Reduce number of KOBJ_REMOVE events
Subsystems (7)
PATCH 1/1 Reduce number of KOBJ_REMOVE events
Author: Zdenek Kabelac <[email protected]> Date: Fri Jul 22 15:02:17 2011 +0200
Reduce number of KOBJ_REMOVE events
For now udev recieves 3 event for removal of DM logical volumes. (1 for bdi and 2 for same block kobject). Reason is dm device generates its own kobject event with approriate env parameter and block layer sends another KOBJ_REMOVE event on its own unconditionaly for the same kobject. As for now only the kobject cleanup checks that the REMOVE event has been already sent and avoids duplicate REMOVE event.
The patch for kobject_uevent_env() which has been testing for duplicate REMOVE event did not passed into the mainline (yet?):
https://www.redhat.com/archives/dm-devel/2010-March/msg00107.html
I'm proposing alternative way around to always use kobject cleanup routine for sending REMOVE event if it was not send by the module - so it makes the code few lines shorter.
This patch removes single appearencies of kobject_uevent() prior kobject removal.
block/blk-integrity.c +0 -1
@@ -439,7 +439,6 @@ void blk_integrity_unregister(struct gendisk *disk) bi = disk->integrity; -kobject_uevent (&bi->kobj,KOBJ_REMOVE );kobject_del (&bi->kobj);kobject_put (&bi->kobj); disk->integrity =NULL ;
block/blk-sysfs.c +0 -2
@@ -521,7 +521,6 @@ int blk_register_queue(struct gendisk *disk) ret =elv_register_queue (q);if (ret) { -kobject_uevent (&q->kobj,KOBJ_REMOVE );kobject_del (&q->kobj);blk_trace_remove_sysfs (dev);kobject_put (&dev->kobj);
@@ -541,7 +540,6 @@ void blk_unregister_queue(struct gendisk *disk)if (q->request_fn)elv_unregister_queue (q); -kobject_uevent (&q->kobj,KOBJ_REMOVE );kobject_del (&q->kobj);blk_trace_remove_sysfs (disk_to_dev (disk));kobject_put (&disk_to_dev (disk)->kobj);
block/elevator.c +0 -1
@@ -892,7 +892,6 @@ EXPORT_SYMBOL(elv_register_queue);static void __elv_unregister_queue (struct elevator_queue *e) { -kobject_uevent (&e->kobj,KOBJ_REMOVE );kobject_del (&e->kobj); e->registered =0 ; }
drivers/base/core.c +0 -2
@@ -1013,7 +1013,6 @@ done: ueventattrError:device_remove_file (dev, &uevent_attr); attrError: -kobject_uevent (&dev->kobj,KOBJ_REMOVE );kobject_del (&dev->kobj); Error:cleanup_device_parent (dev);
@@ -1132,7 +1131,6 @@ void device_del(struct device *dev)*/ if (platform_notify_remove)platform_notify_remove (dev); -kobject_uevent (&dev->kobj,KOBJ_REMOVE );cleanup_device_parent (dev);kobject_del (&dev->kobj);put_device (parent);
mm/slub.c +0 -1
@@ -4720,7 +4720,6 @@ static void sysfs_slab_remove(struct kmem_cache *s)*/ return ; -kobject_uevent (&s->kobj,KOBJ_REMOVE );kobject_del (&s->kobj);kobject_put (&s->kobj); }
net/bridge/br_if.c +0 -1
@@ -153,7 +153,6 @@ static void del_nbp(struct net_bridge_port *p)br_multicast_del_port (p); -kobject_uevent (&p->kobj,KOBJ_REMOVE );kobject_del (&p->kobj);br_netpoll_disable (p);
Quoted 9 lines · Zdenek Kabelac
For now udev recieves 3 event for removal of DM logical volumes. (1 for bdi and 2 for same block kobject). Reason is dm device generates its own kobject event with approriate env parameter and block layer sends another
KOBJ_REMOVEevent on its own unconditionaly for the same kobject. As for now only the kobject cleanup checks that the REMOVE event has been already sent and avoids duplicate REMOVE event.The patch for
kobject_uevent_env()which has been testing for duplicate REMOVE event did not passed into the mainline (yet?):
No, it's wasn't merged. Subsystems should really not send their own 'add' or 'remove' events. These are properties of the driver core.
Quoted 3 lines
I'm proposing alternative way around to always use kobject cleanup routine for sending REMOVE event if it was not send by the module - so it makes the code few lines shorter.
The events the core creates are only sent out at release() not at del(), so we would delay 'remove' events when we keep the device pinned but it's not valid anymore. We can not do that today, we would need to move the core-created 'remove' events to del().
For device-mapper, I would prefer to add a '.dev_uevent' callback to the 'block' class let this callback check 'struct block_device_operations' for a possibly specified '.uevent' callback and call it.
Then have 'dm_blk_dops' add '.uevent' and let the core call into the dm code to the needed properties to the 'remove' event, instead of sending its own, and see the duplication.
Let me know if you need more details.
Thanks,
Kay
Dne 25.7.2011 02:18, Kay Sievers napsal(a):
Quoted 25 lines · Zdenek Kabelac
Quoted 10 lines
For now udev recieves 3 event for removal of DM logical volumes. (1 for bdi and 2 for same block kobject). Reason is dm device generates its own kobject event with approriate env parameter and block layer sends another
KOBJ_REMOVEevent on its own unconditionaly for the same kobject. As for now only the kobject cleanup checks that the REMOVE event has been already sent and avoids duplicate REMOVE event.The patch for
kobject_uevent_env()which has been testing for duplicate REMOVE event did not passed into the mainline (yet?):No, it's wasn't merged. Subsystems should really not send their own 'add' or 'remove' events. These are properties of the driver core.
Quoted 3 lines
I'm proposing alternative way around to always use kobject cleanup routine for sending REMOVE event if it was not send by the module - so it makes the code few lines shorter.
The events the core creates are only sent out at
release()not atdel(), so we would delay 'remove' events when we keep the device pinned but it's not valid anymore. We can not do that today, we would need to move the core-created 'remove' events todel().
Quoted 7 lines
For device-mapper, I would prefer to add a '.dev_uevent' callback to the 'block' class let this callback check 'struct
block_device_operations' for a possibly specified '.uevent' callback and call it.Then have '
dm_blk_dops' add '.uevent' and let the core call into the dm code to the needed properties to the 'remove' event, instead of sending its own, and see the duplication.
Sounds like complex solution - maybe it would be easier to just register some environment variable on dm code side - like kobject_add_env() - so it would take envs from this internal kobject list and after sending uevent it would implicitly clear this list.
So in dm case dm-uevent would just register env(cookie) for KOBJ_REMOVE and would left kobject_uevent() on block layer ?
Also I'm aware that remove event would be delayed by leaving it on kobject_cleanup(), but since you mentioned 'del()' as a better place - why not move this implicit uvent call there - since most kernel driver writers probably do not want to be bothered with uvents?
Zdenek
Quoted 29 lines · Zdenek Kabelac
Dne 25.7.2011 02:18, Kay Sievers napsal(a):
Quoted 28 lines · Zdenek Kabelac
Quoted 10 lines
For now udev recieves 3 event for removal of DM logical volumes. (1 for bdi and 2 for same block kobject). Reason is dm device generates its own kobject event with approriate env parameter and block layer sends another
KOBJ_REMOVEevent on its own unconditionaly for the same kobject. As for now only the kobject cleanup checks that the REMOVE event has been already sent and avoids duplicate REMOVE event.The patch for
kobject_uevent_env()which has been testing for duplicate REMOVE event did not passed into the mainline (yet?):No, it's wasn't merged. Subsystems should really not send their own 'add' or 'remove' events. These are properties of the driver core.
Quoted 3 lines
I'm proposing alternative way around to always use kobject cleanup routine for sending REMOVE event if it was not send by the module - so it makes the code few lines shorter.
The events the core creates are only sent out at
release()not atdel(), so we would delay 'remove' events when we keep the device pinned but it's not valid anymore. We can not do that today, we would need to move the core-created 'remove' events todel().For device-mapper, I would prefer to add a '.dev_uevent' callback to the 'block' class let this callback check 'struct
block_device_operations' for a possibly specified '.uevent' callback and call it.
Quoted 5 lines
Quoted 3 lines
Then have '
dm_blk_dops' add '.uevent' and let the core call into the dm code to the needed properties to the 'remove' event, instead of sending its own, and see the duplication.Sounds like complex solution
I don't think so, It's clean, ~30 lines long, and technically correct, I expect.
Quoted 4 lines
maybe it would be easier to just register some
environment variable on dm code side - likekobject_add_env()- so it would take envs from this internal kobject list and after sending uevent it would implicitly clear this list.
So we would keep allocated per-event-type variables in the kobject, to send when 'remove' is finally called? The callbacks are just much simpler , I guess.
Quoted 6 lines
So in dm case dm-uevent would just register env(cookie) for
KOBJ_REMOVEand would leftkobject_uevent()on block layer ?Also I'm aware that remove event would be delayed by leaving it on
kobject_cleanup(), but since you mentioned 'del()' as a better place - why not move this implicit uvent call there.
It's probably not wrong to do that, but I don't remember now why we added it to release() that time.
Quoted 2 lines
since most kernel driver writers
probably do not want to be bothered with uvents?
I guess, most drivers use devices not raw kobjects, which send events on their own, and don't need anything here.
Kay
Dne 25.7.2011 14:17, Kay Sievers napsal(a):
Quoted 30 lines · Zdenek Kabelac
Quoted 29 lines
Dne 25.7.2011 02:18, Kay Sievers napsal(a):
Quoted 28 lines · Zdenek Kabelac
Quoted 10 lines
For now udev recieves 3 event for removal of DM logical volumes. (1 for bdi and 2 for same block kobject). Reason is dm device generates its own kobject event with approriate env parameter and block layer sends another
KOBJ_REMOVEevent on its own unconditionaly for the same kobject. As for now only the kobject cleanup checks that the REMOVE event has been already sent and avoids duplicate REMOVE event.The patch for
kobject_uevent_env()which has been testing for duplicate REMOVE event did not passed into the mainline (yet?):No, it's wasn't merged. Subsystems should really not send their own 'add' or 'remove' events. These are properties of the driver core.
Quoted 3 lines
I'm proposing alternative way around to always use kobject cleanup routine for sending REMOVE event if it was not send by the module - so it makes the code few lines shorter.
The events the core creates are only sent out at
release()not atdel(), so we would delay 'remove' events when we keep the device pinned but it's not valid anymore. We can not do that today, we would need to move the core-created 'remove' events todel().For device-mapper, I would prefer to add a '.dev_uevent' callback to the 'block' class let this callback check 'struct
block_device_operations' for a possibly specified '.uevent' callback and call it.
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Then have '
dm_blk_dops' add '.uevent' and let the core call into the dm code to the needed properties to the 'remove' event, instead of sending its own, and see the duplication.Sounds like complex solution
I don't think so, It's clean, ~30 lines long, and technically correct, I expect.
Well then I've probably not fully understand your idea here - I guess it would then simpler written by you?
Quoted 8 lines
Quoted 4 lines
maybe it would be easier to just register some
environment variable on dm code side - likekobject_add_env()- so it would take envs from this internal kobject list and after sending uevent it would implicitly clear this list.So we would keep allocated per-event-type variables in the kobject, to send when 'remove' is finally called? The callbacks are just much simpler , I guess.
No - nothing so complex - kobject would have the list - and you would be able to add some env param to this list - the nearest kobject_uevent() would just splice those parameters to the env list it wants to send (something like 10 lines of code). The only user would be probably dm so far - and it would check it wants to send REMOVE - and in this case it would add env to kobject and would skip kobject_uvent.
On the other hand, it would probably extended kobject struct size without big use case - so Milan's solution that checks whether REMOVE has been already sent and skip all future REMOVE events seems by far the simplest here.
I think your proposal also requires struct extension to store callback somewhere ?
Quoted 9 lines
Quoted 6 lines
So in dm case dm-uevent would just register env(cookie) for
KOBJ_REMOVEand would leftkobject_uevent()on block layer ?Also I'm aware that remove event would be delayed by leaving it on
kobject_cleanup(), but since you mentioned 'del()' as a better place - why not move this implicit uvent call there.It's probably not wrong to do that, but I don't remember now why we added it to
release()that time.
del() looks like the best natural place here - and safe few lines of code ;)
Zdenek
Quoted 9 lines · Zdenek Kabelac
Dne 25.7.2011 14:17, Kay Sievers napsal(a):
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Then have '
dm_blk_dops' add '.uevent' and let the core call into the dm code to the needed properties to the 'remove' event, instead of sending its own, and see the duplication.Sounds like complex solution
I don't think so, It's clean, ~30 lines long, and technically correct, I expect.
In include/linux/blkdev.h add:
int (*uevent) (struct gendisk *, struct kobj_uevent_env *env);
to
struct block_device_operations
In block/genhd.c add:
.uevent = block_uevent;
to
struct class block_class
and have the block_uevent() callback check for an existing uevent() in the block_device_operations, call it when it's specified.
In drivers/md/dm.c add:
.dm_uevent;
to
struct block_device_operations dm_blk_dops
and add all the needed variables there, instead of sending out a faked device 'remove' event.
Quoted 9 lines
Quoted 8 lines
Quoted 4 lines
maybe it would be easier to just register some
environment variable on dm code side - likekobject_add_env()- so it would take envs from this internal kobject list and after sending uevent it would implicitly clear this list.So we would keep allocated per-event-type variables in the kobject, to send when 'remove' is finally called? The callbacks are just much simpler , I guess.
Quoted 6 lines
No - nothing so complex - kobject would have the list - and you would be able to add some env param to this list - the nearest kobject_uevent() would just splice those parameters to the env list it wants to send (something like 10 lines of code). The only user would be probably dm so far - and it would check it wants to send REMOVE - and in this case it would add env to kobject and would skip
kobject_uvent.
We don't want to encourage anybody to mess around with events that way. We also don't want to keep track of subsystem specific event properties in the core kobject. It's just wrong to send 'remove' from a driver that uses 'struct device'. It should be changed in dm, not worked-around in the core.
Quoted 3 lines
On the other hand, it would probably extended kobject struct size without big use case - so Milan's solution that checks whether REMOVE has been already sent and skip all future REMOVE events seems by far the simplest here.
Yeah, I see the logic, but it doesn't make sending additional 'remove' events any more correct. We don't want to encourage anybody with code in the core to do this. 'Add' and 'remove' is nothing struct device user should send on their own.
Quoted 1 line
I think your proposal also requires struct extension to store callback somewhere ?
Yeah, it's a block subsys callback, a pointer per block driver, not per device, nothing that really matters, and it's looks generally useful.
Quoted 11 lines
Quoted 9 lines
Quoted 6 lines
So in dm case dm-uevent would just register env(cookie) for
KOBJ_REMOVEand would leftkobject_uevent()on block layer ?Also I'm aware that remove event would be delayed by leaving it on
kobject_cleanup(), but since you mentioned 'del()' as a better place - why not move this implicit uvent call there.It's probably not wrong to do that, but I don't remember now why we added it to
release()that time.
del()looks like the best natural place here - and safe few lines of code ;)
Yeah, but again, raw kobjects never send any event, not even 'add'. It's up the caller to take care of that, if these events are needed. Many things rightfully never send any event.
The code in the core exists only for the un-common use case that the caller does send 'add' but does not clean-up the objects properly, and would introduce an asymmetry with that. We don't want to encourage anything like that.
The code isn't there to allow additional 'remove' events or to leave the cleanup to the core in general.
Thanks,
Kay
Quoted 23 lines · Kay Sievers
Quoted 9 lines · Zdenek Kabelac
Dne 25.7.2011 14:17, Kay Sievers napsal(a):
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Then have '
dm_blk_dops' add '.uevent' and let the core call into the dm code to the needed properties to the 'remove' event, instead of sending its own, and see the duplication.Sounds like complex solution
I don't think so, It's clean, ~30 lines long, and technically correct, I expect.
In
include/linux/blkdev.hadd:
int (*uevent) (struct gendisk *, structkobj_uevent_env*env);
to
structblock_device_operationsIn
block/genhd.cadd:
.uevent =block_uevent;
to
struct classblock_class
and have theblock_uevent()callback check for an existinguevent()in theblock_device_operations, call it when it's specified.
Quoted 6 lines
In
drivers/md/dm.cadd:
.dm_uevent;
to
structblock_device_operationsdm_blk_dops
and add all the needed variables there, instead of sending out a faked device 'remove' event.
Here is the code to illustrate, and fix the issue of needing to send any 'remove' for a device from a driver.
Thanks,
Kay
block/genhd.c +13 -0
@@ -1097,8 +1097,21 @@ static void disk_release(struct device *dev)free_part_info (&disk->part0);kfree (disk); } + +static int block_uevent (struct device *dev,struct kobj_uevent_env *env) +{ +struct gendisk *disk =dev_to_disk (dev); + +if (!disk->fops) +return 0 ; +if (disk->fops->uevent) +return disk->fops->uevent (disk, env); +return 0 ; +} +struct class block_class = { .name ="block" , + .dev_uevent = block_uevent, };static char *block_devnode (struct device *dev,mode_t *mode)
drivers/md/dm.c +6 -0
@@ -2682,11 +2682,17 @@ void dm_free_md_mempools(struct dm_md_mempools *pools)kfree (pools); } +static int dm_uevent (struct gendisk *disk,struct kobj_uevent_env *env) +{ +return add_uevent_var (env,"COOKIE_MONSTER=1" ); +} +static const struct block_device_operations dm_blk_dops = { .open = dm_blk_open, .release = dm_blk_close, .ioctl = dm_blk_ioctl, .getgeo = dm_blk_getgeo, + .uevent = dm_uevent, .owner =THIS_MODULE };
include/linux/blkdev.h +1 -0
@@ -1300,6 +1300,7 @@ struct block_device_operations { void **, unsigned long *);unsigned int (*check_events) (struct gendisk *disk,unsigned int clearing); +int (*uevent) (struct gendisk *,struct kobj_uevent_env *env);/* ->media_changed() is DEPRECATED, use ->check_events() instead */ int (*media_changed) (struct gendisk *);void (*unlock_native_capacity) (struct gendisk *);