[dpdk-dev] [PATCH] kni: dynamically allocate memory for each KNI

[Ferruh Yigit]

On 8/29/2018 10:52 AM, Igor Ryzhov wrote:
> Hello Ferruh,
> 
> Thanks for the review, comments inline.
> 
> On Mon, Aug 27, 2018 at 8:06 PM, Ferruh Yigit <ferruh.yigit at intel.com
> <mailto:ferruh.yigit at intel.com>> wrote:
> 
>     On 8/2/2018 3:25 PM, Igor Ryzhov wrote:
>     > Long time ago preallocation of memory for KNI was introduced in commit
>     > 0c6bc8e. It was done because of lack of ability to free previously
>     > allocated memzones, which led to memzone exhaustion. Currently memzones
>     > can be freed and this patch uses this ability for dynamic KNI memory
>     > allocation.
> 
>     Hi Igor,
> 
>     It is good to be able to allocate memory dynamically and get rid of the
>     "max_kni_ifaces" and "kni_memzone_pool", thanks for the patch.
> 
>     Overall looks good, a few comments below.
> 
>     > 
>     > Signed-off-by: Igor Ryzhov <iryzhov at nfware.com <mailto:iryzhov at nfware.com>>
>     > ---
>     >  lib/librte_kni/rte_kni.c | 392 ++++++++++++---------------------------
>     >  lib/librte_kni/rte_kni.h |   6 +-
>     >  test/test/test_kni.c     |   6 -
>     >  3 files changed, 128 insertions(+), 276 deletions(-)
>     > 
>     > diff --git a/lib/librte_kni/rte_kni.c b/lib/librte_kni/rte_kni.c
>     > index 8a8f6c1cc..028b44bfd 100644
>     > --- a/lib/librte_kni/rte_kni.c
>     > +++ b/lib/librte_kni/rte_kni.c
>     > @@ -36,24 +36,33 @@
>     >   * KNI context
>     >   */
>     >  struct rte_kni {
>     > +     const struct rte_memzone *mz;       /**< KNI context memzone */
> 
>     I was thinking remove the context memzone and use rte_zmalloc() to create kni
>     objects but updated rte_kni_get() API seems relaying this.
>     If you see any other way to get kni object from name in rte_kni_get(), I am for
>     removing above *mz variable from rte_kni struct.
> 
> 
> I had absolutely the same thought but didn't find a way to save rte_kni_get() API.
> Maybe someone has any ideas?
> Or maybe this API can be marked deprecated and deleted in future?

I suggest keep API, there may be some users, we don't know. And it doesn't sound
right to remove an API because of internal implementation details, so looks like
will need to keep memzone.

>  
> 
> 
>     <...>
> 
>     > +static void
>     > +kni_ctx_release_mz(struct rte_kni *ctx)
>     > +{
>     > +     rte_memzone_free(ctx->m_tx_q);
>     > +     rte_memzone_free(ctx->m_rx_q);
>     > +     rte_memzone_free(ctx->m_alloc_q);
>     > +     rte_memzone_free(ctx->m_free_q);
>     > +     rte_memzone_free(ctx->m_req_q);
>     > +     rte_memzone_free(ctx->m_resp_q);
>     > +     rte_memzone_free(ctx->m_sync_addr);
> 
> 
>     "ctx" sounds confusing to me, isn't this "rte_kni" object instance, why not just
>     call it "kni" or if it is too generic "kni_obj" or similar? For other APIs
>     as well.
> 
> 
> "ctx" was already used in the code, so I didn't change it.
> I also think that it's better to use "kni" – will change it in v2.
>  
> 
> 
>     And this is just a detail but about order of APIs would you mind having first
>     reserve() one, later release() one?
> 
> 
> Ok.
>  
> 
> 
>     <...>
> 
>     > -/* Shall be called before any allocation happens */
>     > -void
>     > -rte_kni_init(unsigned int max_kni_ifaces)
>     > +static struct rte_kni *
>     > +kni_ctx_reserve(const char *name)
>     >  {
>     > -     uint32_t i;
>     > -     struct rte_kni_memzone_slot *it;
>     > +     struct rte_kni *ctx;
>     >       const struct rte_memzone *mz;
>     > -#define OBJNAMSIZ 32
>     > -     char obj_name[OBJNAMSIZ];
>     >       char mz_name[RTE_MEMZONE_NAMESIZE];
>     >  
>     > -     /* Immediately return if KNI is already initialized */
>     > -     if (kni_memzone_pool.initialized) {
>     > -             RTE_LOG(WARNING, KNI, "Double call to rte_kni_init()");
>     > -             return;
>     > -     }
>     > +     snprintf(mz_name, RTE_MEMZONE_NAMESIZE, "kni_info_%s", name);
> 
>     Can you please convert memzone names, like "kni_info" to defines, for all of
>     them?
> 
> 
> Ok.
>  
> 
> 
>     <...>
> 
>     > @@ -81,8 +81,12 @@ struct rte_kni_conf {
>     >   *
>     >   * @param max_kni_ifaces
>     >   *  The maximum number of KNI interfaces that can coexist concurrently
>     > + *
>     > + * @return
>     > + *  - 0 indicates success.
>     > + *  - negative value indicates failure.
>     >   */
>     > -void rte_kni_init(unsigned int max_kni_ifaces);
>     > +int rte_kni_init(unsigned int max_kni_ifaces);
> 
>     This changes the API. Return type changes from "void" to "int". I agree "int"
>     makes more sense since API can fail, but this changes the ABI/API.
> 
>     Since existing binaries doesn't check the return type at all there may be no
>     issue from ABI point of view but from API point of view some apps may get return
>     value not checked warnings, not sure though.
> 
>     And the need of the API is questionable at this stage, it may be possible to
>     move rte_kni_alloc() where it already has "kni_fd" check.
> 
>     What do you think keep API signature same for now, but add a deprecation notice
>     to remove the API. Next release (v19.02) remove rte_kni_init() completely?
> 
> 
> As I know, warnings can only be returned if the warn_unused_result attribute is
> used, which is not the case here.
> So I think that changing from void to int should not break anything. Can change
> it back in v2 if I'm wrong.
> 
> Regarding the API removal – I think it's better to save that function, to have a
> more clear API.

OK, fair enough. Lets keep it with same name.

> As we have rte_kni_close to close KNI device, we should have a function to open it.
> Maybe it should be renamed to rte_kni_open :)
>  
> 
> 
>     <...>
> 
>     >  /**
>     > diff --git a/test/test/test_kni.c b/test/test/test_kni.c
>     > index 1b876719a..56c98513a 100644
>     > --- a/test/test/test_kni.c
>     > +++ b/test/test/test_kni.c
>     > @@ -429,12 +429,6 @@ test_kni_processing(uint16_t port_id, struct rte_mempool *mp)
>     >       }
>     >       test_kni_ctx = NULL;
>     >  
>     > -     /* test of releasing a released kni device */
>     > -     if (rte_kni_release(kni) == 0) {
>     > -             printf("should not release a released kni device\n");
>     > -             return -1;
>     > -     }
> 
>     Why need to remove this?
> 
> 
> Previously, rte_kni_release didn't free any memory, and the second call with the
> same argument just checked "in_use" flag and returned.
> After my changes, memory is actually freed, and rte_kni_release must not be
> called twice with the same argument.

OK.

> Will send v2 with approved changes in a couple of days.
> At the same time, I'll think what can we do with context memzone.

Thanks.