Thanks Anatoly for prompt response. Sorry for the delayed response, I
took some time to reverify with SPDK.
Infact, I do want the iommu mapping to be performed. I don't want it to be
bypassed by type1_map() [lib/librte_eal/linuxapp/eal/eal_vfio.c] for
external memory.
Then, if I understood correctly, you suggested to call rte_vfio_dma_map()
as an alternative.
*Context & clarification*
1) We 're using DPDK to manage/allocate memory for SPDK through heap API.
The only DPDK APIs we 're calling are:
A)rte_malloc_heap_memory_add() to add external memory to heap.
B)rte_malloc_heap_get_socket() & rte_malloc_socket() to allocate memory
* Are you suggesting to make a call to rte_vfio_dma_map() from spdk, in
addition to the APIs listed under 1)A & 1)B instead of modifying DPDK vfio
functions?*
Please confirm, Probably I missed the context and might not have
understood fully.
2) .dma_user_map_func=vfio_type1_dma_mem_map() is called from 2 paths in
dpdk. In either case call to dma_user_map_func() is skipped.
A) *During the startup, as you said before:*
rte_vfio_setup_device() --> type1_map()
B)During allocation event:
vfio_mem_event_callback() (type=RTE_MEM_EVENT_ALLOC,..)
-->vfio_dma_mem_map() -->dma_user_map_func()
----------------------------------------------------------------------------------------------------------------------------------------------------------------------
*Conclusion*
So we have 2 alternatives:
A) make additional call to rte_vfio_dma_map() API after adding external
memory using rte_malloc_heap_memory_add() API.
B) remove msl->external check which bypasses call to dma_user_map_func()
in DPDK.
I modified DPDK functions [Option B) ]. I guess you 're suggesting option A)
Please confirm.
Regards,
Rajesh
On Fri, Oct 18, 2019 at 10:23 PM Burakov, Anatoly <anatoly.bura...@intel.com>
wrote:
> On 18-Oct-19 11:54 AM, Rajesh Ravi wrote:
> > + Srinath
> >
> > Thanks Anatoly for reviewing this. Please find my reply inline below:
> >
> > [Anatoly] First of all, what is "iso-cmem"? It doesn't seem to have any
> > defined
> > meaning nor any relation to any existing functionality, and it's not
> > explained anywhere what is "isolated cmem".
> > [Rajesh] I 'll correct commit message to include clearer explanation.
> > _
> > _
> > _Context & purpose_
> > We 're using this to improve SPDK performance. When NVMe completion
> > queues are allocated from a certain memory range,
> > out of order competions completions are enabled with our PCIe and it
> > improves performance.
> >
> > _Usage_
> > spdk_env_init()-->(calls rte_eal_init(), and then calls
> > )spdk_env_dpdk_post_init() [file: spdk/lib/env_dpdk/init.c]
> > --> spdk_mem_map_init()
> > [lib/env_dpdk/memory.c]-->map_cmem_virtual_area() [this is our custom
> > function]
> > In map_cmem_virtual_area(): we 're mmaping anonymous & private region
> > and then creating iova table which makes iova addresses which fall in
> > custom memory region.
> > Then we call rte_malloc_heap_memory_add() and then allocate NVMe
> > completion queues from this heap.
> >
> > [Anatoly] More importantly, why is this necessary? Type1 map only
> bypasses
> > external segments when adding memory at startup - it doesn't stop you
> > from calling rte_vfio_dma_map() to map the memory with VFIO when you
> > create the segment.
> > [Rajesh] Please correct me if I'm wrong or missing some thing here.
> > A) We wanted to create a heap from which we can allocate dynamically
> >
> > B) I see that type1_map()
> > [file:dpdk/lib/librte_eal/linuxapp/eal/eal_vfio.c] getting called once
> > rte_malloc_heap_memory_add()() was invoked.
> > SPDK uses type1 dma map [spdk/lib/env_dpdk/memory.c]
> > vfio_type1_dma_map() is registered for .dma_map_func member
> > [dpdk/lib/librte_eal/linuxapp/eal/eal_vfio.c]
> > So type1_map is called. Not sure whether we can change this.
> >
>
> I'm not sure i'm following. You mean you *don't* want this mapping to be
> performed?
>
> --
> Thanks,
> Anatoly
>
--
Regards,
Rajesh
