Christoph, >>>> +/* >>>> + * Pass bio_write_zeroes_limit() return value in @limit, as the return >>>> + * value may change after a REQ_OP_WRITE_ZEROES is issued. >>>> + */ >>> I don't think that really helps all that much to explain the issue, >>> which is about SCSI not having an ahead of time flag that reliably >>> works for write same support, which makes it clear the limit to 0 on >>> the first I/O completion. Maybe you can actually spell this out? >> >> Please just tell me what you would like to see, and I will copy verbatim. > > Probably just what I just wrote. Unless Martin can come up with > better language. There is no reliable way for the SCSI subsystem to determine whether a device supports a WRITE SAME operation without actually performing a write to media. As a result, write_zeroes is enabled by default and will be disabled if a zeroing operation subsequently fails. This means that this queue limit is likely to change at runtime. Something like that, perhaps? -- Martin K. Petersen Oracle Linux Engineering
