2 * Block driver for s390 storage class memory.
4 * Copyright IBM Corp. 2012
5 * Author(s): Sebastian Ott <sebott@linux.vnet.ibm.com>
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>
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.");
22 #define CLUSTER_SIZE (write_cluster_size * PAGE_SIZE)
24 void __scm_free_rq_cluster(struct scm_request *scmrq)
28 if (!scmrq->cluster.buf)
31 for (i = 0; i < 2 * write_cluster_size; i++)
32 free_page((unsigned long) scmrq->cluster.buf[i]);
34 kfree(scmrq->cluster.buf);
37 int __scm_alloc_rq_cluster(struct scm_request *scmrq)
41 scmrq->cluster.buf = kzalloc(sizeof(void *) * 2 * write_cluster_size,
43 if (!scmrq->cluster.buf)
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])
51 INIT_LIST_HEAD(&scmrq->cluster.list);
55 void scm_request_cluster_init(struct scm_request *scmrq)
57 scmrq->cluster.state = CLUSTER_NONE;
60 static bool clusters_intersect(struct request *A, struct request *B)
62 unsigned long firstA, lastA, firstB, lastB;
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;
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;
72 return (firstB <= lastA && firstA <= lastB);
75 bool scm_reserve_cluster(struct scm_request *scmrq)
77 struct request *req = scmrq->request[scmrq->aob->request.msb_count];
78 struct scm_blk_dev *bdev = scmrq->bdev;
79 struct scm_request *iter;
82 if (write_cluster_size == 0)
85 spin_lock(&bdev->lock);
86 list_for_each_entry(iter, &bdev->cluster_list, cluster.list) {
89 * We don't have to use clusters_intersect here, since
90 * cluster requests are always started separately.
95 for (pos = 0; pos < iter->aob->request.msb_count; pos++) {
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);
105 list_add(&scmrq->cluster.list, &bdev->cluster_list);
106 spin_unlock(&bdev->lock);
111 void scm_release_cluster(struct scm_request *scmrq)
113 struct scm_blk_dev *bdev = scmrq->bdev;
116 if (write_cluster_size == 0)
119 spin_lock_irqsave(&bdev->lock, flags);
120 list_del(&scmrq->cluster.list);
121 spin_unlock_irqrestore(&bdev->lock, flags);
124 void scm_blk_dev_cluster_setup(struct scm_blk_dev *bdev)
126 INIT_LIST_HEAD(&bdev->cluster_list);
127 blk_queue_io_opt(bdev->rq, CLUSTER_SIZE);
130 static int scm_prepare_cluster_request(struct scm_request *scmrq)
132 struct scm_blk_dev *bdev = scmrq->bdev;
133 struct scm_device *scmdev = bdev->gendisk->private_data;
134 struct request *req = scmrq->request[0];
135 struct msb *msb = &scmrq->aob->msb[0];
136 struct req_iterator iter;
142 switch (scmrq->cluster.state) {
144 scmrq->cluster.state = CLUSTER_READ;
148 msb->oc = MSB_OC_READ;
149 msb->flags = MSB_FLAG_IDA;
150 msb->blk_count = write_cluster_size;
152 addr = scmdev->address + ((u64) blk_rq_pos(req) << 9);
153 msb->scm_addr = round_down(addr, CLUSTER_SIZE);
156 round_down(addr + (u64) blk_rq_bytes(req) - 1,
158 msb->blk_count = 2 * write_cluster_size;
160 aidaw = scm_aidaw_fetch(scmrq, msb->blk_count * PAGE_SIZE);
164 scmrq->aob->request.msb_count = 1;
165 msb->data_addr = (u64) aidaw;
166 for (i = 0; i < msb->blk_count; i++) {
167 aidaw->data_addr = (u64) scmrq->cluster.buf[i];
173 aidaw = (void *) msb->data_addr;
174 msb->oc = MSB_OC_WRITE;
176 for (addr = msb->scm_addr;
177 addr < scmdev->address + ((u64) blk_rq_pos(req) << 9);
179 aidaw->data_addr = (u64) scmrq->cluster.buf[i];
183 rq_for_each_segment(bv, req, iter) {
184 aidaw->data_addr = (u64) page_address(bv.bv_page);
188 for (; i < msb->blk_count; i++) {
189 aidaw->data_addr = (u64) scmrq->cluster.buf[i];
197 bool scm_need_cluster_request(struct scm_request *scmrq)
199 int pos = scmrq->aob->request.msb_count;
201 if (rq_data_dir(scmrq->request[pos]) == READ)
204 return blk_rq_bytes(scmrq->request[pos]) < CLUSTER_SIZE;
207 /* Called with queue lock held. */
208 void scm_initiate_cluster_request(struct scm_request *scmrq)
210 if (scm_prepare_cluster_request(scmrq))
212 if (eadm_start_aob(scmrq->aob))
216 scm_request_requeue(scmrq);
219 bool scm_test_cluster_request(struct scm_request *scmrq)
221 return scmrq->cluster.state != CLUSTER_NONE;
224 void scm_cluster_request_irq(struct scm_request *scmrq)
226 struct scm_blk_dev *bdev = scmrq->bdev;
229 switch (scmrq->cluster.state) {
235 scm_request_finish(scmrq);
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);
244 scm_request_finish(scmrq);
249 bool scm_cluster_size_valid(void)
251 if (write_cluster_size == 1 || write_cluster_size > 128)
254 return !(write_cluster_size & (write_cluster_size - 1));