diff options
Diffstat (limited to 'arch/x86/lguest')
| -rw-r--r-- | arch/x86/lguest/boot.c | 428 | ||||
| -rw-r--r-- | arch/x86/lguest/i386_head.S | 110 |
2 files changed, 347 insertions, 191 deletions
diff --git a/arch/x86/lguest/boot.c b/arch/x86/lguest/boot.c index f2bf1f73d468..025c04d18f2b 100644 --- a/arch/x86/lguest/boot.c +++ b/arch/x86/lguest/boot.c @@ -22,7 +22,8 @@ * * So how does the kernel know it's a Guest? We'll see that later, but let's * just say that we end up here where we replace the native functions various - * "paravirt" structures with our Guest versions, then boot like normal. :*/ + * "paravirt" structures with our Guest versions, then boot like normal. +:*/ /* * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation. @@ -74,7 +75,8 @@ * * The Guest in our tale is a simple creature: identical to the Host but * behaving in simplified but equivalent ways. In particular, the Guest is the - * same kernel as the Host (or at least, built from the same source code). :*/ + * same kernel as the Host (or at least, built from the same source code). +:*/ struct lguest_data lguest_data = { .hcall_status = { [0 ... LHCALL_RING_SIZE-1] = 0xFF }, @@ -85,7 +87,8 @@ struct lguest_data lguest_data = { .syscall_vec = SYSCALL_VECTOR, }; -/*G:037 async_hcall() is pretty simple: I'm quite proud of it really. We have a +/*G:037 + * async_hcall() is pretty simple: I'm quite proud of it really. We have a * ring buffer of stored hypercalls which the Host will run though next time we * do a normal hypercall. Each entry in the ring has 5 slots for the hypercall * arguments, and a "hcall_status" word which is 0 if the call is ready to go, @@ -94,7 +97,8 @@ struct lguest_data lguest_data = { * If we come around to a slot which hasn't been finished, then the table is * full and we just make the hypercall directly. This has the nice side * effect of causing the Host to run all the stored calls in the ring buffer - * which empties it for next time! */ + * which empties it for next time! + */ static void async_hcall(unsigned long call, unsigned long arg1, unsigned long arg2, unsigned long arg3, unsigned long arg4) @@ -103,9 +107,11 @@ static void async_hcall(unsigned long call, unsigned long arg1, static unsigned int next_call; unsigned long flags; - /* Disable interrupts if not already disabled: we don't want an + /* + * Disable interrupts if not already disabled: we don't want an * interrupt handler making a hypercall while we're already doing - * one! */ + * one! + */ local_irq_save(flags); if (lguest_data.hcall_status[next_call] != 0xFF) { /* Table full, so do normal hcall which will flush table. */ @@ -125,8 +131,9 @@ static void async_hcall(unsigned long call, unsigned long arg1, local_irq_restore(flags); } -/*G:035 Notice the lazy_hcall() above, rather than hcall(). This is our first - * real optimization trick! +/*G:035 + * Notice the lazy_hcall() above, rather than hcall(). This is our first real + * optimization trick! * * When lazy_mode is set, it means we're allowed to defer all hypercalls and do * them as a batch when lazy_mode is eventually turned off. Because hypercalls @@ -136,7 +143,8 @@ static void async_hcall(unsigned long call, unsigned long arg1, * lguest_leave_lazy_mode(). * * So, when we're in lazy mode, we call async_hcall() to store the call for - * future processing: */ + * future processing: + */ static void lazy_hcall1(unsigned long call, unsigned long arg1) { @@ -208,9 +216,11 @@ static void lguest_end_context_switch(struct task_struct *next) * check there before it tries to deliver an interrupt. */ -/* save_flags() is expected to return the processor state (ie. "flags"). The +/* + * save_flags() is expected to return the processor state (ie. "flags"). The * flags word contains all kind of stuff, but in practice Linux only cares - * about the interrupt flag. Our "save_flags()" just returns that. */ + * about the interrupt flag. Our "save_flags()" just returns that. + */ static unsigned long save_fl(void) { return lguest_data.irq_enabled; @@ -222,13 +232,15 @@ static void irq_disable(void) lguest_data.irq_enabled = 0; } -/* Let's pause a moment. Remember how I said these are called so often? +/* + * Let's pause a moment. Remember how I said these are called so often? * Jeremy Fitzhardinge optimized them so hard early in 2009 that he had to * break some rules. In particular, these functions are assumed to save their * own registers if they need to: normal C functions assume they can trash the * eax register. To use normal C functions, we use * PV_CALLEE_SAVE_REGS_THUNK(), which pushes %eax onto the stack, calls the - * C function, then restores it. */ + * C function, then restores it. + */ PV_CALLEE_SAVE_REGS_THUNK(save_fl); PV_CALLEE_SAVE_REGS_THUNK(irq_disable); /*:*/ @@ -237,18 +249,20 @@ PV_CALLEE_SAVE_REGS_THUNK(irq_disable); extern void lg_irq_enable(void); extern void lg_restore_fl(unsigned long flags); -/*M:003 Note that we don't check for outstanding interrupts when we re-enable - * them (or when we unmask an interrupt). This seems to work for the moment, - * since interrupts are rare and we'll just get the interrupt on the next timer - * tick, but now we can run with CONFIG_NO_HZ, we should revisit this. One way - * would be to put the "irq_enabled" field in a page by itself, and have the - * Host write-protect it when an interrupt comes in when irqs are disabled. - * There will then be a page fault as soon as interrupts are re-enabled. +/*M:003 + * Note that we don't check for outstanding interrupts when we re-enable them + * (or when we unmask an interrupt). This seems to work for the moment, since + * interrupts are rare and we'll just get the interrupt on the next timer tick, + * but now we can run with CONFIG_NO_HZ, we should revisit this. One way would + * be to put the "irq_enabled" field in a page by itself, and have the Host + * write-protect it when an interrupt comes in when irqs are disabled. There + * will then be a page fault as soon as interrupts are re-enabled. * * A better method is to implement soft interrupt disable generally for x86: * instead of disabling interrupts, we set a flag. If an interrupt does come * in, we then disable them for real. This is uncommon, so we could simply use - * a hypercall for interrupt control and not worry about efficiency. :*/ + * a hypercall for interrupt control and not worry about efficiency. +:*/ /*G:034 * The Interrupt Descriptor Table (IDT). @@ -261,10 +275,12 @@ extern void lg_restore_fl(unsigned long flags); static void lguest_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g) { - /* The gate_desc structure is 8 bytes long: we hand it to the Host in + /* + * The gate_desc structure is 8 bytes long: we hand it to the Host in * two 32-bit chunks. The whole 32-bit kernel used to hand descriptors * around like this; typesafety wasn't a big concern in Linux's early - * years. */ + * years. + */ u32 *desc = (u32 *)g; /* Keep the local copy up to date. */ native_write_idt_entry(dt, entrynum, g); @@ -272,9 +288,11 @@ static void lguest_write_idt_entry(gate_desc *dt, kvm_hypercall3(LHCALL_LOAD_IDT_ENTRY, entrynum, desc[0], desc[1]); } -/* Changing to a different IDT is very rare: we keep the IDT up-to-date every +/* + * Changing to a different IDT is very rare: we keep the IDT up-to-date every * time it is written, so we can simply loop through all entries and tell the - * Host about them. */ + * Host about them. + */ static void lguest_load_idt(const struct desc_ptr *desc) { unsigned int i; @@ -305,9 +323,11 @@ static void lguest_load_gdt(const struct desc_ptr *desc) kvm_hypercall3(LHCALL_LOAD_GDT_ENTRY, i, gdt[i].a, gdt[i].b); } -/* For a single GDT entry which changes, we do the lazy thing: alter our GDT, +/* + * For a single GDT entry which changes, we do the lazy thing: alter our GDT, * then tell the Host to reload the entire thing. This operation is so rare - * that this naive implementation is reasonable. */ + * that this naive implementation is reasonable. + */ static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum, const void *desc, int type) { @@ -317,29 +337,36 @@ static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum, dt[entrynum].a, dt[entrynum].b); } -/* OK, I lied. There are three "thread local storage" GDT entries which change +/* + * OK, I lied. There are three "thread local storage" GDT entries which change * on every context switch (these three entries are how glibc implements - * __thread variables). So we have a hypercall specifically for this case. */ + * __thread variables). So we have a hypercall specifically for this case. + */ static void lguest_load_tls(struct thread_struct *t, unsigned int cpu) { - /* There's one problem which normal hardware doesn't have: the Host + /* + * There's one problem which normal hardware doesn't have: the Host * can't handle us removing entries we're currently using. So we clear - * the GS register here: if it's needed it'll be reloaded anyway. */ + * the GS register here: if it's needed it'll be reloaded anyway. + */ lazy_load_gs(0); lazy_hcall2(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu); } -/*G:038 That's enough excitement for now, back to ploughing through each of - * the different pv_ops structures (we're about 1/3 of the way through). +/*G:038 + * That's enough excitement for now, back to ploughing through each of the + * different pv_ops structures (we're about 1/3 of the way through). * * This is the Local Descriptor Table, another weird Intel thingy. Linux only * uses this for some strange applications like Wine. We don't do anything - * here, so they'll get an informative and friendly Segmentation Fault. */ + * here, so they'll get an informative and friendly Segmentation Fault. + */ static void lguest_set_ldt(const void *addr, unsigned entries) { } -/* This loads a GDT entry into the "Task Register": that entry points to a +/* + * This loads a GDT entry into the "Task Register": that entry points to a * structure called the Task State Segment. Some comments scattered though the * kernel code indicate that this used for task switching in ages past, along * with blood sacrifice and astrology. @@ -347,19 +374,21 @@ static void lguest_set_ldt(const void *addr, unsigned entries) * Now there's nothing interesting in here that we don't get told elsewhere. * But the native version uses the "ltr" instruction, which makes the Host * complain to the Guest about a Segmentation Fault and it'll oops. So we - * override the native version with a do-nothing version. */ + * override the native version with a do-nothing version. + */ static void lguest_load_tr_desc(void) { } -/* The "cpuid" instruction is a way of querying both the CPU identity +/* + * The "cpuid" instruction is a way of querying both the CPU identity * (manufacturer, model, etc) and its features. It was introduced before the * Pentium in 1993 and keeps getting extended by both Intel, AMD and others. * As you might imagine, after a decade and a half this treatment, it is now a * giant ball of hair. Its entry in the current Intel manual runs to 28 pages. * * This instruction even it has its own Wikipedia entry. The Wikipedia entry - * has been translated into 4 languages. I am not making this up! + * has been translated into 5 languages. I am not making this up! * * We could get funky here and identify ourselves as "GenuineLguest", but * instead we just use the real "cpuid" instruction. Then I pretty much turned @@ -371,7 +400,8 @@ static void lguest_load_tr_desc(void) * Replacing the cpuid so we can turn features off is great for the kernel, but * anyone (including userspace) can just use the raw "cpuid" instruction and * the Host won't even notice since it isn't privileged. So we try not to get - * too worked up about it. */ + * too worked up about it. + */ static void lguest_cpuid(unsigned int *ax, unsigned int *bx, unsigned int *cx, unsigned int *dx) { @@ -379,43 +409,63 @@ static void lguest_cpuid(unsigned int *ax, unsigned int *bx, native_cpuid(ax, bx, cx, dx); switch (function) { - case 0: /* ID and highest CPUID. Futureproof a little by sticking to - * older ones. */ + /* + * CPUID 0 gives the highest legal CPUID number (and the ID string). + * We futureproof our code a little by sticking to known CPUID values. + */ + case 0: if (*ax > 5) *ax = 5; break; - case 1: /* Basic feature request. */ - /* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */ + + /* + * CPUID 1 is a basic feature request. + * + * CX: we only allow kernel to see SSE3, CMPXCHG16B and SSSE3 + * DX: SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU and PAE. + */ + case 1: *cx &= 0x00002201; - /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU, PAE. */ *dx &= 0x07808151; - /* The Host can do a nice optimization if it knows that the + /* + * The Host can do a nice optimization if it knows that the * kernel mappings (addresses above 0xC0000000 or whatever * PAGE_OFFSET is set to) haven't changed. But Linux calls * flush_tlb_user() for both user and kernel mappings unless - * the Page Global Enable (PGE) feature bit is set. */ + * the Page Global Enable (PGE) feature bit is set. + */ *dx |= 0x00002000; - /* We also lie, and say we're family id 5. 6 or greater + /* + * We also lie, and say we're family id 5. 6 or greater * leads to a rdmsr in early_init_intel which we can't handle. - * Family ID is returned as bits 8-12 in ax. */ + * Family ID is returned as bits 8-12 in ax. + */ *ax &= 0xFFFFF0FF; *ax |= 0x00000500; break; + /* + * 0x80000000 returns the highest Extended Function, so we futureproof + * like we do above by limiting it to known fields. + */ case 0x80000000: - /* Futureproof this a little: if they ask how much extended - * processor information there is, limit it to known fields. */ if (*ax > 0x80000008) *ax = 0x80000008; break; + + /* + * PAE systems can mark pages as non-executable. Linux calls this the + * NX bit. Intel calls it XD (eXecute Disable), AMD EVP (Enhanced + * Virus Protection). We just switch turn if off here, since we don't + * support it. + */ case 0x80000001: - /* Here we should fix nx cap depending on host. */ - /* For this version of PAE, we just clear NX bit. */ *dx &= ~(1 << 20); break; } } -/* Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4. +/* + * Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4. * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother * it. The Host needs to know when the Guest wants to change them, so we have * a whole series of functions like read_cr0() and write_cr0(). @@ -430,7 +480,8 @@ static void lguest_cpuid(unsigned int *ax, unsigned int *bx, * name like "FPUTRAP bit" be a little less cryptic? * * We store cr0 locally because the Host never changes it. The Guest sometimes - * wants to read it and we'd prefer not to bother the Host unnecessarily. */ + * wants to read it and we'd prefer not to bother the Host unnecessarily. + */ static unsigned long current_cr0; static void lguest_write_cr0(unsigned long val) { @@ -443,18 +494,22 @@ static unsigned long lguest_read_cr0(void) return current_cr0; } -/* Intel provided a special instruction to clear the TS bit for people too cool +/* + * Intel provided a special instruction to clear the TS bit for people too cool * to use write_cr0() to do it. This "clts" instruction is faster, because all - * the vowels have been optimized out. */ + * the vowels have been optimized out. + */ static void lguest_clts(void) { lazy_hcall1(LHCALL_TS, 0); current_cr0 &= ~X86_CR0_TS; } -/* cr2 is the virtual address of the last page fault, which the Guest only ever +/* + * cr2 is the virtual address of the last page fault, which the Guest only ever * reads. The Host kindly writes this into our "struct lguest_data", so we - * just read it out of there. */ + * just read it out of there. + */ static unsigned long lguest_read_cr2(void) { return lguest_data.cr2; @@ -463,10 +518,12 @@ static unsigned long lguest_read_cr2(void) /* See lguest_set_pte() below. */ static bool cr3_changed = false; -/* cr3 is the current toplevel pagetable page: the principle is the same as +/* + * cr3 is the current toplevel pagetable page: the principle is the same as * cr0. Keep a local copy, and tell the Host when it changes. The only * difference is that our local copy is in lguest_data because the Host needs - * to set it upon our initial hypercall. */ + * to set it upon our initial hypercall. + */ static void lguest_write_cr3(unsigned long cr3) { lguest_data.pgdir = cr3; @@ -538,10 +595,12 @@ static void lguest_write_cr4(unsigned long val) * the real page tables based on the Guests'. */ -/* The Guest calls this to set a second-level entry (pte), ie. to map a page +/* + * The Guest calls this to set a second-level entry (pte), ie. to map a page * into a process' address space. We set the entry then tell the Host the * toplevel and address this corresponds to. The Guest uses one pagetable per - * process, so we need to tell the Host which one we're changing (mm->pgd). */ + * process, so we need to tell the Host which one we're changing (mm->pgd). + */ static void lguest_pte_update(struct mm_struct *mm, unsigned long addr, pte_t *ptep) { @@ -560,10 +619,13 @@ static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr, lguest_pte_update(mm, addr, ptep); } -/* The Guest calls lguest_set_pud to set a top-level entry and lguest_set_pmd +/* + * The Guest calls lguest_set_pud to set a top-level entry and lguest_set_pmd * to set a middle-level entry when PAE is activated. + * * Again, we set the entry then tell the Host which page we changed, - * and the index of the entry we changed. */ + * and the index of the entry we changed. + */ #ifdef CONFIG_X86_PAE static void lguest_set_pud(pud_t *pudp, pud_t pudval) { @@ -582,8 +644,7 @@ static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) } #else -/* The Guest calls lguest_set_pmd to set a top-level entry when PAE is not - * activated. */ +/* The Guest calls lguest_set_pmd to set a top-level entry when !PAE. */ static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) { native_set_pmd(pmdp, pmdval); @@ -592,7 +653,8 @@ static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) } #endif -/* There are a couple of legacy places where the kernel sets a PTE, but we +/* + * There are a couple of legacy places where the kernel sets a PTE, but we * don't know the top level any more. This is useless for us, since we don't * know which pagetable is changing or what address, so we just tell the Host * to forget all of them. Fortunately, this is very rare. @@ -600,7 +662,8 @@ static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) * ... except in early boot when the kernel sets up the initial pagetables, * which makes booting astonishingly slow: 1.83 seconds! So we don't even tell * the Host anything changed until we've done the first page table switch, - * which brings boot back to 0.25 seconds. */ + * which brings boot back to 0.25 seconds. + */ static void lguest_set_pte(pte_t *ptep, pte_t pteval) { native_set_pte(ptep, pteval); @@ -628,7 +691,8 @@ void lguest_pmd_clear(pmd_t *pmdp) } #endif -/* Unfortunately for Lguest, the pv_mmu_ops for page tables were based on +/* + * Unfortunately for Lguest, the pv_mmu_ops for page tables were based on * native page table operations. On native hardware you can set a new page * table entry whenever you want, but if you want to remove one you have to do * a TLB flush (a TLB is a little cache of page table entries kept by the CPU). @@ -637,24 +701,29 @@ void lguest_pmd_clear(pmd_t *pmdp) * called when a valid entry is written, not when it's removed (ie. marked not * present). Instead, this is where we come when the Guest wants to remove a * page table entry: we tell the Host to set that entry to 0 (ie. the present - * bit is zero). */ + * bit is zero). + */ static void lguest_flush_tlb_single(unsigned long addr) { /* Simply set it to zero: if it was not, it will fault back in. */ lazy_hcall3(LHCALL_SET_PTE, lguest_data.pgdir, addr, 0); } -/* This is what happens after the Guest has removed a large number of entries. +/* + * This is what happens after the Guest has removed a large number of entries. * This tells the Host that any of the page table entries for userspace might - * have changed, ie. virtual addresses below PAGE_OFFSET. */ + * have changed, ie. virtual addresses below PAGE_OFFSET. + */ static void lguest_flush_tlb_user(void) { lazy_hcall1(LHCALL_FLUSH_TLB, 0); } -/* This is called when the kernel page tables have changed. That's not very +/* + * This is called when the kernel page tables have changed. That's not very * common (unless the Guest is using highmem, which makes the Guest extremely - * slow), so it's worth separating this from the user flushing above. */ + * slow), so it's worth separating this from the user flushing above. + */ static void lguest_flush_tlb_kernel(void) { lazy_hcall1(LHCALL_FLUSH_TLB, 1); @@ -691,23 +760,27 @@ static struct irq_chip lguest_irq_controller = { .unmask = enable_lguest_irq, }; -/* This sets up the Interrupt Descriptor Table (IDT) entry for each hardware +/* + * This sets up the Interrupt Descriptor Table (IDT) entry for each hardware * interrupt (except 128, which is used for system calls), and then tells the * Linux infrastructure that each interrupt is controlled by our level-based - * lguest interrupt controller. */ + * lguest interrupt controller. + */ static void __init lguest_init_IRQ(void) { unsigned int i; for (i = FIRST_EXTERNAL_VECTOR; i < NR_VECTORS; i++) { - /* Some systems map "vectors" to interrupts weirdly. Lguest has - * a straightforward 1 to 1 mapping, so force that here. */ + /* Some systems map "vectors" to interrupts weirdly. Not us! */ __get_cpu_var(vector_irq)[i] = i - FIRST_EXTERNAL_VECTOR; if (i != SYSCALL_VECTOR) set_intr_gate(i, interrupt[i - FIRST_EXTERNAL_VECTOR]); } - /* This call is required to set up for 4k stacks, where we have - * separate stacks for hard and soft interrupts. */ + + /* + * This call is required to set up for 4k stacks, where we have + * separate stacks for hard and soft interrupts. + */ irq_ctx_init(smp_processor_id()); } @@ -729,31 +802,39 @@ static unsigned long lguest_get_wallclock(void) return lguest_data.time.tv_sec; } -/* The TSC is an Intel thing called the Time Stamp Counter. The Host tells us +/* + * The TSC is an Intel thing called the Time Stamp Counter. The Host tells us * what speed it runs at, or 0 if it's unusable as a reliable clock source. * This matches what we want here: if we return 0 from this function, the x86 - * TSC clock will give up and not register itself. */ + * TSC clock will give up and not register itself. + */ static unsigned long lguest_tsc_khz(void) { return lguest_data.tsc_khz; } -/* If we can't use the TSC, the kernel falls back to our lower-priority - * "lguest_clock", where we read the time value given to us by the Host. */ +/* + * If we can't use the TSC, the kernel falls back to our lower-priority + * "lguest_clock", where we read the time value given to us by the Host. + */ static cycle_t lguest_clock_read(struct clocksource *cs) { unsigned long sec, nsec; - /* Since the time is in two parts (seconds and nanoseconds), we risk + /* + * Since the time is in two parts (seconds and nanoseconds), we risk * reading it just as it's changing from 99 & 0.999999999 to 100 and 0, * and getting 99 and 0. As Linux tends to come apart under the stress - * of time travel, we must be careful: */ + * of time travel, we must be careful: + */ do { /* First we read the seconds part. */ sec = lguest_data.time.tv_sec; - /* This read memory barrier tells the compiler and the CPU that + /* + * This read memory barrier tells the compiler and the CPU that * this can't be reordered: we have to complete the above - * before going on. */ + * before going on. + */ rmb(); /* Now we read the nanoseconds part. */ nsec = lguest_data.time.tv_nsec; @@ -777,9 +858,11 @@ static struct clocksource lguest_clock = { .flags = CLOCK_SOURCE_IS_CONTINUOUS, }; -/* We also need a "struct clock_event_device": Linux asks us to set it to go +/* + * We also need a "struct clock_event_device": Linux asks us to set it to go * off some time in the future. Actually, James Morris figured all this out, I - * just applied the patch. */ + * just applied the patch. + */ static int lguest_clockevent_set_next_event(unsigned long delta, struct clock_event_device *evt) { @@ -829,8 +912,10 @@ static struct clock_event_device lguest_clockevent = { .max_delta_ns = LG_CLOCK_MAX_DELTA, }; -/* This is the Guest timer interrupt handler (hardware interrupt 0). We just - * call the clockevent infrastructure and it does whatever needs doing. */ +/* + * This is the Guest timer interrupt handler (hardware interrupt 0). We just + * call the clockevent infrastructure and it does whatever needs doing. + */ static void lguest_time_irq(unsigned int irq, struct irq_desc *desc) { unsigned long flags; @@ -841,10 +926,12 @@ static void lguest_time_irq(unsigned int irq, struct irq_desc *desc) local_irq_restore(flags); } -/* At some point in the boot process, we get asked to set up our timing +/* + * At some point in the boot process, we get asked to set up our timing * infrastructure. The kernel doesn't expect timer interrupts before this, but * we cleverly initialized the "blocked_interrupts" field of "struct - * lguest_data" so that timer interrupts were blocked until now. */ + * lguest_data" so that timer interrupts were blocked until now. + */ static void lguest_time_init(void) { /* Set up the timer interrupt (0) to go to our simple timer routine */ @@ -868,14 +955,16 @@ static void lguest_time_init(void) * to work. They're pretty simple. */ -/* The Guest needs to tell the Host what stack it expects traps to use. For +/* + * The Guest needs to tell the Host what stack it expects traps to use. For * native hardware, this is part of the Task State Segment mentioned above in * lguest_load_tr_desc(), but to help hypervisors there's this special call. * * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data * segment), the privilege level (we're privilege level 1, the Host is 0 and * will not tolerate us trying to use that), the stack pointer, and the number - * of pages in the stack. */ + * of pages in the stack. + */ static void lguest_load_sp0(struct tss_struct *tss, struct thread_struct *thread) { @@ -889,7 +978,8 @@ static void lguest_set_debugreg(int regno, unsigned long value) /* FIXME: Implement */ } -/* There are times when the kernel wants to make sure that no memory writes are +/* + * There are times when the kernel wants to make sure that no memory writes are * caught in the cache (that they've all reached real hardware devices). This * doesn't matter for the Guest which has virtual hardware. * @@ -903,11 +993,13 @@ static void lguest_wbinvd(void) { } -/* If the Guest expects to have an Advanced Programmable Interrupt Controller, +/* + * If the Guest expects to have an Advanced Programmable Interrupt Controller, * we play dumb by ignoring writes and returning 0 for reads. So it's no * longer Programmable nor Controlling anything, and I don't think 8 lines of * code qualifies for Advanced. It will also never interrupt anything. It - * does, however, allow us to get through the Linux boot code. */ + * does, however, allow us to get through the Linux boot code. + */ #ifdef CONFIG_X86_LOCAL_APIC static void lguest_apic_write(u32 reg, u32 v) { @@ -956,11 +1048,13 @@ static void lguest_safe_halt(void) kvm_hypercall0(LHCALL_HALT); } -/* The SHUTDOWN hypercall takes a string to describe what's happening, and +/* + * The SHUTDOWN hypercall takes a string to describe what's happening, and * an argument which says whether this to restart (reboot) the Guest or not. * * Note that the Host always prefers that the Guest speak in physical addresses - * rather than virtual addresses, so we use __pa() here. */ + * rather than virtual addresses, so we use __pa() here. + */ static void lguest_power_off(void) { kvm_hypercall2(LHCALL_SHUTDOWN, __pa("Power down"), @@ -991,8 +1085,10 @@ static __init char *lguest_memory_setup(void) * nice to move it back to lguest_init. Patch welcome... */ atomic_notifier_chain_register(&panic_notifier_list, &paniced); - /* The Linux bootloader header contains an "e820" memory map: the - * Launcher populated the first entry with our memory limit. */ + /* + *The Linux bootloader header contains an "e820" memory map: the + * Launcher populated the first entry with our memory limit. + */ e820_add_region(boot_params.e820_map[0].addr, boot_params.e820_map[0].size, boot_params.e820_map[0].type); @@ -1001,16 +1097,17 @@ static __init char *lguest_memory_setup(void) return "LGUEST"; } -/* We will eventually use the virtio console device to produce console output, +/* + * We will eventually use the virtio console device to produce console output, * but before that is set up we use LHCALL_NOTIFY on normal memory to produce - * console output. */ + * console output. + */ static __init int early_put_chars(u32 vtermno, const char *buf, int count) { char scratch[17]; unsigned int len = count; - /* We use a nul-terminated string, so we have to make a copy. Icky, - * huh? */ + /* We use a nul-terminated string, so we make a copy. Icky, huh? */ if (len > sizeof(scratch) - 1) len = sizeof(scratch) - 1; scratch[len] = '\0'; @@ -1021,8 +1118,10 @@ static __init int early_put_chars(u32 vtermno, const char *buf, int count) return len; } -/* Rebooting also tells the Host we're finished, but the RESTART flag tells the - * Launcher to reboot us. */ +/* + * Rebooting also tells the Host we're finished, but the RESTART flag tells the + * Launcher to reboot us. + */ static void lguest_restart(char *reason) { kvm_hypercall2(LHCALL_SHUTDOWN, __pa(reason), LGUEST_SHUTDOWN_RESTART); @@ -1049,7 +1148,8 @@ static void lguest_restart(char *reason) * fit comfortably. * * First we need assembly templates of each of the patchable Guest operations, - * and these are in i386_head.S. */ + * and these are in i386_head.S. + */ /*G:060 We construct a table from the assembler templates: */ static const struct lguest_insns @@ -1060,9 +1160,11 @@ static const struct lguest_insns [PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf }, }; -/* Now our patch routine is fairly simple (based on the native one in +/* + * Now our patch routine is fairly simple (based on the native one in * paravirt.c). If we have a replacement, we copy it in and return how much of - * the available space we used. */ + * the available space we used. + */ static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf, unsigned long addr, unsigned len) { @@ -1074,8 +1176,7 @@ static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf, insn_len = lguest_insns[type].end - lguest_insns[type].start; - /* Similarly if we can't fit replacement (shouldn't happen, but let's - * be thorough). */ + /* Similarly if it can't fit (doesn't happen, but let's be thorough). */ if (len < insn_len) return paravirt_patch_default(type, clobber, ibuf, addr, len); @@ -1084,22 +1185,28 @@ static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf, return insn_len; } -/*G:029 Once we get to lguest_init(), we know we're a Guest. The various +/*G:029 + * Once we get to lguest_init(), we know we're a Guest. The various * pv_ops structures in the kernel provide points for (almost) every routine we - * have to override to avoid privileged instructions. */ + * have to override to avoid privileged instructions. + */ __init void lguest_init(void) { - /* We're under lguest, paravirt is enabled, and we're running at - * privilege level 1, not 0 as normal. */ + /* We're under lguest. */ pv_info.name = "lguest"; + /* Paravirt is enabled. */ pv_info.paravirt_enabled = 1; + /* We're running at privilege level 1, not 0 as normal. */ pv_info.kernel_rpl = 1; + /* Everyone except Xen runs with this set. */ pv_info.shared_kernel_pmd = 1; - /* We set up all the lguest overrides for sensitive operations. These - * are detailed with the operations themselves. */ + /* + * We set up all the lguest overrides for sensitive operations. These + * are detailed with the operations themselves. + */ - /* interrupt-related operations */ + /* Interrupt-related operations */ pv_irq_ops.init_IRQ = lguest_init_IRQ; pv_irq_ops.save_fl = PV_CALLEE_SAVE(save_fl); pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(lg_restore_fl); @@ -1107,11 +1214,11 @@ __init void lguest_init(void) pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(lg_irq_enable); pv_irq_ops.safe_halt = lguest_safe_halt; - /* init-time operations */ + /* Setup operations */ pv_init_ops.memory_setup = lguest_memory_setup; pv_init_ops.patch = lguest_patch; - /* Intercepts of various cpu instructions */ + /* Intercepts of various CPU instructions */ pv_cpu_ops.load_gdt = lguest_load_gdt; pv_cpu_ops.cpuid = lguest_cpuid; pv_cpu_ops.load_idt = lguest_load_idt; @@ -1132,7 +1239,7 @@ __init void lguest_init(void) pv_cpu_ops.start_context_switch = paravirt_start_context_switch; pv_cpu_ops.end_context_switch = lguest_end_context_switch; - /* pagetable management */ + /* Pagetable management */ pv_mmu_ops.write_cr3 = lguest_write_cr3; pv_mmu_ops.flush_tlb_user = lguest_flush_tlb_user; pv_mmu_ops.flush_tlb_single = lguest_flush_tlb_single; @@ -1154,54 +1261,71 @@ __init void lguest_init(void) pv_mmu_ops.pte_update_defer = lguest_pte_update; #ifdef CONFIG_X86_LOCAL_APIC - /* apic read/write intercepts */ + /* APIC read/write intercepts */ set_lguest_basic_apic_ops(); #endif - /* time operations */ + /* Time operations */ pv_time_ops.get_wallclock = lguest_get_wallclock; pv_time_ops.time_init = lguest_time_init; pv_time_ops.get_tsc_khz = lguest_tsc_khz; - /* Now is a good time to look at the implementations of these functions - * before returning to the rest of lguest_init(). */ + /* + * Now is a good time to look at the implementations of these functions + * before returning to the rest of lguest_init(). + */ - /*G:070 Now we've seen all the paravirt_ops, we return to + /*G:070 + * Now we've seen all the paravirt_ops, we return to * lguest_init() where the rest of the fairly chaotic boot setup - * occurs. */ + * occurs. + */ - /* The stack protector is a weird thing where gcc places a canary + /* + * The stack protector is a weird thing where gcc places a canary * value on the stack and then checks it on return. This file is * compiled with -fno-stack-protector it, so we got this far without * problems. The value of the canary is kept at offset 20 from the * %gs register, so we need to set that up before calling C functions - * in other files. */ + * in other files. + */ setup_stack_canary_segment(0); - /* We could just call load_stack_canary_segment(), but we might as - * call switch_to_new_gdt() which loads the whole table and sets up - * the per-cpu segment descriptor register %fs as well. */ + + /* + * We could just call load_stack_canary_segment(), but we might as well + * call switch_to_new_gdt() which loads the whole table and sets up the + * per-cpu segment descriptor register %fs as well. + */ switch_to_new_gdt(0); /* As described in head_32.S, we map the first 128M of memory. */ max_pfn_mapped = (128*1024*1024) >> PAGE_SHIFT; - /* The Host<->Guest Switcher lives at the top of our address space, and + /* + * The Host<->Guest Switcher lives at the top of our address space, and * the Host told us how big it is when we made LGUEST_INIT hypercall: - * it put the answer in lguest_data.reserve_mem */ + * it put the answer in lguest_data.reserve_mem + */ reserve_top_address(lguest_data.reserve_mem); - /* If we don't initialize the lock dependency checker now, it crashes - * paravirt_disable_iospace. */ + /* + * If we don't initialize the lock dependency checker now, it crashes + * paravirt_disable_iospace. + */ lockdep_init(); - /* The IDE code spends about 3 seconds probing for disks: if we reserve + /* + * The IDE code spends about 3 seconds probing for disks: if we reserve * all the I/O ports up front it can't get them and so doesn't probe. * Other device drivers are similar (but less severe). This cuts the - * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. */ + * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. + */ paravirt_disable_iospace(); - /* This is messy CPU setup stuff which the native boot code does before - * start_kernel, so we have to do, too: */ + /* + * This is messy CPU setup stuff which the native boot code does before + * start_kernel, so we have to do, too: + */ cpu_detect(&new_cpu_data); /* head.S usually sets up the first capability word, so do it here. */ new_cpu_data.x86_capability[0] = cpuid_edx(1); @@ -1218,22 +1342,28 @@ __init void lguest_init(void) acpi_ht = 0; #endif - /* We set the preferred console to "hvc". This is the "hypervisor + /* + * We set the preferred console to "hvc". This is the "hypervisor * virtual console" driver written by the PowerPC people, which we also - * adapted for lguest's use. */ + * adapted for lguest's use. + */ add_preferred_console("hvc", 0, NULL); /* Register our very early console. */ virtio_cons_early_init(early_put_chars); - /* Last of all, we set the power management poweroff hook to point to + /* + * Last of all, we set the power management poweroff hook to point to * the Guest routine to power off, and the reboot hook to our restart - * routine. */ + * routine. + */ pm_power_off = lguest_power_off; machine_ops.restart = lguest_restart; - /* Now we're set up, call i386_start_kernel() in head32.c and we proceed - * to boot as normal. It never returns. */ + /* + * Now we're set up, call i386_start_kernel() in head32.c and we proceed + * to boot as normal. It never returns. + */ i386_start_kernel(); } /* diff --git a/arch/x86/lguest/i386_head.S b/arch/x86/lguest/i386_head.S index a9c8cfe61cd4..db6aa95eb054 100644 --- a/arch/x86/lguest/i386_head.S +++ b/arch/x86/lguest/i386_head.S @@ -5,7 +5,8 @@ #include <asm/thread_info.h> #include <asm/processor-flags.h> -/*G:020 Our story starts with the kernel booting into startup_32 in +/*G:020 + * Our story starts with the kernel booting into startup_32 in * arch/x86/kernel/head_32.S. It expects a boot header, which is created by * the bootloader (the Launcher in our case). * @@ -21,11 +22,14 @@ * data without remembering to subtract __PAGE_OFFSET! * * The .section line puts this code in .init.text so it will be discarded after - * boot. */ + * boot. + */ .section .init.text, "ax", @progbits ENTRY(lguest_entry) - /* We make the "initialization" hypercall now to tell the Host about - * us, and also find out where it put our page tables. */ + /* + * We make the "initialization" hypercall now to tell the Host about + * us, and also find out where it put our page tables. + */ movl $LHCALL_LGUEST_INIT, %eax movl $lguest_data - __PAGE_OFFSET, %ebx .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */ @@ -33,13 +37,14 @@ ENTRY(lguest_entry) /* Set up the initial stack so we can run C code. */ movl $(init_thread_union+THREAD_SIZE),%esp - /* Jumps are relative, and we're running __PAGE_OFFSET too low at the - * moment. */ + /* Jumps are relative: we're running __PAGE_OFFSET too low. */ jmp lguest_init+__PAGE_OFFSET -/*G:055 We create a macro which puts the assembler code between lgstart_ and - * lgend_ markers. These templates are put in the .text section: they can't be - * discarded after boot as we may need to patch modules, too. */ +/*G:055 + * We create a macro which puts the assembler code between lgstart_ and lgend_ + * markers. These templates are put in the .text section: they can't be + * discarded after boot as we may need to patch modules, too. + */ .text #define LGUEST_PATCH(name, insns...) \ lgstart_##name: insns; lgend_##name:; \ @@ -48,58 +53,74 @@ ENTRY(lguest_entry) LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled) LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax) -/*G:033 But using those wrappers is inefficient (we'll see why that doesn't - * matter for save_fl and irq_disable later). If we write our routines - * carefully in assembler, we can avoid clobbering any registers and avoid - * jumping through the wrapper functions. +/*G:033 + * But using those wrappers is inefficient (we'll see why that doesn't matter + * for save_fl and irq_disable later). If we write our routines carefully in + * assembler, we can avoid clobbering any registers and avoid jumping through + * the wrapper functions. * * I skipped over our first piece of assembler, but this one is worth studying - * in a bit more detail so I'll describe in easy stages. First, the routine - * to enable interrupts: */ + * in a bit more detail so I'll describe in easy stages. First, the routine to + * enable interrupts: + */ ENTRY(lg_irq_enable) - /* The reverse of irq_disable, this sets lguest_data.irq_enabled to - * X86_EFLAGS_IF (ie. "Interrupts enabled"). */ + /* + * The reverse of irq_disable, this sets lguest_data.irq_enabled to + * X86_EFLAGS_IF (ie. "Interrupts enabled"). + */ movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled - /* But now we need to check if the Host wants to know: there might have + /* + * But now we need to check if the Host wants to know: there might have * been interrupts waiting to be delivered, in which case it will have * set lguest_data.irq_pending to X86_EFLAGS_IF. If it's not zero, we - * jump to send_interrupts, otherwise we're done. */ + * jump to send_interrupts, otherwise we're done. + */ testl $0, lguest_data+LGUEST_DATA_irq_pending jnz send_interrupts - /* One cool thing about x86 is that you can do many things without using + /* + * One cool thing about x86 is that you can do many things without using * a register. In this case, the normal path hasn't needed to save or - * restore any registers at all! */ + * restore any registers at all! + */ ret send_interrupts: - /* OK, now we need a register: eax is used for the hypercall number, + /* + * OK, now we need a register: eax is used for the hypercall number, * which is LHCALL_SEND_INTERRUPTS. * * We used not to bother with this pending detection at all, which was * much simpler. Sooner or later the Host would realize it had to * send us an interrupt. But that turns out to make performance 7 * times worse on a simple tcp benchmark. So now we do this the hard - * way. */ + * way. + */ pushl %eax movl $LHCALL_SEND_INTERRUPTS, %eax - /* This is a vmcall instruction (same thing that KVM uses). Older + /* + * This is a vmcall instruction (same thing that KVM uses). Older * assembler versions might not know the "vmcall" instruction, so we - * create one manually here. */ + * create one manually here. + */ .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */ popl %eax ret -/* Finally, the "popf" or "restore flags" routine. The %eax register holds the +/* + * Finally, the "popf" or "restore flags" routine. The %eax register holds the * flags (in practice, either X86_EFLAGS_IF or 0): if it's X86_EFLAGS_IF we're - * enabling interrupts again, if it's 0 we're leaving them off. */ + * enabling interrupts again, if it's 0 we're leaving them off. + */ ENTRY(lg_restore_fl) /* This is just "lguest_data.irq_enabled = flags;" */ movl %eax, lguest_data+LGUEST_DATA_irq_enabled - /* Now, if the %eax value has enabled interrupts and + /* + * Now, if the %eax value has enabled interrupts and * lguest_data.irq_pending is set, we want to tell the Host so it can * deliver any outstanding interrupts. Fortunately, both values will * be X86_EFLAGS_IF (ie. 512) in that case, and the "testl" * instruction will AND them together for us. If both are set, we - * jump to send_interrupts. */ + * jump to send_interrupts. + */ testl lguest_data+LGUEST_DATA_irq_pending, %eax jnz send_interrupts /* Again, the normal path has used no extra registers. Clever, huh? */ @@ -109,22 +130,24 @@ ENTRY(lg_restore_fl) .global lguest_noirq_start .global lguest_noirq_end -/*M:004 When the Host reflects a trap or injects an interrupt into the Guest, - * it sets the eflags interrupt bit on the stack based on - * lguest_data.irq_enabled, so the Guest iret logic does the right thing when - * restoring it. However, when the Host sets the Guest up for direct traps, - * such as system calls, the processor is the one to push eflags onto the - * stack, and the interrupt bit will be 1 (in reality, interrupts are always - * enabled in the Guest). +/*M:004 + * When the Host reflects a trap or injects an interrupt into the Guest, it + * sets the eflags interrupt bit on the stack based on lguest_data.irq_enabled, + * so the Guest iret logic does the right thing when restoring it. However, + * when the Host sets the Guest up for direct traps, such as system calls, the + * processor is the one to push eflags onto the stack, and the interrupt bit + * will be 1 (in reality, interrupts are always enabled in the Guest). * * This turns out to be harmless: the only trap which should happen under Linux * with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc * regions), which has to be reflected through the Host anyway. If another * trap *does* go off when interrupts are disabled, the Guest will panic, and - * we'll never get to this iret! :*/ + * we'll never get to this iret! +:*/ -/*G:045 There is one final paravirt_op that the Guest implements, and glancing - * at it you can see why I left it to last. It's *cool*! It's in *assembler*! +/*G:045 + * There is one final paravirt_op that the Guest implements, and glancing at it + * you can see why I left it to last. It's *cool*! It's in *assembler*! * * The "iret" instruction is used to return from an interrupt or trap. The * stack looks like this: @@ -148,15 +171,18 @@ ENTRY(lg_restore_fl) * return to userspace or wherever. Our solution to this is to surround the * code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the * Host that it is *never* to interrupt us there, even if interrupts seem to be - * enabled. */ + * enabled. + */ ENTRY(lguest_iret) pushl %eax movl 12(%esp), %eax lguest_noirq_start: - /* Note the %ss: segment prefix here. Normal data accesses use the + /* + * Note the %ss: segment prefix here. Normal data accesses use the * "ds" segment, but that will have already been restored for whatever * we're returning to (such as userspace): we can't trust it. The %ss: - * prefix makes sure we use the stack segment, which is still valid. */ + * prefix makes sure we use the stack segment, which is still valid. + */ movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled popl %eax iret |