From: Nicholas Bellinger <[email protected]>
Greetings Jens and co,
This patch adds the missing support to block/bsg.c:bsg_map_hdr() to accept
struct sg_io_v4->d[out,in]_iovec_count and process struct sg_io_v4->d[out,in]_xferp memory
containing userspace iovecs for kernel level BSG. It adds a new wrapper bsg_rq_map_user_iov()
that will call copy_from_user() and blk_rq_map_user_iov() following the original SG_IO logic in
drivers/scsi/sg.c:sg_start_req().
So far this has been tested on a x86_64 v2.6.34 KVM Host with TCM_Loop Virtual SAS Port/LUNs
into a x86_64 v2.6.26 KVM Guest with Megasas 8707EM2 HBA Emulation + my new scsi-bsg backstore code.
Please consider this for v2.6.36 as it will be required in order for QEMU-KVM MegaSAS and VirtIO HBA
emulation using QEMU scatterlist memory and BSG backstores.
Signed-off-by: Nicholas A. Bellinger <[email protected]>
---
block/bsg.c | 55 +++++++++++++++++++++++++++++++++++++++++++++++--------
1 files changed, 47 insertions(+), 8 deletions(-)
diff --git a/block/bsg.c b/block/bsg.c
index 82d5882..3f789cf 100644
--- a/block/bsg.c
+++ b/block/bsg.c
@@ -239,6 +239,33 @@ bsg_validate_sgv4_hdr(struct request_queue *q, struct sg_io_v4 *hdr, int *rw)
return ret;
}
+static int
+bsg_rq_map_user_iov(struct request_queue *q, struct request *rq, void *dxferp,
+ unsigned int dxfer_len, int iovec_count, int gfp)
+{
+ struct iovec *iov;
+ int len, ret, size = sizeof(struct sg_iovec) * iovec_count;
+
+ iov = kzalloc(size, GFP_ATOMIC);
+ if (!iov)
+ return -ENOMEM;
+
+ if (copy_from_user(iov, dxferp, size)) {
+ kfree(iov);
+ return -EFAULT;
+ }
+
+ len = iov_length(iov, iovec_count);
+ if (dxfer_len < len) {
+ iovec_count = iov_shorten(iov, iovec_count, dxfer_len);
+ len = dxfer_len;
+ }
+ ret = blk_rq_map_user_iov(q, rq, NULL, dxferp, iovec_count,
+ len, gfp);
+ kfree(iov);
+ return ret;
+}
+
/*
* map sg_io_v4 to a request.
*/
@@ -248,7 +275,7 @@ bsg_map_hdr(struct bsg_device *bd, struct sg_io_v4 *hdr, fmode_t has_write_perm,
{
struct request_queue *q = bd->queue;
struct request *rq, *next_rq = NULL;
- int ret, rw;
+ int ret, rw, iovec_count;
unsigned int dxfer_len;
void *dxferp = NULL;
@@ -284,28 +311,40 @@ bsg_map_hdr(struct bsg_device *bd, struct sg_io_v4 *hdr, fmode_t has_write_perm,
rq->next_rq = next_rq;
next_rq->cmd_type = rq->cmd_type;
+ iovec_count = hdr->din_iovec_count;
dxferp = (void*)(unsigned long)hdr->din_xferp;
- ret = blk_rq_map_user(q, next_rq, NULL, dxferp,
- hdr->din_xfer_len, GFP_KERNEL);
+
+ if (iovec_count)
+ ret = bsg_rq_map_user_iov(q, next_rq, dxferp,
+ hdr->din_xfer_len, iovec_count,
+ GFP_KERNEL);
+ else
+ ret = blk_rq_map_user(q, next_rq, NULL, dxferp,
+ hdr->din_xfer_len, GFP_KERNEL);
if (ret)
goto out;
}
if (hdr->dout_xfer_len) {
dxfer_len = hdr->dout_xfer_len;
+ iovec_count = hdr->dout_iovec_count;
dxferp = (void*)(unsigned long)hdr->dout_xferp;
} else if (hdr->din_xfer_len) {
dxfer_len = hdr->din_xfer_len;
+ iovec_count = hdr->din_iovec_count;
dxferp = (void*)(unsigned long)hdr->din_xferp;
} else
dxfer_len = 0;
- if (dxfer_len) {
+ if (iovec_count && dxfer_len)
+ ret = bsg_rq_map_user_iov(q, rq, dxferp, dxfer_len,
+ iovec_count, GFP_KERNEL);
+ else if (dxfer_len)
ret = blk_rq_map_user(q, rq, NULL, dxferp, dxfer_len,
- GFP_KERNEL);
- if (ret)
- goto out;
- }
+ GFP_KERNEL);
+
+ if (ret)
+ goto out;
rq->sense = sense;
rq->sense_len = 0;
--
1.5.6.5
On Sun, 13 Jun 2010 23:11:29 -0700
"Nicholas A. Bellinger" <[email protected]> wrote:
> From: Nicholas Bellinger <[email protected]>
>
> Greetings Jens and co,
>
> This patch adds the missing support to block/bsg.c:bsg_map_hdr() to accept
> struct sg_io_v4->d[out,in]_iovec_count and process struct sg_io_v4->d[out,in]_xferp memory
> containing userspace iovecs for kernel level BSG. It adds a new wrapper bsg_rq_map_user_iov()
> that will call copy_from_user() and blk_rq_map_user_iov() following the original SG_IO logic in
> drivers/scsi/sg.c:sg_start_req().
>
> So far this has been tested on a x86_64 v2.6.34 KVM Host with TCM_Loop Virtual SAS Port/LUNs
> into a x86_64 v2.6.26 KVM Guest with Megasas 8707EM2 HBA Emulation + my new scsi-bsg backstore code.
>
> Please consider this for v2.6.36 as it will be required in order for QEMU-KVM MegaSAS and VirtIO HBA
> emulation using QEMU scatterlist memory and BSG backstores.
>
> Signed-off-by: Nicholas A. Bellinger <[email protected]>
> ---
> block/bsg.c | 55 +++++++++++++++++++++++++++++++++++++++++++++++--------
> 1 files changed, 47 insertions(+), 8 deletions(-)
This have been rejected several times.
This doesn't work on 32bit user/64bit kernel.
The compat code doesn't work for the read/write interface (does for
the ioctl interface though). So we can't support this feature cleanly.
On Mon, 2010-06-14 at 15:22 +0900, FUJITA Tomonori wrote:
> On Sun, 13 Jun 2010 23:11:29 -0700
> "Nicholas A. Bellinger" <[email protected]> wrote:
>
> > From: Nicholas Bellinger <[email protected]>
> >
> > Greetings Jens and co,
> >
> > This patch adds the missing support to block/bsg.c:bsg_map_hdr() to accept
> > struct sg_io_v4->d[out,in]_iovec_count and process struct sg_io_v4->d[out,in]_xferp memory
> > containing userspace iovecs for kernel level BSG. It adds a new wrapper bsg_rq_map_user_iov()
> > that will call copy_from_user() and blk_rq_map_user_iov() following the original SG_IO logic in
> > drivers/scsi/sg.c:sg_start_req().
> >
> > So far this has been tested on a x86_64 v2.6.34 KVM Host with TCM_Loop Virtual SAS Port/LUNs
> > into a x86_64 v2.6.26 KVM Guest with Megasas 8707EM2 HBA Emulation + my new scsi-bsg backstore code.
> >
> > Please consider this for v2.6.36 as it will be required in order for QEMU-KVM MegaSAS and VirtIO HBA
> > emulation using QEMU scatterlist memory and BSG backstores.
> >
> > Signed-off-by: Nicholas A. Bellinger <[email protected]>
> > ---
> > block/bsg.c | 55 +++++++++++++++++++++++++++++++++++++++++++++++--------
> > 1 files changed, 47 insertions(+), 8 deletions(-)
>
> This have been rejected several times.
>
I guess that would explain why the iovec_count structure members are
part of struct sg_io_v4 and block/bsg.c:bsg_map_hdr() silently accepts
and overwrites them.. :(
> This doesn't work on 32bit user/64bit kernel.
Hmmm, being able to use QEMU-KVM HBA emulation with userspace
scatterlist memory for the various HBA emulation with a matched
user/kernel is still useful though, yes..? Being able to handle TMR
emulation from QEMU-KVM HBA emulation would also be useful I think.
Having QEMU-KVM fall back to SG_IO for the 32bit/64bit kernel would be
accpetable for me if we can still make BSG + iovecs work for the typical
case.
>
> The compat code doesn't work for the read/write interface (does for
> the ioctl interface though). So we can't support this feature cleanly.
So I am curious to see if there will be a performance improvement
between QEMU-KVM + scsi-bsg w/ AIO IOVECs compared to the legacy SG_IO
ioctl(). Anyways, I will post my QEMU-KVM scsi-bsg patches and run some
benchmarks this week with this patch on 5500 series Nehalem so see aside
from the userspace TMR case how useful having a BSG backstore for
QEMU-KVM may be.
Best,
--nab