1 #include <linux/threads.h>
2 #include <linux/cpumask.h>
3 #include <linux/string.h>
4 #include <linux/kernel.h>
5 #include <linux/ctype.h>
6 #include <linux/dmar.h>
7 #include <linux/cpu.h>
8 
9 #include <asm/smp.h>
10 #include <asm/x2apic.h>
11 
12 static DEFINE_PER_CPU(u32, x86_cpu_to_logical_apicid);
13 static DEFINE_PER_CPU(cpumask_var_t, cpus_in_cluster);
14 static DEFINE_PER_CPU(cpumask_var_t, ipi_mask);
15 
16 static int x2apic_acpi_madt_oem_check(char *oem_id, char *oem_table_id)
17 {
18 	return x2apic_enabled();
19 }
20 
21 static inline u32 x2apic_cluster(int cpu)
22 {
23 	return per_cpu(x86_cpu_to_logical_apicid, cpu) >> 16;
24 }
25 
26 static void
27 __x2apic_send_IPI_mask(const struct cpumask *mask, int vector, int apic_dest)
28 {
29 	struct cpumask *cpus_in_cluster_ptr;
30 	struct cpumask *ipi_mask_ptr;
31 	unsigned int cpu, this_cpu;
32 	unsigned long flags;
33 	u32 dest;
34 
35 	x2apic_wrmsr_fence();
36 
37 	local_irq_save(flags);
38 
39 	this_cpu = smp_processor_id();
40 
41 	/*
42 	 * We are to modify mask, so we need an own copy
43 	 * and be sure it's manipulated with irq off.
44 	 */
45 	ipi_mask_ptr = this_cpu_cpumask_var_ptr(ipi_mask);
46 	cpumask_copy(ipi_mask_ptr, mask);
47 
48 	/*
49 	 * The idea is to send one IPI per cluster.
50 	 */
51 	for_each_cpu(cpu, ipi_mask_ptr) {
52 		unsigned long i;
53 
54 		cpus_in_cluster_ptr = per_cpu(cpus_in_cluster, cpu);
55 		dest = 0;
56 
57 		/* Collect cpus in cluster. */
58 		for_each_cpu_and(i, ipi_mask_ptr, cpus_in_cluster_ptr) {
59 			if (apic_dest == APIC_DEST_ALLINC || i != this_cpu)
60 				dest |= per_cpu(x86_cpu_to_logical_apicid, i);
61 		}
62 
63 		if (!dest)
64 			continue;
65 
66 		__x2apic_send_IPI_dest(dest, vector, apic->dest_logical);
67 		/*
68 		 * Cluster sibling cpus should be discared now so
69 		 * we would not send IPI them second time.
70 		 */
71 		cpumask_andnot(ipi_mask_ptr, ipi_mask_ptr, cpus_in_cluster_ptr);
72 	}
73 
74 	local_irq_restore(flags);
75 }
76 
77 static void x2apic_send_IPI_mask(const struct cpumask *mask, int vector)
78 {
79 	__x2apic_send_IPI_mask(mask, vector, APIC_DEST_ALLINC);
80 }
81 
82 static void
83 x2apic_send_IPI_mask_allbutself(const struct cpumask *mask, int vector)
84 {
85 	__x2apic_send_IPI_mask(mask, vector, APIC_DEST_ALLBUT);
86 }
87 
88 static void x2apic_send_IPI_allbutself(int vector)
89 {
90 	__x2apic_send_IPI_mask(cpu_online_mask, vector, APIC_DEST_ALLBUT);
91 }
92 
93 static void x2apic_send_IPI_all(int vector)
94 {
95 	__x2apic_send_IPI_mask(cpu_online_mask, vector, APIC_DEST_ALLINC);
96 }
97 
98 static int
99 x2apic_cpu_mask_to_apicid_and(const struct cpumask *cpumask,
100 			      const struct cpumask *andmask,
101 			      unsigned int *apicid)
102 {
103 	u32 dest = 0;
104 	u16 cluster;
105 	int i;
106 
107 	for_each_cpu_and(i, cpumask, andmask) {
108 		if (!cpumask_test_cpu(i, cpu_online_mask))
109 			continue;
110 		dest = per_cpu(x86_cpu_to_logical_apicid, i);
111 		cluster = x2apic_cluster(i);
112 		break;
113 	}
114 
115 	if (!dest)
116 		return -EINVAL;
117 
118 	for_each_cpu_and(i, cpumask, andmask) {
119 		if (!cpumask_test_cpu(i, cpu_online_mask))
120 			continue;
121 		if (cluster != x2apic_cluster(i))
122 			continue;
123 		dest |= per_cpu(x86_cpu_to_logical_apicid, i);
124 	}
125 
126 	*apicid = dest;
127 
128 	return 0;
129 }
130 
131 static void init_x2apic_ldr(void)
132 {
133 	unsigned int this_cpu = smp_processor_id();
134 	unsigned int cpu;
135 
136 	per_cpu(x86_cpu_to_logical_apicid, this_cpu) = apic_read(APIC_LDR);
137 
138 	__cpu_set(this_cpu, per_cpu(cpus_in_cluster, this_cpu));
139 	for_each_online_cpu(cpu) {
140 		if (x2apic_cluster(this_cpu) != x2apic_cluster(cpu))
141 			continue;
142 		__cpu_set(this_cpu, per_cpu(cpus_in_cluster, cpu));
143 		__cpu_set(cpu, per_cpu(cpus_in_cluster, this_cpu));
144 	}
145 }
146 
147  /*
148   * At CPU state changes, update the x2apic cluster sibling info.
149   */
150 static int
151 update_clusterinfo(struct notifier_block *nfb, unsigned long action, void *hcpu)
152 {
153 	unsigned int this_cpu = (unsigned long)hcpu;
154 	unsigned int cpu;
155 	int err = 0;
156 
157 	switch (action) {
158 	case CPU_UP_PREPARE:
159 		if (!zalloc_cpumask_var(&per_cpu(cpus_in_cluster, this_cpu),
160 					GFP_KERNEL)) {
161 			err = -ENOMEM;
162 		} else if (!zalloc_cpumask_var(&per_cpu(ipi_mask, this_cpu),
163 					       GFP_KERNEL)) {
164 			free_cpumask_var(per_cpu(cpus_in_cluster, this_cpu));
165 			err = -ENOMEM;
166 		}
167 		break;
168 	case CPU_UP_CANCELED:
169 	case CPU_UP_CANCELED_FROZEN:
170 	case CPU_DEAD:
171 		for_each_online_cpu(cpu) {
172 			if (x2apic_cluster(this_cpu) != x2apic_cluster(cpu))
173 				continue;
174 			__cpu_clear(this_cpu, per_cpu(cpus_in_cluster, cpu));
175 			__cpu_clear(cpu, per_cpu(cpus_in_cluster, this_cpu));
176 		}
177 		free_cpumask_var(per_cpu(cpus_in_cluster, this_cpu));
178 		free_cpumask_var(per_cpu(ipi_mask, this_cpu));
179 		break;
180 	}
181 
182 	return notifier_from_errno(err);
183 }
184 
185 static struct notifier_block __refdata x2apic_cpu_notifier = {
186 	.notifier_call = update_clusterinfo,
187 };
188 
189 static int x2apic_init_cpu_notifier(void)
190 {
191 	int cpu = smp_processor_id();
192 
193 	zalloc_cpumask_var(&per_cpu(cpus_in_cluster, cpu), GFP_KERNEL);
194 	zalloc_cpumask_var(&per_cpu(ipi_mask, cpu), GFP_KERNEL);
195 
196 	BUG_ON(!per_cpu(cpus_in_cluster, cpu) || !per_cpu(ipi_mask, cpu));
197 
198 	__cpu_set(cpu, per_cpu(cpus_in_cluster, cpu));
199 	register_hotcpu_notifier(&x2apic_cpu_notifier);
200 	return 1;
201 }
202 
203 static int x2apic_cluster_probe(void)
204 {
205 	if (x2apic_mode)
206 		return x2apic_init_cpu_notifier();
207 	else
208 		return 0;
209 }
210 
211 static const struct cpumask *x2apic_cluster_target_cpus(void)
212 {
213 	return cpu_all_mask;
214 }
215 
216 /*
217  * Each x2apic cluster is an allocation domain.
218  */
219 static void cluster_vector_allocation_domain(int cpu, struct cpumask *retmask,
220 					     const struct cpumask *mask)
221 {
222 	/*
223 	 * To minimize vector pressure, default case of boot, device bringup
224 	 * etc will use a single cpu for the interrupt destination.
225 	 *
226 	 * On explicit migration requests coming from irqbalance etc,
227 	 * interrupts will be routed to the x2apic cluster (cluster-id
228 	 * derived from the first cpu in the mask) members specified
229 	 * in the mask.
230 	 */
231 	if (mask == x2apic_cluster_target_cpus())
232 		cpumask_copy(retmask, cpumask_of(cpu));
233 	else
234 		cpumask_and(retmask, mask, per_cpu(cpus_in_cluster, cpu));
235 }
236 
237 static struct apic apic_x2apic_cluster = {
238 
239 	.name				= "cluster x2apic",
240 	.probe				= x2apic_cluster_probe,
241 	.acpi_madt_oem_check		= x2apic_acpi_madt_oem_check,
242 	.apic_id_valid			= x2apic_apic_id_valid,
243 	.apic_id_registered		= x2apic_apic_id_registered,
244 
245 	.irq_delivery_mode		= dest_LowestPrio,
246 	.irq_dest_mode			= 1, /* logical */
247 
248 	.target_cpus			= x2apic_cluster_target_cpus,
249 	.disable_esr			= 0,
250 	.dest_logical			= APIC_DEST_LOGICAL,
251 	.check_apicid_used		= NULL,
252 
253 	.vector_allocation_domain	= cluster_vector_allocation_domain,
254 	.init_apic_ldr			= init_x2apic_ldr,
255 
256 	.ioapic_phys_id_map		= NULL,
257 	.setup_apic_routing		= NULL,
258 	.cpu_present_to_apicid		= default_cpu_present_to_apicid,
259 	.apicid_to_cpu_present		= NULL,
260 	.check_phys_apicid_present	= default_check_phys_apicid_present,
261 	.phys_pkg_id			= x2apic_phys_pkg_id,
262 
263 	.get_apic_id			= x2apic_get_apic_id,
264 	.set_apic_id			= x2apic_set_apic_id,
265 	.apic_id_mask			= 0xFFFFFFFFu,
266 
267 	.cpu_mask_to_apicid_and		= x2apic_cpu_mask_to_apicid_and,
268 
269 	.send_IPI_mask			= x2apic_send_IPI_mask,
270 	.send_IPI_mask_allbutself	= x2apic_send_IPI_mask_allbutself,
271 	.send_IPI_allbutself		= x2apic_send_IPI_allbutself,
272 	.send_IPI_all			= x2apic_send_IPI_all,
273 	.send_IPI_self			= x2apic_send_IPI_self,
274 
275 	.wait_for_init_deassert		= false,
276 	.inquire_remote_apic		= NULL,
277 
278 	.read				= native_apic_msr_read,
279 	.write				= native_apic_msr_write,
280 	.eoi_write			= native_apic_msr_eoi_write,
281 	.icr_read			= native_x2apic_icr_read,
282 	.icr_write			= native_x2apic_icr_write,
283 	.wait_icr_idle			= native_x2apic_wait_icr_idle,
284 	.safe_wait_icr_idle		= native_safe_x2apic_wait_icr_idle,
285 };
286 
287 apic_driver(apic_x2apic_cluster);
288