drivers/s390/block/scm_blk_cluster.c

   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
 117 static void scm_prepare_cluster_request(struct scm_request *scmrq)
 118 {
 119         struct scm_blk_dev *bdev = scmrq->bdev;
 120         struct scm_device *scmdev = bdev->gendisk->private_data;
 121         struct request *req = scmrq->request;
 122         struct aidaw *aidaw = scmrq->aidaw;
 123         struct msb *msb = &scmrq->aob->msb[0];
 124         struct req_iterator iter;
 125         struct bio_vec bv;
 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:
 134                 scmrq->aob->request.msb_count = 1;
 135                 msb->bs = MSB_BS_4K;
 136                 msb->oc = MSB_OC_READ;
 137                 msb->flags = MSB_FLAG_IDA;
 138                 msb->data_addr = (u64) aidaw;
 139                 msb->blk_count = write_cluster_size;
 140
 141                 addr = scmdev->address + ((u64) blk_rq_pos(req) << 9);
 142                 msb->scm_addr = round_down(addr, CLUSTER_SIZE);
 143
 144                 if (msb->scm_addr !=
 145                     round_down(addr + (u64) blk_rq_bytes(req) - 1,
 146                                CLUSTER_SIZE))
 147                         msb->blk_count = 2 * write_cluster_size;
 148
 149                 for (i = 0; i < msb->blk_count; i++) {
 150                         aidaw->data_addr = (u64) scmrq->cluster.buf[i];
 151                         aidaw++;
 152                 }
 153
 154                 break;
 155         case CLUSTER_WRITE:
 156                 msb->oc = MSB_OC_WRITE;
 157
 158                 for (addr = msb->scm_addr;
 159                      addr < scmdev->address + ((u64) blk_rq_pos(req) << 9);
 160                      addr += PAGE_SIZE) {
 161                         aidaw->data_addr = (u64) scmrq->cluster.buf[i];
 162                         aidaw++;
 163                         i++;
 164                 }
 165                 rq_for_each_segment(bv, req, iter) {
 166                         aidaw->data_addr = (u64) page_address(bv.bv_page);
 167                         aidaw++;
 168                         i++;
 169                 }
 170                 for (; i < msb->blk_count; i++) {
 171                         aidaw->data_addr = (u64) scmrq->cluster.buf[i];
 172                         aidaw++;
 173                 }
 174                 break;
 175         }
 176 }
 177
 178 bool scm_need_cluster_request(struct scm_request *scmrq)
 179 {
 180         if (rq_data_dir(scmrq->request) == READ)
 181                 return false;
 182
 183         return blk_rq_bytes(scmrq->request) < CLUSTER_SIZE;
 184 }
 185
 186 /* Called with queue lock held. */
 187 void scm_initiate_cluster_request(struct scm_request *scmrq)
 188 {
 189         scm_prepare_cluster_request(scmrq);
 190         if (eadm_start_aob(scmrq->aob))
 191                 scm_request_requeue(scmrq);
 192 }
 193
 194 bool scm_test_cluster_request(struct scm_request *scmrq)
 195 {
 196         return scmrq->cluster.state != CLUSTER_NONE;
 197 }
 198
 199 void scm_cluster_request_irq(struct scm_request *scmrq)
 200 {
 201         struct scm_blk_dev *bdev = scmrq->bdev;
 202         unsigned long flags;
 203
 204         switch (scmrq->cluster.state) {
 205         case CLUSTER_NONE:
 206                 BUG();
 207                 break;
 208         case CLUSTER_READ:
 209                 if (scmrq->error) {
 210                         scm_request_finish(scmrq);
 211                         break;
 212                 }
 213                 scmrq->cluster.state = CLUSTER_WRITE;
 214                 spin_lock_irqsave(&bdev->rq_lock, flags);
 215                 scm_initiate_cluster_request(scmrq);
 216                 spin_unlock_irqrestore(&bdev->rq_lock, flags);
 217                 break;
 218         case CLUSTER_WRITE:
 219                 scm_request_finish(scmrq);
 220                 break;
 221         }
 222 }
 223
 224 bool scm_cluster_size_valid(void)
 225 {
 226         if (write_cluster_size == 1 || write_cluster_size > 128)
 227                 return false;
 228
 229         return !(write_cluster_size & (write_cluster_size - 1));
 230 }