Message-ID: <e057f2e4-70d6-5078-62c5-748bfe16d865@linux.ibm.com>
Date:   Tue, 29 Sep 2020 17:59:42 -0400
From:   Tony Krowiak <akrowiak@...ux.ibm.com>
To:     Halil Pasic <pasic@...ux.ibm.com>
Cc:     linux-s390@...r.kernel.org, linux-kernel@...r.kernel.org,
        kvm@...r.kernel.org, freude@...ux.ibm.com, borntraeger@...ibm.com,
        cohuck@...hat.com, mjrosato@...ux.ibm.com,
        alex.williamson@...hat.com, kwankhede@...dia.com,
        fiuczy@...ux.ibm.com, frankja@...ux.ibm.com, david@...hat.com,
        imbrenda@...ux.ibm.com, hca@...ux.ibm.com, gor@...ux.ibm.com
Subject: Re: [PATCH v10 08/16] s390/vfio-ap: filter matrix for unavailable
 queue devices



On 9/26/20 4:24 AM, Halil Pasic wrote:
> On Fri, 21 Aug 2020 15:56:08 -0400
> Tony Krowiak <akrowiak@...ux.ibm.com> wrote:
>
>> Even though APQNs for queues that are not in the host's AP configuration
>> may be assigned to a matrix mdev, we do not want to set bits in the guest's
>> APCB for APQNs that do not reference AP queue devices bound to the vfio_ap
>> device driver. Ideally, it would be great if such APQNs could be filtered
>> out before setting the bits in the guest's APCB; however, the architecture
>> precludes filtering individual APQNs. Consequently, either the APID or APQI
>> must be filtered.
>>
>> This patch introduces code to filter the APIDs or APQIs assigned to the
>> matrix mdev's AP configuration before assigning them to the guest's AP
>> configuration (i.e., APCB). We'll start by filtering the APIDs:
>>
>>     If an APQN assigned to the matrix mdev's AP configuration does not
>>     reference a queue device bound to the vfio_ap device driver, the APID
>>     will be filtered out (i.e., not assigned to the guest's APCB).
>>
>> If every APID assigned to the matrix mdev is filtered out, then we'll try
>> filtering the APQI's:
>>
>>     If an APQN assigned to the matrix mdev's AP configuration does not
>>     reference a queue device bound to the vfio_ap device driver, the APQI
>>     will be filtered out (i.e., not assigned to the guest's APCB).
>>
>> In any case, if after filtering either the APIDs or APQIs there are any
>> APQNs that can be assigned to the guest's APCB, they will be assigned and
>> the CRYCB will be hot plugged into the guest.
>>
>> Example
>> =======
>>
>> APQNs bound to vfio_ap device driver:
>>     04.0004
>>     04.0047
>>     04.0054
>>
>>     05.0005
>>     05.0047
>>     05.0054
>>
>> Assignments to matrix mdev:
>>     APIDs  APQIs  -> APQNs
>>     04     0004      04.0004
>>     05     0005      04.0005
>>            0047      04.0047
>>            0054      04.0054
>>                      05.0004
>>                      05.0005
>>                      05.0047
>>                      04.0054
>>
>> Filter APIDs:
>>     APID 04 will be filtered because APQN 04.0005 is not bound.
>>     APID 05 will be filtered because APQN 05.0004 is not bound.
>>     APQNs remaining: None
>>
>> Filter APQIs:
>>     APQI 04 will be filtered because APQN 05.0004 is not bound.
>>     APQI 05 will be filtered because APQN 04.0005 is not bound.
>>     APQNs remaining: 04.0047, 04.0054, 05.0047, 05.0054
>>
>> APQNs 04.0047, 04.0054, 05.0047, 05.0054 will be assigned to the CRYCB and
>> hot plugged into the KVM guest.
>>
> I find this logic where we first do one strategy, and if nothing remains
> do the other strategy a little confusing. I will ramble on about it some
> more in the code.
>
>> Signed-off-by: Tony Krowiak <akrowiak@...ux.ibm.com>
>> ---
>>   drivers/s390/crypto/vfio_ap_ops.c | 159 +++++++++++++++++++++++++++++-
>>   1 file changed, 155 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
>> index 30bf23734af6..eaf4e9eab6cb 100644
>> --- a/drivers/s390/crypto/vfio_ap_ops.c
>> +++ b/drivers/s390/crypto/vfio_ap_ops.c
>> @@ -326,7 +326,7 @@ static bool vfio_ap_mdev_has_crycb(struct ap_matrix_mdev *matrix_mdev)
>>   	return (matrix_mdev->kvm && matrix_mdev->kvm->arch.crypto.crycbd);
>>   }
>>   
>> -static void vfio_ap_mdev_commit_crycb(struct ap_matrix_mdev *matrix_mdev)
>> +static void vfio_ap_mdev_commit_shadow_apcb(struct ap_matrix_mdev *matrix_mdev)
>>   {
>>   	kvm_arch_crypto_set_masks(matrix_mdev->kvm,
>>   				  matrix_mdev->shadow_apcb.apm,
>> @@ -597,6 +597,157 @@ static int vfio_ap_mdev_verify_no_sharing(struct ap_matrix_mdev *matrix_mdev,
>>   	return 0;
>>   }
>>   
>> +/**
>> + * vfio_ap_mdev_filter_matrix
>> + *
>> + * Filter APQNs assigned to the matrix mdev that do not reference an AP queue
>> + * device bound to the vfio_ap device driver.
>> + *
>> + * @matrix_mdev:  the matrix mdev whose AP configuration is to be filtered
>> + * @shadow_apcb:  the shadow of the KVM guest's APCB (contains AP configuration
>> + *		  for guest)
>> + * @filter_apids: boolean value indicating whether the APQNs shall be filtered
>> + *		  by APID (true) or by APQI (false).
>> + *
>> + * Returns the number of APQNs remaining after filtering is complete.
>> + */
>> +static int vfio_ap_mdev_filter_matrix(struct ap_matrix_mdev *matrix_mdev,
>> +				      struct ap_matrix *shadow_apcb,
>> +				      bool filter_apids)
>> +{
>> +	unsigned long apid, apqi, apqn;
>> +
>> +	memcpy(shadow_apcb, &matrix_mdev->matrix, sizeof(*shadow_apcb));
>> +
>> +	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->info.apm)) {
> The patch description and the code seem to be out of sync. Here you do
> some filtering based on the host's  AP config info read at module read at
> module initialization time.
>
>> +			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->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 if we're
>> +			 * filtering APIDs, then filter the APID; otherwise,
>> +			 * filter the APQI.
>> +			 */
>> +			apqn = AP_MKQID(apid, apqi);
>> +			if (!vfio_ap_get_queue(apqn)) {
> Is this really gonna give NULL if the queue is not bound to vfio-ap? I
> don't think so. This will get NULL if the queue is not known to the AP
> bus, or has no driver-data assigned. In the current state it should give
> you non-NULL if another driver has the queue, and maintains it's own
> driver specific data in drvdata.

It will not give you a NULL if the zcrypt driver has the queue because
no zcrypt driver sets any drvdata. You do bring up a good point though
because there is no guarantee that another driver will never set
the driver data for a queue device. Consequently, I will be changing the
vfio_ap_get_queue(apqn) function to check the driver associated with
the device and return NULL if it is not the vfio_ap driver.

>
>> +				if (filter_apids)
>> +					clear_bit_inv(apid, shadow_apcb->apm);
>> +				else
>> +					clear_bit_inv(apqi, shadow_apcb->aqm);
>> +				break;
>> +			}
>> +		}
>> +
>> +		/*
>> +		 * If we're filtering APQIs and all of them have been filtered,
>> +		 * there's no need to continue filtering.
>> +		 */
>> +		if (!filter_apids)
>> +			if (bitmap_empty(shadow_apcb->aqm, AP_DOMAINS))
>> +				break;
>> +	}
>> +
>> +	return bitmap_weight(shadow_apcb->apm, AP_DEVICES) *
>> +	       bitmap_weight(shadow_apcb->aqm, AP_DOMAINS);
>> +}
>> +
>> +/**
>> + * vfio_ap_mdev_config_shadow_apcb
>> + *
>> + * Configure the shadow of a KVM guest's APCB specifying the adapters, domains
>> + * and control domains to be assigned to the guest. The shadow APCB will be
>> + * configured after filtering the APQNs assigned to the matrix mdev that do not
>> + * reference a queue device bound to the vfio_ap device driver.
>> + *
>> + * @matrix_mdev: the matrix mdev whose shadow APCB is to be configured.
>> + *
>> + * Returns true if the shadow APCB contents have been changed; otherwise,
>> + * returns false.
>> + */
>> +static bool vfio_ap_mdev_config_shadow_apcb(struct ap_matrix_mdev *matrix_mdev)
>> +{
>> +	int napm, naqm;
>> +	struct ap_matrix shadow_apcb;
>> +
>> +	vfio_ap_matrix_init(&matrix_dev->info, &shadow_apcb);
>> +	napm = bitmap_weight(matrix_mdev->matrix.apm, AP_DEVICES);
>> +	naqm = bitmap_weight(matrix_mdev->matrix.aqm, AP_DOMAINS);
>> +
>> +	/*
>> +	 * If there are no APIDs or no APQIs assigned to the matrix mdev,
>> +	 * then no APQNs shall be assigned to the guest CRYCB.
>> +	 */
>> +	if ((napm != 0) || (naqm != 0)) {
>> +		/*
>> +		 * Filter the APIDs assigned to the matrix mdev for APQNs that
>> +		 * do not reference an AP queue device bound to the driver.
>> +		 */
>> +		napm = vfio_ap_mdev_filter_matrix(matrix_mdev, &shadow_apcb,
>> +						  true);
>> +		/*
>> +		 * If there are no APQNs that can be assigned to the guest's
>> +		 * CRYCB after filtering, then try filtering the APQIs.
>> +		 */
>> +		if (napm == 0) {
> When do we expect this to happen? Currently we don't assign queues that
> are not bound to us, and we have ->in_use() that inhibits disappearance
> of queues due to re-partitioning.

This will happen when domains are over-provisioned for a matrix
mdev. Suppose the following APQNs are assigned to the matrix
mdev:

00.0000
00.0004
00.0047

If queue 00.0047 is not bound to the vfio_ap device driver, then
the filtering code will filter APID 00.

Is your objection that this patch occurs prior to the
patch that implements over-provisioning? I guess your confusion
makes sense given over-provisioning is not introduced until after
this patch. Maybe I should re-order these to patches.

>
> So what we are left with is queue becomes unavailable to the host
> because of a config change, and maybe manual unbind -- not sure about
> that.

A queue can be removed from the vfio_ap device driver for the
following reasons:

1. The apmask or aqmask change can result in a queue device
     being unbound from vfio_ap.

2. A queue device can be manually unbound from vfio_ap.

3. A queue device can be unbound due to dynamic deconfiguration of
     the adapter via the SE or SCLP Deconfigure Adjunct Processor
     command (i.e., a configuration change)

>
> Now if matrix_dev->info was to reflect the config the bus acts by, which
> seems to the idea behind patch 12 we could react accordingly (if the
> domain is gone filter aqm).

That is handled in patch 13 which is the callback that handles the
notification introduced in patch 12. That patch does not use this
filtering code, however.

>
> I mean, the purpose of this callback seems to be getting us out of
> trouble when domains are missing across all cards (i.e. some domains
> were assigned away from us on the lower level).
>
> Or am I missing something?

I think you are missing the fact that there are other reasons
why a queue device may not be bound to vfio_ap (see reasons
above).

>
>> +			naqm = vfio_ap_mdev_filter_matrix(matrix_mdev,
>> +							  &shadow_apcb, false);
>> +
>> +			/*
>> +			 * If there are no APQNs that can be assigned to the
>> +			 * matrix mdev after filtering the APQIs, then no APQNs
>> +			 * shall be assigned to the guest's CRYCB.
>> +			 */
>> +			if (naqm == 0) {
>> +				bitmap_clear(shadow_apcb.apm, 0, AP_DEVICES);
>> +				bitmap_clear(shadow_apcb.aqm, 0, AP_DOMAINS);
>> +			}
>> +		}
>> +	}
>> +
>> +	/*
>> +	 * If the guest's AP configuration has not changed, then return
>> +	 * indicating such.
>> +	 */
>> +	if (bitmap_equal(matrix_mdev->shadow_apcb.apm, shadow_apcb.apm,
>> +			 AP_DEVICES) &&
>> +	    bitmap_equal(matrix_mdev->shadow_apcb.aqm, shadow_apcb.aqm,
>> +			 AP_DOMAINS) &&
>> +	    bitmap_equal(matrix_mdev->shadow_apcb.adm, shadow_apcb.adm,
>> +			 AP_DOMAINS))
>> +		return false;
>> +
>> +	/*
>> +	 * Copy the changes to the guest's CRYCB, then return indicating that
>> +	 * the guest's AP configuration has changed.
>> +	 */
>> +	memcpy(&matrix_mdev->shadow_apcb, &shadow_apcb, sizeof(shadow_apcb));
>> +
>> +	return true;
>> +}
>> +
>>   enum qlink_type {
>>   	LINK_APID,
>>   	LINK_APQI,
>> @@ -1284,9 +1435,8 @@ static int vfio_ap_mdev_group_notifier(struct notifier_block *nb,
>>   	if (!vfio_ap_mdev_has_crycb(matrix_mdev))
>>   		return NOTIFY_DONE;
>>   
>> -	memcpy(&matrix_mdev->shadow_apcb, &matrix_mdev->matrix,
>> -	       sizeof(matrix_mdev->shadow_apcb));
>> -	vfio_ap_mdev_commit_crycb(matrix_mdev);
>> +	if (vfio_ap_mdev_config_shadow_apcb(matrix_mdev))
>> +		vfio_ap_mdev_commit_shadow_apcb(matrix_mdev);
>>   
>>   	return NOTIFY_OK;
>>   }
>> @@ -1396,6 +1546,7 @@ static void vfio_ap_mdev_release(struct mdev_device *mdev)
>>   	mutex_lock(&matrix_dev->lock);
>>   	if (matrix_mdev->kvm) {
>>   		kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
>> +		vfio_ap_matrix_clear_masks(&matrix_mdev->shadow_apcb);
>>   		matrix_mdev->kvm->arch.crypto.pqap_hook = NULL;
>>   		vfio_ap_mdev_reset_queues(mdev);
>>   		kvm_put_kvm(matrix_mdev->kvm);

Powered by blists - more mailing lists