Found via inspection during spec review of latest HW. We were using the
wrong stride for the WQCFG regsiter when configuring but it just so
happened to be the right value for the current DSA version. We were
mixing up the size of the WQCFG register with the stride value used to
configure the next WQCFG regsiter as they are not contiguous in HW, we
need to read another capabilities bit to determine the address of the
next wqcfg to configure..
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I14d1ff95e0131fd30121aa955bfbc8c8fb3fc512
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10968
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Convert the batch to the single command inside of it.
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Change-Id: Ia117175ef3f4a8290d313e0bdc794f6a3276e042
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11166
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Increase the batch size and with it the effective queue depth per
channel to 512.
Change-Id: Ide665e92d47ee753c141f34dd6a8bc4d040fe8db
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11031
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: John Kariuki <John.K.Kariuki@intel.com>
Transparently group independent operations into idxd batch operations
between polls. This increases the effective queue depth.
Change-Id: Ic09b21ed29aaefe2eccef9a6ae0e1b05990ef631
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10998
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
dynamically
This can be done once on allocation rather than every time the batch is
submitted.
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Change-Id: I8092f65f1b864cc3cc78db9fdee085d8bb0471df
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10293
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Prior a regular round robin could result in strange performance
if an idxd device from another socket was used.
Signed-off-by: peluse <peluse@localhost.localdomain>
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
Change-Id: Id863c79067beabe73ef89d92b3fb3c436821b97a
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9367
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Monica Kenguva <monica.kenguva@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Used to be in the lib directory but an upcoming patch needs
access to it so move it to a more appropriate location.
The changes in the .h file were needed to address compile
error when in the /include dir (didn't get errors elsewhere)
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I44d11fc620f213b13683d62dca899e2792ca6ed5
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9450
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Ziye Yang <ziye.yang@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Small but noticable on perf top. We were using a list to track
valid batches and checking it on every submission. Instead just
put a valid channel ptr in the batch struct and clear it when the
batch is freed. There was no other reason for the list.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: Ia26ceee578be4a82f4fd8abb2358ff18c56271a4
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9149
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Was set at a pretty high number during early development.
Instead of a #define, lets use the same math we use to
determine the size of the operation and descriptor pools as
that's the max number of batches that will be needed.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: Ide5311b4b6c931010413d0f82baed0e05d5fcde7
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9099
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
We can instead use a combination of the op code and the batch
element in the op structure to determine if the op that is
completing is part of a batch or not so we know whether to
return it to the free list or not.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I8e27c7b991f5a20e82394a25a52ece560cab5543
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9071
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Tested-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
The bit arrays were used for dynamic flow control in a previous
implementation. They are no longer needed as flow control is
now static and managed solely in the idxd plug in module. Use
simple lists of descriptors and completion records instead.
This is a simpler implementation and will allow for some future
clean up of structures as well.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I8c4cce12e88ac5416e3fe29a416487759214ec9f
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/8922
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
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: Ziye Yang <ziye.yang@intel.com>
This batch_op field is not necessary because we can
use the comp_ctx->desc->opcode to judge whether it is related
a batched task or not.
Change-Id: Id329221ccf272c4c3bb8c1b5ec08433029a9a1f8
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/8865
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Previously we used a counter of our own to make sure all batch
elements plus the batch itself were done before we freed the batch.
This was due to some observations early on that the batch desc
could complete before the individual elements and a lack of clarity
as to whether this was due to the simulator or the fact that
we poll on completions and could therefore "see" completions in
a different order at that time (we were using bit arrays to poll).
Now we use an ordered (in time) list to poll locations so if we
instead put the elements on the list first and then the batch desc
itself we are assured to always "see" them in order provided the
underlying device meets spec which there's no reason to assume it
does not.
This simplifies things a bit at the same time and still assures
that we call list calbacks in order and then the batch callback
without "special" handling.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I4d9e3997786f2116ce6515682b8117799c645f51
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/8397
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Allows for better performance by not hitting the same portal
address with every submission.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
Change-Id: I1ec8eae6f3acec9e98161029cd5406ec08603aa6
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/8190
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>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Just remove this function pointer and add a new one,i.e.,
dump_sw_error.
Because this function pointer is only used to
read a sw err info. We can hide it in the detailed
idxd implemenation.
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
Change-Id: I42fe2220dae85df307b5af64e37acfd7f748915b
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/8707
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Recent work identified race conditions having to do with the
dynamic flow control mechanism for the idxd engine. In order
to both address the issue and simplify the code a new scheme
is now in place. Essentially every DSA device will be allowed
to accomodate 8 channels and each channel will get a fixed 1/8
the number of work queue entries regardless of how many
channels there are. Assignment of channels to devices is round
robin and if/when no more channels can be accommodated the get
channel request will fail.
The performance tests also revealed another issue that was
masked before, it's a one-line so is in this patch for convenience.
In the idxd poller we limit the number of completions allowed
during one run to avoid the poller thread from starving other
threads since as operations complete on this thread they are
immediately replaced up to the limit for the channel.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I913e809a934b562feb495815a9b9c605d622285c
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/8171
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Was using reserved field in CRC to store the final address of where
to put the result, this not legal. Move to the completion record
and slightly re-arrange the struct to keep it at 96 bytes.
Refacorted the IO prep function so the caller can udpate both the
descriptor and completion records instead of continuing to add
parameters to the prep function for opcodes that need something
unique in the completion record. This also allowed for a minor
fix where the prep function was returning NULL when vtophys failed
which would have indicated busy as opposed to failire. Now we can
proprely fail that path.
fixes#1929
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: Ic23bc7b68bdd5757c30b7963880677f423368e20
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/7735
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>
Purpose: We will also support the kernel idxd driver, so we do not
need export this feature in the module file.
Change-Id: I965e031497920f527962ba187bccd81de6977b8f
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/7336
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Purpose: This patch is used to prepare to add the kernel
idxd support later.
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
Change-Id: If89665f95d622c7342ab75050664158ec6fc615a
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/7330
Community-CI: Broadcom CI
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Add print to confirm how groups/queues/engines are being
programmed based on the init RPC used.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: Ic9462c19c6899478a803433f90d9db9249dd5ca1
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/6325
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
For flow control reasons we have to resize the bit arrays we
use to manage flow as channels come and go. However since
channels are assigned to devices, until the channel count
reaches the device count there's no sharing so no resize of
the array is needed. So, when we use a device for the first
time there's no need to run through the rest of the channels
and re-balance.
Same thing is done on destruction. The code to free idxd
specific resources was moved from the rebalance function to
the idxd put channel function which is a much more logical
place for it as well.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: Ib4df163286906f413dd6429dc6833af7b68e208c
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/5846
Community-CI: Broadcom CI
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Suggestions from a prior review... able to remove a boolean by changing
how the batch elements 'index' and 'remaining' are used.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I273e5e231bb30d51eb3ae0a59eec110377d49ab7
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/4813
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>
Earlier refactoring enables us to not have to keep track of batch completions in
the batch struct as they're always used sequentially now so we can just add
the addresses from the start up to the number of elements in the batch.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I00cdcdec3376a1c32c9dab72c68fea868c1cb540
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/4810
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>
Community-CI: Mellanox Build Bot
And to eliminate an artificial constraint on # of user descriptors.
The main idea here was to move from a single ring that covered all
user descriptors to a pre-allocated ring per pre-allocated batch.
In addition, the other major change here is in how we poll for
completions. We used to poll the batch rings then the main ring.
Now when commands are prepared their completion address is added to
a per channel list and the poller simply runs through that list
not caring which ring the completion address belongs too. This
simplifies the completion logic considerably and will avoid
polling locations that can't potentially have a completion.
Some minor rework was included as well, mainly getting rid of the
ring_ctrl struct as it didn't serve much of a purpose anyway and
with how things are setup now its easier to read with all the
elements in the channel struct.
Also, a change that came in while this was WIP needed a few fixes
to function correctly. Addressed those and moved them to a
helper function so we have one point of control for xlations.
Added support for NOP in cases where a batch is submitted with
only 1 descriptor.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: Ie201b28118823100e908e0d1b08e7c10bb8fa9e7
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/3654
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This patch only includes the basic framework for batching and the
ability to batch one type of command, copy. Follow-on patches will
add the ability to batch other commands and include an example of
how to do so via the accel perf tool. SW engine support for batching
will also come in a future patch. Documentation will also be coming.
Batching allows the application to submit a list of independent
descriptors to DSA with one single "batch" descriptor. This is beneficial
when the application is in a position to have several operations ready
at once; batching saves the overhead of submitting each one separately.
The way batching works in SPDK is as follows:
1) The app gets a handle to a new batch with spdk_accel_batch_create()
2) The app uses that handle to prepare a command to be included in the
batch. For copy the command is spdk_accel_batch_prep_copy(). The
app many continue to prep commands for the batch up to the max via
calling spdk_accel_batch_get_max()
3) The app then submits the batch with spdk_accel_batch_submit()
4) The callback provided for each command in the batch will be called as
they complete, the callback provided to the batch submit itself will be
called then the entire batch is done.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I4102e9291fe59a245cedde6888f42a923b6dbafd
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/2248
Community-CI: Mellanox Build Bot
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>
Module, etc., will follow. Notes:
* IDXD is an Intel silicon feature available in future Intel CPUs.
Initial development is being done on a simulator. Once HW is
available and the code fully tested the experimental label will be
lifted. Spec can be found here: https://software.intel.com/en-us/download/intel-data-streaming-accelerator-preliminary-architecture-specification
* The current implementation will only work with VFIO.
* DSA has a number of engines that can be grouped based on application
need such as type of memory being served or QoS. Engines are processing
units and are assigned to groups. Work queues are on device structures
that act as front-end groups for queueing descriptors. Full details on
what is configurable & how will come in later doc patches.
* There is a finite number of work queue slots that are divided amongst
the number of desired work queues in some fashion (ie evenly).
* SW (outside of the idxd lib) is required to manage flow control, to not
over-run the work queues.This is provided in the accel plug-in module.
The upper layers use public API to manage this.
* Work queue submissions are done with a 64 byte atomic instruction
* The design here creates a set of descriptor rings per channel that match
the size of the work queues. Then, an spdk_bit_array is used to make sure
we don't overrun a queue. If there are not slots available, the operation
is put on a linked list to be retried later from the poller.
* As we need to support any number of channels (we can't limit ourselves
to the number of work queues) we need to dynamically size/resize our
per channel descriptor rings based on the number of current channels. This
is done from upper layers via public API into the lib.
* As channels are created, the total number of work queue slots is divided
across the channels evenly. Same thing when they are destroyed, remaining
channels with see the ring sizes increase. This is done from upper layers
via public API into the lib.
* The sim has 64 total work queue entries (WQE) that get dolled out to the
work queues (WQ) evenly.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I899bbeda3cef3db05bea4197b8757e89dddb579d
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/1809
Community-CI: Mellanox Build Bot
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Vitaliy Mysak <vitaliy.mysak@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
