Re: [PATCH -next 0/6] md: remvoe rcu protection to access rdev from conf

[Yu Kuai]

Hi,

在 2023/10/19 9:04, Yu Kuai 写道:
> Hi,
>
> 在 2023/10/19 1:58, Song Liu 写道:
> > On Sun, Oct 15, 2023 at 6:28 PM Yu Kuai <yukuai1@xxxxxxxxxxxxxxx> wrote:
> > > From: Yu Kuai <yukuai3@xxxxxxxxxx>
> > >
> > > Yu Kuai (6):
> > >    md: remove useless debug code to print configuration
> > >    md: remove flag RemoveSynchronized
> > >    md/raid1: remove rcu protection to access rdev from conf
> > >    md/raid10: remove rcu protection to access rdev from conf
> > >    md/raid5: remove rcu protection to access rdev from conf
> > >    md/md-multipath: remove rcu protection to access rdev from conf
> >
> > While a cover letter is highly recommended for a patchset. An empty
> > cover letter like this doesn't really add value. If the description of 
> > each
> > patch is clear enough, it is OK to not have a cover-letter.
>
> Yes, I was hoping commit message in each patch is clear enough,I'll try
> to explain everything more detailed.

I'll add following cover letter in the next version:

The lifetime of rdev:

1. md_import_device() generate a rdev based on underlying disk;

   mddev_lock()
   rdev = kzalloc();
   rdev->bdev = blkdev_get_by_dev();
   mddev_unlock()

2. bind_rdev_to_array() add this rdev to mddev->disks;

   mddev_lock()
   kobject_add(&rdev->kobj, &mddev->kobj, ...);
   list_add_rcu(&rdev->same_set, &mddev->disks);
   mddev_unlock()

3. remove_and_add_spares() add this rdev to conf;

   mddev_lock()
   rdev_addable();
   pers->hot_add_disk();
   rcu_assign_pointer(conf->rdev, rdev);
   mddev_unlock()

4. Use this array with rdev;

5. remove_and_add_spares() remove rdev from conf;

   // triggered by sysfs/ioctl
   mddev_lock()
   rdev_removeable();
   pers->hot_remove_disk();
    rcu_assign_pointer(conf->rdev, NULL);
    synchronize_rcu();
   mddev_unlock()

   // triggered by deamon
   mddev_lock()
   rdev_removeable();
   synchronize_rcu(); -> this can't protect accessing rdev from conf
   pers->hot_remove_disk();
    rcu_assign_pointer(conf->rdev, NULL);
   mddev_unlock()

6. md_kick_rdev_from_array() remove rdev from mddev->disks;

   mddev_lock()
   list_del_rcu(&rdev->same_set);
   synchronize_rcu();
   list_add(&rdev->same_set, &mddev->deleting)
   mddev_unlock()
    export_rdev

There are two seperate rcu protection for rdev, and this pathset remove
the protection of conf(step 3 and 5), because it's safe to access rdev
from conf in following cases:

 - If 'reconfig_mutex' is held, because rdev can't be added or rmoved to
 conf;
 - If there is normal IO inflight, because mddev_suspend() will prevent
 rdev to be added or removed to conf;
 - If sync thread is running, because remove_and_add_spares() can only
 be called from daemon thread when sync thread is done, and
 'MD_RECOVERY_RUNNING' is also checked for ioctl/sysfs;
 - if rcu_read_lock() or any spinlock is held, because synchronize_rcu()
 from step 6 prevent rdev to be freed until rcu_read_unlock() or
 spinlock is released;

Thanks,
Kuai

> Thanks,
> Kuai
>
> > Thanks,
> > Song
> >
> > >   drivers/md/md-multipath.c |  29 ++---
> > >   drivers/md/md.c           |  37 +-----
> > >   drivers/md/raid1.c        |  94 ++++-----------
> > >   drivers/md/raid10.c       | 248 +++++++++-----------------------------
> > >   drivers/md/raid5-cache.c  |  11 +-
> > >   drivers/md/raid5-ppl.c    |  16 +--
> > >   drivers/md/raid5.c        | 225 ++++++++++------------------------
> > >   drivers/md/raid5.h        |   4 +-
> > >   8 files changed, 163 insertions(+), 501 deletions(-)
> > >
> > > --
> > > 2.39.2
> > .
>
> .