On Tue, 2010-09-28 at 08:46 +0200, Boaz Harrosh wrote:
> On 09/28/2010 12:51 AM, Nicholas A. Bellinger wrote:
> > From: Nicholas Bellinger <nab@xxxxxxxxxxxxxxx>
> >
> > This patch adds UNMAP emulation support using Block layer DISCARD in fd_emulate_scsi_cdb()
> > -> transport_generic_unmap(). This includes the use of blk_queue_discard() in
> > fd_create_virtdevice() to determine when to set DEV_ATTRIB(dev)->emulate_tpe=1 to
> > signal to TCM Core to perform the necessary control CDB emulation for TPE=1 / UNMAP.
> >
> > Note this patch igrab() + iput() when accessing struct block_device via
> > I_BDEV(file->f_mapping->host) that is required for transport_generic_unmap()
> > within fd_emulate_scsi_cdb().
> >
> > Also note that when (S_ISBLK(inode->i_mode) is NULL, the TPE=1 and UNMAP emulation
> > are disabled for non BD struct file I/O.
> >
> > Signed-off-by: Nicholas A. Bellinger <nab@xxxxxxxxxxxxxxx>
> > ---
> > drivers/target/target_core_file.c | 53 ++++++++++++++++++++++++++++++++++++-
> > 1 files changed, 52 insertions(+), 1 deletions(-)
> >
> > diff --git a/drivers/target/target_core_file.c b/drivers/target/target_core_file.c
> > index 93d3f63..3b91556 100644
> > --- a/drivers/target/target_core_file.c
> > +++ b/drivers/target/target_core_file.c
> > @@ -235,6 +235,17 @@ static struct se_device *fd_create_virtdevice(
> >
> > fd_dev->fd_dev_id = fd_host->fd_host_dev_id_count++;
> > fd_dev->fd_queue_depth = dev->queue_depth;
> > + /*
> > + * Check for QUEUE_FLAG_DISCARD and enable TCM TPE emulation
> > + * if present for struct block_device backend
> > + */
> > + if (S_ISBLK(inode->i_mode)) {
> > + if (blk_queue_discard(bdev_get_queue(inode->i_bdev))) {
> > + DEV_ATTRIB(dev)->emulate_tpe = 1;
> > + printk(KERN_INFO "FILEIO: Enabling BLOCK Discard"
> > + " and TPE=1 emulation\n");
> > + }
> > + }
> >
> > printk(KERN_INFO "CORE_FILE[%u] - Added TCM FILEIO Device ID: %u at %s,"
> > " %llu total bytes\n", fd_host->fd_host_id, fd_dev->fd_dev_id,
> > @@ -388,9 +399,13 @@ static int fd_emulate_read_cap16(struct se_task *task)
> > */
> > static int fd_emulate_scsi_cdb(struct se_task *task)
> > {
> > - int ret;
> > struct se_cmd *cmd = TASK_CMD(task);
> > + struct fd_dev *fd_dev = (struct fd_dev *) task->se_dev->dev_ptr;
>
> No need to typecast a void pointer
Whoops, this was a cut and paste but these unnecessary typecasts for
'struct sub_dev' and 'struct sub_task' have been around forever and can
go. Removing these in target_core_file.c, target_core_iblock,
target_core_pscsi.c target_core_rd.c, and target_core_stgt.c now.
>
> > struct fd_request *fd_req = (struct fd_request *) task->transport_req;
> > + struct file *f = fd_dev->fd_file;
> > + struct inode *i;
> > + struct block_device *bd;
> > + int ret;
> >
> > switch (fd_req->fd_scsi_cdb[0]) {
> > case INQUIRY:
> > @@ -433,6 +448,42 @@ static int fd_emulate_scsi_cdb(struct se_task *task)
> > if (ret < 0)
> > return ret;
> > break;
> > + case UNMAP:
> > + i = igrab(f->f_mapping->host);
> > + if (!(i)) {
>
> You must have a reference on your fileio backend file, right?
> This could be a BUG_ON()
Yes, this was an extra bit of paranoia as we always expect to have a
valid struct file here.
>
> > + printk(KERN_ERR "FILEIO: Unable to locate inode for"
> > + " backend for UNMAP\n");
> > + return PYX_TRANSPORT_LU_COMM_FAILURE;
> > + }
> > + /*
> > + * Currently for struct file w/o a struct block_device
> > + * backend we return a success..
> > + */
> > + if (!(S_ISBLK(i->i_mode))) {
>
> I did not see in the first patch that you do this check when actually sending
> the VPD page. I only saw that it is enabled if set from configfs. (Sorry for
> the noise if I missed it). So you might want to call a target specific function
> when sending the VPD page. This way this will never happen right?
That is correct, once we determine that fd_create_virtdevice() ->
blk_queue_discard() is going to set DEV_ATTRIB(dev)->emulate_tpe=1, this
is what will signal TPE=1 in SA READ_CAPACITY_16 and other locations in
Patch 1 for the emulated case with TCM/IBLOCK and TCM/FILEIO.
>
> > + printk(KERN_WARNING "Ignoring UNMAP for non BD"
> > + " backend for struct file\n");
>
> TODO: Lots of extent based filesystems could enjoy if you punch an hole
> in the file. (Which will also send a proper discard).
> So how to punch an hole in a file via vfs?
Good question, as all of the current block discard support assumes a
struct block_device * pointer currently.
I know that hch has mentioned that accesing a struct block_device from
struct file is very dangerous, but I am assuming for the TCM/FILEIO case
that this code will be protected by igrab() and iput().
hch and jaxboe, any comments here..?
Thanks!
--nab
>
> > + iput(i);
> > + return PYX_TRANSPORT_LU_COMM_FAILURE;
> > + }
> > + bd = I_BDEV(f->f_mapping->host);
> > + if (!(bd)) {
> > + printk(KERN_ERR "FILEIO: Unable to locate struct"
> > + " block_device for UNMAP\n");
> > + iput(i);
> > + return PYX_TRANSPORT_LU_COMM_FAILURE;
> > + }
> > + /*
> > + * Now call the transport_generic_unmap() -> blkdev_issue_discard()
> > + * wrapper to translate SCSI UNMAP into Linux/BLOCK discards on
> > + * LBA+Range descriptors in the UNMAP write paylaod.
> > + */
> > + ret = transport_generic_unmap(cmd, bd);
> > + if (ret < 0) {
> > + iput(i);
> > + return ret;
> > + }
> > + iput(i);
> > + break;
> > case ALLOW_MEDIUM_REMOVAL:
> > case ERASE:
> > case REZERO_UNIT:
>
> Boaz
>
--
To unsubscribe from this list: 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
