diff options
Diffstat (limited to 'arch/x86/lguest/boot.c')
| -rw-r--r-- | arch/x86/lguest/boot.c | 510 |
1 files changed, 350 insertions, 160 deletions
diff --git a/arch/x86/lguest/boot.c b/arch/x86/lguest/boot.c index 7bc65f0f62c4..d677fa9ca650 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) { @@ -146,6 +154,7 @@ static void lazy_hcall1(unsigned long call, async_hcall(call, arg1, 0, 0, 0); } +/* You can imagine what lazy_hcall2, 3 and 4 look like. :*/ static void lazy_hcall2(unsigned long call, unsigned long arg1, unsigned long arg2) @@ -181,8 +190,10 @@ static void lazy_hcall4(unsigned long call, } #endif -/* When lazy mode is turned off reset the per-cpu lazy mode variable and then - * issue the do-nothing hypercall to flush any stored calls. */ +/*G:036 + * When lazy mode is turned off reset the per-cpu lazy mode variable and then + * issue the do-nothing hypercall to flush any stored calls. +:*/ static void lguest_leave_lazy_mmu_mode(void) { kvm_hypercall0(LHCALL_FLUSH_ASYNC); @@ -208,9 +219,11 @@ 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 +235,15 @@ 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 +252,18 @@ 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 + * We could be more efficient in our checking of outstanding interrupts, rather + * than using a branch. 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 +276,12 @@ 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 +289,11 @@ 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 +324,11 @@ 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 +338,36 @@ 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 +375,21 @@ 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 +401,8 @@ 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) { @@ -379,38 +410,63 @@ static void lguest_cpuid(unsigned int *ax, unsigned int *bx, native_cpuid(ax, bx, cx, dx); switch (function) { - case 1: /* Basic feature request. */ - /* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */ + /* + * 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; + + /* + * 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(). @@ -425,7 +481,8 @@ 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) { @@ -438,18 +495,22 @@ 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; @@ -458,10 +519,12 @@ 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; @@ -506,7 +569,7 @@ static void lguest_write_cr4(unsigned long val) * cr3 ---> +---------+ * | --------->+---------+ * | | | PADDR1 | - * Top-level | | PADDR2 | + * Mid-level | | PADDR2 | * (PMD) page | | | * | | Lower-level | * | | (PTE) page | @@ -526,21 +589,62 @@ static void lguest_write_cr4(unsigned long val) * Index into top Index into second Offset within page * page directory page pagetable page * - * The kernel spends a lot of time changing both the top-level page directory - * and lower-level pagetable pages. The Guest doesn't know physical addresses, - * so while it maintains these page tables exactly like normal, it also needs - * to keep the Host informed whenever it makes a change: the Host will create - * the real page tables based on the Guests'. + * Now, unfortunately, this isn't the whole story: Intel added Physical Address + * Extension (PAE) to allow 32 bit systems to use 64GB of memory (ie. 36 bits). + * These are held in 64-bit page table entries, so we can now only fit 512 + * entries in a page, and the neat three-level tree breaks down. + * + * The result is a four level page table: + * + * cr3 --> [ 4 Upper ] + * [ Level ] + * [ Entries ] + * [(PUD Page)]---> +---------+ + * | --------->+---------+ + * | | | PADDR1 | + * Mid-level | | PADDR2 | + * (PMD) page | | | + * | | Lower-level | + * | | (PTE) page | + * | | | | + * .... .... + * + * + * And the virtual address is decoded as: + * + * 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 + * |<-2->|<--- 9 bits ---->|<---- 9 bits --->|<------ 12 bits ------>| + * Index into Index into mid Index into lower Offset within page + * top entries directory page pagetable page + * + * It's too hard to switch between these two formats at runtime, so Linux only + * supports one or the other depending on whether CONFIG_X86_PAE is set. Many + * distributions turn it on, and not just for people with silly amounts of + * memory: the larger PTE entries allow room for the NX bit, which lets the + * kernel disable execution of pages and increase security. + * + * This was a problem for lguest, which couldn't run on these distributions; + * then Matias Zabaljauregui figured it all out and implemented it, and only a + * handful of puppies were crushed in the process! + * + * Back to our point: the kernel spends a lot of time changing both the + * top-level page directory and lower-level pagetable pages. The Guest doesn't + * know physical addresses, so while it maintains these page tables exactly + * like normal, it also needs to keep the Host informed whenever it makes a + * change: the Host will create the real page tables based on the Guests'. */ -/* 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). */ +/* + * The Guest calls this after it has set a second-level entry (pte), ie. to map + * a page into a process' address space. Wetell 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). + */ static void lguest_pte_update(struct mm_struct *mm, unsigned long addr, pte_t *ptep) { #ifdef CONFIG_X86_PAE + /* PAE needs to hand a 64 bit page table entry, so it uses two args. */ lazy_hcall4(LHCALL_SET_PTE, __pa(mm->pgd), addr, ptep->pte_low, ptep->pte_high); #else @@ -548,6 +652,7 @@ static void lguest_pte_update(struct mm_struct *mm, unsigned long addr, #endif } +/* This is the "set and update" combo-meal-deal version. */ static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pteval) { @@ -555,10 +660,13 @@ 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) { @@ -577,8 +685,7 @@ 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); @@ -587,7 +694,8 @@ 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. @@ -595,7 +703,8 @@ 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); @@ -604,6 +713,11 @@ static void lguest_set_pte(pte_t *ptep, pte_t pteval) } #ifdef CONFIG_X86_PAE +/* + * With 64-bit PTE values, we need to be careful setting them: if we set 32 + * bits at a time, the hardware could see a weird half-set entry. These + * versions ensure we update all 64 bits at once. + */ static void lguest_set_pte_atomic(pte_t *ptep, pte_t pte) { native_set_pte_atomic(ptep, pte); @@ -611,19 +725,21 @@ static void lguest_set_pte_atomic(pte_t *ptep, pte_t pte) lazy_hcall1(LHCALL_FLUSH_TLB, 1); } -void lguest_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep) +static void lguest_pte_clear(struct mm_struct *mm, unsigned long addr, + pte_t *ptep) { native_pte_clear(mm, addr, ptep); lguest_pte_update(mm, addr, ptep); } -void lguest_pmd_clear(pmd_t *pmdp) +static void lguest_pmd_clear(pmd_t *pmdp) { lguest_set_pmd(pmdp, __pmd(0)); } #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). @@ -632,24 +748,29 @@ 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); @@ -686,26 +807,38 @@ 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()); } +/* + * With CONFIG_SPARSE_IRQ, interrupt descriptors are allocated as-needed, so + * rather than set them in lguest_init_IRQ we are called here every time an + * lguest device needs an interrupt. + * + * FIXME: irq_to_desc_alloc_node() can fail due to lack of memory, we should + * pass that up! + */ void lguest_setup_irq(unsigned int irq) { irq_to_desc_alloc_node(irq, 0); @@ -724,31 +857,39 @@ 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; @@ -772,9 +913,11 @@ 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) { @@ -824,8 +967,10 @@ 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; @@ -836,10 +981,12 @@ 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 */ @@ -863,14 +1010,16 @@ 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) { @@ -884,7 +1033,8 @@ 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. * @@ -898,11 +1048,13 @@ 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) { @@ -951,11 +1103,13 @@ 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"), @@ -986,8 +1140,10 @@ 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); @@ -996,16 +1152,17 @@ 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'; @@ -1016,8 +1173,10 @@ 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); @@ -1044,7 +1203,8 @@ 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 @@ -1055,9 +1215,11 @@ 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) { @@ -1069,8 +1231,7 @@ 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); @@ -1079,22 +1240,28 @@ static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf, return insn_len; } -/*G:030 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); @@ -1102,11 +1269,11 @@ __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; @@ -1127,7 +1294,7 @@ __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; @@ -1149,54 +1316,71 @@ __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. */ + /* We actually boot with all memory mapped, but let's say 128MB. */ 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); @@ -1213,22 +1397,28 @@ __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(); } /* |