Rusty Russell | f938d2c | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 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." |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 5 | * |
Rusty Russell | f938d2c | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 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/lguest/lguest.c) is called the |
| 11 | * Launcher. |
| 12 | * |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 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. |
Rusty Russell | f938d2c | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 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 | * |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 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. :*/ |
Rusty Russell | f938d2c | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 26 | |
| 27 | /* |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 28 | * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation. |
| 29 | * |
| 30 | * This program is free software; you can redistribute it and/or modify |
| 31 | * it under the terms of the GNU General Public License as published by |
| 32 | * the Free Software Foundation; either version 2 of the License, or |
| 33 | * (at your option) any later version. |
| 34 | * |
| 35 | * This program is distributed in the hope that it will be useful, but |
| 36 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
| 37 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or |
| 38 | * NON INFRINGEMENT. See the GNU General Public License for more |
| 39 | * details. |
| 40 | * |
| 41 | * You should have received a copy of the GNU General Public License |
| 42 | * along with this program; if not, write to the Free Software |
| 43 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
| 44 | */ |
| 45 | #include <linux/kernel.h> |
| 46 | #include <linux/start_kernel.h> |
| 47 | #include <linux/string.h> |
| 48 | #include <linux/console.h> |
| 49 | #include <linux/screen_info.h> |
| 50 | #include <linux/irq.h> |
| 51 | #include <linux/interrupt.h> |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 52 | #include <linux/clocksource.h> |
| 53 | #include <linux/clockchips.h> |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 54 | #include <linux/lguest.h> |
| 55 | #include <linux/lguest_launcher.h> |
Rusty Russell | 19f1537 | 2007-10-22 11:24:21 +1000 | [diff] [blame] | 56 | #include <linux/virtio_console.h> |
Jeff Garzik | 4cfe6c3 | 2007-10-25 14:15:09 +1000 | [diff] [blame] | 57 | #include <linux/pm.h> |
Ingo Molnar | 7b6aa33 | 2009-02-17 13:58:15 +0100 | [diff] [blame] | 58 | #include <asm/apic.h> |
Harvey Harrison | cbc3497 | 2008-02-13 13:14:35 -0800 | [diff] [blame] | 59 | #include <asm/lguest.h> |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 60 | #include <asm/paravirt.h> |
| 61 | #include <asm/param.h> |
| 62 | #include <asm/page.h> |
| 63 | #include <asm/pgtable.h> |
| 64 | #include <asm/desc.h> |
| 65 | #include <asm/setup.h> |
| 66 | #include <asm/e820.h> |
| 67 | #include <asm/mce.h> |
| 68 | #include <asm/io.h> |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 69 | #include <asm/i387.h> |
Rusty Russell | 2cb7878 | 2009-06-03 14:52:24 +0930 | [diff] [blame] | 70 | #include <asm/stackprotector.h> |
Balaji Rao | ec04b13 | 2007-12-28 14:26:24 +0530 | [diff] [blame] | 71 | #include <asm/reboot.h> /* for struct machine_ops */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 72 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 73 | /*G:010 Welcome to the Guest! |
| 74 | * |
| 75 | * The Guest in our tale is a simple creature: identical to the Host but |
| 76 | * behaving in simplified but equivalent ways. In particular, the Guest is the |
| 77 | * same kernel as the Host (or at least, built from the same source code). :*/ |
| 78 | |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 79 | struct lguest_data lguest_data = { |
| 80 | .hcall_status = { [0 ... LHCALL_RING_SIZE-1] = 0xFF }, |
| 81 | .noirq_start = (u32)lguest_noirq_start, |
| 82 | .noirq_end = (u32)lguest_noirq_end, |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 83 | .kernel_address = PAGE_OFFSET, |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 84 | .blocked_interrupts = { 1 }, /* Block timer interrupts */ |
Rusty Russell | c18acd7 | 2007-10-22 11:03:35 +1000 | [diff] [blame] | 85 | .syscall_vec = SYSCALL_VECTOR, |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 86 | }; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 87 | |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 88 | /*G:037 async_hcall() is pretty simple: I'm quite proud of it really. We have a |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 89 | * ring buffer of stored hypercalls which the Host will run though next time we |
Matias Zabaljauregui | cefcad1 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 90 | * do a normal hypercall. Each entry in the ring has 5 slots for the hypercall |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 91 | * arguments, and a "hcall_status" word which is 0 if the call is ready to go, |
| 92 | * and 255 once the Host has finished with it. |
| 93 | * |
| 94 | * If we come around to a slot which hasn't been finished, then the table is |
| 95 | * full and we just make the hypercall directly. This has the nice side |
| 96 | * effect of causing the Host to run all the stored calls in the ring buffer |
| 97 | * which empties it for next time! */ |
Adrian Bunk | 9b56fdb | 2007-11-02 16:43:10 +0100 | [diff] [blame] | 98 | static void async_hcall(unsigned long call, unsigned long arg1, |
Matias Zabaljauregui | cefcad1 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 99 | unsigned long arg2, unsigned long arg3, |
| 100 | unsigned long arg4) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 101 | { |
| 102 | /* Note: This code assumes we're uniprocessor. */ |
| 103 | static unsigned int next_call; |
| 104 | unsigned long flags; |
| 105 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 106 | /* Disable interrupts if not already disabled: we don't want an |
| 107 | * interrupt handler making a hypercall while we're already doing |
| 108 | * one! */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 109 | local_irq_save(flags); |
| 110 | if (lguest_data.hcall_status[next_call] != 0xFF) { |
| 111 | /* Table full, so do normal hcall which will flush table. */ |
Matias Zabaljauregui | cefcad1 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 112 | kvm_hypercall4(call, arg1, arg2, arg3, arg4); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 113 | } else { |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 114 | lguest_data.hcalls[next_call].arg0 = call; |
| 115 | lguest_data.hcalls[next_call].arg1 = arg1; |
| 116 | lguest_data.hcalls[next_call].arg2 = arg2; |
| 117 | lguest_data.hcalls[next_call].arg3 = arg3; |
Matias Zabaljauregui | cefcad1 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 118 | lguest_data.hcalls[next_call].arg4 = arg4; |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 119 | /* Arguments must all be written before we mark it to go */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 120 | wmb(); |
| 121 | lguest_data.hcall_status[next_call] = 0; |
| 122 | if (++next_call == LHCALL_RING_SIZE) |
| 123 | next_call = 0; |
| 124 | } |
| 125 | local_irq_restore(flags); |
| 126 | } |
Adrian Bunk | 9b56fdb | 2007-11-02 16:43:10 +0100 | [diff] [blame] | 127 | |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 128 | /*G:035 Notice the lazy_hcall() above, rather than hcall(). This is our first |
| 129 | * real optimization trick! |
| 130 | * |
| 131 | * When lazy_mode is set, it means we're allowed to defer all hypercalls and do |
| 132 | * them as a batch when lazy_mode is eventually turned off. Because hypercalls |
| 133 | * are reasonably expensive, batching them up makes sense. For example, a |
| 134 | * large munmap might update dozens of page table entries: that code calls |
| 135 | * paravirt_enter_lazy_mmu(), does the dozen updates, then calls |
| 136 | * lguest_leave_lazy_mode(). |
| 137 | * |
| 138 | * So, when we're in lazy mode, we call async_hcall() to store the call for |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 139 | * future processing: */ |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 140 | static void lazy_hcall1(unsigned long call, |
| 141 | unsigned long arg1) |
| 142 | { |
| 143 | if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE) |
| 144 | kvm_hypercall1(call, arg1); |
| 145 | else |
Matias Zabaljauregui | cefcad1 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 146 | async_hcall(call, arg1, 0, 0, 0); |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 147 | } |
| 148 | |
| 149 | static void lazy_hcall2(unsigned long call, |
| 150 | unsigned long arg1, |
| 151 | unsigned long arg2) |
| 152 | { |
| 153 | if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE) |
| 154 | kvm_hypercall2(call, arg1, arg2); |
| 155 | else |
Matias Zabaljauregui | cefcad1 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 156 | async_hcall(call, arg1, arg2, 0, 0); |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 157 | } |
| 158 | |
| 159 | static void lazy_hcall3(unsigned long call, |
Adrian Bunk | 9b56fdb | 2007-11-02 16:43:10 +0100 | [diff] [blame] | 160 | unsigned long arg1, |
| 161 | unsigned long arg2, |
| 162 | unsigned long arg3) |
| 163 | { |
| 164 | if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE) |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 165 | kvm_hypercall3(call, arg1, arg2, arg3); |
Adrian Bunk | 9b56fdb | 2007-11-02 16:43:10 +0100 | [diff] [blame] | 166 | else |
Matias Zabaljauregui | cefcad1 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 167 | async_hcall(call, arg1, arg2, arg3, 0); |
| 168 | } |
| 169 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 170 | #ifdef CONFIG_X86_PAE |
Matias Zabaljauregui | cefcad1 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 171 | static void lazy_hcall4(unsigned long call, |
| 172 | unsigned long arg1, |
| 173 | unsigned long arg2, |
| 174 | unsigned long arg3, |
| 175 | unsigned long arg4) |
| 176 | { |
| 177 | if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE) |
| 178 | kvm_hypercall4(call, arg1, arg2, arg3, arg4); |
| 179 | else |
| 180 | async_hcall(call, arg1, arg2, arg3, arg4); |
Adrian Bunk | 9b56fdb | 2007-11-02 16:43:10 +0100 | [diff] [blame] | 181 | } |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 182 | #endif |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 183 | |
| 184 | /* When lazy mode is turned off reset the per-cpu lazy mode variable and then |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 185 | * issue the do-nothing hypercall to flush any stored calls. */ |
Jeremy Fitzhardinge | b407fc5 | 2009-02-17 23:46:21 -0800 | [diff] [blame] | 186 | static void lguest_leave_lazy_mmu_mode(void) |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 187 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 188 | kvm_hypercall0(LHCALL_FLUSH_ASYNC); |
Jeremy Fitzhardinge | b407fc5 | 2009-02-17 23:46:21 -0800 | [diff] [blame] | 189 | paravirt_leave_lazy_mmu(); |
| 190 | } |
| 191 | |
Jeremy Fitzhardinge | 224101e | 2009-02-18 11:18:57 -0800 | [diff] [blame] | 192 | static void lguest_end_context_switch(struct task_struct *next) |
Jeremy Fitzhardinge | b407fc5 | 2009-02-17 23:46:21 -0800 | [diff] [blame] | 193 | { |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 194 | kvm_hypercall0(LHCALL_FLUSH_ASYNC); |
Jeremy Fitzhardinge | 224101e | 2009-02-18 11:18:57 -0800 | [diff] [blame] | 195 | paravirt_end_context_switch(next); |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 196 | } |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 197 | |
Rusty Russell | 61f4bc8 | 2009-06-12 22:27:03 -0600 | [diff] [blame] | 198 | /*G:032 |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 199 | * After that diversion we return to our first native-instruction |
| 200 | * replacements: four functions for interrupt control. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 201 | * |
| 202 | * The simplest way of implementing these would be to have "turn interrupts |
| 203 | * off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow: |
| 204 | * these are by far the most commonly called functions of those we override. |
| 205 | * |
| 206 | * So instead we keep an "irq_enabled" field inside our "struct lguest_data", |
| 207 | * which the Guest can update with a single instruction. The Host knows to |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 208 | * check there before it tries to deliver an interrupt. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 209 | */ |
| 210 | |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 211 | /* save_flags() is expected to return the processor state (ie. "flags"). The |
| 212 | * flags word contains all kind of stuff, but in practice Linux only cares |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 213 | * about the interrupt flag. Our "save_flags()" just returns that. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 214 | static unsigned long save_fl(void) |
| 215 | { |
| 216 | return lguest_data.irq_enabled; |
| 217 | } |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 218 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 219 | /* Interrupts go off... */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 220 | static void irq_disable(void) |
| 221 | { |
| 222 | lguest_data.irq_enabled = 0; |
| 223 | } |
Rusty Russell | 61f4bc8 | 2009-06-12 22:27:03 -0600 | [diff] [blame] | 224 | |
| 225 | /* Let's pause a moment. Remember how I said these are called so often? |
| 226 | * Jeremy Fitzhardinge optimized them so hard early in 2009 that he had to |
| 227 | * break some rules. In particular, these functions are assumed to save their |
| 228 | * own registers if they need to: normal C functions assume they can trash the |
| 229 | * eax register. To use normal C functions, we use |
| 230 | * PV_CALLEE_SAVE_REGS_THUNK(), which pushes %eax onto the stack, calls the |
| 231 | * C function, then restores it. */ |
| 232 | PV_CALLEE_SAVE_REGS_THUNK(save_fl); |
Jeremy Fitzhardinge | ecb93d1 | 2009-01-28 14:35:05 -0800 | [diff] [blame] | 233 | PV_CALLEE_SAVE_REGS_THUNK(irq_disable); |
Rusty Russell | f56a384 | 2007-07-26 10:41:05 -0700 | [diff] [blame] | 234 | /*:*/ |
Rusty Russell | 61f4bc8 | 2009-06-12 22:27:03 -0600 | [diff] [blame] | 235 | |
| 236 | /* These are in i386_head.S */ |
| 237 | extern void lg_irq_enable(void); |
| 238 | extern void lg_restore_fl(unsigned long flags); |
| 239 | |
Rusty Russell | f56a384 | 2007-07-26 10:41:05 -0700 | [diff] [blame] | 240 | /*M:003 Note that we don't check for outstanding interrupts when we re-enable |
| 241 | * them (or when we unmask an interrupt). This seems to work for the moment, |
| 242 | * since interrupts are rare and we'll just get the interrupt on the next timer |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 243 | * tick, but now we can run with CONFIG_NO_HZ, we should revisit this. One way |
Rusty Russell | f56a384 | 2007-07-26 10:41:05 -0700 | [diff] [blame] | 244 | * would be to put the "irq_enabled" field in a page by itself, and have the |
| 245 | * Host write-protect it when an interrupt comes in when irqs are disabled. |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 246 | * There will then be a page fault as soon as interrupts are re-enabled. |
| 247 | * |
| 248 | * A better method is to implement soft interrupt disable generally for x86: |
| 249 | * instead of disabling interrupts, we set a flag. If an interrupt does come |
| 250 | * in, we then disable them for real. This is uncommon, so we could simply use |
| 251 | * a hypercall for interrupt control and not worry about efficiency. :*/ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 252 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 253 | /*G:034 |
| 254 | * The Interrupt Descriptor Table (IDT). |
| 255 | * |
| 256 | * The IDT tells the processor what to do when an interrupt comes in. Each |
| 257 | * entry in the table is a 64-bit descriptor: this holds the privilege level, |
| 258 | * address of the handler, and... well, who cares? The Guest just asks the |
| 259 | * Host to make the change anyway, because the Host controls the real IDT. |
| 260 | */ |
Glauber de Oliveira Costa | 8d94734 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 261 | static void lguest_write_idt_entry(gate_desc *dt, |
| 262 | int entrynum, const gate_desc *g) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 263 | { |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 264 | /* The gate_desc structure is 8 bytes long: we hand it to the Host in |
| 265 | * two 32-bit chunks. The whole 32-bit kernel used to hand descriptors |
| 266 | * around like this; typesafety wasn't a big concern in Linux's early |
| 267 | * years. */ |
Glauber de Oliveira Costa | 8d94734 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 268 | u32 *desc = (u32 *)g; |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 269 | /* Keep the local copy up to date. */ |
Glauber de Oliveira Costa | 8d94734 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 270 | native_write_idt_entry(dt, entrynum, g); |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 271 | /* Tell Host about this new entry. */ |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 272 | kvm_hypercall3(LHCALL_LOAD_IDT_ENTRY, entrynum, desc[0], desc[1]); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 273 | } |
| 274 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 275 | /* Changing to a different IDT is very rare: we keep the IDT up-to-date every |
| 276 | * time it is written, so we can simply loop through all entries and tell the |
| 277 | * Host about them. */ |
Glauber de Oliveira Costa | 6b68f01 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 278 | static void lguest_load_idt(const struct desc_ptr *desc) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 279 | { |
| 280 | unsigned int i; |
| 281 | struct desc_struct *idt = (void *)desc->address; |
| 282 | |
| 283 | for (i = 0; i < (desc->size+1)/8; i++) |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 284 | kvm_hypercall3(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 285 | } |
| 286 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 287 | /* |
| 288 | * The Global Descriptor Table. |
| 289 | * |
| 290 | * The Intel architecture defines another table, called the Global Descriptor |
| 291 | * Table (GDT). You tell the CPU where it is (and its size) using the "lgdt" |
| 292 | * instruction, and then several other instructions refer to entries in the |
| 293 | * table. There are three entries which the Switcher needs, so the Host simply |
| 294 | * controls the entire thing and the Guest asks it to make changes using the |
| 295 | * LOAD_GDT hypercall. |
| 296 | * |
Rusty Russell | a489f0b | 2009-04-19 23:14:00 -0600 | [diff] [blame] | 297 | * This is the exactly like the IDT code. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 298 | */ |
Glauber de Oliveira Costa | 6b68f01 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 299 | static void lguest_load_gdt(const struct desc_ptr *desc) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 300 | { |
Rusty Russell | a489f0b | 2009-04-19 23:14:00 -0600 | [diff] [blame] | 301 | unsigned int i; |
| 302 | struct desc_struct *gdt = (void *)desc->address; |
| 303 | |
| 304 | for (i = 0; i < (desc->size+1)/8; i++) |
| 305 | kvm_hypercall3(LHCALL_LOAD_GDT_ENTRY, i, gdt[i].a, gdt[i].b); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 306 | } |
| 307 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 308 | /* For a single GDT entry which changes, we do the lazy thing: alter our GDT, |
| 309 | * then tell the Host to reload the entire thing. This operation is so rare |
| 310 | * that this naive implementation is reasonable. */ |
Glauber de Oliveira Costa | 014b15b | 2008-01-30 13:31:13 +0100 | [diff] [blame] | 311 | static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum, |
| 312 | const void *desc, int type) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 313 | { |
Glauber de Oliveira Costa | 014b15b | 2008-01-30 13:31:13 +0100 | [diff] [blame] | 314 | native_write_gdt_entry(dt, entrynum, desc, type); |
Rusty Russell | a489f0b | 2009-04-19 23:14:00 -0600 | [diff] [blame] | 315 | /* Tell Host about this new entry. */ |
| 316 | kvm_hypercall3(LHCALL_LOAD_GDT_ENTRY, entrynum, |
| 317 | dt[entrynum].a, dt[entrynum].b); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 318 | } |
| 319 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 320 | /* OK, I lied. There are three "thread local storage" GDT entries which change |
| 321 | * on every context switch (these three entries are how glibc implements |
| 322 | * __thread variables). So we have a hypercall specifically for this case. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 323 | static void lguest_load_tls(struct thread_struct *t, unsigned int cpu) |
| 324 | { |
Rusty Russell | 0d027c0 | 2007-08-09 20:57:13 +1000 | [diff] [blame] | 325 | /* There's one problem which normal hardware doesn't have: the Host |
| 326 | * can't handle us removing entries we're currently using. So we clear |
| 327 | * the GS register here: if it's needed it'll be reloaded anyway. */ |
Tejun Heo | ccbeed3 | 2009-02-09 22:17:40 +0900 | [diff] [blame] | 328 | lazy_load_gs(0); |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 329 | lazy_hcall2(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 330 | } |
| 331 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 332 | /*G:038 That's enough excitement for now, back to ploughing through each of |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 333 | * the different pv_ops structures (we're about 1/3 of the way through). |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 334 | * |
| 335 | * This is the Local Descriptor Table, another weird Intel thingy. Linux only |
| 336 | * uses this for some strange applications like Wine. We don't do anything |
| 337 | * here, so they'll get an informative and friendly Segmentation Fault. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 338 | static void lguest_set_ldt(const void *addr, unsigned entries) |
| 339 | { |
| 340 | } |
| 341 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 342 | /* This loads a GDT entry into the "Task Register": that entry points to a |
| 343 | * structure called the Task State Segment. Some comments scattered though the |
| 344 | * kernel code indicate that this used for task switching in ages past, along |
| 345 | * with blood sacrifice and astrology. |
| 346 | * |
| 347 | * Now there's nothing interesting in here that we don't get told elsewhere. |
| 348 | * But the native version uses the "ltr" instruction, which makes the Host |
| 349 | * complain to the Guest about a Segmentation Fault and it'll oops. So we |
| 350 | * override the native version with a do-nothing version. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 351 | static void lguest_load_tr_desc(void) |
| 352 | { |
| 353 | } |
| 354 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 355 | /* The "cpuid" instruction is a way of querying both the CPU identity |
| 356 | * (manufacturer, model, etc) and its features. It was introduced before the |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 357 | * Pentium in 1993 and keeps getting extended by both Intel, AMD and others. |
| 358 | * As you might imagine, after a decade and a half this treatment, it is now a |
| 359 | * giant ball of hair. Its entry in the current Intel manual runs to 28 pages. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 360 | * |
| 361 | * This instruction even it has its own Wikipedia entry. The Wikipedia entry |
| 362 | * has been translated into 4 languages. I am not making this up! |
| 363 | * |
| 364 | * We could get funky here and identify ourselves as "GenuineLguest", but |
| 365 | * instead we just use the real "cpuid" instruction. Then I pretty much turned |
| 366 | * off feature bits until the Guest booted. (Don't say that: you'll damage |
| 367 | * lguest sales!) Shut up, inner voice! (Hey, just pointing out that this is |
| 368 | * hardly future proof.) Noone's listening! They don't like you anyway, |
| 369 | * parenthetic weirdo! |
| 370 | * |
| 371 | * Replacing the cpuid so we can turn features off is great for the kernel, but |
| 372 | * anyone (including userspace) can just use the raw "cpuid" instruction and |
| 373 | * the Host won't even notice since it isn't privileged. So we try not to get |
| 374 | * too worked up about it. */ |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 375 | static void lguest_cpuid(unsigned int *ax, unsigned int *bx, |
| 376 | unsigned int *cx, unsigned int *dx) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 377 | { |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 378 | int function = *ax; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 379 | |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 380 | native_cpuid(ax, bx, cx, dx); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 381 | switch (function) { |
| 382 | case 1: /* Basic feature request. */ |
| 383 | /* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */ |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 384 | *cx &= 0x00002201; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 385 | /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU, PAE. */ |
| 386 | *dx &= 0x07808151; |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 387 | /* The Host can do a nice optimization if it knows that the |
| 388 | * kernel mappings (addresses above 0xC0000000 or whatever |
| 389 | * PAGE_OFFSET is set to) haven't changed. But Linux calls |
| 390 | * flush_tlb_user() for both user and kernel mappings unless |
| 391 | * the Page Global Enable (PGE) feature bit is set. */ |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 392 | *dx |= 0x00002000; |
Rusty Russell | cbd88c8 | 2009-03-09 10:06:22 -0600 | [diff] [blame] | 393 | /* We also lie, and say we're family id 5. 6 or greater |
| 394 | * leads to a rdmsr in early_init_intel which we can't handle. |
| 395 | * Family ID is returned as bits 8-12 in ax. */ |
| 396 | *ax &= 0xFFFFF0FF; |
| 397 | *ax |= 0x00000500; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 398 | break; |
| 399 | case 0x80000000: |
| 400 | /* Futureproof this a little: if they ask how much extended |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 401 | * processor information there is, limit it to known fields. */ |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 402 | if (*ax > 0x80000008) |
| 403 | *ax = 0x80000008; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 404 | break; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 405 | case 0x80000001: |
| 406 | /* Here we should fix nx cap depending on host. */ |
| 407 | /* For this version of PAE, we just clear NX bit. */ |
| 408 | *dx &= ~(1 << 20); |
| 409 | break; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 410 | } |
| 411 | } |
| 412 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 413 | /* Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4. |
| 414 | * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother |
| 415 | * it. The Host needs to know when the Guest wants to change them, so we have |
| 416 | * a whole series of functions like read_cr0() and write_cr0(). |
| 417 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 418 | * We start with cr0. cr0 allows you to turn on and off all kinds of basic |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 419 | * features, but Linux only really cares about one: the horrifically-named Task |
| 420 | * Switched (TS) bit at bit 3 (ie. 8) |
| 421 | * |
| 422 | * What does the TS bit do? Well, it causes the CPU to trap (interrupt 7) if |
| 423 | * the floating point unit is used. Which allows us to restore FPU state |
| 424 | * lazily after a task switch, and Linux uses that gratefully, but wouldn't a |
| 425 | * name like "FPUTRAP bit" be a little less cryptic? |
| 426 | * |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 427 | * We store cr0 locally because the Host never changes it. The Guest sometimes |
| 428 | * wants to read it and we'd prefer not to bother the Host unnecessarily. */ |
| 429 | static unsigned long current_cr0; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 430 | static void lguest_write_cr0(unsigned long val) |
| 431 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 432 | lazy_hcall1(LHCALL_TS, val & X86_CR0_TS); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 433 | current_cr0 = val; |
| 434 | } |
| 435 | |
| 436 | static unsigned long lguest_read_cr0(void) |
| 437 | { |
| 438 | return current_cr0; |
| 439 | } |
| 440 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 441 | /* Intel provided a special instruction to clear the TS bit for people too cool |
| 442 | * to use write_cr0() to do it. This "clts" instruction is faster, because all |
| 443 | * the vowels have been optimized out. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 444 | static void lguest_clts(void) |
| 445 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 446 | lazy_hcall1(LHCALL_TS, 0); |
Rusty Russell | 25c47bb | 2007-10-25 14:09:53 +1000 | [diff] [blame] | 447 | current_cr0 &= ~X86_CR0_TS; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 448 | } |
| 449 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 450 | /* cr2 is the virtual address of the last page fault, which the Guest only ever |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 451 | * reads. The Host kindly writes this into our "struct lguest_data", so we |
| 452 | * just read it out of there. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 453 | static unsigned long lguest_read_cr2(void) |
| 454 | { |
| 455 | return lguest_data.cr2; |
| 456 | } |
| 457 | |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 458 | /* See lguest_set_pte() below. */ |
| 459 | static bool cr3_changed = false; |
| 460 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 461 | /* cr3 is the current toplevel pagetable page: the principle is the same as |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 462 | * cr0. Keep a local copy, and tell the Host when it changes. The only |
| 463 | * difference is that our local copy is in lguest_data because the Host needs |
| 464 | * to set it upon our initial hypercall. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 465 | static void lguest_write_cr3(unsigned long cr3) |
| 466 | { |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 467 | lguest_data.pgdir = cr3; |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 468 | lazy_hcall1(LHCALL_NEW_PGTABLE, cr3); |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 469 | cr3_changed = true; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 470 | } |
| 471 | |
| 472 | static unsigned long lguest_read_cr3(void) |
| 473 | { |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 474 | return lguest_data.pgdir; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 475 | } |
| 476 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 477 | /* cr4 is used to enable and disable PGE, but we don't care. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 478 | static unsigned long lguest_read_cr4(void) |
| 479 | { |
| 480 | return 0; |
| 481 | } |
| 482 | |
| 483 | static void lguest_write_cr4(unsigned long val) |
| 484 | { |
| 485 | } |
| 486 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 487 | /* |
| 488 | * Page Table Handling. |
| 489 | * |
| 490 | * Now would be a good time to take a rest and grab a coffee or similarly |
| 491 | * relaxing stimulant. The easy parts are behind us, and the trek gradually |
| 492 | * winds uphill from here. |
| 493 | * |
| 494 | * Quick refresher: memory is divided into "pages" of 4096 bytes each. The CPU |
| 495 | * maps virtual addresses to physical addresses using "page tables". We could |
| 496 | * use one huge index of 1 million entries: each address is 4 bytes, so that's |
| 497 | * 1024 pages just to hold the page tables. But since most virtual addresses |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 498 | * are unused, we use a two level index which saves space. The cr3 register |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 499 | * contains the physical address of the top level "page directory" page, which |
| 500 | * contains physical addresses of up to 1024 second-level pages. Each of these |
| 501 | * second level pages contains up to 1024 physical addresses of actual pages, |
| 502 | * or Page Table Entries (PTEs). |
| 503 | * |
| 504 | * Here's a diagram, where arrows indicate physical addresses: |
| 505 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 506 | * cr3 ---> +---------+ |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 507 | * | --------->+---------+ |
| 508 | * | | | PADDR1 | |
| 509 | * Top-level | | PADDR2 | |
| 510 | * (PMD) page | | | |
| 511 | * | | Lower-level | |
| 512 | * | | (PTE) page | |
| 513 | * | | | | |
| 514 | * .... .... |
| 515 | * |
| 516 | * So to convert a virtual address to a physical address, we look up the top |
| 517 | * level, which points us to the second level, which gives us the physical |
| 518 | * address of that page. If the top level entry was not present, or the second |
| 519 | * level entry was not present, then the virtual address is invalid (we |
| 520 | * say "the page was not mapped"). |
| 521 | * |
| 522 | * Put another way, a 32-bit virtual address is divided up like so: |
| 523 | * |
| 524 | * 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 |
| 525 | * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>| |
| 526 | * Index into top Index into second Offset within page |
| 527 | * page directory page pagetable page |
| 528 | * |
| 529 | * The kernel spends a lot of time changing both the top-level page directory |
| 530 | * and lower-level pagetable pages. The Guest doesn't know physical addresses, |
| 531 | * so while it maintains these page tables exactly like normal, it also needs |
| 532 | * to keep the Host informed whenever it makes a change: the Host will create |
| 533 | * the real page tables based on the Guests'. |
| 534 | */ |
| 535 | |
| 536 | /* The Guest calls this to set a second-level entry (pte), ie. to map a page |
| 537 | * into a process' address space. We set the entry then tell the Host the |
| 538 | * toplevel and address this corresponds to. The Guest uses one pagetable per |
| 539 | * process, so we need to tell the Host which one we're changing (mm->pgd). */ |
Rusty Russell | b7ff99e | 2009-03-30 21:55:23 -0600 | [diff] [blame] | 540 | static void lguest_pte_update(struct mm_struct *mm, unsigned long addr, |
| 541 | pte_t *ptep) |
| 542 | { |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 543 | #ifdef CONFIG_X86_PAE |
| 544 | lazy_hcall4(LHCALL_SET_PTE, __pa(mm->pgd), addr, |
| 545 | ptep->pte_low, ptep->pte_high); |
| 546 | #else |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 547 | lazy_hcall3(LHCALL_SET_PTE, __pa(mm->pgd), addr, ptep->pte_low); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 548 | #endif |
Rusty Russell | b7ff99e | 2009-03-30 21:55:23 -0600 | [diff] [blame] | 549 | } |
| 550 | |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 551 | static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr, |
| 552 | pte_t *ptep, pte_t pteval) |
| 553 | { |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 554 | native_set_pte(ptep, pteval); |
Rusty Russell | b7ff99e | 2009-03-30 21:55:23 -0600 | [diff] [blame] | 555 | lguest_pte_update(mm, addr, ptep); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 556 | } |
| 557 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 558 | /* The Guest calls lguest_set_pud to set a top-level entry and lguest_set_pmd |
| 559 | * to set a middle-level entry when PAE is activated. |
| 560 | * Again, we set the entry then tell the Host which page we changed, |
| 561 | * and the index of the entry we changed. */ |
| 562 | #ifdef CONFIG_X86_PAE |
| 563 | static void lguest_set_pud(pud_t *pudp, pud_t pudval) |
| 564 | { |
| 565 | native_set_pud(pudp, pudval); |
| 566 | |
| 567 | /* 32 bytes aligned pdpt address and the index. */ |
| 568 | lazy_hcall2(LHCALL_SET_PGD, __pa(pudp) & 0xFFFFFFE0, |
| 569 | (__pa(pudp) & 0x1F) / sizeof(pud_t)); |
| 570 | } |
| 571 | |
| 572 | static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) |
| 573 | { |
| 574 | native_set_pmd(pmdp, pmdval); |
| 575 | lazy_hcall2(LHCALL_SET_PMD, __pa(pmdp) & PAGE_MASK, |
| 576 | (__pa(pmdp) & (PAGE_SIZE - 1)) / sizeof(pmd_t)); |
| 577 | } |
| 578 | #else |
| 579 | |
| 580 | /* The Guest calls lguest_set_pmd to set a top-level entry when PAE is not |
| 581 | * activated. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 582 | static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) |
| 583 | { |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 584 | native_set_pmd(pmdp, pmdval); |
Matias Zabaljauregui | ebe0ba8 | 2009-05-30 15:48:08 -0300 | [diff] [blame] | 585 | lazy_hcall2(LHCALL_SET_PGD, __pa(pmdp) & PAGE_MASK, |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 586 | (__pa(pmdp) & (PAGE_SIZE - 1)) / sizeof(pmd_t)); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 587 | } |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 588 | #endif |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 589 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 590 | /* There are a couple of legacy places where the kernel sets a PTE, but we |
| 591 | * don't know the top level any more. This is useless for us, since we don't |
| 592 | * know which pagetable is changing or what address, so we just tell the Host |
| 593 | * to forget all of them. Fortunately, this is very rare. |
| 594 | * |
| 595 | * ... except in early boot when the kernel sets up the initial pagetables, |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 596 | * which makes booting astonishingly slow: 1.83 seconds! So we don't even tell |
| 597 | * the Host anything changed until we've done the first page table switch, |
| 598 | * which brings boot back to 0.25 seconds. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 599 | static void lguest_set_pte(pte_t *ptep, pte_t pteval) |
| 600 | { |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 601 | native_set_pte(ptep, pteval); |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 602 | if (cr3_changed) |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 603 | lazy_hcall1(LHCALL_FLUSH_TLB, 1); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 604 | } |
| 605 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 606 | #ifdef CONFIG_X86_PAE |
| 607 | static void lguest_set_pte_atomic(pte_t *ptep, pte_t pte) |
| 608 | { |
| 609 | native_set_pte_atomic(ptep, pte); |
| 610 | if (cr3_changed) |
| 611 | lazy_hcall1(LHCALL_FLUSH_TLB, 1); |
| 612 | } |
| 613 | |
| 614 | void lguest_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep) |
| 615 | { |
| 616 | native_pte_clear(mm, addr, ptep); |
| 617 | lguest_pte_update(mm, addr, ptep); |
| 618 | } |
| 619 | |
| 620 | void lguest_pmd_clear(pmd_t *pmdp) |
| 621 | { |
| 622 | lguest_set_pmd(pmdp, __pmd(0)); |
| 623 | } |
| 624 | #endif |
| 625 | |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 626 | /* Unfortunately for Lguest, the pv_mmu_ops for page tables were based on |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 627 | * native page table operations. On native hardware you can set a new page |
| 628 | * table entry whenever you want, but if you want to remove one you have to do |
| 629 | * a TLB flush (a TLB is a little cache of page table entries kept by the CPU). |
| 630 | * |
| 631 | * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only |
| 632 | * called when a valid entry is written, not when it's removed (ie. marked not |
| 633 | * present). Instead, this is where we come when the Guest wants to remove a |
| 634 | * page table entry: we tell the Host to set that entry to 0 (ie. the present |
| 635 | * bit is zero). */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 636 | static void lguest_flush_tlb_single(unsigned long addr) |
| 637 | { |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 638 | /* Simply set it to zero: if it was not, it will fault back in. */ |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 639 | lazy_hcall3(LHCALL_SET_PTE, lguest_data.pgdir, addr, 0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 640 | } |
| 641 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 642 | /* This is what happens after the Guest has removed a large number of entries. |
| 643 | * This tells the Host that any of the page table entries for userspace might |
| 644 | * have changed, ie. virtual addresses below PAGE_OFFSET. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 645 | static void lguest_flush_tlb_user(void) |
| 646 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 647 | lazy_hcall1(LHCALL_FLUSH_TLB, 0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 648 | } |
| 649 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 650 | /* This is called when the kernel page tables have changed. That's not very |
| 651 | * common (unless the Guest is using highmem, which makes the Guest extremely |
| 652 | * slow), so it's worth separating this from the user flushing above. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 653 | static void lguest_flush_tlb_kernel(void) |
| 654 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 655 | lazy_hcall1(LHCALL_FLUSH_TLB, 1); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 656 | } |
| 657 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 658 | /* |
| 659 | * The Unadvanced Programmable Interrupt Controller. |
| 660 | * |
| 661 | * This is an attempt to implement the simplest possible interrupt controller. |
| 662 | * I spent some time looking though routines like set_irq_chip_and_handler, |
| 663 | * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and |
| 664 | * I *think* this is as simple as it gets. |
| 665 | * |
| 666 | * We can tell the Host what interrupts we want blocked ready for using the |
| 667 | * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as |
| 668 | * simple as setting a bit. We don't actually "ack" interrupts as such, we |
| 669 | * just mask and unmask them. I wonder if we should be cleverer? |
| 670 | */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 671 | static void disable_lguest_irq(unsigned int irq) |
| 672 | { |
| 673 | set_bit(irq, lguest_data.blocked_interrupts); |
| 674 | } |
| 675 | |
| 676 | static void enable_lguest_irq(unsigned int irq) |
| 677 | { |
| 678 | clear_bit(irq, lguest_data.blocked_interrupts); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 679 | } |
| 680 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 681 | /* This structure describes the lguest IRQ controller. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 682 | static struct irq_chip lguest_irq_controller = { |
| 683 | .name = "lguest", |
| 684 | .mask = disable_lguest_irq, |
| 685 | .mask_ack = disable_lguest_irq, |
| 686 | .unmask = enable_lguest_irq, |
| 687 | }; |
| 688 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 689 | /* This sets up the Interrupt Descriptor Table (IDT) entry for each hardware |
| 690 | * interrupt (except 128, which is used for system calls), and then tells the |
| 691 | * Linux infrastructure that each interrupt is controlled by our level-based |
| 692 | * lguest interrupt controller. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 693 | static void __init lguest_init_IRQ(void) |
| 694 | { |
| 695 | unsigned int i; |
| 696 | |
Rusty Russell | 1028375 | 2009-06-12 22:26:59 -0600 | [diff] [blame] | 697 | for (i = FIRST_EXTERNAL_VECTOR; i < NR_VECTORS; i++) { |
Rusty Russell | 526e5ab | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 698 | /* Some systems map "vectors" to interrupts weirdly. Lguest has |
| 699 | * a straightforward 1 to 1 mapping, so force that here. */ |
Rusty Russell | 1028375 | 2009-06-12 22:26:59 -0600 | [diff] [blame] | 700 | __get_cpu_var(vector_irq)[i] = i - FIRST_EXTERNAL_VECTOR; |
| 701 | if (i != SYSCALL_VECTOR) |
| 702 | set_intr_gate(i, interrupt[i - FIRST_EXTERNAL_VECTOR]); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 703 | } |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 704 | /* This call is required to set up for 4k stacks, where we have |
| 705 | * separate stacks for hard and soft interrupts. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 706 | irq_ctx_init(smp_processor_id()); |
| 707 | } |
| 708 | |
Rusty Russell | 6db6a5f | 2009-03-09 10:06:28 -0600 | [diff] [blame] | 709 | void lguest_setup_irq(unsigned int irq) |
| 710 | { |
Yinghai Lu | 85ac16d | 2009-04-27 18:00:38 -0700 | [diff] [blame] | 711 | irq_to_desc_alloc_node(irq, 0); |
Rusty Russell | 6db6a5f | 2009-03-09 10:06:28 -0600 | [diff] [blame] | 712 | set_irq_chip_and_handler_name(irq, &lguest_irq_controller, |
| 713 | handle_level_irq, "level"); |
| 714 | } |
| 715 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 716 | /* |
| 717 | * Time. |
| 718 | * |
| 719 | * It would be far better for everyone if the Guest had its own clock, but |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 720 | * until then the Host gives us the time on every interrupt. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 721 | */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 722 | static unsigned long lguest_get_wallclock(void) |
| 723 | { |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 724 | return lguest_data.time.tv_sec; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 725 | } |
| 726 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 727 | /* The TSC is an Intel thing called the Time Stamp Counter. The Host tells us |
| 728 | * what speed it runs at, or 0 if it's unusable as a reliable clock source. |
| 729 | * This matches what we want here: if we return 0 from this function, the x86 |
| 730 | * TSC clock will give up and not register itself. */ |
Alok Kataria | e93ef94 | 2008-07-01 11:43:36 -0700 | [diff] [blame] | 731 | static unsigned long lguest_tsc_khz(void) |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 732 | { |
| 733 | return lguest_data.tsc_khz; |
| 734 | } |
| 735 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 736 | /* If we can't use the TSC, the kernel falls back to our lower-priority |
| 737 | * "lguest_clock", where we read the time value given to us by the Host. */ |
Magnus Damm | 8e19608 | 2009-04-21 12:24:00 -0700 | [diff] [blame] | 738 | static cycle_t lguest_clock_read(struct clocksource *cs) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 739 | { |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 740 | unsigned long sec, nsec; |
| 741 | |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 742 | /* Since the time is in two parts (seconds and nanoseconds), we risk |
| 743 | * reading it just as it's changing from 99 & 0.999999999 to 100 and 0, |
| 744 | * and getting 99 and 0. As Linux tends to come apart under the stress |
| 745 | * of time travel, we must be careful: */ |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 746 | do { |
| 747 | /* First we read the seconds part. */ |
| 748 | sec = lguest_data.time.tv_sec; |
| 749 | /* This read memory barrier tells the compiler and the CPU that |
| 750 | * this can't be reordered: we have to complete the above |
| 751 | * before going on. */ |
| 752 | rmb(); |
| 753 | /* Now we read the nanoseconds part. */ |
| 754 | nsec = lguest_data.time.tv_nsec; |
| 755 | /* Make sure we've done that. */ |
| 756 | rmb(); |
| 757 | /* Now if the seconds part has changed, try again. */ |
| 758 | } while (unlikely(lguest_data.time.tv_sec != sec)); |
| 759 | |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 760 | /* Our lguest clock is in real nanoseconds. */ |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 761 | return sec*1000000000ULL + nsec; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 762 | } |
| 763 | |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 764 | /* This is the fallback clocksource: lower priority than the TSC clocksource. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 765 | static struct clocksource lguest_clock = { |
| 766 | .name = "lguest", |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 767 | .rating = 200, |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 768 | .read = lguest_clock_read, |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 769 | .mask = CLOCKSOURCE_MASK(64), |
Rusty Russell | 3725009 | 2007-08-09 20:52:35 +1000 | [diff] [blame] | 770 | .mult = 1 << 22, |
| 771 | .shift = 22, |
Tony Breeds | 05aa026 | 2007-10-22 10:56:25 +1000 | [diff] [blame] | 772 | .flags = CLOCK_SOURCE_IS_CONTINUOUS, |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 773 | }; |
| 774 | |
| 775 | /* We also need a "struct clock_event_device": Linux asks us to set it to go |
| 776 | * off some time in the future. Actually, James Morris figured all this out, I |
| 777 | * just applied the patch. */ |
| 778 | static int lguest_clockevent_set_next_event(unsigned long delta, |
| 779 | struct clock_event_device *evt) |
| 780 | { |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 781 | /* FIXME: I don't think this can ever happen, but James tells me he had |
| 782 | * to put this code in. Maybe we should remove it now. Anyone? */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 783 | if (delta < LG_CLOCK_MIN_DELTA) { |
| 784 | if (printk_ratelimit()) |
| 785 | printk(KERN_DEBUG "%s: small delta %lu ns\n", |
Harvey Harrison | 77bf90e | 2008-03-03 11:37:23 -0800 | [diff] [blame] | 786 | __func__, delta); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 787 | return -ETIME; |
| 788 | } |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 789 | |
| 790 | /* Please wake us this far in the future. */ |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 791 | kvm_hypercall1(LHCALL_SET_CLOCKEVENT, delta); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 792 | return 0; |
| 793 | } |
| 794 | |
| 795 | static void lguest_clockevent_set_mode(enum clock_event_mode mode, |
| 796 | struct clock_event_device *evt) |
| 797 | { |
| 798 | switch (mode) { |
| 799 | case CLOCK_EVT_MODE_UNUSED: |
| 800 | case CLOCK_EVT_MODE_SHUTDOWN: |
| 801 | /* A 0 argument shuts the clock down. */ |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 802 | kvm_hypercall0(LHCALL_SET_CLOCKEVENT); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 803 | break; |
| 804 | case CLOCK_EVT_MODE_ONESHOT: |
| 805 | /* This is what we expect. */ |
| 806 | break; |
| 807 | case CLOCK_EVT_MODE_PERIODIC: |
| 808 | BUG(); |
Thomas Gleixner | 18de5bc | 2007-07-21 04:37:34 -0700 | [diff] [blame] | 809 | case CLOCK_EVT_MODE_RESUME: |
| 810 | break; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 811 | } |
| 812 | } |
| 813 | |
| 814 | /* This describes our primitive timer chip. */ |
| 815 | static struct clock_event_device lguest_clockevent = { |
| 816 | .name = "lguest", |
| 817 | .features = CLOCK_EVT_FEAT_ONESHOT, |
| 818 | .set_next_event = lguest_clockevent_set_next_event, |
| 819 | .set_mode = lguest_clockevent_set_mode, |
| 820 | .rating = INT_MAX, |
| 821 | .mult = 1, |
| 822 | .shift = 0, |
| 823 | .min_delta_ns = LG_CLOCK_MIN_DELTA, |
| 824 | .max_delta_ns = LG_CLOCK_MAX_DELTA, |
| 825 | }; |
| 826 | |
| 827 | /* This is the Guest timer interrupt handler (hardware interrupt 0). We just |
| 828 | * call the clockevent infrastructure and it does whatever needs doing. */ |
| 829 | static void lguest_time_irq(unsigned int irq, struct irq_desc *desc) |
| 830 | { |
| 831 | unsigned long flags; |
| 832 | |
| 833 | /* Don't interrupt us while this is running. */ |
| 834 | local_irq_save(flags); |
| 835 | lguest_clockevent.event_handler(&lguest_clockevent); |
| 836 | local_irq_restore(flags); |
| 837 | } |
| 838 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 839 | /* At some point in the boot process, we get asked to set up our timing |
| 840 | * infrastructure. The kernel doesn't expect timer interrupts before this, but |
| 841 | * we cleverly initialized the "blocked_interrupts" field of "struct |
| 842 | * lguest_data" so that timer interrupts were blocked until now. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 843 | static void lguest_time_init(void) |
| 844 | { |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 845 | /* Set up the timer interrupt (0) to go to our simple timer routine */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 846 | set_irq_handler(0, lguest_time_irq); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 847 | |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 848 | clocksource_register(&lguest_clock); |
| 849 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 850 | /* We can't set cpumask in the initializer: damn C limitations! Set it |
| 851 | * here and register our timer device. */ |
Rusty Russell | 320ab2b | 2008-12-13 21:20:26 +1030 | [diff] [blame] | 852 | lguest_clockevent.cpumask = cpumask_of(0); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 853 | clockevents_register_device(&lguest_clockevent); |
| 854 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 855 | /* Finally, we unblock the timer interrupt. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 856 | enable_lguest_irq(0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 857 | } |
| 858 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 859 | /* |
| 860 | * Miscellaneous bits and pieces. |
| 861 | * |
| 862 | * Here is an oddball collection of functions which the Guest needs for things |
| 863 | * to work. They're pretty simple. |
| 864 | */ |
| 865 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 866 | /* The Guest needs to tell the Host what stack it expects traps to use. For |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 867 | * native hardware, this is part of the Task State Segment mentioned above in |
| 868 | * lguest_load_tr_desc(), but to help hypervisors there's this special call. |
| 869 | * |
| 870 | * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data |
| 871 | * segment), the privilege level (we're privilege level 1, the Host is 0 and |
| 872 | * will not tolerate us trying to use that), the stack pointer, and the number |
| 873 | * of pages in the stack. */ |
H. Peter Anvin | faca622 | 2008-01-30 13:31:02 +0100 | [diff] [blame] | 874 | static void lguest_load_sp0(struct tss_struct *tss, |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 875 | struct thread_struct *thread) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 876 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 877 | lazy_hcall3(LHCALL_SET_STACK, __KERNEL_DS | 0x1, thread->sp0, |
| 878 | THREAD_SIZE / PAGE_SIZE); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 879 | } |
| 880 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 881 | /* Let's just say, I wouldn't do debugging under a Guest. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 882 | static void lguest_set_debugreg(int regno, unsigned long value) |
| 883 | { |
| 884 | /* FIXME: Implement */ |
| 885 | } |
| 886 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 887 | /* There are times when the kernel wants to make sure that no memory writes are |
| 888 | * caught in the cache (that they've all reached real hardware devices). This |
| 889 | * doesn't matter for the Guest which has virtual hardware. |
| 890 | * |
| 891 | * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush |
| 892 | * (clflush) instruction is available and the kernel uses that. Otherwise, it |
| 893 | * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction. |
| 894 | * Unlike clflush, wbinvd can only be run at privilege level 0. So we can |
| 895 | * ignore clflush, but replace wbinvd. |
| 896 | */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 897 | static void lguest_wbinvd(void) |
| 898 | { |
| 899 | } |
| 900 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 901 | /* If the Guest expects to have an Advanced Programmable Interrupt Controller, |
| 902 | * we play dumb by ignoring writes and returning 0 for reads. So it's no |
| 903 | * longer Programmable nor Controlling anything, and I don't think 8 lines of |
| 904 | * code qualifies for Advanced. It will also never interrupt anything. It |
| 905 | * does, however, allow us to get through the Linux boot code. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 906 | #ifdef CONFIG_X86_LOCAL_APIC |
Suresh Siddha | ad66dd3 | 2008-07-11 13:11:56 -0700 | [diff] [blame] | 907 | static void lguest_apic_write(u32 reg, u32 v) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 908 | { |
| 909 | } |
| 910 | |
Suresh Siddha | ad66dd3 | 2008-07-11 13:11:56 -0700 | [diff] [blame] | 911 | static u32 lguest_apic_read(u32 reg) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 912 | { |
| 913 | return 0; |
| 914 | } |
Suresh Siddha | 511d9d3 | 2008-07-14 09:49:14 -0700 | [diff] [blame] | 915 | |
| 916 | static u64 lguest_apic_icr_read(void) |
| 917 | { |
| 918 | return 0; |
| 919 | } |
| 920 | |
| 921 | static void lguest_apic_icr_write(u32 low, u32 id) |
| 922 | { |
| 923 | /* Warn to see if there's any stray references */ |
| 924 | WARN_ON(1); |
| 925 | } |
| 926 | |
| 927 | static void lguest_apic_wait_icr_idle(void) |
| 928 | { |
| 929 | return; |
| 930 | } |
| 931 | |
| 932 | static u32 lguest_apic_safe_wait_icr_idle(void) |
| 933 | { |
| 934 | return 0; |
| 935 | } |
| 936 | |
Yinghai Lu | c1eeb2d | 2009-02-16 23:02:14 -0800 | [diff] [blame] | 937 | static void set_lguest_basic_apic_ops(void) |
| 938 | { |
| 939 | apic->read = lguest_apic_read; |
| 940 | apic->write = lguest_apic_write; |
| 941 | apic->icr_read = lguest_apic_icr_read; |
| 942 | apic->icr_write = lguest_apic_icr_write; |
| 943 | apic->wait_icr_idle = lguest_apic_wait_icr_idle; |
| 944 | apic->safe_wait_icr_idle = lguest_apic_safe_wait_icr_idle; |
Suresh Siddha | 511d9d3 | 2008-07-14 09:49:14 -0700 | [diff] [blame] | 945 | }; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 946 | #endif |
| 947 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 948 | /* STOP! Until an interrupt comes in. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 949 | static void lguest_safe_halt(void) |
| 950 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 951 | kvm_hypercall0(LHCALL_HALT); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 952 | } |
| 953 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 954 | /* The SHUTDOWN hypercall takes a string to describe what's happening, and |
| 955 | * an argument which says whether this to restart (reboot) the Guest or not. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 956 | * |
| 957 | * Note that the Host always prefers that the Guest speak in physical addresses |
| 958 | * rather than virtual addresses, so we use __pa() here. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 959 | static void lguest_power_off(void) |
| 960 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 961 | kvm_hypercall2(LHCALL_SHUTDOWN, __pa("Power down"), |
| 962 | LGUEST_SHUTDOWN_POWEROFF); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 963 | } |
| 964 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 965 | /* |
| 966 | * Panicing. |
| 967 | * |
| 968 | * Don't. But if you did, this is what happens. |
| 969 | */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 970 | static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p) |
| 971 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 972 | kvm_hypercall2(LHCALL_SHUTDOWN, __pa(p), LGUEST_SHUTDOWN_POWEROFF); |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 973 | /* The hcall won't return, but to keep gcc happy, we're "done". */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 974 | return NOTIFY_DONE; |
| 975 | } |
| 976 | |
| 977 | static struct notifier_block paniced = { |
| 978 | .notifier_call = lguest_panic |
| 979 | }; |
| 980 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 981 | /* Setting up memory is fairly easy. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 982 | static __init char *lguest_memory_setup(void) |
| 983 | { |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 984 | /* We do this here and not earlier because lockcheck used to barf if we |
| 985 | * did it before start_kernel(). I think we fixed that, so it'd be |
| 986 | * nice to move it back to lguest_init. Patch welcome... */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 987 | atomic_notifier_chain_register(&panic_notifier_list, &paniced); |
| 988 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 989 | /* The Linux bootloader header contains an "e820" memory map: the |
| 990 | * Launcher populated the first entry with our memory limit. */ |
Yinghai Lu | d0be6bd | 2008-06-15 18:58:51 -0700 | [diff] [blame] | 991 | e820_add_region(boot_params.e820_map[0].addr, |
H. Peter Anvin | 30c8264 | 2007-10-15 17:13:22 -0700 | [diff] [blame] | 992 | boot_params.e820_map[0].size, |
| 993 | boot_params.e820_map[0].type); |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 994 | |
| 995 | /* This string is for the boot messages. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 996 | return "LGUEST"; |
| 997 | } |
| 998 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 999 | /* We will eventually use the virtio console device to produce console output, |
| 1000 | * but before that is set up we use LHCALL_NOTIFY on normal memory to produce |
| 1001 | * console output. */ |
Rusty Russell | 19f1537 | 2007-10-22 11:24:21 +1000 | [diff] [blame] | 1002 | static __init int early_put_chars(u32 vtermno, const char *buf, int count) |
| 1003 | { |
| 1004 | char scratch[17]; |
| 1005 | unsigned int len = count; |
| 1006 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 1007 | /* We use a nul-terminated string, so we have to make a copy. Icky, |
| 1008 | * huh? */ |
Rusty Russell | 19f1537 | 2007-10-22 11:24:21 +1000 | [diff] [blame] | 1009 | if (len > sizeof(scratch) - 1) |
| 1010 | len = sizeof(scratch) - 1; |
| 1011 | scratch[len] = '\0'; |
| 1012 | memcpy(scratch, buf, len); |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 1013 | kvm_hypercall1(LHCALL_NOTIFY, __pa(scratch)); |
Rusty Russell | 19f1537 | 2007-10-22 11:24:21 +1000 | [diff] [blame] | 1014 | |
| 1015 | /* This routine returns the number of bytes actually written. */ |
| 1016 | return len; |
| 1017 | } |
| 1018 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1019 | /* Rebooting also tells the Host we're finished, but the RESTART flag tells the |
| 1020 | * Launcher to reboot us. */ |
| 1021 | static void lguest_restart(char *reason) |
| 1022 | { |
Matias Zabaljauregui | 4cd8b5e | 2009-03-14 13:37:52 -0200 | [diff] [blame] | 1023 | kvm_hypercall2(LHCALL_SHUTDOWN, __pa(reason), LGUEST_SHUTDOWN_RESTART); |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1024 | } |
| 1025 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1026 | /*G:050 |
| 1027 | * Patching (Powerfully Placating Performance Pedants) |
| 1028 | * |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1029 | * We have already seen that pv_ops structures let us replace simple native |
| 1030 | * instructions with calls to the appropriate back end all throughout the |
| 1031 | * kernel. This allows the same kernel to run as a Guest and as a native |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1032 | * kernel, but it's slow because of all the indirect branches. |
| 1033 | * |
| 1034 | * Remember that David Wheeler quote about "Any problem in computer science can |
| 1035 | * be solved with another layer of indirection"? The rest of that quote is |
| 1036 | * "... But that usually will create another problem." This is the first of |
| 1037 | * those problems. |
| 1038 | * |
| 1039 | * Our current solution is to allow the paravirt back end to optionally patch |
| 1040 | * over the indirect calls to replace them with something more efficient. We |
Rusty Russell | a32a881 | 2009-06-12 22:27:02 -0600 | [diff] [blame] | 1041 | * patch two of the simplest of the most commonly called functions: disable |
| 1042 | * interrupts and save interrupts. We usually have 6 or 10 bytes to patch |
| 1043 | * into: the Guest versions of these operations are small enough that we can |
| 1044 | * fit comfortably. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1045 | * |
| 1046 | * First we need assembly templates of each of the patchable Guest operations, |
Atsushi SAKAI | 72410af | 2009-01-16 20:39:14 +0900 | [diff] [blame] | 1047 | * and these are in i386_head.S. */ |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1048 | |
| 1049 | /*G:060 We construct a table from the assembler templates: */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1050 | static const struct lguest_insns |
| 1051 | { |
| 1052 | const char *start, *end; |
| 1053 | } lguest_insns[] = { |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1054 | [PARAVIRT_PATCH(pv_irq_ops.irq_disable)] = { lgstart_cli, lgend_cli }, |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1055 | [PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf }, |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1056 | }; |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1057 | |
| 1058 | /* Now our patch routine is fairly simple (based on the native one in |
| 1059 | * paravirt.c). If we have a replacement, we copy it in and return how much of |
| 1060 | * the available space we used. */ |
Andi Kleen | ab144f5 | 2007-08-10 22:31:03 +0200 | [diff] [blame] | 1061 | static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf, |
| 1062 | unsigned long addr, unsigned len) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1063 | { |
| 1064 | unsigned int insn_len; |
| 1065 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1066 | /* Don't do anything special if we don't have a replacement */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1067 | if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start) |
Andi Kleen | ab144f5 | 2007-08-10 22:31:03 +0200 | [diff] [blame] | 1068 | return paravirt_patch_default(type, clobber, ibuf, addr, len); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1069 | |
| 1070 | insn_len = lguest_insns[type].end - lguest_insns[type].start; |
| 1071 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1072 | /* Similarly if we can't fit replacement (shouldn't happen, but let's |
| 1073 | * be thorough). */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1074 | if (len < insn_len) |
Andi Kleen | ab144f5 | 2007-08-10 22:31:03 +0200 | [diff] [blame] | 1075 | return paravirt_patch_default(type, clobber, ibuf, addr, len); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1076 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1077 | /* Copy in our instructions. */ |
Andi Kleen | ab144f5 | 2007-08-10 22:31:03 +0200 | [diff] [blame] | 1078 | memcpy(ibuf, lguest_insns[type].start, insn_len); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1079 | return insn_len; |
| 1080 | } |
| 1081 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1082 | /*G:030 Once we get to lguest_init(), we know we're a Guest. The various |
| 1083 | * pv_ops structures in the kernel provide points for (almost) every routine we |
| 1084 | * have to override to avoid privileged instructions. */ |
Rusty Russell | 814a0e5 | 2007-10-22 11:29:44 +1000 | [diff] [blame] | 1085 | __init void lguest_init(void) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1086 | { |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1087 | /* We're under lguest, paravirt is enabled, and we're running at |
| 1088 | * privilege level 1, not 0 as normal. */ |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1089 | pv_info.name = "lguest"; |
| 1090 | pv_info.paravirt_enabled = 1; |
| 1091 | pv_info.kernel_rpl = 1; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 1092 | pv_info.shared_kernel_pmd = 1; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1093 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1094 | /* We set up all the lguest overrides for sensitive operations. These |
| 1095 | * are detailed with the operations themselves. */ |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1096 | |
| 1097 | /* interrupt-related operations */ |
| 1098 | pv_irq_ops.init_IRQ = lguest_init_IRQ; |
Jeremy Fitzhardinge | ecb93d1 | 2009-01-28 14:35:05 -0800 | [diff] [blame] | 1099 | pv_irq_ops.save_fl = PV_CALLEE_SAVE(save_fl); |
Rusty Russell | 61f4bc8 | 2009-06-12 22:27:03 -0600 | [diff] [blame] | 1100 | pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(lg_restore_fl); |
Jeremy Fitzhardinge | ecb93d1 | 2009-01-28 14:35:05 -0800 | [diff] [blame] | 1101 | pv_irq_ops.irq_disable = PV_CALLEE_SAVE(irq_disable); |
Rusty Russell | 61f4bc8 | 2009-06-12 22:27:03 -0600 | [diff] [blame] | 1102 | pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(lg_irq_enable); |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1103 | pv_irq_ops.safe_halt = lguest_safe_halt; |
| 1104 | |
| 1105 | /* init-time operations */ |
| 1106 | pv_init_ops.memory_setup = lguest_memory_setup; |
| 1107 | pv_init_ops.patch = lguest_patch; |
| 1108 | |
| 1109 | /* Intercepts of various cpu instructions */ |
| 1110 | pv_cpu_ops.load_gdt = lguest_load_gdt; |
| 1111 | pv_cpu_ops.cpuid = lguest_cpuid; |
| 1112 | pv_cpu_ops.load_idt = lguest_load_idt; |
| 1113 | pv_cpu_ops.iret = lguest_iret; |
H. Peter Anvin | faca622 | 2008-01-30 13:31:02 +0100 | [diff] [blame] | 1114 | pv_cpu_ops.load_sp0 = lguest_load_sp0; |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1115 | pv_cpu_ops.load_tr_desc = lguest_load_tr_desc; |
| 1116 | pv_cpu_ops.set_ldt = lguest_set_ldt; |
| 1117 | pv_cpu_ops.load_tls = lguest_load_tls; |
| 1118 | pv_cpu_ops.set_debugreg = lguest_set_debugreg; |
| 1119 | pv_cpu_ops.clts = lguest_clts; |
| 1120 | pv_cpu_ops.read_cr0 = lguest_read_cr0; |
| 1121 | pv_cpu_ops.write_cr0 = lguest_write_cr0; |
| 1122 | pv_cpu_ops.read_cr4 = lguest_read_cr4; |
| 1123 | pv_cpu_ops.write_cr4 = lguest_write_cr4; |
| 1124 | pv_cpu_ops.write_gdt_entry = lguest_write_gdt_entry; |
| 1125 | pv_cpu_ops.write_idt_entry = lguest_write_idt_entry; |
| 1126 | pv_cpu_ops.wbinvd = lguest_wbinvd; |
Jeremy Fitzhardinge | 224101e | 2009-02-18 11:18:57 -0800 | [diff] [blame] | 1127 | pv_cpu_ops.start_context_switch = paravirt_start_context_switch; |
| 1128 | pv_cpu_ops.end_context_switch = lguest_end_context_switch; |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1129 | |
| 1130 | /* pagetable management */ |
| 1131 | pv_mmu_ops.write_cr3 = lguest_write_cr3; |
| 1132 | pv_mmu_ops.flush_tlb_user = lguest_flush_tlb_user; |
| 1133 | pv_mmu_ops.flush_tlb_single = lguest_flush_tlb_single; |
| 1134 | pv_mmu_ops.flush_tlb_kernel = lguest_flush_tlb_kernel; |
| 1135 | pv_mmu_ops.set_pte = lguest_set_pte; |
| 1136 | pv_mmu_ops.set_pte_at = lguest_set_pte_at; |
| 1137 | pv_mmu_ops.set_pmd = lguest_set_pmd; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 1138 | #ifdef CONFIG_X86_PAE |
| 1139 | pv_mmu_ops.set_pte_atomic = lguest_set_pte_atomic; |
| 1140 | pv_mmu_ops.pte_clear = lguest_pte_clear; |
| 1141 | pv_mmu_ops.pmd_clear = lguest_pmd_clear; |
| 1142 | pv_mmu_ops.set_pud = lguest_set_pud; |
| 1143 | #endif |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1144 | pv_mmu_ops.read_cr2 = lguest_read_cr2; |
| 1145 | pv_mmu_ops.read_cr3 = lguest_read_cr3; |
Jeremy Fitzhardinge | 8965c1c | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1146 | pv_mmu_ops.lazy_mode.enter = paravirt_enter_lazy_mmu; |
Jeremy Fitzhardinge | b407fc5 | 2009-02-17 23:46:21 -0800 | [diff] [blame] | 1147 | pv_mmu_ops.lazy_mode.leave = lguest_leave_lazy_mmu_mode; |
Rusty Russell | b7ff99e | 2009-03-30 21:55:23 -0600 | [diff] [blame] | 1148 | pv_mmu_ops.pte_update = lguest_pte_update; |
| 1149 | pv_mmu_ops.pte_update_defer = lguest_pte_update; |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1150 | |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1151 | #ifdef CONFIG_X86_LOCAL_APIC |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1152 | /* apic read/write intercepts */ |
Yinghai Lu | c1eeb2d | 2009-02-16 23:02:14 -0800 | [diff] [blame] | 1153 | set_lguest_basic_apic_ops(); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1154 | #endif |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1155 | |
| 1156 | /* time operations */ |
| 1157 | pv_time_ops.get_wallclock = lguest_get_wallclock; |
| 1158 | pv_time_ops.time_init = lguest_time_init; |
Alok Kataria | e93ef94 | 2008-07-01 11:43:36 -0700 | [diff] [blame] | 1159 | pv_time_ops.get_tsc_khz = lguest_tsc_khz; |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1160 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1161 | /* Now is a good time to look at the implementations of these functions |
| 1162 | * before returning to the rest of lguest_init(). */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1163 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1164 | /*G:070 Now we've seen all the paravirt_ops, we return to |
| 1165 | * lguest_init() where the rest of the fairly chaotic boot setup |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 1166 | * occurs. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1167 | |
Rusty Russell | 2cb7878 | 2009-06-03 14:52:24 +0930 | [diff] [blame] | 1168 | /* The stack protector is a weird thing where gcc places a canary |
| 1169 | * value on the stack and then checks it on return. This file is |
| 1170 | * compiled with -fno-stack-protector it, so we got this far without |
| 1171 | * problems. The value of the canary is kept at offset 20 from the |
| 1172 | * %gs register, so we need to set that up before calling C functions |
| 1173 | * in other files. */ |
| 1174 | setup_stack_canary_segment(0); |
| 1175 | /* We could just call load_stack_canary_segment(), but we might as |
| 1176 | * call switch_to_new_gdt() which loads the whole table and sets up |
| 1177 | * the per-cpu segment descriptor register %fs as well. */ |
| 1178 | switch_to_new_gdt(0); |
| 1179 | |
Rusty Russell | 5d006d8 | 2008-07-29 09:58:29 -0500 | [diff] [blame] | 1180 | /* As described in head_32.S, we map the first 128M of memory. */ |
| 1181 | max_pfn_mapped = (128*1024*1024) >> PAGE_SHIFT; |
| 1182 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1183 | /* The Host<->Guest Switcher lives at the top of our address space, and |
| 1184 | * the Host told us how big it is when we made LGUEST_INIT hypercall: |
| 1185 | * it put the answer in lguest_data.reserve_mem */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1186 | reserve_top_address(lguest_data.reserve_mem); |
| 1187 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1188 | /* If we don't initialize the lock dependency checker now, it crashes |
| 1189 | * paravirt_disable_iospace. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1190 | lockdep_init(); |
| 1191 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1192 | /* The IDE code spends about 3 seconds probing for disks: if we reserve |
| 1193 | * all the I/O ports up front it can't get them and so doesn't probe. |
| 1194 | * Other device drivers are similar (but less severe). This cuts the |
| 1195 | * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1196 | paravirt_disable_iospace(); |
| 1197 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1198 | /* This is messy CPU setup stuff which the native boot code does before |
| 1199 | * start_kernel, so we have to do, too: */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1200 | cpu_detect(&new_cpu_data); |
| 1201 | /* head.S usually sets up the first capability word, so do it here. */ |
| 1202 | new_cpu_data.x86_capability[0] = cpuid_edx(1); |
| 1203 | |
| 1204 | /* Math is always hard! */ |
| 1205 | new_cpu_data.hard_math = 1; |
| 1206 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1207 | /* We don't have features. We have puppies! Puppies! */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1208 | #ifdef CONFIG_X86_MCE |
| 1209 | mce_disabled = 1; |
| 1210 | #endif |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1211 | #ifdef CONFIG_ACPI |
| 1212 | acpi_disabled = 1; |
| 1213 | acpi_ht = 0; |
| 1214 | #endif |
| 1215 | |
Atsushi SAKAI | 72410af | 2009-01-16 20:39:14 +0900 | [diff] [blame] | 1216 | /* We set the preferred console to "hvc". This is the "hypervisor |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1217 | * virtual console" driver written by the PowerPC people, which we also |
| 1218 | * adapted for lguest's use. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1219 | add_preferred_console("hvc", 0, NULL); |
| 1220 | |
Rusty Russell | 19f1537 | 2007-10-22 11:24:21 +1000 | [diff] [blame] | 1221 | /* Register our very early console. */ |
| 1222 | virtio_cons_early_init(early_put_chars); |
| 1223 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1224 | /* Last of all, we set the power management poweroff hook to point to |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1225 | * the Guest routine to power off, and the reboot hook to our restart |
| 1226 | * routine. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1227 | pm_power_off = lguest_power_off; |
Balaji Rao | ec04b13 | 2007-12-28 14:26:24 +0530 | [diff] [blame] | 1228 | machine_ops.restart = lguest_restart; |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1229 | |
Yinghai Lu | f0d4310 | 2008-05-29 12:56:36 -0700 | [diff] [blame] | 1230 | /* Now we're set up, call i386_start_kernel() in head32.c and we proceed |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1231 | * to boot as normal. It never returns. */ |
Yinghai Lu | f0d4310 | 2008-05-29 12:56:36 -0700 | [diff] [blame] | 1232 | i386_start_kernel(); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1233 | } |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1234 | /* |
| 1235 | * This marks the end of stage II of our journey, The Guest. |
| 1236 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 1237 | * It is now time for us to explore the layer of virtual drivers and complete |
| 1238 | * our understanding of the Guest in "make Drivers". |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1239 | */ |