Hi Ferruh,
I've just send a new corrected version of the patchset, which takes into
account the comments of Wie and Jeff, hope it is OK now.
Kind regards and thank you,
Renata
On 06/19/2020 06:54 PM, Ferruh Yigit wrote:
On 5/18/2020 10:48 AM, Renata Saiakhova wrote:
Hi Ferruh,
thanks for comments,
are the rte_eth_dma_zone_reserve() calls always used to allocate HW rings? It is
not totally clear to me. That was partly the reason I don't do the fix for every
driver which uses this API. I made fixes in the drivers which uses the same
pattern to allocate / release queues, for other drivers I was not sure but
anyway I couldn't spend more time for further investigations. In the company I
work for we use dpdk for our project and maintain it in separate tree, and the
vulnerability with HW rings is a real issue for igb and ixgbe drivers and needs
to be fixed. Therefore I would like this patch to be accepted in order to not
maintain the fix ourselves. But unfortunately I don't have resources (e.g. time)
to fix the issue for all the drivers, because, as I mentioned, they are not
following the same pattern to release their queues. So my proposal is that I fix
it in this patch in a number of drivers (including igb, ixgbe and i40e) and
others can take over and improve other drivers, if they see the same issue. This
is also a reason why the drivers' changes are not in one commit for all the
drivers.
For the proposal adding pmd name as prefix to queue memzone name or update
the 'rte_eth_dma_zone_reserve()' to check the size & alignment instead of just a
name, I don't know, as an external person, how sensitive dpdk project to change
an internal API and existing code (the call should be changed in all the
drivers). But anyway, I think the real problem is more an absence of memzone
pointer, and in long term it should be solved in this way, rather than search by
name.
Hi Renata,
'rte_eth_dma_zone_reserve()' returning existing memzone just based on 'name'
without checking the size, alignment, socket_id seems wrong to me, let me check
it, it shouldn't be hard to update.
But even 'rte_eth_dma_zone_reserve()' checks won't be enough (what to do if
check fails and not returns a memzone?), the proper solution is PMD free the
memzone as they removed, as done in this patch.
Also I suggested prefixing the memzone name as workaround but instead we can go
with this patchset with the implemented PMDs and other PMD owners can implement
the free when they need instead of workaround.
There is a comment from Wie to add more frees to i40e, and there is another from
Jeff on 'rte_eth_dma_zone_free()' return type, can you please check them? So we
can proceed with this patch for the release.
Thanks,
ferruh
Kind regards,
Renata
--------------------------------------------------------------------------------
*From:* Ferruh Yigit <ferruh.yi...@intel.com>
*Sent:* Wednesday, May 13, 2020 5:22 PM
*To:* Renata Saiakhova <renata.saiakh...@ekinops.com>; dev@dpdk.org
<dev@dpdk.org>
*Cc:* Anatoly Burakov <anatoly.bura...@intel.com>; Thomas Monjalon
<tho...@monjalon.net>; Neil Horman <nhor...@tuxdriver.com>
*Subject:* Re: [dpdk-dev] [PATCH v3 0/4] Memory corruption due to HW rings
allocation
On 5/13/2020 2:14 PM, Renata Saiakhova wrote:
igb and ixgbe and some other drivers allocate HW rings using
rte_eth_dma_zone_reserve(),
which checks first if the memzone exists for a given name, consisting of port
id, queue_id, rx/tx direction, but not for the size, alignment, and socket_id.
If the memzone with a given name exists it is returned, otherwise it is
allocated.
Disconnecting dpdk port from one type of interface (igb) and connecting it
to another type of interface (ixgbe) for the same port id, potentially creates
memory overlap and corruption, because it may require memzone of bigger size.
That's what is happening from switching from igb to ixgbe having the same port
id.
v2->v3: Remove #undef ETH_DMA_MZONE_NAME and minor changes in code standard
v1->v2: Minor changes on code standard and additional fixes in i40e em and ice
drivers
Renata Saiakhova (4):
librte_ethdev: Introduce a function to release HW rings
drivers/net: Fix in igb and ixgbe HW rings memory
drivers/net: Fix in i40e HW rings memory overlap
drivers/net: Fix in em and ice HW rings memory overlap
I think all driver patches can be squashed into single patch, overall they are
implementing same logic.
But as mentioned before, there are multiple other drivers allocating HW rings
with exact same name. At least I can see:
iavf
nfp
fm10k
axgbe
Or how can we know if a new PMD won't cause exact same behavior? What to you
think adding pmd name as prefix to queue memzone name for all PMDs? This can
help new PMDs using existing code as sample.
I don't know if it has been discussed before, but wouldn't update the
'rte_eth_dma_zone_reserve()' to check the size & alignment instead of just name
fix the issue for all drivers without needing to update them?
