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/blkdev.h>
10#include <linux/interrupt.h>
11#include <linux/cpu.h>
12#include <linux/blk-iopoll.h>
13#include <linux/delay.h>
14
15#include "blk.h"
16
17static unsigned int blk_iopoll_budget __read_mostly = 256;
18
19static DEFINE_PER_CPU(struct list_head, blk_cpu_iopoll);
20
21/**
22 * blk_iopoll_sched - Schedule a run of the iopoll handler
23 * @iop:      The parent iopoll structure
24 *
25 * Description:
26 *     Add this blk_iopoll structure to the pending poll list and trigger the
27 *     raise of the blk iopoll softirq. The driver must already have gotten a
28 *     successful return from blk_iopoll_sched_prep() before calling this.
29 **/
30void blk_iopoll_sched(struct blk_iopoll *iop)
31{
32	unsigned long flags;
33
34	local_irq_save(flags);
35	list_add_tail(&iop->list, this_cpu_ptr(&blk_cpu_iopoll));
36	__raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
37	local_irq_restore(flags);
38}
39EXPORT_SYMBOL(blk_iopoll_sched);
40
41/**
42 * __blk_iopoll_complete - Mark this @iop as un-polled again
43 * @iop:      The parent iopoll structure
44 *
45 * Description:
46 *     See blk_iopoll_complete(). This function must be called with interrupts
47 *     disabled.
48 **/
49void __blk_iopoll_complete(struct blk_iopoll *iop)
50{
51	list_del(&iop->list);
52	smp_mb__before_atomic();
53	clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
54}
55EXPORT_SYMBOL(__blk_iopoll_complete);
56
57/**
58 * blk_iopoll_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 blk_iopoll_sched_prep()
65 *     is called.
66 **/
67void blk_iopoll_complete(struct blk_iopoll *iop)
68{
69	unsigned long flags;
70
71	local_irq_save(flags);
72	__blk_iopoll_complete(iop);
73	local_irq_restore(flags);
74}
75EXPORT_SYMBOL(blk_iopoll_complete);
76
77static void blk_iopoll_softirq(struct softirq_action *h)
78{
79	struct list_head *list = this_cpu_ptr(&blk_cpu_iopoll);
80	int rearm = 0, budget = blk_iopoll_budget;
81	unsigned long start_time = jiffies;
82
83	local_irq_disable();
84
85	while (!list_empty(list)) {
86		struct blk_iopoll *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 blk_iopoll, list);
105
106		weight = iop->weight;
107		work = 0;
108		if (test_bit(IOPOLL_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 (blk_iopoll_disable_pending(iop))
125				__blk_iopoll_complete(iop);
126			else
127				list_move_tail(&iop->list, list);
128		}
129	}
130
131	if (rearm)
132		__raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
133
134	local_irq_enable();
135}
136
137/**
138 * blk_iopoll_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 blk_iopoll_disable(struct blk_iopoll *iop)
145{
146	set_bit(IOPOLL_F_DISABLE, &iop->state);
147	while (test_and_set_bit(IOPOLL_F_SCHED, &iop->state))
148		msleep(1);
149	clear_bit(IOPOLL_F_DISABLE, &iop->state);
150}
151EXPORT_SYMBOL(blk_iopoll_disable);
152
153/**
154 * blk_iopoll_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 blk_iopoll_enable(struct blk_iopoll *iop)
162{
163	BUG_ON(!test_bit(IOPOLL_F_SCHED, &iop->state));
164	smp_mb__before_atomic();
165	clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
166}
167EXPORT_SYMBOL(blk_iopoll_enable);
168
169/**
170 * blk_iopoll_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 this blk_iopoll structure. Before being actively used, the
177 *     driver must call blk_iopoll_enable().
178 **/
179void blk_iopoll_init(struct blk_iopoll *iop, int weight, blk_iopoll_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	set_bit(IOPOLL_F_SCHED, &iop->state);
186}
187EXPORT_SYMBOL(blk_iopoll_init);
188
189static int blk_iopoll_cpu_notify(struct notifier_block *self,
190				 unsigned long action, void *hcpu)
191{
192	/*
193	 * If a CPU goes away, splice its entries to the current CPU
194	 * and trigger a run of the softirq
195	 */
196	if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
197		int cpu = (unsigned long) hcpu;
198
199		local_irq_disable();
200		list_splice_init(&per_cpu(blk_cpu_iopoll, cpu),
201				 this_cpu_ptr(&blk_cpu_iopoll));
202		__raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
203		local_irq_enable();
204	}
205
206	return NOTIFY_OK;
207}
208
209static struct notifier_block blk_iopoll_cpu_notifier = {
210	.notifier_call	= blk_iopoll_cpu_notify,
211};
212
213static __init int blk_iopoll_setup(void)
214{
215	int i;
216
217	for_each_possible_cpu(i)
218		INIT_LIST_HEAD(&per_cpu(blk_cpu_iopoll, i));
219
220	open_softirq(BLOCK_IOPOLL_SOFTIRQ, blk_iopoll_softirq);
221	register_hotcpu_notifier(&blk_iopoll_cpu_notifier);
222	return 0;
223}
224subsys_initcall(blk_iopoll_setup);
225