fix md disk_name lifetime problems v3
Subsystems (2)
Hi all,
this series tries to fix a problem repored by Logan where we see duplicate sysfs file name in md. It is due to the fact that the md driver only checks for duplicates on currently live mddevs, while the sysfs name can live on longer. It is an old problem, but the race window got longer due to waiting for the device freeze earlier in del_gendisk.
Changes since v2:
- delay mddex->kobj initialization
Changes since v1:
- rebased to the md-next branch
- fixed error handling in md_alloc for real
- add a little cleanup patch
Diffstat:
md.c | 304 +++++++++++++++++++++++++++++++++++-------------------------------- md.h | 1 2 files changed, 162 insertions(+), 143 deletions(-)
PATCH 1/10 md: fix mddev->kobj lifetime
Once a kobject is initialized, the containing object should not be directly freed. So delay initialization until it is added. Also remove the kobject_del call as the last put will remove the kobject as well. The explicitly delete isn't needed here, and dropping it will simplify further fixes.
--- drivers/md/md.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/md/md.c +1 -2
@@ -678,7 +678,6 @@ static void md_safemode_timeout(struct timer_list *t);void mddev_init (struct mddev *mddev) { -kobject_init (&mddev->kobj, &md_ktype);mutex_init (&mddev->open_mutex);mutex_init (&mddev->reconfig_mutex);mutex_init (&mddev->bitmap_info.mutex);
@@ -5617,7 +5616,6 @@ static void mddev_delayed_delete(struct work_struct *ws) { struct mddev *mddev = container_of(ws, struct mddev, del_work); -kobject_del (&mddev->kobj);kobject_put (&mddev->kobj); }
@@ -5719,6 +5717,7 @@ int md_alloc(dev_t dev, char *name)if (error) goto out_cleanup_disk; +kobject_init (&mddev->kobj, &md_ktype); error =kobject_add (&mddev->kobj, &disk_to_dev (disk)->kobj,"%s" ,"md" );if (error) goto out_del_gendisk;
Signature · 1 line
2.30.2
Re: [PATCH 01/10] md: fix mddev->kobj lifetime
Quoted 32 lines · Christoph Hellwig
Once a kobject is initialized, the containing object should not be directly freed. So delay initialization until it is added. Also remove the
kobject_delcall as the last put will remove the kobject as well. The explicitly delete isn't needed here, and dropping it will simplify further fixes.Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)drivers/md/md.c +0
-2@@ -678,7 +678,6 @@ static void md_safemode_timeout(struct timer_list *t);void mddev_init (struct mddev *mddev) { -kobject_init (&mddev->kobj, &md_ktype);mutex_init (&mddev->open_mutex);mutex_init (&mddev->reconfig_mutex);mutex_init (&mddev->bitmap_info.mutex);@@ -5617,7 +5616,6 @@ static void mddev_delayed_delete(struct work_struct *ws) { struct mddev *mddev = container_of(ws, struct mddev, del_work); -kobject_del (&mddev->kobj);kobject_put (&mddev->kobj); }
drivers/md/md.c int md_alloc(dev_t dev, char *name) · line 5723
Quoted 7 lines
drivers/md/md.c +1
-0@@ -5719,6 +5717,7 @@ int md_alloc(dev_t dev, char *name)if (error) goto out_cleanup_disk; +kobject_init (&mddev->kobj, &md_ktype); error =kobject_add (&mddev->kobj, &disk_to_dev (disk)->kobj,"%s" ,"md" );if (error) goto out_del_gendisk;
Cheers,
Hannes
Re: [PATCH 01/10] md: fix mddev->kobj lifetime
In reply to Christoph HellwigQuoted 7 lines · Christoph Hellwig
Once a kobject is initialized, the containing object should not be directly freed. So delay initialization until it is added. Also remove the
kobject_delcall as the last put will remove the kobject as well. The explicitly delete isn't needed here, and dropping it will simplify further fixes.Signed-off-by Christoph Hellwig <[email protected]>
Logan
PATCH 2/10 md: fix error handling in md_alloc
In reply to Christoph HellwigError handling in md_alloc is a mess. Untangle it to just free the mddev directly before add_disk is called and thus the gendisk is globally visible. After that clear the hold flag and let the mddev_put take care of cleaning up the mddev through the usual mechanisms.
--- drivers/md/md.c | 45 ++++++++++++++++++++++++++++++++------------- 1 file changed, 32 insertions(+), 13 deletions(-)
drivers/md/md.c +32 -13
@@ -790,6 +790,15 @@ static struct mddev *mddev_alloc(dev_t unit)return ERR_PTR (error); } +static void mddev_free (struct mddev *mddev) +{ +spin_lock (&all_mddevs_lock); +list_del (&mddev->all_mddevs); +spin_unlock (&all_mddevs_lock); + +kfree (mddev); +} +static const struct attribute_group md_redundancy_group;void mddev_unlock (struct mddev *mddev)
@@ -5662,8 +5671,8 @@ int md_alloc(dev_t dev, char *name)mutex_lock (&disks_mutex); mddev =mddev_alloc (dev);if (IS_ERR (mddev)) { -mutex_unlock (&disks_mutex); -return PTR_ERR (mddev); + error =PTR_ERR (mddev); + goto out_unlock; } partitioned = (MAJOR (mddev->unit) !=MD_MAJOR );
@@ -5681,7 +5690,7 @@ int md_alloc(dev_t dev, char *name) strcmp(mddev2->gendisk->disk_name, name) == 0) {spin_unlock (&all_mddevs_lock); error = -EEXIST ; - goto out_unlock_disks_mutex; + goto out_free_mddev; }spin_unlock (&all_mddevs_lock); }
@@ -5694,7 +5703,7 @@ int md_alloc(dev_t dev, char *name) error = -ENOMEM ; disk =blk_alloc_disk (NUMA_NO_NODE );if (!disk) - goto out_unlock_disks_mutex; + goto out_free_mddev; disk->major =MAJOR (mddev->unit); disk->first_minor = unit << shift;
@@ -5715,26 +5724,36 @@ int md_alloc(dev_t dev, char *name) mddev->gendisk = disk; error =add_disk (disk);if (error) - goto out_cleanup_disk; + goto out_put_disk;kobject_init (&mddev->kobj, &md_ktype); error =kobject_add (&mddev->kobj, &disk_to_dev (disk)->kobj,"%s" ,"md" ); -if (error) - goto out_del_gendisk; +if (error) { +/* +* The disk is already live at this point. Clear the hold flag +* and let mddev_put take care of the deletion, as it isn't any +* different from a normal close on last release now. +*/ + mddev->hold_active =0 ; + goto done; + }kobject_uevent (&mddev->kobj,KOBJ_ADD ); mddev->sysfs_state =sysfs_get_dirent_safe (mddev->kobj.sd,"array_state" ); mddev->sysfs_level =sysfs_get_dirent_safe (mddev->kobj.sd,"level" ); - goto out_unlock_disks_mutex; -out_del_gendisk: -del_gendisk (disk); -out_cleanup_disk: -blk_cleanup_disk (disk); -out_unlock_disks_mutex: +done:mutex_unlock (&disks_mutex);mddev_put (mddev);return error; + +out_put_disk: +put_disk (disk); +out_free_mddev: +mddev_free (mddev); +out_unlock: +mutex_unlock (&disks_mutex); +return error; }static void md_probe (dev_t dev)
Signature · 1 line
2.30.2
Re: [PATCH 02/10] md: fix error handling in md_alloc
Quoted 5 lines · Christoph Hellwig
Error handling in
md_allocis a mess. Untangle it to just free the mddev directly beforeadd_diskis called and thus the gendisk is globally visible. After that clear the hold flag and let themddev_puttake care of cleaning up the mddev through the usual mechanisms.
Quoted 8 lines
Fixes 5e55e2f5fc95 ("[PATCH] md: convert compile time warnings into runtime warnings")Fixes 9be68dd7ac0e ("md: add error handling support for add_disk()")Fixes 7ad1069166c0 ("md: properly unwind when failing to add the kobject in md_alloc")Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 45 ++++++++++++++++++++++++++++++++------------- 1 file changed, 32 insertions(+), 13 deletions(-)
Cheers,
Hannes
Re: [PATCH 02/10] md: fix error handling in md_alloc
In reply to Christoph HellwigQuoted 9 lines · Christoph Hellwig
Error handling in
md_allocis a mess. Untangle it to just free the mddev directly beforeadd_diskis called and thus the gendisk is globally visible. After that clear the hold flag and let themddev_puttake care of cleaning up the mddev through the usual mechanisms.Fixes 5e55e2f5fc95 ("[PATCH] md: convert compile time warnings into runtime warnings")Fixes 9be68dd7ac0e ("md: add error handling support for add_disk()")Fixes 7ad1069166c0 ("md: properly unwind when failing to add the kobject in md_alloc")Signed-off-by Christoph Hellwig <[email protected]>
Logan
PATCH 3/10 md: implement ->free_disk
In reply to Christoph HellwigEnsure that all private data is only freed once all accesses are done.
--- drivers/md/md.c | 17 ++++++++++++----- 1 file changed, 12 insertions(+), 5 deletions(-)
drivers/md/md.c +12 -5
@@ -5602,11 +5602,6 @@ static void md_free(struct kobject *ko)del_gendisk (mddev->gendisk);blk_cleanup_disk (mddev->gendisk); } -percpu_ref_exit (&mddev->writes_pending); - -bioset_exit (&mddev->bio_set); -bioset_exit (&mddev->sync_set); -kfree (mddev); }static const struct sysfs_ops md_sysfs_ops = {
@@ -7876,6 +7871,17 @@ static unsigned int md_check_events(struct gendisk *disk, unsigned int clearing)return ret; } +static void md_free_disk (struct gendisk *disk) +{ +struct mddev *mddev = disk->private_data; + +percpu_ref_exit (&mddev->writes_pending); +bioset_exit (&mddev->bio_set); +bioset_exit (&mddev->sync_set); + +kfree (mddev); +} +const struct block_device_operations md_fops = { .owner =THIS_MODULE ,
@@ -7889,6 +7895,7 @@ const struct block_device_operations md_fops = .getgeo = md_getgeo, .check_events = md_check_events, .set_read_only = md_set_read_only, + .free_disk = md_free_disk, };static int md_thread (void *arg)
Signature · 1 line
2.30.2
Re: [PATCH 03/10] md: implement ->free_disk
Quoted 7 lines · Christoph Hellwig
Ensure that all private data is only freed once all accesses are done.
Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 17 ++++++++++++----- 1 file changed, 12 insertions(+), 5 deletions(-)
Cheers,
Hannes
Re: [PATCH 03/10] md: implement ->free_disk
In reply to Christoph HellwigQuoted 12 lines · Christoph Hellwig
Ensure that all private data is only freed once all accesses are done.
Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 17 ++++++++++++----- 1 file changed, 12 insertions(+), 5 deletions(-)drivers/md/md.c +0
-0
drivers/md/md.c static void md_free(struct kobject *ko) · line 5604
Quoted 3 lines
drivers/md/md.c +0
-0@@ -5602,11 +5602,6 @@ static void md_free(struct kobject *ko)del_gendisk (mddev->gendisk);blk_cleanup_disk (mddev->gendisk); }
Looks like, with the previous patch, we can now remove the conditional on gendisk. Other than that it looks good:
Logan
PATCH 4/10 md: rename md_free to md_kobj_release
In reply to Christoph HellwigThe md_free name is rather misleading, so pick a better one.
--- drivers/md/md.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/md/md.c +2 -2
@@ -5589,7 +5589,7 @@ md_attr_store(struct kobject *kobj, struct attribute *attr,return rv; } -static void md_free (struct kobject *ko) +static void md_kobj_release (struct kobject *ko) { struct mddev *mddev = container_of(ko, struct mddev, kobj);
@@ -5609,7 +5609,7 @@ static const struct sysfs_ops md_sysfs_ops = { .store = md_attr_store, };static struct kobj_type md_ktype = { - .release = md_free, + .release = md_kobj_release, .sysfs_ops = &md_sysfs_ops, .default_groups = md_attr_groups, };
Signature · 1 line
2.30.2
Re: [PATCH 04/10] md: rename md_free to md_kobj_release
Quoted 21 lines · Christoph Hellwig
The
md_freename is rather misleading, so pick a better one.Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)drivers/md/md.c +1
-1@@ -5589,7 +5589,7 @@ md_attr_store(struct kobject *kobj, struct attribute *attr,return rv; } -static void md_free (struct kobject *ko) +static void md_kobj_release (struct kobject *ko) { struct mddev *mddev = container_of(ko, struct mddev, kobj);
drivers/md/md.c static const struct sysfs_ops md_sysfs_ops = { · line 5615
Quoted 8 lines
drivers/md/md.c +1
-1@@ -5609,7 +5609,7 @@ static const struct sysfs_ops md_sysfs_ops = { .store = md_attr_store, };static struct kobj_type md_ktype = { - .release = md_free, + .release = md_kobj_release, .sysfs_ops = &md_sysfs_ops, .default_groups = md_attr_groups, };
Cheers,
Hannes
Re: [PATCH 04/10] md: rename md_free to md_kobj_release
In reply to Christoph HellwigQuoted 3 lines · Christoph Hellwig
The
md_freename is rather misleading, so pick a better one.Signed-off-by Christoph Hellwig <[email protected]>
Logan
PATCH 5/10 md: factor out the rdev overlaps check from rdev_size_store
In reply to Christoph HellwigThis splits the code into nicely readable chunks and also avoids the refcount inc/dec manipulations.
--- drivers/md/md.c | 84 +++++++++++++++++++++++-------------------------- 1 file changed, 39 insertions(+), 45 deletions(-)
drivers/md/md.c +39 -45
@@ -3343,14 +3343,33 @@ rdev_size_show(struct md_rdev *rdev, char *page)return sprintf (page,"%llu\n" , (unsigned long long )rdev->sectors /2 ); } -static int overlaps (sector_t s1,sector_t l1,sector_t s2,sector_t l2) +static int md_rdevs_overlap (struct md_rdev *a,struct md_rdev *b) {/* check if two start/length pairs overlap */ -if (s1+l1 <= s2) -return 0 ; -if (s2+l2 <= s1) -return 0 ; -return 1 ; +if (a->data_offset + a->sectors <= b->data_offset) +return false; +if (b->data_offset + b->sectors <= a->data_offset) +return false; +return true; +} + +static bool md_rdev_overlaps (struct md_rdev *rdev) +{ +struct mddev *mddev; +struct md_rdev *rdev2; + +spin_lock (&all_mddevs_lock); +list_for_each_entry (mddev, &all_mddevs, all_mddevs) { +rdev_for_each (rdev2, mddev) { +if (rdev != rdev2 && rdev->bdev == rdev2->bdev && +md_rdevs_overlap (rdev, rdev2)) { +spin_unlock (&all_mddevs_lock); +return true; + } + } + } +spin_unlock (&all_mddevs_lock); +return false; }static int strict_blocks_to_sectors (const char *buf,sector_t *sectors)
@@ -3402,46 +3421,21 @@ rdev_size_store(struct md_rdev *rdev, const char *buf, size_t len)return -EINVAL ;/* component must fit device */ rdev->sectors = sectors; -if (sectors > oldsectors && my_mddev->external) { -/* Need to check that all other rdevs with the same -* ->bdev do not overlap. 'rcu' is sufficient to walk -* the rdev lists safely. -* This check does not provide a hard guarantee, it -* just helps avoid dangerous mistakes. -*/ -struct mddev *mddev; -int overlap =0 ; -struct list_head *tmp; -rcu_read_lock (); -for_each_mddev (mddev, tmp) { -struct md_rdev *rdev2; - -rdev_for_each (rdev2, mddev) -if (rdev->bdev == rdev2->bdev && - rdev != rdev2 && -overlaps (rdev->data_offset, rdev->sectors, - rdev2->data_offset, - rdev2->sectors)) { - overlap =1 ; -break ; - } -if (overlap) { -mddev_put (mddev); -break ; - } - } -rcu_read_unlock (); -if (overlap) { -/* Someone else could have slipped in a size -* change here, but doing so is just silly. -* We put oldsectors back because we *know* it is -* safe, and trust userspace not to race with -* itself -*/ - rdev->sectors = oldsectors; -return -EBUSY ; - } +/* +* Check that all other rdevs with the same bdev do not overlap. This +* check does not provide a hard guarantee, it just helps avoid +* dangerous mistakes. +*/ +if (sectors > oldsectors && my_mddev->external && +md_rdev_overlaps (rdev)) { +/* +* Someone else could have slipped in a size change here, but +* doing so is just silly. We put oldsectors back because we +* know it is safe, and trust userspace not to race with itself. +*/ + rdev->sectors = oldsectors; +return -EBUSY ; }return len; }
Signature · 1 line
2.30.2
Re: [PATCH 05/10] md: factor out the rdev overlaps check from rdev_size_store
Quoted 8 lines · Christoph Hellwig
This splits the code into nicely readable chunks and also avoids the refcount inc/dec manipulations.
Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 84 +++++++++++++++++++++++-------------------------- 1 file changed, 39 insertions(+), 45 deletions(-)
Cheers,
Hannes
Re: [PATCH 05/10] md: factor out the rdev overlaps check from rdev_size_store
In reply to Christoph HellwigQuoted 4 lines · Christoph Hellwig
This splits the code into nicely readable chunks and also avoids the refcount inc/dec manipulations.
Signed-off-by Christoph Hellwig <[email protected]>
Logan
PATCH 6/10 md: stop using for_each_mddev in md_do_sync
In reply to Christoph HellwigJust do a plain list_for_each that only grabs a mddev reference in the case where the thread sleeps and restarts the list iteration.
--- drivers/md/md.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/md/md.c +5 -3
@@ -8725,7 +8725,6 @@ void md_do_sync(struct md_thread *thread)unsigned long update_time;sector_t mark_cnt[SYNC_MARKS ];int last_mark,m; -struct list_head *tmp;sector_t last_check;int skipped =0 ;struct md_rdev *rdev;
@@ -8789,7 +8788,8 @@ void md_do_sync(struct md_thread *thread) try_again:if (test_bit (MD_RECOVERY_INTR , &mddev->recovery)) goto skip; -for_each_mddev (mddev2, tmp) { +spin_lock (&all_mddevs_lock); +list_for_each_entry (mddev2, &all_mddevs, all_mddevs) {if (mddev2 == mddev)continue ; if (!mddev->parallel_resync
@@ -8821,7 +8821,8 @@ void md_do_sync(struct md_thread *thread) desc,mdname (mddev), mdname(mddev2)); } -mddev_put (mddev2); +spin_unlock (&all_mddevs_lock); +if (signal_pending (current))flush_signals (current);schedule ();
@@ -8831,6 +8832,7 @@ void md_do_sync(struct md_thread *thread)finish_wait (&resync_wait, &wq); } } +spin_unlock (&all_mddevs_lock); }while (mddev->curr_resync <MD_RESYNC_DELAYED ); j =0 ;
Signature · 1 line
2.30.2
Re: [PATCH 06/10] md: stop using for_each_mddev in md_do_sync
Quoted 8 lines · Christoph Hellwig
Just do a plain
list_for_eachthat only grabs a mddev reference in the case where the thread sleeps and restarts the list iteration.Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
Cheers,
Hannes
Re: [PATCH 06/10] md: stop using for_each_mddev in md_do_sync
In reply to Christoph HellwigQuoted 4 lines · Christoph Hellwig
Just do a plain
list_for_eachthat only grabs a mddev reference in the case where the thread sleeps and restarts the list iteration.Signed-off-by Christoph Hellwig <[email protected]>
Logan
PATCH 7/10 md: stop using for_each_mddev in md_notify_reboot
In reply to Christoph HellwigJust do a simple list_for_each_entry_safe on all_mddevs, and only grab a reference when we drop the lock.
--- drivers/md/md.c | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-)
drivers/md/md.c +9 -3
@@ -9587,11 +9587,13 @@ EXPORT_SYMBOL_GPL(rdev_clear_badblocks);static int md_notify_reboot (struct notifier_block *this,unsigned long code,void *x) { -struct list_head *tmp; -struct mddev *mddev; +struct mddev *mddev, *n;int need_delay =0 ; -for_each_mddev (mddev, tmp) { +spin_lock (&all_mddevs_lock); +list_for_each_entry_safe (mddev, n, &all_mddevs, all_mddevs) { +mddev_get (mddev); +spin_unlock (&all_mddevs_lock);if (mddev_trylock (mddev)) {if (mddev->pers)__md_stop_writes (mddev);
@@ -9600,7 +9602,11 @@ static int md_notify_reboot(struct notifier_block *this,mddev_unlock (mddev); } need_delay =1 ; +mddev_put (mddev); +spin_lock (&all_mddevs_lock); } +spin_unlock (&all_mddevs_lock); +/* * certain more exotic SCSI devices are known to be * volatile wrt too early system reboots. While the
Signature · 1 line
2.30.2
Re: [PATCH 07/10] md: stop using for_each_mddev in md_notify_reboot
Quoted 31 lines · Christoph Hellwig
Just do a simple
list_for_each_entry_safeonall_mddevs, and only grab a reference when we drop the lock.Reviewed-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-)drivers/md/md.c +5
-3@@ -9587,11 +9587,13 @@ EXPORT_SYMBOL_GPL(rdev_clear_badblocks);static int md_notify_reboot (struct notifier_block *this,unsigned long code,void *x) { -struct list_head *tmp; -struct mddev *mddev; +struct mddev *mddev, *n;int need_delay =0 ; -for_each_mddev (mddev, tmp) { +spin_lock (&all_mddevs_lock); +list_for_each_entry_safe (mddev, n, &all_mddevs, all_mddevs) { +mddev_get (mddev); +spin_unlock (&all_mddevs_lock);if (mddev_trylock (mddev)) {if (mddev->pers)__md_stop_writes (mddev);@@ -9600,7 +9602,11 @@ static int md_notify_reboot(struct notifier_block *this,mddev_unlock (mddev);
drivers/md/md.c static int md_notify_reboot(struct notifier_block *this, · line 9612
Quoted 10 lines
drivers/md/md.c +4
-0} need_delay =1 ; +mddev_put (mddev); +spin_lock (&all_mddevs_lock); } +spin_unlock (&all_mddevs_lock); +/* * certain more exotic SCSI devices are known to be * volatile wrt too early system reboots. While the
I wish 'mddev_get()' could fail if 'mddev->active' is zero when we call it ... but maybe for another time.
Cheers,
Hannes
Re: [PATCH 07/10] md: stop using for_each_mddev in md_notify_reboot
In reply to Christoph HellwigQuoted 4 lines · Christoph Hellwig
Just do a simple
list_for_each_entry_safeonall_mddevs, and only grab a reference when we drop the lock.Reviewed-by Christoph Hellwig <[email protected]>
Logan
PATCH 8/10 md: stop using for_each_mddev in md_exit
In reply to Christoph HellwigJust do a simple list_for_each_entry_safe on all_mddevs, and only grab a reference when we drop the lock and delete the now unused for_each_mddev macro.
--- drivers/md/md.c | 39 +++++++++++---------------------------- 1 file changed, 11 insertions(+), 28 deletions(-)
drivers/md/md.c +11 -28
@@ -368,28 +368,6 @@ EXPORT_SYMBOL_GPL(md_new_event);static LIST_HEAD (all_mddevs);static DEFINE_SPINLOCK (all_mddevs_lock); -/* -* iterates through all used mddevs in the system. -* We take care to grab the all_mddevs_lock whenever navigating -* the list, and to always hold a refcount when unlocked. -* Any code which breaks out of this loop while own -* a reference to the current mddev and must mddev_put it. -*/ -#define for_each_mddev (_mddev,_tmp) \ - \ - for (({ spin_lock(&all_mddevs_lock); \ - _tmp = all_mddevs.next; \ - _mddev = NULL;}); \ - ({ if (_tmp != &all_mddevs) \ - mddev_get(list_entry(_tmp, struct mddev, all_mddevs));\ - spin_unlock(&all_mddevs_lock); \ - if (_mddev) mddev_put(_mddev); \ - _mddev = list_entry(_tmp, struct mddev, all_mddevs); \ - _tmp != &all_mddevs;}); \ - ({ spin_lock(&all_mddevs_lock); \ - _tmp = _tmp->next;}) \ - ) -/* Rather than calling directly into the personality make_request function, * IO requests come here first so that we can check if the device is * being suspended pending a reconfiguration.
@@ -9925,8 +9903,7 @@ void md_autostart_arrays(int part) static __exit void md_exit(void) { -struct mddev *mddev; -struct list_head *tmp; +struct mddev *mddev, *n;int delay =1 ;unregister_blkdev (MD_MAJOR ,"md" );
@@ -9946,17 +9923,23 @@ static __exit void md_exit(void) }remove_proc_entry ("mdstat" ,NULL ); -for_each_mddev (mddev, tmp) { +spin_lock (&all_mddevs_lock); +list_for_each_entry_safe (mddev, n, &all_mddevs, all_mddevs) { +mddev_get (mddev); +spin_unlock (&all_mddevs_lock);export_array (mddev); mddev->ctime =0 ; mddev->hold_active =0 ;/* -* for_each_mddev() will call mddev_put() at the end of each -* iteration. As the mddev is now fully clear, this will -* schedule the mddev for destruction by a workqueue, and the +* As the mddev is now fully clear, mddev_put will schedule +* the mddev for destruction by a workqueue, and the * destroy_workqueue() below will wait for that to complete. */ +mddev_put (mddev); +spin_lock (&all_mddevs_lock); } +spin_unlock (&all_mddevs_lock); +destroy_workqueue (md_rdev_misc_wq);destroy_workqueue (md_misc_wq);destroy_workqueue (md_wq);
Signature · 1 line
2.30.2
Re: [PATCH 08/10] md: stop using for_each_mddev in md_exit
Quoted 70 lines · Christoph Hellwig
Just do a simple
list_for_each_entry_safeonall_mddevs, and only grab a reference when we drop the lock and delete the now unusedfor_each_mddevmacro.Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 39 +++++++++++---------------------------- 1 file changed, 11 insertions(+), 28 deletions(-)drivers/md/md.c +7
-28@@ -368,28 +368,6 @@ EXPORT_SYMBOL_GPL(md_new_event);static LIST_HEAD (all_mddevs);static DEFINE_SPINLOCK (all_mddevs_lock); -/* -* iterates through all used mddevs in the system. -* We take care to grab the all_mddevs_lock whenever navigating -* the list, and to always hold a refcount when unlocked. -* Any code which breaks out of this loop while own -* a reference to the current mddev and must mddev_put it. -*/ -#define for_each_mddev (_mddev,_tmp) \ - \ - for (({ spin_lock(&all_mddevs_lock); \ - _tmp = all_mddevs.next; \ - _mddev = NULL;}); \ - ({ if (_tmp != &all_mddevs) \ - mddev_get(list_entry(_tmp, struct mddev, all_mddevs));\ - spin_unlock(&all_mddevs_lock); \ - if (_mddev) mddev_put(_mddev); \ - _mddev = list_entry(_tmp, struct mddev, all_mddevs); \ - _tmp != &all_mddevs;}); \ - ({ spin_lock(&all_mddevs_lock); \ - _tmp = _tmp->next;}) \ - ) -/* Rather than calling directly into the personality make_request function, * IO requests come here first so that we can check if the device is * being suspended pending a reconfiguration. @@ -9925,8 +9903,7 @@ void md_autostart_arrays(int part) static __exit void md_exit(void) { -struct mddev *mddev; -struct list_head *tmp; +struct mddev *mddev, *n;int delay =1 ;unregister_blkdev (MD_MAJOR ,"md" );@@ -9946,17 +9923,23 @@ static __exit void md_exit(void) }remove_proc_entry ("mdstat" ,NULL ); -for_each_mddev (mddev, tmp) { +spin_lock (&all_mddevs_lock); +list_for_each_entry_safe (mddev, n, &all_mddevs, all_mddevs) { +mddev_get (mddev); +spin_unlock (&all_mddevs_lock);export_array (mddev); mddev->ctime =0 ; mddev->hold_active =0 ;/* -* for_each_mddev() will call mddev_put() at the end of each -* iteration. As the mddev is now fully clear, this will -* schedule the mddev for destruction by a workqueue, and the +* As the mddev is now fully clear, mddev_put will schedule +* the mddev for destruction by a workqueue, and the
drivers/md/md.c static __exit void md_exit(void) · line 9945
Quoted 10 lines
drivers/md/md.c +4
-0* destroy_workqueue() below will wait for that to complete. */ +mddev_put (mddev); +spin_lock (&all_mddevs_lock); } +spin_unlock (&all_mddevs_lock); +destroy_workqueue (md_rdev_misc_wq);destroy_workqueue (md_misc_wq);destroy_workqueue (md_wq);
Having thought about it some more ... wouldn't it make sense to modify mddev_get() to
if (atomic_inc_not_zero(&mddev->active))
return NULL;
to ensure we're not missing any use-after-free issues, which previously would have been caught by the 'for_each_mddev()' macro?
Cheers,
Hannes
Re: [PATCH 08/10] md: stop using for_each_mddev in md_exit
Quoted 9 lines · Hannes Reinecke
Having thought about it some more ... wouldn't it make sense to modify
mddev_get()toif (
atomic_inc_not_zero(&mddev->active))return NULL;to ensure we're not missing any use-after-free issues, which previously would have been caught by the '
for_each_mddev()' macro?
No. mddev->active = 0 can still be a perfectly valid mddev and they are not automatically deleted. Check the conditions in mddev_put.
Hannes Reinecke
Reviewed-byRe: [PATCH 08/10] md: stop using for_each_mddev in md_exit
Quoted 5 lines · Christoph Hellwig
Quoted 4 lines · Hannes Reinecke
Having thought about it some more ... wouldn't it make sense to modify
mddev_get()to
Quoted 8 lines
Quoted 5 lines
if (
atomic_inc_not_zero(&mddev->active))return NULL;to ensure we're not missing any use-after-free issues, which previously would have been caught by the '
for_each_mddev()' macro?No.
mddev->active= 0 can still be a perfectly valid mddev and they are not automatically deleted. Check the conditions inmddev_put.
Sigh. Okay.
MD is a mess anyway.
Cheers,
Hannes
Re: [PATCH 08/10] md: stop using for_each_mddev in md_exit
In reply to Christoph HellwigQuoted 5 lines · Christoph Hellwig
Just do a simple
list_for_each_entry_safeonall_mddevs, and only grab a reference when we drop the lock and delete the now unusedfor_each_mddevmacro.Signed-off-by Christoph Hellwig <[email protected]>
Logan
PATCH 9/10 md: only delete entries from all_mddevs when the disk is freed
In reply to Christoph HellwigThis ensures device names don't get prematurely reused. Instead add a deleted flag to skip already deleted devices in mddev_get and other places that only want to see live mddevs.
Reported-by; Logan Gunthorpe <[email protected]>
--- drivers/md/md.c | 56 +++++++++++++++++++++++++++++++++---------------- drivers/md/md.h | 1 + 2 files changed, 39 insertions(+), 18 deletions(-)
drivers/md/md.c +38 -18
@@ -625,6 +625,10 @@ EXPORT_SYMBOL(md_flush_request);static inline struct mddev *mddev_get (struct mddev *mddev) { +lockdep_assert_held (&all_mddevs_lock); + +if (mddev->deleted) +return NULL ;atomic_inc (&mddev->active);return mddev; }
@@ -639,7 +643,7 @@ static void mddev_put(struct mddev *mddev) mddev->ctime == 0 && !mddev->hold_active) {/* Array is not configured at all, and not held active, * so destroy it */ -list_del_init (&mddev->all_mddevs); + mddev->deleted = true;/* * Call queue_work inside the spinlock so that
@@ -719,8 +723,8 @@ static struct mddev *mddev_find(dev_t unit)spin_lock (&all_mddevs_lock); mddev =mddev_find_locked (unit); -if (mddev) -mddev_get (mddev); +if (mddev && !mddev_get (mddev)) + mddev =NULL ;spin_unlock (&all_mddevs_lock);return mddev;
@@ -3338,6 +3342,8 @@ static bool md_rdev_overlaps(struct md_rdev *rdev)spin_lock (&all_mddevs_lock);list_for_each_entry (mddev, &all_mddevs, all_mddevs) { +if (mddev->deleted) +continue ;rdev_for_each (rdev2, mddev) {if (rdev != rdev2 && rdev->bdev == rdev2->bdev &&md_rdevs_overlap (rdev, rdev2)) {
@@ -5525,11 +5531,10 @@ md_attr_show(struct kobject *kobj, struct attribute *attr, char *page)if (!entry->show)return -EIO ;spin_lock (&all_mddevs_lock); -if (list_empty (&mddev->all_mddevs)) { +if (!mddev_get (mddev)) {spin_unlock (&all_mddevs_lock);return -EBUSY ; } -mddev_get (mddev);spin_unlock (&all_mddevs_lock); rv = entry->show (mddev, page);
@@ -5550,11 +5555,10 @@ md_attr_store(struct kobject *kobj, struct attribute *attr,if (!capable (CAP_SYS_ADMIN ))return -EACCES ;spin_lock (&all_mddevs_lock); -if (list_empty (&mddev->all_mddevs)) { +if (!mddev_get (mddev)) {spin_unlock (&all_mddevs_lock);return -EBUSY ; } -mddev_get (mddev);spin_unlock (&all_mddevs_lock); rv = entry->store (mddev, page, length);mddev_put (mddev);
@@ -7851,7 +7855,7 @@ static void md_free_disk(struct gendisk *disk)bioset_exit (&mddev->bio_set);bioset_exit (&mddev->sync_set); -kfree (mddev); +mddev_free (mddev); } const struct block_device_operations md_fops =
@@ -8173,6 +8177,8 @@ static void *md_seq_start(struct seq_file *seq, loff_t *pos)if (!l--) { mddev = list_entry(tmp, struct mddev, all_mddevs);mddev_get (mddev); +if (!mddev_get (mddev)) +continue ;spin_unlock (&all_mddevs_lock);return mddev; }
@@ -8186,25 +8192,35 @@ static void *md_seq_next(struct seq_file *seq, void *v, loff_t *pos) {struct list_head *tmp;struct mddev *next_mddev, *mddev = v; +struct mddev *to_put =NULL ; ++*pos;if (v == (void *)2 )return NULL ;spin_lock (&all_mddevs_lock); -if (v == (void *)1 ) +if (v == (void *)1 ) { tmp = all_mddevs.next; -else + }else { + to_put = mddev; tmp = mddev->all_mddevs.next; -if (tmp != &all_mddevs) - next_mddev = mddev_get(list_entry(tmp,struct mddev,all_mddevs)); -else { - next_mddev = (void *)2 ; - *pos =0x10000 ; } + +for (;;) { +if (tmp == &all_mddevs) { + next_mddev = (void *)2 ; + *pos =0x10000 ; +break ; + } + next_mddev = list_entry(tmp, struct mddev, all_mddevs); +if (mddev_get (next_mddev)) +break ; + mddev = next_mddev; + tmp = mddev->all_mddevs.next; + };spin_unlock (&all_mddevs_lock); -if (v != (void *)1 ) +if (to_put)mddev_put (mddev);return next_mddev;
@@ -8768,6 +8784,8 @@ void md_do_sync(struct md_thread *thread) goto skip;spin_lock (&all_mddevs_lock);list_for_each_entry (mddev2, &all_mddevs, all_mddevs) { +if (mddev2->deleted) +continue ;if (mddev2 == mddev)continue ; if (!mddev->parallel_resync
@@ -9570,7 +9588,8 @@ static int md_notify_reboot(struct notifier_block *this,spin_lock (&all_mddevs_lock);list_for_each_entry_safe (mddev, n, &all_mddevs, all_mddevs) { -mddev_get (mddev); +if (!mddev_get (mddev)) +continue ;spin_unlock (&all_mddevs_lock);if (mddev_trylock (mddev)) {if (mddev->pers)
@@ -9925,7 +9944,8 @@ static __exit void md_exit(void)spin_lock (&all_mddevs_lock);list_for_each_entry_safe (mddev, n, &all_mddevs, all_mddevs) { -mddev_get (mddev); +if (!mddev_get (mddev)) +continue ;spin_unlock (&all_mddevs_lock);export_array (mddev); mddev->ctime =0 ;
drivers/md/md.h +1 -0
@@ -503,6 +503,7 @@ struct mddev {atomic_t max_corr_read_errors;/* max read retries */ struct list_head all_mddevs; +bool deleted;const struct attribute_group *to_remove;
Signature · 1 line
2.30.2
Re: [PATCH 09/10] md: only delete entries from all_mddevs when the disk is freed
Quoted 16 lines · Christoph Hellwig
This ensures device names don't get prematurely reused. Instead add a deleted flag to skip already deleted devices in
mddev_getand other places that only want to see live mddevs.Reported-by; Logan Gunthorpe <[email protected]>
Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 56 +++++++++++++++++++++++++++++++++---------------- drivers/md/md.h | 1 + 2 files changed, 39 insertions(+), 18 deletions(-)drivers/md/md.c +0
-0
drivers/md/md.c EXPORT_SYMBOL(md_flush_request); · line 634
Quoted 10 lines
drivers/md/md.c +4
-0@@ -625,6 +625,10 @@ EXPORT_SYMBOL(md_flush_request);static inline struct mddev *mddev_get (struct mddev *mddev) { +lockdep_assert_held (&all_mddevs_lock); + +if (mddev->deleted) +return NULL ;atomic_inc (&mddev->active);return mddev; }
Why can't we use 'atomic_inc_unless_zero' here and do away with the additional 'deleted' flag?
Quoted 21 lines
drivers/md/md.c +3
-3@@ -639,7 +643,7 @@ static void mddev_put(struct mddev *mddev) mddev->ctime == 0 && !mddev->hold_active) {/* Array is not configured at all, and not held active, * so destroy it */ -list_del_init (&mddev->all_mddevs); + mddev->deleted = true;/* * Call queue_work inside the spinlock so that @@ -719,8 +723,8 @@ static struct mddev *mddev_find(dev_t unit)spin_lock (&all_mddevs_lock); mddev =mddev_find_locked (unit); -if (mddev) -mddev_get (mddev); +if (mddev && !mddev_get (mddev)) + mddev =NULL ;spin_unlock (&all_mddevs_lock);return mddev;
drivers/md/md.c static bool md_rdev_overlaps(struct md_rdev *rdev) · line 3346
Quoted 5 lines
drivers/md/md.c +2
-0@@ -3338,6 +3342,8 @@ static bool md_rdev_overlaps(struct md_rdev *rdev)spin_lock (&all_mddevs_lock);list_for_each_entry (mddev, &all_mddevs, all_mddevs) { +if (mddev->deleted) +continue ;
Any particular reason why you can't use the 'mddev_get()' / 'mddev_put()' sequence here? After all, I don't think that 'mddev' should vanish while being in this loop, no?
Quoted 124 lines
drivers/md/md.c +29
-15rdev_for_each (rdev2, mddev) {if (rdev != rdev2 && rdev->bdev == rdev2->bdev &&md_rdevs_overlap (rdev, rdev2)) {@@ -5525,11 +5531,10 @@ md_attr_show(struct kobject *kobj, struct attribute *attr, char *page)if (!entry->show)return -EIO ;spin_lock (&all_mddevs_lock); -if (list_empty (&mddev->all_mddevs)) { + if (!mddev_get(mddev)) { > spin_unlock(&all_mddevs_lock);return -EBUSY ; } -mddev_get (mddev);spin_unlock (&all_mddevs_lock); rv = entry->show (mddev, page);@@ -5550,11 +5555,10 @@ md_attr_store(struct kobject *kobj, struct attribute *attr,if (!capable (CAP_SYS_ADMIN ))return -EACCES ;spin_lock (&all_mddevs_lock); -if (list_empty (&mddev->all_mddevs)) { +if (!mddev_get (mddev)) {spin_unlock (&all_mddevs_lock);return -EBUSY ; } -mddev_get (mddev);spin_unlock (&all_mddevs_lock); rv = entry->store (mddev, page, length);mddev_put (mddev);@@ -7851,7 +7855,7 @@ static void md_free_disk(struct gendisk *disk)bioset_exit (&mddev->bio_set);bioset_exit (&mddev->sync_set); -kfree (mddev); +mddev_free (mddev); } const struct block_device_operations md_fops =@@ -8173,6 +8177,8 @@ static void *md_seq_start(struct seq_file *seq, loff_t *pos)if (!l--) { mddev = list_entry(tmp, struct mddev, all_mddevs);mddev_get (mddev); +if (!mddev_get (mddev)) +continue ;spin_unlock (&all_mddevs_lock);return mddev; }@@ -8186,25 +8192,35 @@ static void *md_seq_next(struct seq_file *seq, void *v, loff_t *pos) {struct list_head *tmp;struct mddev *next_mddev, *mddev = v; +struct mddev *to_put =NULL ; ++*pos;if (v == (void *)2 )return NULL ;spin_lock (&all_mddevs_lock); -if (v == (void *)1 ) +if (v == (void *)1 ) { tmp = all_mddevs.next; -else + }else { + to_put = mddev; tmp = mddev->all_mddevs.next; -if (tmp != &all_mddevs) - next_mddev = mddev_get(list_entry(tmp,struct mddev,all_mddevs)); -else { - next_mddev = (void *)2 ; - *pos =0x10000 ; } + +for (;;) { +if (tmp == &all_mddevs) { + next_mddev = (void *)2 ; + *pos =0x10000 ; +break ; + } + next_mddev = list_entry(tmp, struct mddev, all_mddevs); +if (mddev_get (next_mddev)) +break ; + mddev = next_mddev; + tmp = mddev->all_mddevs.next; + };spin_unlock (&all_mddevs_lock); -if (v != (void *)1 ) +if (to_put)mddev_put (mddev);return next_mddev;@@ -8768,6 +8784,8 @@ void md_do_sync(struct md_thread *thread) goto skip;spin_lock (&all_mddevs_lock);list_for_each_entry (mddev2, &all_mddevs, all_mddevs) { +if (mddev2->deleted) +continue ;if (mddev2 == mddev)continue ; if (!mddev->parallel_resync@@ -9570,7 +9588,8 @@ static int md_notify_reboot(struct notifier_block *this,spin_lock (&all_mddevs_lock);list_for_each_entry_safe (mddev, n, &all_mddevs, all_mddevs) { -mddev_get (mddev); +if (!mddev_get (mddev)) +continue ;spin_unlock (&all_mddevs_lock);if (mddev_trylock (mddev)) {if (mddev->pers)@@ -9925,7 +9944,8 @@ static __exit void md_exit(void)spin_lock (&all_mddevs_lock);list_for_each_entry_safe (mddev, n, &all_mddevs, all_mddevs) { -mddev_get (mddev); +if (!mddev_get (mddev)) +continue ;spin_unlock (&all_mddevs_lock);export_array (mddev); mddev->ctime =0 ;drivers/md/md.h +0
-0
drivers/md/md.h struct mddev { · line 509
Quoted 7 lines
drivers/md/md.h +1
-0@@ -503,6 +503,7 @@ struct mddev {atomic_t max_corr_read_errors;/* max read retries */ struct list_head all_mddevs; +bool deleted;const struct attribute_group *to_remove;
Cheers,
Hannes
Song Liu
maintainerRe: [PATCH 09/10] md: only delete entries from all_mddevs when the disk is freed
Quoted 19 lines · Hannes Reinecke
Quoted 17 lines · Christoph Hellwig
This ensures device names don't get prematurely reused. Instead add a deleted flag to skip already deleted devices in
mddev_getand other places that only want to see live mddevs.Reported-by; Logan Gunthorpe <[email protected]>
Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 56 +++++++++++++++++++++++++++++++++---------------- drivers/md/md.h | 1 + 2 files changed, 39 insertions(+), 18 deletions(-)drivers/md/md.c +0
-0@@ -625,6 +625,10 @@ EXPORT_SYMBOL(md_flush_request);
Quoted 12 lines
drivers/md/md.c EXPORT_SYMBOL(md_flush_request); · line 634
Quoted 9 lines
drivers/md/md.c +4
-0static inline struct mddev *mddev_get (struct mddev *mddev) { +lockdep_assert_held (&all_mddevs_lock); + +if (mddev->deleted) +return NULL ;atomic_inc (&mddev->active);return mddev; }Why can't we use '
atomic_inc_unless_zero' here and do away with the additional 'deleted' flag?
Thanks Christoph for the set. And thanks Hannes for the quick review!
I think atomic_inc_unless_zero() should work here. Alternatively, we can also grab a big from mddev->flags as MD_DELETED.
Thanks,
Song
[...]
Re: [PATCH 09/10] md: only delete entries from all_mddevs when the disk is freed
Quoted 10 lines · Song Liu
Quoted 10 lines
Quoted 9 lines
static inline struct mddev *mddev_get(struct mddev *mddev) {
drivers/md/md.c +4
-0+lockdep_assert_held (&all_mddevs_lock); + +if (mddev->deleted) +return NULL ;atomic_inc (&mddev->active);return mddev; }
Quoted 5 lines
Quoted 2 lines
Why can't we use '
atomic_inc_unless_zero' here and do away with the additional 'deleted' flag?I think
atomic_inc_unless_zero()should work here. Alternatively, we can also grab a big frommddev->flagsasMD_DELETED.
s/big/bit/ ?
Yes, this could use mddev->flags. I don't think atomic_inc_unless_zero would work, as an array can have a refcount of 0 but we still allow to take new references to it, the deletion only happens if the other conditions in mddev_put are met.
Do you want me to repin the series using a flag?
Song Liu
maintainerRe: [PATCH 09/10] md: only delete entries from all_mddevs when the disk is freed
Quoted 12 lines · Christoph Hellwig
Quoted 10 lines · Song Liu
Quoted 10 lines
Quoted 9 lines
static inline struct mddev *mddev_get(struct mddev *mddev) {
drivers/md/md.c +4
-0+lockdep_assert_held (&all_mddevs_lock); + +if (mddev->deleted) +return NULL ;atomic_inc (&mddev->active);return mddev; }
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Why can't we use '
atomic_inc_unless_zero' here and do away with the additional 'deleted' flag?I think
atomic_inc_unless_zero()should work here. Alternatively, we can also grab a big frommddev->flagsasMD_DELETED.s/big/bit/ ?
yeah, bit...
Quoted 7 lines
Yes, this could use
mddev->flags. I don't thinkatomic_inc_unless_zerowould work, as an array can have a refcount of 0 but we still allow to take new references to it, the deletion only happens if the other conditions inmddev_putare met.Do you want me to repin the series using a flag?
Respin the series sounds good. Please also address Hannes' feedback on 8/10.
Thanks,
Song
Re: [PATCH 09/10] md: only delete entries from all_mddevs when the disk is freed
In reply to Christoph HellwigQuoted 17 lines · Christoph Hellwig
Quoted 10 lines · Song Liu
Quoted 10 lines
Quoted 9 lines
static inline struct mddev *mddev_get(struct mddev *mddev) {
drivers/md/md.c +4
-0+lockdep_assert_held (&all_mddevs_lock); + +if (mddev->deleted) +return NULL ;atomic_inc (&mddev->active);return mddev; }Quoted 5 lines
Quoted 2 lines
Why can't we use '
atomic_inc_unless_zero' here and do away with the additional 'deleted' flag?I think
atomic_inc_unless_zero()should work here. Alternatively, we can also grab a big frommddev->flagsasMD_DELETED.
Quoted 8 lines
s/big/bit/ ?
Yes, this could use
mddev->flags. I don't thinkatomic_inc_unless_zerowould work, as an array can have a refcount of 0 but we still allow to take new references to it, the deletion only happens if the other conditions inmddev_putare met.Do you want me to repin the series using a flag?
I would vote for it.
Cheers,
Hannes
Re: [PATCH 09/10] md: only delete entries from all_mddevs when the disk is freed
In reply to Hannes ReineckeQuoted 2 lines · Hannes Reinecke
Why can't we use '
atomic_inc_unless_zero' here and do away with the additional 'deleted' flag?
See my previous reply.
Quoted 8 lines
drivers/md/md.c static bool md_rdev_overlaps(struct md_rdev *rdev) · line 3345
Quoted 5 lines
drivers/md/md.c +2
-0@@ -3338,6 +3342,8 @@ static bool md_rdev_overlaps(struct md_rdev *rdev)spin_lock (&all_mddevs_lock);list_for_each_entry (mddev, &all_mddevs, all_mddevs) { +if (mddev->deleted) +continue ;Any particular reason why you can't use the '
mddev_get()' / 'mddev_put()' sequence here?
Mostly because it would be pretty mess. mdev_put takes all_mddevs_lock, so we'd have to add an unlock/relock cycle for each loop iteration.
Quoted 2 lines
After all, I don't think that 'mddev' should vanish while being in this loop, no?
It won't, at least without the call to mddev_put. Once mddev_put is in the game things aren't that easy, and I suspect the exising and new code might have bugs in that area in the reboot notifier and md_exit for extreme corner cases.
Re: [PATCH 09/10] md: only delete entries from all_mddevs when the disk is freed
Quoted 18 lines · Christoph Hellwig
Quoted 2 lines · Hannes Reinecke
Why can't we use '
atomic_inc_unless_zero' here and do away with the additional 'deleted' flag?See my previous reply.
Quoted 8 lines
drivers/md/md.c static bool md_rdev_overlaps(struct md_rdev *rdev) · line 3345
Quoted 5 lines
drivers/md/md.c +2
-0@@ -3338,6 +3342,8 @@ static bool md_rdev_overlaps(struct md_rdev *rdev)spin_lock (&all_mddevs_lock);list_for_each_entry (mddev, &all_mddevs, all_mddevs) { +if (mddev->deleted) +continue ;Any particular reason why you can't use the '
mddev_get()' / 'mddev_put()' sequence here?Mostly because it would be pretty mess.
mdev_puttakesall_mddevs_lock, so we'd have to add an unlock/relock cycle for each loop iteration.
Quoted 7 lines
Quoted 2 lines
After all, I don't think that 'mddev' should vanish while being in this loop, no?
It won't, at least without the call to
mddev_put. Oncemddev_putis in the game things aren't that easy, and I suspect the exising and new code might have bugs in that area in the reboot notifier andmd_exitfor extreme corner cases.
And again, MD mess.
But thanks for start clearing it up.
Cheers,
Hannes
PATCH 10/10 md: simplify md_open
In reply to Christoph HellwigNow that devices are on the all_mddevs list until the gendisk is freed, there can't be any duplicates. Remove the global list lookup and just grab a reference.
--- drivers/md/md.c | 42 +++++++++++++++--------------------------- 1 file changed, 15 insertions(+), 27 deletions(-)
drivers/md/md.c +15 -27
@@ -7785,45 +7785,33 @@ static int md_set_read_only(struct block_device *bdev, bool ro)static int md_open (struct block_device *bdev,fmode_t mode) { -/* -* Succeed if we can lock the mddev, which confirms that -* it isn't being stopped right now. -*/ -struct mddev *mddev =mddev_find (bdev->bd_dev); +struct mddev *mddev;int err; +spin_lock (&all_mddevs_lock); + mddev =mddev_get (bdev->bd_disk->private_data); +spin_unlock (&all_mddevs_lock);if (!mddev)return -ENODEV ; -if (mddev->gendisk != bdev->bd_disk) { -/* we are racing with mddev_put which is discarding this -* bd_disk. -*/ -mddev_put (mddev); -/* Wait until bdev->bd_disk is definitely gone */ -if (work_pending (&mddev->del_work)) -flush_workqueue (md_misc_wq); -return -EBUSY ; - } -BUG_ON (mddev != bdev->bd_disk->private_data); - -if ((err =mutex_lock_interruptible (&mddev->open_mutex))) + err =mutex_lock_interruptible (&mddev->open_mutex); +if (err) goto out; -if (test_bit (MD_CLOSING , &mddev->flags)) { -mutex_unlock (&mddev->open_mutex); - err = -ENODEV ; - goto out; - } + err = -ENODEV ; +if (test_bit (MD_CLOSING , &mddev->flags)) + goto out_unlock; - err =0 ;atomic_inc (&mddev->openers);mutex_unlock (&mddev->open_mutex);bdev_check_media_change (bdev); - out: -if (err) -mddev_put (mddev); +return 0 ; + +out_unlock: +mutex_unlock (&mddev->open_mutex); +out: +mddev_put (mddev);return err; }
Signature · 1 line
2.30.2
Re: [PATCH 10/10] md: simplify md_open
Quoted 64 lines · Christoph Hellwig
Now that devices are on the
all_mddevslist until the gendisk is freed, there can't be any duplicates. Remove the global list lookup and just grab a reference.Signed-off-by Christoph Hellwig <[email protected]>--- drivers/md/md.c | 42 +++++++++++++++--------------------------- 1 file changed, 15 insertions(+), 27 deletions(-)drivers/md/md.c +9
-26@@ -7785,45 +7785,33 @@ static int md_set_read_only(struct block_device *bdev, bool ro)static int md_open (struct block_device *bdev,fmode_t mode) { -/* -* Succeed if we can lock the mddev, which confirms that -* it isn't being stopped right now. -*/ -struct mddev *mddev =mddev_find (bdev->bd_dev); +struct mddev *mddev;int err; +spin_lock (&all_mddevs_lock); + mddev =mddev_get (bdev->bd_disk->private_data); +spin_unlock (&all_mddevs_lock);if (!mddev)return -ENODEV ; -if (mddev->gendisk != bdev->bd_disk) { -/* we are racing with mddev_put which is discarding this -* bd_disk. -*/ -mddev_put (mddev); -/* Wait until bdev->bd_disk is definitely gone */ -if (work_pending (&mddev->del_work)) -flush_workqueue (md_misc_wq); -return -EBUSY ; - } -BUG_ON (mddev != bdev->bd_disk->private_data); - -if ((err =mutex_lock_interruptible (&mddev->open_mutex))) + err =mutex_lock_interruptible (&mddev->open_mutex); +if (err) goto out; -if (test_bit (MD_CLOSING , &mddev->flags)) { -mutex_unlock (&mddev->open_mutex); - err = -ENODEV ; - goto out; - } + err = -ENODEV ; +if (test_bit (MD_CLOSING , &mddev->flags)) + goto out_unlock; - err =0 ;atomic_inc (&mddev->openers);mutex_unlock (&mddev->open_mutex);bdev_check_media_change (bdev); - out: -if (err)
drivers/md/md.c static int md_set_read_only(struct block_device *bdev, bool ro) · line 7817
Quoted 10 lines
drivers/md/md.c +6
-1-mddev_put (mddev); +return 0 ; + +out_unlock: +mutex_unlock (&mddev->open_mutex); +out: +mddev_put (mddev);return err; }
Cheers,
Hannes