On Sat, May 12, 2018 at 8:13 AM, Amir Goldstein <amir73il@xxxxxxxxx> wrote: > On Sat, May 12, 2018 at 4:29 AM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: >> [this one really, really needs a review by overlayfs folks; >> all I can promise is that it compiles.] >> >> vfs_mkdir() may succeed and leave the dentry passed to it unhashed and >> negative. ovl_create_real() is the last caller breaking when that >> happens. >> >> Make ovl_create_real() return the dentry to be used or ERR_PTR(-E...) >> in case of error; usually that'll be the dentry passed to it, but >> it might be a different one - result of lookup in case of vfs_mkdir() >> leaving looking the inode, etc. up to the next lookup to come. >> >> In that case (as well as in case of an error), original dentry is >> dropped. That simplifies the callers. Moreover, passing ERR_PTR() >> as dentry leads to immediate return of the same ERR_PTR(). That >> simplifies the callers even more. > > So in the beginning I had mixed feelings about the internal interface > change.. it feels shady, but it does simplify the callers... but then > I realized the correct way to simplify the callers would be a helper > ovl_create_real_temp(), because in most of the call sites > ovl_lookup_temp() is the argument to ovl_create_real(), so > your patch shouldn't bother with it. > > With that in mind, I think your mkdir fix should go into ovl_do_mkdir(), > hopefully, calling a vfs helper vfs_mkdir_hashed(), where all > of the above, including ovl_create_real() pass in a struct dentry ** > > I can prep and test the ovl patches if you like. > My tested version is at: https://github.com/amir73il/linux/commits/ovl-fixes The branch is based on a merge of your vfs/fixes and Miklos' vfs/overlayfs-rorw branches. It includes also a patch for overlayfs to use d_instantiate_new(). Will post the patches to overlayfs list shortly. Thanks, Amir.
