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