On 6/18/19 9:48 PM, Gage Eads wrote:
The mempool library assigns handler ops indexes based on the dynamic load
order of mempool handlers. Indexes are used so a mempool can be used by
multiple processes, but this only works if all processes agree on the
mapping from index to mempool handler.
When using the '-d' argument, it's possible for different processes to load
mempool handlers in different orders, and thus have different
index->handler mappings. Using a mempool in multiple of such processes will
result in undefined behavior.
This commit adds a note to the mempool library programmer's guide warning
users against this.
Fixes: 449c49b93a6b ("mempool: support handler operations")
Cc: sta...@dpdk.org
Signed-off-by: Gage Eads <gage.e...@intel.com>
---
doc/guides/prog_guide/mempool_lib.rst | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/doc/guides/prog_guide/mempool_lib.rst
b/doc/guides/prog_guide/mempool_lib.rst
index 52a569f57..4470f6b38 100644
--- a/doc/guides/prog_guide/mempool_lib.rst
+++ b/doc/guides/prog_guide/mempool_lib.rst
@@ -133,6 +133,13 @@ For applications that use ``rte_pktmbuf_create()``, there
is a config setting
(``RTE_MBUF_DEFAULT_MEMPOOL_OPS``) that allows the application to make use of
an alternative mempool handler.
+ .. note::
+
+ When running a DPDK application with shared libraries, mempool handler
+ shared objects specified with the '-d' EAL command-line parameter are
+ dynamically loaded. When running a multi-process application with shared
+ libraries, the -d arguments for mempool handlers *must be specified in the
+ same order for all processes* to ensure correct operation.
One more empty line is required here, other than that:
Acked-by: Andrew Rybchenko <arybche...@solarflare.com>
I think it is OK as a fix for stable branches.
In theory I think it is still technically possible to guarantee ops
indices correctness
using dedicated memzone (as Olivier suggested), but with sync up on
rte_mempool_{get,set}_ops() (which should be called when EAL is
initialized) and
reordering. However, it requires API breakage to avoid returning of
ops_index
on register (since it may change in secondary process on resync when EAL is
already initialized).