Lustre devel - Dec 2006 - [Bug 11312] FMR/OFED-1.1 is broken

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=11312



Created an attachment (id=9189)
Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
 --> (https://bugzilla.lustre.org/attachment.cgi?id=9189&action=view)
o2iblnd patch to support OFED-1.1 FMR

Improved patch in case IB_USER_VERBS_ABI_VERSION disappears in future releases.

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=11312



Eric

I realize there is another issue with this API change.
Since the read addr (rd->rd_addr) is not an output argument anymore it needs
to
be initialized the same we were doing it with VIB. (kibnal_map_tx/kiblnd_map_tx)

But I see another piece of code in VIB and I wonder if this should apply to o2ib
as well (kibnal_map_tx):

        rd->rd_addr = tx->tx_md.md_addr;

        /* Compensate for adaptor-tavor''s munging of gatherlist
addresses */
        if (active)
                rd->rd_addr += PAGE_OFFSET;

For now I have just added in kiblnd_map_tx:

#if !USE_OLD_OFED_FMR_API
        rd->rd_addr = page_offset;
#endif

I discovered this issue while using zeroconf to mount. Seems to be the single 
case where we do large non bulk IO (no kiovs).

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=11312



I have tested the code with the "rd->rd_addr = page_offset;" fix
but without the "if (active) rd->rd_addr += PAGE_OFFSET;" one.

Works fine on a x86_64 platform.
So you think it should be OK ?