s390/scm_block: force cluster writes

Force writes to Storage Class Memory (SCM) to be in done in clusters.

Signed-off-by: Sebastian Ott <sebott@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

5 files changed, 310 insertions, 3 deletions

diff --git a/drivers/s390/block/Kconfig b/drivers/s390/block/Kconfig
index 18178b64e89a..4a3b62326183 100644
--- a/drivers/s390/block/Kconfig
+++ b/drivers/s390/block/Kconfig
@@ -81,3 +81,10 @@ config SCM_BLOCK
 
 	  To compile this driver as a module, choose M here: the
 	  module will be called scm_block.
+
+config SCM_BLOCK_CLUSTER_WRITE
+	def_bool y
+	prompt "SCM force cluster writes"
+	depends on SCM_BLOCK
+	help
+	  Force writes to Storage Class Memory (SCM) to be in done in clusters.
diff --git a/drivers/s390/block/Makefile b/drivers/s390/block/Makefile
index b64e2b32c753..c2f4e673e031 100644
--- a/drivers/s390/block/Makefile
+++ b/drivers/s390/block/Makefile
@@ -19,4 +19,7 @@ obj-$(CONFIG_BLK_DEV_XPRAM) += xpram.o
 obj-$(CONFIG_DCSSBLK) += dcssblk.o
 
 scm_block-objs := scm_drv.o scm_blk.o
+ifdef CONFIG_SCM_BLOCK_CLUSTER_WRITE
+scm_block-objs += scm_blk_cluster.o
+endif
 obj-$(CONFIG_SCM_BLOCK) += scm_block.o
diff --git a/drivers/s390/block/scm_blk.c b/drivers/s390/block/scm_blk.c
index 634ad58cbef6..9978ad4433cb 100644
--- a/drivers/s390/block/scm_blk.c
+++ b/drivers/s390/block/scm_blk.c
@@ -37,6 +37,7 @@ static void __scm_free_rq(struct scm_request *scmrq)
 
 	free_page((unsigned long) scmrq->aob);
 	free_page((unsigned long) scmrq->aidaw);
+	__scm_free_rq_cluster(scmrq);
 	kfree(aobrq);
 }
 
@@ -70,6 +71,12 @@ static int __scm_alloc_rq(void)
 		__scm_free_rq(scmrq);
 		return -ENOMEM;
 	}
+
+	if (__scm_alloc_rq_cluster(scmrq)) {
+		__scm_free_rq(scmrq);
+		return -ENOMEM;
+	}
+
 	INIT_LIST_HEAD(&scmrq->list);
 	spin_lock_irq(&list_lock);
 	list_add(&scmrq->list, &inactive_requests);
@@ -170,6 +177,7 @@ static inline void scm_request_init(struct scm_blk_dev *bdev,
 	scmrq->bdev = bdev;
 	scmrq->retries = 4;
 	scmrq->error = 0;
+	scm_request_cluster_init(scmrq);
 }
 
 static void scm_ensure_queue_restart(struct scm_blk_dev *bdev)
@@ -181,17 +189,19 @@ static void scm_ensure_queue_restart(struct scm_blk_dev *bdev)
 	blk_delay_queue(bdev->rq, SCM_QUEUE_DELAY);
 }
 
-static void scm_request_requeue(struct scm_request *scmrq)
+void scm_request_requeue(struct scm_request *scmrq)
 {
 	struct scm_blk_dev *bdev = scmrq->bdev;
 
+	scm_release_cluster(scmrq);
 	blk_requeue_request(bdev->rq, scmrq->request);
 	scm_request_done(scmrq);
 	scm_ensure_queue_restart(bdev);
 }
 
-static void scm_request_finish(struct scm_request *scmrq)
+void scm_request_finish(struct scm_request *scmrq)
 {
+	scm_release_cluster(scmrq);
 	blk_end_request_all(scmrq->request, scmrq->error);
 	scm_request_done(scmrq);
 }
@@ -215,6 +225,16 @@ static void scm_blk_request(struct request_queue *rq)
 			return;
 		}
 		scm_request_init(bdev, scmrq, req);
+		if (!scm_reserve_cluster(scmrq)) {
+			SCM_LOG(5, "cluster busy");
+			scm_request_done(scmrq);
+			return;
+		}
+		if (scm_need_cluster_request(scmrq)) {
+			blk_start_request(req);
+			scm_initiate_cluster_request(scmrq);
+			return;
+		}
 		scm_request_prepare(scmrq);
 		blk_start_request(req);
 
@@ -282,6 +302,13 @@ static void scm_blk_tasklet(struct scm_blk_dev *bdev)
 			spin_lock_irqsave(&bdev->lock, flags);
 			continue;
 		}
+
+		if (scm_test_cluster_request(scmrq)) {
+			scm_cluster_request_irq(scmrq);
+			spin_lock_irqsave(&bdev->lock, flags);
+			continue;
+		}
+
 		scm_request_finish(scmrq);
 		atomic_dec(&bdev->queued_reqs);
 		spin_lock_irqsave(&bdev->lock, flags);
@@ -325,6 +352,7 @@ int scm_blk_dev_setup(struct scm_blk_dev *bdev, struct scm_device *scmdev)
 	blk_queue_max_hw_sectors(rq, nr_max_blk << 3); /* 8 * 512 = blk_size */
 	blk_queue_max_segments(rq, nr_max_blk);
 	queue_flag_set_unlocked(QUEUE_FLAG_NONROT, rq);
+	scm_blk_dev_cluster_setup(bdev);
 
 	bdev->gendisk = alloc_disk(SCM_NR_PARTS);
 	if (!bdev->gendisk)
@@ -370,7 +398,10 @@ void scm_blk_dev_cleanup(struct scm_blk_dev *bdev)
 
 static int __init scm_blk_init(void)
 {
-	int ret;
+	int ret = -EINVAL;
+
+	if (!scm_cluster_size_valid())
+		goto out;
 
 	ret = register_blkdev(0, "scm");
 	if (ret < 0)
diff --git a/drivers/s390/block/scm_blk.h b/drivers/s390/block/scm_blk.h
index 5aba5612588f..7ac6bad919ef 100644
--- a/drivers/s390/block/scm_blk.h
+++ b/drivers/s390/block/scm_blk.h
@@ -22,6 +22,9 @@ struct scm_blk_dev {
 	spinlock_t lock;	/* guard the rest of the blockdev */
 	atomic_t queued_reqs;
 	struct list_head finished_requests;
+#ifdef CONFIG_SCM_BLOCK_CLUSTER_WRITE
+	struct list_head cluster_list;
+#endif
 };
 
 struct scm_request {
@@ -32,6 +35,13 @@ struct scm_request {
 	struct list_head list;
 	u8 retries;
 	int error;
+#ifdef CONFIG_SCM_BLOCK_CLUSTER_WRITE
+	struct {
+		enum {CLUSTER_NONE, CLUSTER_READ, CLUSTER_WRITE} state;
+		struct list_head list;
+		void **buf;
+	} cluster;
+#endif
 };
 
 #define to_aobrq(rq) container_of((void *) rq, struct aob_rq_header, data)
@@ -40,9 +50,37 @@ int scm_blk_dev_setup(struct scm_blk_dev *, struct scm_device *);
 void scm_blk_dev_cleanup(struct scm_blk_dev *);
 void scm_blk_irq(struct scm_device *, void *, int);
 
+void scm_request_finish(struct scm_request *);
+void scm_request_requeue(struct scm_request *);
+
 int scm_drv_init(void);
 void scm_drv_cleanup(void);
 
+#ifdef CONFIG_SCM_BLOCK_CLUSTER_WRITE
+void __scm_free_rq_cluster(struct scm_request *);
+int __scm_alloc_rq_cluster(struct scm_request *);
+void scm_request_cluster_init(struct scm_request *);
+bool scm_reserve_cluster(struct scm_request *);
+void scm_release_cluster(struct scm_request *);
+void scm_blk_dev_cluster_setup(struct scm_blk_dev *);
+bool scm_need_cluster_request(struct scm_request *);
+void scm_initiate_cluster_request(struct scm_request *);
+void scm_cluster_request_irq(struct scm_request *);
+bool scm_test_cluster_request(struct scm_request *);
+bool scm_cluster_size_valid(void);
+#else
+#define __scm_free_rq_cluster(scmrq) {}
+#define __scm_alloc_rq_cluster(scmrq) 0
+#define scm_request_cluster_init(scmrq) {}
+#define scm_reserve_cluster(scmrq) true
+#define scm_release_cluster(scmrq) {}
+#define scm_blk_dev_cluster_setup(bdev) {}
+#define scm_need_cluster_request(scmrq) false
+#define scm_initiate_cluster_request(scmrq) {}
+#define scm_cluster_request_irq(scmrq) {}
+#define scm_test_cluster_request(scmrq) false
+#define scm_cluster_size_valid() true
+#endif
 
 extern debug_info_t *scm_debug;
 
diff --git a/drivers/s390/block/scm_blk_cluster.c b/drivers/s390/block/scm_blk_cluster.c
new file mode 100644
index 000000000000..f4bb61b0cea1
--- /dev/null
+++ b/drivers/s390/block/scm_blk_cluster.c
@@ -0,0 +1,228 @@
+/*
+ * 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 void 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 aidaw *aidaw = scmrq->aidaw;
+	struct msb *msb = &scmrq->aob->msb[0];
+	struct req_iterator iter;
+	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:
+		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:
+		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;
+	}
+}
+
+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)
+{
+	scm_prepare_cluster_request(scmrq);
+	if (scm_start_aob(scmrq->aob))
+		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)
+{
+	return write_cluster_size == 0 || write_cluster_size == 32 ||
+		write_cluster_size == 64 || write_cluster_size == 128;
+}