On Thu, 01/19 15:49, Daniel P. Berrange wrote: > On Thu, Jan 19, 2017 at 10:38:14PM +0800, Fam Zheng wrote: > > This implements open flag sensible image locking for local file > > and host device protocol. > > > > virtlockd in libvirt locks the first byte, so we start looking at the > > file bytes from 1. > > > > Quoting what was proposed by Kevin Wolf <kw...@redhat.com>, there are > > four locking modes by combining two bits (BDRV_O_RDWR and > > BDRV_O_SHARE_RW), and implemented by taking two locks: > > > > Lock bytes: > > > > * byte 1: I can't allow other processes to write to the image > > * byte 2: I am writing to the image > > > > Lock modes: > > > > * shared writer (BDRV_O_RDWR | BDRV_O_SHARE_RW): Take shared lock on > > byte 2. Test whether byte 1 is locked using an exclusive lock, and > > fail if so. > > > > * exclusive writer (BDRV_O_RDWR only): Take shared lock on byte 2. Test > > whether byte 1 is locked using an exclusive lock, and fail if so. Then > > take shared lock on byte 1. I suppose this is racy, but we can > > probably tolerate that. > > > > * reader that can tolerate writers (BDRV_O_SHARE_RW only): Don't do anything > > > > * reader that can't tolerate writers (neither bit is set): Take shared > > lock on byte 1. Test whether byte 2 is locked, and fail if so. > > Ahh, using two bytes is an interesting technique for mapping the four > different access methods onto the more limit fcntl lock semantics. We > might want to copy this approach in libvirt too.... > > > +/* Posix file locking bytes. Libvirt takes byte 0, so start from byte 1. */ > > +#define RAW_LOCK_BYTE_MIN 1 > > +#define RAW_LOCK_BYTE_NO_OTHER_WRITER 1 > > +#define RAW_LOCK_BYTE_WRITE 2 > > ...would you mind if QEMU started from say byte 10, leaving the first 10 > reserved for libvirt uses. This lets libvirt have a continuous space for > its own usage if we want to use more bytes
That's easy, will do it. (Actually the descriptions above are a bit stale because exclusive writers now take two bytes exclusively and the lock testings are done with F_OFD_GETLK so that RO readers can test against shared locks without acquiring a write lock, which requires O_RDWR). Fam
