[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 scm_request *A, struct scm_request *B)
{
	unsigned long firstA, lastA, firstB, lastB;

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

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

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

bool scm_reserve_cluster(struct scm_request *scmrq)
{
	struct scm_blk_dev *bdev = scmrq->bdev;
	struct scm_request *iter;

	if (write_cluster_size == 0)
		return true;

	spin_lock(&bdev->lock);
	list_for_each_entry(iter, &bdev->cluster_list, cluster.list) {
		if (clusters_intersect(scmrq, iter) &&
		    (rq_data_dir(scmrq->request) == WRITE ||
		     rq_data_dir(iter->request) == WRITE)) {
			spin_unlock(&bdev->lock);
			return false;
		}
	}
	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;
	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:
		aidaw = scm_aidaw_alloc();
		if (!aidaw)
			return -ENOMEM;

		memset(aidaw, 0, PAGE_SIZE);
		scmrq->aob->request.msb_count = 1;
		msb->bs = MSB_BS_4K;
		msb->oc = MSB_OC_READ;
		msb->flags = MSB_FLAG_IDA;
		msb->data_addr = (u64) aidaw;
		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;

		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)
{
	if (rq_data_dir(scmrq->request) == READ)
		return false;

	return blk_rq_bytes(scmrq->request) < 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
	60	static bool clusters_intersect(struct scm_request A, struct scm_request B)
	61	{
	62	unsigned long firstA, lastA, firstB, lastB;
	63
	64	firstA = ((u64) blk_rq_pos(A->request) << 9) / CLUSTER_SIZE;
65	lastA = (((u64) blk_rq_pos(A->request) << 9) +
66	blk_rq_bytes(A->request) - 1) / CLUSTER_SIZE;
67
68	firstB = ((u64) blk_rq_pos(B->request) << 9) / CLUSTER_SIZE;
69	lastB = (((u64) blk_rq_pos(B->request) << 9) +
70	blk_rq_bytes(B->request) - 1) / CLUSTER_SIZE;
71
72	return (firstB <= lastA && firstA <= lastB);
73	}
74
75	bool scm_reserve_cluster(struct scm_request *scmrq)
76	{
77	struct scm_blk_dev *bdev = scmrq->bdev;
78	struct scm_request *iter;
79
80	if (write_cluster_size == 0)
81	return true;
82
83	spin_lock(&bdev->lock);
84	list_for_each_entry(iter, &bdev->cluster_list, cluster.list) {
85	if (clusters_intersect(scmrq, iter) &&
86	(rq_data_dir(scmrq->request) == WRITE \|\|
87	rq_data_dir(iter->request) == WRITE)) {
88	spin_unlock(&bdev->lock);
89	return false;
90	}
91	}
92	list_add(&scmrq->cluster.list, &bdev->cluster_list);
93	spin_unlock(&bdev->lock);
94
95	return true;
96	}
97
98	void scm_release_cluster(struct scm_request *scmrq)
99	{
100	struct scm_blk_dev *bdev = scmrq->bdev;
101	unsigned long flags;
102
103	if (write_cluster_size == 0)
104	return;
105
106	spin_lock_irqsave(&bdev->lock, flags);
107	list_del(&scmrq->cluster.list);
108	spin_unlock_irqrestore(&bdev->lock, flags);
109	}
110
111	void scm_blk_dev_cluster_setup(struct scm_blk_dev *bdev)
112	{
113	INIT_LIST_HEAD(&bdev->cluster_list);
114	blk_queue_io_opt(bdev->rq, CLUSTER_SIZE);
115	}
116
9d4df77f	117	static int scm_prepare_cluster_request(struct scm_request *scmrq)
0d804b20 SO	118	{
	119	struct scm_blk_dev *bdev = scmrq->bdev;
	120	struct scm_device *scmdev = bdev->gendisk->private_data;
	121	struct request *req = scmrq->request;
0d804b20 SO	122	struct msb *msb = &scmrq->aob->msb[0];
0d804b20 SO	123	struct req_iterator iter;
9d4df77f	124	struct aidaw *aidaw;
7988613b	125	struct bio_vec bv;
0d804b20 SO	126	int i = 0;
	127	u64 addr;
	128
	129	switch (scmrq->cluster.state) {
	130	case CLUSTER_NONE:
	131	scmrq->cluster.state = CLUSTER_READ;
	132	/* fall through */
	133	case CLUSTER_READ:
9d4df77f SO	134	aidaw = scm_aidaw_alloc();
	135	if (!aidaw)
	136	return -ENOMEM;
	137
	138	memset(aidaw, 0, PAGE_SIZE);
0d804b20 SO	139	scmrq->aob->request.msb_count = 1;
	140	msb->bs = MSB_BS_4K;
	141	msb->oc = MSB_OC_READ;
	142	msb->flags = MSB_FLAG_IDA;
	143	msb->data_addr = (u64) aidaw;
	144	msb->blk_count = write_cluster_size;
	145
	146	addr = scmdev->address + ((u64) blk_rq_pos(req) << 9);
	147	msb->scm_addr = round_down(addr, CLUSTER_SIZE);
	148
	149	if (msb->scm_addr !=
	150	round_down(addr + (u64) blk_rq_bytes(req) - 1,
	151	CLUSTER_SIZE))
	152	msb->blk_count = 2 * write_cluster_size;
	153
	154	for (i = 0; i < msb->blk_count; i++) {
	155	aidaw->data_addr = (u64) scmrq->cluster.buf[i];
	156	aidaw++;
	157	}
	158
	159	break;
	160	case CLUSTER_WRITE:
9d4df77f	161	aidaw = (void *) msb->data_addr;
0d804b20 SO	162	msb->oc = MSB_OC_WRITE;
	163
	164	for (addr = msb->scm_addr;
	165	addr < scmdev->address + ((u64) blk_rq_pos(req) << 9);
	166	addr += PAGE_SIZE) {
	167	aidaw->data_addr = (u64) scmrq->cluster.buf[i];
	168	aidaw++;
	169	i++;
	170	}
	171	rq_for_each_segment(bv, req, iter) {
7988613b	172	aidaw->data_addr = (u64) page_address(bv.bv_page);
0d804b20 SO	173	aidaw++;
	174	i++;
	175	}
	176	for (; i < msb->blk_count; i++) {
	177	aidaw->data_addr = (u64) scmrq->cluster.buf[i];
	178	aidaw++;
	179	}
	180	break;
	181	}
9d4df77f	182	return 0;
0d804b20 SO	183	}
	184
	185	bool scm_need_cluster_request(struct scm_request *scmrq)
	186	{
	187	if (rq_data_dir(scmrq->request) == READ)
	188	return false;
	189
	190	return blk_rq_bytes(scmrq->request) < CLUSTER_SIZE;
	191	}
	192
	193	/* Called with queue lock held. */
	194	void scm_initiate_cluster_request(struct scm_request *scmrq)
	195	{
9d4df77f SO	196	if (scm_prepare_cluster_request(scmrq))
9d4df77f SO	197	goto requeue;
605c3698	198	if (eadm_start_aob(scmrq->aob))
9d4df77f SO	199	goto requeue;
	200	return;
	201	requeue:
	202	scm_request_requeue(scmrq);
0d804b20 SO	203	}
	204
	205	bool scm_test_cluster_request(struct scm_request *scmrq)
	206	{
	207	return scmrq->cluster.state != CLUSTER_NONE;
	208	}
	209
	210	void scm_cluster_request_irq(struct scm_request *scmrq)
	211	{
	212	struct scm_blk_dev *bdev = scmrq->bdev;
	213	unsigned long flags;
	214
	215	switch (scmrq->cluster.state) {
	216	case CLUSTER_NONE:
	217	BUG();
	218	break;
	219	case CLUSTER_READ:
	220	if (scmrq->error) {
	221	scm_request_finish(scmrq);
	222	break;
	223	}
	224	scmrq->cluster.state = CLUSTER_WRITE;
	225	spin_lock_irqsave(&bdev->rq_lock, flags);
	226	scm_initiate_cluster_request(scmrq);
	227	spin_unlock_irqrestore(&bdev->rq_lock, flags);
	228	break;
	229	case CLUSTER_WRITE:
	230	scm_request_finish(scmrq);
	231	break;
	232	}
	233	}
	234
	235	bool scm_cluster_size_valid(void)
	236	{
bd86055f SO	237	if (write_cluster_size == 1 \|\| write_cluster_size > 128)
	238	return false;
	239
	240	return !(write_cluster_size & (write_cluster_size - 1));
0d804b20	241	}