[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH] xen/sched: fix locking in sched_tick_[suspend|resume]()


  • To: Jürgen Groß <jgross@xxxxxxxx>, <xen-devel@xxxxxxxxxxxxxxxxxxxx>
  • From: George Dunlap <george.dunlap@xxxxxxxxxx>
  • Date: Fri, 4 Oct 2019 15:56:15 +0100
  • Authentication-results: esa1.hc3370-68.iphmx.com; dkim=none (message not signed) header.i=none; spf=None smtp.pra=george.dunlap@xxxxxxxxxx; spf=Pass smtp.mailfrom=George.Dunlap@xxxxxxxxxx; spf=None smtp.helo=postmaster@xxxxxxxxxxxxxxx
  • Autocrypt: addr=george.dunlap@xxxxxxxxxx; prefer-encrypt=mutual; keydata= mQINBFPqG+MBEACwPYTQpHepyshcufo0dVmqxDo917iWPslB8lauFxVf4WZtGvQSsKStHJSj 92Qkxp4CH2DwudI8qpVbnWCXsZxodDWac9c3PordLwz5/XL41LevEoM3NWRm5TNgJ3ckPA+J K5OfSK04QtmwSHFP3G/SXDJpGs+oDJgASta2AOl9vPV+t3xG6xyfa2NMGn9wmEvvVMD44Z7R W3RhZPn/NEZ5gaJhIUMgTChGwwWDOX0YPY19vcy5fT4bTIxvoZsLOkLSGoZb/jHIzkAAznug Q7PPeZJ1kXpbW9EHHaUHiCD9C87dMyty0N3TmWfp0VvBCaw32yFtM9jUgB7UVneoZUMUKeHA fgIXhJ7I7JFmw3J0PjGLxCLHf2Q5JOD8jeEXpdxugqF7B/fWYYmyIgwKutiGZeoPhl9c/7RE Bf6f9Qv4AtQoJwtLw6+5pDXsTD5q/GwhPjt7ohF7aQZTMMHhZuS52/izKhDzIufl6uiqUBge 0lqG+/ViLKwCkxHDREuSUTtfjRc9/AoAt2V2HOfgKORSCjFC1eI0+8UMxlfdq2z1AAchinU0 eSkRpX2An3CPEjgGFmu2Je4a/R/Kd6nGU8AFaE8ta0oq5BSFDRYdcKchw4TSxetkG6iUtqOO ZFS7VAdF00eqFJNQpi6IUQryhnrOByw+zSobqlOPUO7XC5fjnwARAQABtCRHZW9yZ2UgVy4g RHVubGFwIDxkdW5sYXBnQHVtaWNoLmVkdT6JAlcEEwEKAEECGwMFCwkIBwMFFQoJCAsFFgID AQACHgECF4ACGQEWIQTXqBy2bTNXPzpOYFimNjwxBZC0bQUCXEowWQUJDCJ7dgAKCRCmNjwx BZC0beKvEACJ75YlJXd7TnNHgFyiCJkm/qPeoQ3sFGSDZuZh7SKcdt9+3V2bFEb0Mii1hQaz 3hRqZb8sYPHJrGP0ljK09k3wf8k3OuNxziLQBJyzvn7WNlE4wBEcy/Ejo9TVBdA4ph5D0YaZ nqdsPmxe/xlTFuSkgu4ep1v9dfVP1TQR0e+JIBa/Ss+cKC5intKm+8JxpOploAHuzaPu0L/X FapzsIXqgT9eIQeBEgO2hge6h9Jov3WeED/vh8kA7f8c6zQ/gs5E7VGALwsiLrhr0LZFcKcw kI3oCCrB/C/wyPZv789Ra8EXbeRSJmTjcnBwHRPjnjwQmetRDD1t+VyrkC6uujT5jmgOBzaj KCqZ8PcMAssOzdzQtKmjUQ2b3ICPs2X13xZ5M5/OVs1W3TG5gkvMh4YoHi4ilFnOk+v3/j7q 65FG6N0JLb94Ndi80HkIOQQ1XVGTyu6bUPaBg3rWK91Csp1682kD/dNVF3FKHrRLmSVtmEQR 5rK0+VGc/FmR6vd4haKGWIRuPxzg+pBR77avIZpU7C7+UXGuZ5CbHwIdY8LojJg2TuUdqaVj yxmEZLOA8rVHipCGrslRNthVbJrGN/pqtKjCClFZHIAYJQ9EGLHXLG9Pj76opfjHij3MpR3o pCGAh6KsCrfrsvjnpDwqSbngGyEVH030irSk4SwIqZ7FwLkBDQRUWmc6AQgAzpc8Ng5Opbrh iZrn69Xr3js28p+b4a+0BOvC48NfrNovZw4eFeKIzmI/t6EkJkSqBIxobWRpBkwGweENsqnd 0qigmsDw4N7J9Xx0h9ARDqiWxX4jr7u9xauI+CRJ1rBNO3VV30QdACwQ4LqhR/WA+IjdhyMH wj3EJGE61NdP/h0zfaLYAbvEg47/TPThFsm4m8Rd6bX7RkrrOgBbL/AOnYOMEivyfZZKX1vv iEemAvLfdk2lZt7Vm6X/fbKbV8tPUuZELzNedJvTTBS3/l1FVz9OUcLDeWhGEdlxqXH0sYWh E9+PXTAfz5JxKH+LMetwEM8DbuOoDIpmIGZKrZ+2fQARAQABiQNbBBgBCgAmAhsCFiEE16gc tm0zVz86TmBYpjY8MQWQtG0FAlxKMJ4FCQnQ/OQBKcBdIAQZAQoABgUCVFpnOgAKCRCyFcen x4Qb7cXrCAC0qQeEWmLa9oEAPa+5U6wvG1t/mi22gZN6uzQXH1faIOoDehr7PPESE6tuR/vI CTTnaSrd4UDPNeqOqVF07YexWD1LDcQG6PnRqC5DIX1RGE3BaSaMl2pFJP8y+chews11yP8G DBbxaIsTcHZI1iVIC9XLhoeegWi84vYc8F4ziADVfowbmbvcVw11gE8tmALCwTeBeZVteXjh 0OELHwrc1/4j4yvENjIXRO+QLIgk43kB57Upr4tP2MEcs0odgPM+Q+oETOJ00xzLgkTnLPim C1FIW2bOZdTj+Uq6ezRS2LKsNmW+PRRvNyA5ojEbA/faxmAjMZtLdSSSeFK8y4SoCRCmNjwx BZC0bevWEACRu+GyQgrdGmorUptniIeO1jQlpTiP5WpVnk9Oe8SiLoXUhXXNj6EtzyLGpYmf kEAbki+S6WAKnzZd3shL58AuMyDxtFNNjNeKJOcl6FL7JPBIIgIp3wR401Ep+/s5pl3Nw8Ii 157f0T7o8CPb54w6S1WsMkU78WzTxIs/1lLblSMcvyz1Jq64g4OqiWI85JfkzPLlloVf1rzy ebIBLrrmjhCE2tL1RONpE/KRVb+Q+PIs5+YcZ+Q1e0vXWA7NhTWFbWx3+N6WW6gaGpbFbopo FkYRpj+2TA5cX5zW148/xU5/ATEb5vdUkFLUFVy5YNUSyeBHuaf6fGmBrDc47rQjAOt1rmyD 56MUBHpLUbvA6NkPezb7T6bQpupyzGRkMUmSwHiLyQNJQhVe+9NiJJvtEE3jol0JVJoQ9WVn FAzPNCgHQyvbsIF3gYkCYKI0w8EhEoH5FHYLoKS6Jg880IY5rXzoAEfPvLXegy6mhYl+mNVN QUBD4h9XtOvcdzR559lZuC0Ksy7Xqw3BMolmKsRO3gWKhXSna3zKl4UuheyZtubVWoNWP/bn vbyiYnLwuiKDfNAinEWERC8nPKlv3PkZw5d3t46F1Dx0TMf16NmP+azsRpnMZyzpY8BL2eur feSGAOB9qjZNyzbo5nEKHldKWCKE7Ye0EPEjECS1gjKDwbkBDQRUWrq9AQgA7aJ0i1pQSmUR 6ZXZD2YEDxia2ByR0uZoTS7N0NYv1OjU8v6p017u0Fco5+Qoju/fZ97ScHhp5xGVAk5kxZBF DT4ovJd0nIeSr3bbWwfNzGx1waztfdzXt6n3MBKr7AhioB1m+vuk31redUdnhbtvN7O40MC+ fgSk5/+jRGxY3IOVPooQKzUO7M51GoOg4wl9ia3H2EzOoGhN2vpTbT8qCcL92ZZZwkBRldoA Wn7c1hEKSTuT3f1VpSmhjnX0J4uvKZ1V2R7rooKJYFBcySC0wa8aTmAtAvLgfcpe+legOtgq DKzLuN45xzEjyjCiI521t8zxNMPJY9FiCPNv0sCkDwARAQABiQI8BBgBCgAmAhsMFiEE16gc tm0zVz86TmBYpjY8MQWQtG0FAlxKNJYFCQnQrVkACgkQpjY8MQWQtG2Xxg//RrRP+PFYuNXt 9C5hec/JoY24TkGPPd2tMC9usWZVImIk7VlHlAeqHeE0lWU0LRGIvOBITbS9izw6fOVQBvCA Fni56S12fKLusWgWhgu03toT9ZGxZ9W22yfw5uThSHQ4y09wRWAIYvhJsKnPGGC2KDxFvtz5 4pYYNe8Icy4bwsxcgbaSFaRh+mYtts6wE9VzyJvyfTqbe8VrvE+3InG5rrlNn51AO6M4Wv20 iFEgYanJXfhicl0WCQrHyTLfdB5p1w+072CL8uryHQVfD0FcDe+J/wl3bmYze+aD1SlPzFoI MaSIXKejC6oh6DAT4rvU8kMAbX90T834Mvbc3jplaWorNJEwjAH/r+v877AI9Vsmptis+rni JwUissjRbcdlkKBisoUZRPmxQeUifxUpqgulZcYwbEC/a49+WvbaYUriaDLHzg9xisijHwD2 yWV8igBeg+cmwnk0mPz8tIVvwi4lICAgXob7HZiaqKnwaDXs4LiS4vdG5s/ElnE3rIc87yru 24n3ypeDZ6f5LkdqL1UNp5/0Aqbr3EiN7/ina4YVyscy9754l944kyHnnMRLVykg0v+kakj0 h0RJ5LbfLAMM8M52KIA3y14g0Fb7kHLcOUMVcgfQ3PrN6chtC+5l6ouDIlSLR3toxH8Aam7E rIFfe2Dk+lD9A9BVd2rfoHA=
  • Cc: George Dunlap <george.dunlap@xxxxxxxxxxxxx>, Dario Faggioli <dfaggioli@xxxxxxxx>
  • Delivery-date: Fri, 04 Oct 2019 14:56:21 +0000
  • Ironport-sdr: ieMC28sULzVT/MCV513V0DXv9TSVwCZVEPs+GWSdK9TW64U03hGn494atH3OCXrr0YL+NuhfCL pyp1QJZ9nPzXQ07zA3AdvDOYKTBB+cxRz4UwdIEF3n1WTDZBv/5Imil69KNiA1Ih2ZCLp6Te8x rcXHW+4hQC3JRHhPqNBqUK0VRpIS6BQILwm7OKae4frFZc8uwG9iC4z1YUct9t7sn+7MvkPit+ tYDfEgE47uDf/Zt0lVTNmKDb5Ec5zebHVIR7GJoKnhL6A/Nt2sJQf7Ax8yjvGnyoIgxu9ynr0U kJg=
  • List-id: Xen developer discussion <xen-devel.lists.xenproject.org>
  • Openpgp: preference=signencrypt

On 10/4/19 3:43 PM, Jürgen Groß wrote:
> On 04.10.19 16:34, George Dunlap wrote:
>> On 10/4/19 3:24 PM, Jürgen Groß wrote:
>>> On 04.10.19 16:08, George Dunlap wrote:
>>>> On 10/4/19 7:40 AM, Juergen Gross wrote:
>>>>> sched_tick_suspend() and sched_tick_resume() should not call the
>>>>> scheduler specific timer handlers in case the cpu they are running on
>>>>> is just being moved to or from a cpupool.
>>>>>
>>>>> Use a new percpu lock for that purpose.
>>>>
>>>> Is there a reason we can't use the pcpu_schedule_lock() instead of
>>>> introducing a new one?  Sorry if this is obvious, but it's been a while
>>>> since I poked around this code.
>>>
>>> Lock contention would be higher especially with credit2 which is using a
>>> per-core or even per-socket lock. We don't care about other scheduling
>>> activity here, all we need is a guard against our per-cpu scheduler
>>> data being changed beneath our feet.
>>
>> Is this code really being called so often that we need to worry about
>> this level of contention?
> 
> Its called each time idle is entered and left again.
> 
> Especially with core scheduling there is a high probability of multiple
> cpus leaving idle at the same time and the per-scheduler lock being used
> in parallel already.

Hrm, that does sound pretty bad.

>> We already have a *lot* of locks; and in this case you're adding a
>> second lock which interacts with the per-scheduler cpu lock.  This just
>> seems like asking for trouble.
> 
> In which way does it interact with the per-scheduler cpu lock?
> 
>> I won't Nack the patch, but I don't think I would ack it without clear
>> evidence that the extra lock has a performance improvement that's worth
>> the cost of the extra complexity.
> 
> I think complexity is lower this way. Especially considering the per-
> scheduler lock changing with moving a cpu to or from a cpupool.

The key aspect of the per-scheduler lock is that once you hold it, the
pointer to the lock can't change.

After this patch, the fact remains that sometimes you need to grab one
lock, sometimes the other, and sometimes both.

And, tick_suspend() lives in the per-scheduler code.  Each scheduler has
to remember that tick_suspend and tick_resume hold a completely
different lock to the rest of the scheduling functions.

 -George

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.