blob: ea192ec029c45bf7354d8184b44e1eb460cdbf77 [file] [log] [blame]
Jens Axboef2298c02013-10-25 11:52:25 +01001#include <linux/module.h>
2#include <linux/moduleparam.h>
3#include <linux/sched.h>
4#include <linux/fs.h>
5#include <linux/blkdev.h>
6#include <linux/init.h>
7#include <linux/slab.h>
8#include <linux/blk-mq.h>
9#include <linux/hrtimer.h>
10
11struct nullb_cmd {
12 struct list_head list;
13 struct llist_node ll_list;
14 struct call_single_data csd;
15 struct request *rq;
16 struct bio *bio;
17 unsigned int tag;
18 struct nullb_queue *nq;
19};
20
21struct nullb_queue {
22 unsigned long *tag_map;
23 wait_queue_head_t wait;
24 unsigned int queue_depth;
25
26 struct nullb_cmd *cmds;
27};
28
29struct nullb {
30 struct list_head list;
31 unsigned int index;
32 struct request_queue *q;
33 struct gendisk *disk;
34 struct hrtimer timer;
35 unsigned int queue_depth;
36 spinlock_t lock;
37
38 struct nullb_queue *queues;
39 unsigned int nr_queues;
40};
41
42static LIST_HEAD(nullb_list);
43static struct mutex lock;
44static int null_major;
45static int nullb_indexes;
46
47struct completion_queue {
48 struct llist_head list;
49 struct hrtimer timer;
50};
51
52/*
53 * These are per-cpu for now, they will need to be configured by the
54 * complete_queues parameter and appropriately mapped.
55 */
56static DEFINE_PER_CPU(struct completion_queue, completion_queues);
57
58enum {
59 NULL_IRQ_NONE = 0,
60 NULL_IRQ_SOFTIRQ = 1,
61 NULL_IRQ_TIMER = 2,
62
63 NULL_Q_BIO = 0,
64 NULL_Q_RQ = 1,
65 NULL_Q_MQ = 2,
66};
67
68static int submit_queues = 1;
69module_param(submit_queues, int, S_IRUGO);
70MODULE_PARM_DESC(submit_queues, "Number of submission queues");
71
72static int home_node = NUMA_NO_NODE;
73module_param(home_node, int, S_IRUGO);
74MODULE_PARM_DESC(home_node, "Home node for the device");
75
76static int queue_mode = NULL_Q_MQ;
77module_param(queue_mode, int, S_IRUGO);
78MODULE_PARM_DESC(use_mq, "Use blk-mq interface (0=bio,1=rq,2=multiqueue)");
79
80static int gb = 250;
81module_param(gb, int, S_IRUGO);
82MODULE_PARM_DESC(gb, "Size in GB");
83
84static int bs = 512;
85module_param(bs, int, S_IRUGO);
86MODULE_PARM_DESC(bs, "Block size (in bytes)");
87
88static int nr_devices = 2;
89module_param(nr_devices, int, S_IRUGO);
90MODULE_PARM_DESC(nr_devices, "Number of devices to register");
91
92static int irqmode = NULL_IRQ_SOFTIRQ;
93module_param(irqmode, int, S_IRUGO);
94MODULE_PARM_DESC(irqmode, "IRQ completion handler. 0-none, 1-softirq, 2-timer");
95
96static int completion_nsec = 10000;
97module_param(completion_nsec, int, S_IRUGO);
98MODULE_PARM_DESC(completion_nsec, "Time in ns to complete a request in hardware. Default: 10,000ns");
99
100static int hw_queue_depth = 64;
101module_param(hw_queue_depth, int, S_IRUGO);
102MODULE_PARM_DESC(hw_queue_depth, "Queue depth for each hardware queue. Default: 64");
103
104static bool use_per_node_hctx = true;
105module_param(use_per_node_hctx, bool, S_IRUGO);
106MODULE_PARM_DESC(use_per_node_hctx, "Use per-node allocation for hardware context queues. Default: true");
107
108static void put_tag(struct nullb_queue *nq, unsigned int tag)
109{
110 clear_bit_unlock(tag, nq->tag_map);
111
112 if (waitqueue_active(&nq->wait))
113 wake_up(&nq->wait);
114}
115
116static unsigned int get_tag(struct nullb_queue *nq)
117{
118 unsigned int tag;
119
120 do {
121 tag = find_first_zero_bit(nq->tag_map, nq->queue_depth);
122 if (tag >= nq->queue_depth)
123 return -1U;
124 } while (test_and_set_bit_lock(tag, nq->tag_map));
125
126 return tag;
127}
128
129static void free_cmd(struct nullb_cmd *cmd)
130{
131 put_tag(cmd->nq, cmd->tag);
132}
133
134static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
135{
136 struct nullb_cmd *cmd;
137 unsigned int tag;
138
139 tag = get_tag(nq);
140 if (tag != -1U) {
141 cmd = &nq->cmds[tag];
142 cmd->tag = tag;
143 cmd->nq = nq;
144 return cmd;
145 }
146
147 return NULL;
148}
149
150static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
151{
152 struct nullb_cmd *cmd;
153 DEFINE_WAIT(wait);
154
155 cmd = __alloc_cmd(nq);
156 if (cmd || !can_wait)
157 return cmd;
158
159 do {
160 prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
161 cmd = __alloc_cmd(nq);
162 if (cmd)
163 break;
164
165 io_schedule();
166 } while (1);
167
168 finish_wait(&nq->wait, &wait);
169 return cmd;
170}
171
172static void end_cmd(struct nullb_cmd *cmd)
173{
174 if (cmd->rq) {
175 if (queue_mode == NULL_Q_MQ)
176 blk_mq_end_io(cmd->rq, 0);
177 else {
178 INIT_LIST_HEAD(&cmd->rq->queuelist);
179 blk_end_request_all(cmd->rq, 0);
180 }
181 } else if (cmd->bio)
182 bio_endio(cmd->bio, 0);
183
184 if (queue_mode != NULL_Q_MQ)
185 free_cmd(cmd);
186}
187
188static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
189{
190 struct completion_queue *cq;
191 struct llist_node *entry;
192 struct nullb_cmd *cmd;
193
194 cq = &per_cpu(completion_queues, smp_processor_id());
195
196 while ((entry = llist_del_all(&cq->list)) != NULL) {
197 do {
198 cmd = container_of(entry, struct nullb_cmd, ll_list);
199 end_cmd(cmd);
200 entry = entry->next;
201 } while (entry);
202 }
203
204 return HRTIMER_NORESTART;
205}
206
207static void null_cmd_end_timer(struct nullb_cmd *cmd)
208{
209 struct completion_queue *cq = &per_cpu(completion_queues, get_cpu());
210
211 cmd->ll_list.next = NULL;
212 if (llist_add(&cmd->ll_list, &cq->list)) {
213 ktime_t kt = ktime_set(0, completion_nsec);
214
215 hrtimer_start(&cq->timer, kt, HRTIMER_MODE_REL);
216 }
217
218 put_cpu();
219}
220
221static void null_softirq_done_fn(struct request *rq)
222{
223 blk_end_request_all(rq, 0);
224}
225
Yuanhan Liu044c8d42013-11-21 14:32:01 -0800226#ifdef CONFIG_SMP
Jens Axboef2298c02013-10-25 11:52:25 +0100227
228static void null_ipi_cmd_end_io(void *data)
229{
230 struct completion_queue *cq;
231 struct llist_node *entry, *next;
232 struct nullb_cmd *cmd;
233
234 cq = &per_cpu(completion_queues, smp_processor_id());
235
236 entry = llist_del_all(&cq->list);
237
238 while (entry) {
239 next = entry->next;
240 cmd = llist_entry(entry, struct nullb_cmd, ll_list);
241 end_cmd(cmd);
242 entry = next;
243 }
244}
245
246static void null_cmd_end_ipi(struct nullb_cmd *cmd)
247{
248 struct call_single_data *data = &cmd->csd;
249 int cpu = get_cpu();
250 struct completion_queue *cq = &per_cpu(completion_queues, cpu);
251
252 cmd->ll_list.next = NULL;
253
254 if (llist_add(&cmd->ll_list, &cq->list)) {
255 data->func = null_ipi_cmd_end_io;
256 data->flags = 0;
257 __smp_call_function_single(cpu, data, 0);
258 }
259
260 put_cpu();
261}
262
Yuanhan Liu044c8d42013-11-21 14:32:01 -0800263#endif /* CONFIG_SMP */
Jens Axboef2298c02013-10-25 11:52:25 +0100264
265static inline void null_handle_cmd(struct nullb_cmd *cmd)
266{
267 /* Complete IO by inline, softirq or timer */
268 switch (irqmode) {
269 case NULL_IRQ_NONE:
270 end_cmd(cmd);
271 break;
272 case NULL_IRQ_SOFTIRQ:
Yuanhan Liu044c8d42013-11-21 14:32:01 -0800273#ifdef CONFIG_SMP
Jens Axboef2298c02013-10-25 11:52:25 +0100274 null_cmd_end_ipi(cmd);
275#else
276 end_cmd(cmd);
277#endif
278 break;
279 case NULL_IRQ_TIMER:
280 null_cmd_end_timer(cmd);
281 break;
282 }
283}
284
285static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
286{
287 int index = 0;
288
289 if (nullb->nr_queues != 1)
290 index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
291
292 return &nullb->queues[index];
293}
294
295static void null_queue_bio(struct request_queue *q, struct bio *bio)
296{
297 struct nullb *nullb = q->queuedata;
298 struct nullb_queue *nq = nullb_to_queue(nullb);
299 struct nullb_cmd *cmd;
300
301 cmd = alloc_cmd(nq, 1);
302 cmd->bio = bio;
303
304 null_handle_cmd(cmd);
305}
306
307static int null_rq_prep_fn(struct request_queue *q, struct request *req)
308{
309 struct nullb *nullb = q->queuedata;
310 struct nullb_queue *nq = nullb_to_queue(nullb);
311 struct nullb_cmd *cmd;
312
313 cmd = alloc_cmd(nq, 0);
314 if (cmd) {
315 cmd->rq = req;
316 req->special = cmd;
317 return BLKPREP_OK;
318 }
319
320 return BLKPREP_DEFER;
321}
322
323static void null_request_fn(struct request_queue *q)
324{
325 struct request *rq;
326
327 while ((rq = blk_fetch_request(q)) != NULL) {
328 struct nullb_cmd *cmd = rq->special;
329
330 spin_unlock_irq(q->queue_lock);
331 null_handle_cmd(cmd);
332 spin_lock_irq(q->queue_lock);
333 }
334}
335
336static int null_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *rq)
337{
338 struct nullb_cmd *cmd = rq->special;
339
340 cmd->rq = rq;
341 cmd->nq = hctx->driver_data;
342
343 null_handle_cmd(cmd);
344 return BLK_MQ_RQ_QUEUE_OK;
345}
346
347static struct blk_mq_hw_ctx *null_alloc_hctx(struct blk_mq_reg *reg, unsigned int hctx_index)
348{
349 return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL,
350 hctx_index);
351}
352
353static void null_free_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_index)
354{
355 kfree(hctx);
356}
357
358static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
359 unsigned int index)
360{
361 struct nullb *nullb = data;
362 struct nullb_queue *nq = &nullb->queues[index];
363
364 init_waitqueue_head(&nq->wait);
365 nq->queue_depth = nullb->queue_depth;
366 nullb->nr_queues++;
367 hctx->driver_data = nq;
368
369 return 0;
370}
371
372static struct blk_mq_ops null_mq_ops = {
373 .queue_rq = null_queue_rq,
374 .map_queue = blk_mq_map_queue,
375 .init_hctx = null_init_hctx,
376};
377
378static struct blk_mq_reg null_mq_reg = {
379 .ops = &null_mq_ops,
380 .queue_depth = 64,
381 .cmd_size = sizeof(struct nullb_cmd),
382 .flags = BLK_MQ_F_SHOULD_MERGE,
383};
384
385static void null_del_dev(struct nullb *nullb)
386{
387 list_del_init(&nullb->list);
388
389 del_gendisk(nullb->disk);
390 if (queue_mode == NULL_Q_MQ)
391 blk_mq_free_queue(nullb->q);
392 else
393 blk_cleanup_queue(nullb->q);
394 put_disk(nullb->disk);
395 kfree(nullb);
396}
397
398static int null_open(struct block_device *bdev, fmode_t mode)
399{
400 return 0;
401}
402
403static void null_release(struct gendisk *disk, fmode_t mode)
404{
405}
406
407static const struct block_device_operations null_fops = {
408 .owner = THIS_MODULE,
409 .open = null_open,
410 .release = null_release,
411};
412
413static int setup_commands(struct nullb_queue *nq)
414{
415 struct nullb_cmd *cmd;
416 int i, tag_size;
417
418 nq->cmds = kzalloc(nq->queue_depth * sizeof(*cmd), GFP_KERNEL);
419 if (!nq->cmds)
420 return 1;
421
422 tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
423 nq->tag_map = kzalloc(tag_size * sizeof(unsigned long), GFP_KERNEL);
424 if (!nq->tag_map) {
425 kfree(nq->cmds);
426 return 1;
427 }
428
429 for (i = 0; i < nq->queue_depth; i++) {
430 cmd = &nq->cmds[i];
431 INIT_LIST_HEAD(&cmd->list);
432 cmd->ll_list.next = NULL;
433 cmd->tag = -1U;
434 }
435
436 return 0;
437}
438
439static void cleanup_queue(struct nullb_queue *nq)
440{
441 kfree(nq->tag_map);
442 kfree(nq->cmds);
443}
444
445static void cleanup_queues(struct nullb *nullb)
446{
447 int i;
448
449 for (i = 0; i < nullb->nr_queues; i++)
450 cleanup_queue(&nullb->queues[i]);
451
452 kfree(nullb->queues);
453}
454
455static int setup_queues(struct nullb *nullb)
456{
457 struct nullb_queue *nq;
458 int i;
459
460 nullb->queues = kzalloc(submit_queues * sizeof(*nq), GFP_KERNEL);
461 if (!nullb->queues)
462 return 1;
463
464 nullb->nr_queues = 0;
465 nullb->queue_depth = hw_queue_depth;
466
467 if (queue_mode == NULL_Q_MQ)
468 return 0;
469
470 for (i = 0; i < submit_queues; i++) {
471 nq = &nullb->queues[i];
472 init_waitqueue_head(&nq->wait);
473 nq->queue_depth = hw_queue_depth;
474 if (setup_commands(nq))
475 break;
476 nullb->nr_queues++;
477 }
478
479 if (i == submit_queues)
480 return 0;
481
482 cleanup_queues(nullb);
483 return 1;
484}
485
486static int null_add_dev(void)
487{
488 struct gendisk *disk;
489 struct nullb *nullb;
490 sector_t size;
491
492 nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, home_node);
493 if (!nullb)
494 return -ENOMEM;
495
496 spin_lock_init(&nullb->lock);
497
498 if (setup_queues(nullb))
499 goto err;
500
501 if (queue_mode == NULL_Q_MQ) {
502 null_mq_reg.numa_node = home_node;
503 null_mq_reg.queue_depth = hw_queue_depth;
504
505 if (use_per_node_hctx) {
506 null_mq_reg.ops->alloc_hctx = null_alloc_hctx;
507 null_mq_reg.ops->free_hctx = null_free_hctx;
508
509 null_mq_reg.nr_hw_queues = nr_online_nodes;
510 } else {
511 null_mq_reg.ops->alloc_hctx = blk_mq_alloc_single_hw_queue;
512 null_mq_reg.ops->free_hctx = blk_mq_free_single_hw_queue;
513
514 null_mq_reg.nr_hw_queues = submit_queues;
515 }
516
517 nullb->q = blk_mq_init_queue(&null_mq_reg, nullb);
518 } else if (queue_mode == NULL_Q_BIO) {
519 nullb->q = blk_alloc_queue_node(GFP_KERNEL, home_node);
520 blk_queue_make_request(nullb->q, null_queue_bio);
521 } else {
522 nullb->q = blk_init_queue_node(null_request_fn, &nullb->lock, home_node);
523 blk_queue_prep_rq(nullb->q, null_rq_prep_fn);
524 if (nullb->q)
525 blk_queue_softirq_done(nullb->q, null_softirq_done_fn);
526 }
527
528 if (!nullb->q)
529 goto queue_fail;
530
531 nullb->q->queuedata = nullb;
532 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, nullb->q);
533
534 disk = nullb->disk = alloc_disk_node(1, home_node);
535 if (!disk) {
536queue_fail:
537 if (queue_mode == NULL_Q_MQ)
538 blk_mq_free_queue(nullb->q);
539 else
540 blk_cleanup_queue(nullb->q);
541 cleanup_queues(nullb);
542err:
543 kfree(nullb);
544 return -ENOMEM;
545 }
546
547 mutex_lock(&lock);
548 list_add_tail(&nullb->list, &nullb_list);
549 nullb->index = nullb_indexes++;
550 mutex_unlock(&lock);
551
552 blk_queue_logical_block_size(nullb->q, bs);
553 blk_queue_physical_block_size(nullb->q, bs);
554
555 size = gb * 1024 * 1024 * 1024ULL;
556 sector_div(size, bs);
557 set_capacity(disk, size);
558
559 disk->flags |= GENHD_FL_EXT_DEVT;
560 disk->major = null_major;
561 disk->first_minor = nullb->index;
562 disk->fops = &null_fops;
563 disk->private_data = nullb;
564 disk->queue = nullb->q;
565 sprintf(disk->disk_name, "nullb%d", nullb->index);
566 add_disk(disk);
567 return 0;
568}
569
570static int __init null_init(void)
571{
572 unsigned int i;
573
Yuanhan Liu044c8d42013-11-21 14:32:01 -0800574#if !defined(CONFIG_SMP)
Jens Axboef2298c02013-10-25 11:52:25 +0100575 if (irqmode == NULL_IRQ_SOFTIRQ) {
576 pr_warn("null_blk: softirq completions not available.\n");
577 pr_warn("null_blk: using direct completions.\n");
578 irqmode = NULL_IRQ_NONE;
579 }
580#endif
581
582 if (submit_queues > nr_cpu_ids)
583 submit_queues = nr_cpu_ids;
584 else if (!submit_queues)
585 submit_queues = 1;
586
587 mutex_init(&lock);
588
589 /* Initialize a separate list for each CPU for issuing softirqs */
590 for_each_possible_cpu(i) {
591 struct completion_queue *cq = &per_cpu(completion_queues, i);
592
593 init_llist_head(&cq->list);
594
595 if (irqmode != NULL_IRQ_TIMER)
596 continue;
597
598 hrtimer_init(&cq->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
599 cq->timer.function = null_cmd_timer_expired;
600 }
601
602 null_major = register_blkdev(0, "nullb");
603 if (null_major < 0)
604 return null_major;
605
606 for (i = 0; i < nr_devices; i++) {
607 if (null_add_dev()) {
608 unregister_blkdev(null_major, "nullb");
609 return -EINVAL;
610 }
611 }
612
613 pr_info("null: module loaded\n");
614 return 0;
615}
616
617static void __exit null_exit(void)
618{
619 struct nullb *nullb;
620
621 unregister_blkdev(null_major, "nullb");
622
623 mutex_lock(&lock);
624 while (!list_empty(&nullb_list)) {
625 nullb = list_entry(nullb_list.next, struct nullb, list);
626 null_del_dev(nullb);
627 }
628 mutex_unlock(&lock);
629}
630
631module_init(null_init);
632module_exit(null_exit);
633
634MODULE_AUTHOR("Jens Axboe <jaxboe@fusionio.com>");
635MODULE_LICENSE("GPL");