On Thu, 4 Jun 2020 20:22:06 +0000
"Wang, Yipeng1" <yipeng1.w...@intel.com> wrote:
> > -----Original Message-----
> > From: Honnappa Nagarahalli <honnappa.nagaraha...@arm.com>
> > Sent: Thursday, June 4, 2020 12:34 PM
> > To: Wang, Yipeng1 <yipeng1.w...@intel.com>; Stephen Hemminger
> > <step...@networkplumber.org>
> > Cc: Gobriel, Sameh <sameh.gobr...@intel.com>; Richardson, Bruce
> > <bruce.richard...@intel.com>; dev@dpdk.org; De Lara Guarch, Pablo
> > <pablo.de.lara.gua...@intel.com>; nd <n...@arm.com>; Honnappa
> > Nagarahalli <honnappa.nagaraha...@arm.com>; nd <n...@arm.com>
> > Subject: RE: [PATCH] hash: document breakage with multi-writer thread
> >
> > <snip>
> >
> > > > > >
> > > > > > > Subject: [PATCH] hash: document breakage with multi-writer
> > > > > > > thread
> > > > > > >
> > > > > > > The code in rte_cuckoo_hash multi-writer support is broken if
> > > > > > > write operations are called from a non-EAL thread.
> > > > > > >
> > > > > > > rte_lcore_id() wil return LCORE_ID_ANY (UINT32_MAX) for non
> > > > > > > EAL thread and that leads to using wrong local cache.
> > > > > > >
> > > > > > > Add error checks and document the restriction.
> > > > > > Having multiple non-EAL writer threads is a valid use case.
> > > > > > Should we fix the
> > > > > issue instead?
> > > > >
> > > > > Discovered this the hard way...
> > > > >
> > > > > Fixing is non-trivial. Basically, the local cache has to be take
> > > > > out and that leads to having to do real locking or atomic operations.
> > > > >
> > > > Looking at rte_hash_create function:
> > > >
> > > > if (params->extra_flag &
> > > > RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
> > > > use_local_cache = 1;
> > > > writer_takes_lock = 1;
> > > > }
> > > >
> > > > The writer locks are in place already. The code to handle the case
> > > > when local cache is taken out is also there.
> > > > What we need is another input flag that says 'multi writer + non-eal
> > > threads'
> > > > which would set 'use_local_cache = 0' and 'writer_takes_lock = 1'.
> > > > Not sure, it would be valuable addition. But looks like this is what
> > > > you were expecting when you had enabled
> > > > 'RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD'. Many other APIs in
> > DPDK
> > > do
> > > > not provide this kind of MT safety.
> > >
> > > [Wang, Yipeng]
> > > If possible, we can try to not add new flags, because there are
> > > already a lot of flag options.
> > > How about in the code, we check if the writer is a non-eal or not by
> > > checking the rte_lcore_id, and operate on the global queue?
> > > Could this work?
> > > If(h->use_local_cache) {
> > > lcore_id = rte_lcore_id();
> > > if(lcore_id == LCORE_ID_ANY) { // this is non-eal threads
> > > <call rte_ring_mp/mc_* to directly operate on global queue>
> > > }
> > > Else {
> > > <original path>
> > > }
> > > }
> > The other thing I wanted to do was saving on the memory allocated for the
> > local cache when the writers are non-eal threads. Without knowing the kind
> > of threads upfront, we might have to create the local cache when a writer
> > adds the entry first time.
>
> I got what you mean. If people only use non-eal threads, we could save the
> space of local cache completely.
> Creating local cache during the first write is one solution. But the current
> rte_hash always allocate things during
> table creation time. This provides guarantee that the program won't fail in
> the middle due to memory allocation issue.
>
> Meanwhile I would rather be wasting some space than adding another option
> flag related to multi-threading.
> In my opinion, all those flags are already confusing enough. It would also be
> harder to maintain in future.
Don't care about exact fix. Just don't want to randomly corrupt memory.
Having it work would be better. But can we just error out for now; existing
code is broken
