On Thu, 2008-01-31 at 11:53 -0600, James Bottomley wrote: > On Thu, 2008-01-31 at 09:28 -0800, Nicholas A. Bellinger wrote: > > On Thu, 2008-01-31 at 10:26 -0600, James Bottomley wrote: > > > On Thu, 2008-01-31 at 08:10 -0800, Nicholas A. Bellinger wrote: > > > > On Thu, 2008-01-31 at 09:34 -0600, James Bottomley wrote: > > > > > On Thu, 2008-01-31 at 06:10 -0800, Nicholas A. Bellinger wrote: > > > > > > Greetings Geert and Co, > > > > > > > > > > > > I have a related patch that I have been using with ps3rom.c for some > > > > > > time that fixes a bug in fs/bio.c that assumes 512 byte sectors for > > > > > > ATAPI operations. This bug actually exists for all non 512 byte sector > > > > > > devices go through this code path (I found it with > > > > > > scsi_execute_async()), but I first ran into this issue with ps3rom.c > > > > > > because max_sectors (32) is small enough to trigger the bug assuming 512 > > > > > > byte sectors during typical ATAPI READ_10 ops with iSCSI/HD. Because > > > > > > typical max_sector settings for libata and USB are much higher, I have > > > > > > never ran into this issue outside of ps3rom.c, but the bug exists > > > > > > nevertheless.. > > > > > > > > > > > > The current patch assumes 512 byte sectors, and adds a sector_size > > > > > > parameter to drivers/scsi/scsi_lib.c:scsi_req_map_sg() to change it for > > > > > > passed struct request. I know that some folks talked about killing > > > > > > scsi_execute_async() and fixing this problem elsewhere, but until then > > > > > > please consider this patch. Any input is also appreciated. > > > > > > > > > > My first reaction is really, no; there's no way we should be doing such > > > > > a nasty layering violation. > > > > > > > > > > > > > I don't care for it either, but without this patch (or something > > > > similar) all SCSI targets that use scsi_execute_async(), for non 512 > > > > byte requests are broken. This causes a problem when a small max_sector > > > > is correctly used by the LLD, and trips the check in > > > > fs/bio.c:__bio_add_page() > > > > > > > > if (((bio->bi_size + len) >> 9) > max_sectors) > > > > > > > > > > Could we rewind this discussion back to an actual problem description > > > then, please? Nothing in the standard path for a CD/DVD should be using > > > scsi_execute_async(), what's the actual problem use case? > > > > > > > The problem case is a SCSI Target Mode engine that receives a 2048 Byte > > single sector ATAPI READ_10 request from the storage fabric, and uses > > scsi_execute_async() (the only option >= 2.6.18) to issue said request > > to the underlying struct scsi_device. Because the underlying bio code > > assumes 512 byte only sectors, the check in __bio_add_page() incorrectly > > determines that max_sectors (max_sectors has to be low, as with 32 from > > ps3rom.c) has been exceeded, and fails the request back up the stack. > > OK, so this is a totally separate issue from the one you actually posted > it as a patch to fix? > > the queue max_sectors parameter is also counted in the block internal of > 512 byte sectors. If you set it to 32 that means you were only > expecting 16k of transfers per command maximum. If that's not right, > then set the limit correctly. > > In short, and to repeat: almost every internal size counter to block is > in units of 512 byte sectors ... that includes capacity, maximum etc ... > Ok, after reading your followup with Geert I see that this looks like a bug in ps3rom.c assuming 2048 byte sectors to calculate .max_sectors (which was originally set to 32 as I mentioned). Using the setting BOUNCE_SIZE << 9 where BOUNCE_SIZE is the request size in bytes looks like this will solve the issue. My misunderstanding was that .max_sectors was allowed to be calcuated in non 512 byte sectors, so please disregard my patch. Geert, .max_sectors for ps3rom.c using 512 byte sectors ends up being 128, yes.? James, could we put something in the SCSI docs stating that .max_sectors MUST be calculated against 512 byte sectors..? Thanks again, --nab > James > > > - To unsubscribe from this list: 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
