Also, make it possible to remove copy operations following a fill
operation if they're using the same buffers.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I7da195ce80650a02c5db99d9400ee692f797b1f8
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15940
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Some of the copy operations can be elided, so they're not the best for
this kind of test. So, use another operation, decompress, that can be
appended to an accel sequence.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: Ic59e7678436bdf1d5ab6eb103de4cc0c0c347b9f
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/16018
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Also, replace src2 with an iovec + iovcnt and rename it to s2 to
keep the naming consistent with the source buffer (s).
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I44787128377addd514818ec5aaec084b1a31f0c3
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15939
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Also, replace dst2 with an iovec + iovcnt and rename it to d2 to
keep the naming consistent with the destination buffer (d).
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: Ib394c127eeb5890451535ff485f96f7edd2897a4
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15938
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
This patch is first in the series of patches aimed to make all accel
operations describe their buffers with iovecs. The intention is to make
it easier to handle tasks in a generic way.
It doesn't mean that we change the API - all function signatures are
preserved. If a function doesn't use iovecs, we use the aux_iovs array.
However, this does mean that each accel module that provides support for
a given operation will need to be adjusted to use iovecs.
Additionally, update the unit test checking copy elision to verify the
buffers of the copy operation that is left.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I9e6d8d1be3b8b9706cb4a6222dad30e8c373d8fb
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15937
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Users can now specify buffers allocated through `spdk_accel_get_buf()`
when appending operations to a sequence. When an operation in a
sequence is executed, we check it if it uses buffers from accel domain,
allocate data buffers and update all operations within a sequence that
were also using those buffers.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I430206158f6a4289e15f04ddb18f0d1a2137f0b4
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15748
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
It will be used for allocating buffers from accel domain and
allocating bounce buffers to push/pull the data from memory domains for
modules that don't support memory domains.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: Idbe4d2129d0aff87d9e517214e9f81e8470c5088
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15745
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
This domain is meant to represent data being transformed by accel
engine. Users will be able to allocate buffers from that memory domain
and use them when appending operations to an accel sequence.
Since these buffers are only meant to be used as placeholders for actual
buffers, none of the push/pull/translate callbacks are implemented. To
access the data after it was transformed by accel, users should make
sure that the final command's destination buffer isn't allocated from
accel memory domain.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: Ia031c7b205e98792d0a93f01513101b86afa9faa
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15744
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reversing a sequence means that the order of its operations is reversed,
i.e. the first operation becomes last and vice versa. It's especially
useful in read paths, as it makes it possible to build the sequence
during submission, then, once the data is read from storage, reverse the
sequence and execute it.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I93d617c1e6d251f8c59b94c50dc4300e51908096
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15636
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Operation sequence should always be treated as a whole, meaning that
users cannot rely on the contents of any intermediate buffers and should
only care about the buffer that's the destination of the whole
operation. This allows us to remove some of those copy operations by
changing source / destination buffer of a preceding / following
operation.
If a sequence is using buffers from non-local memory domain, users can
append a copy operation to a sequence to specify a local destination
buffer. If the module executing the operations is aware of memory
domains, this can avoid doing an extra spdk_memory_domain_pull_data().
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I93b94d46ee32700819e9e6f1c55350692db8a67a
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15530
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
This patch introduces the concept of chaining multiple accel operations
and executing them all at once in a single step. This means that it
will be possible to schedule accel operations at different layers of the
stack (e.g. copy in NVMe-oF transport, crypto in bdev_crypto), but
execute them all in a single place. Thanks to this, we can take
advantage of hardware accelerators that supports executing multiple
operations as a single operation (e.g. copy + crypto).
This operation group is called spdk_accel_sequence and operations can be
appended to that object via one of the spdk_accel_append_* functions.
New operations are always added at the end of a sequence. Users can
specify a callback to be notified when a particular operation in a
sequence is completed, but they don't receive the status of whether it
was successful or not. This is by design, as they shouldn't care about
the status of an individual operation and should rely on other means to
receive the status of the whole sequence. It's also important to note
that any intermediate steps within a sequence may not produce observable
results. For instance, appending a copy from A to B and then a copy
from B to C, it's indeterminate whether A's data will be in B after a
sequence is executed. It is only guaranteed that A's data will be in C.
A sequence can also be reversed using spdk_accel_sequence_reverse(),
meaning that the first operation becomes last and vice versa. It's
especially useful in read paths, as it makes it possible to build the
sequence during submission, then, once the data is read from storage,
reverse the sequence and execute it.
Finally, there are two ways to terminate a sequence: aborting or
executing. It can be aborted via spdk_accel_sequence_abort() which will
execute individual operations' callbacks and free any allocated
resources. To execute it, one must use spdk_accel_sequence_finish().
For now, each operation is executed one by one and is submitted to the
appropriate accel module. Executing multiple operations as a single one
will be added in the future.
Also, currently, only fill and copy operations can be appended to a
sequence. Support for more operations will be added in subsequent
patches.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: Id35d093e14feb59b996f780ef77e000e10bfcd20
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15529
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
per Intel policy to include file commit date using git cmd
below. The policy does not apply to non-Intel (C) notices.
git log --follow -C90% --format=%ad --date default <file> | tail -1
and then pull just the 4 digit year from the result.
Intel copyrights were not added to files where Intel either had
no contribution ot the contribution lacked substance (ie license
header updates, formatting changes, etc). Contribution date used
"--follow -C95%" to get the most accurate date.
Note that several files in this patch didn't end the license/(c)
block with a blank comment line so these were added as the vast
majority of files do have this last blank line. Simply there for
consistency.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: Id5b7ce4f658fe87132f14139ead58d6e285c04d4
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15192
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Community-CI: Mellanox Build Bot
In prep for adding both src and dst iovec support for compression.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I704b8d2bd459de03deb7f8ee45d76261910a3727
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13746
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
This is consistent with the use of terms in other parts of SPDK and fits
with the code living under module/
Change-Id: If182f7cf2d160d57443a1b5f24e0065f191b59b2
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13919
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
The word engine was both used (interchangeably with module) to refer to
the things that plug into the framework and to the framework itself.
This patch eliminates all use of the word engine that meant the
framework. It leaves uses of the word that meant "module".
Change-Id: I6b9b50e2f045ac39f2a74d0152ee8d6269be4bd1
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13918
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
