tlb_single(unsigned long addr)
324{ 561{
325 /* Simply set it to zero, and it will fault back in. */ 562 /* Simply set it to zero: if it was not, it will fault back in. */
326 lazy_hcall(LHCALL_SET_PTE, current_cr3, addr, 0); 563 lazy_hcall(LHCALL_SET_PTE, current_cr3, addr, 0);
327} 564}
328 565
566/* This is what happens after the Guest has removed a large number of entries.
567 * This tells the Host that any of the page table entries for userspace might
568 * have changed, ie. virtual addresses below PAGE_OFFSET. */
329static void lguest_flush_tlb_user(void) 569static void lguest_flush_tlb_user(void)
330{ 570{
331 lazy_hcall(LHCALL_FLUSH_TLB, 0, 0, 0); 571 lazy_hcall(LHCALL_FLUSH_TLB, 0, 0, 0);
332} 572}
333 573
574/* This is called when the kernel page tables have changed. That's not very
575 * common (unless the Guest is using highmem, which makes the Guest extremely
576 * slow), so it's worth separating this from the user flushing above. */
334static void lguest_flush_tlb_kernel(void) 577static void lguest_flush_tlb_kernel(void)
335{ 578{
336 lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0); 579 lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0);
337} 580}
338 581
582/*
583 * The Unadvanced Programmable Interrupt Controller.
584 *
585 * This is an attempt to implement the simplest possible interrupt controller.
586 * I spent some time looking though routines like set_irq_chip_and_handler,
587 * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and
588 * I *think* this is as simple as it gets.
589 *
590 * We can tell the Host what interrupts we want blocked ready for using the
591 * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as
592 * simple as setting a bit. We don't actually "ack" interrupts as such, we
593 * just mask and unmask them. I wonder if we should be cleverer?
594 */
339static void disable_lguest_irq(unsigned int irq) 595static void disable_lguest_irq(unsigned int irq)
340{ 596{
341 set_bit(irq, lguest_data.blocked_interrupts); 597 set_bit(irq, lguest_data.blocked_interrupts);
@@ -344,9 +600,9 @@ static void disable_lguest_irq(unsigned int irq)
344static void enable_lguest_irq(unsigned int irq) 600static void enable_lguest_irq(unsigned int irq)
345{ 601{
346 clear_bit(irq, lguest_data.blocked_interrupts); 602 clear_bit(irq, lguest_data.blocked_interrupts);
347 /* FIXME: If it's pending? */
348} 603}
349 604
605/* This structure describes the lguest IRQ controller. */
350static struct irq_chip lguest_irq_controller = { 606static struct irq_chip lguest_irq_controller = {
351 .name = "lguest", 607 .name = "lguest",
352 .mask = disable_lguest_irq, 608 .mask = disable_lguest_irq,
@@ -354,6 +610,10 @@ static struct irq_chip lguest_irq_controller = {
354 .unmask = enable_lguest_irq, 610 .unmask = enable_lguest_irq,
355}; 611};
356 612
613/* This sets up the Interrupt Descriptor Table (IDT) entry for each hardware
614 * interrupt (except 128, which is used for system calls), and then tells the
615 * Linux infrastructure that each interrupt is controlled by our level-based
616 * lguest interrupt controller. */
357static void __init lguest_init_IRQ(void) 617static void __init lguest_init_IRQ(void)
358{ 618{
359 unsigned int i; 619 unsigned int i;
@@ -366,14 +626,24 @@ static void __init lguest_init_IRQ(void)
366 handle_level_irq); 626 handle_level_irq);
367 } 627 }
368 } 628 }
629 /* This call is required to set up for 4k stacks, where we have
630 * separate stacks for hard and soft interrupts. */
369 irq_ctx_init(smp_processor_id()); 631 irq_ctx_init(smp_processor_id());
370} 632}
371 633
634/*
635 * Time.
636 *
637 * It would be far better for everyone if the Guest had its own clock, but
638 * until then it must ask the Host for the time.
639 */
372static unsigned long lguest_get_wallclock(void) 640static unsigned long lguest_get_wallclock(void)
373{ 641{
374 return hcall(LHCALL_GET_WALLCLOCK, 0, 0, 0); 642 return hcall(LHCALL_GET_WALLCLOCK, 0, 0, 0);
375} 643}
376 644
645/* If the Host tells us we can trust the TSC, we use that, otherwise we simply
646 * use the imprecise but reliable "jiffies" counter. */
377static cycle_t lguest_clock_read(void) 647static cycle_t lguest_clock_read(void)
378{ 648{
379 if (lguest_data.tsc_khz) 649 if (lguest_data.tsc_khz)
@@ -454,12 +724,19 @@ static void lguest_time_irq(unsigned int irq, struct irq_desc *desc)
454 local_irq_restore(flags); 724 local_irq_restore(flags);
455} 725}
456 726
727/* At some point in the boot process, we get asked to set up our timing
728 * infrastructure. The kernel doesn't expect timer interrupts before this, but
729 * we cleverly initialized the "blocked_interrupts" field of "struct
730 * lguest_data" so that timer interrupts were blocked until now. */
457static void lguest_time_init(void) 731static void lguest_time_init(void)
458{ 732{
733 /* Set up the timer interrupt (0) to go to our simple timer routine */
459 set_irq_handler(0, lguest_time_irq); 734 set_irq_handler(0, lguest_time_irq);
460 735
461 /* We use the TSC if the Host tells us we can, otherwise a dumb 736 /* Our clock structure look like arch/i386/kernel/tsc.c if we can use
462 * jiffies-based clock. */ 737 * the TSC, otherwise it looks like kernel/time/jiffies.c. Either way,
738 * the "rating" is initialized so high that it's always chosen over any
739 * other clocksource. */
463 if (lguest_data.tsc_khz) { 740 if (lguest_data.tsc_khz) {
464 lguest_clock.shift = 22; 741 lguest_clock.shift = 22;
465 lguest_clock.mult = clocksource_khz2mult(lguest_data.tsc_khz, 742 lguest_clock.mult = clocksource_khz2mult(lguest_data.tsc_khz,
@@ -475,13 +752,30 @@ static void lguest_time_init(void)
475 clock_base = lguest_clock_read(); 752 clock_base = lguest_clock_read();
476 clocksource_register(&lguest_clock); 753 clocksource_register(&lguest_clock);
477 754
478 /* We can't set cpumask in the initializer: damn C limitations! */ 755 /* We can't set cpumask in the initializer: damn C limitations! Set it
756 * here and register our timer device. */
479 lguest_clockevent.cpumask = cpumask_of_cpu(0); 757 lguest_clockevent.cpumask = cpumask_of_cpu(0);
480 clockevents_register_device(&lguest_clockevent); 758 clockevents_register_device(&lguest_clockevent);
481 759
760 /* Finally, we unblock the timer interrupt. */
482 enable_lguest_irq(0); 761 enable_lguest_irq(0);
483} 762}
484 763
764/*
765 * Miscellaneous bits and pieces.
766 *
767 * Here is an oddball collection of functions which the Guest needs for things
768 * to work. They're pretty simple.
769 */
770
771/* The Guest needs to tell the host what stack it expects traps to use. For
772 * native hardware, this is part of the Task State Segment mentioned above in
773 * lguest_load_tr_desc(), but to help hypervisors there's this special call.
774 *
775 * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data
776 * segment), the privilege level (we're privilege level 1, the Host is 0 and
777 * will not tolerate us trying to use that), the stack pointer, and the number
778 * of pages in the stack. */
485static void lguest_load_esp0(struct tss_struct *tss, 779static void lguest_load_esp0(struct tss_struct *tss,
486 struct thread_struct *thread) 780 struct thread_struct *thread)
487{ 781{
@@ -489,15 +783,31 @@ static void lguest_load_esp0(struct tss_struct *tss,
489 THREAD_SIZE/PAGE_SIZE); 783 THREAD_SIZE/PAGE_SIZE);
490} 784}
491 785
786/* Let's just say, I wouldn't do debugging under a Guest. */
492static void lguest_set_debugreg(int regno, unsigned long value) 787static void lguest_set_debugreg(int regno, unsigned long value)
493{ 788{
494 /* FIXME: Implement */ 789 /* FIXME: Implement */
495} 790}
496 791
792/* There are times when the kernel wants to make sure that no memory writes are
793 * caught in the cache (that they've all reached real hardware devices). This
794 * doesn't matter for the Guest which has virtual hardware.
795 *
796 * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush
797 * (clflush) instruction is available and the kernel uses that. Otherwise, it
798 * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction.
799 * Unlike clflush, wbinvd can only be run at privilege level 0. So we can
800 * ignore clflush, but replace wbinvd.
801 */
497static void lguest_wbinvd(void) 802static void lguest_wbinvd(void)
498{ 803{
499} 804}
500 805
806/* If the Guest expects to have an Advanced Programmable Interrupt Controller,
807 * we play dumb by ignoring writes and returning 0 for reads. So it's no
808 * longer Programmable nor Controlling anything, and I don't think 8 lines of
809 * code qualifies for Advanced. It will also never interrupt anything. It
810 * does, however, allow us to get through the Linux boot code. */
501#ifdef CONFIG_X86_LOCAL_APIC 811#ifdef CONFIG_X86_LOCAL_APIC
502static void lguest_apic_write(unsigned long reg, unsigned long v) 812static void lguest_apic_write(unsigned long reg, unsigned long v)
503{ 813{
@@ -509,19 +819,32 @@ static unsigned long lguest_apic_read(unsigned long reg)
509} 819}
510#endif 820#endif
511 821
822/* STOP! Until an interrupt comes in. */
512static void lguest_safe_halt(void) 823static void lguest_safe_halt(void)
513{ 824{
514 hcall(LHCALL_HALT, 0, 0, 0); 825 hcall(LHCALL_HALT, 0, 0, 0);
515} 826}
516 827
828/* Perhaps CRASH isn't the best name for this hypercall, but we use it to get a
829 * message out when we're crashing as well as elegant termination like powering
830 * off.
831 *
832 * Note that the Host always prefers that the Guest speak in physical addresses
833 * rather than virtual addresses, so we use __pa() here. */
517static void lguest_power_off(void) 834static void lguest_power_off(void)
518{ 835{
519 hcall(LHCALL_CRASH, __pa("Power down"), 0, 0); 836 hcall(LHCALL_CRASH, __pa("Power down"), 0, 0);
520} 837}
521 838
839/*
840 * Panicing.
841 *
842 * Don't. But if you did, this is what happens.
843 */
522static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p) 844static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p)
523{ 845{
524 hcall(LHCALL_CRASH, __pa(p), 0, 0); 846 hcall(LHCALL_CRASH, __pa(p), 0, 0);
847 /* The hcall won't return, but to keep gcc happy, we're "done". */
525 return NOTIFY_DONE; 848 return NOTIFY_DONE;
526} 849}
527 850
@@ -529,15 +852,45 @@ static struct notifier_block paniced = {
529 .notifier_call = lguest_panic 852 .notifier_call = lguest_panic
530}; 853};
531 854
855/* Setting up memory is fairly easy. */
532static __init char *lguest_memory_setup(void) 856static __init char *lguest_memory_setup(void)
533{ 857{
534 /* We do this here because lockcheck barfs if before start_kernel */ 858 /* We do this here and not earlier because lockcheck barfs if we do it
859 * before start_kernel() */
535 atomic_notifier_chain_register(&panic_notifier_list, &paniced); 860 atomic_notifier_chain_register(&panic_notifier_list, &paniced);
536 861
862 /* The Linux bootloader header contains an "e820" memory map: the
863 * Launcher populated the first entry with our memory limit. */
537 add_memory_region(E820_MAP->addr, E820_MAP->size, E820_MAP->type); 864 add_memory_region(E820_MAP->addr, E820_MAP->size, E820_MAP->type);
865
866 /* This string is for the boot messages. */
538 return "LGUEST"; 867 return "LGUEST";
539} 868}
540 869
870/*G:050
871 * Patching (Powerfully Placating Performance Pedants)
872 *
873 * We have already seen that "struct paravirt_ops" lets us replace simple
874 * native instructions with calls to the appropriate back end all throughout
875 * the kernel. This allows the same kernel to run as a Guest and as a native
876 * kernel, but it's slow because of all the indirect branches.
877 *
878 * Remember that David Wheeler quote about "Any problem in computer science can
879 * be solved with another layer of indirection"? The rest of that quote is
880 * "... But that usually will create another problem." This is the first of
881 * those problems.
882 *
883 * Our current solution is to allow the paravirt back end to optionally patch
884 * over the indirect calls to replace them with something more efficient. We
885 * patch the four most commonly called functions: disable interrupts, enable
886 * interrupts, restore interrupts and save interrupts. We usually have 10
887 * bytes to patch into: the Guest versions of these operations are small enough
888 * that we can fit comfortably.
889 *
890 * First we need assembly templates of each of the patchable Guest operations,
891 * and these are in lguest_asm.S. */
892
893/*G:060 We construct a table from the assembler templates: */
541static const struct lguest_insns 894static const struct lguest_insns
542{ 895{
543 const char *start, *end; 896 const char *start, *end;
@@ -547,35 +900,52 @@ static const struct lguest_insns
547 [PARAVIRT_PATCH(restore_fl)] = { lgstart_popf, lgend_popf }, 900 [PARAVIRT_PATCH(restore_fl)] = { lgstart_popf, lgend_popf },
548 [PARAVIRT_PATCH(save_fl)] = { lgstart_pushf, lgend_pushf }, 901 [PARAVIRT_PATCH(save_fl)] = { lgstart_pushf, lgend_pushf },
549}; 902};
903
904/* Now our patch routine is fairly simple (based on the native one in
905 * paravirt.c). If we have a replacement, we copy it in and return how much of
906 * the available space we used. */
550static unsigned lguest_patch(u8 type, u16 clobber, void *insns, unsigned len) 907static unsigned lguest_patch(u8 type, u16 clobber, void *insns, unsigned len)
551{ 908{
552 unsigned int insn_len; 909 unsigned int insn_len;
553 910
554 /* Don't touch it if we don't have a replacement */ 911 /* Don't do anything special if we don't have a replacement */
555 if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start) 912 if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start)
556 return paravirt_patch_default(type, clobber, insns, len); 913 return paravirt_patch_default(type, clobber, insns, len);
557 914
558 insn_len = lguest_insns[type].end - lguest_insns[type].start; 915 insn_len = lguest_insns[type].end - lguest_insns[type].start;
559 916
560 /* Similarly if we can't fit replacement. */ 917 /* Similarly if we can't fit replacement (shouldn't happen, but let's
918 * be thorough). */
561 if (len < insn_len) 919 if (len < insn_len)
562 return paravirt_patch_default(type, clobber, insns, len); 920 return paravirt_patch_default(type, clobber, insns, len);
563 921
922 /* Copy in our instructions. */
564 memcpy(insns, lguest_insns[type].start, insn_len); 923 memcpy(insns, lguest_insns[type].start, insn_len);
565 return insn_len; 924 return insn_len;
566} 925}
567 926
927/*G:030 Once we get to lguest_init(), we know we're a Guest. The paravirt_ops
928 * structure in the kernel provides a single point for (almost) every routine
929 * we have to override to avoid privileged instructions. */
568__init void lguest_init(void *boot) 930__init void lguest_init(void *boot)
569{ 931{
570 /* Copy boot parameters first. */ 932 /* Copy boot parameters first: the Launcher put the physical location
933 * in %esi, and head.S converted that to a virtual address and handed
934 * it to us. */
571 memcpy(&boot_params, boot, PARAM_SIZE); 935 memcpy(&boot_params, boot, PARAM_SIZE);
936 /* The boot parameters also tell us where the command-line is: save
937 * that, too. */
572 memcpy(boot_command_line, __va(boot_params.hdr.cmd_line_ptr), 938 memcpy(boot_command_line, __va(boot_params.hdr.cmd_line_ptr),
573 COMMAND_LINE_SIZE); 939 COMMAND_LINE_SIZE);
574 940
941 /* We're under lguest, paravirt is enabled, and we're running at
942 * privilege level 1, not 0 as normal. */
575 paravirt_ops.name = "lguest"; 943 paravirt_ops.name = "lguest";
576 paravirt_ops.paravirt_enabled = 1; 944 paravirt_ops.paravirt_enabled = 1;
577 paravirt_ops.kernel_rpl = 1; 945 paravirt_ops.kernel_rpl = 1;
578 946
947 /* We set up all the lguest overrides for sensitive operations. These
948 * are detailed with the operations themselves. */
579 paravirt_ops.save_fl = save_fl; 949 paravirt_ops.save_fl = save_fl;
580 paravirt_ops.restore_fl = restore_fl; 950 paravirt_ops.restore_fl = restore_fl;
581 paravirt_ops.irq_disable = irq_disable; 951 paravirt_ops.irq_disable = irq_disable;
@@ -619,20 +989,45 @@ __init void lguest_init(void *boot)
619 paravirt_ops.set_lazy_mode = lguest_lazy_mode; 989 paravirt_ops.set_lazy_mode = lguest_lazy_mode;
620 paravirt_ops.wbinvd = lguest_wbinvd; 990 paravirt_ops.wbinvd = lguest_wbinvd;
621 paravirt_ops.sched_clock = lguest_sched_clock; 991 paravirt_ops.sched_clock = lguest_sched_clock;
622 992 /* Now is a good time to look at the implementations of these functions
993 * before returning to the rest of lguest_init(). */
994
995 /*G:070 Now we've seen all the paravirt_ops, we return to
996 * lguest_init() where the rest of the fairly chaotic boot setup
997 * occurs.
998 *
999 * The Host expects our first hypercall to tell it where our "struct
1000 * lguest_data" is, so we do that first. */
623 hcall(LHCALL_LGUEST_INIT, __pa(&lguest_data), 0, 0); 1001 hcall(LHCALL_LGUEST_INIT, __pa(&lguest_data), 0, 0);
624 1002
625 /* We use top of mem for initial pagetables. */ 1003 /* The native boot code sets up initial page tables immediately after
1004 * the kernel itself, and sets init_pg_tables_end so they're not
1005 * clobbered. The Launcher places our initial pagetables somewhere at
1006 * the top of our physical memory, so we don't need extra space: set
1007 * init_pg_tables_end to the end of the kernel. */
626 init_pg_tables_end = __pa(pg0); 1008 init_pg_tables_end = __pa(pg0);
627 1009
1010 /* Load the %fs segment register (the per-cpu segment register) with
1011 * the normal data segment to get through booting. */
628 asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS) : "memory"); 1012 asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS) : "memory");
629 1013
1014 /* The Host uses the top of the Guest's virtual address space for the
1015 * Host<->Guest Switcher, and it tells us how much it needs in
1016 * lguest_data.reserve_mem, set up on the LGUEST_INIT hypercall. */
630 reserve_top_address(lguest_data.reserve_mem); 1017 reserve_top_address(lguest_data.reserve_mem);
631 1018
1019 /* If we don't initialize the lock dependency checker now, it crashes
1020 * paravirt_disable_iospace. */
632 lockdep_init(); 1021 lockdep_init();
633 1022
1023 /* The IDE code spends about 3 seconds probing for disks: if we reserve
1024 * all the I/O ports up front it can't get them and so doesn't probe.
1025 * Other device drivers are similar (but less severe). This cuts the
1026 * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. */
634 paravirt_disable_iospace(); 1027 paravirt_disable_iospace();
635 1028
1029 /* This is messy CPU setup stuff which the native boot code does before
1030 * start_kernel, so we have to do, too: */
636 cpu_detect(&new_cpu_data); 1031 cpu_detect(&new_cpu_data);
637 /* head.S usually sets up the first capability word, so do it here. */ 1032 /* head.S usually sets up the first capability word, so do it here. */
638 new_cpu_data.x86_capability[0] = cpuid_edx(1); 1033 new_cpu_data.x86_capability[0] = cpuid_edx(1);
@@ -643,14 +1038,27 @@ __init void lguest_init(void *boot)
643#ifdef CONFIG_X86_MCE 1038#ifdef CONFIG_X86_MCE
644 mce_disabled = 1; 1039 mce_disabled = 1;
645#endif 1040#endif
646
647#ifdef CONFIG_ACPI 1041#ifdef CONFIG_ACPI
648 acpi_disabled = 1; 1042 acpi_disabled = 1;
649 acpi_ht = 0; 1043 acpi_ht = 0;
650#endif 1044#endif
651 1045
1046 /* We set the perferred console to "hvc". This is the "hypervisor
1047 * virtual console" driver written by the PowerPC people, which we also
1048 * adapted for lguest's use. */
652 add_preferred_console("hvc", 0, NULL); 1049 add_preferred_console("hvc", 0, NULL);
653 1050
1051 /* Last of all, we set the power management poweroff hook to point to
1052 * the Guest routine to power off. */
654 pm_power_off = lguest_power_off; 1053 pm_power_off = lguest_power_off;
1054
1055 /* Now we're set up, call start_kernel() in init/main.c and we proceed
1056 * to boot as normal. It never returns. */
655 start_kernel(); 1057 start_kernel();
656} 1058}
1059/*
1060 * This marks the end of stage II of our journey, The Guest.
1061 *
1062 * It is now time for us to explore the nooks and crannies of the three Guest
1063 * devices and complete our understanding of the Guest in "make Drivers".
1064 */