1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Functions related to interrupt-poll handling in the block layer. This
4 * is similar to NAPI for network devices.
5 */
6#include <linux/kernel.h>
7#include <linux/module.h>
8#include <linux/init.h>
9#include <linux/bio.h>
10#include <linux/interrupt.h>
11#include <linux/cpu.h>
12#include <linux/irq_poll.h>
13#include <linux/delay.h>
14
15static unsigned int irq_poll_budget __read_mostly = 256;
16
17static DEFINE_PER_CPU(struct list_head, blk_cpu_iopoll);
18
19/**
20 * irq_poll_sched - Schedule a run of the iopoll handler
21 * @iop:      The parent iopoll structure
22 *
23 * Description:
24 *     Add this irq_poll structure to the pending poll list and trigger the
25 *     raise of the blk iopoll softirq.
26 **/
27void irq_poll_sched(struct irq_poll *iop)
28{
29	unsigned long flags;
30
31	if (test_bit(IRQ_POLL_F_DISABLE, &iop->state))
32		return;
33	if (test_and_set_bit(IRQ_POLL_F_SCHED, &iop->state))
34		return;
35
36	local_irq_save(flags);
37	list_add_tail(&iop->list, this_cpu_ptr(&blk_cpu_iopoll));
38	raise_softirq_irqoff(IRQ_POLL_SOFTIRQ);
39	local_irq_restore(flags);
40}
41EXPORT_SYMBOL(irq_poll_sched);
42
43/**
44 * __irq_poll_complete - Mark this @iop as un-polled again
45 * @iop:      The parent iopoll structure
46 *
47 * Description:
48 *     See irq_poll_complete(). This function must be called with interrupts
49 *     disabled.
50 **/
51static void __irq_poll_complete(struct irq_poll *iop)
52{
53	list_del(&iop->list);
54	smp_mb__before_atomic();
55	clear_bit_unlock(IRQ_POLL_F_SCHED, &iop->state);
56}
57
58/**
59 * irq_poll_complete - Mark this @iop as un-polled again
60 * @iop:      The parent iopoll structure
61 *
62 * Description:
63 *     If a driver consumes less than the assigned budget in its run of the
64 *     iopoll handler, it'll end the polled mode by calling this function. The
65 *     iopoll handler will not be invoked again before irq_poll_sched()
66 *     is called.
67 **/
68void irq_poll_complete(struct irq_poll *iop)
69{
70	unsigned long flags;
71
72	local_irq_save(flags);
73	__irq_poll_complete(iop);
74	local_irq_restore(flags);
75}
76EXPORT_SYMBOL(irq_poll_complete);
77
78static void __latent_entropy irq_poll_softirq(struct softirq_action *h)
79{
80	struct list_head *list = this_cpu_ptr(&blk_cpu_iopoll);
81	int rearm = 0, budget = irq_poll_budget;
82	unsigned long start_time = jiffies;
83
84	local_irq_disable();
85
86	while (!list_empty(list)) {
87		struct irq_poll *iop;
88		int work, weight;
89
90		/*
91		 * If softirq window is exhausted then punt.
92		 */
93		if (budget <= 0 || time_after(jiffies, start_time)) {
94			rearm = 1;
95			break;
96		}
97
98		local_irq_enable();
99
100		/* Even though interrupts have been re-enabled, this
101		 * access is safe because interrupts can only add new
102		 * entries to the tail of this list, and only ->poll()
103		 * calls can remove this head entry from the list.
104		 */
105		iop = list_entry(list->next, struct irq_poll, list);
106
107		weight = iop->weight;
108		work = 0;
109		if (test_bit(IRQ_POLL_F_SCHED, &iop->state))
110			work = iop->poll(iop, weight);
111
112		budget -= work;
113
114		local_irq_disable();
115
116		/*
117		 * Drivers must not modify the iopoll state, if they
118		 * consume their assigned weight (or more, some drivers can't
119		 * easily just stop processing, they have to complete an
120		 * entire mask of commands).In such cases this code
121		 * still "owns" the iopoll instance and therefore can
122		 * move the instance around on the list at-will.
123		 */
124		if (work >= weight) {
125			if (test_bit(IRQ_POLL_F_DISABLE, &iop->state))
126				__irq_poll_complete(iop);
127			else
128				list_move_tail(&iop->list, list);
129		}
130	}
131
132	if (rearm)
133		__raise_softirq_irqoff(IRQ_POLL_SOFTIRQ);
134
135	local_irq_enable();
136}
137
138/**
139 * irq_poll_disable - Disable iopoll on this @iop
140 * @iop:      The parent iopoll structure
141 *
142 * Description:
143 *     Disable io polling and wait for any pending callbacks to have completed.
144 **/
145void irq_poll_disable(struct irq_poll *iop)
146{
147	set_bit(IRQ_POLL_F_DISABLE, &iop->state);
148	while (test_and_set_bit(IRQ_POLL_F_SCHED, &iop->state))
149		msleep(1);
150	clear_bit(IRQ_POLL_F_DISABLE, &iop->state);
151}
152EXPORT_SYMBOL(irq_poll_disable);
153
154/**
155 * irq_poll_enable - Enable iopoll on this @iop
156 * @iop:      The parent iopoll structure
157 *
158 * Description:
159 *     Enable iopoll on this @iop. Note that the handler run will not be
160 *     scheduled, it will only mark it as active.
161 **/
162void irq_poll_enable(struct irq_poll *iop)
163{
164	BUG_ON(!test_bit(IRQ_POLL_F_SCHED, &iop->state));
165	smp_mb__before_atomic();
166	clear_bit_unlock(IRQ_POLL_F_SCHED, &iop->state);
167}
168EXPORT_SYMBOL(irq_poll_enable);
169
170/**
171 * irq_poll_init - Initialize this @iop
172 * @iop:      The parent iopoll structure
173 * @weight:   The default weight (or command completion budget)
174 * @poll_fn:  The handler to invoke
175 *
176 * Description:
177 *     Initialize and enable this irq_poll structure.
178 **/
179void irq_poll_init(struct irq_poll *iop, int weight, irq_poll_fn *poll_fn)
180{
181	memset(iop, 0, sizeof(*iop));
182	INIT_LIST_HEAD(&iop->list);
183	iop->weight = weight;
184	iop->poll = poll_fn;
185}
186EXPORT_SYMBOL(irq_poll_init);
187
188static int irq_poll_cpu_dead(unsigned int cpu)
189{
190	/*
191	 * If a CPU goes away, splice its entries to the current CPU and
192	 * set the POLL softirq bit. The local_bh_disable()/enable() pair
193	 * ensures that it is handled. Otherwise the current CPU could
194	 * reach idle with the POLL softirq pending.
195	 */
196	local_bh_disable();
197	local_irq_disable();
198	list_splice_init(&per_cpu(blk_cpu_iopoll, cpu),
199			 this_cpu_ptr(&blk_cpu_iopoll));
200	__raise_softirq_irqoff(IRQ_POLL_SOFTIRQ);
201	local_irq_enable();
202	local_bh_enable();
203
204	return 0;
205}
206
207static __init int irq_poll_setup(void)
208{
209	int i;
210
211	for_each_possible_cpu(i)
212		INIT_LIST_HEAD(&per_cpu(blk_cpu_iopoll, i));
213
214	open_softirq(IRQ_POLL_SOFTIRQ, irq_poll_softirq);
215	cpuhp_setup_state_nocalls(CPUHP_IRQ_POLL_DEAD, "irq_poll:dead", NULL,
216				  irq_poll_cpu_dead);
217	return 0;
218}
219subsys_initcall(irq_poll_setup);
220