On Mon, Dec 04, 2023 at 09:31:21PM -0700, Keith Busch wrote:
> On Tue, Dec 05, 2023 at 12:14:22PM +0800, Ming Lei wrote:
> > On Mon, Dec 04, 2023 at 11:57:55AM -0700, Keith Busch wrote:
> > > On Mon, Dec 04, 2023 at 01:40:58PM -0500, Jeff Moyer wrote:
> > > > I added a CC: linux-security-module@vger
> > > > Keith Busch <kbusch@xxxxxxxx> writes:
> > > > > From: Keith Busch <kbusch@xxxxxxxxxx>
> > > > >
> > > > > The uring_cmd operation is often used for privileged actions, so drivers
> > > > > subscribing to this interface check capable() for each command. The
> > > > > capable() function is not fast path friendly for many kernel configs,
> > > > > and this can really harm performance. Stash the capable sys admin
> > > > > attribute in the io_uring context and set a new issue_flag for the
> > > > > uring_cmd interface.
> > > >
> > > > I have a few questions. What privileged actions are performance
> > > > sensitive? I would hope that anything requiring privileges would not
> > > > be in a fast path (but clearly that's not the case).
> > >
> > > Protocol specifics that don't have a generic equivalent. For example,
> > > NVMe FDP is reachable only through the uring_cmd and ioctl interfaces,
> > > but you use it like normal reads and writes so has to be as fast as the
> > > generic interfaces.
> >
> > But normal read/write pt command doesn't require ADMIN any more since
> > commit 855b7717f44b ("nvme: fine-granular CAP_SYS_ADMIN for nvme io commands"),
> > why do you have to pay the cost of checking capable(CAP_SYS_ADMIN)?
>
> Good question. The "capable" check had always been first so even with
> the relaxed permissions, it was still paying the price. I have changed
> that order in commit staged here (not yet upstream):
>
> http://git.infradead.org/nvme.git/commitdiff/7be866b1cf0bf1dfa74480fe8097daeceda68622
With this change, I guess you shouldn't see the following big gap, right?
> Before: 970k IOPs
> After: 1750k IOPs
>
> Note that only prevents the costly capable() check if the inexpensive
> checks could make a determination. That's still not solving the problem
> long term since we aim for forward compatibility where we have no idea
> which opcodes, admin identifications, or vendor specifics could be
> deemed "safe" for non-root users in the future, so those conditions
> would always fall back to the more expensive check that this patch was
> trying to mitigate for admin processes.
Not sure I get the idea, it is related with nvme's permission model for
user pt command, and:
1) it should be always checked in entry of nvme user pt command
2) only the following two types of commands require ADMIN, per commit
855b7717f44b ("nvme: fine-granular CAP_SYS_ADMIN for nvme io commands")
- any admin-cmd is not allowed
- vendor-specific and fabric commmand are not allowed
Can you provide more details why the expensive check can't be avoided for
fast read/write user IO commands?
Thanks,
Ming
