[PATCH] sg: recheck MMAP_IO request length with lock held

[PATCH] sg: recheck MMAP_IO request length with lock held

[Todd Poynor]

Commit 1bc0eb044615 ("scsi: sg: protect accesses to 'reserved' page
array") adds needed concurrency protection for the "reserve" buffer.
Some checks that are initially made outside the lock are replicated once
the lock is taken to ensure the checks and resulting decisions are made
using consistent state.

The check that a request with flag SG_FLAG_MMAP_IO set fits in the
reserve buffer also needs to be performed again under the lock to
ensure the reserve buffer length compared against matches the value in
effect when the request is linked to the reserve buffer.  An -ENOMEM
should be returned in this case, instead of switching over to an
indirect buffer as for non-MMAP_IO requests.

Signed-off-by: Todd Poynor <toddpoynor@google.com>
---
 drivers/scsi/sg.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
index d7ff71e0c85c..3a44b4bc872b 100644
--- a/drivers/scsi/sg.c
+++ b/drivers/scsi/sg.c
@@ -1735,9 +1735,12 @@ sg_start_req(Sg_request *srp, unsigned char *cmd)
 		    !sfp->res_in_use) {
 			sfp->res_in_use = 1;
 			sg_link_reserve(sfp, srp, dxfer_len);
-		} else if ((hp->flags & SG_FLAG_MMAP_IO) && sfp->res_in_use) {
+		} else if (hp->flags & SG_FLAG_MMAP_IO) {
+			res = -EBUSY; /* sfp->res_in_use == 1 */
+			if (dxfer_len > rsv_schp->bufflen)
+				res = -ENOMEM;
 			mutex_unlock(&sfp->f_mutex);
-			return -EBUSY;
+			return res;
 		} else {
 			res = sg_build_indirect(req_schp, sfp, dxfer_len);
 			if (res) {
-- 
2.14.1.480.gb18f417b89-goog

Re: [PATCH] sg: recheck MMAP_IO request length with lock held

[Douglas Gilbert]

On 2017-08-16 12:48 AM, Todd Poynor wrote:
> Commit 1bc0eb044615 ("scsi: sg: protect accesses to 'reserved' page
> array") adds needed concurrency protection for the "reserve" buffer.
> Some checks that are initially made outside the lock are replicated once
> the lock is taken to ensure the checks and resulting decisions are made
> using consistent state.
> 
> The check that a request with flag SG_FLAG_MMAP_IO set fits in the
> reserve buffer also needs to be performed again under the lock to
> ensure the reserve buffer length compared against matches the value in
> effect when the request is linked to the reserve buffer.  An -ENOMEM
> should be returned in this case, instead of switching over to an
> indirect buffer as for non-MMAP_IO requests.
> 
> Signed-off-by: Todd Poynor <toddpoynor@google.com>
Acked-by: Douglas Gilbert <dgilbert@interlog.com>

Thanks.

> ---
>   drivers/scsi/sg.c | 7 +++++--
>   1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
> index d7ff71e0c85c..3a44b4bc872b 100644
> --- a/drivers/scsi/sg.c
> +++ b/drivers/scsi/sg.c
> @@ -1735,9 +1735,12 @@ sg_start_req(Sg_request *srp, unsigned char *cmd)
>   		    !sfp->res_in_use) {
>   			sfp->res_in_use = 1;
>   			sg_link_reserve(sfp, srp, dxfer_len);
> -		} else if ((hp->flags & SG_FLAG_MMAP_IO) && sfp->res_in_use) {
> +		} else if (hp->flags & SG_FLAG_MMAP_IO) {
> +			res = -EBUSY; /* sfp->res_in_use == 1 */
> +			if (dxfer_len > rsv_schp->bufflen)
> +				res = -ENOMEM;
>   			mutex_unlock(&sfp->f_mutex);
> -			return -EBUSY;
> +			return res;
>   		} else {
>   			res = sg_build_indirect(req_schp, sfp, dxfer_len);
>   			if (res) {
> 

Re: [PATCH] sg: recheck MMAP_IO request length with lock held

[Martin K. Petersen]

Todd,

> Commit 1bc0eb044615 ("scsi: sg: protect accesses to 'reserved' page
> array") adds needed concurrency protection for the "reserve" buffer.
> Some checks that are initially made outside the lock are replicated once
> the lock is taken to ensure the checks and resulting decisions are made
> using consistent state.
>
> The check that a request with flag SG_FLAG_MMAP_IO set fits in the
> reserve buffer also needs to be performed again under the lock to
> ensure the reserve buffer length compared against matches the value in
> effect when the request is linked to the reserve buffer.  An -ENOMEM
> should be returned in this case, instead of switching over to an
> indirect buffer as for non-MMAP_IO requests.

Applied to 4.14/scsi-queue, thank you!

-- 
Martin K. Petersen	Oracle Linux Engineering