x86/amd-iommu: Add per IOMMU reference counting
[linux/fpc-iii.git] / drivers / lguest / segments.c
blob951c57b0a7e075a9d927193d3c54d46ccb682545
1 /*P:600
2 * The x86 architecture has segments, which involve a table of descriptors
3 * 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.
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
12 * from frolicking through its parklike serenity.
13 :*/
14 #include "lg.h"
16 /*H:600
17 * Segments & The Global Descriptor Table
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).
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".
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).
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:
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
42 * As a result, this file contains a certain amount of magic numeracy. Let's
43 * begin.
47 * There are several entries we don't let the Guest set. The TSS entry is the
48 * "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
50 * the Guest can't be trusted to deal with double faults.
52 static bool ignored_gdt(unsigned int num)
54 return (num == GDT_ENTRY_TSS
55 || num == GDT_ENTRY_LGUEST_CS
56 || num == GDT_ENTRY_LGUEST_DS
57 || num == GDT_ENTRY_DOUBLEFAULT_TSS);
60 /*H:630
61 * Once the Guest gave us new GDT entries, we fix them up a little. We
62 * 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
65 * mess: the message will be "unhandled trap 256".
67 static void fixup_gdt_table(struct lg_cpu *cpu, unsigned start, unsigned end)
69 unsigned int i;
71 for (i = start; i < end; i++) {
73 * We never copy these ones to real GDT, so we don't care what
74 * they say
76 if (ignored_gdt(i))
77 continue;
80 * Segment descriptors contain a privilege level: the Guest is
81 * sometimes careless and leaves this as 0, even though it's
82 * running at privilege level 1. If so, we fix it here.
84 if ((cpu->arch.gdt[i].b & 0x00006000) == 0)
85 cpu->arch.gdt[i].b |= (GUEST_PL << 13);
88 * Each descriptor has an "accessed" bit. If we don't set it
89 * 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
91 * writable by the Guest, so bad things can happen.
93 cpu->arch.gdt[i].b |= 0x00000100;
97 /*H:610
98 * Like the IDT, we never simply use the GDT the Guest gives us. We keep
99 * a GDT for each CPU, and copy across the Guest's entries each time we want to
100 * run the Guest on that CPU.
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
104 * running.
106 void setup_default_gdt_entries(struct lguest_ro_state *state)
108 struct desc_struct *gdt = state->guest_gdt;
109 unsigned long tss = (unsigned long)&state->guest_tss;
111 /* The Switcher segments are full 0-4G segments, privilege level 0 */
112 gdt[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT;
113 gdt[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT;
116 * The TSS segment refers to the TSS entry for this particular CPU.
117 * Forgive the magic flags: the 0x8900 means the entry is Present, it's
118 * privilege level 0 Available 386 TSS system segment, and the 0x67
119 * means Saturn is eclipsed by Mercury in the twelfth house.
121 gdt[GDT_ENTRY_TSS].a = 0x00000067 | (tss << 16);
122 gdt[GDT_ENTRY_TSS].b = 0x00008900 | (tss & 0xFF000000)
123 | ((tss >> 16) & 0x000000FF);
127 * This routine sets up the initial Guest GDT for booting. All entries start
128 * as 0 (unusable).
130 void setup_guest_gdt(struct lg_cpu *cpu)
133 * Start with full 0-4G segments...except the Guest is allowed to use
134 * them, so set the privilege level appropriately in the flags.
136 cpu->arch.gdt[GDT_ENTRY_KERNEL_CS] = FULL_EXEC_SEGMENT;
137 cpu->arch.gdt[GDT_ENTRY_KERNEL_DS] = FULL_SEGMENT;
138 cpu->arch.gdt[GDT_ENTRY_KERNEL_CS].b |= (GUEST_PL << 13);
139 cpu->arch.gdt[GDT_ENTRY_KERNEL_DS].b |= (GUEST_PL << 13);
142 /*H:650
143 * An optimization of copy_gdt(), for just the three "thead-local storage"
144 * entries.
146 void copy_gdt_tls(const struct lg_cpu *cpu, struct desc_struct *gdt)
148 unsigned int i;
150 for (i = GDT_ENTRY_TLS_MIN; i <= GDT_ENTRY_TLS_MAX; i++)
151 gdt[i] = cpu->arch.gdt[i];
154 /*H:640
155 * When the Guest is run on a different CPU, or the GDT entries have changed,
156 * copy_gdt() is called to copy the Guest's GDT entries across to this CPU's
157 * GDT.
159 void copy_gdt(const struct lg_cpu *cpu, struct desc_struct *gdt)
161 unsigned int i;
164 * The default entries from setup_default_gdt_entries() are not
165 * replaced. See ignored_gdt() above.
167 for (i = 0; i < GDT_ENTRIES; i++)
168 if (!ignored_gdt(i))
169 gdt[i] = cpu->arch.gdt[i];
172 /*H:620
173 * This is where the Guest asks us to load a new GDT entry
174 * (LHCALL_LOAD_GDT_ENTRY). We tweak the entry and copy it in.
176 void load_guest_gdt_entry(struct lg_cpu *cpu, u32 num, u32 lo, u32 hi)
179 * We assume the Guest has the same number of GDT entries as the
180 * Host, otherwise we'd have to dynamically allocate the Guest GDT.
182 if (num >= ARRAY_SIZE(cpu->arch.gdt))
183 kill_guest(cpu, "too many gdt entries %i", num);
185 /* Set it up, then fix it. */
186 cpu->arch.gdt[num].a = lo;
187 cpu->arch.gdt[num].b = hi;
188 fixup_gdt_table(cpu, num, num+1);
190 * Mark that the GDT changed so the core knows it has to copy it again,
191 * even if the Guest is run on the same CPU.
193 cpu->changed |= CHANGED_GDT;
197 * This is the fast-track version for just changing the three TLS entries.
198 * Remember that this happens on every context switch, so it's worth
199 * optimizing. But wouldn't it be neater to have a single hypercall to cover
200 * both cases?
202 void guest_load_tls(struct lg_cpu *cpu, unsigned long gtls)
204 struct desc_struct *tls = &cpu->arch.gdt[GDT_ENTRY_TLS_MIN];
206 __lgread(cpu, tls, gtls, sizeof(*tls)*GDT_ENTRY_TLS_ENTRIES);
207 fixup_gdt_table(cpu, GDT_ENTRY_TLS_MIN, GDT_ENTRY_TLS_MAX+1);
208 /* Note that just the TLS entries have changed. */
209 cpu->changed |= CHANGED_GDT_TLS;
212 /*H:660
213 * With this, we have finished the Host.
215 * Five of the seven parts of our task are complete. You have made it through
216 * the Bit of Despair (I think that's somewhere in the page table code,
217 * myself).
219 * Next, we examine "make Switcher". It's short, but intense.