Re: [dpdk-dev] [PATCH] eal/ppc: remove fix of memory barrier for IBM POWER

[Shahaf Shuler <shahafs@mellanox.com>]

Pradeep Satyanarayana <pradeep@us.ibm.com> wrote on Thu 3/21/2019 12:41 AM:
>Thomas Monjalon <thomas@monjalon.net> wrote on 03/19/2019 01:45:01 PM:
>
>> From: Thomas Monjalon <thomas@monjalon.net>
>> To: Shahaf Shuler <shahafs@mellanox.com>
>> Cc: Dekel Peled <dekelp@mellanox.com>, Chao Zhu
>> <chaozhu@linux.vnet.ibm.com>, Yongseok Koh <yskoh@mellanox.com>,
>> "dev@dpdk.org" <dev@dpdk.org>, Ori Kam <orika@mellanox.com>,
>> "stable@dpdk.org" <stable@dpdk.org>, pradeep@us.ibm.com
>> Date: 03/19/2019 01:45 PM
>> Subject: Re: [PATCH] eal/ppc: remove fix of memory barrier for IBM POWER
>>

[...]

>> >
>> > So far, when not running on power, we used the rte_wmb for that.
>> On x86 and ARM systems it provided the needed guarantees.
>> > It is also mentioned in the barrier doxygen on ARM arch:
>> > "
>> > Write memory barrier.
>> >
>> > Guarantees that the STORE operations generated before the barrier
>> > occur before the STORE operations generated after.
>> > "
>> >
>> > It doesn't restrict to store to system memory only.
>> > w/ power is on somewhat different and in fact rte_mb is required.
>> It obviously miss the point of those barrier if we will need to use
>> a different barrier based on the system arch.
>> >
>> > We need to align the definition of the different barriers in DPDK:
>> > 1. need a clear documentation of each. this should be global and
>> not part of the specific implementation on each arch.
>
>A single approach may not work for all architectures. Power is different
>from others, so we need to be able to accommodate that. More comments below.

it don't get this claim.
It is ok to have some differences between the different arch, but here you implement a well-defined barrier - rte_wmb.
if you see a need we can discuss to define a **new** barrier which sync STORE only to system memory, and will be able to utilize the lwsync command.

>
>>
>> The global definition is in lib/librte_eal/common/include/generic/rte_atomic.h
>>
>> There are some copy/paste in Arm32 and PPC that I will remove.
>>
>> > 2. either modify ppc rte_wmb to match ARM and x86 ones or to
>> define a new type of barrier which will sync between both I/O and
>> stores to systems memory.
>>
>> The basic memory barrier of DPDK does not mention
>> a difference between I/O and system memory.
>
>In the case of Power, sync will cater to both I/O and system memory. However, that
>is too big a hammer in all cases.

rte_wmb requires such sync. you propose to have the wrong barrier in favor of performance.
to mitigate this you can take my suggestion above and define a new, more lightweight one.

>
>> It is not explicit (yet) but I assume it is protecting both.
>> So, in my opinion, we need to make it explicit in the doc,
>> and fix the PPC implementation to comply with this definition.
>>
>> Anyway, I don't see any significant effort from IBM to move from
>> the alpha support stage to a real Open Source support.
>> PS: sending a mail every two months, to promise improvements, is not enough!
>

[...]

>
>We should retain lwsync, should not be removed as discussed in here:
>
>http://mails.dpdk.org/archives/dev/2019-March/126746.html

i don't agree.
it is very clear the rte_wmb implementation in power is broken and we need to fix this right away before other customers will hit the same issue.