Flush mtdblock buffer after writes to hnvram
The mtdblock driver in Linux caches erase blocks indefinitely until the
last file descriptor to /dev/mtdblockX has been closed. This is not
acceptable because we assume that data has been flushed to flash storage
when the hnvram tool returns.
Steps to reproduce:
Open a fd to /dev/mtdblockX and leave it open:
cat > /dev/mtd/hnvram
On a different tty, run
hnvram -w ACTIVATED_KERNEL_NAME=kernel7 ; reboot
On the next boot, the main parition of HNVRAM will contain
ACTIVATED_KERNEL_NAME=kernel7, but the backup partition still has the
old value. Hence, the main and backup HNVRAM partition are inconsistent.
Hint: The hnvram tool internally subdivides the HNVRAM partition in the
NOR into a main and backup partition. (Look for NVRAM_RW_OFFSET and
NVRAM_RWB_OFFSET in hmx_upgrade_nvram.c)
Change-Id: Ifd3c095a5bfe966b3e512ded81164b710f506fd2
diff --git a/libupgrade/hmx_upgrade_flash.c b/libupgrade/hmx_upgrade_flash.c
index 6153d01..9f9ae2c 100644
--- a/libupgrade/hmx_upgrade_flash.c
+++ b/libupgrade/hmx_upgrade_flash.c
@@ -142,6 +142,16 @@
return -1;
}
+ /* Writes to /dev/mtdblockX are cached indefinitely until the last file
+ * descriptor is closed. Flush this cache after writing. The user depends on
+ * data being physically written to flash when this function returns.
+ * Requires CAP_SYS_ADMIN. */
+ ret = ioctl(fd_nvram, BLKFLSBUF, 0);
+ if(ret < 0) {
+ perror("ioctl(hnvram, BLKFLSBUF, 0)");
+ return -1;
+ }
+
return 0;
}
