[PATCH] sg: protect against races between mmap() and SG_SET_RESERVED_SIZE

[Todd Poynor]

Take f_mutex around mmap() processing to protect against races with
the SG_SET_RESERVED_SIZE ioctl.  Ensure the reserve buffer length
remains consistent during the mapping operation, and set the
"mmap called" flag to prevent further changes to the reserved buffer
size as an atomic operation with the mapping.

Signed-off-by: Todd Poynor <toddpoy...@google.com>
---
 drivers/scsi/sg.c | 12 +---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
index 3a44b4bc872b..a20718e9f1f4 100644
--- a/drivers/scsi/sg.c
+++ b/drivers/scsi/sg.c
@@ -1233,6 +1233,7 @@ sg_mmap(struct file *filp, struct vm_area_struct *vma)
unsigned long req_sz, len, sa;
Sg_scatter_hold *rsv_schp;
int k, length;
+int ret = 0;
 
if ((!filp) || (!vma) || (!(sfp = (Sg_fd *) filp->private_data)))
return -ENXIO;
@@ -1243,8 +1244,11 @@ sg_mmap(struct file *filp, struct vm_area_struct *vma)
if (vma->vm_pgoff)
return -EINVAL; /* want no offset */
rsv_schp = >reserve;
-   if (req_sz > rsv_schp->bufflen)
-   return -ENOMEM; /* cannot map more than reserved buffer */
+   mutex_lock(>f_mutex);
+   if (req_sz > rsv_schp->bufflen) {
+   ret = -ENOMEM;  /* cannot map more than reserved buffer */
+   goto out;
+   }
 
sa = vma->vm_start;
length = 1 << (PAGE_SHIFT + rsv_schp->page_order);
@@ -1258,7 +1262,9 @@ sg_mmap(struct file *filp, struct vm_area_struct *vma)
vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
vma->vm_private_data = sfp;
vma->vm_ops = _mmap_vm_ops;
-   return 0;
+out:
+   mutex_unlock(>f_mutex);
+   return ret;
 }
 
 static void
-- 
2.14.1.480.gb18f417b89-goog

[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 <toddpoy...@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(>f_mutex);
-   return -EBUSY;
+   return res;
} else {
res = sg_build_indirect(req_schp, sfp, dxfer_len);
if (res) {
-- 
2.14.1.480.gb18f417b89-goog