On 1/17/06, Michael S. Tsirkin <[EMAIL PROTECTED]> wrote: > Quoting Or Gerlitz <[EMAIL PROTECTED]>: > > Second and indeed more important, from our experience, there are > > eventually IB consumers such as the Linux SCSI Mid-Layer which sometimes > > generate Scatter-Gather lists that are "RDMA aligned" when treated in a > > resolution different from the system PAGE_SHIFT. Example to that we saw > > with ia64 SLES9 SP1/2 kernels. So if you work in PAGE_SHIFT you can not > > produce one VA for many of the SG submitted by the mid-layer. > > Interesting. Where does the mid-layer get the 4K (not PAGE_SIZE) aligned > buffers? Any idea? > Not really. We suspect a different allocation unit in the buffer cache or some file systems not doing what we think they do.. We saw it also on Itanium with 16K page size that was sending 4K sg elements. What it made obvious is that the memory registration restrictions are the HCA restrictions, and those have nothing to do with kernel restrictions, yet the driver code relies on kernel restrictions.
Dan > > -- > MST > _______________________________________________ > openib-general mailing list > [email protected] > http://openib.org/mailman/listinfo/openib-general > > To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general > _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
