On Tue 16-07-24 16:51:17, Matthew Wilcox wrote: > On Tue, Jul 16, 2024 at 06:41:49PM +0300, Sergey Shtylyov wrote: > > And here we'll have at least one potential problem (that you neglected > > to describe): with 1 << 31, that 1 will land in a sign bit and then, when > > it's implicitly cast to *unsigned long*, the 32-bit value will be sign- > > extended to 64-bit on 64-bit arches) and then we'll have an incorrect size > > (0xffffffff80000000) passed to create_empty_buffers()... > > Tell me more about this block device you have with a 2GB block size ... ? > > (ie note that this is a purely theoretical issue) Yeah, this just does not make huge amount of sense. Maybe a proper fix would be to just make blocksize uint? There are a lot of places where blocksize is actually stored in a 32-bit type... Honza > > > > to use 1UL instead. > > > > Perphas was worth noting that using 1UL saves us 1 movsx instruction on > > x86_64... > > That is a worthwhile addition to the change log. > > Also, you should cc the person who wrote that code, ie me. > -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
