~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/arch/x86/lguest/boot.c

Version: ~ [ linux-5.18 ] ~ [ linux-5.17.9 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.41 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.117 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.195 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.244 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.280 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.315 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*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."
  5  *
  6  * We keep things simple in two ways.  First, we start with a normal Linux
  7  * kernel and insert a module (lg.ko) which allows us to run other Linux
  8  * kernels the same way we'd run processes.  We call the first kernel the Host,
  9  * and the others the Guests.  The program which sets up and configures Guests
 10  * (such as the example in Documentation/virtual/lguest/lguest.c) is called the
 11  * Launcher.
 12  *
 13  * Secondly, we only run specially modified Guests, not normal kernels: setting
 14  * CONFIG_LGUEST_GUEST to "y" compiles this file into the kernel so it knows
 15  * how to be a Guest at boot time.  This means that you can use the same kernel
 16  * you boot normally (ie. as a Host) as a Guest.
 17  *
 18  * These Guests know that they cannot do privileged operations, such as disable
 19  * interrupts, and that they have to ask the Host to do such things explicitly.
 20  * This file consists of all the replacements for such low-level native
 21  * hardware operations: these special Guest versions call the Host.
 22  *
 23  * So how does the kernel know it's a Guest?  We'll see that later, but let's
 24  * just say that we end up here where we replace the native functions various
 25  * "paravirt" structures with our Guest versions, then boot like normal.
 26 :*/
 27 
 28 /*
 29  * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation.
 30  *
 31  * This program is free software; you can redistribute it and/or modify
 32  * it under the terms of the GNU General Public License as published by
 33  * the Free Software Foundation; either version 2 of the License, or
 34  * (at your option) any later version.
 35  *
 36  * This program is distributed in the hope that it will be useful, but
 37  * WITHOUT ANY WARRANTY; without even the implied warranty of
 38  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
 39  * NON INFRINGEMENT.  See the GNU General Public License for more
 40  * details.
 41  *
 42  * You should have received a copy of the GNU General Public License
 43  * along with this program; if not, write to the Free Software
 44  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
 45  */
 46 #include <linux/kernel.h>
 47 #include <linux/start_kernel.h>
 48 #include <linux/string.h>
 49 #include <linux/console.h>
 50 #include <linux/screen_info.h>
 51 #include <linux/irq.h>
 52 #include <linux/interrupt.h>
 53 #include <linux/clocksource.h>
 54 #include <linux/clockchips.h>
 55 #include <linux/lguest.h>
 56 #include <linux/lguest_launcher.h>
 57 #include <linux/virtio_console.h>
 58 #include <linux/pm.h>
 59 #include <linux/export.h>
 60 #include <asm/apic.h>
 61 #include <asm/lguest.h>
 62 #include <asm/paravirt.h>
 63 #include <asm/param.h>
 64 #include <asm/page.h>
 65 #include <asm/pgtable.h>
 66 #include <asm/desc.h>
 67 #include <asm/setup.h>
 68 #include <asm/e820.h>
 69 #include <asm/mce.h>
 70 #include <asm/io.h>
 71 #include <asm/i387.h>
 72 #include <asm/stackprotector.h>
 73 #include <asm/reboot.h>         /* for struct machine_ops */
 74 #include <asm/kvm_para.h>
 75 
 76 /*G:010
 77  * Welcome to the Guest!
 78  *
 79  * The Guest in our tale is a simple creature: identical to the Host but
 80  * behaving in simplified but equivalent ways.  In particular, the Guest is the
 81  * same kernel as the Host (or at least, built from the same source code).
 82 :*/
 83 
 84 struct lguest_data lguest_data = {
 85         .hcall_status = { [0 ... LHCALL_RING_SIZE-1] = 0xFF },
 86         .noirq_start = (u32)lguest_noirq_start,
 87         .noirq_end = (u32)lguest_noirq_end,
 88         .kernel_address = PAGE_OFFSET,
 89         .blocked_interrupts = { 1 }, /* Block timer interrupts */
 90         .syscall_vec = SYSCALL_VECTOR,
 91 };
 92 
 93 /*G:037
 94  * async_hcall() is pretty simple: I'm quite proud of it really.  We have a
 95  * ring buffer of stored hypercalls which the Host will run though next time we
 96  * do a normal hypercall.  Each entry in the ring has 5 slots for the hypercall
 97  * arguments, and a "hcall_status" word which is 0 if the call is ready to go,
 98  * and 255 once the Host has finished with it.
 99  *
100  * If we come around to a slot which hasn't been finished, then the table is
101  * full and we just make the hypercall directly.  This has the nice side
102  * effect of causing the Host to run all the stored calls in the ring buffer
103  * which empties it for next time!
104  */
105 static void async_hcall(unsigned long call, unsigned long arg1,
106                         unsigned long arg2, unsigned long arg3,
107                         unsigned long arg4)
108 {
109         /* Note: This code assumes we're uniprocessor. */
110         static unsigned int next_call;
111         unsigned long flags;
112 
113         /*
114          * Disable interrupts if not already disabled: we don't want an
115          * interrupt handler making a hypercall while we're already doing
116          * one!
117          */
118         local_irq_save(flags);
119         if (lguest_data.hcall_status[next_call] != 0xFF) {
120                 /* Table full, so do normal hcall which will flush table. */
121                 hcall(call, arg1, arg2, arg3, arg4);
122         } else {
123                 lguest_data.hcalls[next_call].arg0 = call;
124                 lguest_data.hcalls[next_call].arg1 = arg1;
125                 lguest_data.hcalls[next_call].arg2 = arg2;
126                 lguest_data.hcalls[next_call].arg3 = arg3;
127                 lguest_data.hcalls[next_call].arg4 = arg4;
128                 /* Arguments must all be written before we mark it to go */
129                 wmb();
130                 lguest_data.hcall_status[next_call] = 0;
131                 if (++next_call == LHCALL_RING_SIZE)
132                         next_call = 0;
133         }
134         local_irq_restore(flags);
135 }
136 
137 /*G:035
138  * Notice the lazy_hcall() above, rather than hcall().  This is our first real
139  * optimization trick!
140  *
141  * When lazy_mode is set, it means we're allowed to defer all hypercalls and do
142  * them as a batch when lazy_mode is eventually turned off.  Because hypercalls
143  * are reasonably expensive, batching them up makes sense.  For example, a
144  * large munmap might update dozens of page table entries: that code calls
145  * paravirt_enter_lazy_mmu(), does the dozen updates, then calls
146  * lguest_leave_lazy_mode().
147  *
148  * So, when we're in lazy mode, we call async_hcall() to store the call for
149  * future processing:
150  */
151 static void lazy_hcall1(unsigned long call, unsigned long arg1)
152 {
153         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
154                 hcall(call, arg1, 0, 0, 0);
155         else
156                 async_hcall(call, arg1, 0, 0, 0);
157 }
158 
159 /* You can imagine what lazy_hcall2, 3 and 4 look like. :*/
160 static void lazy_hcall2(unsigned long call,
161                         unsigned long arg1,
162                         unsigned long arg2)
163 {
164         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
165                 hcall(call, arg1, arg2, 0, 0);
166         else
167                 async_hcall(call, arg1, arg2, 0, 0);
168 }
169 
170 static void lazy_hcall3(unsigned long call,
171                         unsigned long arg1,
172                         unsigned long arg2,
173                         unsigned long arg3)
174 {
175         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
176                 hcall(call, arg1, arg2, arg3, 0);
177         else
178                 async_hcall(call, arg1, arg2, arg3, 0);
179 }
180 
181 #ifdef CONFIG_X86_PAE
182 static void lazy_hcall4(unsigned long call,
183                         unsigned long arg1,
184                         unsigned long arg2,
185                         unsigned long arg3,
186                         unsigned long arg4)
187 {
188         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
189                 hcall(call, arg1, arg2, arg3, arg4);
190         else
191                 async_hcall(call, arg1, arg2, arg3, arg4);
192 }
193 #endif
194 
195 /*G:036
196  * When lazy mode is turned off, we issue the do-nothing hypercall to
197  * flush any stored calls, and call the generic helper to reset the
198  * per-cpu lazy mode variable.
199  */
200 static void lguest_leave_lazy_mmu_mode(void)
201 {
202         hcall(LHCALL_FLUSH_ASYNC, 0, 0, 0, 0);
203         paravirt_leave_lazy_mmu();
204 }
205 
206 /*
207  * We also catch the end of context switch; we enter lazy mode for much of
208  * that too, so again we need to flush here.
209  *
210  * (Technically, this is lazy CPU mode, and normally we're in lazy MMU
211  * mode, but unlike Xen, lguest doesn't care about the difference).
212  */
213 static void lguest_end_context_switch(struct task_struct *next)
214 {
215         hcall(LHCALL_FLUSH_ASYNC, 0, 0, 0, 0);
216         paravirt_end_context_switch(next);
217 }
218 
219 /*G:032
220  * After that diversion we return to our first native-instruction
221  * replacements: four functions for interrupt control.
222  *
223  * The simplest way of implementing these would be to have "turn interrupts
224  * off" and "turn interrupts on" hypercalls.  Unfortunately, this is too slow:
225  * these are by far the most commonly called functions of those we override.
226  *
227  * So instead we keep an "irq_enabled" field inside our "struct lguest_data",
228  * which the Guest can update with a single instruction.  The Host knows to
229  * check there before it tries to deliver an interrupt.
230  */
231 
232 /*
233  * save_flags() is expected to return the processor state (ie. "flags").  The
234  * flags word contains all kind of stuff, but in practice Linux only cares
235  * about the interrupt flag.  Our "save_flags()" just returns that.
236  */
237 static unsigned long save_fl(void)
238 {
239         return lguest_data.irq_enabled;
240 }
241 
242 /* Interrupts go off... */
243 static void irq_disable(void)
244 {
245         lguest_data.irq_enabled = 0;
246 }
247 
248 /*
249  * Let's pause a moment.  Remember how I said these are called so often?
250  * Jeremy Fitzhardinge optimized them so hard early in 2009 that he had to
251  * break some rules.  In particular, these functions are assumed to save their
252  * own registers if they need to: normal C functions assume they can trash the
253  * eax register.  To use normal C functions, we use
254  * PV_CALLEE_SAVE_REGS_THUNK(), which pushes %eax onto the stack, calls the
255  * C function, then restores it.
256  */
257 PV_CALLEE_SAVE_REGS_THUNK(save_fl);
258 PV_CALLEE_SAVE_REGS_THUNK(irq_disable);
259 /*:*/
260 
261 /* These are in i386_head.S */
262 extern void lg_irq_enable(void);
263 extern void lg_restore_fl(unsigned long flags);
264 
265 /*M:003
266  * We could be more efficient in our checking of outstanding interrupts, rather
267  * than using a branch.  One way would be to put the "irq_enabled" field in a
268  * page by itself, and have the Host write-protect it when an interrupt comes
269  * in when irqs are disabled.  There will then be a page fault as soon as
270  * interrupts are re-enabled.
271  *
272  * A better method is to implement soft interrupt disable generally for x86:
273  * instead of disabling interrupts, we set a flag.  If an interrupt does come
274  * in, we then disable them for real.  This is uncommon, so we could simply use
275  * a hypercall for interrupt control and not worry about efficiency.
276 :*/
277 
278 /*G:034
279  * The Interrupt Descriptor Table (IDT).
280  *
281  * The IDT tells the processor what to do when an interrupt comes in.  Each
282  * entry in the table is a 64-bit descriptor: this holds the privilege level,
283  * address of the handler, and... well, who cares?  The Guest just asks the
284  * Host to make the change anyway, because the Host controls the real IDT.
285  */
286 static void lguest_write_idt_entry(gate_desc *dt,
287                                    int entrynum, const gate_desc *g)
288 {
289         /*
290          * The gate_desc structure is 8 bytes long: we hand it to the Host in
291          * two 32-bit chunks.  The whole 32-bit kernel used to hand descriptors
292          * around like this; typesafety wasn't a big concern in Linux's early
293          * years.
294          */
295         u32 *desc = (u32 *)g;
296         /* Keep the local copy up to date. */
297         native_write_idt_entry(dt, entrynum, g);
298         /* Tell Host about this new entry. */
299         hcall(LHCALL_LOAD_IDT_ENTRY, entrynum, desc[0], desc[1], 0);
300 }
301 
302 /*
303  * Changing to a different IDT is very rare: we keep the IDT up-to-date every
304  * time it is written, so we can simply loop through all entries and tell the
305  * Host about them.
306  */
307 static void lguest_load_idt(const struct desc_ptr *desc)
308 {
309         unsigned int i;
310         struct desc_struct *idt = (void *)desc->address;
311 
312         for (i = 0; i < (desc->size+1)/8; i++)
313                 hcall(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b, 0);
314 }
315 
316 /*
317  * The Global Descriptor Table.
318  *
319  * The Intel architecture defines another table, called the Global Descriptor
320  * Table (GDT).  You tell the CPU where it is (and its size) using the "lgdt"
321  * instruction, and then several other instructions refer to entries in the
322  * table.  There are three entries which the Switcher needs, so the Host simply
323  * controls the entire thing and the Guest asks it to make changes using the
324  * LOAD_GDT hypercall.
325  *
326  * This is the exactly like the IDT code.
327  */
328 static void lguest_load_gdt(const struct desc_ptr *desc)
329 {
330         unsigned int i;
331         struct desc_struct *gdt = (void *)desc->address;
332 
333         for (i = 0; i < (desc->size+1)/8; i++)
334                 hcall(LHCALL_LOAD_GDT_ENTRY, i, gdt[i].a, gdt[i].b, 0);
335 }
336 
337 /*
338  * For a single GDT entry which changes, we simply change our copy and
339  * then tell the host about it.
340  */
341 static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum,
342                                    const void *desc, int type)
343 {
344         native_write_gdt_entry(dt, entrynum, desc, type);
345         /* Tell Host about this new entry. */
346         hcall(LHCALL_LOAD_GDT_ENTRY, entrynum,
347               dt[entrynum].a, dt[entrynum].b, 0);
348 }
349 
350 /*
351  * There are three "thread local storage" GDT entries which change
352  * on every context switch (these three entries are how glibc implements
353  * __thread variables).  As an optimization, we have a hypercall
354  * specifically for this case.
355  *
356  * Wouldn't it be nicer to have a general LOAD_GDT_ENTRIES hypercall
357  * which took a range of entries?
358  */
359 static void lguest_load_tls(struct thread_struct *t, unsigned int cpu)
360 {
361         /*
362          * There's one problem which normal hardware doesn't have: the Host
363          * can't handle us removing entries we're currently using.  So we clear
364          * the GS register here: if it's needed it'll be reloaded anyway.
365          */
366         lazy_load_gs(0);
367         lazy_hcall2(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu);
368 }
369 
370 /*G:038
371  * That's enough excitement for now, back to ploughing through each of the
372  * different pv_ops structures (we're about 1/3 of the way through).
373  *
374  * This is the Local Descriptor Table, another weird Intel thingy.  Linux only
375  * uses this for some strange applications like Wine.  We don't do anything
376  * here, so they'll get an informative and friendly Segmentation Fault.
377  */
378 static void lguest_set_ldt(const void *addr, unsigned entries)
379 {
380 }
381 
382 /*
383  * This loads a GDT entry into the "Task Register": that entry points to a
384  * structure called the Task State Segment.  Some comments scattered though the
385  * kernel code indicate that this used for task switching in ages past, along
386  * with blood sacrifice and astrology.
387  *
388  * Now there's nothing interesting in here that we don't get told elsewhere.
389  * But the native version uses the "ltr" instruction, which makes the Host
390  * complain to the Guest about a Segmentation Fault and it'll oops.  So we
391  * override the native version with a do-nothing version.
392  */
393 static void lguest_load_tr_desc(void)
394 {
395 }
396 
397 /*
398  * The "cpuid" instruction is a way of querying both the CPU identity
399  * (manufacturer, model, etc) and its features.  It was introduced before the
400  * Pentium in 1993 and keeps getting extended by both Intel, AMD and others.
401  * As you might imagine, after a decade and a half this treatment, it is now a
402  * giant ball of hair.  Its entry in the current Intel manual runs to 28 pages.
403  *
404  * This instruction even it has its own Wikipedia entry.  The Wikipedia entry
405  * has been translated into 6 languages.  I am not making this up!
406  *
407  * We could get funky here and identify ourselves as "GenuineLguest", but
408  * instead we just use the real "cpuid" instruction.  Then I pretty much turned
409  * off feature bits until the Guest booted.  (Don't say that: you'll damage
410  * lguest sales!)  Shut up, inner voice!  (Hey, just pointing out that this is
411  * hardly future proof.)  No one's listening!  They don't like you anyway,
412  * parenthetic weirdo!
413  *
414  * Replacing the cpuid so we can turn features off is great for the kernel, but
415  * anyone (including userspace) can just use the raw "cpuid" instruction and
416  * the Host won't even notice since it isn't privileged.  So we try not to get
417  * too worked up about it.
418  */
419 static void lguest_cpuid(unsigned int *ax, unsigned int *bx,
420                          unsigned int *cx, unsigned int *dx)
421 {
422         int function = *ax;
423 
424         native_cpuid(ax, bx, cx, dx);
425         switch (function) {
426         /*
427          * CPUID 0 gives the highest legal CPUID number (and the ID string).
428          * We futureproof our code a little by sticking to known CPUID values.
429          */
430         case 0:
431                 if (*ax > 5)
432                         *ax = 5;
433                 break;
434 
435         /*
436          * CPUID 1 is a basic feature request.
437          *
438          * CX: we only allow kernel to see SSE3, CMPXCHG16B and SSSE3
439          * DX: SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU and PAE.
440          */
441         case 1:
442                 *cx &= 0x00002201;
443                 *dx &= 0x07808151;
444                 /*
445                  * The Host can do a nice optimization if it knows that the
446                  * kernel mappings (addresses above 0xC0000000 or whatever
447                  * PAGE_OFFSET is set to) haven't changed.  But Linux calls
448                  * flush_tlb_user() for both user and kernel mappings unless
449                  * the Page Global Enable (PGE) feature bit is set.
450                  */
451                 *dx |= 0x00002000;
452                 /*
453                  * We also lie, and say we're family id 5.  6 or greater
454                  * leads to a rdmsr in early_init_intel which we can't handle.
455                  * Family ID is returned as bits 8-12 in ax.
456                  */
457                 *ax &= 0xFFFFF0FF;
458                 *ax |= 0x00000500;
459                 break;
460 
461         /*
462          * This is used to detect if we're running under KVM.  We might be,
463          * but that's a Host matter, not us.  So say we're not.
464          */
465         case KVM_CPUID_SIGNATURE:
466                 *bx = *cx = *dx = 0;
467                 break;
468 
469         /*
470          * 0x80000000 returns the highest Extended Function, so we futureproof
471          * like we do above by limiting it to known fields.
472          */
473         case 0x80000000:
474                 if (*ax > 0x80000008)
475                         *ax = 0x80000008;
476                 break;
477 
478         /*
479          * PAE systems can mark pages as non-executable.  Linux calls this the
480          * NX bit.  Intel calls it XD (eXecute Disable), AMD EVP (Enhanced
481          * Virus Protection).  We just switch it off here, since we don't
482          * support it.
483          */
484         case 0x80000001:
485                 *dx &= ~(1 << 20);
486                 break;
487         }
488 }
489 
490 /*
491  * Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4.
492  * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother
493  * it.  The Host needs to know when the Guest wants to change them, so we have
494  * a whole series of functions like read_cr0() and write_cr0().
495  *
496  * We start with cr0.  cr0 allows you to turn on and off all kinds of basic
497  * features, but Linux only really cares about one: the horrifically-named Task
498  * Switched (TS) bit at bit 3 (ie. 8)
499  *
500  * What does the TS bit do?  Well, it causes the CPU to trap (interrupt 7) if
501  * the floating point unit is used.  Which allows us to restore FPU state
502  * lazily after a task switch, and Linux uses that gratefully, but wouldn't a
503  * name like "FPUTRAP bit" be a little less cryptic?
504  *
505  * We store cr0 locally because the Host never changes it.  The Guest sometimes
506  * wants to read it and we'd prefer not to bother the Host unnecessarily.
507  */
508 static unsigned long current_cr0;
509 static void lguest_write_cr0(unsigned long val)
510 {
511         lazy_hcall1(LHCALL_TS, val & X86_CR0_TS);
512         current_cr0 = val;
513 }
514 
515 static unsigned long lguest_read_cr0(void)
516 {
517         return current_cr0;
518 }
519 
520 /*
521  * Intel provided a special instruction to clear the TS bit for people too cool
522  * to use write_cr0() to do it.  This "clts" instruction is faster, because all
523  * the vowels have been optimized out.
524  */
525 static void lguest_clts(void)
526 {
527         lazy_hcall1(LHCALL_TS, 0);
528         current_cr0 &= ~X86_CR0_TS;
529 }
530 
531 /*
532  * cr2 is the virtual address of the last page fault, which the Guest only ever
533  * reads.  The Host kindly writes this into our "struct lguest_data", so we
534  * just read it out of there.
535  */
536 static unsigned long lguest_read_cr2(void)
537 {
538         return lguest_data.cr2;
539 }
540 
541 /* See lguest_set_pte() below. */
542 static bool cr3_changed = false;
543 static unsigned long current_cr3;
544 
545 /*
546  * cr3 is the current toplevel pagetable page: the principle is the same as
547  * cr0.  Keep a local copy, and tell the Host when it changes.
548  */
549 static void lguest_write_cr3(unsigned long cr3)
550 {
551         lazy_hcall1(LHCALL_NEW_PGTABLE, cr3);
552         current_cr3 = cr3;
553 
554         /* These two page tables are simple, linear, and used during boot */
555         if (cr3 != __pa_symbol(swapper_pg_dir) &&
556             cr3 != __pa_symbol(initial_page_table))
557                 cr3_changed = true;
558 }
559 
560 static unsigned long lguest_read_cr3(void)
561 {
562         return current_cr3;
563 }
564 
565 /* cr4 is used to enable and disable PGE, but we don't care. */
566 static unsigned long lguest_read_cr4(void)
567 {
568         return 0;
569 }
570 
571 static void lguest_write_cr4(unsigned long val)
572 {
573 }
574 
575 /*
576  * Page Table Handling.
577  *
578  * Now would be a good time to take a rest and grab a coffee or similarly
579  * relaxing stimulant.  The easy parts are behind us, and the trek gradually
580  * winds uphill from here.
581  *
582  * Quick refresher: memory is divided into "pages" of 4096 bytes each.  The CPU
583  * maps virtual addresses to physical addresses using "page tables".  We could
584  * use one huge index of 1 million entries: each address is 4 bytes, so that's
585  * 1024 pages just to hold the page tables.   But since most virtual addresses
586  * are unused, we use a two level index which saves space.  The cr3 register
587  * contains the physical address of the top level "page directory" page, which
588  * contains physical addresses of up to 1024 second-level pages.  Each of these
589  * second level pages contains up to 1024 physical addresses of actual pages,
590  * or Page Table Entries (PTEs).
591  *
592  * Here's a diagram, where arrows indicate physical addresses:
593  *
594  * cr3 ---> +---------+
595  *          |      --------->+---------+
596  *          |         |      | PADDR1  |
597  *        Mid-level   |      | PADDR2  |
598  *        (PMD) page  |      |         |
599  *          |         |    Lower-level |
600  *          |         |    (PTE) page  |
601  *          |         |      |         |
602  *            ....               ....
603  *
604  * So to convert a virtual address to a physical address, we look up the top
605  * level, which points us to the second level, which gives us the physical
606  * address of that page.  If the top level entry was not present, or the second
607  * level entry was not present, then the virtual address is invalid (we
608  * say "the page was not mapped").
609  *
610  * Put another way, a 32-bit virtual address is divided up like so:
611  *
612  *  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
613  * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>|
614  *    Index into top     Index into second      Offset within page
615  *  page directory page    pagetable page
616  *
617  * Now, unfortunately, this isn't the whole story: Intel added Physical Address
618  * Extension (PAE) to allow 32 bit systems to use 64GB of memory (ie. 36 bits).
619  * These are held in 64-bit page table entries, so we can now only fit 512
620  * entries in a page, and the neat three-level tree breaks down.
621  *
622  * The result is a four level page table:
623  *
624  * cr3 --> [ 4 Upper  ]
625  *         [   Level  ]
626  *         [  Entries ]
627  *         [(PUD Page)]---> +---------+
628  *                          |      --------->+---------+
629  *                          |         |      | PADDR1  |
630  *                        Mid-level   |      | PADDR2  |
631  *                        (PMD) page  |      |         |
632  *                          |         |    Lower-level |
633  *                          |         |    (PTE) page  |
634  *                          |         |      |         |
635  *                            ....               ....
636  *
637  *
638  * And the virtual address is decoded as:
639  *
640  *         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
641  *      |<-2->|<--- 9 bits ---->|<---- 9 bits --->|<------ 12 bits ------>|
642  * Index into    Index into mid    Index into lower    Offset within page
643  * top entries   directory page     pagetable page
644  *
645  * It's too hard to switch between these two formats at runtime, so Linux only
646  * supports one or the other depending on whether CONFIG_X86_PAE is set.  Many
647  * distributions turn it on, and not just for people with silly amounts of
648  * memory: the larger PTE entries allow room for the NX bit, which lets the
649  * kernel disable execution of pages and increase security.
650  *
651  * This was a problem for lguest, which couldn't run on these distributions;
652  * then Matias Zabaljauregui figured it all out and implemented it, and only a
653  * handful of puppies were crushed in the process!
654  *
655  * Back to our point: the kernel spends a lot of time changing both the
656  * top-level page directory and lower-level pagetable pages.  The Guest doesn't
657  * know physical addresses, so while it maintains these page tables exactly
658  * like normal, it also needs to keep the Host informed whenever it makes a
659  * change: the Host will create the real page tables based on the Guests'.
660  */
661 
662 /*
663  * The Guest calls this after it has set a second-level entry (pte), ie. to map
664  * a page into a process' address space.  We tell the Host the toplevel and
665  * address this corresponds to.  The Guest uses one pagetable per process, so
666  * we need to tell the Host which one we're changing (mm->pgd).
667  */
668 static void lguest_pte_update(struct mm_struct *mm, unsigned long addr,
669                                pte_t *ptep)
670 {
671 #ifdef CONFIG_X86_PAE
672         /* PAE needs to hand a 64 bit page table entry, so it uses two args. */
673         lazy_hcall4(LHCALL_SET_PTE, __pa(mm->pgd), addr,
674                     ptep->pte_low, ptep->pte_high);
675 #else
676         lazy_hcall3(LHCALL_SET_PTE, __pa(mm->pgd), addr, ptep->pte_low);
677 #endif
678 }
679 
680 /* This is the "set and update" combo-meal-deal version. */
681 static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr,
682                               pte_t *ptep, pte_t pteval)
683 {
684         native_set_pte(ptep, pteval);
685         lguest_pte_update(mm, addr, ptep);
686 }
687 
688 /*
689  * The Guest calls lguest_set_pud to set a top-level entry and lguest_set_pmd
690  * to set a middle-level entry when PAE is activated.
691  *
692  * Again, we set the entry then tell the Host which page we changed,
693  * and the index of the entry we changed.
694  */
695 #ifdef CONFIG_X86_PAE
696 static void lguest_set_pud(pud_t *pudp, pud_t pudval)
697 {
698         native_set_pud(pudp, pudval);
699 
700         /* 32 bytes aligned pdpt address and the index. */
701         lazy_hcall2(LHCALL_SET_PGD, __pa(pudp) & 0xFFFFFFE0,
702                    (__pa(pudp) & 0x1F) / sizeof(pud_t));
703 }
704 
705 static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
706 {
707         native_set_pmd(pmdp, pmdval);
708         lazy_hcall2(LHCALL_SET_PMD, __pa(pmdp) & PAGE_MASK,
709                    (__pa(pmdp) & (PAGE_SIZE - 1)) / sizeof(pmd_t));
710 }
711 #else
712 
713 /* The Guest calls lguest_set_pmd to set a top-level entry when !PAE. */
714 static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
715 {
716         native_set_pmd(pmdp, pmdval);
717         lazy_hcall2(LHCALL_SET_PGD, __pa(pmdp) & PAGE_MASK,
718                    (__pa(pmdp) & (PAGE_SIZE - 1)) / sizeof(pmd_t));
719 }
720 #endif
721 
722 /*
723  * There are a couple of legacy places where the kernel sets a PTE, but we
724  * don't know the top level any more.  This is useless for us, since we don't
725  * know which pagetable is changing or what address, so we just tell the Host
726  * to forget all of them.  Fortunately, this is very rare.
727  *
728  * ... except in early boot when the kernel sets up the initial pagetables,
729  * which makes booting astonishingly slow: 48 seconds!  So we don't even tell
730  * the Host anything changed until we've done the first real page table switch,
731  * which brings boot back to 4.3 seconds.
732  */
733 static void lguest_set_pte(pte_t *ptep, pte_t pteval)
734 {
735         native_set_pte(ptep, pteval);
736         if (cr3_changed)
737                 lazy_hcall1(LHCALL_FLUSH_TLB, 1);
738 }
739 
740 #ifdef CONFIG_X86_PAE
741 /*
742  * With 64-bit PTE values, we need to be careful setting them: if we set 32
743  * bits at a time, the hardware could see a weird half-set entry.  These
744  * versions ensure we update all 64 bits at once.
745  */
746 static void lguest_set_pte_atomic(pte_t *ptep, pte_t pte)
747 {
748         native_set_pte_atomic(ptep, pte);
749         if (cr3_changed)
750                 lazy_hcall1(LHCALL_FLUSH_TLB, 1);
751 }
752 
753 static void lguest_pte_clear(struct mm_struct *mm, unsigned long addr,
754                              pte_t *ptep)
755 {
756         native_pte_clear(mm, addr, ptep);
757         lguest_pte_update(mm, addr, ptep);
758 }
759 
760 static void lguest_pmd_clear(pmd_t *pmdp)
761 {
762         lguest_set_pmd(pmdp, __pmd(0));
763 }
764 #endif
765 
766 /*
767  * Unfortunately for Lguest, the pv_mmu_ops for page tables were based on
768  * native page table operations.  On native hardware you can set a new page
769  * table entry whenever you want, but if you want to remove one you have to do
770  * a TLB flush (a TLB is a little cache of page table entries kept by the CPU).
771  *
772  * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only
773  * called when a valid entry is written, not when it's removed (ie. marked not
774  * present).  Instead, this is where we come when the Guest wants to remove a
775  * page table entry: we tell the Host to set that entry to 0 (ie. the present
776  * bit is zero).
777  */
778 static void lguest_flush_tlb_single(unsigned long addr)
779 {
780         /* Simply set it to zero: if it was not, it will fault back in. */
781         lazy_hcall3(LHCALL_SET_PTE, current_cr3, addr, 0);
782 }
783 
784 /*
785  * This is what happens after the Guest has removed a large number of entries.
786  * This tells the Host that any of the page table entries for userspace might
787  * have changed, ie. virtual addresses below PAGE_OFFSET.
788  */
789 static void lguest_flush_tlb_user(void)
790 {
791         lazy_hcall1(LHCALL_FLUSH_TLB, 0);
792 }
793 
794 /*
795  * This is called when the kernel page tables have changed.  That's not very
796  * common (unless the Guest is using highmem, which makes the Guest extremely
797  * slow), so it's worth separating this from the user flushing above.
798  */
799 static void lguest_flush_tlb_kernel(void)
800 {
801         lazy_hcall1(LHCALL_FLUSH_TLB, 1);
802 }
803 
804 /*
805  * The Unadvanced Programmable Interrupt Controller.
806  *
807  * This is an attempt to implement the simplest possible interrupt controller.
808  * I spent some time looking though routines like set_irq_chip_and_handler,
809  * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and
810  * I *think* this is as simple as it gets.
811  *
812  * We can tell the Host what interrupts we want blocked ready for using the
813  * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as
814  * simple as setting a bit.  We don't actually "ack" interrupts as such, we
815  * just mask and unmask them.  I wonder if we should be cleverer?
816  */
817 static void disable_lguest_irq(struct irq_data *data)
818 {
819         set_bit(data->irq, lguest_data.blocked_interrupts);
820 }
821 
822 static void enable_lguest_irq(struct irq_data *data)
823 {
824         clear_bit(data->irq, lguest_data.blocked_interrupts);
825 }
826 
827 /* This structure describes the lguest IRQ controller. */
828 static struct irq_chip lguest_irq_controller = {
829         .name           = "lguest",
830         .irq_mask       = disable_lguest_irq,
831         .irq_mask_ack   = disable_lguest_irq,
832         .irq_unmask     = enable_lguest_irq,
833 };
834 
835 /*
836  * This sets up the Interrupt Descriptor Table (IDT) entry for each hardware
837  * interrupt (except 128, which is used for system calls), and then tells the
838  * Linux infrastructure that each interrupt is controlled by our level-based
839  * lguest interrupt controller.
840  */
841 static void __init lguest_init_IRQ(void)
842 {
843         unsigned int i;
844 
845         for (i = FIRST_EXTERNAL_VECTOR; i < NR_VECTORS; i++) {
846                 /* Some systems map "vectors" to interrupts weirdly.  Not us! */
847                 __this_cpu_write(vector_irq[i], i - FIRST_EXTERNAL_VECTOR);
848                 if (i != SYSCALL_VECTOR)
849                         set_intr_gate(i, interrupt[i - FIRST_EXTERNAL_VECTOR]);
850         }
851 
852         /*
853          * This call is required to set up for 4k stacks, where we have
854          * separate stacks for hard and soft interrupts.
855          */
856         irq_ctx_init(smp_processor_id());
857 }
858 
859 /*
860  * Interrupt descriptors are allocated as-needed, but low-numbered ones are
861  * reserved by the generic x86 code.  So we ignore irq_alloc_desc_at if it
862  * tells us the irq is already used: other errors (ie. ENOMEM) we take
863  * seriously.
864  */
865 int lguest_setup_irq(unsigned int irq)
866 {
867         int err;
868 
869         /* Returns -ve error or vector number. */
870         err = irq_alloc_desc_at(irq, 0);
871         if (err < 0 && err != -EEXIST)
872                 return err;
873 
874         irq_set_chip_and_handler_name(irq, &lguest_irq_controller,
875                                       handle_level_irq, "level");
876         return 0;
877 }
878 
879 /*
880  * Time.
881  *
882  * It would be far better for everyone if the Guest had its own clock, but
883  * until then the Host gives us the time on every interrupt.
884  */
885 static unsigned long lguest_get_wallclock(void)
886 {
887         return lguest_data.time.tv_sec;
888 }
889 
890 /*
891  * The TSC is an Intel thing called the Time Stamp Counter.  The Host tells us
892  * what speed it runs at, or 0 if it's unusable as a reliable clock source.
893  * This matches what we want here: if we return 0 from this function, the x86
894  * TSC clock will give up and not register itself.
895  */
896 static unsigned long lguest_tsc_khz(void)
897 {
898         return lguest_data.tsc_khz;
899 }
900 
901 /*
902  * If we can't use the TSC, the kernel falls back to our lower-priority
903  * "lguest_clock", where we read the time value given to us by the Host.
904  */
905 static cycle_t lguest_clock_read(struct clocksource *cs)
906 {
907         unsigned long sec, nsec;
908 
909         /*
910          * Since the time is in two parts (seconds and nanoseconds), we risk
911          * reading it just as it's changing from 99 & 0.999999999 to 100 and 0,
912          * and getting 99 and 0.  As Linux tends to come apart under the stress
913          * of time travel, we must be careful:
914          */
915         do {
916                 /* First we read the seconds part. */
917                 sec = lguest_data.time.tv_sec;
918                 /*
919                  * This read memory barrier tells the compiler and the CPU that
920                  * this can't be reordered: we have to complete the above
921                  * before going on.
922                  */
923                 rmb();
924                 /* Now we read the nanoseconds part. */
925                 nsec = lguest_data.time.tv_nsec;
926                 /* Make sure we've done that. */
927                 rmb();
928                 /* Now if the seconds part has changed, try again. */
929         } while (unlikely(lguest_data.time.tv_sec != sec));
930 
931         /* Our lguest clock is in real nanoseconds. */
932         return sec*1000000000ULL + nsec;
933 }
934 
935 /* This is the fallback clocksource: lower priority than the TSC clocksource. */
936 static struct clocksource lguest_clock = {
937         .name           = "lguest",
938         .rating         = 200,
939         .read           = lguest_clock_read,
940         .mask           = CLOCKSOURCE_MASK(64),
941         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
942 };
943 
944 /*
945  * We also need a "struct clock_event_device": Linux asks us to set it to go
946  * off some time in the future.  Actually, James Morris figured all this out, I
947  * just applied the patch.
948  */
949 static int lguest_clockevent_set_next_event(unsigned long delta,
950                                            struct clock_event_device *evt)
951 {
952         /* FIXME: I don't think this can ever happen, but James tells me he had
953          * to put this code in.  Maybe we should remove it now.  Anyone? */
954         if (delta < LG_CLOCK_MIN_DELTA) {
955                 if (printk_ratelimit())
956                         printk(KERN_DEBUG "%s: small delta %lu ns\n",
957                                __func__, delta);
958                 return -ETIME;
959         }
960 
961         /* Please wake us this far in the future. */
962         hcall(LHCALL_SET_CLOCKEVENT, delta, 0, 0, 0);
963         return 0;
964 }
965 
966 static void lguest_clockevent_set_mode(enum clock_event_mode mode,
967                                       struct clock_event_device *evt)
968 {
969         switch (mode) {
970         case CLOCK_EVT_MODE_UNUSED:
971         case CLOCK_EVT_MODE_SHUTDOWN:
972                 /* A 0 argument shuts the clock down. */
973                 hcall(LHCALL_SET_CLOCKEVENT, 0, 0, 0, 0);
974                 break;
975         case CLOCK_EVT_MODE_ONESHOT:
976                 /* This is what we expect. */
977                 break;
978         case CLOCK_EVT_MODE_PERIODIC:
979                 BUG();
980         case CLOCK_EVT_MODE_RESUME:
981                 break;
982         }
983 }
984 
985 /* This describes our primitive timer chip. */
986 static struct clock_event_device lguest_clockevent = {
987         .name                   = "lguest",
988         .features               = CLOCK_EVT_FEAT_ONESHOT,
989         .set_next_event         = lguest_clockevent_set_next_event,
990         .set_mode               = lguest_clockevent_set_mode,
991         .rating                 = INT_MAX,
992         .mult                   = 1,
993         .shift                  = 0,
994         .min_delta_ns           = LG_CLOCK_MIN_DELTA,
995         .max_delta_ns           = LG_CLOCK_MAX_DELTA,
996 };
997 
998 /*
999  * This is the Guest timer interrupt handler (hardware interrupt 0).  We just
1000  * call the clockevent infrastructure and it does whatever needs doing.
1001  */
1002 static void lguest_time_irq(unsigned int irq, struct irq_desc *desc)
1003 {
1004         unsigned long flags;
1005 
1006         /* Don't interrupt us while this is running. */
1007         local_irq_save(flags);
1008         lguest_clockevent.event_handler(&lguest_clockevent);
1009         local_irq_restore(flags);
1010 }
1011 
1012 /*
1013  * At some point in the boot process, we get asked to set up our timing
1014  * infrastructure.  The kernel doesn't expect timer interrupts before this, but
1015  * we cleverly initialized the "blocked_interrupts" field of "struct
1016  * lguest_data" so that timer interrupts were blocked until now.
1017  */
1018 static void lguest_time_init(void)
1019 {
1020         /* Set up the timer interrupt (0) to go to our simple timer routine */
1021         lguest_setup_irq(0);
1022         irq_set_handler(0, lguest_time_irq);
1023 
1024         clocksource_register_hz(&lguest_clock, NSEC_PER_SEC);
1025 
1026         /* We can't set cpumask in the initializer: damn C limitations!  Set it
1027          * here and register our timer device. */
1028         lguest_clockevent.cpumask = cpumask_of(0);
1029         clockevents_register_device(&lguest_clockevent);
1030 
1031         /* Finally, we unblock the timer interrupt. */
1032         clear_bit(0, lguest_data.blocked_interrupts);
1033 }
1034 
1035 /*
1036  * Miscellaneous bits and pieces.
1037  *
1038  * Here is an oddball collection of functions which the Guest needs for things
1039  * to work.  They're pretty simple.
1040  */
1041 
1042 /*
1043  * The Guest needs to tell the Host what stack it expects traps to use.  For
1044  * native hardware, this is part of the Task State Segment mentioned above in
1045  * lguest_load_tr_desc(), but to help hypervisors there's this special call.
1046  *
1047  * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data
1048  * segment), the privilege level (we're privilege level 1, the Host is 0 and
1049  * will not tolerate us trying to use that), the stack pointer, and the number
1050  * of pages in the stack.
1051  */
1052 static void lguest_load_sp0(struct tss_struct *tss,
1053                             struct thread_struct *thread)
1054 {
1055         lazy_hcall3(LHCALL_SET_STACK, __KERNEL_DS | 0x1, thread->sp0,
1056                    THREAD_SIZE / PAGE_SIZE);
1057 }
1058 
1059 /* Let's just say, I wouldn't do debugging under a Guest. */
1060 static void lguest_set_debugreg(int regno, unsigned long value)
1061 {
1062         /* FIXME: Implement */
1063 }
1064 
1065 /*
1066  * There are times when the kernel wants to make sure that no memory writes are
1067  * caught in the cache (that they've all reached real hardware devices).  This
1068  * doesn't matter for the Guest which has virtual hardware.
1069  *
1070  * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush
1071  * (clflush) instruction is available and the kernel uses that.  Otherwise, it
1072  * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction.
1073  * Unlike clflush, wbinvd can only be run at privilege level 0.  So we can
1074  * ignore clflush, but replace wbinvd.
1075  */
1076 static void lguest_wbinvd(void)
1077 {
1078 }
1079 
1080 /*
1081  * If the Guest expects to have an Advanced Programmable Interrupt Controller,
1082  * we play dumb by ignoring writes and returning 0 for reads.  So it's no
1083  * longer Programmable nor Controlling anything, and I don't think 8 lines of
1084  * code qualifies for Advanced.  It will also never interrupt anything.  It
1085  * does, however, allow us to get through the Linux boot code.
1086  */
1087 #ifdef CONFIG_X86_LOCAL_APIC
1088 static void lguest_apic_write(u32 reg, u32 v)
1089 {
1090 }
1091 
1092 static u32 lguest_apic_read(u32 reg)
1093 {
1094         return 0;
1095 }
1096 
1097 static u64 lguest_apic_icr_read(void)
1098 {
1099         return 0;
1100 }
1101 
1102 static void lguest_apic_icr_write(u32 low, u32 id)
1103 {
1104         /* Warn to see if there's any stray references */
1105         WARN_ON(1);
1106 }
1107 
1108 static void lguest_apic_wait_icr_idle(void)
1109 {
1110         return;
1111 }
1112 
1113 static u32 lguest_apic_safe_wait_icr_idle(void)
1114 {
1115         return 0;
1116 }
1117 
1118 static void set_lguest_basic_apic_ops(void)
1119 {
1120         apic->read = lguest_apic_read;
1121         apic->write = lguest_apic_write;
1122         apic->icr_read = lguest_apic_icr_read;
1123         apic->icr_write = lguest_apic_icr_write;
1124         apic->wait_icr_idle = lguest_apic_wait_icr_idle;
1125         apic->safe_wait_icr_idle = lguest_apic_safe_wait_icr_idle;
1126 };
1127 #endif
1128 
1129 /* STOP!  Until an interrupt comes in. */
1130 static void lguest_safe_halt(void)
1131 {
1132         hcall(LHCALL_HALT, 0, 0, 0, 0);
1133 }
1134 
1135 /*
1136  * The SHUTDOWN hypercall takes a string to describe what's happening, and
1137  * an argument which says whether this to restart (reboot) the Guest or not.
1138  *
1139  * Note that the Host always prefers that the Guest speak in physical addresses
1140  * rather than virtual addresses, so we use __pa() here.
1141  */
1142 static void lguest_power_off(void)
1143 {
1144         hcall(LHCALL_SHUTDOWN, __pa("Power down"),
1145               LGUEST_SHUTDOWN_POWEROFF, 0, 0);
1146 }
1147 
1148 /*
1149  * Panicing.
1150  *
1151  * Don't.  But if you did, this is what happens.
1152  */
1153 static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p)
1154 {
1155         hcall(LHCALL_SHUTDOWN, __pa(p), LGUEST_SHUTDOWN_POWEROFF, 0, 0);
1156         /* The hcall won't return, but to keep gcc happy, we're "done". */
1157         return NOTIFY_DONE;
1158 }
1159 
1160 static struct notifier_block paniced = {
1161         .notifier_call = lguest_panic
1162 };
1163 
1164 /* Setting up memory is fairly easy. */
1165 static __init char *lguest_memory_setup(void)
1166 {
1167         /*
1168          * The Linux bootloader header contains an "e820" memory map: the
1169          * Launcher populated the first entry with our memory limit.
1170          */
1171         e820_add_region(boot_params.e820_map[0].addr,
1172                           boot_params.e820_map[0].size,
1173                           boot_params.e820_map[0].type);
1174 
1175         /* This string is for the boot messages. */
1176         return "LGUEST";
1177 }
1178 
1179 /*
1180  * We will eventually use the virtio console device to produce console output,
1181  * but before that is set up we use LHCALL_NOTIFY on normal memory to produce
1182  * console output.
1183  */
1184 static __init int early_put_chars(u32 vtermno, const char *buf, int count)
1185 {
1186         char scratch[17];
1187         unsigned int len = count;
1188 
1189         /* We use a nul-terminated string, so we make a copy.  Icky, huh? */
1190         if (len > sizeof(scratch) - 1)
1191                 len = sizeof(scratch) - 1;
1192         scratch[len] = '\0';
1193         memcpy(scratch, buf, len);
1194         hcall(LHCALL_NOTIFY, __pa(scratch), 0, 0, 0);
1195 
1196         /* This routine returns the number of bytes actually written. */
1197         return len;
1198 }
1199 
1200 /*
1201  * Rebooting also tells the Host we're finished, but the RESTART flag tells the
1202  * Launcher to reboot us.
1203  */
1204 static void lguest_restart(char *reason)
1205 {
1206         hcall(LHCALL_SHUTDOWN, __pa(reason), LGUEST_SHUTDOWN_RESTART, 0, 0);
1207 }
1208 
1209 /*G:050
1210  * Patching (Powerfully Placating Performance Pedants)
1211  *
1212  * We have already seen that pv_ops structures let us replace simple native
1213  * instructions with calls to the appropriate back end all throughout the
1214  * kernel.  This allows the same kernel to run as a Guest and as a native
1215  * kernel, but it's slow because of all the indirect branches.
1216  *
1217  * Remember that David Wheeler quote about "Any problem in computer science can
1218  * be solved with another layer of indirection"?  The rest of that quote is
1219  * "... But that usually will create another problem."  This is the first of
1220  * those problems.
1221  *
1222  * Our current solution is to allow the paravirt back end to optionally patch
1223  * over the indirect calls to replace them with something more efficient.  We
1224  * patch two of the simplest of the most commonly called functions: disable
1225  * interrupts and save interrupts.  We usually have 6 or 10 bytes to patch
1226  * into: the Guest versions of these operations are small enough that we can
1227  * fit comfortably.
1228  *
1229  * First we need assembly templates of each of the patchable Guest operations,
1230  * and these are in i386_head.S.
1231  */
1232 
1233 /*G:060 We construct a table from the assembler templates: */
1234 static const struct lguest_insns
1235 {
1236         const char *start, *end;
1237 } lguest_insns[] = {
1238         [PARAVIRT_PATCH(pv_irq_ops.irq_disable)] = { lgstart_cli, lgend_cli },
1239         [PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf },
1240 };
1241 
1242 /*
1243  * Now our patch routine is fairly simple (based on the native one in
1244  * paravirt.c).  If we have a replacement, we copy it in and return how much of
1245  * the available space we used.
1246  */
1247 static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf,
1248                              unsigned long addr, unsigned len)
1249 {
1250         unsigned int insn_len;
1251 
1252         /* Don't do anything special if we don't have a replacement */
1253         if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start)
1254                 return paravirt_patch_default(type, clobber, ibuf, addr, len);
1255 
1256         insn_len = lguest_insns[type].end - lguest_insns[type].start;
1257 
1258         /* Similarly if it can't fit (doesn't happen, but let's be thorough). */
1259         if (len < insn_len)
1260                 return paravirt_patch_default(type, clobber, ibuf, addr, len);
1261 
1262         /* Copy in our instructions. */
1263         memcpy(ibuf, lguest_insns[type].start, insn_len);
1264         return insn_len;
1265 }
1266 
1267 /*G:029
1268  * Once we get to lguest_init(), we know we're a Guest.  The various
1269  * pv_ops structures in the kernel provide points for (almost) every routine we
1270  * have to override to avoid privileged instructions.
1271  */
1272 __init void lguest_init(void)
1273 {
1274         /* We're under lguest. */
1275         pv_info.name = "lguest";
1276         /* Paravirt is enabled. */
1277         pv_info.paravirt_enabled = 1;
1278         /* We're running at privilege level 1, not 0 as normal. */
1279         pv_info.kernel_rpl = 1;
1280         /* Everyone except Xen runs with this set. */
1281         pv_info.shared_kernel_pmd = 1;
1282 
1283         /*
1284          * We set up all the lguest overrides for sensitive operations.  These
1285          * are detailed with the operations themselves.
1286          */
1287 
1288         /* Interrupt-related operations */
1289         pv_irq_ops.save_fl = PV_CALLEE_SAVE(save_fl);
1290         pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(lg_restore_fl);
1291         pv_irq_ops.irq_disable = PV_CALLEE_SAVE(irq_disable);
1292         pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(lg_irq_enable);
1293         pv_irq_ops.safe_halt = lguest_safe_halt;
1294 
1295         /* Setup operations */
1296         pv_init_ops.patch = lguest_patch;
1297 
1298         /* Intercepts of various CPU instructions */
1299         pv_cpu_ops.load_gdt = lguest_load_gdt;
1300         pv_cpu_ops.cpuid = lguest_cpuid;
1301         pv_cpu_ops.load_idt = lguest_load_idt;
1302         pv_cpu_ops.iret = lguest_iret;
1303         pv_cpu_ops.load_sp0 = lguest_load_sp0;
1304         pv_cpu_ops.load_tr_desc = lguest_load_tr_desc;
1305         pv_cpu_ops.set_ldt = lguest_set_ldt;
1306         pv_cpu_ops.load_tls = lguest_load_tls;
1307         pv_cpu_ops.set_debugreg = lguest_set_debugreg;
1308         pv_cpu_ops.clts = lguest_clts;
1309         pv_cpu_ops.read_cr0 = lguest_read_cr0;
1310         pv_cpu_ops.write_cr0 = lguest_write_cr0;
1311         pv_cpu_ops.read_cr4 = lguest_read_cr4;
1312         pv_cpu_ops.write_cr4 = lguest_write_cr4;
1313         pv_cpu_ops.write_gdt_entry = lguest_write_gdt_entry;
1314         pv_cpu_ops.write_idt_entry = lguest_write_idt_entry;
1315         pv_cpu_ops.wbinvd = lguest_wbinvd;
1316         pv_cpu_ops.start_context_switch = paravirt_start_context_switch;
1317         pv_cpu_ops.end_context_switch = lguest_end_context_switch;
1318 
1319         /* Pagetable management */
1320         pv_mmu_ops.write_cr3 = lguest_write_cr3;
1321         pv_mmu_ops.flush_tlb_user = lguest_flush_tlb_user;
1322         pv_mmu_ops.flush_tlb_single = lguest_flush_tlb_single;
1323         pv_mmu_ops.flush_tlb_kernel = lguest_flush_tlb_kernel;
1324         pv_mmu_ops.set_pte = lguest_set_pte;
1325         pv_mmu_ops.set_pte_at = lguest_set_pte_at;
1326         pv_mmu_ops.set_pmd = lguest_set_pmd;
1327 #ifdef CONFIG_X86_PAE
1328         pv_mmu_ops.set_pte_atomic = lguest_set_pte_atomic;
1329         pv_mmu_ops.pte_clear = lguest_pte_clear;
1330         pv_mmu_ops.pmd_clear = lguest_pmd_clear;
1331         pv_mmu_ops.set_pud = lguest_set_pud;
1332 #endif
1333         pv_mmu_ops.read_cr2 = lguest_read_cr2;
1334         pv_mmu_ops.read_cr3 = lguest_read_cr3;
1335         pv_mmu_ops.lazy_mode.enter = paravirt_enter_lazy_mmu;
1336         pv_mmu_ops.lazy_mode.leave = lguest_leave_lazy_mmu_mode;
1337         pv_mmu_ops.lazy_mode.flush = paravirt_flush_lazy_mmu;
1338         pv_mmu_ops.pte_update = lguest_pte_update;
1339         pv_mmu_ops.pte_update_defer = lguest_pte_update;
1340 
1341 #ifdef CONFIG_X86_LOCAL_APIC
1342         /* APIC read/write intercepts */
1343         set_lguest_basic_apic_ops();
1344 #endif
1345 
1346         x86_init.resources.memory_setup = lguest_memory_setup;
1347         x86_init.irqs.intr_init = lguest_init_IRQ;
1348         x86_init.timers.timer_init = lguest_time_init;
1349         x86_platform.calibrate_tsc = lguest_tsc_khz;
1350         x86_platform.get_wallclock =  lguest_get_wallclock;
1351 
1352         /*
1353          * Now is a good time to look at the implementations of these functions
1354          * before returning to the rest of lguest_init().
1355          */
1356 
1357         /*G:070
1358          * Now we've seen all the paravirt_ops, we return to
1359          * lguest_init() where the rest of the fairly chaotic boot setup
1360          * occurs.
1361          */
1362 
1363         /*
1364          * The stack protector is a weird thing where gcc places a canary
1365          * value on the stack and then checks it on return.  This file is
1366          * compiled with -fno-stack-protector it, so we got this far without
1367          * problems.  The value of the canary is kept at offset 20 from the
1368          * %gs register, so we need to set that up before calling C functions
1369          * in other files.
1370          */
1371         setup_stack_canary_segment(0);
1372 
1373         /*
1374          * We could just call load_stack_canary_segment(), but we might as well
1375          * call switch_to_new_gdt() which loads the whole table and sets up the
1376          * per-cpu segment descriptor register %fs as well.
1377          */
1378         switch_to_new_gdt(0);
1379 
1380         /*
1381          * The Host<->Guest Switcher lives at the top of our address space, and
1382          * the Host told us how big it is when we made LGUEST_INIT hypercall:
1383          * it put the answer in lguest_data.reserve_mem
1384          */
1385         reserve_top_address(lguest_data.reserve_mem);
1386 
1387         /*
1388          * If we don't initialize the lock dependency checker now, it crashes
1389          * atomic_notifier_chain_register, then paravirt_disable_iospace.
1390          */
1391         lockdep_init();
1392 
1393         /* Hook in our special panic hypercall code. */
1394         atomic_notifier_chain_register(&panic_notifier_list, &paniced);
1395 
1396         /*
1397          * The IDE code spends about 3 seconds probing for disks: if we reserve
1398          * all the I/O ports up front it can't get them and so doesn't probe.
1399          * Other device drivers are similar (but less severe).  This cuts the
1400          * kernel boot time on my machine from 4.1 seconds to 0.45 seconds.
1401          */
1402         paravirt_disable_iospace();
1403 
1404         /*
1405          * This is messy CPU setup stuff which the native boot code does before
1406          * start_kernel, so we have to do, too:
1407          */
1408         cpu_detect(&new_cpu_data);
1409         /* head.S usually sets up the first capability word, so do it here. */
1410         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1411 
1412         /* Math is always hard! */
1413         new_cpu_data.hard_math = 1;
1414 
1415         /* We don't have features.  We have puppies!  Puppies! */
1416 #ifdef CONFIG_X86_MCE
1417         mca_cfg.disabled = true;
1418 #endif
1419 #ifdef CONFIG_ACPI
1420         acpi_disabled = 1;
1421 #endif
1422 
1423         /*
1424          * We set the preferred console to "hvc".  This is the "hypervisor
1425          * virtual console" driver written by the PowerPC people, which we also
1426          * adapted for lguest's use.
1427          */
1428         add_preferred_console("hvc", 0, NULL);
1429 
1430         /* Register our very early console. */
1431         virtio_cons_early_init(early_put_chars);
1432 
1433         /*
1434          * Last of all, we set the power management poweroff hook to point to
1435          * the Guest routine to power off, and the reboot hook to our restart
1436          * routine.
1437          */
1438         pm_power_off = lguest_power_off;
1439         machine_ops.restart = lguest_restart;
1440 
1441         /*
1442          * Now we're set up, call i386_start_kernel() in head32.c and we proceed
1443          * to boot as normal.  It never returns.
1444          */
1445         i386_start_kernel();
1446 }
1447 /*
1448  * This marks the end of stage II of our journey, The Guest.
1449  *
1450  * It is now time for us to explore the layer of virtual drivers and complete
1451  * our understanding of the Guest in "make Drivers".
1452  */
1453 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp