1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Multiplex several virtual IPIs over a single HW IPI.
4 *
5 * Copyright The Asahi Linux Contributors
6 * Copyright (c) 2022 Ventana Micro Systems Inc.
7 */
8
9 #define pr_fmt(fmt) "ipi-mux: " fmt
10 #include <linux/cpu.h>
11 #include <linux/init.h>
12 #include <linux/irq.h>
13 #include <linux/irqchip.h>
14 #include <linux/irqchip/chained_irq.h>
15 #include <linux/irqdomain.h>
16 #include <linux/jump_label.h>
17 #include <linux/percpu.h>
18 #include <linux/smp.h>
19
20 struct ipi_mux_cpu {
21 atomic_t enable;
22 atomic_t bits;
23 };
24
25 static struct ipi_mux_cpu __percpu *ipi_mux_pcpu;
26 static struct irq_domain *ipi_mux_domain;
27 static void (*ipi_mux_send)(unsigned int cpu);
28
ipi_mux_mask(struct irq_data * d)29 static void ipi_mux_mask(struct irq_data *d)
30 {
31 struct ipi_mux_cpu *icpu = this_cpu_ptr(ipi_mux_pcpu);
32
33 atomic_andnot(BIT(irqd_to_hwirq(d)), &icpu->enable);
34 }
35
ipi_mux_unmask(struct irq_data * d)36 static void ipi_mux_unmask(struct irq_data *d)
37 {
38 struct ipi_mux_cpu *icpu = this_cpu_ptr(ipi_mux_pcpu);
39 u32 ibit = BIT(irqd_to_hwirq(d));
40
41 atomic_or(ibit, &icpu->enable);
42
43 /*
44 * The atomic_or() above must complete before the atomic_read()
45 * below to avoid racing ipi_mux_send_mask().
46 */
47 smp_mb__after_atomic();
48
49 /* If a pending IPI was unmasked, raise a parent IPI immediately. */
50 if (atomic_read(&icpu->bits) & ibit)
51 ipi_mux_send(smp_processor_id());
52 }
53
ipi_mux_send_mask(struct irq_data * d,const struct cpumask * mask)54 static void ipi_mux_send_mask(struct irq_data *d, const struct cpumask *mask)
55 {
56 struct ipi_mux_cpu *icpu = this_cpu_ptr(ipi_mux_pcpu);
57 u32 ibit = BIT(irqd_to_hwirq(d));
58 unsigned long pending;
59 int cpu;
60
61 for_each_cpu(cpu, mask) {
62 icpu = per_cpu_ptr(ipi_mux_pcpu, cpu);
63
64 /*
65 * This sequence is the mirror of the one in ipi_mux_unmask();
66 * see the comment there. Additionally, release semantics
67 * ensure that the vIPI flag set is ordered after any shared
68 * memory accesses that precede it. This therefore also pairs
69 * with the atomic_fetch_andnot in ipi_mux_process().
70 */
71 pending = atomic_fetch_or_release(ibit, &icpu->bits);
72
73 /*
74 * The atomic_fetch_or_release() above must complete
75 * before the atomic_read() below to avoid racing with
76 * ipi_mux_unmask().
77 */
78 smp_mb__after_atomic();
79
80 /*
81 * The flag writes must complete before the physical IPI is
82 * issued to another CPU. This is implied by the control
83 * dependency on the result of atomic_read() below, which is
84 * itself already ordered after the vIPI flag write.
85 */
86 if (!(pending & ibit) && (atomic_read(&icpu->enable) & ibit))
87 ipi_mux_send(cpu);
88 }
89 }
90
91 static const struct irq_chip ipi_mux_chip = {
92 .name = "IPI Mux",
93 .irq_mask = ipi_mux_mask,
94 .irq_unmask = ipi_mux_unmask,
95 .ipi_send_mask = ipi_mux_send_mask,
96 };
97
ipi_mux_domain_alloc(struct irq_domain * d,unsigned int virq,unsigned int nr_irqs,void * arg)98 static int ipi_mux_domain_alloc(struct irq_domain *d, unsigned int virq,
99 unsigned int nr_irqs, void *arg)
100 {
101 int i;
102
103 for (i = 0; i < nr_irqs; i++) {
104 irq_set_percpu_devid(virq + i);
105 irq_domain_set_info(d, virq + i, i, &ipi_mux_chip, NULL,
106 handle_percpu_devid_irq, NULL, NULL);
107 }
108
109 return 0;
110 }
111
112 static const struct irq_domain_ops ipi_mux_domain_ops = {
113 .alloc = ipi_mux_domain_alloc,
114 .free = irq_domain_free_irqs_top,
115 };
116
117 /**
118 * ipi_mux_process - Process multiplexed virtual IPIs
119 */
ipi_mux_process(void)120 void ipi_mux_process(void)
121 {
122 struct ipi_mux_cpu *icpu = this_cpu_ptr(ipi_mux_pcpu);
123 irq_hw_number_t hwirq;
124 unsigned long ipis;
125 unsigned int en;
126
127 /*
128 * Reading enable mask does not need to be ordered as long as
129 * this function is called from interrupt handler because only
130 * the CPU itself can change it's own enable mask.
131 */
132 en = atomic_read(&icpu->enable);
133
134 /*
135 * Clear the IPIs we are about to handle. This pairs with the
136 * atomic_fetch_or_release() in ipi_mux_send_mask().
137 */
138 ipis = atomic_fetch_andnot(en, &icpu->bits) & en;
139
140 for_each_set_bit(hwirq, &ipis, BITS_PER_TYPE(int))
141 generic_handle_domain_irq(ipi_mux_domain, hwirq);
142 }
143
144 /**
145 * ipi_mux_create - Create virtual IPIs multiplexed on top of a single
146 * parent IPI.
147 * @nr_ipi: number of virtual IPIs to create. This should
148 * be <= BITS_PER_TYPE(int)
149 * @mux_send: callback to trigger parent IPI for a particular CPU
150 *
151 * Returns first virq of the newly created virtual IPIs upon success
152 * or <=0 upon failure
153 */
ipi_mux_create(unsigned int nr_ipi,void (* mux_send)(unsigned int cpu))154 int ipi_mux_create(unsigned int nr_ipi, void (*mux_send)(unsigned int cpu))
155 {
156 struct fwnode_handle *fwnode;
157 struct irq_domain *domain;
158 int rc;
159
160 if (ipi_mux_domain)
161 return -EEXIST;
162
163 if (BITS_PER_TYPE(int) < nr_ipi || !mux_send)
164 return -EINVAL;
165
166 ipi_mux_pcpu = alloc_percpu(typeof(*ipi_mux_pcpu));
167 if (!ipi_mux_pcpu)
168 return -ENOMEM;
169
170 fwnode = irq_domain_alloc_named_fwnode("IPI-Mux");
171 if (!fwnode) {
172 pr_err("unable to create IPI Mux fwnode\n");
173 rc = -ENOMEM;
174 goto fail_free_cpu;
175 }
176
177 domain = irq_domain_create_linear(fwnode, nr_ipi,
178 &ipi_mux_domain_ops, NULL);
179 if (!domain) {
180 pr_err("unable to add IPI Mux domain\n");
181 rc = -ENOMEM;
182 goto fail_free_fwnode;
183 }
184
185 domain->flags |= IRQ_DOMAIN_FLAG_IPI_SINGLE;
186 irq_domain_update_bus_token(domain, DOMAIN_BUS_IPI);
187
188 rc = irq_domain_alloc_irqs(domain, nr_ipi, NUMA_NO_NODE, NULL);
189 if (rc <= 0) {
190 pr_err("unable to alloc IRQs from IPI Mux domain\n");
191 goto fail_free_domain;
192 }
193
194 ipi_mux_domain = domain;
195 ipi_mux_send = mux_send;
196
197 return rc;
198
199 fail_free_domain:
200 irq_domain_remove(domain);
201 fail_free_fwnode:
202 irq_domain_free_fwnode(fwnode);
203 fail_free_cpu:
204 free_percpu(ipi_mux_pcpu);
205 return rc;
206 }
207