On 15-Nov-19 8:46 AM, Olivier Matz wrote:
On Thu, Nov 14, 2019 at 01:58:20PM +0000, Anatoly Burakov wrote:
Currently, when mempool is being populated, we get IOVA address
of every segment using rte_mem_virt2iova(). This works for internal
memory, but does not really work for external memory, and does not
work on platforms which return RTE_BAD_IOVA as a result of this
call (such as FreeBSD). Moreover, even when it works, the function
in question will do unnecessary pagewalks in IOVA as PA mode, as
it falls back to rte_mem_virt2phy() instead of just doing a lookup in
internal memseg table.
To fix it, replace the call to first attempt to look through the
internal memseg table (this takes care of internal and external memory),
and fall back to rte_mem_virt2iova() when unable to perform VA->IOVA
translation via memseg table.
Fixes: 66cc45e293ed ("mem: replace memseg with memseg lists")
Shouldn't we also add
Fixes: 035ee5bea5ef ("mempool: remove optimistic IOVA-contiguous allocation")
?
From what I understand, even if the problem existed in populate_virt()
before, this is the commit that makes the problem visible in most cases.
The commit didn't introduce the issue - only revealed it. So the commit
itself isn't buggy, and there's nothing to fix :) However, i'm not
opposed to sharing the blame!
Cc: sta...@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.bura...@intel.com>
Acked-by: Olivier Matz <olivier.m...@6wind.com>
--
Thanks,
Anatoly
