On Sun, Oct 04, 2020 at 10:31:19AM -0400, Vivek Goyal wrote: > On Fri, Oct 02, 2020 at 10:44:37PM -0400, Qian Cai wrote: > > On Fri, 2020-10-02 at 12:28 -0400, Qian Cai wrote: > > > Running some fuzzing on virtiofs from a non-privileged user could trigger a > > > warning in virtio_fs_enqueue_req(): > > > > > > WARN_ON(out_sgs + in_sgs != total_sgs); > > > > Okay, I can reproduce this after running for a few hours: > > > > out_sgs = 3, in_sgs = 2, total_sgs = 6 > > Thanks. I can also reproduce it simply by calling. > > ioctl(fd, 0x5a004000, buf); > > I think following WARN_ON() is not correct. > > WARN_ON(out_sgs + in_sgs != total_sgs) > > toal_sgs should actually be max sgs. It looks at ap->num_pages and > counts one sg for each page. And it assumes that same number of > pages will be used both for input and output. > > But there are no such guarantees. With above ioctl() call, I noticed > we are using 2 pages for input (out_sgs) and one page for output (in_sgs). > > So out_sgs=4, in_sgs=3 and total_sgs=8 and warning triggers. > > I think total sgs is actually max number of sgs and warning > should probably be. > > WARN_ON(out_sgs + in_sgs > total_sgs) > > Stefan, WDYT? It should be possible to calculate total_sgs precisely (not a maximum). Treating it as a maximum could hide bugs. Maybe sg_count_fuse_req() should count in_args/out_args[numargs - 1].size pages instead of adding ap->num_pages. Do you have the details of struct fuse_req and struct fuse_args_pages fields for the ioctl in question? Thanks, Stefan
Attachment:
signature.asc
Description: PGP signature
