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