LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
From: Douglas Gilbert <dougg@torque.net>
To: Alan Stern <stern@rowland.harvard.edu>
Cc: Joerg Schilling <Joerg.Schilling@fokus.fraunhofer.de>,
linux-kernel@vger.kernel.org, jens.axboe@oracle.com,
James.Bottomley@steeleye.com
Subject: Re: [PATCH] Block layer: separate out queue-oriented ioctls
Date: Mon, 19 Feb 2007 17:25:11 -0500 [thread overview]
Message-ID: <45DA23C7.6090800@torque.net> (raw)
In-Reply-To: <Pine.LNX.4.44L0.0702191148421.13197-100000@netrider.rowland.org>
Alan Stern wrote:
> On Mon, 19 Feb 2007, Joerg Schilling wrote:
>
>> Alan Stern <stern@rowland.harvard.edu> wrote:
>>
>>> Well, if Doug wants to reduce the value returned by SG_GET_RESERVED_SIZE,
>>> it's okay with me. An advantage of doing this is that older versions of
>>> cdrecord would then work correctly.
>>>
>>> However you don't seem to realize that people can use programs like
>>> cdrecord with devices whose drivers don't support SG_GET_RESERVED_SIZE --
>>> because that ioctl works only with sg. Programs would have to try
>>> SG_GET_RESERVED_SIZE and if it faied, then try BLKSECTGET.
>> Is there any reason not to have one single ioctl for one basic feature?
>
> Indeed there is not. That's what I wrote in an earlier email:
>
> "There should be one single ioctl which can be applied uniformly to all
> CD-type devices (in fact, to all devices using a request_queue) to learn
> max_sectors. This rules out using SG_GET_RESERVED_SIZE."
>
>>> Remember also, the "reserved size" is _not_ the maximum allowed size of a
>>> DMA transfer. Rather, it is the size of an internal buffer maintained by
>>> sg. It's legal to do an I/O transfer larger than the "reserved size", but
>>> it is not legal to do an I/O transfer larger than max_sectors.
>> At the time the call SG_GET_RESERVED_SIZE has been discussed/defined, we did
>> originally agree that the max value should be limited to what the HW allows
>> as DMA size. This is why I did originally files a bug against
>> SG_GET_RESERVED_SIZE.
>
> How do you feel about the patch below, either in addition to or instead of
> the previous patch?
Alan,
The SG_GET_RESERVED_SIZE ioctl is also defined in
the block layer, see block/scsi_ioctl.c .
I suspect it is just a kludge to fool cdrecord that it
is talking to a sg device. [One of many kludges in the
block SG_IO ioctl implementation to that end.]
So perhaps the block layer versions of SG_SET_RESERVED_SIZE
and SG_GET_RESERVED_SIZE need to be similarly capped.
Actually I think that I would default SG_GET_RESERVED_SIZE to
request_queue->max_sectors * 512 in the block layer
implementation (as there is no "reserve buffer" associated
with a block device).
<aside>
The idea of a reserved buffer may live on in bsg as experience
with sg has shown that it is the fastest way to do (mmap-ed) IO.
Having one reserved buffer per file descriptor means not
having to create and tear down a scatter gather list
per IO. [Having a pool of such lists would be even better.]
Until optical storage needs 10 times its current datarates
then cdrecord will not need this mechanism.
</aside>
Doug Gilbert
> Index: usb-2.6/drivers/scsi/sg.c
> ===================================================================
> --- usb-2.6.orig/drivers/scsi/sg.c
> +++ usb-2.6/drivers/scsi/sg.c
> @@ -917,6 +917,8 @@ sg_ioctl(struct inode *inode, struct fil
> return result;
> if (val < 0)
> return -EINVAL;
> + if (val > sdp->device->request_queue->max_sectors * 512)
> + return -EOVERFLOW;
> if (val != sfp->reserve.bufflen) {
> if (sg_res_in_use(sfp) || sfp->mmap_called)
> return -EBUSY;
> @@ -925,7 +927,8 @@ sg_ioctl(struct inode *inode, struct fil
> }
> return 0;
> case SG_GET_RESERVED_SIZE:
> - val = (int) sfp->reserve.bufflen;
> + val = min_t(int, sfp->reserve.bufflen,
> + sdp->device->request_queue->max_sectors * 512);
> return put_user(val, ip);
> case SG_SET_COMMAND_Q:
> result = get_user(val, ip);
>
next prev parent reply other threads:[~2007-02-19 22:25 UTC|newest]
Thread overview: 17+ messages / expand[flat|nested] mbox.gz Atom feed top
2007-02-16 19:37 Alan Stern
2007-02-17 6:28 ` Jens Axboe
2007-02-17 21:18 ` Joerg Schilling
2007-02-18 3:43 ` Douglas Gilbert
2007-02-18 12:37 ` Joerg Schilling
2007-02-18 16:44 ` Alan Stern
2007-02-18 18:27 ` Joerg Schilling
2007-02-19 16:06 ` Alan Stern
2007-02-19 16:08 ` Joerg Schilling
2007-02-19 17:06 ` Alan Stern
2007-02-19 22:25 ` Douglas Gilbert [this message]
2007-02-20 3:48 ` Alan Stern
2007-02-20 4:47 ` Douglas Gilbert
2007-02-20 15:55 ` Alan Stern
2007-04-26 9:19 ` Joerg Schilling
2007-04-26 15:04 ` Alan Stern
2007-04-26 15:08 ` James Bottomley
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=45DA23C7.6090800@torque.net \
--to=dougg@torque.net \
--cc=James.Bottomley@steeleye.com \
--cc=Joerg.Schilling@fokus.fraunhofer.de \
--cc=jens.axboe@oracle.com \
--cc=linux-kernel@vger.kernel.org \
--cc=stern@rowland.harvard.edu \
--subject='Re: [PATCH] Block layer: separate out queue-oriented ioctls' \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).