On 11/5/20 7:27 AM, Halil Pasic wrote:
On Wed, 4 Nov 2020 16:20:26 -0500
Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:
But I'm sure the code is suggesting it can, because
vfio_ap_mdev_filter_guest_matrix() has a third parameter called filter_apid,
which governs whether the apm or the aqm bit should be removed. And
vfio_ap_mdev_filter_guest_matrix() does get called with filter_apid=false in
assign_domain_store() and I don't see subsequent unlink operations that would
severe q->mdev_matrix.
I think you may be conflating two different things. The q in q->matrix_mdev
represents a queue device bound to the driver. The link to matrix_mdev
indicates the APQN of the queue device is assigned to the matrix_mdev.
When a new domain is assigned to matrix_mdev, we know that
all APQNS currently assigned to the shadow_apcb are bound to the vfio
driver
because of previous filtering, so we are only concerned with those APQNs
with the APQI of the new domain being assigned.
1. Queues bound to vfio_ap:
04.0004
04.0047
2. APQNs assigned to matrix_mdev:
04.0004
04.0047
3. shadow_apcb:
04.0004
04.0047
4. Assign domain 0054 to matrix_mdev
5. APQI 0054 gets filtered because 04.0054 not bound to vfio_ap
6. no change to shadow_apcb:
04.0004
04.0047
Let me please expand on your example. For reference see the filtering
code after the example.
Since our face to face discussion, I've made changes which
affect the scenario you laid out. The following shows the difference
in results using your scenario. Let me know what you think.
1. Queues bound to vfio_ap:
04.0004
04.0047
05.0004
05.0047
05.0054
2. APQNs assigned to matrix_mdev:
04.0004
04.0047
3. shadow_apcb:
04.0004
04.0047
4. Assign domain 0054 to matrix_mdev
5. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0054
5. APQI 0054 gets filtered because 04.0054 not bound to vfio_ap
6. no change to shadow_apcb:
04.0004
04.0047
7. assign adapter 05
8. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0054
05.0004
05.0047
05.0054
9. shadow_apcb changes to:
04.0004
04.0047
05.0004
05.0047
because adapter 05 is checked against the APQIs in the
shadow_apcb (0004, 0047) and since 05.0004 and
05.0047 are bound to the driver, adapter 05 is
hot plugged.
10. assign domain 0052
11. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0053
04.0054
05.0004
05.0047
05.0053
05.0054
11. shadow_apcb remains
04.0004
04.0047
05.0004
05.0047
because domain 0052 is checked against adapters assigned to
shadow_apcb and rejected because neither 04.0052 nor 05.0052
is bound to the vfio_ap driver.
12. 05.0054 gets removed (unbound)
13. Nothing is removed because 05.0054 is not assigned to shadow_apcb
14. unassign adapter 05
15. unassign domain 0053
16. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0054
17. shadow apcb is
04.0004
04.0047
16. assign adapter 05
15. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0054
05.0004
05.0047
05.0054
16. shadow_apcb changes to:
04.0004
04.0047
05.0004
05.0047
because adapter 05 is checked against APQIs (0004, 0047)
in shadow_apcb and since queues 05.0004 and 05.0047
are bound to vfio_ap, the adapter is hot plugged
1. Queues bound to vfio_ap:
04.0004
04.0047
05.0004
05.0047
05.0054
2. APQNs assigned to matrix_mdev:
04.0004
04.0047
3. shadow_apcb:
04.0004
04.0047
4. Assign domain 0054 to matrix_mdev
5. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0054
5. APQI 0054 gets filtered because 04.0054 not bound to vfio_ap
6. no change to shadow_apcb:
04.0004
04.0047
7. assign adapter 05
8. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0054
05.0004
05.0047
05.0054
9. shadow_apcb changes to:
05.0004
05.0047
05.0054
because now vfio_ap_mdev_filter_guest_matrix() is called with filter_apid=true
10. assign domain 0052
11. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0053
04.0054
05.0004
05.0047
05.0053
05.0054
11. shadow_apcb changes to
04.0004
04.0047
05.0004
05.0047
because now filter_guest_matrix() is called with filter_apid=false
and apqis 0053 and 0054 get filtered
12. 05.0054 gets removed (unbound)
13. with your current code we unplug adapter 05 from shadow_apcb
despite the fact that 05.0054 was not in the shadow_apcb in
the first place
14. unassign adapter 05
15. unassign domain 0053
16. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0054
17. shadow apcb is
04.0004
04.0047
16. assign adapter 05
15. APQNs assigned to matrix_mdev:
04.0004
04.0047
04.0054
05.0004
05.0047
05.0054
16. shadow_apcb changes to
<empty>
because now filter_guest_matrix() is called with filter_apid=true
and apqn 04 gets filtered because queues 04.0053 are not bound
and apqn 05 gets filtered because queues 05.0053 are not bound
static int vfio_ap_mdev_filter_guest_matrix(struct ap_matrix_mdev *matrix_mdev,
bool filter_apid)
{
struct ap_matrix shadow_apcb;
unsigned long apid, apqi, apqn;
memcpy(&shadow_apcb, &matrix_mdev->matrix, sizeof(struct ap_matrix));
for_each_set_bit_inv(apid, matrix_mdev->matrix.apm, AP_DEVICES) {
/*
* If the APID is not assigned to the host AP configuration,
* we can not assign it to the guest's AP configuration
*/
if (!test_bit_inv(apid, (unsigned long *)
matrix_dev->config_info.apm)) {
clear_bit_inv(apid, shadow_apcb.apm);
continue;
}
for_each_set_bit_inv(apqi, matrix_mdev->matrix.aqm,
AP_DOMAINS) {
/*
* If the APQI is not assigned to the host AP
* configuration, then it can not be assigned to the
* guest's AP configuration
*/
if (!test_bit_inv(apqi, (unsigned long *)
matrix_dev->config_info.aqm)) {
clear_bit_inv(apqi, shadow_apcb.aqm);
continue;
}
/*
* If the APQN is not bound to the vfio_ap device
* driver, then we can't assign it to the guest's
* AP configuration. The AP architecture won't
* allow filtering of a single APQN, so let's filter
* the APID.
*/
apqn = AP_MKQID(apid, apqi);
if (!vfio_ap_mdev_get_queue(matrix_mdev, apqn)) {
if (filter_apid) {
clear_bit_inv(apid, shadow_apcb.apm);
break;
}
clear_bit_inv(apqi, shadow_apcb.aqm);
continue;
}
}
I realize this scenario (to play through to the end) requires
manually unbound queue (more precisely queue missing not because
of host ap config or because of a[pq]mask), but just one 'hole' suffices.
I'm afraid, that I might be bitching around, because last time it was me
who downplayed the effects of such 'holes'.
Nevertheless, I would like to ask you to verify the scenario I've
sketched, or complain if I've gotten something wrong.
Your scenario looks correct and convinced me to change
the filtering logic giving the results I pointed out above.
It was a good catch on your part, so I thank you for
the review.
Regarding solutions to the problem. It makes no sense to talk about a
solution, before agreeing on the existence of the problem. Nevertheless
I will write down two sentences, mostly as a reminder to myself, for the
case we do agree on the existence of the problem. The simplest approach
is to always filter by apid. That way we get a quirky adapter unplug
right at steps 4, but it won't create the complicated mess we have in
the rest of the points. Another idea is to restrict the overprovisioning
of domains. Basically we would make the step 4 fail because we detected
a 'hole'. But this idea has its own problems, and in some scenarios
it does boil down to the unplug the adapter rule.
I made the following changes that I believe rectify this problem:
1. On guest startup, the shadow_apcb is initialized by filtering the
mdev's matrix by APID (i.e., if an APQN derived from a particular
APID and the APQIs assigned to the mdev's matrix does not
reference a queue device bound to vfio_ap, that APID is not
assigned to the shadow_apcb).
2. On adapter assignment, if each APQN derived from the APID
being assigned and the APQIs assigned to the shadow_apcb
does not reference a queue bound to vfio_ap, the adapter
will not be hot plugged.
3. On adapter unassignment, if the APID is set in the shadow_apcb,
the adapter will be hot unplugged.
4. On domain assignment, if each APQN derived from the APQI
being assigned and the APIDs assigned to the shadow_apcb
does not reference a queue bound to vfio_ap, the domain
will not be hot plugged.
5. On domain unassignment, if the APQI is set in the shadow_apcb,
the domain will be hot unplugged.
6. On probe:
a. For the queue's APID, the same logic as #2 will be used.
b. For the queue's APQI, the same logic as #4 will be used.
7. On remove, if the APQN of the queue being unbound is assigned
to the shadow_apcb, the adapter will be hot unplugged.
[..]
I'm not sure why you are bringing up unlinking in the context of assigning
a new domain. Unlinking only occurs when an APID or APQI is unassigned.
Are you certain? What about vfio_ap_mdev_on_cfg_remove()? I believe it
unplugs from the shadow_apcb, but it does not change the
assignments to the matrix_mdev. We do that so we know in remove that the
queue was already cleaned up, and does not need more cleanup.
What you say is true; however, that is not related to my comment.
I asked why you were bringing up unlinking in the context of
assigning a new domain. The point you just made above has to
do with unassigning adapters/domains.
Regards,
Halil
