xref: /openbmc/linux/arch/x86/hyperv/mmu.c (revision fb574682)
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 void hyperv_flush_tlb_others(const struct cpumask *cpus,
56 				    const struct flush_tlb_info *info)
57 {
58 	int cpu, vcpu, gva_n, max_gvas;
59 	struct hv_tlb_flush **flush_pcpu;
60 	struct hv_tlb_flush *flush;
61 	u64 status = U64_MAX;
62 	unsigned long flags;
63 
64 	trace_hyperv_mmu_flush_tlb_others(cpus, info);
65 
66 	if (!hv_hypercall_pg)
67 		goto do_native;
68 
69 	if (cpumask_empty(cpus))
70 		return;
71 
72 	local_irq_save(flags);
73 
74 	flush_pcpu = (struct hv_tlb_flush **)
75 		     this_cpu_ptr(hyperv_pcpu_input_arg);
76 
77 	flush = *flush_pcpu;
78 
79 	if (unlikely(!flush)) {
80 		local_irq_restore(flags);
81 		goto do_native;
82 	}
83 
84 	if (info->mm) {
85 		/*
86 		 * AddressSpace argument must match the CR3 with PCID bits
87 		 * stripped out.
88 		 */
89 		flush->address_space = virt_to_phys(info->mm->pgd);
90 		flush->address_space &= CR3_ADDR_MASK;
91 		flush->flags = 0;
92 	} else {
93 		flush->address_space = 0;
94 		flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
95 	}
96 
97 	flush->processor_mask = 0;
98 	if (cpumask_equal(cpus, cpu_present_mask)) {
99 		flush->flags |= HV_FLUSH_ALL_PROCESSORS;
100 	} else {
101 		/*
102 		 * From the supplied CPU set we need to figure out if we can get
103 		 * away with cheaper HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE}
104 		 * hypercalls. This is possible when the highest VP number in
105 		 * the set is < 64. As VP numbers are usually in ascending order
106 		 * and match Linux CPU ids, here is an optimization: we check
107 		 * the VP number for the highest bit in the supplied set first
108 		 * so we can quickly find out if using *_EX hypercalls is a
109 		 * must. We will also check all VP numbers when walking the
110 		 * supplied CPU set to remain correct in all cases.
111 		 */
112 		if (hv_cpu_number_to_vp_number(cpumask_last(cpus)) >= 64)
113 			goto do_ex_hypercall;
114 
115 		for_each_cpu(cpu, cpus) {
116 			vcpu = hv_cpu_number_to_vp_number(cpu);
117 			if (vcpu == VP_INVAL) {
118 				local_irq_restore(flags);
119 				goto do_native;
120 			}
121 
122 			if (vcpu >= 64)
123 				goto do_ex_hypercall;
124 
125 			__set_bit(vcpu, (unsigned long *)
126 				  &flush->processor_mask);
127 		}
128 	}
129 
130 	/*
131 	 * We can flush not more than max_gvas with one hypercall. Flush the
132 	 * whole address space if we were asked to do more.
133 	 */
134 	max_gvas = (PAGE_SIZE - sizeof(*flush)) / sizeof(flush->gva_list[0]);
135 
136 	if (info->end == TLB_FLUSH_ALL) {
137 		flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
138 		status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
139 					 flush, NULL);
140 	} else if (info->end &&
141 		   ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
142 		status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
143 					 flush, NULL);
144 	} else {
145 		gva_n = fill_gva_list(flush->gva_list, 0,
146 				      info->start, info->end);
147 		status = hv_do_rep_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST,
148 					     gva_n, 0, flush, NULL);
149 	}
150 	goto check_status;
151 
152 do_ex_hypercall:
153 	status = hyperv_flush_tlb_others_ex(cpus, info);
154 
155 check_status:
156 	local_irq_restore(flags);
157 
158 	if (!(status & HV_HYPERCALL_RESULT_MASK))
159 		return;
160 do_native:
161 	native_flush_tlb_others(cpus, info);
162 }
163 
164 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
165 				      const struct flush_tlb_info *info)
166 {
167 	int nr_bank = 0, max_gvas, gva_n;
168 	struct hv_tlb_flush_ex **flush_pcpu;
169 	struct hv_tlb_flush_ex *flush;
170 	u64 status;
171 
172 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
173 		return U64_MAX;
174 
175 	flush_pcpu = (struct hv_tlb_flush_ex **)
176 		     this_cpu_ptr(hyperv_pcpu_input_arg);
177 
178 	flush = *flush_pcpu;
179 
180 	if (info->mm) {
181 		/*
182 		 * AddressSpace argument must match the CR3 with PCID bits
183 		 * stripped out.
184 		 */
185 		flush->address_space = virt_to_phys(info->mm->pgd);
186 		flush->address_space &= CR3_ADDR_MASK;
187 		flush->flags = 0;
188 	} else {
189 		flush->address_space = 0;
190 		flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
191 	}
192 
193 	flush->hv_vp_set.valid_bank_mask = 0;
194 
195 	flush->hv_vp_set.format = HV_GENERIC_SET_SPARSE_4K;
196 	nr_bank = cpumask_to_vpset(&(flush->hv_vp_set), cpus);
197 	if (nr_bank < 0)
198 		return U64_MAX;
199 
200 	/*
201 	 * We can flush not more than max_gvas with one hypercall. Flush the
202 	 * whole address space if we were asked to do more.
203 	 */
204 	max_gvas =
205 		(PAGE_SIZE - sizeof(*flush) - nr_bank *
206 		 sizeof(flush->hv_vp_set.bank_contents[0])) /
207 		sizeof(flush->gva_list[0]);
208 
209 	if (info->end == TLB_FLUSH_ALL) {
210 		flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
211 		status = hv_do_rep_hypercall(
212 			HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
213 			0, nr_bank, flush, NULL);
214 	} else if (info->end &&
215 		   ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
216 		status = hv_do_rep_hypercall(
217 			HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
218 			0, nr_bank, flush, NULL);
219 	} else {
220 		gva_n = fill_gva_list(flush->gva_list, nr_bank,
221 				      info->start, info->end);
222 		status = hv_do_rep_hypercall(
223 			HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX,
224 			gva_n, nr_bank, flush, NULL);
225 	}
226 
227 	return status;
228 }
229 
230 void hyperv_setup_mmu_ops(void)
231 {
232 	if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
233 		return;
234 
235 	pr_info("Using hypercall for remote TLB flush\n");
236 	pv_ops.mmu.flush_tlb_others = hyperv_flush_tlb_others;
237 	pv_ops.mmu.tlb_remove_table = tlb_remove_table;
238 }
239