mtd: sf: Make sf_mtd.c more robust
authorBoris Brezillon <boris.brezillon@bootlin.com>
Sun, 2 Dec 2018 09:54:32 +0000 (10:54 +0100)
committerJagan Teki <jagan@amarulasolutions.com>
Wed, 5 Dec 2018 19:15:36 +0000 (00:45 +0530)
SPI flash based MTD devs can be registered/unregistered at any time
through the sf probe command or the spi_flash_free() function.

This commit does not try to fix the root cause as it would probably
require rewriting most of the code and have an mtd_info object
instance per spi_flash object (not to mention that the the spi-flash
layer is likely to be replaced by a spi-nor layer ported from Linux).

Instead, we try to be as safe as can be by checking the code returned
by del_mtd_device() and complain loudly when there's nothing we can
do about the deregistration failure. When that happens we also reset
sf_mtd_info.priv to NULL, and check for NULL pointer in the mtd hooks
so that -ENODEV is returned instead of hitting a NULL pointer
dereference exception when the MTD instance is later accessed by a user.

Signed-off-by: Boris Brezillon <boris.brezillon@bootlin.com>
Tested-by: Heiko Schocher <hs@denx.de>
drivers/mtd/spi/sf_mtd.c

index aabbc35..68c3600 100644 (file)
@@ -18,6 +18,9 @@ static int spi_flash_mtd_erase(struct mtd_info *mtd, struct erase_info *instr)
        struct spi_flash *flash = mtd->priv;
        int err;
 
+       if (!flash)
+               return -ENODEV;
+
        instr->state = MTD_ERASING;
 
        err = spi_flash_erase(flash, instr->addr, instr->len);
@@ -39,6 +42,9 @@ static int spi_flash_mtd_read(struct mtd_info *mtd, loff_t from, size_t len,
        struct spi_flash *flash = mtd->priv;
        int err;
 
+       if (!flash)
+               return -ENODEV;
+
        err = spi_flash_read(flash, from, len, buf);
        if (!err)
                *retlen = len;
@@ -52,6 +58,9 @@ static int spi_flash_mtd_write(struct mtd_info *mtd, loff_t to, size_t len,
        struct spi_flash *flash = mtd->priv;
        int err;
 
+       if (!flash)
+               return -ENODEV;
+
        err = spi_flash_write(flash, to, len, buf);
        if (!err)
                *retlen = len;
@@ -76,8 +85,13 @@ int spi_flash_mtd_register(struct spi_flash *flash)
 {
        int ret;
 
-       if (sf_mtd_registered)
-               del_mtd_device(&sf_mtd_info);
+       if (sf_mtd_registered) {
+               ret = del_mtd_device(&sf_mtd_info);
+               if (ret)
+                       return ret;
+
+               sf_mtd_registered = false;
+       }
 
        sf_mtd_registered = false;
        memset(&sf_mtd_info, 0, sizeof(sf_mtd_info));
@@ -110,5 +124,24 @@ int spi_flash_mtd_register(struct spi_flash *flash)
 
 void spi_flash_mtd_unregister(void)
 {
-       del_mtd_device(&sf_mtd_info);
+       int ret;
+
+       if (!sf_mtd_registered)
+               return;
+
+       ret = del_mtd_device(&sf_mtd_info);
+       if (!ret) {
+               sf_mtd_registered = false;
+               return;
+       }
+
+       /*
+        * Setting mtd->priv to NULL is the best we can do. Thanks to that,
+        * the MTD layer can still call mtd hooks without risking a
+        * use-after-free bug. Still, things should be fixed to prevent the
+        * spi_flash object from being destroyed when del_mtd_device() fails.
+        */
+       sf_mtd_info.priv = NULL;
+       printf("Failed to unregister MTD %s and the spi_flash object is going away: you're in deep trouble!",
+              sf_mtd_info.name);
 }