Message-ID: <edb488a7-c89f-a8c9-48ff-b49a856cbcbc@linux.ibm.com>
Date:   Wed, 29 Apr 2020 09:56:05 +0200
From:   Harald Freudenberger <freude@...ux.ibm.com>
To:     Tony Krowiak <akrowiak@...ux.ibm.com>,
        Halil Pasic <pasic@...ux.ibm.com>
Cc:     linux-s390@...r.kernel.org, linux-kernel@...r.kernel.org,
        kvm@...r.kernel.org, borntraeger@...ibm.com, cohuck@...hat.com,
        mjrosato@...ux.ibm.com, pmorel@...ux.ibm.com,
        alex.williamson@...hat.com, kwankhede@...dia.com,
        jjherne@...ux.ibm.com, fiuczy@...ux.ibm.com
Subject: Re: [PATCH v7 01/15] s390/vfio-ap: store queue struct in hash table
 for quick access

On 29.04.20 00:30, Tony Krowiak wrote:
>
>
> On 4/28/20 6:57 AM, Harald Freudenberger wrote:
>> On 28.04.20 12:07, Halil Pasic wrote:
>>> On Mon, 27 Apr 2020 17:48:58 -0400
>>> Tony Krowiak <akrowiak@...ux.ibm.com> wrote:
>>>
>>>> On 4/27/20 11:17 AM, Halil Pasic wrote:
>>>>> On Mon, 27 Apr 2020 15:05:23 +0200
>>>>> Harald Freudenberger <freude@...ux.ibm.com> wrote:
>>>>>
>>>>>> On 24.04.20 05:57, Halil Pasic wrote:
>>>>>>> On Tue,  7 Apr 2020 15:20:01 -0400
>>>>>>> Tony Krowiak <akrowiak@...ux.ibm.com> wrote:
>>>>>>>   
>>>>>>>> Rather than looping over potentially 65535 objects, let's store the
>>>>>>>> structures for caching information about queue devices bound to the
>>>>>>>> vfio_ap device driver in a hash table keyed by APQN.
>>>>>>> @Harald:
>>>>>>> Would it make sense to make the efficient lookup of an apqueue base
>>>>>>> on its APQN core AP functionality instead of each driver figuring it out
>>>>>>> on it's own?
>>>>>>>
>>>>>>> If I'm not wrong the zcrypt device/driver(s) must the problem of
>>>>>>> looking up a queue based on its APQN as well.
>>>>>>>
>>>>>>> For instance struct ep11_cprb has a target_id filed
>>>>>>> (arch/s390/include/uapi/asm/zcrypt.h).
>>>>>>>
>>>>>>> Regards,
>>>>>>> Halil
>>>>>> Hi Halil
>>>>>>
>>>>>> no, the zcrypt drivers don't have this problem. They build up their own device object which
>>>>>> includes a pointer to the base ap device.
>>>>> I'm a bit confused. Doesn't your code loop first trough the ap_card
>>>>> objects to find the APID portion of the APQN, and then loop the queue
>>>>> list of the matching card to find the right ap_queue object? Or did I
>>>>> miss something? Isn't that what _zcrypt_send_ep11_cprb() does? Can you
>>>>> point me to the code that avoids the lookup (by apqn) for zcrypt?
>>>> The code you reference, _zcrypt_send_ep11_cprb(), does loop through
>>>> each queue associated with each card, but it doesn't appear to be
>>>> looking for
>>>> a queue with a particular APQN. It appears to be looking for a queue
>>>> meeting a specific set of conditions. At least that's my take after
>>>> taking a very
>>>> brief look at the code, so I'm not sure that applies here.
>>>>
>>> One of the possible conditions is that the APQN is in the targets array.
>>> Please have another look at the code below, is_desired_ep11_queue()
>>> and is_desired_ep11_card() do APQI and APID part of the check
>>> respectively:
>>>
>>>          for_each_zcrypt_card(zc) {
>>>                  /* Check for online EP11 cards */
>>>                  if (!zc->online || !(zc->card->functions & 0x04000000))
>>>                          continue;
>>>                  /* Check for user selected EP11 card */
>>>                  if (targets &&
>>>                      !is_desired_ep11_card(zc->card->id, target_num, targets))
>>>                          continue;
>>>                  /* check if device node has admission for this card */
>>>                  if (!zcrypt_check_card(perms, zc->card->id))
>>>                          continue;
>>>                  /* get weight index of the card device  */
>>>                  weight = speed_idx_ep11(func_code) * zc->speed_rating[SECKEY];
>>>                  if (zcrypt_card_compare(zc, pref_zc, weight, pref_weight))
>>>                          continue;
>>>                  for_each_zcrypt_queue(zq, zc) {
>>>                          /* check if device is online and eligible */
>>>                          if (!zq->online ||
>>>                              !zq->ops->send_ep11_cprb ||
>>>                              (targets &&
>>>                               !is_desired_ep11_queue(zq->queue->qid,
>>>                                                      target_num, targets)))
>>>
>>>
>>> Yes the size of targets may or may not be 1 (example for size == 1 is
>>> the invocation form ep11_cryptsingle()) and the respective costs
>>> depend on the usual size of the array. Since the goal of the whole
>>> exercise seems to be to pick a single queue, and we settle with the first
>>> suitable (first not in the input array, but in our lists) that is
>>> suitable, I assumed we wouldn't need many hashtable lookups.
>>>
>>> Regards,
>>> Halil
>> again, this is all code related to zcrypt card and queues and has nothing directly to do with ap queue and ap cards.
>> If you want to have a look how this works for ap devices, have a look into the scan routines for the ap bus in ap_bus.c
>> There you can find a bus_for_each_device() which would fit together with the right matching function for your needs.
>> And this is exactly what Tony implemented in the first shot. However, as written I can provide something like that
>> for you.
>> One note for the improvement via hash list with the argument about the max 65535 objects.
>> Think about a real big machine which has currently up to 30 crypto cards (z15 GA1.5) which when CEX7S are
>> plugged appear as 60 crypto adapters and have up to 85 domains each. When all these crypto resources
>> are assigned to one LPAR we end up in 60x85 = 5100 APQNs. Well, of course with a hash you can improve
>> the linear search through an array or list but can you measure the performance gain and then compare this
>> to the complexity.  ... just some thoughts about beautifying code ...
>
> I set up a test case to compare searching using a hashtable verses using a list.
> I created both a hashtable and a list of 5100 objects. Each structure had a single
> APQN field. I then randomly searched both the hashtable and the list for
> each APQN. The following table contains the result of 5 test runs. The elapsed
> times are in nanoseconds.
>
> Test:                              List Search    Hashtable Search
> ------                              ----------- ----------------
> Avg. Per APQN:             11651           81
> Total per 5500 APQNs:  60164268     1085368
>
> Avg. Per APQN:              10925           78
> Total per 5500 APQNs:   56482780    1084590
>
> Avg. Per APQN:              10190           80
> Total per 5500 APQNs:   52714920    1123205
>
> Avg. Per APQN:              8431             76
> Total per 5500 APQNs:   43748838    1061414
>
> Avg. Per APQN:              9678             75
> Total per 5500 APQNs:   50103437    1044427
> -----------------------------------------------
> Per APQN Search Avg:   10175          78            Hashtable is 130 times faster
> Total Search 5500 Avg:  52642848    1079800  Hashtable is 49 times faster
>
> Note that the list search was just a straight search of an object in a list, not
> a device attached to a bus. I don't know if that would add time, but it seems
> that the savings using a hashtable are significant.

Halil, I did not say that a hashtable is not faster than a linear list. The only thing
I wanted to express is that we are adding complexity and performance improving
code which is not even integrated somewhere. We are beautifying here.

>
> So I have two questions:
>
> 1. Would it make more sense to provide AP bus interfaces to search for
>     queue devices by APQN?
>
> 2. If so, shall we store the queue devices in a hashtable to make the
>     searches more efficient?
If there is a decision to implement this as a feature function within the AP
bus base code, I will use the bus functions provided by the kernel common
code. So here this will be a bus_for_each_device() together with a filter
function. I don't know how this is implemented within the common bus code.
>
>
>>>>> If you look at the new function of vfio_ap_get_queue(unsigned long apqn)
>>>>> it basically about finding the queue based on the apqn, with the
>>>>> difference that it is vfio specific.
>>>>>
>>>>> Regards,
>>>>> Halil
>>>>>
>>>>>> However, this is not a big issue, as the ap_bus holds a list of ap_card objects and within each
>>>>>> ap_card object there exists a list of ap_queues.
>>>>>
>

Powered by blists - more mailing lists