xref: /linux/arch/x86/hyperv/mmu.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
1 #define pr_fmt(fmt)  "Hyper-V: " fmt
2 
3 #include <linux/hyperv.h>
4 #include <linux/log2.h>
5 #include <linux/slab.h>
6 #include <linux/types.h>
7 
8 #include <asm/fpu/api.h>
9 #include <asm/mshyperv.h>
10 #include <asm/msr.h>
11 #include <asm/tlbflush.h>
12 
13 #define CREATE_TRACE_POINTS
14 #include <asm/trace/hyperv.h>
15 
16 /* HvFlushVirtualAddressSpace, HvFlushVirtualAddressList hypercalls */
17 struct hv_flush_pcpu {
18 	u64 address_space;
19 	u64 flags;
20 	u64 processor_mask;
21 	u64 gva_list[];
22 };
23 
24 /* HvFlushVirtualAddressSpaceEx, HvFlushVirtualAddressListEx hypercalls */
25 struct hv_flush_pcpu_ex {
26 	u64 address_space;
27 	u64 flags;
28 	struct {
29 		u64 format;
30 		u64 valid_bank_mask;
31 		u64 bank_contents[];
32 	} hv_vp_set;
33 	u64 gva_list[];
34 };
35 
36 /* Each gva in gva_list encodes up to 4096 pages to flush */
37 #define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)
38 
39 static struct hv_flush_pcpu __percpu **pcpu_flush;
40 
41 static struct hv_flush_pcpu_ex __percpu **pcpu_flush_ex;
42 
43 /*
44  * Fills in gva_list starting from offset. Returns the number of items added.
45  */
46 static inline int fill_gva_list(u64 gva_list[], int offset,
47 				unsigned long start, unsigned long end)
48 {
49 	int gva_n = offset;
50 	unsigned long cur = start, diff;
51 
52 	do {
53 		diff = end > cur ? end - cur : 0;
54 
55 		gva_list[gva_n] = cur & PAGE_MASK;
56 		/*
57 		 * Lower 12 bits encode the number of additional
58 		 * pages to flush (in addition to the 'cur' page).
59 		 */
60 		if (diff >= HV_TLB_FLUSH_UNIT)
61 			gva_list[gva_n] |= ~PAGE_MASK;
62 		else if (diff)
63 			gva_list[gva_n] |= (diff - 1) >> PAGE_SHIFT;
64 
65 		cur += HV_TLB_FLUSH_UNIT;
66 		gva_n++;
67 
68 	} while (cur < end);
69 
70 	return gva_n - offset;
71 }
72 
73 /* Return the number of banks in the resulting vp_set */
74 static inline int cpumask_to_vp_set(struct hv_flush_pcpu_ex *flush,
75 				    const struct cpumask *cpus)
76 {
77 	int cpu, vcpu, vcpu_bank, vcpu_offset, nr_bank = 1;
78 
79 	/* valid_bank_mask can represent up to 64 banks */
80 	if (hv_max_vp_index / 64 >= 64)
81 		return 0;
82 
83 	/*
84 	 * Clear all banks up to the maximum possible bank as hv_flush_pcpu_ex
85 	 * structs are not cleared between calls, we risk flushing unneeded
86 	 * vCPUs otherwise.
87 	 */
88 	for (vcpu_bank = 0; vcpu_bank <= hv_max_vp_index / 64; vcpu_bank++)
89 		flush->hv_vp_set.bank_contents[vcpu_bank] = 0;
90 
91 	/*
92 	 * Some banks may end up being empty but this is acceptable.
93 	 */
94 	for_each_cpu(cpu, cpus) {
95 		vcpu = hv_cpu_number_to_vp_number(cpu);
96 		vcpu_bank = vcpu / 64;
97 		vcpu_offset = vcpu % 64;
98 		__set_bit(vcpu_offset, (unsigned long *)
99 			  &flush->hv_vp_set.bank_contents[vcpu_bank]);
100 		if (vcpu_bank >= nr_bank)
101 			nr_bank = vcpu_bank + 1;
102 	}
103 	flush->hv_vp_set.valid_bank_mask = GENMASK_ULL(nr_bank - 1, 0);
104 
105 	return nr_bank;
106 }
107 
108 static void hyperv_flush_tlb_others(const struct cpumask *cpus,
109 				    const struct flush_tlb_info *info)
110 {
111 	int cpu, vcpu, gva_n, max_gvas;
112 	struct hv_flush_pcpu **flush_pcpu;
113 	struct hv_flush_pcpu *flush;
114 	u64 status = U64_MAX;
115 	unsigned long flags;
116 
117 	trace_hyperv_mmu_flush_tlb_others(cpus, info);
118 
119 	if (!pcpu_flush || !hv_hypercall_pg)
120 		goto do_native;
121 
122 	if (cpumask_empty(cpus))
123 		return;
124 
125 	local_irq_save(flags);
126 
127 	flush_pcpu = this_cpu_ptr(pcpu_flush);
128 
129 	if (unlikely(!*flush_pcpu))
130 		*flush_pcpu = page_address(alloc_page(GFP_ATOMIC));
131 
132 	flush = *flush_pcpu;
133 
134 	if (unlikely(!flush)) {
135 		local_irq_restore(flags);
136 		goto do_native;
137 	}
138 
139 	if (info->mm) {
140 		flush->address_space = virt_to_phys(info->mm->pgd);
141 		flush->flags = 0;
142 	} else {
143 		flush->address_space = 0;
144 		flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
145 	}
146 
147 	flush->processor_mask = 0;
148 	if (cpumask_equal(cpus, cpu_present_mask)) {
149 		flush->flags |= HV_FLUSH_ALL_PROCESSORS;
150 	} else {
151 		for_each_cpu(cpu, cpus) {
152 			vcpu = hv_cpu_number_to_vp_number(cpu);
153 			if (vcpu >= 64)
154 				goto do_native;
155 
156 			__set_bit(vcpu, (unsigned long *)
157 				  &flush->processor_mask);
158 		}
159 	}
160 
161 	/*
162 	 * We can flush not more than max_gvas with one hypercall. Flush the
163 	 * whole address space if we were asked to do more.
164 	 */
165 	max_gvas = (PAGE_SIZE - sizeof(*flush)) / sizeof(flush->gva_list[0]);
166 
167 	if (info->end == TLB_FLUSH_ALL) {
168 		flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
169 		status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
170 					 flush, NULL);
171 	} else if (info->end &&
172 		   ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
173 		status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
174 					 flush, NULL);
175 	} else {
176 		gva_n = fill_gva_list(flush->gva_list, 0,
177 				      info->start, info->end);
178 		status = hv_do_rep_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST,
179 					     gva_n, 0, flush, NULL);
180 	}
181 
182 	local_irq_restore(flags);
183 
184 	if (!(status & HV_HYPERCALL_RESULT_MASK))
185 		return;
186 do_native:
187 	native_flush_tlb_others(cpus, info);
188 }
189 
190 static void hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
191 				       const struct flush_tlb_info *info)
192 {
193 	int nr_bank = 0, max_gvas, gva_n;
194 	struct hv_flush_pcpu_ex **flush_pcpu;
195 	struct hv_flush_pcpu_ex *flush;
196 	u64 status = U64_MAX;
197 	unsigned long flags;
198 
199 	trace_hyperv_mmu_flush_tlb_others(cpus, info);
200 
201 	if (!pcpu_flush_ex || !hv_hypercall_pg)
202 		goto do_native;
203 
204 	if (cpumask_empty(cpus))
205 		return;
206 
207 	local_irq_save(flags);
208 
209 	flush_pcpu = this_cpu_ptr(pcpu_flush_ex);
210 
211 	if (unlikely(!*flush_pcpu))
212 		*flush_pcpu = page_address(alloc_page(GFP_ATOMIC));
213 
214 	flush = *flush_pcpu;
215 
216 	if (unlikely(!flush)) {
217 		local_irq_restore(flags);
218 		goto do_native;
219 	}
220 
221 	if (info->mm) {
222 		flush->address_space = virt_to_phys(info->mm->pgd);
223 		flush->flags = 0;
224 	} else {
225 		flush->address_space = 0;
226 		flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
227 	}
228 
229 	flush->hv_vp_set.valid_bank_mask = 0;
230 
231 	if (!cpumask_equal(cpus, cpu_present_mask)) {
232 		flush->hv_vp_set.format = HV_GENERIC_SET_SPARCE_4K;
233 		nr_bank = cpumask_to_vp_set(flush, cpus);
234 	}
235 
236 	if (!nr_bank) {
237 		flush->hv_vp_set.format = HV_GENERIC_SET_ALL;
238 		flush->flags |= HV_FLUSH_ALL_PROCESSORS;
239 	}
240 
241 	/*
242 	 * We can flush not more than max_gvas with one hypercall. Flush the
243 	 * whole address space if we were asked to do more.
244 	 */
245 	max_gvas =
246 		(PAGE_SIZE - sizeof(*flush) - nr_bank *
247 		 sizeof(flush->hv_vp_set.bank_contents[0])) /
248 		sizeof(flush->gva_list[0]);
249 
250 	if (info->end == TLB_FLUSH_ALL) {
251 		flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
252 		status = hv_do_rep_hypercall(
253 			HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
254 			0, nr_bank, flush, NULL);
255 	} else if (info->end &&
256 		   ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
257 		status = hv_do_rep_hypercall(
258 			HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
259 			0, nr_bank, flush, NULL);
260 	} else {
261 		gva_n = fill_gva_list(flush->gva_list, nr_bank,
262 				      info->start, info->end);
263 		status = hv_do_rep_hypercall(
264 			HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX,
265 			gva_n, nr_bank, flush, NULL);
266 	}
267 
268 	local_irq_restore(flags);
269 
270 	if (!(status & HV_HYPERCALL_RESULT_MASK))
271 		return;
272 do_native:
273 	native_flush_tlb_others(cpus, info);
274 }
275 
276 void hyperv_setup_mmu_ops(void)
277 {
278 	if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
279 		return;
280 
281 	setup_clear_cpu_cap(X86_FEATURE_PCID);
282 
283 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED)) {
284 		pr_info("Using hypercall for remote TLB flush\n");
285 		pv_mmu_ops.flush_tlb_others = hyperv_flush_tlb_others;
286 	} else {
287 		pr_info("Using ext hypercall for remote TLB flush\n");
288 		pv_mmu_ops.flush_tlb_others = hyperv_flush_tlb_others_ex;
289 	}
290 }
291 
292 void hyper_alloc_mmu(void)
293 {
294 	if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
295 		return;
296 
297 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
298 		pcpu_flush = alloc_percpu(struct hv_flush_pcpu *);
299 	else
300 		pcpu_flush_ex = alloc_percpu(struct hv_flush_pcpu_ex *);
301 }
302