mtd locking fix and cleanups
Hi mtd maintainers,
the first series in this patch fixes a lock order reversal reported by Guenter based on a recent commit, although from code inspection it should have been around for much longer.
The rest is drive-by cleanups in the area that I noticed while trying to understand the locking changes. These are untested and need to be handled with care!
Diffstat:
ftl.c | 2 - mtd_blkdevs.c | 60 ++++++---------------------------------------------------- rfd_ftl.c | 2 - 3 files changed, 9 insertions(+), 55 deletions(-)
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
PATCH 1/8 mtd_blkdevs: don't hold del_mtd_blktrans_dev in blktrans_{open, release}
There is nothing that this protects against except for slightly reducing the window when new opens can appear just before calling del_gendisk.
--- drivers/mtd/mtd_blkdevs.c | 5 ----- 1 file changed, 5 deletions(-)
drivers/mtd/mtd_blkdevs.c +0 -5
@@ -207,7 +207,6 @@ static int blktrans_open(struct block_device *bdev, fmode_t mode)if (!dev)return -ERESTARTSYS ;/* FIXME: busy loop! -arnd*/ -mutex_lock (&mtd_table_mutex);mutex_lock (&dev->lock);if (dev->open)
@@ -233,7 +232,6 @@ static int blktrans_open(struct block_device *bdev, fmode_t mode) unlock: dev->open++;mutex_unlock (&dev->lock); -mutex_unlock (&mtd_table_mutex);blktrans_dev_put (dev);return ret;
@@ -244,7 +242,6 @@ static int blktrans_open(struct block_device *bdev, fmode_t mode)module_put (dev->tr->owner);kref_put (&dev->ref, blktrans_dev_release);mutex_unlock (&dev->lock); -mutex_unlock (&mtd_table_mutex);blktrans_dev_put (dev);return ret; }
@@ -256,7 +253,6 @@ static void blktrans_release(struct gendisk *disk, fmode_t mode)if (!dev)return ; -mutex_lock (&mtd_table_mutex);mutex_lock (&dev->lock);if (--dev->open)
@@ -272,7 +268,6 @@ static void blktrans_release(struct gendisk *disk, fmode_t mode) } unlock:mutex_unlock (&dev->lock); -mutex_unlock (&mtd_table_mutex);blktrans_dev_put (dev); }
Signature · 6 lines
2.30.2 ______________________________________________________ Linux MTD discussion mailing list http://lists.infradead.org/mailman/listinfo/linux-mtd/
Re: [PATCH 1/8] mtd_blkdevs: don't hold del_mtd_blktrans_dev in blktrans_{open, release}
Quoted 6 lines · Christoph Hellwig
There is nothing that this protects against except for slightly reducing the window when new opens can appear just before calling
del_gendisk.Reported-by Guenter Roeck <[email protected]>Signed-off-by Christoph Hellwig <[email protected]>Tested-by Guenter Roeck <[email protected]>
Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git mtd/next, thanks.
Miquel
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
PATCH 2/8 mtd_blkdevs: use lockdep_assert_held
In reply to Christoph HellwigUse lockdep_assert_held to ensure mtd_table_mutex is held instead of mutex_trylock games.
--- drivers/mtd/mtd_blkdevs.c | 10 ++-------- 1 file changed, 2 insertions(+), 8 deletions(-)
drivers/mtd/mtd_blkdevs.c +2 -8
@@ -310,10 +310,7 @@ int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)struct gendisk *gd;int ret; -if (mutex_trylock (&mtd_table_mutex)) { -mutex_unlock (&mtd_table_mutex); -BUG (); - } +lockdep_assert_held (&mtd_table_mutex);mutex_lock (&blktrans_ref_mutex);list_for_each_entry (d, &tr->devs, list) {
@@ -444,10 +441,7 @@ int del_mtd_blktrans_dev(struct mtd_blktrans_dev *old) {unsigned long flags; -if (mutex_trylock (&mtd_table_mutex)) { -mutex_unlock (&mtd_table_mutex); -BUG (); - } +lockdep_assert_held (&mtd_table_mutex);if (old->disk_attributes) sysfs_remove_group(&disk_to_dev(old->disk)->kobj,
Signature · 6 lines
2.30.2 ______________________________________________________ Linux MTD discussion mailing list http://lists.infradead.org/mailman/listinfo/linux-mtd/
Re: [PATCH 2/8] mtd_blkdevs: use lockdep_assert_held
Quoted 4 lines · Christoph Hellwig
Use
lockdep_assert_heldto ensuremtd_table_mutexis held instead ofmutex_trylockgames.Signed-off-by Christoph Hellwig <[email protected]>
Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git mtd/next, thanks.
Miquel
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
PATCH 3/8 mtd/ftl: don't cast away the type when calling add_mtd_blktrans_dev
In reply to Christoph HellwigPass the actual mtd_blktrans_dev instead of casting the containing structure to void *.
--- drivers/mtd/ftl.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/mtd/ftl.c +1 -1
@@ -1029,7 +1029,7 @@ static void ftl_add_mtd(struct mtd_blktrans_ops *tr, struct mtd_info *mtd) partition->mbd.tr = tr; partition->mbd.devnum =-1 ; -if (!add_mtd_blktrans_dev ((void *)partition)) +if (!add_mtd_blktrans_dev (&partition->mbd))return ; }
Signature · 6 lines
2.30.2 ______________________________________________________ Linux MTD discussion mailing list http://lists.infradead.org/mailman/listinfo/linux-mtd/
Re: [PATCH 3/8] mtd/ftl: don't cast away the type when calling add_mtd_blktrans_dev
Quoted 4 lines · Christoph Hellwig
Pass the actual
mtd_blktrans_devinstead of casting the containing structure to void *.Signed-off-by Christoph Hellwig <[email protected]>
Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git mtd/next, thanks.
Miquel
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
PATCH 4/8 mtd/rfd_ftl: don't cast away the type when calling add_mtd_blktrans_dev
In reply to Christoph HellwigPass the actual mtd_blktrans_dev instead of casting the containing structure to void *.
--- drivers/mtd/rfd_ftl.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/mtd/rfd_ftl.c +1 -1
@@ -754,7 +754,7 @@ static void rfd_ftl_add_mtd(struct mtd_blktrans_ops *tr, struct mtd_info *mtd) printk(KERN_INFO PREFIX "name: '%s' type: %d flags %x\n", mtd->name, mtd->type, mtd->flags); -if (!add_mtd_blktrans_dev ((void *)part)) +if (!add_mtd_blktrans_dev (&part->mbd))return ; } out:
Signature · 6 lines
2.30.2 ______________________________________________________ Linux MTD discussion mailing list http://lists.infradead.org/mailman/listinfo/linux-mtd/
Re: [PATCH 4/8] mtd/rfd_ftl: don't cast away the type when calling add_mtd_blktrans_dev
Quoted 4 lines · Christoph Hellwig
Pass the actual
mtd_blktrans_devinstead of casting the containing structure to void *.Signed-off-by Christoph Hellwig <[email protected]>
Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git mtd/next, thanks.
Miquel
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
PATCH 5/8 mtd_blkdevs: simplify blktrans_dev_get
In reply to Christoph Hellwig->private_data is set before the disk is added and never cleared, so don't bother trying to handle a NULL pointer there.
--- drivers/mtd/mtd_blkdevs.c | 17 ++--------------- 1 file changed, 2 insertions(+), 15 deletions(-)
drivers/mtd/mtd_blkdevs.c +2 -15
@@ -39,16 +39,12 @@ static void blktrans_dev_release(struct kref *kref)static struct mtd_blktrans_dev *blktrans_dev_get (struct gendisk *disk) { -struct mtd_blktrans_dev *dev; +struct mtd_blktrans_dev *dev = disk->private_data;mutex_lock (&blktrans_ref_mutex); - dev = disk->private_data; - -if (!dev) - goto unlock;kref_get (&dev->ref); -unlock:mutex_unlock (&blktrans_ref_mutex); +return dev; }
@@ -204,9 +200,6 @@ static int blktrans_open(struct block_device *bdev, fmode_t mode)struct mtd_blktrans_dev *dev =blktrans_dev_get (bdev->bd_disk);int ret =0 ; -if (!dev) -return -ERESTARTSYS ;/* FIXME: busy loop! -arnd*/ -mutex_lock (&dev->lock);if (dev->open)
@@ -250,9 +243,6 @@ static void blktrans_release(struct gendisk *disk, fmode_t mode) {struct mtd_blktrans_dev *dev =blktrans_dev_get (disk); -if (!dev) -return ; -mutex_lock (&dev->lock);if (--dev->open)
@@ -276,9 +266,6 @@ static int blktrans_getgeo(struct block_device *bdev, struct hd_geometry *geo)struct mtd_blktrans_dev *dev =blktrans_dev_get (bdev->bd_disk);int ret = -ENXIO ; -if (!dev) -return ret; -mutex_lock (&dev->lock);if (!dev->mtd)
Signature · 6 lines
2.30.2 ______________________________________________________ Linux MTD discussion mailing list http://lists.infradead.org/mailman/listinfo/linux-mtd/
Re: [PATCH 5/8] mtd_blkdevs: simplify blktrans_dev_get
Quoted 4 lines · Christoph Hellwig
->private_data is set before the disk is added and never cleared, so don't bother trying to handle a NULL pointer there.
Signed-off-by Christoph Hellwig <[email protected]>
Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git mtd/next, thanks.
Miquel
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
PATCH 6/8 mtd_blkdevs: remove blktrans_ref_mutex
In reply to Christoph Hellwigblktrans_ref_mutex is not actually needed. The kref is serialized internally, and devnum assignment in add_mtd_blktrans_dev happens before the disk is added and thus any of the block_device_operations methods otherwise using it are called. It is also already serialized by the global mtd_table_mutex.
--- drivers/mtd/mtd_blkdevs.c | 13 +------------ 1 file changed, 1 insertion(+), 12 deletions(-)
drivers/mtd/mtd_blkdevs.c +1 -12
@@ -23,7 +23,6 @@#include "mtdcore.h" static LIST_HEAD (blktrans_majors); -static DEFINE_MUTEX (blktrans_ref_mutex);static void blktrans_dev_release (struct kref *kref) {
@@ -41,18 +40,13 @@ static struct mtd_blktrans_dev *blktrans_dev_get(struct gendisk *disk) {struct mtd_blktrans_dev *dev = disk->private_data; -mutex_lock (&blktrans_ref_mutex);kref_get (&dev->ref); -mutex_unlock (&blktrans_ref_mutex); -return dev; }static void blktrans_dev_put (struct mtd_blktrans_dev *dev) { -mutex_lock (&blktrans_ref_mutex);kref_put (&dev->ref, blktrans_dev_release); -mutex_unlock (&blktrans_ref_mutex); }
@@ -299,7 +293,6 @@ int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)lockdep_assert_held (&mtd_table_mutex); -mutex_lock (&blktrans_ref_mutex);list_for_each_entry (d, &tr->devs, list) {if (new->devnum ==-1 ) {/* Use first free number */
@@ -311,7 +304,6 @@ int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new) } }else if (d->devnum == new->devnum) {/* Required number taken */ -mutex_unlock (&blktrans_ref_mutex);return -EBUSY ; }else if (d->devnum > new->devnum) {/* Required number was free */
@@ -329,14 +321,11 @@ int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)* minor numbers and that the disk naming code below can cope * with this number. */ if (new->devnum > (MINORMASK >> tr->part_bits) || - (tr->part_bits && new->devnum >=27 *26 )) { -mutex_unlock (&blktrans_ref_mutex); + (tr->part_bits && new->devnum >=27 *26 ))return ret; - }list_add_tail (&new->list, &tr->devs); added: -mutex_unlock (&blktrans_ref_mutex);mutex_init (&new->lock);kref_init (&new->ref);
Signature · 6 lines
2.30.2 ______________________________________________________ Linux MTD discussion mailing list http://lists.infradead.org/mailman/listinfo/linux-mtd/
Re: [PATCH 6/8] mtd_blkdevs: remove blktrans_ref_mutex
Quoted 7 lines · Christoph Hellwig
blktrans_ref_mutexis not actually needed. The kref is serialized internally, and devnum assignment inadd_mtd_blktrans_devhappens before the disk is added and thus any of theblock_device_operationsmethods otherwise using it are called. It is also already serialized by the globalmtd_table_mutex.Signed-off-by Christoph Hellwig <[email protected]>
Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git mtd/next, thanks.
Miquel
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
PATCH 7/8 mtd_blkdevs: simplify blktrans_getgeo
In reply to Christoph HellwigNo need to grab a mtd_blktrans_dev given that ->open already holds one and ->getgeo can only be called on an open disk.
--- drivers/mtd/mtd_blkdevs.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/mtd/mtd_blkdevs.c +1 -2
@@ -257,7 +257,7 @@ static void blktrans_release(struct gendisk *disk, fmode_t mode)static int blktrans_getgeo (struct block_device *bdev,struct hd_geometry *geo) { -struct mtd_blktrans_dev *dev =blktrans_dev_get (bdev->bd_disk); +struct mtd_blktrans_dev *dev = bdev->bd_disk->private_data;int ret = -ENXIO ;mutex_lock (&dev->lock);
@@ -268,7 +268,6 @@ static int blktrans_getgeo(struct block_device *bdev, struct hd_geometry *geo) ret = dev->tr->getgeo ? dev->tr->getgeo (dev, geo) : -ENOTTY ; unlock:mutex_unlock (&dev->lock); -blktrans_dev_put (dev);return ret; }
Signature · 6 lines
2.30.2 ______________________________________________________ Linux MTD discussion mailing list http://lists.infradead.org/mailman/listinfo/linux-mtd/
Re: [PATCH 7/8] mtd_blkdevs: simplify blktrans_getgeo
Quoted 4 lines · Christoph Hellwig
No need to grab a
mtd_blktrans_devgiven that ->open already holds one and ->getgeo can only be called on an open disk.Signed-off-by Christoph Hellwig <[email protected]>
Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git mtd/next, thanks.
Miquel
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
PATCH 8/8 mtd_blkdevs: simplify the refcounting in blktrans_{open, release}
In reply to Christoph HellwigAlways grab a reference to the mtd_blktrans_dev in ->open instead of just on the first open, and do away with the additional temporary references in ->open and ->release.
--- drivers/mtd/mtd_blkdevs.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)
drivers/mtd/mtd_blkdevs.c +4 -14
@@ -36,14 +36,6 @@ static void blktrans_dev_release(struct kref *kref)kfree (dev); } -static struct mtd_blktrans_dev *blktrans_dev_get (struct gendisk *disk) -{ -struct mtd_blktrans_dev *dev = disk->private_data; - -kref_get (&dev->ref); -return dev; -} -static void blktrans_dev_put (struct mtd_blktrans_dev *dev) {kref_put (&dev->ref, blktrans_dev_release);
@@ -191,15 +183,16 @@ static blk_status_t mtd_queue_rq(struct blk_mq_hw_ctx *hctx,static int blktrans_open (struct block_device *bdev,fmode_t mode) { -struct mtd_blktrans_dev *dev =blktrans_dev_get (bdev->bd_disk); +struct mtd_blktrans_dev *dev = bdev->bd_disk->private_data;int ret =0 ; +kref_get (&dev->ref); +mutex_lock (&dev->lock);if (dev->open) goto unlock; -kref_get (&dev->ref);__module_get (dev->tr->owner);if (!dev->mtd)
@@ -219,7 +212,6 @@ static int blktrans_open(struct block_device *bdev, fmode_t mode) unlock: dev->open++;mutex_unlock (&dev->lock); -blktrans_dev_put (dev);return ret; error_release:
@@ -227,7 +219,6 @@ static int blktrans_open(struct block_device *bdev, fmode_t mode) dev->tr->release (dev); error_put:module_put (dev->tr->owner); -kref_put (&dev->ref, blktrans_dev_release);mutex_unlock (&dev->lock);blktrans_dev_put (dev);return ret;
@@ -235,14 +226,13 @@ static int blktrans_open(struct block_device *bdev, fmode_t mode)static void blktrans_release (struct gendisk *disk,fmode_t mode) { -struct mtd_blktrans_dev *dev =blktrans_dev_get (disk); +struct mtd_blktrans_dev *dev = disk->private_data;mutex_lock (&dev->lock);if (--dev->open) goto unlock; -kref_put (&dev->ref, blktrans_dev_release);module_put (dev->tr->owner);if (dev->mtd) {
Signature · 6 lines
2.30.2 ______________________________________________________ Linux MTD discussion mailing list http://lists.infradead.org/mailman/listinfo/linux-mtd/
Re: [PATCH 8/8] mtd_blkdevs: simplify the refcounting in blktrans_{open, release}
Quoted 5 lines · Christoph Hellwig
Always grab a reference to the
mtd_blktrans_devin ->open instead of just on the first open, and do away with the additional temporary references in ->open and ->release.Signed-off-by Christoph Hellwig <[email protected]>
Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git mtd/next, thanks.
Miquel
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/
Miquel Raynal
maintainerHi Christoph,
Christoph Hellwig <[email protected]> wrote on Mon, 23 Aug 2021 09:33:51 +0200:
Quoted 9 lines
Hi mtd maintainers,
the first series in this patch fixes a lock order reversal reported by Guenter based on a recent commit, although from code inspection it should have been around for much longer.
The rest is drive-by cleanups in the area that I noticed while trying to understand the locking changes. These are untested and need to be handled with care!
Thanks for the fix and the various cleanups, they look good to me. I don't know how important these drivers are nor if they are still actually used, it would be good to have some feedback from people still following. In the mean time I'll apply the series right away, I don't want to wait an additional month or so because of the merge window.
Thanks,
Miquèl
______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/