mtd: nftl: Remove VLA usage

On the quest to remove all VLAs from the kernel[1] this changes the check_free_sectors() routine to use the same stack buffer for both data byte checks (SECTORSIZE) and oob byte checks (oobsize). Since these regions aren't needed at the same time, they don't need to be consecutively allocated. Additionally, while it's possible for oobsize to be large, it is unlikely to be larger than the actual SECTORSIZE. As such, remove the VLA, adjust offsets and add a sanity check to make sure we never get a pathological oobsize.

[1] https://lkml.kernel.org/r/CA+55aFzCG-zNmZwX4A2FQpadafLfEzK6CC=qPXydAacU1RqZWA@mail.gmail.com

Signed-off-by Kees Cook <[email protected]>

---
 drivers/mtd/inftlmount.c | 11 ++++++++---
 drivers/mtd/nftlmount.c  | 11 ++++++++---
 2 files changed, 16 insertions(+), 6 deletions(-)

drivers/mtd/inftlmount.c +0 -0

index aab4f68bd36f..9cdae7f0fc2e 100644

drivers/mtd/inftlmount.c +7 -2

 	struct mtd_info *mtd = inftl->mbd.mtd;
 	size_t retlen;
 	int i;

@@ -346,10 +346,15 @@ static int check_free_sectors(struct INFTLrecord *inftl, unsigned int address,
 			return -1;
 
 		if (check_oob) {
+			if (mtd->oobsize > sizeof(buf)) {
+				pr_warn("MTD oobsize > SECTORSIZE: %d\n",
+					mtd->oobsize);
+				return -1;
+			}
 			if(inftl_read_oob(mtd, address, mtd->oobsize,
-					  &retlen, &buf[SECTORSIZE]) < 0)
+					  &retlen, buf) < 0)
 				return -1;
-			if (memcmpb(buf + SECTORSIZE, 0xff, mtd->oobsize) != 0)
+			if (memcmpb(buf, 0xff, mtd->oobsize) != 0)
 				return -1;
 		}
 		address += SECTORSIZE;

drivers/mtd/nftlmount.c +8 -3

index a6fbfa4e5799..e6eba7f3fdf5 100644
@@ -272,7 +272,7 @@ static int memcmpb(void *a, int c, int n)
 static int check_free_sectors(struct NFTLrecord *nftl, unsigned int address, int len,
 			      int check_oob)
 {
-	u8 buf[SECTORSIZE + nftl->mbd.mtd->oobsize];
+	u8 buf[SECTORSIZE];
 	struct mtd_info *mtd = nftl->mbd.mtd;
 	size_t retlen;
 	int i;

@@ -284,10 +284,15 @@ static int check_free_sectors(struct NFTLrecord *nftl, unsigned int address, int
 			return -1;
 
 		if (check_oob) {
+			if (mtd->oobsize > sizeof(buf)) {
+				pr_warn("MTD oobsize > SECTORSIZE: %d\n",
+					mtd->oobsize);
+				return -1;
+			}
 			if(nftl_read_oob(mtd, address, mtd->oobsize,
-					 &retlen, &buf[SECTORSIZE]) < 0)
+					 &retlen, buf) < 0)
 				return -1;
-			if (memcmpb(buf + SECTORSIZE, 0xff, mtd->oobsize) != 0)
+			if (memcmpb(buf, 0xff, mtd->oobsize) != 0)
 				return -1;
 		}
 		address += SECTORSIZE;

Hi Kees,

Quoted 17 lines · On Mon, 23 Apr 2018 13:35:00 -0700 Kees Cook <[email protected]> wrote:

On the quest to remove all VLAs from the kernel[1] this changes the check_free_sectors() routine to use the same stack buffer for both data byte checks (SECTORSIZE) and oob byte checks (oobsize). Since these regions aren't needed at the same time, they don't need to be consecutively allocated. Additionally, while it's possible for oobsize to be large, it is unlikely to be larger than the actual SECTORSIZE. As such, remove the VLA, adjust offsets and add a sanity check to make sure we never get a pathological oobsize.

[1] https://lkml.kernel.org/r/CA+55aFzCG-zNmZwX4A2FQpadafLfEzK6CC=qPXydAacU1RqZWA@mail.gmail.com

Signed-off-by Kees Cook <[email protected]>

---
 drivers/mtd/inftlmount.c | 11 ++++++++---
 drivers/mtd/nftlmount.c  | 11 ++++++++---
 2 files changed, 16 insertions(+), 6 deletions(-)

drivers/mtd/inftlmount.c static int memcmpb(void *a, int c, int n) · line 337

Quoted 10 lines

drivers/mtd/inftlmount.c +1 -1

index aab4f68bd36f..9cdae7f0fc2e 100644
@@ -334,7 +334,7 @@ static int memcmpb(void *a, int c, int n)
 static int check_free_sectors(struct INFTLrecord *inftl, unsigned int address,
      int len, int check_oob)
 {
-     u8 buf[SECTORSIZE + inftl->mbd.mtd->oobsize];
+     u8 buf[SECTORSIZE];

Could we instead move to dynamic allocation. I mean, SECTORSIZE is 512 bytes, so only with this function we consume 1/16 of the stack. Not to mention that some MTD drivers might want to do DMA on buffer passed by the MTD user, and if the buffer is on the stack they'll have to use a bounce buffer instead.

Quoted 22 lines · Boris Brezillon

Hi Kees,

Quoted 17 lines · On Mon, 23 Apr 2018 13:35:00 -0700 Kees Cook <[email protected]> wrote:

On the quest to remove all VLAs from the kernel[1] this changes the check_free_sectors() routine to use the same stack buffer for both data byte checks (SECTORSIZE) and oob byte checks (oobsize). Since these regions aren't needed at the same time, they don't need to be consecutively allocated. Additionally, while it's possible for oobsize to be large, it is unlikely to be larger than the actual SECTORSIZE. As such, remove the VLA, adjust offsets and add a sanity check to make sure we never get a pathological oobsize.

[1] https://lkml.kernel.org/r/CA+55aFzCG-zNmZwX4A2FQpadafLfEzK6CC=qPXydAacU1RqZWA@mail.gmail.com

Signed-off-by Kees Cook <[email protected]>

---
 drivers/mtd/inftlmount.c | 11 ++++++++---
 drivers/mtd/nftlmount.c  | 11 ++++++++---
 2 files changed, 16 insertions(+), 6 deletions(-)

Quoted 16 lines

drivers/mtd/inftlmount.c static int memcmpb(void *a, int c, int n) · line 337

Quoted 10 lines

drivers/mtd/inftlmount.c +1 -1

index aab4f68bd36f..9cdae7f0fc2e 100644
@@ -334,7 +334,7 @@ static int memcmpb(void *a, int c, int n)
 static int check_free_sectors(struct INFTLrecord *inftl, unsigned int address,
      int len, int check_oob)
 {
-     u8 buf[SECTORSIZE + inftl->mbd.mtd->oobsize];
+     u8 buf[SECTORSIZE];  

Could we instead move to dynamic allocation. I mean, SECTORSIZE is 512 bytes, so only with this function we consume 1/16 of the stack. Not to mention that some MTD drivers might want to do DMA on buffer passed by the MTD user, and if the buffer is on the stack they'll have to use a bounce buffer instead.

Sure! I can rework it to do that. Is GFP_KERNEL okay for that, or does it need something else?