mtd: spi-nor: take the flash lock in spi_nor_restore()

spi_nor_shutdown() and spi_nor_remove() reach the chip through spi_nor_restore() without holding nor->lock, which every other path to the flash takes through spi_nor_prep_and_lock(). Both run with the MTD device still registered, so another thread can be mid-operation.

Neither outcome of that race is good. While the chip is busy the address mode exit is typically ignored, so the restore is silently dropped and the chip is left in 4-byte addressing, which a kexec then hands to a kernel that cannot read it. The soft reset is typically accepted even while busy, so it instead aborts an in-flight erase part way through.

Take the lock in spi_nor_restore(); both callers are teardown paths that can sleep and neither already holds it. Waiting is deliberate: a partially erased sector cannot be undone from here, and the wait is bounded by the one operation already in flight. If the lock cannot be taken the restore is skipped with a warning.

Requests arriving after the restore are a separate problem: they race with device removal itself rather than with the restore.

---
Changes in v4:
- Correct what v3 claimed a busy chip does. v3 said it "ignores everything but status reads", which is roughly right for the address mode exit but wrong for the soft reset: that one is accepted while busy and aborts an in-flight erase. This strengthens the case for taking the lock rather than weakening it, but the old wording was not accurate and the comment now says both halves.
- Lock inside spi_nor_restore() instead of adding a spi_nor_restore_locked() wrapper around it. The _locked suffix conventionally means "the caller holds the lock", which is the opposite of what that wrapper did, and spi_nor_restore() is static with only the two teardown callers.
- Handle the spi_nor_prep_and_lock() error return instead of ignoring it, and warn when the restore is skipped.
- Say what happens to an operation that is already in flight, which Michael asked about on v3. Short answer: it is waited for, because an erase cannot be cancelled safely.
- Rebased on mainline. v3 depended on the spi_nor_rww_start_exclusive() mutex leak fix, which has since merged as 44b8a0bf5f96 ("mtd:
spi-nor: core: Fix mutex leak in spi_nor_rww_start_exclusive()"), so this no longer depends on anything unmerged. Note that it is not yet in a released tag: v7.3-rc5 does not have it.

Link to v3:
https://lore.kernel.org/r/[email protected]
Link to v2:
https://lore.kernel.org/r/[email protected]

 drivers/mtd/spi-nor/core.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

index 8bc117b..94dbf1b 100644
@@ -3455,6 +3455,21 @@ static void spi_nor_restore(struct spi_nor *nor)
 {
 	int ret;
 
+	/*
+	 * The MTD device is still registered here, so another thread may
+	 * have an operation in flight. Wait for it rather than cut in:
+	 * while the chip is busy the address mode exit below is typically
+	 * ignored, which silently loses the restore, and the soft reset is
+	 * typically accepted, which aborts an erase part way through.
+	 */
+	ret = spi_nor_prep_and_lock(nor);
+	if (ret) {
+		dev_warn(nor->dev,
+			 "Failed to lock flash, skipping restore, err = %d\n",
+			 ret);
+		return;
+	}
+
 	/* restore the addressing mode */
 	if (nor->addr_nbytes == 4 && !(nor->flags & SNOR_F_4B_OPCODES) &&
 	    nor->flags & SNOR_F_BROKEN_RESET) {

@@ -3470,6 +3485,8 @@ static void spi_nor_restore(struct spi_nor *nor)
 
 	if (nor->flags & SNOR_F_SOFT_RESET)
 		spi_nor_soft_reset(nor);
+
+	spi_nor_unlock_and_unprep(nor);
 }
 
 static const struct flash_info *spi_nor_match_name(struct spi_nor *nor,

base-commit: 6addb4f385570ebc11c4eb499a4f1c149f313e84

There are two mistakes here.

The wait is not bounded the way this suggests. spi_nor_erase() holds the lock for the whole request, so shutdown or unbind can end up waiting for a full chip erase.

And v4 still skips the restore when the lock cannot be taken, as v3 did. In my reply to Michael on v3 [1] I offered to do a best-effort restore with a warning instead, but v4 only adds the warning: