struct scsi_cmnd is a fairly large data structure and is used
to construct scsi command. On x86-64 arch, it is 448 bytes.
In scsi_get_command(), it is a bit too paranoid in zeroing the
data structure. Since most of the member variables will be
initialized in various stage of I/O submission, i.e., in
scsi_prep_fn, scsi_init_io, sd_init_command, scsi_init_cmd_errh,
etc. So instead of blindly zeroing the whole structure, initialize
to some other value later on. All it needs in scsi_get_command
is to zero out a few member variables that aren't initialized in
the I/O path.
I'm proposing the following patch to optimize scsi_cmnd initialization.
The only part that I'm not 100% sure is struct scsi_pointer SCp and
eh_eflags. It appears to be safe, but since I don't have every scsi
controllers in the world to test, I can't say so for sure. I'm willing
to take on the responsibility to fix any hiccup people might have with
this patch. Comments?
p.s. de-referencing jiffies_at_alloc ought be in the if(cmd!=NULL) block.
Signed-off-by: Ken Chen <kenneth.w.chen@xxxxxxxxx>
--- ./drivers/scsi/scsi.c.orig 2005-11-23 13:04:46.607198199 -0800
+++ ./drivers/scsi/scsi.c 2005-11-23 13:05:38.154072568 -0800
@@ -258,17 +258,18 @@ struct scsi_cmnd *scsi_get_command(struc
if (likely(cmd != NULL)) {
unsigned long flags;
- memset(cmd, 0, sizeof(*cmd));
+ cmd->cmd_len = 0;
+ cmd->sc_request = 0;
+ cmd->retries = 0;
cmd->device = dev;
init_timer(&cmd->eh_timeout);
- INIT_LIST_HEAD(&cmd->list);
spin_lock_irqsave(&dev->list_lock, flags);
list_add_tail(&cmd->list, &dev->cmd_list);
spin_unlock_irqrestore(&dev->list_lock, flags);
+ cmd->jiffies_at_alloc = jiffies;
} else
put_device(&dev->sdev_gendev);
- cmd->jiffies_at_alloc = jiffies;
return cmd;
}
EXPORT_SYMBOL(scsi_get_command);
-
: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
