Re: [PATCH v8 1/2] btrfs: introduce device dynamic state transition to failed

[Liu Bo <bo.li.liu@oracle.com>]

On Tue, Oct 03, 2017 at 11:59:19PM +0800, Anand Jain wrote:
> From: Anand Jain <Anand.Jain@oracle.com>
> 
> This patch provides helper functions to force a device to failed,
> and we need it for the following reasons,
> 1) a. It can be reported that device has failed when it does and
>    b. Close the device when it goes offline so that blocklayer can
>       cleanup
> 2) Identify the candidate for the auto replace
> 3) Stop further RW to the failing device and
> 4) A device in the multi device btrfs may fail, but as of now in
>    some system config whole of btrfs gets unmounted.
> 
> Signed-off-by: Anand Jain <anand.jain@oracle.com>
> Tested-by: Austin S. Hemmelgarn <ahferroin7@gmail.com>
> ---
> V8: General misc cleanup. Based on v4.14-rc2
> 
>  fs/btrfs/volumes.c | 104 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>  fs/btrfs/volumes.h |  15 +++++++-
>  2 files changed, 118 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
> index 0e8f16c305df..06e7cf4cef81 100644
> --- a/fs/btrfs/volumes.c
> +++ b/fs/btrfs/volumes.c
> @@ -7255,3 +7255,107 @@ void btrfs_reset_fs_info_ptr(struct btrfs_fs_info *fs_info)
>  		fs_devices = fs_devices->seed;
>  	}
>  }
> +
> +static void do_close_device(struct work_struct *work)
> +{
> +	struct btrfs_device *device;
> +
> +	device = container_of(work, struct btrfs_device, rcu_work);
> +
> +	if (device->closing_bdev)
> +		blkdev_put(device->closing_bdev, device->mode);
> +
> +	device->closing_bdev = NULL;
> +}
> +
> +static void btrfs_close_one_device(struct rcu_head *head)
> +{
> +	struct btrfs_device *device;
> +
> +	device = container_of(head, struct btrfs_device, rcu);
> +
> +	INIT_WORK(&device->rcu_work, do_close_device);
> +	schedule_work(&device->rcu_work);
> +}

No need to schedule the job to a worker.

thanks,

-liubo

> +
> +void btrfs_force_device_close(struct btrfs_device *device)
> +{
> +	struct btrfs_fs_info *fs_info;
> +	struct btrfs_fs_devices *fs_devices;
> +
> +	fs_devices = device->fs_devices;
> +	fs_info = fs_devices->fs_info;
> +
> +	btrfs_sysfs_rm_device_link(fs_devices, device);
> +
> +	mutex_lock(&fs_devices->device_list_mutex);
> +	mutex_lock(&fs_devices->fs_info->chunk_mutex);
> +
> +	btrfs_assign_next_active_device(fs_devices->fs_info, device, NULL);
> +
> +	if (device->bdev)
> +		fs_devices->open_devices--;
> +
> +	if (device->writeable) {
> +		list_del_init(&device->dev_alloc_list);
> +		fs_devices->rw_devices--;
> +	}
> +	device->writeable = 0;
> +
> +	/*
> +	 * Todo: We have miss-used missing flag all around, and here
> +	 * too for now. (In the long run I want to keep missing to only
> +	 * indicate that it was not present when RAID was assembled.)
> +	 */
> +	device->missing = 1;
> +	fs_devices->missing_devices++;
> +	device->closing_bdev = device->bdev;
> +	device->bdev = NULL;
> +
> +	call_rcu(&device->rcu, btrfs_close_one_device);
> +
> +	mutex_unlock(&fs_devices->fs_info->chunk_mutex);
> +	mutex_unlock(&fs_devices->device_list_mutex);
> +
> +	rcu_barrier();
> +
> +	btrfs_warn_in_rcu(fs_info, "device %s failed",
> +				rcu_str_deref(device->name));
> +
> +	/*
> +	 * We lost one/more disk, which means its not as it
> +	 * was configured by the user. Show mount should show
> +	 * degraded.
> +	 */
> +	btrfs_set_opt(fs_info->mount_opt, DEGRADED);
> +
> +	/*
> +	 * Now having lost one of the device, check if chunk stripe
> +	 * is incomplete and handle fatal error if needed.
> +	 */
> +	if (!btrfs_check_rw_degradable(fs_info))
> +		btrfs_handle_fs_error(fs_info, -EIO,
> +				"devices below critical level");
> +}
> +
> +void btrfs_mark_device_failed(struct btrfs_device *dev)
> +{
> +	struct btrfs_fs_devices *fs_devices = dev->fs_devices;
> +
> +	/* This shouldn't be called if device is already missing */
> +	if (dev->missing || !dev->bdev)
> +		return;
> +	if (dev->failed)
> +		return;
> +	dev->failed = 1;
> +
> +	/* Last RW device is requested to force close let FS handle it. */
> +	if (fs_devices->rw_devices == 1) {
> +		btrfs_handle_fs_error(fs_devices->fs_info, -EIO,
> +					"Last RW device failed");
> +		return;
> +	}
> +
> +	/* Point of no return start here. */
> +	btrfs_force_device_close(dev);
> +}
> diff --git a/fs/btrfs/volumes.h b/fs/btrfs/volumes.h
> index 6108fdfec67f..05b150c03995 100644
> --- a/fs/btrfs/volumes.h
> +++ b/fs/btrfs/volumes.h
> @@ -65,13 +65,26 @@ struct btrfs_device {
>  	struct btrfs_pending_bios pending_sync_bios;
>  
>  	struct block_device *bdev;
> +	struct block_device *closing_bdev;
>  
>  	/* the mode sent to blkdev_get */
>  	fmode_t mode;
>  
>  	int writeable;
>  	int in_fs_metadata;
> +	/* missing: device not found at the time of mount */
>  	int missing;
> +	/* failed: device confirmed to have experienced critical io failure */
> +	int failed;
> +	/*
> +	* offline: system or user or block layer transport has removed
> +	* offlined the device which was once present and without going
> +	* through unmount. Implies an intriem communication break down
> +	* and not necessarily a candidate for the device replace. And
> +	* device might be online after user intervention or after
> +	* block transport layer error recovery.
> +	*/
> +	int offline;
>  	int can_discard;
>  	int is_tgtdev_for_dev_replace;
>  	blk_status_t last_flush_error;
> @@ -544,5 +557,5 @@ void btrfs_reset_fs_info_ptr(struct btrfs_fs_info *fs_info);
>  bool btrfs_check_rw_degradable(struct btrfs_fs_info *fs_info);
>  void btrfs_report_missing_device(struct btrfs_fs_info *fs_info, u64 devid,
>  				 u8 *uuid);
> -
> +void btrfs_mark_device_failed(struct btrfs_device *dev);
>  #endif
> -- 
> 2.7.0
>