target: stop splitting commands into multiple tasks
authorChristoph Hellwig <hch@infradead.org>
Mon, 26 Mar 2012 21:53:17 +0000 (17:53 -0400)
committerNicholas Bellinger <nab@linux-iscsi.org>
Sun, 15 Apr 2012 00:40:31 +0000 (17:40 -0700)
From hch:

The high-performance backends (iblock and rd) support tasks of unlimited
size.  With that there is no reason to keep a complex infrastructure for
splitting up commands in place.  Stop doing so and only submit a single
task per data direction.  Once this is in place we can slowly remove fields
from the task that duplicate things in the command, or move other fields
into the command.

From nab:

The benefit to IBLOCK performance by removing the additional
fast-path allocation overhead + SGL mapping to se_task->task_sg[] is now
greater than transparently supporting an received CDB I/O length that
exceeds what is allowed by backend pSCSI LLD hardware max_sectors, that
was originally supported for all backend export cases.

This change may effect some users of pSCSI users on legacy hardware, but
I think most folks are now using TYPE_DISK struct scsi_device export
with IBLOCK.  The only other place where this may can issues that cannot
be resolved with IBLOCK TYPE_DISK is using TYPE_ROM, TYPE_TAPE or other
pSCSI non TYPE_DISK export with an SCSI LLDs using a smaller
max_sectors.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
drivers/target/target_core_cdb.c
drivers/target/target_core_transport.c

index 30a6770..f6f6059 100644 (file)
@@ -432,6 +432,7 @@ static int
 target_emulate_evpd_b0(struct se_cmd *cmd, unsigned char *buf)
 {
        struct se_device *dev = cmd->se_dev;
+       u32 max_sectors;
        int have_tp = 0;
 
        /*
@@ -456,7 +457,9 @@ target_emulate_evpd_b0(struct se_cmd *cmd, unsigned char *buf)
        /*
         * Set MAXIMUM TRANSFER LENGTH
         */
-       put_unaligned_be32(dev->se_sub_dev->se_dev_attrib.fabric_max_sectors, &buf[8]);
+       max_sectors = min(dev->se_sub_dev->se_dev_attrib.fabric_max_sectors,
+                         dev->se_sub_dev->se_dev_attrib.max_sectors);
+       put_unaligned_be32(max_sectors, &buf[8]);
 
        /*
         * Set OPTIMAL TRANSFER LENGTH
index 25c67c8..db05c75 100644 (file)
@@ -3183,7 +3183,8 @@ static int transport_generic_cmd_sequencer(
        }
 
        if (cmd->se_cmd_flags & SCF_SCSI_DATA_SG_IO_CDB &&
-           sectors > dev->se_sub_dev->se_dev_attrib.fabric_max_sectors) {
+           (sectors > dev->se_sub_dev->se_dev_attrib.fabric_max_sectors ||
+            sectors > dev->se_sub_dev->se_dev_attrib.max_sectors)) {
                printk_ratelimited(KERN_ERR "SCSI OP %02xh with too big sectors %u\n",
                                   cdb[0], sectors);
                goto out_invalid_cdb_field;
@@ -3446,12 +3447,7 @@ static void transport_free_dev_tasks(struct se_cmd *cmd)
        while (!list_empty(&dispose_list)) {
                task = list_first_entry(&dispose_list, struct se_task, t_list);
 
-               if (task->task_sg != cmd->t_data_sg &&
-                   task->task_sg != cmd->t_bidi_data_sg)
-                       kfree(task->task_sg);
-
                list_del(&task->t_list);
-
                cmd->se_dev->transport->free_task(task);
        }
 }
@@ -3782,112 +3778,35 @@ transport_allocate_data_tasks(struct se_cmd *cmd,
        struct scatterlist *cmd_sg, unsigned int sgl_nents)
 {
        struct se_device *dev = cmd->se_dev;
-       int task_count, i;
-       unsigned long long lba;
-       sector_t sectors, dev_max_sectors;
-       u32 sector_size;
+       struct se_dev_attrib *attr = &dev->se_sub_dev->se_dev_attrib;
+       sector_t sectors;
+       struct se_task *task;
+       unsigned long flags;
 
        if (transport_cmd_get_valid_sectors(cmd) < 0)
                return -EINVAL;
 
-       dev_max_sectors = dev->se_sub_dev->se_dev_attrib.max_sectors;
-       sector_size = dev->se_sub_dev->se_dev_attrib.block_size;
-
-       WARN_ON(cmd->data_length % sector_size);
-
-       lba = cmd->t_task_lba;
-       sectors = DIV_ROUND_UP(cmd->data_length, sector_size);
-       task_count = DIV_ROUND_UP_SECTOR_T(sectors, dev_max_sectors);
-
-       /*
-        * If we need just a single task reuse the SG list in the command
-        * and avoid a lot of work.
-        */
-       if (task_count == 1) {
-               struct se_task *task;
-               unsigned long flags;
-
-               task = transport_generic_get_task(cmd, data_direction);
-               if (!task)
-                       return -ENOMEM;
-
-               task->task_sg = cmd_sg;
-               task->task_sg_nents = sgl_nents;
-
-               task->task_lba = lba;
-               task->task_sectors = sectors;
-               task->task_size = task->task_sectors * sector_size;
-
-               spin_lock_irqsave(&cmd->t_state_lock, flags);
-               list_add_tail(&task->t_list, &cmd->t_task_list);
-               spin_unlock_irqrestore(&cmd->t_state_lock, flags);
-
-               return task_count;
-       }
-
-       for (i = 0; i < task_count; i++) {
-               struct se_task *task;
-               unsigned int task_size, task_sg_nents_padded;
-               struct scatterlist *sg;
-               unsigned long flags;
-               int count;
-
-               task = transport_generic_get_task(cmd, data_direction);
-               if (!task)
-                       return -ENOMEM;
-
-               task->task_lba = lba;
-               task->task_sectors = min(sectors, dev_max_sectors);
-               task->task_size = task->task_sectors * sector_size;
+       sectors = DIV_ROUND_UP(cmd->data_length, attr->block_size);
 
-               /*
-                * This now assumes that passed sg_ents are in PAGE_SIZE chunks
-                * in order to calculate the number per task SGL entries
-                */
-               task->task_sg_nents = DIV_ROUND_UP(task->task_size, PAGE_SIZE);
-               /*
-                * Check if the fabric module driver is requesting that all
-                * struct se_task->task_sg[] be chained together..  If so,
-                * then allocate an extra padding SG entry for linking and
-                * marking the end of the chained SGL for every task except
-                * the last one for (task_count > 1) operation, or skipping
-                * the extra padding for the (task_count == 1) case.
-                */
-               if (cmd->se_tfo->task_sg_chaining && (i < (task_count - 1))) {
-                       task_sg_nents_padded = (task->task_sg_nents + 1);
-               } else
-                       task_sg_nents_padded = task->task_sg_nents;
-
-               task->task_sg = kmalloc(sizeof(struct scatterlist) *
-                                       task_sg_nents_padded, GFP_KERNEL);
-               if (!task->task_sg) {
-                       cmd->se_dev->transport->free_task(task);
-                       return -ENOMEM;
-               }
-
-               sg_init_table(task->task_sg, task_sg_nents_padded);
-
-               task_size = task->task_size;
+       BUG_ON(cmd->data_length % attr->block_size);
+       BUG_ON(sectors > attr->max_sectors);
 
-               /* Build new sgl, only up to task_size */
-               for_each_sg(task->task_sg, sg, task->task_sg_nents, count) {
-                       if (cmd_sg->length > task_size)
-                               break;
+       task = transport_generic_get_task(cmd, data_direction);
+       if (!task)
+               return -ENOMEM;
 
-                       *sg = *cmd_sg;
-                       task_size -= cmd_sg->length;
-                       cmd_sg = sg_next(cmd_sg);
-               }
+       task->task_sg = cmd_sg;
+       task->task_sg_nents = sgl_nents;
+       task->task_size = cmd->data_length;
 
-               lba += task->task_sectors;
-               sectors -= task->task_sectors;
+       task->task_lba = cmd->t_task_lba;
+       task->task_sectors = sectors;
 
-               spin_lock_irqsave(&cmd->t_state_lock, flags);
-               list_add_tail(&task->t_list, &cmd->t_task_list);
-               spin_unlock_irqrestore(&cmd->t_state_lock, flags);
-       }
+       spin_lock_irqsave(&cmd->t_state_lock, flags);
+       list_add_tail(&task->t_list, &cmd->t_task_list);
+       spin_unlock_irqrestore(&cmd->t_state_lock, flags);
 
-       return task_count;
+       return 1;
 }
 
 static int