| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
DMA batch status was set by hardware. Its value may be variable between
cpus twice accesses. Saving the value of status can fix it.
Type: fix
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Change-Id: Ibc9337239555744a571685b486c986991c3e9b18
|
|
Allocate and initialize dma batch structure when adding dma config.
The number of required dma batches is set by max_batches parameter.
Thus dma batches are not allocated dynamically in worker thread.
Application need to check the return value of vlib_dma_batch_new.
Type: improvement
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Change-Id: I5d05a67b59634cf2862a377d5ab77cb1040343ce
|
|
This patch introduces DMA infrastructure into vlib. This is well known
that large amount of memory movements will drain core resource. Nowadays
more and more hardware accelerators were designed out for freeing core
from this burden. Meanwhile some restrictions still remained when
utilizing hardware accelerators, e.g. cross numa throughput will have a
significant drop compared to same node. Normally the number of hardware
accelerator instances will less than cores number, not to mention that
applications number will even beyond the number of cores. Some hardware
may support share virtual address with cores, while others are not.
Here we introduce new DMA infrastructure which can fulfill the
requirements of vpp applications like session and memif and in the
meantime dealing with hardware limitations.
Here is some design backgrounds:
Backend is the abstract of resource which allocated from DMA device
and can do some basic operations like configuration, DMA copy and
result query.
Config is the abstract of application DMA requirement. Application
need to request an unique config index from DMA infrastructure. This
unique config index is associated with backend resource. Two options
cpu fallback and barrier before last can be specified in config.
DMA transfer will be performed by CPU when backend is busy if cpu
fallback option is enabled. DMA transfer callback will be in order
if barrier before last option is enabled.
We constructs all the stuffs that DMA transfer request needed into
DMA batch. It contains the pattern of DMA descriptors and function
pointers for submission and callback. One DMA transfer request need
multiple times batch update and one time batch submission.
DMA backends will assigned to config's workers threads equally. Lock
will be used for thread-safety if same backends assigned to multiple
threads. Backend node will check all the pending requests in worker
thread and do callback with the pointer of DMA batch if transfer
completed. Application can utilize cookie in DMA batch for selves
usage.
DMA architecture:
+----------+ +----------+ +----------+ +----------+
| Config1 | | Config2 | | Config1 | | Config2 |
+----------+ +----------+ +----------+ +----------+
|| || || ||
+-------------------------+ +-------------------------+
| DMA polling thread A | | DMA polling thread B |
+-------------------------+ +-------------------------+
|| ||
+----------+ +----------+
| Backend1 | | Backend2 |
+----------+ +----------+
Type: feature
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Change-Id: I1725e0c26687985aac29618c9abe4f5e0de08ebf
|
Marvin Liu