On Tue, Apr 12, 2011 at 5:51 AM, Amit Shah <amit.shah@xxxxxxxxxx> wrote: > On (Fri) 08 Apr 2011 [09:52:07], Tejun Heo wrote: >> Hello, >> >> On Fri, Apr 08, 2011 at 05:43:16PM +0100, Stefan Hajnoczi wrote: >> > >> I think it would make sense to refresh the inode size on media change >> > >> so that even open file descriptors see the new size and a single >> > >> process cannot force a stale value for all other userspace processes >> > >> on the system. >> > > >> > > Hmmm... I don't know. Maybe we can but I'm not sure whether there's a >> > > good reason for it. cdrom is locked while opened after all. Are >> > > there actual problems? >> > >> > Yeah, sorry I didn't explain what the use case was. With QEMU you can >> > pass through the physical CD-ROM into the virtual machine. >> > >> > QEMU opens /dev/cdrom with O_NONBLOCK | O_RDONLY. The guest can test >> > if the medium is present and QEMU will do ioctl(fd, >> > CDROM_DRIVE_STATUS, CDSL_CURRENT). The guest can also lock the tray >> > and eject, again using the respective ioctls. Read operations are >> > serviced by performing a read on the file descriptor in QEMU. And >> > finally the medium size is queried by QEMU using lseek(fd, 0, >> > SEEK_END). >> > >> > Today QEMU cannot keep /dev/cdrom open across media change because it >> > will have an outdated inode size returned from lseek(fd, 0, SEEK_END). >> > But if the cdrom driver (or sr) refresh the inode size on media >> > change then there is no need to work around this from userspace. >> >> Hmmm... ISTR there was some discussion about changing inode size on >> the fly quite a while ago. I didn't follow the discussion but it >> seemed to have rather nasty/delicate implications. > > I don't necessarily agree with having to modify inode sizes on the > fly, but the main bug here is that the inode doesn't get invalidated > if a CDROM is ejected while a process has an fd to the CDROM device > opened. So as in the original case, if a CD is swapped with one > having more data, lseek continues to report the original media's size > in the process that keeps the fd open across eject/insert. > > I haven't tried this on physical systems, so can't count out it being > a qemu bug as well. Amit, This sounds exactly like the bug that I described and it happens on bare metal without virtualization. Stefan -- 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