If only one descriptor is needed, there is a special case in the spec for
SGL1 using the Data Block descriptor type. Add a comment to make it
clear what is going on.
Also tweak the SGL1 setup to copy from the first SGL descriptor list
element instead of relying on the last value from the loop above, since
that could be easily broken by accident.
Change-Id: I49ef97fe5bf18d2bf1d86b4310a7d3abdfd03e57
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
tr->u.sgl is already a struct spdk_nvme_sgl_descriptor pointer.
Change-Id: Ie2c8c052fc28e6369d1d095b8d566acae47975d1
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
_nvme_qpair_build_hw_sgl_request() will only be called for payload_size
!= 0, so every SGL will have at least one segment. Drop the 'else' that
was handling nseg == 0, and add an assert to document the payload_size
requirement.
Change-Id: I48e2a862a7657ba85605c0d35c0b65dfac072167
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
The assert is checking a variable that is invariant within the loop, so
move the assert up to the top of the function.
Change-Id: Iee7eea1736bc7f953665feb390c3d6340dbeffbc
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This makes it easier to find the larger doc comments that produce separate
pages.
It also allows removing the lib/nvme directory from the Doxyfile, so
only the public API headers are used to generate documentation.
Change-Id: I8c46edb8067a91dda5b23fb0864efd3dd8aaeba5
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
spdk_nvme_qpair_process_completions() is already documented in
spdk/nvme.h, so merge the doc comment from nvme_qpair.c into the public
header.
Change-Id: Id7722d99d209852ee64286e0a3fa127b863e10aa
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
If the controller is failed, attempting to submit additional I/O is
futile - it will be immediately failed using the completion callback,
which can result in infinite recursion if the application code resubmits
I/Os on failure.
Instead, provide a way for request submission to indicate failure, and
use it to exit early if the controller is failed; this can only happen
when a reset failed (timed out).
If a request is submitted directly by the user when the controller has
failed, we can return an error code directly. For the case where I/O
was queued and is being resubmitted after a reset, we still need to call
the completion handler via _nvme_fail_request_ctrlr_failed().
Change-Id: I9e144328d524b25db2acf48e923b584746e8d0b6
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Replace the previous code that allocated each tracker individually with
one large allocation per queue pair.
struct nvme_tracker is now explicitly padded to reach exactly 4096 bytes
to allow normal array indexing to work correctly while maintaining the
alignment requirement that ensures each tracker's PRP list does not
cross a page boundary.
This also allows removal of the act_tr array, since the tr array can be
indexed directly now, and each tracker can store its own active state.
Change-Id: Ia7c51735b96594d12f7f478cefcc4aedc84207ad
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
The previous method for registering I/O queues did not allow the user
to specify queue priority for weighted round robin arbitration, and it
limited the application to one queue per controller per thread.
Change the API to require explicit allocation of each queue for each
controller using the new function spdk_nvme_ctrlr_alloc_io_qpair().
Each function that submits a command on an I/O queue now takes an
explicit qpair parameter rather than implicitly using the thread-local
queue.
This also allows the application to allocate different numbers of
threads per controller; previously, the number of queues was capped at
the smallest value supported by any attached controller.
Weighted round robin arbitration is not supported yet; additional
changes to the controller startup process are required to enable
alternate arbitration methods.
Change-Id: Ia33be1050a6953bc5a3cca9284aefcd95b01116e
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
When I/O spans a stripe boundary, the driver splits the request into
multiple requests, so for 1 segment memory larger than the stripe
size, we also need to split the segment memory.
Change-Id: I22ea5734d7066865a57a3c90fe18d5f76f373f1d
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
For those NVMe controllers which can support SGL feature in
firmware, we will use SGL for scattered payloads.
Change-Id: If688e6494ed62e8cba1d55fc6372c6e162cc09c3
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
This will be exposed in the public API. This rename is in a separate
commit to ease review.
Change-Id: I1b7fef36f85265db27935ac4d22ceef3c7282502
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Set up the infrastructure for creating I/O submission queues with
variable queue priority (QPRIO in Create I/O SQ command).
Currently, this is unused, since we always use the default arbitration
method (round robin), but it will allow reinitializing submission queues
with the correct priority once weighted round robin is supported.
Change-Id: I425003879e624cfcc9687bdc495b5c1726b5a8af
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Prepare for qpair to be exposed as part of the public API.
Change-Id: Ia63e863e95554adceeade20c829f12fe346375d5
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Allows simplification of _nvme_qpair_build_sgl_request(), which does not
need to know whether a request is a child or not.
This also removes a read of req->parent for non-split I/Os; the parent
field is in the section of nvme_request that is only intended to be
initialized for split I/Os, which should be detected by looking at
num_children.
Additionally, this fixes a potential problem if requests were nested
more than one level deep (e.g. req->parent was not the original user
request).
Change-Id: I3ea1dc134bbb1e3b8c6b5a479f5d760bd97ea848
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Instead of writing the completion doorbell once per completion,
just write it once at the end of the completion while loop.
This reduces the number of mmio writes by coalescing several
writes into one when we get multiple completions at a time.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I3cc3864dcfe43186bec51be1a732e84ef3be05ae
Also add a space between Copyright and (c).
The copyright year can be determined using git metadata.
Also remove the duplicated "All rights reserved." - every instance of
this line already has a corresponding "All rights reserved" immediately
below it, except for examples/ioat/kperf/kmod/dma_perf.c, where I have
added it manually.
Performed using this command:
git ls-files | xargs sed -i -e 's/Copyright(c) \(.*\) Intel Corporation. All rights reserved./Copyright (c) Intel Corporation./'
Change-Id: I3779f404966800709024eb1eb66a50068af2716c
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This will allow replacing these _nvme_fail_request_bad_vtophys() calls
with the correct error later. vtophys is not actually used within the
SGL request builder, so this is the wrong error.
Change-Id: Ibc2a3b029a8abad1d563b9df200325d7d64498da
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
No code change, just moved into a function for readability.
Change-Id: I883443c06d961c6dbeffed1a6fb153177e6e3fcd
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This cleans up the I/O splitting code somewhat.
It also moves the SGL payload function pointers up into the hot cache
section of struct nvme_request without pushing the other important
members past the cacheline boundary (because payload is now a union).
Change-Id: I14a5c24f579d57bb84d845147d03aa53bb4bb209
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
For the purpose to support different types of input scattered payloads,
such as iovs or scattered list, we define common method in the NVMe
driver, users should implement their own functions to iterate each
segment memory.
Change-Id: Id2765747296a66997518281af0db04888ffc4b53
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
nvme_ctrlr_process_io_completions() and
nvme_ctrlr_process_admin_completions() now return the number of
completions processed.
This also adds the possibility of returning an error from the
process_*_completions functions (currently unused, but this at least
gets the API ready in case error conditions are added later).
Change-Id: I1b32ee4f2f3c1c474d646fa2d6b8b7bbb769785f
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
nvme_qpair_submit_tracker() and nvme_qpair_manual_complete_request() are
only used from within nvme_qpair.c, so they can be static.
nvme_qpair_submit_tracker() is moved up to avoid needing a declaration
(no other code change).
nvme_ctrlr_hw_reset() is only used from within nvme_ctrlr.c, so it can
be static.
Change-Id: I9a7953d7baaec76e875dd535daf557ea24bef801
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
NVMe doesn't require the specific 64-bit MMIO ordering on 32-bit
platforms performed in spdk_mmio_read_8(), but it doesn't hurt.
We have to pick one of the two possible orderings, so pick the one
required by I/OAT.
Change-Id: I2b909d64d0c077b797d0f64a11d78d1ecc55eec7
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
nvme_ctrlr_process_io_completions() now takes a second parameter,
max_completions, to let the user limit the number of I/Os completed on
each poll.
If there are many I/Os waiting to be completed, the
nvme_ctrlr_process_io_completions() function could run for a long time
before returning control to the user, so the max_completions parameter
lets the user have more control of latency.
Change-Id: I3173059d94ec1cc5dbb636fc0ffd3dc09f3bfe4b
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
GCC generates a series of 64-bit MOV instructions for the memcpy() into
the submission queue. We can do better with 128-bit SSE2 instructions.
DPDK already has a memcpy implementation that is optimized for small
inline copies, so use it instead of memcpy.
Change-Id: I5f09259b4d5cb089ace4a8ea6d2078c03fee84f3
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
No change in behavior, just a simplification.
We already have a check for retry, so pull the cb_fn check out and put
it under the !retry branch.
This makes it clearer that requests that are going to be retried will
not get their callbacks called.
Change-Id: I70c7067e550c7fca78b0441b5474833f73863315
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Since it is empty function. It can be deleted.
Add nvme_assert to fix issue reported by scan-build.
Change-Id: Ia0e8f656e1dac0da7ec72f8404469ea1b0dcb40e
Signed-off-by: Liang Yan <liangx.yan@intel.com>
In nvme_qpair_complete_tracker, make sure we got a valid request in the
tracker that is being completed.
This should never occur in practice, but safeguard against it in case of
programmer error. Fixes a scan-build warning about potential NULL
dereference.
Change-Id: Id82af604d2a5ed5de0aeccf3affa1900f6712ebe
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Document the assumptions made by qpair_construct using asserts.
These values can't actually be 0 in practice due to the way they are
derived, but scan-build can't see that. It is also useful to have these
asserts in case of future modifications.
Change-Id: I546c057f5cbe7ccc62acd90b595e423cd450d86a
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
nvme_dump_command is totally unused aside from the unit test.
nvme_dump_completion was used in qpair, but it can be replaced with the
equivalent nvme_qpair_print_completion.
Also added the missing nvme_completion fields to nvme_qpair_print_completion
that had been printed by nvme_dump_command.
Change-Id: Ia5ee66f3553df06febe8f465d42e49a84c555dd2
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
