Re: [PATCH] bcache: avoid redundant access RB tree in read_dirty

> 2025年10月21日 00:39，Pierre Juhen <[email protected]> 写道：
> 
> Hi 
> I am on kernel 6.16.12.
> I have had errors with bcache recently, And I lost my fronted 3 or 4 times :
> oct. 20 15:37:40 pierre.juhen kernel: bcache: journal_read_bucket() 128: bad csum, 32768 bytes, offset 0 
> oct. 20 15:37:40 pierre.juhen kernel: bcache: journal_read_bucket() 64: bad csum, 22928 bytes, offset 0 
> oct. 20 15:37:40 pierre.juhen kernel: bcache: journal_read_bucket() 32: bad csum, 4848 bytes, offset 2 
> oct. 20 15:37:40 pierre.juhen kernel: bcache: journal_read_bucket() 48: bad csum, 14096 bytes, offset 0 
> oct. 20 15:37:40 pierre.juhen (udev-worker)[461]: nvme0n1p3: Process 'bcache-register /dev/nvme0n1p3' failed with exit code 1. 
> oct. 20 15:37:40 pierre.juhen kernel: bcache: prio_read() bad csum reading priorities 
> oct. 20 15:37:40 pierre.juhen kernel: bcache: bch_cache_set_error() error on 448f191c-28df-4396-bc44-14d1f77c9005: IO error reading priorities, disabling caching 
> oct. 20 15:37:40 pierre.juhen kernel: bcache: register_bcache() error : failed to register device
> 

I assume this email is irrelevant to the patch “bcache: avoid redundant access RB tree in read_dirty”, am I correct?


> I had to reconfigure everything after a disk problem.
> I have been running bcache for years now, without any problems.
> The only difference might be that I configured the frontend with the discard option.

The discard option is not recommended. Indeed in next merge window I will submit a patch series to drop the discard option.


> The logical volume using bcache have also a discard option in fstab.
> The frontend is on a Samsung 980 nvme disk. 

Try not to enable discard on cache device. This option will disappear soon.

I don’t know whether discard option of Samsung 980 nvme disk may change the content of discarded LBA or not, from NVMe spec, it could be zero-filled or undefined.
Anyway in current code discard doesn’t help performance, I suggest to not enable discard and see whether the issue still shows up.

My suggestion is: always use default configuration, all our test case and performance optimization are for default configurations.

Thanks.

Coly Li