On Tuesday, April 27, 2021 12:03:27 AM CEST Matthew Wilcox wrote:
> On Mon, Apr 26, 2021 at 11:15:40PM +0200, Fabio M. De Francesco wrote:
> > On Monday, April 26, 2021 9:21:01 PM CEST Matthew Wilcox wrote:
> > > On Mon, Apr 26, 2021 at 08:42:45PM +0200, Fabio M. De Francesco wrote:
> > > > +static void setup_scsitaskmgmt_handles(struct xarray *xa, struct
> >
> > uiscmdrsp *cmdrsp,
> >
> > > > wait_queue_head_t *event,
> >
> > int *result)
> >
> > > > {
> > > >
> > > > - /* specify the event that has to be triggered when this */
> > > > - /* cmd is complete */
> > > > - cmdrsp->scsitaskmgmt.notify_handle =
> > > > - simple_idr_get(idrtable, event, lock);
> > > > - cmdrsp->scsitaskmgmt.notifyresult_handle =
> > > > - simple_idr_get(idrtable, result, lock);
> > > > + u32 id;
> > > > + int ret;
> > > > +
> > > > + /* specify the event that has to be triggered when this cmd is
> >
> > complete */
> >
> > > > + id = (u32)cmdrsp->scsitaskmgmt.notify_handle;
> > > > + ret = xa_alloc_irq(xa, &id, event, XA_LIMIT(1, INT_MAX),
> >
> > GFP_KERNEL);
> >
> > > OK, think this one through a bit. When xa_alloc_irq() stores the ID
that
> > > it assigned into 'id', what happens to it next?
> >
> > Oh, I overlooked that... The ID in 'id' is lost when the function exits
and
> > the stack frame is unwound.
> >
> > Now I have another problem: xa_alloc_irq() writes id u32* but I have an
u64*
> > in 'id'. What happens if I cast 'id' to an u32* when passing it to
> > xa_alloc_irq()?
> >
> > u64 *id;
> > int ret;
> > id = &cmdrsp->scsitaskmgmt.notify_handle;
> > ret = xa_alloc_irq(xa, (u32 *)id, event, XA_LIMIT(1, INT_MAX),
GFP_KERNEL);
> >
> > Do I destroy the information stored in 'id' with that cast?
>
> That is a great question! That would be a really serious bug because
> it behaves differently on big and little endian systems. That is, on a
> little endian system, a pointer to a u64 can be treated as a pointer to a
> u32 and it will write to the bottom 32 bits of the u64. On a big endian
> system, treating a pointer to a u64 as if it's a pointer to a u32 means
> you write to the _top_ 32 bits of the u64, and things go wrong from there!
>
> Similarly, if you have a u16, you can't pass a pointer to it, because
> the called function has no idea that it's only 16 bits, and will do a
> 32-bit store to it, overwriting the 16 bits after it.
>
> So you need to pass a pointer to a u32 on the stack, and then copy the
> id out of it afterwards.
>
As far as I understand, in setup_scsitaskmgmt_handles(), the task of
xa_alloc_irq() is: find two empty entries in 'xa', store the indexes into the
'id' pointer (before the first call this is set to &cmdrsp-
>scsitaskmgmt.notify_handle, and before the second call it is changed to
&cmdrsp->scsitaskmgmt.notifyresult_handle), then store the entries at their
respective indexes.
Indexes that xa_alloc_* set are of type u32*, so why not just change the type
of notify_handle and notifyresult_handle from u64 to u32?
Furthermore, I cannot understand why those indexes should be passed in and out
as arguments of the function . It seems that they are not needed anywhere else
in that file. Are they?
Maybe that I'm still missing something...
Thanks,
Fabio
