[linux-2.6-block.git] / drivers / s390 / block / scm_blk_cluster.c

/*
 * Block driver for s390 storage class memory.
 *
 * Copyright IBM Corp. 2012
 * Author(s): Sebastian Ott <sebott@linux.vnet.ibm.com>
 */

#include <linux/spinlock.h>
#include <linux/module.h>
#include <linux/blkdev.h>
#include <linux/genhd.h>
#include <linux/slab.h>
#include <linux/list.h>
#include <asm/eadm.h>
#include "scm_blk.h"

static unsigned int write_cluster_size = 64;
module_param(write_cluster_size, uint, S_IRUGO);
MODULE_PARM_DESC(write_cluster_size,
		 "Number of pages used for contiguous writes.");

#define CLUSTER_SIZE (write_cluster_size * PAGE_SIZE)

void __scm_free_rq_cluster(struct scm_request *scmrq)
{
	int i;

	if (!scmrq->cluster.buf)
		return;

	for (i = 0; i < 2 * write_cluster_size; i++)
		free_page((unsigned long) scmrq->cluster.buf[i]);

	kfree(scmrq->cluster.buf);
}

int __scm_alloc_rq_cluster(struct scm_request *scmrq)
{
	int i;

	scmrq->cluster.buf = kzalloc(sizeof(void *) * 2 * write_cluster_size,
				 GFP_KERNEL);
	if (!scmrq->cluster.buf)
		return -ENOMEM;

	for (i = 0; i < 2 * write_cluster_size; i++) {
		scmrq->cluster.buf[i] = (void *) get_zeroed_page(GFP_DMA);
		if (!scmrq->cluster.buf[i])
			return -ENOMEM;
	}
	INIT_LIST_HEAD(&scmrq->cluster.list);
	return 0;
}

void scm_request_cluster_init(struct scm_request *scmrq)
{
	scmrq->cluster.state = CLUSTER_NONE;
}

static bool clusters_intersect(struct request *A, struct request *B)
{
	unsigned long firstA, lastA, firstB, lastB;

	firstA = ((u64) blk_rq_pos(A) << 9) / CLUSTER_SIZE;
	lastA = (((u64) blk_rq_pos(A) << 9) +
		    blk_rq_bytes(A) - 1) / CLUSTER_SIZE;

	firstB = ((u64) blk_rq_pos(B) << 9) / CLUSTER_SIZE;
	lastB = (((u64) blk_rq_pos(B) << 9) +
		    blk_rq_bytes(B) - 1) / CLUSTER_SIZE;

	return (firstB <= lastA && firstA <= lastB);
}

bool scm_reserve_cluster(struct scm_request *scmrq)
{
	struct request *req = scmrq->request[scmrq->aob->request.msb_count];
	struct scm_blk_dev *bdev = scmrq->bdev;
	struct scm_request *iter;
	int pos, add = 1;

	if (write_cluster_size == 0)
		return true;

	spin_lock(&bdev->lock);
	list_for_each_entry(iter, &bdev->cluster_list, cluster.list) {
		if (iter == scmrq) {
			/*
			 * We don't have to use clusters_intersect here, since
			 * cluster requests are always started separately.
			 */
			add = 0;
			continue;
		}
		for (pos = 0; pos < iter->aob->request.msb_count; pos++) {
			if (clusters_intersect(req, iter->request[pos]) &&
			    (rq_data_dir(req) == WRITE ||
			     rq_data_dir(iter->request[pos]) == WRITE)) {
				spin_unlock(&bdev->lock);
				return false;
			}
		}
	}
	if (add)
		list_add(&scmrq->cluster.list, &bdev->cluster_list);
	spin_unlock(&bdev->lock);

	return true;
}

void scm_release_cluster(struct scm_request *scmrq)
{
	struct scm_blk_dev *bdev = scmrq->bdev;
	unsigned long flags;

	if (write_cluster_size == 0)
		return;

	spin_lock_irqsave(&bdev->lock, flags);
	list_del(&scmrq->cluster.list);
	spin_unlock_irqrestore(&bdev->lock, flags);
}

void scm_blk_dev_cluster_setup(struct scm_blk_dev *bdev)
{
	INIT_LIST_HEAD(&bdev->cluster_list);
	blk_queue_io_opt(bdev->rq, CLUSTER_SIZE);
}

static int scm_prepare_cluster_request(struct scm_request *scmrq)
{
	struct scm_blk_dev *bdev = scmrq->bdev;
	struct scm_device *scmdev = bdev->gendisk->private_data;
	struct request *req = scmrq->request[0];
	struct msb *msb = &scmrq->aob->msb[0];
	struct req_iterator iter;
	struct aidaw *aidaw;
	struct bio_vec bv;
	int i = 0;
	u64 addr;

	switch (scmrq->cluster.state) {
	case CLUSTER_NONE:
		scmrq->cluster.state = CLUSTER_READ;
		/* fall through */
	case CLUSTER_READ:
		msb->bs = MSB_BS_4K;
		msb->oc = MSB_OC_READ;
		msb->flags = MSB_FLAG_IDA;
		msb->blk_count = write_cluster_size;

		addr = scmdev->address + ((u64) blk_rq_pos(req) << 9);
		msb->scm_addr = round_down(addr, CLUSTER_SIZE);

		if (msb->scm_addr !=
		    round_down(addr + (u64) blk_rq_bytes(req) - 1,
			       CLUSTER_SIZE))
			msb->blk_count = 2 * write_cluster_size;

		aidaw = scm_aidaw_fetch(scmrq, msb->blk_count * PAGE_SIZE);
		if (!aidaw)
			return -ENOMEM;

		scmrq->aob->request.msb_count = 1;
		msb->data_addr = (u64) aidaw;
		for (i = 0; i < msb->blk_count; i++) {
			aidaw->data_addr = (u64) scmrq->cluster.buf[i];
			aidaw++;
		}

		break;
	case CLUSTER_WRITE:
		aidaw = (void *) msb->data_addr;
		msb->oc = MSB_OC_WRITE;

		for (addr = msb->scm_addr;
		     addr < scmdev->address + ((u64) blk_rq_pos(req) << 9);
		     addr += PAGE_SIZE) {
			aidaw->data_addr = (u64) scmrq->cluster.buf[i];
			aidaw++;
			i++;
		}
		rq_for_each_segment(bv, req, iter) {
			aidaw->data_addr = (u64) page_address(bv.bv_page);
			aidaw++;
			i++;
		}
		for (; i < msb->blk_count; i++) {
			aidaw->data_addr = (u64) scmrq->cluster.buf[i];
			aidaw++;
		}
		break;
	}
	return 0;
}

bool scm_need_cluster_request(struct scm_request *scmrq)
{
	int pos = scmrq->aob->request.msb_count;

	if (rq_data_dir(scmrq->request[pos]) == READ)
		return false;

	return blk_rq_bytes(scmrq->request[pos]) < CLUSTER_SIZE;
}

/* Called with queue lock held. */
void scm_initiate_cluster_request(struct scm_request *scmrq)
{
	if (scm_prepare_cluster_request(scmrq))
		goto requeue;
	if (eadm_start_aob(scmrq->aob))
		goto requeue;
	return;
requeue:
	scm_request_requeue(scmrq);
}

bool scm_test_cluster_request(struct scm_request *scmrq)
{
	return scmrq->cluster.state != CLUSTER_NONE;
}

void scm_cluster_request_irq(struct scm_request *scmrq)
{
	struct scm_blk_dev *bdev = scmrq->bdev;
	unsigned long flags;

	switch (scmrq->cluster.state) {
	case CLUSTER_NONE:
		BUG();
		break;
	case CLUSTER_READ:
		if (scmrq->error) {
			scm_request_finish(scmrq);
			break;
		}
		scmrq->cluster.state = CLUSTER_WRITE;
		spin_lock_irqsave(&bdev->rq_lock, flags);
		scm_initiate_cluster_request(scmrq);
		spin_unlock_irqrestore(&bdev->rq_lock, flags);
		break;
	case CLUSTER_WRITE:
		scm_request_finish(scmrq);
		break;
	}
}

bool scm_cluster_size_valid(void)
{
	if (write_cluster_size == 1 || write_cluster_size > 128)
		return false;

	return !(write_cluster_size & (write_cluster_size - 1));
}
Commit	Line	Data
0d804b20 SO	1	/*
	2	* Block driver for s390 storage class memory.
	3	*
	4	* Copyright IBM Corp. 2012
	5	* Author(s): Sebastian Ott <sebott@linux.vnet.ibm.com>
	6	*/
	7
	8	#include <linux/spinlock.h>
	9	#include <linux/module.h>
	10	#include <linux/blkdev.h>
	11	#include <linux/genhd.h>
	12	#include <linux/slab.h>
	13	#include <linux/list.h>
	14	#include <asm/eadm.h>
	15	#include "scm_blk.h"
	16
	17	static unsigned int write_cluster_size = 64;
	18	module_param(write_cluster_size, uint, S_IRUGO);
	19	MODULE_PARM_DESC(write_cluster_size,
	20	"Number of pages used for contiguous writes.");
	21
	22	#define CLUSTER_SIZE (write_cluster_size * PAGE_SIZE)
	23
	24	void __scm_free_rq_cluster(struct scm_request *scmrq)
	25	{
	26	int i;
	27
	28	if (!scmrq->cluster.buf)
	29	return;
	30
	31	for (i = 0; i < 2 * write_cluster_size; i++)
	32	free_page((unsigned long) scmrq->cluster.buf[i]);
	33
	34	kfree(scmrq->cluster.buf);
	35	}
	36
	37	int __scm_alloc_rq_cluster(struct scm_request *scmrq)
	38	{
	39	int i;
	40
	41	scmrq->cluster.buf = kzalloc(sizeof(void ) 2 * write_cluster_size,
	42	GFP_KERNEL);
	43	if (!scmrq->cluster.buf)
	44	return -ENOMEM;
	45
	46	for (i = 0; i < 2 * write_cluster_size; i++) {
	47	scmrq->cluster.buf[i] = (void *) get_zeroed_page(GFP_DMA);
	48	if (!scmrq->cluster.buf[i])
	49	return -ENOMEM;
	50	}
	51	INIT_LIST_HEAD(&scmrq->cluster.list);
	52	return 0;
	53	}
	54
	55	void scm_request_cluster_init(struct scm_request *scmrq)
	56	{
	57	scmrq->cluster.state = CLUSTER_NONE;
	58	}
	59
bbc610a9	60	static bool clusters_intersect(struct request A, struct request B)
0d804b20 SO	61	{
	62	unsigned long firstA, lastA, firstB, lastB;
	63
bbc610a9 SO	64	firstA = ((u64) blk_rq_pos(A) << 9) / CLUSTER_SIZE;
	65	lastA = (((u64) blk_rq_pos(A) << 9) +
	66	blk_rq_bytes(A) - 1) / CLUSTER_SIZE;
0d804b20	67
bbc610a9 SO	68	firstB = ((u64) blk_rq_pos(B) << 9) / CLUSTER_SIZE;
	69	lastB = (((u64) blk_rq_pos(B) << 9) +
	70	blk_rq_bytes(B) - 1) / CLUSTER_SIZE;
0d804b20 SO	71
	72	return (firstB <= lastA && firstA <= lastB);
	73	}
	74
	75	bool scm_reserve_cluster(struct scm_request *scmrq)
	76	{
bbc610a9	77	struct request *req = scmrq->request[scmrq->aob->request.msb_count];
0d804b20 SO	78	struct scm_blk_dev *bdev = scmrq->bdev;
0d804b20 SO	79	struct scm_request *iter;
bbc610a9	80	int pos, add = 1;
0d804b20 SO	81
	82	if (write_cluster_size == 0)
	83	return true;
	84
	85	spin_lock(&bdev->lock);
	86	list_for_each_entry(iter, &bdev->cluster_list, cluster.list) {
bbc610a9 SO	87	if (iter == scmrq) {
	88	/*
	89	* We don't have to use clusters_intersect here, since
	90	* cluster requests are always started separately.
	91	*/
	92	add = 0;
	93	continue;
	94	}
a13ccb04	95	for (pos = 0; pos < iter->aob->request.msb_count; pos++) {
bbc610a9 SO	96	if (clusters_intersect(req, iter->request[pos]) &&
	97	(rq_data_dir(req) == WRITE \|\|
	98	rq_data_dir(iter->request[pos]) == WRITE)) {
	99	spin_unlock(&bdev->lock);
	100	return false;
	101	}
0d804b20 SO	102	}
0d804b20 SO	103	}
bbc610a9 SO	104	if (add)
bbc610a9 SO	105	list_add(&scmrq->cluster.list, &bdev->cluster_list);
0d804b20 SO	106	spin_unlock(&bdev->lock);
	107
	108	return true;
	109	}
	110
	111	void scm_release_cluster(struct scm_request *scmrq)
	112	{
	113	struct scm_blk_dev *bdev = scmrq->bdev;
	114	unsigned long flags;
	115
	116	if (write_cluster_size == 0)
	117	return;
	118
	119	spin_lock_irqsave(&bdev->lock, flags);
	120	list_del(&scmrq->cluster.list);
	121	spin_unlock_irqrestore(&bdev->lock, flags);
	122	}
	123
	124	void scm_blk_dev_cluster_setup(struct scm_blk_dev *bdev)
	125	{
	126	INIT_LIST_HEAD(&bdev->cluster_list);
	127	blk_queue_io_opt(bdev->rq, CLUSTER_SIZE);
	128	}
	129
9d4df77f	130	static int scm_prepare_cluster_request(struct scm_request *scmrq)
0d804b20 SO	131	{
	132	struct scm_blk_dev *bdev = scmrq->bdev;
	133	struct scm_device *scmdev = bdev->gendisk->private_data;
bbc610a9	134	struct request *req = scmrq->request[0];
0d804b20 SO	135	struct msb *msb = &scmrq->aob->msb[0];
0d804b20 SO	136	struct req_iterator iter;
9d4df77f	137	struct aidaw *aidaw;
7988613b	138	struct bio_vec bv;
0d804b20 SO	139	int i = 0;
	140	u64 addr;
	141
	142	switch (scmrq->cluster.state) {
	143	case CLUSTER_NONE:
	144	scmrq->cluster.state = CLUSTER_READ;
	145	/* fall through */
	146	case CLUSTER_READ:
0d804b20 SO	147	msb->bs = MSB_BS_4K;
	148	msb->oc = MSB_OC_READ;
	149	msb->flags = MSB_FLAG_IDA;
0d804b20 SO	150	msb->blk_count = write_cluster_size;
	151
	152	addr = scmdev->address + ((u64) blk_rq_pos(req) << 9);
	153	msb->scm_addr = round_down(addr, CLUSTER_SIZE);
	154
	155	if (msb->scm_addr !=
	156	round_down(addr + (u64) blk_rq_bytes(req) - 1,
	157	CLUSTER_SIZE))
	158	msb->blk_count = 2 * write_cluster_size;
	159
de88d0d2 SO	160	aidaw = scm_aidaw_fetch(scmrq, msb->blk_count * PAGE_SIZE);
	161	if (!aidaw)
	162	return -ENOMEM;
	163
	164	scmrq->aob->request.msb_count = 1;
	165	msb->data_addr = (u64) aidaw;
0d804b20 SO	166	for (i = 0; i < msb->blk_count; i++) {
	167	aidaw->data_addr = (u64) scmrq->cluster.buf[i];
	168	aidaw++;
	169	}
	170
	171	break;
	172	case CLUSTER_WRITE:
9d4df77f	173	aidaw = (void *) msb->data_addr;
0d804b20 SO	174	msb->oc = MSB_OC_WRITE;
	175
	176	for (addr = msb->scm_addr;
	177	addr < scmdev->address + ((u64) blk_rq_pos(req) << 9);
	178	addr += PAGE_SIZE) {
	179	aidaw->data_addr = (u64) scmrq->cluster.buf[i];
	180	aidaw++;
	181	i++;
	182	}
	183	rq_for_each_segment(bv, req, iter) {
7988613b	184	aidaw->data_addr = (u64) page_address(bv.bv_page);
0d804b20 SO	185	aidaw++;
	186	i++;
	187	}
	188	for (; i < msb->blk_count; i++) {
	189	aidaw->data_addr = (u64) scmrq->cluster.buf[i];
	190	aidaw++;
	191	}
	192	break;
	193	}
9d4df77f	194	return 0;
0d804b20 SO	195	}
	196
	197	bool scm_need_cluster_request(struct scm_request *scmrq)
	198	{
bbc610a9 SO	199	int pos = scmrq->aob->request.msb_count;
	200
	201	if (rq_data_dir(scmrq->request[pos]) == READ)
0d804b20 SO	202	return false;
0d804b20 SO	203
bbc610a9	204	return blk_rq_bytes(scmrq->request[pos]) < CLUSTER_SIZE;
0d804b20 SO	205	}
	206
	207	/* Called with queue lock held. */
	208	void scm_initiate_cluster_request(struct scm_request *scmrq)
	209	{
9d4df77f SO	210	if (scm_prepare_cluster_request(scmrq))
9d4df77f SO	211	goto requeue;
605c3698	212	if (eadm_start_aob(scmrq->aob))
9d4df77f SO	213	goto requeue;
	214	return;
	215	requeue:
	216	scm_request_requeue(scmrq);
0d804b20 SO	217	}
	218
	219	bool scm_test_cluster_request(struct scm_request *scmrq)
	220	{
	221	return scmrq->cluster.state != CLUSTER_NONE;
	222	}
	223
	224	void scm_cluster_request_irq(struct scm_request *scmrq)
	225	{
	226	struct scm_blk_dev *bdev = scmrq->bdev;
	227	unsigned long flags;
	228
	229	switch (scmrq->cluster.state) {
	230	case CLUSTER_NONE:
	231	BUG();
	232	break;
	233	case CLUSTER_READ:
	234	if (scmrq->error) {
	235	scm_request_finish(scmrq);
	236	break;
	237	}
	238	scmrq->cluster.state = CLUSTER_WRITE;
	239	spin_lock_irqsave(&bdev->rq_lock, flags);
	240	scm_initiate_cluster_request(scmrq);
	241	spin_unlock_irqrestore(&bdev->rq_lock, flags);
	242	break;
	243	case CLUSTER_WRITE:
	244	scm_request_finish(scmrq);
	245	break;
	246	}
	247	}
	248
	249	bool scm_cluster_size_valid(void)
	250	{
bd86055f SO	251	if (write_cluster_size == 1 \|\| write_cluster_size > 128)
	252	return false;
	253
	254	return !(write_cluster_size & (write_cluster_size - 1));
0d804b20	255	}