On Fri, 2022-12-16 at 15:10 -0500, Matthew Rosato wrote:
> On 11/21/22 4:40 PM, Eric Farman wrote:
> > Everything about this allocation is harder than necessary,
> > since the memory allocation is already aligned to our needs.
> > Break them apart for readability, instead of doing the
> > funky artithmetic.
> >
> > Of the structures that are involved, only ch_ccw needs the
> > GFP_DMA flag, so the others can be allocated without it.
> >
> > Signed-off-by: Eric Farman <farman@xxxxxxxxxxxxx>
> > ---
> > drivers/s390/cio/vfio_ccw_cp.c | 39 ++++++++++++++++++------------
> > ----
> > 1 file changed, 21 insertions(+), 18 deletions(-)
> >
> > diff --git a/drivers/s390/cio/vfio_ccw_cp.c
> > b/drivers/s390/cio/vfio_ccw_cp.c
> > index d41d94cecdf8..4b6b5f9dc92d 100644
> > --- a/drivers/s390/cio/vfio_ccw_cp.c
> > +++ b/drivers/s390/cio/vfio_ccw_cp.c
> > @@ -311,40 +311,41 @@ static inline int is_tic_within_range(struct
> > ccw1 *ccw, u32 head, int len)
> > static struct ccwchain *ccwchain_alloc(struct channel_program *cp,
> > int len)
> > {
> > struct ccwchain *chain;
> > - void *data;
> > - size_t size;
> > -
> > - /* Make ccw address aligned to 8. */
> > - size = ((sizeof(*chain) + 7L) & -8L) +
> > - sizeof(*chain->ch_ccw) * len +
> > - sizeof(*chain->ch_pa) * len;
> > - chain = kzalloc(size, GFP_DMA | GFP_KERNEL);
> > +
> > + chain = kzalloc(sizeof(*chain), GFP_KERNEL);
>
> I suppose you could consider a WARN_ONCE here if one of these
> kzalloc'd addresses has something in the low-order 3 bits; would
> probably make it more obvious if for some reason the alignment
> guarantee was broken vs some status after-the-fact in the IRB. But
> as per our discussion off-list I think that can only happen if
> ARCH_KMALLOC_MINALIGN were to change.
Yeah, maybe, but the "status after-the-fact" is a program check that
would be generated by the channel, just as would be done if the ORB was
located in a similarly-weird location (which we don't check for
either). Since this is all mainline paths, I don't think it makes sense
to re-check all those possible permutations here.
(And, for what it's worth, it's not this allocation that matters, but
rather the one that gets stuffed into the ORB below [1])
>
> > if (!chain)
> > return NULL;
> >
> > - data = (u8 *)chain + ((sizeof(*chain) + 7L) & -8L);
> > - chain->ch_ccw = (struct ccw1 *)data;
> > -
> > - data = (u8 *)(chain->ch_ccw) + sizeof(*chain->ch_ccw) *
> > len;
> > - chain->ch_pa = (struct page_array *)data;
> > + chain->ch_ccw = kcalloc(len, sizeof(*chain->ch_ccw),
> > GFP_DMA | GFP_KERNEL);
[1]
> > + if (!chain->ch_ccw)
> > + goto out_err;
> >
> > - chain->ch_len = len;
> > + chain->ch_pa = kcalloc(len, sizeof(*chain->ch_pa),
> > GFP_KERNEL);
> > + if (!chain->ch_pa)
> > + goto out_err;
> >
> > list_add_tail(&chain->next, &cp->ccwchain_list);
> >
> > return chain;
> > +
> > +out_err:
> > + kfree(chain->ch_ccw);
> > + kfree(chain);
> > + return NULL;
> > }
> >
> > static void ccwchain_free(struct ccwchain *chain)
> > {
> > list_del(&chain->next);
> > + kfree(chain->ch_pa);
> > + kfree(chain->ch_ccw);
> > kfree(chain);
> > }
> >
> > /* Free resource for a ccw that allocated memory for its cda. */
> > static void ccwchain_cda_free(struct ccwchain *chain, int idx)
> > {
> > - struct ccw1 *ccw = chain->ch_ccw + idx;
> > + struct ccw1 *ccw = &chain->ch_ccw[idx];
> >
> > if (ccw_is_tic(ccw))
> > return;
> > @@ -443,6 +444,8 @@ static int ccwchain_handle_ccw(u32 cda, struct
> > channel_program *cp)
> > chain = ccwchain_alloc(cp, len);
> > if (!chain)
> > return -ENOMEM;
> > +
> > + chain->ch_len = len;
> > chain->ch_iova = cda;
> >
> > /* Copy the actual CCWs into the new chain */
> > @@ -464,7 +467,7 @@ static int ccwchain_loop_tic(struct ccwchain
> > *chain, struct channel_program *cp)
> > int i, ret;
> >
> > for (i = 0; i < chain->ch_len; i++) {
> > - tic = chain->ch_ccw + i;
> > + tic = &chain->ch_ccw[i];
>
> These don't seem equivalent... Before at each iteration you'd offset
> tic by i bytes, now you're treating i as an index of 8B ccw1 structs,
> so it seems like this went from tic = x + i to tic = x + (8 * i)?
> Was the old code broken or am I missing something?
I think the latter. :) The old code did one allocation measured in
bytes, stored it in chain, and then calculated locations within that
for ch_ccw and ch_pa, cast to the respective pointer types. (See the
reference [1] above.)
So any use of "i" was an index into the pointer types and thus already
a "8 * i" addition from your example. My intention here was to remove
the pseudo-assembly above, and changed these along the way as I was un-
tangling everything. Looking at the resulting assembly before/after,
these hunks don't end up changing at all so I'll back these changes
back out. Especially since...
>
> >
> > if (!ccw_is_tic(tic))
> > continue;
> > @@ -739,8 +742,8 @@ int cp_prefetch(struct channel_program *cp)
> > list_for_each_entry(chain, &cp->ccwchain_list, next) {
> > len = chain->ch_len;
> > for (idx = 0; idx < len; idx++) {
> > - ccw = chain->ch_ccw + idx;
> > - pa = chain->ch_pa + idx;
> > + ccw = &chain->ch_ccw[idx];
> > + pa = &chain->ch_pa[idx];
>
> Same sort of question re: ch_pa
...this prompted me to notice that I didn't change the users of "chain-
>ch_pa + i" when calling page_array_unpin_free(), so now we have both
flavors which isn't ideal.
BEFORE:
ccw = chain->ch_ccw + idx;
pa = chain->ch_pa + idx;
1536: eb 3b 00 01 00 0d sllg %r3,%r11,1
153c: b9 08 00 3b agr %r3,%r11
1540: eb 33 00 03 00 0d sllg %r3,%r3,3
1546: e3 30 80 28 00 08 ag %r3,40(%r8)
ccw = chain->ch_ccw + idx;
154c: eb 2b 00 03 00 0d sllg %r2,%r11,3
1552: e3 20 80 10 00 08 ag %r2,16(%r8)
AFTER
ccw = &chain->ch_ccw[idx];
pa = &chain->ch_pa[idx];
15be: eb 3b 00 01 00 0d sllg %r3,%r11,1
15c4: b9 08 00 3b agr %r3,%r11
15c8: eb 33 00 03 00 0d sllg %r3,%r3,3
15ce: e3 30 80 28 00 08 ag %r3,40(%r8)
ccw = &chain->ch_ccw[idx];
15d4: eb 2b 00 03 00 0d sllg %r2,%r11,3
15da: e3 20 80 10 00 08 ag %r2,16(%r8)
>
> >
> > ret = ccwchain_fetch_one(ccw, pa, cp);
> > if (ret)
>
