blob: c4fb424dfddbac93d4a191f99dd93627136f5aa6 [file] [log] [blame]
Rusty Russell2e04ef72009-07-30 16:03:45 -06001/*P:600
2 * The x86 architecture has segments, which involve a table of descriptors
Rusty Russellf938d2c2007-07-26 10:41:02 -07003 * which can be used to do funky things with virtual address interpretation.
4 * We originally used to use segments so the Guest couldn't alter the
5 * Guest<->Host Switcher, and then we had to trim Guest segments, and restore
6 * for userspace per-thread segments, but trim again for on userspace->kernel
7 * transitions... This nightmarish creation was contained within this file,
8 * where we knew not to tread without heavy armament and a change of underwear.
9 *
10 * In these modern times, the segment handling code consists of simple sanity
11 * checks, and the worst you'll experience reading this code is butterfly-rash
Rusty Russell2e04ef72009-07-30 16:03:45 -060012 * from frolicking through its parklike serenity.
13:*/
Rusty Russelld7e28ff2007-07-19 01:49:23 -070014#include "lg.h"
15
Rusty Russellbff672e2007-07-26 10:41:04 -070016/*H:600
Rusty Russellbff672e2007-07-26 10:41:04 -070017 * Segments & The Global Descriptor Table
18 *
19 * (That title sounds like a bad Nerdcore group. Not to suggest that there are
20 * any good Nerdcore groups, but in high school a friend of mine had a band
21 * called Joe Fish and the Chips, so there are definitely worse band names).
22 *
23 * To refresh: the GDT is a table of 8-byte values describing segments. Once
24 * set up, these segments can be loaded into one of the 6 "segment registers".
25 *
26 * GDT entries are passed around as "struct desc_struct"s, which like IDT
27 * entries are split into two 32-bit members, "a" and "b". One day, someone
28 * will clean that up, and be declared a Hero. (No pressure, I'm just saying).
29 *
30 * Anyway, the GDT entry contains a base (the start address of the segment), a
31 * limit (the size of the segment - 1), and some flags. Sounds simple, and it
32 * would be, except those zany Intel engineers decided that it was too boring
33 * to put the base at one end, the limit at the other, and the flags in
34 * between. They decided to shotgun the bits at random throughout the 8 bytes,
35 * like so:
36 *
37 * 0 16 40 48 52 56 63
38 * [ limit part 1 ][ base part 1 ][ flags ][li][fl][base ]
39 * mit ags part 2
40 * part 2
41 *
42 * As a result, this file contains a certain amount of magic numeracy. Let's
43 * begin.
44 */
45
Rusty Russell2e04ef72009-07-30 16:03:45 -060046/*
47 * There are several entries we don't let the Guest set. The TSS entry is the
Rusty Russellbff672e2007-07-26 10:41:04 -070048 * "Task State Segment" which controls all kinds of delicate things. The
49 * LGUEST_CS and LGUEST_DS entries are reserved for the Switcher, and the
Rusty Russell2e04ef72009-07-30 16:03:45 -060050 * the Guest can't be trusted to deal with double faults.
51 */
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -030052static bool ignored_gdt(unsigned int num)
Rusty Russelld7e28ff2007-07-19 01:49:23 -070053{
54 return (num == GDT_ENTRY_TSS
55 || num == GDT_ENTRY_LGUEST_CS
56 || num == GDT_ENTRY_LGUEST_DS
57 || num == GDT_ENTRY_DOUBLEFAULT_TSS);
58}
59
Rusty Russell2e04ef72009-07-30 16:03:45 -060060/*H:630
61 * Once the Guest gave us new GDT entries, we fix them up a little. We
Rusty Russell0d027c02007-08-09 20:57:13 +100062 * don't care if they're invalid: the worst that can happen is a General
63 * Protection Fault in the Switcher when it restores a Guest segment register
64 * which tries to use that entry. Then we kill the Guest for causing such a
Rusty Russell2e04ef72009-07-30 16:03:45 -060065 * mess: the message will be "unhandled trap 256".
66 */
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -020067static void fixup_gdt_table(struct lg_cpu *cpu, unsigned start, unsigned end)
Rusty Russelld7e28ff2007-07-19 01:49:23 -070068{
69 unsigned int i;
70
71 for (i = start; i < end; i++) {
Rusty Russell2e04ef72009-07-30 16:03:45 -060072 /*
73 * We never copy these ones to real GDT, so we don't care what
74 * they say
75 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -070076 if (ignored_gdt(i))
77 continue;
78
Rusty Russell2e04ef72009-07-30 16:03:45 -060079 /*
80 * Segment descriptors contain a privilege level: the Guest is
Rusty Russellbff672e2007-07-26 10:41:04 -070081 * sometimes careless and leaves this as 0, even though it's
Rusty Russell2e04ef72009-07-30 16:03:45 -060082 * running at privilege level 1. If so, we fix it here.
83 */
Jacek Galowicz39082f72012-01-12 15:44:47 +103084 if (cpu->arch.gdt[i].dpl == 0)
85 cpu->arch.gdt[i].dpl |= GUEST_PL;
Rusty Russelld7e28ff2007-07-19 01:49:23 -070086
Rusty Russell2e04ef72009-07-30 16:03:45 -060087 /*
88 * Each descriptor has an "accessed" bit. If we don't set it
Rusty Russellbff672e2007-07-26 10:41:04 -070089 * now, the CPU will try to set it when the Guest first loads
90 * that entry into a segment register. But the GDT isn't
Rusty Russell2e04ef72009-07-30 16:03:45 -060091 * writable by the Guest, so bad things can happen.
92 */
Jacek Galowicz39082f72012-01-12 15:44:47 +103093 cpu->arch.gdt[i].type |= 0x1;
Rusty Russelld7e28ff2007-07-19 01:49:23 -070094 }
95}
96
Rusty Russell2e04ef72009-07-30 16:03:45 -060097/*H:610
98 * Like the IDT, we never simply use the GDT the Guest gives us. We keep
Rusty Russelle1e72962007-10-25 15:02:50 +100099 * a GDT for each CPU, and copy across the Guest's entries each time we want to
100 * run the Guest on that CPU.
101 *
102 * This routine is called at boot or modprobe time for each CPU to set up the
103 * constant GDT entries: the ones which are the same no matter what Guest we're
Rusty Russell2e04ef72009-07-30 16:03:45 -0600104 * running.
105 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700106void setup_default_gdt_entries(struct lguest_ro_state *state)
107{
108 struct desc_struct *gdt = state->guest_gdt;
109 unsigned long tss = (unsigned long)&state->guest_tss;
110
Rusty Russelle1e72962007-10-25 15:02:50 +1000111 /* The Switcher segments are full 0-4G segments, privilege level 0 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700112 gdt[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT;
113 gdt[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT;
114
Rusty Russell2e04ef72009-07-30 16:03:45 -0600115 /*
116 * The TSS segment refers to the TSS entry for this particular CPU.
Rusty Russell2e04ef72009-07-30 16:03:45 -0600117 */
Jacek Galowicz39082f72012-01-12 15:44:47 +1030118 gdt[GDT_ENTRY_TSS].a = 0;
119 gdt[GDT_ENTRY_TSS].b = 0;
120
121 gdt[GDT_ENTRY_TSS].limit0 = 0x67;
122 gdt[GDT_ENTRY_TSS].base0 = tss & 0xFFFF;
123 gdt[GDT_ENTRY_TSS].base1 = (tss >> 16) & 0xFF;
124 gdt[GDT_ENTRY_TSS].base2 = tss >> 24;
125 gdt[GDT_ENTRY_TSS].type = 0x9; /* 32-bit TSS (available) */
126 gdt[GDT_ENTRY_TSS].p = 0x1; /* Entry is present */
127 gdt[GDT_ENTRY_TSS].dpl = 0x0; /* Privilege level 0 */
128 gdt[GDT_ENTRY_TSS].s = 0x0; /* system segment */
129
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700130}
131
Rusty Russell2e04ef72009-07-30 16:03:45 -0600132/*
133 * This routine sets up the initial Guest GDT for booting. All entries start
134 * as 0 (unusable).
135 */
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200136void setup_guest_gdt(struct lg_cpu *cpu)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700137{
Rusty Russell2e04ef72009-07-30 16:03:45 -0600138 /*
139 * Start with full 0-4G segments...except the Guest is allowed to use
140 * them, so set the privilege level appropriately in the flags.
141 */
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200142 cpu->arch.gdt[GDT_ENTRY_KERNEL_CS] = FULL_EXEC_SEGMENT;
143 cpu->arch.gdt[GDT_ENTRY_KERNEL_DS] = FULL_SEGMENT;
Jacek Galowicz39082f72012-01-12 15:44:47 +1030144 cpu->arch.gdt[GDT_ENTRY_KERNEL_CS].dpl |= GUEST_PL;
145 cpu->arch.gdt[GDT_ENTRY_KERNEL_DS].dpl |= GUEST_PL;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700146}
147
Rusty Russell2e04ef72009-07-30 16:03:45 -0600148/*H:650
149 * An optimization of copy_gdt(), for just the three "thead-local storage"
150 * entries.
151 */
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200152void copy_gdt_tls(const struct lg_cpu *cpu, struct desc_struct *gdt)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700153{
154 unsigned int i;
155
156 for (i = GDT_ENTRY_TLS_MIN; i <= GDT_ENTRY_TLS_MAX; i++)
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200157 gdt[i] = cpu->arch.gdt[i];
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700158}
159
Rusty Russell2e04ef72009-07-30 16:03:45 -0600160/*H:640
161 * When the Guest is run on a different CPU, or the GDT entries have changed,
162 * copy_gdt() is called to copy the Guest's GDT entries across to this CPU's
163 * GDT.
164 */
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200165void copy_gdt(const struct lg_cpu *cpu, struct desc_struct *gdt)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700166{
167 unsigned int i;
168
Rusty Russell2e04ef72009-07-30 16:03:45 -0600169 /*
170 * The default entries from setup_default_gdt_entries() are not
171 * replaced. See ignored_gdt() above.
172 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700173 for (i = 0; i < GDT_ENTRIES; i++)
174 if (!ignored_gdt(i))
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200175 gdt[i] = cpu->arch.gdt[i];
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700176}
177
Rusty Russell2e04ef72009-07-30 16:03:45 -0600178/*H:620
179 * This is where the Guest asks us to load a new GDT entry
180 * (LHCALL_LOAD_GDT_ENTRY). We tweak the entry and copy it in.
181 */
Rusty Russella489f0b2009-04-19 23:14:00 -0600182void load_guest_gdt_entry(struct lg_cpu *cpu, u32 num, u32 lo, u32 hi)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700183{
Rusty Russell2e04ef72009-07-30 16:03:45 -0600184 /*
185 * We assume the Guest has the same number of GDT entries as the
186 * Host, otherwise we'd have to dynamically allocate the Guest GDT.
187 */
Rusty Russell3e272492010-01-04 19:26:14 +1030188 if (num >= ARRAY_SIZE(cpu->arch.gdt)) {
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200189 kill_guest(cpu, "too many gdt entries %i", num);
Rusty Russell3e272492010-01-04 19:26:14 +1030190 return;
191 }
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700192
Rusty Russella489f0b2009-04-19 23:14:00 -0600193 /* Set it up, then fix it. */
194 cpu->arch.gdt[num].a = lo;
195 cpu->arch.gdt[num].b = hi;
196 fixup_gdt_table(cpu, num, num+1);
Rusty Russell2e04ef72009-07-30 16:03:45 -0600197 /*
198 * Mark that the GDT changed so the core knows it has to copy it again,
199 * even if the Guest is run on the same CPU.
200 */
Glauber de Oliveira Costaae3749d2008-01-17 19:14:46 -0200201 cpu->changed |= CHANGED_GDT;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700202}
203
Rusty Russell2e04ef72009-07-30 16:03:45 -0600204/*
205 * This is the fast-track version for just changing the three TLS entries.
Rusty Russelle1e72962007-10-25 15:02:50 +1000206 * Remember that this happens on every context switch, so it's worth
207 * optimizing. But wouldn't it be neater to have a single hypercall to cover
Rusty Russell2e04ef72009-07-30 16:03:45 -0600208 * both cases?
209 */
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200210void guest_load_tls(struct lg_cpu *cpu, unsigned long gtls)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700211{
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200212 struct desc_struct *tls = &cpu->arch.gdt[GDT_ENTRY_TLS_MIN];
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700213
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200214 __lgread(cpu, tls, gtls, sizeof(*tls)*GDT_ENTRY_TLS_ENTRIES);
Glauber de Oliveira Costafc708b32008-01-07 11:05:33 -0200215 fixup_gdt_table(cpu, GDT_ENTRY_TLS_MIN, GDT_ENTRY_TLS_MAX+1);
Rusty Russelle1e72962007-10-25 15:02:50 +1000216 /* Note that just the TLS entries have changed. */
Glauber de Oliveira Costaae3749d2008-01-17 19:14:46 -0200217 cpu->changed |= CHANGED_GDT_TLS;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700218}
Rusty Russellbff672e2007-07-26 10:41:04 -0700219
Rusty Russelle1e72962007-10-25 15:02:50 +1000220/*H:660
Rusty Russellbff672e2007-07-26 10:41:04 -0700221 * With this, we have finished the Host.
222 *
223 * Five of the seven parts of our task are complete. You have made it through
224 * the Bit of Despair (I think that's somewhere in the page table code,
225 * myself).
226 *
227 * Next, we examine "make Switcher". It's short, but intense.
228 */