On 12-Dec-18 12:55 PM, Yongseok Koh wrote:
On Mon, Dec 03, 2018 at 10:23:03AM +0000, Burakov, Anatoly wrote:
On 02-Dec-18 11:28 PM, Yongseok Koh wrote:
On Dec 1, 2018, at 9:48 PM, Shahaf Shuler <shah...@mellanox.com> wrote:
Hi Anatoly,
Thursday, November 29, 2018 3:49 PM, Anatoly Burakov:
Subject: [PATCH 0/4] Allow using external memory without malloc
Currently, the only way to use externally allocated memory is through
rte_malloc API's. While this is fine for a lot of use cases, it may not be
suitable
for certain other use cases like manual memory management, etc.
This patchset adds another API to register memory segments with DPDK (so
that API's like ``rte_mem_virt2memseg`` could be relied on by PMD's and
such), but not create a malloc heap out of them.
Aside from the obvious (not adding memory to a heap), the other major
difference between this API and the ``rte_malloc_heap_*`` external memory
functions is the fact that no DMA mapping is performed automatically.
This really draws a line in the sand, and there are now two ways of doing
things - do everything automatically (using the ``rte_malloc_heap_*`` API's),
or do everything manually (``rte_extmem_*`` and future DMA mapping API
[1] that would replace ``rte_vfio_dma_map``). This way, the consistency of
API is kept, and flexibility is also allowed.
As you know I like the idea.
One question though, do you see a use case for application to have externally
allocated memory which needs to be registered to the DPDK subsystem however not
being used for DMA?
My only guess would be so helper libraries which requires the memory allocation
from user (however it doesn't seems like a good API).
If no use case, maybe it is better to merge between the two (rte_extmem_* and
rte_dma_map) to have a single call for app to register and DMA map the memory.
The rte_mem_virt2memseg is not something application needs to understand, it is
used internally by PMDs or other libs.
Just FYI.
My implementation for mlx4/5 doesn't need to have a separate registration for
DMA by rte_dma_map() as long as it is included in the memseg list. Registration
is done only if Lkey lookup misses and only mem free event is taken to release
it. From my end, the reason why we wanted to have a generic DMA registration was
because some people doesn't want to use the new API to make the ext mem included
in the memseg list but want to simply call the API for DMA registration.
In a nutshell, mlx4/5 needs users use either rte_extmem_register() or
rte_dma_map(). However, it is no problem to call both.
It would be good to create a segment when using rte_dma_map().
Unfortunately, that's not realistic :)
Registering memory within DPDK does not necessarily have to be performed by
the primary process - whichever process that wants to create the table, can
do so, and later processes have to attach to the memory area. There's also
no way to know if memory segment can be attached to - this is a question
only application can answer.
In other words, there's no way to combine rte_dma_map() and
rte_extmem_register() into one call and keep multiprocess support.
Sorry for late reply. I was away for a while.
I understood your point that rte_dma_map() can't create a segment but isn't the
opposite possible? I still have a question about
rte_extmem_register/unregister/attach/detach(). Why don't these APIs generate
memory events? Do you define the memory events are limited to memories for
malloc? What if some app wants to know the events even if it is extmem? What
makes difference between two types of extmem (one for malloc heap and the other
for just memseg) in generating the events?
I've reviewed your patches and all look good :-) But it is still unclear to me.
Hi Yongseok,
Idealistically speaking, my view is, if you want the luxury of DPDK
doing everything for you, use malloc heaps. If you don't - you're on
your own :) Any callbacks etc. that you might want to get if you're
*not* using malloc, is not really my problem - it's yours, because *you*
don't want to use built-in DPDK facilities, for whatever reason.
Pragmatically speaking, right now the reason to not do so is to avoid
that memory being automatically mapped for DMA due to VFIO currently
using callbacks mechanism to subscribe to notifications. When they are
decoupled - we can talk about making it so that registered external
memory still triggers callbacks (although i do not see why, to be honest).
Thanks,
Yongseok
[1]
https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fma
ils.dpdk.org%2Farchives%2Fdev%2F2018-
November%2F118175.html&data=02%7C01%7Cshahafs%40mellanox.co
m%7C007a9234feaf42c82f6508d656015eb1%7Ca652971c7d2e4d9ba6a4d1492
56f461b%7C0%7C0%7C636790961244424277&sdata=YqwcPEEhJM3I7Toe
Ne%2BGcbeo%2FmPbYEnNFckoA7ES2Hg%3D&reserved=0
Note: at the time of this writing, there's no release notes
template, so no release notes updates in this patchset.
They will be added in later revisions.
Anatoly Burakov (4):
malloc: separate creating memseg list and malloc heap
malloc: separate destroying memseg list and heap data
mem: allow registering external memory areas
mem: allow usage of non-heap external memory in multiprocess
.../prog_guide/env_abstraction_layer.rst | 63 +++++++--
lib/librte_eal/common/eal_common_memory.c | 116
+++++++++++++++++
lib/librte_eal/common/include/rte_memory.h | 122
++++++++++++++++++
lib/librte_eal/common/malloc_heap.c | 104 +++++++++++----
lib/librte_eal/common/malloc_heap.h | 15 ++-
lib/librte_eal/common/rte_malloc.c | 115 +++++++----------
lib/librte_eal/rte_eal_version.map | 4 +
7 files changed, 434 insertions(+), 105 deletions(-)
--
2.17.1
--
Thanks,
Anatoly
--
Thanks,
Anatoly
