Re: [PATCH v3 02/11] pstore/blk: new support logger for block devices

[WeiXiong Liao]

hi Kees Cook,

On 2020/5/8 PM 2:07, Kees Cook wrote:
> On Wed, Mar 25, 2020 at 04:54:57PM +0800, WeiXiong Liao wrote:
>> pstore/blk is similar to pstore/ram, but dump log to block device
>> rather than persistent ram.
>> [...]
>> +int psblk_register_blkdev(unsigned int major, psblk_panic_write_op panic_write)
>> +{
>> +	struct block_device *bdev;
>> +	struct psblk_device dev = {0};
>> +	struct bdev_info *binfo;
>> +	int ret = -ENODEV;
>> +	void *holder = blkdev;
>> +
>> +	binfo = psblk_get_bdev_info();
>> +	if (IS_ERR(binfo))
>> +		return PTR_ERR(binfo);
>> +
>> +	/* only allow driver matching the @blkdev */
>> +	if (!binfo->devt || MAJOR(binfo->devt) != major) {
>> +		pr_debug("invalid major %u (expect %u)\n",
>> +				major, MAJOR(binfo->devt));
>> +		return -ENODEV;
>> +	}
>> +
>> +	/* hold bdev exclusively */
>> +	bdev = psblk_get_bdev(holder);
>> +	if (IS_ERR(bdev)) {
>> +		pr_err("failed to open '%s'!\n", blkdev);
>> +		return PTR_ERR(bdev);
>> +	}
>> +
>> +	/* psblk_bdev must be assigned before register to pstore/blk */
>> +	psblk_bdev = bdev;
>> +	blkdev_panic_write = panic_write;
>> +
>> +	dev.total_size = psblk_bdev_size(bdev);
>> +	dev.panic_write = panic_write ? psblk_blk_panic_write : NULL;
>> +	dev.read = psblk_generic_blk_read;
>> +	dev.write = psblk_generic_blk_write;
>> +
>> +	ret = psblk_register_do(&dev);
>> +	if (ret)
>> +		goto err_put_bdev;
>> +
>> +	pr_info("using '%s'\n", blkdev);
>> +	return 0;
>> +
>> +err_put_bdev:
>> +	psblk_bdev = NULL;
>> +	blkdev_panic_write = NULL;
>> +	psblk_put_bdev(bdev, holder);
>> +	return ret;
>> +}
>> +EXPORT_SYMBOL_GPL(psblk_register_blkdev);
> 
> I've gotten this series refactored on top of current pstore, and I've
> been making various bikeshed changes to names, etc, and as I went to go
> start testing, I realized that nothing actually uses
> psblk_register_blkdev().
> 
> It seems like it should be possible to just start using this on any
> block device of the user's choosing. I assume the idea is to allow for
> drivers to register panic_write handlers, but even without that, it'd be
> nice to just be able to test this with something like /dev/loop0.
> 

Yes. psblk_register_blkdev() is there for block drivers to register
panic_write()
handlers. The panic_wrire() is used only when panic occurs. Not only the
panic
log, but also all data on dirty zones. I implement the panic_write() of mmc
and nand on the platform of Allwinner, but I think it is not ready to
submit to
community.

All other front-ends but dmesg for panic are available since pstore/blk
provides
the general write/read through IO stack. That's why /dev/loop0 seemed to
works well.

> What's your thinking on how this would happen? It seems like if
> pstore/blk uses pstore/zone, and mtdpstore uses pstore/blk, there should
> be a blkoops that uses pstore/blk too?  I guess I need to learn a bit> more about how block device probing works so pstore/blk can notice
> devices as they're brought online, etc.
> 

pstore/blk provides all  user options and register function for device
drivers.
The mtdpstore is the implementation case using pstore/blk. How about
'mmc_pstore' for mmc and *_pstore for others block device?

I guess I need to learn more about how pstore/blk can notice devices too.
I think pstore/blk can be better if block device can do like mtd device that
not only notifies but also provides generic panic_write().

-- 
WeiXiong Liao