On Fri, 3 Feb 2006, Mike Christie wrote: > Hugh Dickins wrote: > > On some architectures, mapping the scatterlist may coalesce entries: > > if that coalesced list is then used for freeing the pages afterwards, > > there's a danger that pages may be doubly freed (and others leaked). > > > > Fix SCSI Tape's sgl_unmap_user_pages by freeing from the pagelist used > > in sgl_map_user_pages. Fixes Ryan Richter's crash on x86_64, with Bad > > page state mapcount 2 from sgl_unmap_user_pages, and consequent mayhem. > > Is this crash occuring with 2.6.16-rc1? I do believe 2.6.15 is the latest release that Ryan has tried (and, strangely, nobody else has ever reported it - identifiably, anyway). Interested in trying unpatched 2.6.16-rc2, Ryan? > I ask becuase in that kernel the scatterlist passed into scsi_execute_async > > if (scsi_execute_async(STp->device, cmd, direction, > &((STp->buffer)->sg[0]), bytes, > > is not the same one that gets send down to the device/HBA. Wow, great info, thanks. > scsi_execute_async takes the scatterlist passed to it from st or sg, uses it > as a hint to build a request + bios, then later when the request is sent to > the device a new scatterlist is sent to the device and the device does the > pci/dma operation on that scatterlist from the block/scsi layer. Very interesting. James can confirm, but I think that means everybody can ignore my drivers/scsi/st.c patch for 2.6.16-rc, and the unposted sg.c patches for the same issue that I was going to send Doug. But the Infiniband, ipr and osst patches still look to be necessary - unless maintainers can point to a similar intervening layer. Hugh - : send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
