Can a non-sg scsi write command be more than PAGE_SIZE length?
...or more importantly, is it allowed. Kernel is FC3 2.6.10-1.766.
The latest iscsi driver[1] blows on a 32K-long request for a tape write which followed this path:
[<f8cef985>] iscsi_queuecommand+0x161/0x2f1 [iscsi_tcp] [<f883f724>] scsi_dispatch_cmd+0x1e9/0x24f [scsi_mod] [<f88440ff>] scsi_request_fn+0x29a/0x310 [scsi_mod] [<c0217487>] blk_insert_request+0x8b/0x9e [<f88430b8>] scsi_insert_special_req+0x20/0x26 [scsi_mod] [<f8cc043e>] st_do_scsi+0x111/0x147 [st] [<f8cc026c>] st_sleep_done+0x0/0xc1 [st] [<f8cc189b>] st_write+0x3eb/0x671 [st] [<c0152484>] vfs_write+0xb6/0xe2 [<c015254e>] sys_write+0x3c/0x62 [<c0103c97>] syscall_call+0x7/0xb
The command is not using scatter-gather (sc->use_sg is 0) but is more than PAGE_SIZE, and the driver hates that. So my question is whether it's allowed (then I'll have to fix the driver, not sure many people have virtual tapes over iscsi) or whether st_write has to be fixed.
OG.
[1] 5.0.0.2.
Quoted 1 line · Olivier Galibert
...or more importantly, is it allowed. Kernel is FC3 2.6.10-1.766.
Yes, it's allowed.
Quoted 4 lines · Christoph Hellwig
Quoted 1 line · Olivier Galibert
...or more importantly, is it allowed. Kernel is FC3 2.6.10-1.766.
Yes, it's allowed.
Thanks. Pages in that case are continuous then, right?
OG.
Quoted 3 lines · Olivier Galibert
Quoted 1 line
Yes, it's allowed.
Thanks. Pages in that case are continuous then, right?
Good question actually. I know XFS does passed vmalloc'ed memory down the block I/O path, but that's as a scatter/gather request. All non-s/g request should be contingous I think.
We really need to write down the rules about what memory can be passed down the block I/O path - XFS for example sends kmalloced memory down which all the iSCSI implementations don't like at all.
Quoted 7 lines · Christoph Hellwig
Good question actually. I know XFS does passed vmalloc'ed memory down the block I/O path, but that's as a scatter/gather request. All non-s/g request should be contingous I think.
We really need to write down the rules about what memory can be passed down the block I/O path - XFS for example sends kmalloced memory down which all the iSCSI implementations don't like at all.
We have two rules and two cases in every driver: sg and non-sg. (i.e. use_sg == 0) this is where we do the distinction between dma_map_sg and dma_map_single.
If someone wants a project, it should be possible to eject our non-sg path. This would mean that all requests go as sg through the block layer. The advantages would be
1) internally we would no longer care about kmalloc vs vmalloc memory
2) we wouldn't need to know the gfp flag when allocating internal requests (block would bounce for us if necessary) thus we could kill the scsi isa dma flag
3) We could strip the special casing out of every driver ...
4) it should simplify the I/O traversal paths
James