bsg: fix bsg_poll() to return POLLOUT properly
[linux-2.6.git] / block / blk-iopoll.c
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
17 int blk_iopoll_enabled = 1;
18 EXPORT_SYMBOL(blk_iopoll_enabled);
19
20 static unsigned int blk_iopoll_budget __read_mostly = 256;
21
22 static DEFINE_PER_CPU(struct list_head, blk_cpu_iopoll);
23
24 /**
25  * blk_iopoll_sched - Schedule a run of the iopoll handler
26  * @iop:      The parent iopoll structure
27  *
28  * Description:
29  *     Add this blk_iopoll structure to the pending poll list and trigger the
30  *     raise of the blk iopoll softirq. The driver must already have gotten a
31  *     successful return from blk_iopoll_sched_prep() before calling this.
32  **/
33 void blk_iopoll_sched(struct blk_iopoll *iop)
34 {
35         unsigned long flags;
36
37         local_irq_save(flags);
38         list_add_tail(&iop->list, &__get_cpu_var(blk_cpu_iopoll));
39         __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
40         local_irq_restore(flags);
41 }
42 EXPORT_SYMBOL(blk_iopoll_sched);
43
44 /**
45  * __blk_iopoll_complete - Mark this @iop as un-polled again
46  * @iop:      The parent iopoll structure
47  *
48  * Description:
49  *     See blk_iopoll_complete(). This function must be called with interrupts
50  *     disabled.
51  **/
52 void __blk_iopoll_complete(struct blk_iopoll *iop)
53 {
54         list_del(&iop->list);
55         smp_mb__before_clear_bit();
56         clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
57 }
58 EXPORT_SYMBOL(__blk_iopoll_complete);
59
60 /**
61  * blk_iopoll_complete - Mark this @iop as un-polled again
62  * @iop:      The parent iopoll structure
63  *
64  * Description:
65  *     If a driver consumes less than the assigned budget in its run of the
66  *     iopoll handler, it'll end the polled mode by calling this function. The
67  *     iopoll handler will not be invoked again before blk_iopoll_sched_prep()
68  *     is called.
69  **/
70 void blk_iopoll_complete(struct blk_iopoll *iopoll)
71 {
72         unsigned long flags;
73
74         local_irq_save(flags);
75         __blk_iopoll_complete(iopoll);
76         local_irq_restore(flags);
77 }
78 EXPORT_SYMBOL(blk_iopoll_complete);
79
80 static void blk_iopoll_softirq(struct softirq_action *h)
81 {
82         struct list_head *list = &__get_cpu_var(blk_cpu_iopoll);
83         int rearm = 0, budget = blk_iopoll_budget;
84         unsigned long start_time = jiffies;
85
86         local_irq_disable();
87
88         while (!list_empty(list)) {
89                 struct blk_iopoll *iop;
90                 int work, weight;
91
92                 /*
93                  * If softirq window is exhausted then punt.
94                  */
95                 if (budget <= 0 || time_after(jiffies, start_time)) {
96                         rearm = 1;
97                         break;
98                 }
99
100                 local_irq_enable();
101
102                 /* Even though interrupts have been re-enabled, this
103                  * access is safe because interrupts can only add new
104                  * entries to the tail of this list, and only ->poll()
105                  * calls can remove this head entry from the list.
106                  */
107                 iop = list_entry(list->next, struct blk_iopoll, list);
108
109                 weight = iop->weight;
110                 work = 0;
111                 if (test_bit(IOPOLL_F_SCHED, &iop->state))
112                         work = iop->poll(iop, weight);
113
114                 budget -= work;
115
116                 local_irq_disable();
117
118                 /*
119                  * Drivers must not modify the iopoll state, if they
120                  * consume their assigned weight (or more, some drivers can't
121                  * easily just stop processing, they have to complete an
122                  * entire mask of commands).In such cases this code
123                  * still "owns" the iopoll instance and therefore can
124                  * move the instance around on the list at-will.
125                  */
126                 if (work >= weight) {
127                         if (blk_iopoll_disable_pending(iop))
128                                 __blk_iopoll_complete(iop);
129                         else
130                                 list_move_tail(&iop->list, list);
131                 }
132         }
133
134         if (rearm)
135                 __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
136
137         local_irq_enable();
138 }
139
140 /**
141  * blk_iopoll_disable - Disable iopoll on this @iop
142  * @iop:      The parent iopoll structure
143  *
144  * Description:
145  *     Disable io polling and wait for any pending callbacks to have completed.
146  **/
147 void blk_iopoll_disable(struct blk_iopoll *iop)
148 {
149         set_bit(IOPOLL_F_DISABLE, &iop->state);
150         while (test_and_set_bit(IOPOLL_F_SCHED, &iop->state))
151                 msleep(1);
152         clear_bit(IOPOLL_F_DISABLE, &iop->state);
153 }
154 EXPORT_SYMBOL(blk_iopoll_disable);
155
156 /**
157  * blk_iopoll_enable - Enable iopoll on this @iop
158  * @iop:      The parent iopoll structure
159  *
160  * Description:
161  *     Enable iopoll on this @iop. Note that the handler run will not be
162  *     scheduled, it will only mark it as active.
163  **/
164 void blk_iopoll_enable(struct blk_iopoll *iop)
165 {
166         BUG_ON(!test_bit(IOPOLL_F_SCHED, &iop->state));
167         smp_mb__before_clear_bit();
168         clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
169 }
170 EXPORT_SYMBOL(blk_iopoll_enable);
171
172 /**
173  * blk_iopoll_init - Initialize this @iop
174  * @iop:      The parent iopoll structure
175  * @weight:   The default weight (or command completion budget)
176  * @poll_fn:  The handler to invoke
177  *
178  * Description:
179  *     Initialize this blk_iopoll structure. Before being actively used, the
180  *     driver must call blk_iopoll_enable().
181  **/
182 void blk_iopoll_init(struct blk_iopoll *iop, int weight, blk_iopoll_fn *poll_fn)
183 {
184         memset(iop, 0, sizeof(*iop));
185         INIT_LIST_HEAD(&iop->list);
186         iop->weight = weight;
187         iop->poll = poll_fn;
188         set_bit(IOPOLL_F_SCHED, &iop->state);
189 }
190 EXPORT_SYMBOL(blk_iopoll_init);
191
192 static int __cpuinit blk_iopoll_cpu_notify(struct notifier_block *self,
193                                           unsigned long action, void *hcpu)
194 {
195         /*
196          * If a CPU goes away, splice its entries to the current CPU
197          * and trigger a run of the softirq
198          */
199         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
200                 int cpu = (unsigned long) hcpu;
201
202                 local_irq_disable();
203                 list_splice_init(&per_cpu(blk_cpu_iopoll, cpu),
204                                  &__get_cpu_var(blk_cpu_iopoll));
205                 __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
206                 local_irq_enable();
207         }
208
209         return NOTIFY_OK;
210 }
211
212 static struct notifier_block __cpuinitdata blk_iopoll_cpu_notifier = {
213         .notifier_call  = blk_iopoll_cpu_notify,
214 };
215
216 static __init int blk_iopoll_setup(void)
217 {
218         int i;
219
220         for_each_possible_cpu(i)
221                 INIT_LIST_HEAD(&per_cpu(blk_cpu_iopoll, i));
222
223         open_softirq(BLOCK_IOPOLL_SOFTIRQ, blk_iopoll_softirq);
224         register_hotcpu_notifier(&blk_iopoll_cpu_notifier);
225         return 0;
226 }
227 subsys_initcall(blk_iopoll_setup);