On Wed, Apr 24, 2019 at 09:14:36AM -0400, Martin K. Petersen wrote: > > Ming, > > I'm traveling today and probably won't be able to take a closer look > until tomorrow. But from a quick glance this looks OK. > > > The integrity buffer can't be very big, for example, the max sectors > > for one bio is 2560, one sector may take at most 8bytes for integrity > > info, so the max size of integrity buffer is just 20k(<=5 pages). > > Just a comment on your rationale about 5 pages. > > buffer_head submissions have traditionally been small, and depending on > your choice of allocator, new allocations would grow backwards in > memory. So there were several common I/O patterns that produced a > single, non-mergeable 8 byte integrity metadata allocation for every 512 > bytes of data in the I/O. It's a pathological corner case. Just make > sure it's something you handle when you muck with this. ext[23] and dd > to the block device used to be able to reproduce this scenario easily. Yeah, you are right. I have realized that 5 pages aren't correct, and it should be one protection data segment for each bio usually. Thanks, Ming
