Hi John,

The new API is intended to allow memory that is shared and not DMA-able, so spdk_dma_malloc() wouldn’t be an appropriate name; also, we did not want to break public API by adding a new flags parameter to the existing function.  We intend to update the internal SPDK callers of spdk_dma_malloc() to all use spdk_malloc() with the appropriate flags (leaving the old names for backwards compatibility for now).

If the user passes flags == 0 to the new spdk_malloc() call, this could be implemented by malloc() or equivalent behind the scenes, since the memory does not need to be DMA-able or shared across multi-process boundaries (it has no special properties).  Does this suffice for your use case?

Thanks,

-- Daniel

From: SPDK [mailto:spdk-bounces@lists.01.org] On Behalf Of Meneghini, John
Sent: Thursday, May 3, 2018 10:53 AM
To: Harris, James R <james.r.harris@intel.com>; Walker, Benjamin <benjamin.walker@intel.com>; z.khatami88@gmail.com; Storage Performance Development Kit <spdk@lists.01.org>
Cc: Meneghini, John <John.Meneghini@netapp.com>
Subject: [SPDK] spdk_malloc vs. malloc

Hi Jim, Ben, and Zahra.

I am sorry I didn’t catch this during the code review, but it’s really hard to keep track of all of the changes that are going on. 

https://review.gerrithub.io/#/c/spdk/spdk/+/402334/

As you may recall, we wanted to use spdk_malloc and spdk_calloc to abstract the POSIX malloc and calloc calls.  This is still an issue in multiple places as we want a non-blocking malloc and calloc call that returns a status rather than setting errno (e.g. to ENOMEM). Remember, not all execution environments support errno.

Can we revert change a77cd3f7, and use the spdk_malloc/calloc/free api for that purpose?

Also, I don’t see the point of introducing an additional API for dma-able memory.    We went through this once before.

* a77cd3f7 2018-04-24  env: add malloc variants with DMA/shareable flags [ Ben Walker / z.khatami88@gmail.com ]

* 8a44220b 2017-05-31  env: Rename spdk_malloc/zmalloc/realloc/free to spdk_dma_(func) [ Jim Harris / johnm@netapp.com ]

I fully support adding the SPDK_MALLOC_DMA and SPDK_MALLOC_SHARE flag and socket_id to spdk_dma_malloc/zmalloc(), and I prefer to do this rather than invent a new API for allocating dma-able memory.

void *spdk_dma_malloc(size_t size, size_t align, uint64_t *phys_addr);

void           *spdk_malloc(size_t size, size_t align, uint64_t *phys_addr, int socket_id, uint32_t flags);

I’m proposing:

void *spdk_dma_malloc(size_t size, size_t align, uint64_t *phys_addr, int socket_id, uint32_t flags);

/John

P.S. What I really need is a filter that allows me to watch all of the headers in include/spdk/{api}.h.  Looking at the GerritHub documentation, I see no way to do this.  It appears all I can do is watch the master branch.

https://review.gerrithub.io/Documentation/intro-user.html#watch