1 /*
2 * Async PF test. For the test to actually do anything it needs to be started
3 * in memory cgroup with 512M of memory and with more than 1G memory provided
4 * to the guest.
5 *
6 * To identify the cgroup version on Linux:
7 * stat -fc %T /sys/fs/cgroup/
8 *
9 * If the output is tmpfs, your system is using cgroup v1:
10 * To create cgroup do as root:
11 * mkdir /dev/cgroup
12 * mount -t cgroup none -omemory /dev/cgroup
13 * chmod a+rxw /dev/cgroup/
14 *
15 * From a shell you will start qemu from:
16 * mkdir /dev/cgroup/1
17 * echo $$ > /dev/cgroup/1/tasks
18 * echo 512M > /dev/cgroup/1/memory.limit_in_bytes
19 *
20 * If the output is cgroup2fs, your system is using cgroup v2:
21 * mkdir /sys/fs/cgroup/cg1
22 * echo $$ > /sys/fs/cgroup/cg1/cgroup.procs
23 * echo 512M > /sys/fs/cgroup/cg1/memory.max
24 *
25 */
26 #include "x86/processor.h"
27 #include "x86/apic.h"
28 #include "x86/isr.h"
29 #include "x86/vm.h"
30 #include "alloc.h"
31 #include "vmalloc.h"
32
33 #define KVM_PV_REASON_PAGE_NOT_PRESENT 1
34
35 #define MSR_KVM_ASYNC_PF_EN 0x4b564d02
36 #define MSR_KVM_ASYNC_PF_INT 0x4b564d06
37 #define MSR_KVM_ASYNC_PF_ACK 0x4b564d07
38
39 #define KVM_ASYNC_PF_ENABLED (1 << 0)
40 #define KVM_ASYNC_PF_SEND_ALWAYS (1 << 1)
41 #define KVM_ASYNC_PF_DELIVERY_AS_INT (1 << 3)
42
43 #define HYPERVISOR_CALLBACK_VECTOR 0xf3
44
45 struct kvm_vcpu_pv_apf_data {
46 /* Used for 'page not present' events delivered via #PF */
47 uint32_t flags;
48
49 /* Used for 'page ready' events delivered via interrupt notification */
50 uint32_t token;
51
52 uint8_t pad[56];
53 } apf_reason __attribute__((aligned(64)));
54
55 char *buf;
56 void* virt;
57 volatile uint64_t i;
58 volatile uint64_t phys;
59 volatile uint32_t saved_token;
60 volatile unsigned asyncpf_num;
61
get_and_clear_apf_reason(void)62 static inline uint32_t get_and_clear_apf_reason(void)
63 {
64 uint32_t r = apf_reason.flags;
65 apf_reason.flags = 0;
66 return r;
67 }
68
handle_interrupt(isr_regs_t * regs)69 static void handle_interrupt(isr_regs_t *regs)
70 {
71 uint32_t apf_token = apf_reason.token;
72
73 apf_reason.token = 0;
74 wrmsr(MSR_KVM_ASYNC_PF_ACK, 1);
75
76 if (apf_token == 0xffffffff) {
77 report_pass("Wakeup all, got token 0x%" PRIx32, apf_token);
78 } else if (apf_token == saved_token) {
79 asyncpf_num++;
80 install_pte(phys_to_virt(read_cr3()), 1, virt, phys | PT_PRESENT_MASK | PT_WRITABLE_MASK, 0);
81 phys = 0;
82 } else {
83 report_fail("unexpected async pf int token 0x%" PRIx32, apf_token);
84 }
85
86 eoi();
87 }
88
handle_pf(struct ex_regs * r)89 static void handle_pf(struct ex_regs *r)
90 {
91 virt = (void*)((ulong)(buf+i) & ~(PAGE_SIZE-1));
92 uint32_t reason = get_and_clear_apf_reason();
93 switch (reason) {
94 case 0:
95 report_fail("unexpected #PF at %#lx", read_cr2());
96 exit(report_summary());
97 case KVM_PV_REASON_PAGE_NOT_PRESENT:
98 phys = virt_to_pte_phys(phys_to_virt(read_cr3()), virt);
99 install_pte(phys_to_virt(read_cr3()), 1, virt, phys, 0);
100 write_cr3(read_cr3());
101 saved_token = read_cr2();
102 while (phys) {
103 safe_halt(); /* enables irq */
104 }
105 break;
106 default:
107 report_fail("unexpected async pf with reason 0x%" PRIx32, reason);
108 exit(report_summary());
109 }
110 }
111
112 #define MEM (1ull*1024*1024*1024)
113
main(int ac,char ** av)114 int main(int ac, char **av)
115 {
116 if (!this_cpu_has(KVM_FEATURE_ASYNC_PF)) {
117 report_skip("KVM_FEATURE_ASYNC_PF is not supported\n");
118 return report_summary();
119 }
120
121 if (!this_cpu_has(KVM_FEATURE_ASYNC_PF_INT)) {
122 report_skip("KVM_FEATURE_ASYNC_PF_INT is not supported\n");
123 return report_summary();
124 }
125
126 setup_vm();
127
128 handle_exception(PF_VECTOR, handle_pf);
129 handle_irq(HYPERVISOR_CALLBACK_VECTOR, handle_interrupt);
130 memset(&apf_reason, 0, sizeof(apf_reason));
131
132 wrmsr(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
133 wrmsr(MSR_KVM_ASYNC_PF_EN, virt_to_phys((void*)&apf_reason) |
134 KVM_ASYNC_PF_SEND_ALWAYS | KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT);
135
136 buf = malloc(MEM);
137 sti();
138
139 /* access a lot of memory to make host swap it out */
140 for (i = 0; i < MEM; i += 4096)
141 buf[i] = 1;
142
143 cli();
144 if (!asyncpf_num)
145 report_skip("No async page fault events, cgroup configuration likely needed");
146 else
147 report_pass("Serviced %d async page faults events (!PRESENT #PF + READY IRQ)",
148 asyncpf_num);
149 return report_summary();
150 }
151