xref: /linux/arch/x86/hyperv/mmu.c (revision 2241f81c91f211b512bd2c3a26a4a74258d0e008)
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 #include <asm/tlb.h>
13 
14 #define CREATE_TRACE_POINTS
15 #include <asm/trace/hyperv.h>
16 
17 /* Each gva in gva_list encodes up to 4096 pages to flush */
18 #define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)
19 
20 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
21 				      const struct flush_tlb_info *info);
22 
23 /*
24  * Fills in gva_list starting from offset. Returns the number of items added.
25  */
26 static inline int fill_gva_list(u64 gva_list[], int offset,
27 				unsigned long start, unsigned long end)
28 {
29 	int gva_n = offset;
30 	unsigned long cur = start, diff;
31 
32 	do {
33 		diff = end > cur ? end - cur : 0;
34 
35 		gva_list[gva_n] = cur & PAGE_MASK;
36 		/*
37 		 * Lower 12 bits encode the number of additional
38 		 * pages to flush (in addition to the 'cur' page).
39 		 */
40 		if (diff >= HV_TLB_FLUSH_UNIT) {
41 			gva_list[gva_n] |= ~PAGE_MASK;
42 			cur += HV_TLB_FLUSH_UNIT;
43 		}  else if (diff) {
44 			gva_list[gva_n] |= (diff - 1) >> PAGE_SHIFT;
45 			cur = end;
46 		}
47 
48 		gva_n++;
49 
50 	} while (cur < end);
51 
52 	return gva_n - offset;
53 }
54 
55 static bool cpu_is_lazy(int cpu)
56 {
57 	return per_cpu(cpu_tlbstate_shared.is_lazy, cpu);
58 }
59 
60 static void hyperv_flush_tlb_multi(const struct cpumask *cpus,
61 				   const struct flush_tlb_info *info)
62 {
63 	int cpu, vcpu, gva_n, max_gvas;
64 	struct hv_tlb_flush *flush;
65 	u64 status;
66 	unsigned long flags;
67 	bool do_lazy = !info->freed_tables;
68 
69 	trace_hyperv_mmu_flush_tlb_multi(cpus, info);
70 
71 	if (!hv_hypercall_pg)
72 		goto do_native;
73 
74 	local_irq_save(flags);
75 
76 	flush = *this_cpu_ptr(hyperv_pcpu_input_arg);
77 
78 	if (unlikely(!flush)) {
79 		local_irq_restore(flags);
80 		goto do_native;
81 	}
82 
83 	if (info->mm) {
84 		/*
85 		 * AddressSpace argument must match the CR3 with PCID bits
86 		 * stripped out.
87 		 */
88 		flush->address_space = virt_to_phys(info->mm->pgd);
89 		flush->address_space &= CR3_ADDR_MASK;
90 		flush->flags = 0;
91 	} else {
92 		flush->address_space = 0;
93 		flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
94 	}
95 
96 	flush->processor_mask = 0;
97 	if (cpumask_equal(cpus, cpu_present_mask)) {
98 		flush->flags |= HV_FLUSH_ALL_PROCESSORS;
99 	} else {
100 		/*
101 		 * From the supplied CPU set we need to figure out if we can get
102 		 * away with cheaper HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE}
103 		 * hypercalls. This is possible when the highest VP number in
104 		 * the set is < 64. As VP numbers are usually in ascending order
105 		 * and match Linux CPU ids, here is an optimization: we check
106 		 * the VP number for the highest bit in the supplied set first
107 		 * so we can quickly find out if using *_EX hypercalls is a
108 		 * must. We will also check all VP numbers when walking the
109 		 * supplied CPU set to remain correct in all cases.
110 		 */
111 		cpu = cpumask_last(cpus);
112 
113 		if (cpu < nr_cpumask_bits && hv_cpu_number_to_vp_number(cpu) >= 64)
114 			goto do_ex_hypercall;
115 
116 		for_each_cpu(cpu, cpus) {
117 			if (do_lazy && cpu_is_lazy(cpu))
118 				continue;
119 			vcpu = hv_cpu_number_to_vp_number(cpu);
120 			if (vcpu == VP_INVAL) {
121 				local_irq_restore(flags);
122 				goto do_native;
123 			}
124 
125 			if (vcpu >= 64)
126 				goto do_ex_hypercall;
127 
128 			__set_bit(vcpu, (unsigned long *)
129 				  &flush->processor_mask);
130 		}
131 
132 		/* nothing to flush if 'processor_mask' ends up being empty */
133 		if (!flush->processor_mask) {
134 			local_irq_restore(flags);
135 			return;
136 		}
137 	}
138 
139 	/*
140 	 * We can flush not more than max_gvas with one hypercall. Flush the
141 	 * whole address space if we were asked to do more.
142 	 */
143 	max_gvas = (PAGE_SIZE - sizeof(*flush)) / sizeof(flush->gva_list[0]);
144 
145 	if (info->end == TLB_FLUSH_ALL) {
146 		flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
147 		status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
148 					 flush, NULL);
149 	} else if (info->end &&
150 		   ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
151 		status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
152 					 flush, NULL);
153 	} else {
154 		gva_n = fill_gva_list(flush->gva_list, 0,
155 				      info->start, info->end);
156 		status = hv_do_rep_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST,
157 					     gva_n, 0, flush, NULL);
158 	}
159 	goto check_status;
160 
161 do_ex_hypercall:
162 	status = hyperv_flush_tlb_others_ex(cpus, info);
163 
164 check_status:
165 	local_irq_restore(flags);
166 
167 	if (hv_result_success(status))
168 		return;
169 do_native:
170 	native_flush_tlb_multi(cpus, info);
171 }
172 
173 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
174 				      const struct flush_tlb_info *info)
175 {
176 	int nr_bank = 0, max_gvas, gva_n;
177 	struct hv_tlb_flush_ex *flush;
178 	u64 status;
179 
180 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
181 		return HV_STATUS_INVALID_PARAMETER;
182 
183 	flush = *this_cpu_ptr(hyperv_pcpu_input_arg);
184 
185 	if (info->mm) {
186 		/*
187 		 * AddressSpace argument must match the CR3 with PCID bits
188 		 * stripped out.
189 		 */
190 		flush->address_space = virt_to_phys(info->mm->pgd);
191 		flush->address_space &= CR3_ADDR_MASK;
192 		flush->flags = 0;
193 	} else {
194 		flush->address_space = 0;
195 		flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
196 	}
197 
198 	flush->hv_vp_set.valid_bank_mask = 0;
199 
200 	flush->hv_vp_set.format = HV_GENERIC_SET_SPARSE_4K;
201 	nr_bank = cpumask_to_vpset_skip(&flush->hv_vp_set, cpus,
202 			info->freed_tables ? NULL : cpu_is_lazy);
203 	if (nr_bank < 0)
204 		return HV_STATUS_INVALID_PARAMETER;
205 
206 	/*
207 	 * We can flush not more than max_gvas with one hypercall. Flush the
208 	 * whole address space if we were asked to do more.
209 	 */
210 	max_gvas =
211 		(PAGE_SIZE - sizeof(*flush) - nr_bank *
212 		 sizeof(flush->hv_vp_set.bank_contents[0])) /
213 		sizeof(flush->gva_list[0]);
214 
215 	if (info->end == TLB_FLUSH_ALL) {
216 		flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
217 		status = hv_do_rep_hypercall(
218 			HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
219 			0, nr_bank, flush, NULL);
220 	} else if (info->end &&
221 		   ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
222 		status = hv_do_rep_hypercall(
223 			HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
224 			0, nr_bank, flush, NULL);
225 	} else {
226 		gva_n = fill_gva_list(flush->gva_list, nr_bank,
227 				      info->start, info->end);
228 		status = hv_do_rep_hypercall(
229 			HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX,
230 			gva_n, nr_bank, flush, NULL);
231 	}
232 
233 	return status;
234 }
235 
236 void hyperv_setup_mmu_ops(void)
237 {
238 	if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
239 		return;
240 
241 	pr_info("Using hypercall for remote TLB flush\n");
242 	pv_ops.mmu.flush_tlb_multi = hyperv_flush_tlb_multi;
243 	pv_ops.mmu.tlb_remove_table = tlb_remove_table;
244 }
245