panic in ffs (Re: hangs in nbufkv)

[Scott Long]

Kris Kennaway wrote:
> On Tue, Oct 12, 2004 at 11:18:16AM -0700, Matthew Dillon wrote:
> 
>>:[...]
>>:
>>:>    But to be absolutely safe, I would follow Bruce's original suggestion
>>:>    and increase BKVASIZE to 64K, for your particular system.
>>:>  
>>:>
>>:After doing this and testing our backup script, the machine panicked two 
>>:hours later (about half-way through the backup) with 
>>:"initiate_write_inodeblock_ufs2: already started" (in 
>>:ufs/ffs/ffs_softdep.c)... I guess, block sizes above 16Kb are just buggy 
>>:and newfs(8) should be honest about it...
>>:
>>:    -mi
>>
>>    Well, it's possible that UFS has bugs related to large block sizes.
>>    People have gotten bitten on and off over the years but usually it 
>>    works ok if you leave the 8:1 blocksize:fragsize ratio intact.  e.g.
>>    if you have a 64KB block size then you should use a 8K frag size.
>>    If you have a 32KB block size then you should use a 4K frag size.
>>
>>    I think the buffer cache itself is is likely not the source of this
>>    particular bug.
> 
> 
> FYI, I ran the package build cluster with 4:1 ratios for a few months
> and did not have problems.  If there are major bugs there I would have
> expected to come across them.
> 
> Kris

IIRC, the fragment allocation code in FFS allocates in terms of bits 
withing a byte, so anything that is 8:1 or less should work.

Scott