Hi, On Mon, Oct 27, 2014 at 10:31:42AM +1100, Andre Wolokita wrote: > I'm noticing some strange behaviour in the gadgetfs driver when > running gadgetfs-test; the program fails with the error "ep0 read > after poll: Invalid argument". > > As far as I understand, an inode is created upon an open() call in > gadgetfs-test and an initial fops is assigned to the struct file that > doesn't contain a .read member. > > The .write member of this first fops points to dev_config() in > usb/gadget/legacy/inode.c. > > When user-space write is called for the first time, dev_config() > performs a bunch of configuration stuff before doing "fd->f_op = > &ep0_io_operations;". This new file_operations struct does contain a > .read member, which points to ep0_read(). > > I'm confused as to why when an error occurs when a subsequent read() > is called in gadgetfs-test. I thought that changes to an fops pointer > is preserved upon returning to the user-space caller. How can it be > that "fd->f_op = &ep0_io_operations;" clearly changes the fops > pointer, yet these changes aren't reflected when returning to > user-space? probably a regression caused by another commit ? Can you try bisecting to pin-point what broke it ? Do you remember which was the last kernel where this worked ? -- balbi
Attachment:
signature.asc
Description: Digital signature
