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> |
Balaji Rao | ec04b13 | 2007-12-28 14:26:24 +0530 | [diff] [blame] | 70 | #include <asm/reboot.h> /* for struct machine_ops */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 71 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 72 | /*G:010 Welcome to the Guest! |
| 73 | * |
| 74 | * The Guest in our tale is a simple creature: identical to the Host but |
| 75 | * behaving in simplified but equivalent ways. In particular, the Guest is the |
| 76 | * same kernel as the Host (or at least, built from the same source code). :*/ |
| 77 | |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 78 | struct lguest_data lguest_data = { |
| 79 | .hcall_status = { [0 ... LHCALL_RING_SIZE-1] = 0xFF }, |
| 80 | .noirq_start = (u32)lguest_noirq_start, |
| 81 | .noirq_end = (u32)lguest_noirq_end, |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 82 | .kernel_address = PAGE_OFFSET, |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 83 | .blocked_interrupts = { 1 }, /* Block timer interrupts */ |
Rusty Russell | c18acd7 | 2007-10-22 11:03:35 +1000 | [diff] [blame] | 84 | .syscall_vec = SYSCALL_VECTOR, |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 85 | }; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 86 | |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 87 | /*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] | 88 | * ring buffer of stored hypercalls which the Host will run though next time we |
| 89 | * do a normal hypercall. Each entry in the ring has 4 slots for the hypercall |
| 90 | * arguments, and a "hcall_status" word which is 0 if the call is ready to go, |
| 91 | * and 255 once the Host has finished with it. |
| 92 | * |
| 93 | * If we come around to a slot which hasn't been finished, then the table is |
| 94 | * full and we just make the hypercall directly. This has the nice side |
| 95 | * effect of causing the Host to run all the stored calls in the ring buffer |
| 96 | * which empties it for next time! */ |
Adrian Bunk | 9b56fdb | 2007-11-02 16:43:10 +0100 | [diff] [blame] | 97 | static void async_hcall(unsigned long call, unsigned long arg1, |
| 98 | unsigned long arg2, unsigned long arg3) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 99 | { |
| 100 | /* Note: This code assumes we're uniprocessor. */ |
| 101 | static unsigned int next_call; |
| 102 | unsigned long flags; |
| 103 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 104 | /* Disable interrupts if not already disabled: we don't want an |
| 105 | * interrupt handler making a hypercall while we're already doing |
| 106 | * one! */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 107 | local_irq_save(flags); |
| 108 | if (lguest_data.hcall_status[next_call] != 0xFF) { |
| 109 | /* Table full, so do normal hcall which will flush table. */ |
| 110 | hcall(call, arg1, arg2, arg3); |
| 111 | } else { |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 112 | lguest_data.hcalls[next_call].arg0 = call; |
| 113 | lguest_data.hcalls[next_call].arg1 = arg1; |
| 114 | lguest_data.hcalls[next_call].arg2 = arg2; |
| 115 | lguest_data.hcalls[next_call].arg3 = arg3; |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 116 | /* Arguments must all be written before we mark it to go */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 117 | wmb(); |
| 118 | lguest_data.hcall_status[next_call] = 0; |
| 119 | if (++next_call == LHCALL_RING_SIZE) |
| 120 | next_call = 0; |
| 121 | } |
| 122 | local_irq_restore(flags); |
| 123 | } |
Adrian Bunk | 9b56fdb | 2007-11-02 16:43:10 +0100 | [diff] [blame] | 124 | |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 125 | /*G:035 Notice the lazy_hcall() above, rather than hcall(). This is our first |
| 126 | * real optimization trick! |
| 127 | * |
| 128 | * When lazy_mode is set, it means we're allowed to defer all hypercalls and do |
| 129 | * them as a batch when lazy_mode is eventually turned off. Because hypercalls |
| 130 | * are reasonably expensive, batching them up makes sense. For example, a |
| 131 | * large munmap might update dozens of page table entries: that code calls |
| 132 | * paravirt_enter_lazy_mmu(), does the dozen updates, then calls |
| 133 | * lguest_leave_lazy_mode(). |
| 134 | * |
| 135 | * 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] | 136 | * future processing: */ |
Adrian Bunk | 9b56fdb | 2007-11-02 16:43:10 +0100 | [diff] [blame] | 137 | static void lazy_hcall(unsigned long call, |
| 138 | unsigned long arg1, |
| 139 | unsigned long arg2, |
| 140 | unsigned long arg3) |
| 141 | { |
| 142 | if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE) |
| 143 | hcall(call, arg1, arg2, arg3); |
| 144 | else |
| 145 | async_hcall(call, arg1, arg2, arg3); |
| 146 | } |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 147 | |
| 148 | /* 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] | 149 | * issue the do-nothing hypercall to flush any stored calls. */ |
Rusty Russell | 633872b | 2007-11-05 21:55:57 +1100 | [diff] [blame] | 150 | static void lguest_leave_lazy_mode(void) |
| 151 | { |
| 152 | paravirt_leave_lazy(paravirt_get_lazy_mode()); |
| 153 | hcall(LHCALL_FLUSH_ASYNC, 0, 0, 0); |
| 154 | } |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 155 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 156 | /*G:033 |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 157 | * After that diversion we return to our first native-instruction |
| 158 | * replacements: four functions for interrupt control. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 159 | * |
| 160 | * The simplest way of implementing these would be to have "turn interrupts |
| 161 | * off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow: |
| 162 | * these are by far the most commonly called functions of those we override. |
| 163 | * |
| 164 | * So instead we keep an "irq_enabled" field inside our "struct lguest_data", |
| 165 | * 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] | 166 | * check there before it tries to deliver an interrupt. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 167 | */ |
| 168 | |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 169 | /* save_flags() is expected to return the processor state (ie. "flags"). The |
| 170 | * 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] | 171 | * about the interrupt flag. Our "save_flags()" just returns that. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 172 | static unsigned long save_fl(void) |
| 173 | { |
| 174 | return lguest_data.irq_enabled; |
| 175 | } |
Jeremy Fitzhardinge | ecb93d1 | 2009-01-28 14:35:05 -0800 | [diff] [blame] | 176 | PV_CALLEE_SAVE_REGS_THUNK(save_fl); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 177 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 178 | /* restore_flags() just sets the flags back to the value given. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 179 | static void restore_fl(unsigned long flags) |
| 180 | { |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 181 | lguest_data.irq_enabled = flags; |
| 182 | } |
Jeremy Fitzhardinge | ecb93d1 | 2009-01-28 14:35:05 -0800 | [diff] [blame] | 183 | PV_CALLEE_SAVE_REGS_THUNK(restore_fl); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 184 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 185 | /* Interrupts go off... */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 186 | static void irq_disable(void) |
| 187 | { |
| 188 | lguest_data.irq_enabled = 0; |
| 189 | } |
Jeremy Fitzhardinge | ecb93d1 | 2009-01-28 14:35:05 -0800 | [diff] [blame] | 190 | PV_CALLEE_SAVE_REGS_THUNK(irq_disable); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 191 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 192 | /* Interrupts go on... */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 193 | static void irq_enable(void) |
| 194 | { |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 195 | lguest_data.irq_enabled = X86_EFLAGS_IF; |
| 196 | } |
Jeremy Fitzhardinge | ecb93d1 | 2009-01-28 14:35:05 -0800 | [diff] [blame] | 197 | PV_CALLEE_SAVE_REGS_THUNK(irq_enable); |
| 198 | |
Rusty Russell | f56a384 | 2007-07-26 10:41:05 -0700 | [diff] [blame] | 199 | /*:*/ |
| 200 | /*M:003 Note that we don't check for outstanding interrupts when we re-enable |
| 201 | * them (or when we unmask an interrupt). This seems to work for the moment, |
| 202 | * 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] | 203 | * 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] | 204 | * would be to put the "irq_enabled" field in a page by itself, and have the |
| 205 | * 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] | 206 | * There will then be a page fault as soon as interrupts are re-enabled. |
| 207 | * |
| 208 | * A better method is to implement soft interrupt disable generally for x86: |
| 209 | * instead of disabling interrupts, we set a flag. If an interrupt does come |
| 210 | * in, we then disable them for real. This is uncommon, so we could simply use |
| 211 | * a hypercall for interrupt control and not worry about efficiency. :*/ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 212 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 213 | /*G:034 |
| 214 | * The Interrupt Descriptor Table (IDT). |
| 215 | * |
| 216 | * The IDT tells the processor what to do when an interrupt comes in. Each |
| 217 | * entry in the table is a 64-bit descriptor: this holds the privilege level, |
| 218 | * address of the handler, and... well, who cares? The Guest just asks the |
| 219 | * Host to make the change anyway, because the Host controls the real IDT. |
| 220 | */ |
Glauber de Oliveira Costa | 8d94734 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 221 | static void lguest_write_idt_entry(gate_desc *dt, |
| 222 | int entrynum, const gate_desc *g) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 223 | { |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 224 | /* The gate_desc structure is 8 bytes long: we hand it to the Host in |
| 225 | * two 32-bit chunks. The whole 32-bit kernel used to hand descriptors |
| 226 | * around like this; typesafety wasn't a big concern in Linux's early |
| 227 | * years. */ |
Glauber de Oliveira Costa | 8d94734 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 228 | u32 *desc = (u32 *)g; |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 229 | /* Keep the local copy up to date. */ |
Glauber de Oliveira Costa | 8d94734 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 230 | native_write_idt_entry(dt, entrynum, g); |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 231 | /* Tell Host about this new entry. */ |
Glauber de Oliveira Costa | 8d94734 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 232 | hcall(LHCALL_LOAD_IDT_ENTRY, entrynum, desc[0], desc[1]); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 233 | } |
| 234 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 235 | /* Changing to a different IDT is very rare: we keep the IDT up-to-date every |
| 236 | * time it is written, so we can simply loop through all entries and tell the |
| 237 | * Host about them. */ |
Glauber de Oliveira Costa | 6b68f01 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 238 | static void lguest_load_idt(const struct desc_ptr *desc) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 239 | { |
| 240 | unsigned int i; |
| 241 | struct desc_struct *idt = (void *)desc->address; |
| 242 | |
| 243 | for (i = 0; i < (desc->size+1)/8; i++) |
| 244 | hcall(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b); |
| 245 | } |
| 246 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 247 | /* |
| 248 | * The Global Descriptor Table. |
| 249 | * |
| 250 | * The Intel architecture defines another table, called the Global Descriptor |
| 251 | * Table (GDT). You tell the CPU where it is (and its size) using the "lgdt" |
| 252 | * instruction, and then several other instructions refer to entries in the |
| 253 | * table. There are three entries which the Switcher needs, so the Host simply |
| 254 | * controls the entire thing and the Guest asks it to make changes using the |
| 255 | * LOAD_GDT hypercall. |
| 256 | * |
| 257 | * This is the opposite of the IDT code where we have a LOAD_IDT_ENTRY |
| 258 | * hypercall and use that repeatedly to load a new IDT. I don't think it |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 259 | * really matters, but wouldn't it be nice if they were the same? Wouldn't |
| 260 | * it be even better if you were the one to send the patch to fix it? |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 261 | */ |
Glauber de Oliveira Costa | 6b68f01 | 2008-01-30 13:31:12 +0100 | [diff] [blame] | 262 | static void lguest_load_gdt(const struct desc_ptr *desc) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 263 | { |
| 264 | BUG_ON((desc->size+1)/8 != GDT_ENTRIES); |
| 265 | hcall(LHCALL_LOAD_GDT, __pa(desc->address), GDT_ENTRIES, 0); |
| 266 | } |
| 267 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 268 | /* For a single GDT entry which changes, we do the lazy thing: alter our GDT, |
| 269 | * then tell the Host to reload the entire thing. This operation is so rare |
| 270 | * that this naive implementation is reasonable. */ |
Glauber de Oliveira Costa | 014b15b | 2008-01-30 13:31:13 +0100 | [diff] [blame] | 271 | static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum, |
| 272 | const void *desc, int type) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 273 | { |
Glauber de Oliveira Costa | 014b15b | 2008-01-30 13:31:13 +0100 | [diff] [blame] | 274 | native_write_gdt_entry(dt, entrynum, desc, type); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 275 | hcall(LHCALL_LOAD_GDT, __pa(dt), GDT_ENTRIES, 0); |
| 276 | } |
| 277 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 278 | /* OK, I lied. There are three "thread local storage" GDT entries which change |
| 279 | * on every context switch (these three entries are how glibc implements |
| 280 | * __thread variables). So we have a hypercall specifically for this case. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 281 | static void lguest_load_tls(struct thread_struct *t, unsigned int cpu) |
| 282 | { |
Rusty Russell | 0d027c0 | 2007-08-09 20:57:13 +1000 | [diff] [blame] | 283 | /* There's one problem which normal hardware doesn't have: the Host |
| 284 | * can't handle us removing entries we're currently using. So we clear |
| 285 | * 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] | 286 | lazy_load_gs(0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 287 | lazy_hcall(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu, 0); |
| 288 | } |
| 289 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 290 | /*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] | 291 | * 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] | 292 | * |
| 293 | * This is the Local Descriptor Table, another weird Intel thingy. Linux only |
| 294 | * uses this for some strange applications like Wine. We don't do anything |
| 295 | * here, so they'll get an informative and friendly Segmentation Fault. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 296 | static void lguest_set_ldt(const void *addr, unsigned entries) |
| 297 | { |
| 298 | } |
| 299 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 300 | /* This loads a GDT entry into the "Task Register": that entry points to a |
| 301 | * structure called the Task State Segment. Some comments scattered though the |
| 302 | * kernel code indicate that this used for task switching in ages past, along |
| 303 | * with blood sacrifice and astrology. |
| 304 | * |
| 305 | * Now there's nothing interesting in here that we don't get told elsewhere. |
| 306 | * But the native version uses the "ltr" instruction, which makes the Host |
| 307 | * complain to the Guest about a Segmentation Fault and it'll oops. So we |
| 308 | * override the native version with a do-nothing version. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 309 | static void lguest_load_tr_desc(void) |
| 310 | { |
| 311 | } |
| 312 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 313 | /* The "cpuid" instruction is a way of querying both the CPU identity |
| 314 | * (manufacturer, model, etc) and its features. It was introduced before the |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 315 | * Pentium in 1993 and keeps getting extended by both Intel, AMD and others. |
| 316 | * As you might imagine, after a decade and a half this treatment, it is now a |
| 317 | * 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] | 318 | * |
| 319 | * This instruction even it has its own Wikipedia entry. The Wikipedia entry |
| 320 | * has been translated into 4 languages. I am not making this up! |
| 321 | * |
| 322 | * We could get funky here and identify ourselves as "GenuineLguest", but |
| 323 | * instead we just use the real "cpuid" instruction. Then I pretty much turned |
| 324 | * off feature bits until the Guest booted. (Don't say that: you'll damage |
| 325 | * lguest sales!) Shut up, inner voice! (Hey, just pointing out that this is |
| 326 | * hardly future proof.) Noone's listening! They don't like you anyway, |
| 327 | * parenthetic weirdo! |
| 328 | * |
| 329 | * Replacing the cpuid so we can turn features off is great for the kernel, but |
| 330 | * anyone (including userspace) can just use the raw "cpuid" instruction and |
| 331 | * the Host won't even notice since it isn't privileged. So we try not to get |
| 332 | * too worked up about it. */ |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 333 | static void lguest_cpuid(unsigned int *ax, unsigned int *bx, |
| 334 | unsigned int *cx, unsigned int *dx) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 335 | { |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 336 | int function = *ax; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 337 | |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 338 | native_cpuid(ax, bx, cx, dx); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 339 | switch (function) { |
| 340 | case 1: /* Basic feature request. */ |
| 341 | /* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */ |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 342 | *cx &= 0x00002201; |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 343 | /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU. */ |
| 344 | *dx &= 0x07808111; |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 345 | /* The Host can do a nice optimization if it knows that the |
| 346 | * kernel mappings (addresses above 0xC0000000 or whatever |
| 347 | * PAGE_OFFSET is set to) haven't changed. But Linux calls |
| 348 | * flush_tlb_user() for both user and kernel mappings unless |
| 349 | * the Page Global Enable (PGE) feature bit is set. */ |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 350 | *dx |= 0x00002000; |
Rusty Russell | cbd88c8 | 2009-03-09 10:06:22 -0600 | [diff] [blame] | 351 | /* We also lie, and say we're family id 5. 6 or greater |
| 352 | * leads to a rdmsr in early_init_intel which we can't handle. |
| 353 | * Family ID is returned as bits 8-12 in ax. */ |
| 354 | *ax &= 0xFFFFF0FF; |
| 355 | *ax |= 0x00000500; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 356 | break; |
| 357 | case 0x80000000: |
| 358 | /* Futureproof this a little: if they ask how much extended |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 359 | * processor information there is, limit it to known fields. */ |
H. Peter Anvin | 65ea5b0 | 2008-01-30 13:30:56 +0100 | [diff] [blame] | 360 | if (*ax > 0x80000008) |
| 361 | *ax = 0x80000008; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 362 | break; |
| 363 | } |
| 364 | } |
| 365 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 366 | /* Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4. |
| 367 | * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother |
| 368 | * it. The Host needs to know when the Guest wants to change them, so we have |
| 369 | * a whole series of functions like read_cr0() and write_cr0(). |
| 370 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 371 | * 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] | 372 | * features, but Linux only really cares about one: the horrifically-named Task |
| 373 | * Switched (TS) bit at bit 3 (ie. 8) |
| 374 | * |
| 375 | * What does the TS bit do? Well, it causes the CPU to trap (interrupt 7) if |
| 376 | * the floating point unit is used. Which allows us to restore FPU state |
| 377 | * lazily after a task switch, and Linux uses that gratefully, but wouldn't a |
| 378 | * name like "FPUTRAP bit" be a little less cryptic? |
| 379 | * |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 380 | * We store cr0 locally because the Host never changes it. The Guest sometimes |
| 381 | * wants to read it and we'd prefer not to bother the Host unnecessarily. */ |
| 382 | static unsigned long current_cr0; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 383 | static void lguest_write_cr0(unsigned long val) |
| 384 | { |
Rusty Russell | 25c47bb | 2007-10-25 14:09:53 +1000 | [diff] [blame] | 385 | lazy_hcall(LHCALL_TS, val & X86_CR0_TS, 0, 0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 386 | current_cr0 = val; |
| 387 | } |
| 388 | |
| 389 | static unsigned long lguest_read_cr0(void) |
| 390 | { |
| 391 | return current_cr0; |
| 392 | } |
| 393 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 394 | /* Intel provided a special instruction to clear the TS bit for people too cool |
| 395 | * to use write_cr0() to do it. This "clts" instruction is faster, because all |
| 396 | * the vowels have been optimized out. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 397 | static void lguest_clts(void) |
| 398 | { |
| 399 | lazy_hcall(LHCALL_TS, 0, 0, 0); |
Rusty Russell | 25c47bb | 2007-10-25 14:09:53 +1000 | [diff] [blame] | 400 | current_cr0 &= ~X86_CR0_TS; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 401 | } |
| 402 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 403 | /* 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] | 404 | * reads. The Host kindly writes this into our "struct lguest_data", so we |
| 405 | * just read it out of there. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 406 | static unsigned long lguest_read_cr2(void) |
| 407 | { |
| 408 | return lguest_data.cr2; |
| 409 | } |
| 410 | |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 411 | /* See lguest_set_pte() below. */ |
| 412 | static bool cr3_changed = false; |
| 413 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 414 | /* 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] | 415 | * cr0. Keep a local copy, and tell the Host when it changes. The only |
| 416 | * difference is that our local copy is in lguest_data because the Host needs |
| 417 | * to set it upon our initial hypercall. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 418 | static void lguest_write_cr3(unsigned long cr3) |
| 419 | { |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 420 | lguest_data.pgdir = cr3; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 421 | lazy_hcall(LHCALL_NEW_PGTABLE, cr3, 0, 0); |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 422 | cr3_changed = true; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 423 | } |
| 424 | |
| 425 | static unsigned long lguest_read_cr3(void) |
| 426 | { |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 427 | return lguest_data.pgdir; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 428 | } |
| 429 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 430 | /* 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] | 431 | static unsigned long lguest_read_cr4(void) |
| 432 | { |
| 433 | return 0; |
| 434 | } |
| 435 | |
| 436 | static void lguest_write_cr4(unsigned long val) |
| 437 | { |
| 438 | } |
| 439 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 440 | /* |
| 441 | * Page Table Handling. |
| 442 | * |
| 443 | * Now would be a good time to take a rest and grab a coffee or similarly |
| 444 | * relaxing stimulant. The easy parts are behind us, and the trek gradually |
| 445 | * winds uphill from here. |
| 446 | * |
| 447 | * Quick refresher: memory is divided into "pages" of 4096 bytes each. The CPU |
| 448 | * maps virtual addresses to physical addresses using "page tables". We could |
| 449 | * use one huge index of 1 million entries: each address is 4 bytes, so that's |
| 450 | * 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] | 451 | * 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] | 452 | * contains the physical address of the top level "page directory" page, which |
| 453 | * contains physical addresses of up to 1024 second-level pages. Each of these |
| 454 | * second level pages contains up to 1024 physical addresses of actual pages, |
| 455 | * or Page Table Entries (PTEs). |
| 456 | * |
| 457 | * Here's a diagram, where arrows indicate physical addresses: |
| 458 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 459 | * cr3 ---> +---------+ |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 460 | * | --------->+---------+ |
| 461 | * | | | PADDR1 | |
| 462 | * Top-level | | PADDR2 | |
| 463 | * (PMD) page | | | |
| 464 | * | | Lower-level | |
| 465 | * | | (PTE) page | |
| 466 | * | | | | |
| 467 | * .... .... |
| 468 | * |
| 469 | * So to convert a virtual address to a physical address, we look up the top |
| 470 | * level, which points us to the second level, which gives us the physical |
| 471 | * address of that page. If the top level entry was not present, or the second |
| 472 | * level entry was not present, then the virtual address is invalid (we |
| 473 | * say "the page was not mapped"). |
| 474 | * |
| 475 | * Put another way, a 32-bit virtual address is divided up like so: |
| 476 | * |
| 477 | * 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 |
| 478 | * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>| |
| 479 | * Index into top Index into second Offset within page |
| 480 | * page directory page pagetable page |
| 481 | * |
| 482 | * The kernel spends a lot of time changing both the top-level page directory |
| 483 | * and lower-level pagetable pages. The Guest doesn't know physical addresses, |
| 484 | * so while it maintains these page tables exactly like normal, it also needs |
| 485 | * to keep the Host informed whenever it makes a change: the Host will create |
| 486 | * the real page tables based on the Guests'. |
| 487 | */ |
| 488 | |
| 489 | /* The Guest calls this to set a second-level entry (pte), ie. to map a page |
| 490 | * into a process' address space. We set the entry then tell the Host the |
| 491 | * toplevel and address this corresponds to. The Guest uses one pagetable per |
| 492 | * process, so we need to tell the Host which one we're changing (mm->pgd). */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 493 | static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr, |
| 494 | pte_t *ptep, pte_t pteval) |
| 495 | { |
| 496 | *ptep = pteval; |
| 497 | lazy_hcall(LHCALL_SET_PTE, __pa(mm->pgd), addr, pteval.pte_low); |
| 498 | } |
| 499 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 500 | /* The Guest calls this to set a top-level entry. Again, we set the entry then |
| 501 | * tell the Host which top-level page we changed, and the index of the entry we |
| 502 | * changed. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 503 | static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) |
| 504 | { |
| 505 | *pmdp = pmdval; |
| 506 | lazy_hcall(LHCALL_SET_PMD, __pa(pmdp)&PAGE_MASK, |
Rusty Russell | 4357bd9 | 2008-03-11 09:35:57 -0500 | [diff] [blame] | 507 | (__pa(pmdp)&(PAGE_SIZE-1))/4, 0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 508 | } |
| 509 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 510 | /* There are a couple of legacy places where the kernel sets a PTE, but we |
| 511 | * don't know the top level any more. This is useless for us, since we don't |
| 512 | * know which pagetable is changing or what address, so we just tell the Host |
| 513 | * to forget all of them. Fortunately, this is very rare. |
| 514 | * |
| 515 | * ... except in early boot when the kernel sets up the initial pagetables, |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 516 | * which makes booting astonishingly slow: 1.83 seconds! So we don't even tell |
| 517 | * the Host anything changed until we've done the first page table switch, |
| 518 | * which brings boot back to 0.25 seconds. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 519 | static void lguest_set_pte(pte_t *ptep, pte_t pteval) |
| 520 | { |
| 521 | *ptep = pteval; |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 522 | if (cr3_changed) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 523 | lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0); |
| 524 | } |
| 525 | |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 526 | /* 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] | 527 | * native page table operations. On native hardware you can set a new page |
| 528 | * table entry whenever you want, but if you want to remove one you have to do |
| 529 | * a TLB flush (a TLB is a little cache of page table entries kept by the CPU). |
| 530 | * |
| 531 | * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only |
| 532 | * called when a valid entry is written, not when it's removed (ie. marked not |
| 533 | * present). Instead, this is where we come when the Guest wants to remove a |
| 534 | * page table entry: we tell the Host to set that entry to 0 (ie. the present |
| 535 | * bit is zero). */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 536 | static void lguest_flush_tlb_single(unsigned long addr) |
| 537 | { |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 538 | /* Simply set it to zero: if it was not, it will fault back in. */ |
Rusty Russell | ad5173f | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 539 | lazy_hcall(LHCALL_SET_PTE, lguest_data.pgdir, addr, 0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 540 | } |
| 541 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 542 | /* This is what happens after the Guest has removed a large number of entries. |
| 543 | * This tells the Host that any of the page table entries for userspace might |
| 544 | * have changed, ie. virtual addresses below PAGE_OFFSET. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 545 | static void lguest_flush_tlb_user(void) |
| 546 | { |
| 547 | lazy_hcall(LHCALL_FLUSH_TLB, 0, 0, 0); |
| 548 | } |
| 549 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 550 | /* This is called when the kernel page tables have changed. That's not very |
| 551 | * common (unless the Guest is using highmem, which makes the Guest extremely |
| 552 | * slow), so it's worth separating this from the user flushing above. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 553 | static void lguest_flush_tlb_kernel(void) |
| 554 | { |
| 555 | lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0); |
| 556 | } |
| 557 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 558 | /* |
| 559 | * The Unadvanced Programmable Interrupt Controller. |
| 560 | * |
| 561 | * This is an attempt to implement the simplest possible interrupt controller. |
| 562 | * I spent some time looking though routines like set_irq_chip_and_handler, |
| 563 | * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and |
| 564 | * I *think* this is as simple as it gets. |
| 565 | * |
| 566 | * We can tell the Host what interrupts we want blocked ready for using the |
| 567 | * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as |
| 568 | * simple as setting a bit. We don't actually "ack" interrupts as such, we |
| 569 | * just mask and unmask them. I wonder if we should be cleverer? |
| 570 | */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 571 | static void disable_lguest_irq(unsigned int irq) |
| 572 | { |
| 573 | set_bit(irq, lguest_data.blocked_interrupts); |
| 574 | } |
| 575 | |
| 576 | static void enable_lguest_irq(unsigned int irq) |
| 577 | { |
| 578 | clear_bit(irq, lguest_data.blocked_interrupts); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 579 | } |
| 580 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 581 | /* This structure describes the lguest IRQ controller. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 582 | static struct irq_chip lguest_irq_controller = { |
| 583 | .name = "lguest", |
| 584 | .mask = disable_lguest_irq, |
| 585 | .mask_ack = disable_lguest_irq, |
| 586 | .unmask = enable_lguest_irq, |
| 587 | }; |
| 588 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 589 | /* This sets up the Interrupt Descriptor Table (IDT) entry for each hardware |
| 590 | * interrupt (except 128, which is used for system calls), and then tells the |
| 591 | * Linux infrastructure that each interrupt is controlled by our level-based |
| 592 | * lguest interrupt controller. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 593 | static void __init lguest_init_IRQ(void) |
| 594 | { |
| 595 | unsigned int i; |
| 596 | |
| 597 | for (i = 0; i < LGUEST_IRQS; i++) { |
| 598 | int vector = FIRST_EXTERNAL_VECTOR + i; |
Rusty Russell | 526e5ab | 2008-10-31 11:24:27 -0500 | [diff] [blame] | 599 | /* Some systems map "vectors" to interrupts weirdly. Lguest has |
| 600 | * a straightforward 1 to 1 mapping, so force that here. */ |
| 601 | __get_cpu_var(vector_irq)[vector] = i; |
Rusty Russell | 6db6a5f | 2009-03-09 10:06:28 -0600 | [diff] [blame] | 602 | if (vector != SYSCALL_VECTOR) |
| 603 | set_intr_gate(vector, interrupt[i]); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 604 | } |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 605 | /* This call is required to set up for 4k stacks, where we have |
| 606 | * separate stacks for hard and soft interrupts. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 607 | irq_ctx_init(smp_processor_id()); |
| 608 | } |
| 609 | |
Rusty Russell | 6db6a5f | 2009-03-09 10:06:28 -0600 | [diff] [blame] | 610 | void lguest_setup_irq(unsigned int irq) |
| 611 | { |
| 612 | irq_to_desc_alloc_cpu(irq, 0); |
| 613 | set_irq_chip_and_handler_name(irq, &lguest_irq_controller, |
| 614 | handle_level_irq, "level"); |
| 615 | } |
| 616 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 617 | /* |
| 618 | * Time. |
| 619 | * |
| 620 | * 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] | 621 | * until then the Host gives us the time on every interrupt. |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 622 | */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 623 | static unsigned long lguest_get_wallclock(void) |
| 624 | { |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 625 | return lguest_data.time.tv_sec; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 626 | } |
| 627 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 628 | /* The TSC is an Intel thing called the Time Stamp Counter. The Host tells us |
| 629 | * what speed it runs at, or 0 if it's unusable as a reliable clock source. |
| 630 | * This matches what we want here: if we return 0 from this function, the x86 |
| 631 | * TSC clock will give up and not register itself. */ |
Alok Kataria | e93ef94 | 2008-07-01 11:43:36 -0700 | [diff] [blame] | 632 | static unsigned long lguest_tsc_khz(void) |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 633 | { |
| 634 | return lguest_data.tsc_khz; |
| 635 | } |
| 636 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 637 | /* If we can't use the TSC, the kernel falls back to our lower-priority |
| 638 | * "lguest_clock", where we read the time value given to us by the Host. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 639 | static cycle_t lguest_clock_read(void) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 640 | { |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 641 | unsigned long sec, nsec; |
| 642 | |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 643 | /* Since the time is in two parts (seconds and nanoseconds), we risk |
| 644 | * reading it just as it's changing from 99 & 0.999999999 to 100 and 0, |
| 645 | * and getting 99 and 0. As Linux tends to come apart under the stress |
| 646 | * of time travel, we must be careful: */ |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 647 | do { |
| 648 | /* First we read the seconds part. */ |
| 649 | sec = lguest_data.time.tv_sec; |
| 650 | /* This read memory barrier tells the compiler and the CPU that |
| 651 | * this can't be reordered: we have to complete the above |
| 652 | * before going on. */ |
| 653 | rmb(); |
| 654 | /* Now we read the nanoseconds part. */ |
| 655 | nsec = lguest_data.time.tv_nsec; |
| 656 | /* Make sure we've done that. */ |
| 657 | rmb(); |
| 658 | /* Now if the seconds part has changed, try again. */ |
| 659 | } while (unlikely(lguest_data.time.tv_sec != sec)); |
| 660 | |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 661 | /* Our lguest clock is in real nanoseconds. */ |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 662 | return sec*1000000000ULL + nsec; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 663 | } |
| 664 | |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 665 | /* This is the fallback clocksource: lower priority than the TSC clocksource. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 666 | static struct clocksource lguest_clock = { |
| 667 | .name = "lguest", |
Rusty Russell | 3fabc55 | 2008-03-11 09:35:56 -0500 | [diff] [blame] | 668 | .rating = 200, |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 669 | .read = lguest_clock_read, |
Rusty Russell | 6c8dca5 | 2007-07-27 13:42:52 +1000 | [diff] [blame] | 670 | .mask = CLOCKSOURCE_MASK(64), |
Rusty Russell | 3725009 | 2007-08-09 20:52:35 +1000 | [diff] [blame] | 671 | .mult = 1 << 22, |
| 672 | .shift = 22, |
Tony Breeds | 05aa026 | 2007-10-22 10:56:25 +1000 | [diff] [blame] | 673 | .flags = CLOCK_SOURCE_IS_CONTINUOUS, |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 674 | }; |
| 675 | |
| 676 | /* We also need a "struct clock_event_device": Linux asks us to set it to go |
| 677 | * off some time in the future. Actually, James Morris figured all this out, I |
| 678 | * just applied the patch. */ |
| 679 | static int lguest_clockevent_set_next_event(unsigned long delta, |
| 680 | struct clock_event_device *evt) |
| 681 | { |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 682 | /* FIXME: I don't think this can ever happen, but James tells me he had |
| 683 | * to put this code in. Maybe we should remove it now. Anyone? */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 684 | if (delta < LG_CLOCK_MIN_DELTA) { |
| 685 | if (printk_ratelimit()) |
| 686 | printk(KERN_DEBUG "%s: small delta %lu ns\n", |
Harvey Harrison | 77bf90e | 2008-03-03 11:37:23 -0800 | [diff] [blame] | 687 | __func__, delta); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 688 | return -ETIME; |
| 689 | } |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 690 | |
| 691 | /* Please wake us this far in the future. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 692 | hcall(LHCALL_SET_CLOCKEVENT, delta, 0, 0); |
| 693 | return 0; |
| 694 | } |
| 695 | |
| 696 | static void lguest_clockevent_set_mode(enum clock_event_mode mode, |
| 697 | struct clock_event_device *evt) |
| 698 | { |
| 699 | switch (mode) { |
| 700 | case CLOCK_EVT_MODE_UNUSED: |
| 701 | case CLOCK_EVT_MODE_SHUTDOWN: |
| 702 | /* A 0 argument shuts the clock down. */ |
| 703 | hcall(LHCALL_SET_CLOCKEVENT, 0, 0, 0); |
| 704 | break; |
| 705 | case CLOCK_EVT_MODE_ONESHOT: |
| 706 | /* This is what we expect. */ |
| 707 | break; |
| 708 | case CLOCK_EVT_MODE_PERIODIC: |
| 709 | BUG(); |
Thomas Gleixner | 18de5bc | 2007-07-21 04:37:34 -0700 | [diff] [blame] | 710 | case CLOCK_EVT_MODE_RESUME: |
| 711 | break; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 712 | } |
| 713 | } |
| 714 | |
| 715 | /* This describes our primitive timer chip. */ |
| 716 | static struct clock_event_device lguest_clockevent = { |
| 717 | .name = "lguest", |
| 718 | .features = CLOCK_EVT_FEAT_ONESHOT, |
| 719 | .set_next_event = lguest_clockevent_set_next_event, |
| 720 | .set_mode = lguest_clockevent_set_mode, |
| 721 | .rating = INT_MAX, |
| 722 | .mult = 1, |
| 723 | .shift = 0, |
| 724 | .min_delta_ns = LG_CLOCK_MIN_DELTA, |
| 725 | .max_delta_ns = LG_CLOCK_MAX_DELTA, |
| 726 | }; |
| 727 | |
| 728 | /* This is the Guest timer interrupt handler (hardware interrupt 0). We just |
| 729 | * call the clockevent infrastructure and it does whatever needs doing. */ |
| 730 | static void lguest_time_irq(unsigned int irq, struct irq_desc *desc) |
| 731 | { |
| 732 | unsigned long flags; |
| 733 | |
| 734 | /* Don't interrupt us while this is running. */ |
| 735 | local_irq_save(flags); |
| 736 | lguest_clockevent.event_handler(&lguest_clockevent); |
| 737 | local_irq_restore(flags); |
| 738 | } |
| 739 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 740 | /* At some point in the boot process, we get asked to set up our timing |
| 741 | * infrastructure. The kernel doesn't expect timer interrupts before this, but |
| 742 | * we cleverly initialized the "blocked_interrupts" field of "struct |
| 743 | * lguest_data" so that timer interrupts were blocked until now. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 744 | static void lguest_time_init(void) |
| 745 | { |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 746 | /* 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] | 747 | set_irq_handler(0, lguest_time_irq); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 748 | |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 749 | clocksource_register(&lguest_clock); |
| 750 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 751 | /* We can't set cpumask in the initializer: damn C limitations! Set it |
| 752 | * here and register our timer device. */ |
Rusty Russell | 320ab2b | 2008-12-13 21:20:26 +1030 | [diff] [blame] | 753 | lguest_clockevent.cpumask = cpumask_of(0); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 754 | clockevents_register_device(&lguest_clockevent); |
| 755 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 756 | /* Finally, we unblock the timer interrupt. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 757 | enable_lguest_irq(0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 758 | } |
| 759 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 760 | /* |
| 761 | * Miscellaneous bits and pieces. |
| 762 | * |
| 763 | * Here is an oddball collection of functions which the Guest needs for things |
| 764 | * to work. They're pretty simple. |
| 765 | */ |
| 766 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 767 | /* 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] | 768 | * native hardware, this is part of the Task State Segment mentioned above in |
| 769 | * lguest_load_tr_desc(), but to help hypervisors there's this special call. |
| 770 | * |
| 771 | * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data |
| 772 | * segment), the privilege level (we're privilege level 1, the Host is 0 and |
| 773 | * will not tolerate us trying to use that), the stack pointer, and the number |
| 774 | * of pages in the stack. */ |
H. Peter Anvin | faca622 | 2008-01-30 13:31:02 +0100 | [diff] [blame] | 775 | static void lguest_load_sp0(struct tss_struct *tss, |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 776 | struct thread_struct *thread) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 777 | { |
H. Peter Anvin | faca622 | 2008-01-30 13:31:02 +0100 | [diff] [blame] | 778 | lazy_hcall(LHCALL_SET_STACK, __KERNEL_DS|0x1, thread->sp0, |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 779 | THREAD_SIZE/PAGE_SIZE); |
| 780 | } |
| 781 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 782 | /* Let's just say, I wouldn't do debugging under a Guest. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 783 | static void lguest_set_debugreg(int regno, unsigned long value) |
| 784 | { |
| 785 | /* FIXME: Implement */ |
| 786 | } |
| 787 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 788 | /* There are times when the kernel wants to make sure that no memory writes are |
| 789 | * caught in the cache (that they've all reached real hardware devices). This |
| 790 | * doesn't matter for the Guest which has virtual hardware. |
| 791 | * |
| 792 | * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush |
| 793 | * (clflush) instruction is available and the kernel uses that. Otherwise, it |
| 794 | * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction. |
| 795 | * Unlike clflush, wbinvd can only be run at privilege level 0. So we can |
| 796 | * ignore clflush, but replace wbinvd. |
| 797 | */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 798 | static void lguest_wbinvd(void) |
| 799 | { |
| 800 | } |
| 801 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 802 | /* If the Guest expects to have an Advanced Programmable Interrupt Controller, |
| 803 | * we play dumb by ignoring writes and returning 0 for reads. So it's no |
| 804 | * longer Programmable nor Controlling anything, and I don't think 8 lines of |
| 805 | * code qualifies for Advanced. It will also never interrupt anything. It |
| 806 | * does, however, allow us to get through the Linux boot code. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 807 | #ifdef CONFIG_X86_LOCAL_APIC |
Suresh Siddha | ad66dd3 | 2008-07-11 13:11:56 -0700 | [diff] [blame] | 808 | static void lguest_apic_write(u32 reg, u32 v) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 809 | { |
| 810 | } |
| 811 | |
Suresh Siddha | ad66dd3 | 2008-07-11 13:11:56 -0700 | [diff] [blame] | 812 | static u32 lguest_apic_read(u32 reg) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 813 | { |
| 814 | return 0; |
| 815 | } |
Suresh Siddha | 511d9d3 | 2008-07-14 09:49:14 -0700 | [diff] [blame] | 816 | |
| 817 | static u64 lguest_apic_icr_read(void) |
| 818 | { |
| 819 | return 0; |
| 820 | } |
| 821 | |
| 822 | static void lguest_apic_icr_write(u32 low, u32 id) |
| 823 | { |
| 824 | /* Warn to see if there's any stray references */ |
| 825 | WARN_ON(1); |
| 826 | } |
| 827 | |
| 828 | static void lguest_apic_wait_icr_idle(void) |
| 829 | { |
| 830 | return; |
| 831 | } |
| 832 | |
| 833 | static u32 lguest_apic_safe_wait_icr_idle(void) |
| 834 | { |
| 835 | return 0; |
| 836 | } |
| 837 | |
Yinghai Lu | c1eeb2d | 2009-02-16 23:02:14 -0800 | [diff] [blame] | 838 | static void set_lguest_basic_apic_ops(void) |
| 839 | { |
| 840 | apic->read = lguest_apic_read; |
| 841 | apic->write = lguest_apic_write; |
| 842 | apic->icr_read = lguest_apic_icr_read; |
| 843 | apic->icr_write = lguest_apic_icr_write; |
| 844 | apic->wait_icr_idle = lguest_apic_wait_icr_idle; |
| 845 | apic->safe_wait_icr_idle = lguest_apic_safe_wait_icr_idle; |
Suresh Siddha | 511d9d3 | 2008-07-14 09:49:14 -0700 | [diff] [blame] | 846 | }; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 847 | #endif |
| 848 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 849 | /* STOP! Until an interrupt comes in. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 850 | static void lguest_safe_halt(void) |
| 851 | { |
| 852 | hcall(LHCALL_HALT, 0, 0, 0); |
| 853 | } |
| 854 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 855 | /* The SHUTDOWN hypercall takes a string to describe what's happening, and |
| 856 | * 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] | 857 | * |
| 858 | * Note that the Host always prefers that the Guest speak in physical addresses |
| 859 | * rather than virtual addresses, so we use __pa() here. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 860 | static void lguest_power_off(void) |
| 861 | { |
Balaji Rao | ec04b13 | 2007-12-28 14:26:24 +0530 | [diff] [blame] | 862 | hcall(LHCALL_SHUTDOWN, __pa("Power down"), LGUEST_SHUTDOWN_POWEROFF, 0); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 863 | } |
| 864 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 865 | /* |
| 866 | * Panicing. |
| 867 | * |
| 868 | * Don't. But if you did, this is what happens. |
| 869 | */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 870 | static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p) |
| 871 | { |
Balaji Rao | ec04b13 | 2007-12-28 14:26:24 +0530 | [diff] [blame] | 872 | hcall(LHCALL_SHUTDOWN, __pa(p), LGUEST_SHUTDOWN_POWEROFF, 0); |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 873 | /* 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] | 874 | return NOTIFY_DONE; |
| 875 | } |
| 876 | |
| 877 | static struct notifier_block paniced = { |
| 878 | .notifier_call = lguest_panic |
| 879 | }; |
| 880 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 881 | /* Setting up memory is fairly easy. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 882 | static __init char *lguest_memory_setup(void) |
| 883 | { |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 884 | /* We do this here and not earlier because lockcheck used to barf if we |
| 885 | * did it before start_kernel(). I think we fixed that, so it'd be |
| 886 | * nice to move it back to lguest_init. Patch welcome... */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 887 | atomic_notifier_chain_register(&panic_notifier_list, &paniced); |
| 888 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 889 | /* The Linux bootloader header contains an "e820" memory map: the |
| 890 | * Launcher populated the first entry with our memory limit. */ |
Yinghai Lu | d0be6bd | 2008-06-15 18:58:51 -0700 | [diff] [blame] | 891 | e820_add_region(boot_params.e820_map[0].addr, |
H. Peter Anvin | 30c8264 | 2007-10-15 17:13:22 -0700 | [diff] [blame] | 892 | boot_params.e820_map[0].size, |
| 893 | boot_params.e820_map[0].type); |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 894 | |
| 895 | /* This string is for the boot messages. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 896 | return "LGUEST"; |
| 897 | } |
| 898 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 899 | /* We will eventually use the virtio console device to produce console output, |
| 900 | * but before that is set up we use LHCALL_NOTIFY on normal memory to produce |
| 901 | * console output. */ |
Rusty Russell | 19f1537 | 2007-10-22 11:24:21 +1000 | [diff] [blame] | 902 | static __init int early_put_chars(u32 vtermno, const char *buf, int count) |
| 903 | { |
| 904 | char scratch[17]; |
| 905 | unsigned int len = count; |
| 906 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 907 | /* We use a nul-terminated string, so we have to make a copy. Icky, |
| 908 | * huh? */ |
Rusty Russell | 19f1537 | 2007-10-22 11:24:21 +1000 | [diff] [blame] | 909 | if (len > sizeof(scratch) - 1) |
| 910 | len = sizeof(scratch) - 1; |
| 911 | scratch[len] = '\0'; |
| 912 | memcpy(scratch, buf, len); |
| 913 | hcall(LHCALL_NOTIFY, __pa(scratch), 0, 0); |
| 914 | |
| 915 | /* This routine returns the number of bytes actually written. */ |
| 916 | return len; |
| 917 | } |
| 918 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 919 | /* Rebooting also tells the Host we're finished, but the RESTART flag tells the |
| 920 | * Launcher to reboot us. */ |
| 921 | static void lguest_restart(char *reason) |
| 922 | { |
| 923 | hcall(LHCALL_SHUTDOWN, __pa(reason), LGUEST_SHUTDOWN_RESTART, 0); |
| 924 | } |
| 925 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 926 | /*G:050 |
| 927 | * Patching (Powerfully Placating Performance Pedants) |
| 928 | * |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 929 | * We have already seen that pv_ops structures let us replace simple native |
| 930 | * instructions with calls to the appropriate back end all throughout the |
| 931 | * 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] | 932 | * kernel, but it's slow because of all the indirect branches. |
| 933 | * |
| 934 | * Remember that David Wheeler quote about "Any problem in computer science can |
| 935 | * be solved with another layer of indirection"? The rest of that quote is |
| 936 | * "... But that usually will create another problem." This is the first of |
| 937 | * those problems. |
| 938 | * |
| 939 | * Our current solution is to allow the paravirt back end to optionally patch |
| 940 | * over the indirect calls to replace them with something more efficient. We |
| 941 | * patch the four most commonly called functions: disable interrupts, enable |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 942 | * interrupts, restore interrupts and save interrupts. We usually have 6 or 10 |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 943 | * bytes to patch into: the Guest versions of these operations are small enough |
| 944 | * that we can fit comfortably. |
| 945 | * |
| 946 | * First we need assembly templates of each of the patchable Guest operations, |
Atsushi SAKAI | 72410af | 2009-01-16 20:39:14 +0900 | [diff] [blame] | 947 | * and these are in i386_head.S. */ |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 948 | |
| 949 | /*G:060 We construct a table from the assembler templates: */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 950 | static const struct lguest_insns |
| 951 | { |
| 952 | const char *start, *end; |
| 953 | } lguest_insns[] = { |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 954 | [PARAVIRT_PATCH(pv_irq_ops.irq_disable)] = { lgstart_cli, lgend_cli }, |
| 955 | [PARAVIRT_PATCH(pv_irq_ops.irq_enable)] = { lgstart_sti, lgend_sti }, |
| 956 | [PARAVIRT_PATCH(pv_irq_ops.restore_fl)] = { lgstart_popf, lgend_popf }, |
| 957 | [PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf }, |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 958 | }; |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 959 | |
| 960 | /* Now our patch routine is fairly simple (based on the native one in |
| 961 | * paravirt.c). If we have a replacement, we copy it in and return how much of |
| 962 | * the available space we used. */ |
Andi Kleen | ab144f5 | 2007-08-10 22:31:03 +0200 | [diff] [blame] | 963 | static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf, |
| 964 | unsigned long addr, unsigned len) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 965 | { |
| 966 | unsigned int insn_len; |
| 967 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 968 | /* Don't do anything special if we don't have a replacement */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 969 | if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start) |
Andi Kleen | ab144f5 | 2007-08-10 22:31:03 +0200 | [diff] [blame] | 970 | return paravirt_patch_default(type, clobber, ibuf, addr, len); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 971 | |
| 972 | insn_len = lguest_insns[type].end - lguest_insns[type].start; |
| 973 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 974 | /* Similarly if we can't fit replacement (shouldn't happen, but let's |
| 975 | * be thorough). */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 976 | if (len < insn_len) |
Andi Kleen | ab144f5 | 2007-08-10 22:31:03 +0200 | [diff] [blame] | 977 | return paravirt_patch_default(type, clobber, ibuf, addr, len); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 978 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 979 | /* Copy in our instructions. */ |
Andi Kleen | ab144f5 | 2007-08-10 22:31:03 +0200 | [diff] [blame] | 980 | memcpy(ibuf, lguest_insns[type].start, insn_len); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 981 | return insn_len; |
| 982 | } |
| 983 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 984 | /*G:030 Once we get to lguest_init(), we know we're a Guest. The various |
| 985 | * pv_ops structures in the kernel provide points for (almost) every routine we |
| 986 | * have to override to avoid privileged instructions. */ |
Rusty Russell | 814a0e5 | 2007-10-22 11:29:44 +1000 | [diff] [blame] | 987 | __init void lguest_init(void) |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 988 | { |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 989 | /* We're under lguest, paravirt is enabled, and we're running at |
| 990 | * privilege level 1, not 0 as normal. */ |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 991 | pv_info.name = "lguest"; |
| 992 | pv_info.paravirt_enabled = 1; |
| 993 | pv_info.kernel_rpl = 1; |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 994 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 995 | /* We set up all the lguest overrides for sensitive operations. These |
| 996 | * are detailed with the operations themselves. */ |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 997 | |
| 998 | /* interrupt-related operations */ |
| 999 | pv_irq_ops.init_IRQ = lguest_init_IRQ; |
Jeremy Fitzhardinge | ecb93d1 | 2009-01-28 14:35:05 -0800 | [diff] [blame] | 1000 | pv_irq_ops.save_fl = PV_CALLEE_SAVE(save_fl); |
| 1001 | pv_irq_ops.restore_fl = PV_CALLEE_SAVE(restore_fl); |
| 1002 | pv_irq_ops.irq_disable = PV_CALLEE_SAVE(irq_disable); |
| 1003 | pv_irq_ops.irq_enable = PV_CALLEE_SAVE(irq_enable); |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1004 | pv_irq_ops.safe_halt = lguest_safe_halt; |
| 1005 | |
| 1006 | /* init-time operations */ |
| 1007 | pv_init_ops.memory_setup = lguest_memory_setup; |
| 1008 | pv_init_ops.patch = lguest_patch; |
| 1009 | |
| 1010 | /* Intercepts of various cpu instructions */ |
| 1011 | pv_cpu_ops.load_gdt = lguest_load_gdt; |
| 1012 | pv_cpu_ops.cpuid = lguest_cpuid; |
| 1013 | pv_cpu_ops.load_idt = lguest_load_idt; |
| 1014 | pv_cpu_ops.iret = lguest_iret; |
H. Peter Anvin | faca622 | 2008-01-30 13:31:02 +0100 | [diff] [blame] | 1015 | pv_cpu_ops.load_sp0 = lguest_load_sp0; |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1016 | pv_cpu_ops.load_tr_desc = lguest_load_tr_desc; |
| 1017 | pv_cpu_ops.set_ldt = lguest_set_ldt; |
| 1018 | pv_cpu_ops.load_tls = lguest_load_tls; |
| 1019 | pv_cpu_ops.set_debugreg = lguest_set_debugreg; |
| 1020 | pv_cpu_ops.clts = lguest_clts; |
| 1021 | pv_cpu_ops.read_cr0 = lguest_read_cr0; |
| 1022 | pv_cpu_ops.write_cr0 = lguest_write_cr0; |
| 1023 | pv_cpu_ops.read_cr4 = lguest_read_cr4; |
| 1024 | pv_cpu_ops.write_cr4 = lguest_write_cr4; |
| 1025 | pv_cpu_ops.write_gdt_entry = lguest_write_gdt_entry; |
| 1026 | pv_cpu_ops.write_idt_entry = lguest_write_idt_entry; |
| 1027 | pv_cpu_ops.wbinvd = lguest_wbinvd; |
Jeremy Fitzhardinge | 8965c1c | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1028 | pv_cpu_ops.lazy_mode.enter = paravirt_enter_lazy_cpu; |
| 1029 | pv_cpu_ops.lazy_mode.leave = lguest_leave_lazy_mode; |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1030 | |
| 1031 | /* pagetable management */ |
| 1032 | pv_mmu_ops.write_cr3 = lguest_write_cr3; |
| 1033 | pv_mmu_ops.flush_tlb_user = lguest_flush_tlb_user; |
| 1034 | pv_mmu_ops.flush_tlb_single = lguest_flush_tlb_single; |
| 1035 | pv_mmu_ops.flush_tlb_kernel = lguest_flush_tlb_kernel; |
| 1036 | pv_mmu_ops.set_pte = lguest_set_pte; |
| 1037 | pv_mmu_ops.set_pte_at = lguest_set_pte_at; |
| 1038 | pv_mmu_ops.set_pmd = lguest_set_pmd; |
| 1039 | pv_mmu_ops.read_cr2 = lguest_read_cr2; |
| 1040 | pv_mmu_ops.read_cr3 = lguest_read_cr3; |
Jeremy Fitzhardinge | 8965c1c | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1041 | pv_mmu_ops.lazy_mode.enter = paravirt_enter_lazy_mmu; |
| 1042 | pv_mmu_ops.lazy_mode.leave = lguest_leave_lazy_mode; |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1043 | |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1044 | #ifdef CONFIG_X86_LOCAL_APIC |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1045 | /* apic read/write intercepts */ |
Yinghai Lu | c1eeb2d | 2009-02-16 23:02:14 -0800 | [diff] [blame] | 1046 | set_lguest_basic_apic_ops(); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1047 | #endif |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1048 | |
| 1049 | /* time operations */ |
| 1050 | pv_time_ops.get_wallclock = lguest_get_wallclock; |
| 1051 | pv_time_ops.time_init = lguest_time_init; |
Alok Kataria | e93ef94 | 2008-07-01 11:43:36 -0700 | [diff] [blame] | 1052 | pv_time_ops.get_tsc_khz = lguest_tsc_khz; |
Jeremy Fitzhardinge | 93b1eab | 2007-10-16 11:51:29 -0700 | [diff] [blame] | 1053 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1054 | /* Now is a good time to look at the implementations of these functions |
| 1055 | * before returning to the rest of lguest_init(). */ |
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 | /*G:070 Now we've seen all the paravirt_ops, we return to |
| 1058 | * lguest_init() where the rest of the fairly chaotic boot setup |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 1059 | * occurs. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1060 | |
Rusty Russell | 5d006d8 | 2008-07-29 09:58:29 -0500 | [diff] [blame] | 1061 | /* As described in head_32.S, we map the first 128M of memory. */ |
| 1062 | max_pfn_mapped = (128*1024*1024) >> PAGE_SHIFT; |
| 1063 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1064 | /* Load the %fs segment register (the per-cpu segment register) with |
| 1065 | * the normal data segment to get through booting. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1066 | asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS) : "memory"); |
| 1067 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1068 | /* The Host<->Guest Switcher lives at the top of our address space, and |
| 1069 | * the Host told us how big it is when we made LGUEST_INIT hypercall: |
| 1070 | * it put the answer in lguest_data.reserve_mem */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1071 | reserve_top_address(lguest_data.reserve_mem); |
| 1072 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1073 | /* If we don't initialize the lock dependency checker now, it crashes |
| 1074 | * paravirt_disable_iospace. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1075 | lockdep_init(); |
| 1076 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1077 | /* The IDE code spends about 3 seconds probing for disks: if we reserve |
| 1078 | * all the I/O ports up front it can't get them and so doesn't probe. |
| 1079 | * Other device drivers are similar (but less severe). This cuts the |
| 1080 | * 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] | 1081 | paravirt_disable_iospace(); |
| 1082 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1083 | /* This is messy CPU setup stuff which the native boot code does before |
| 1084 | * start_kernel, so we have to do, too: */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1085 | cpu_detect(&new_cpu_data); |
| 1086 | /* head.S usually sets up the first capability word, so do it here. */ |
| 1087 | new_cpu_data.x86_capability[0] = cpuid_edx(1); |
| 1088 | |
| 1089 | /* Math is always hard! */ |
| 1090 | new_cpu_data.hard_math = 1; |
| 1091 | |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1092 | /* We don't have features. We have puppies! Puppies! */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1093 | #ifdef CONFIG_X86_MCE |
| 1094 | mce_disabled = 1; |
| 1095 | #endif |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1096 | #ifdef CONFIG_ACPI |
| 1097 | acpi_disabled = 1; |
| 1098 | acpi_ht = 0; |
| 1099 | #endif |
| 1100 | |
Atsushi SAKAI | 72410af | 2009-01-16 20:39:14 +0900 | [diff] [blame] | 1101 | /* We set the preferred console to "hvc". This is the "hypervisor |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1102 | * virtual console" driver written by the PowerPC people, which we also |
| 1103 | * adapted for lguest's use. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1104 | add_preferred_console("hvc", 0, NULL); |
| 1105 | |
Rusty Russell | 19f1537 | 2007-10-22 11:24:21 +1000 | [diff] [blame] | 1106 | /* Register our very early console. */ |
| 1107 | virtio_cons_early_init(early_put_chars); |
| 1108 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1109 | /* 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] | 1110 | * the Guest routine to power off, and the reboot hook to our restart |
| 1111 | * routine. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1112 | pm_power_off = lguest_power_off; |
Balaji Rao | ec04b13 | 2007-12-28 14:26:24 +0530 | [diff] [blame] | 1113 | machine_ops.restart = lguest_restart; |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1114 | |
Yinghai Lu | f0d4310 | 2008-05-29 12:56:36 -0700 | [diff] [blame] | 1115 | /* 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] | 1116 | * to boot as normal. It never returns. */ |
Yinghai Lu | f0d4310 | 2008-05-29 12:56:36 -0700 | [diff] [blame] | 1117 | i386_start_kernel(); |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1118 | } |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1119 | /* |
| 1120 | * This marks the end of stage II of our journey, The Guest. |
| 1121 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 1122 | * It is now time for us to explore the layer of virtual drivers and complete |
| 1123 | * our understanding of the Guest in "make Drivers". |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1124 | */ |