For namespaces with end-to-end protection information, metadata size
of exactly 8 bytes, and extended LBA configured, the NVMe driver would
calculate the size of the data block incorrectly. The NVMe spec has a
special provision for this specific case (8-byte metadata only) and
PRACT = 1 that requires that the host does not send the metadata as part
of the host memory buffer.
To fix this, clean up the calculation of the per-block data transfer
size by adding a new extended_lba_size field in the namespace, which
represents the total size of data to be transferred per block based on
the namespace's configured metadata size and whether it transfers
metadata as part of the data buffer. Then add the special case for
PRACT = 1 and PI configured and extended LBA in the R/W helper
functions.
Change-Id: I0b383a58c773cac06e6c018858b57129064c6059
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This prepares for using this in tests for other libraries.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Ibd4d12006d602f772cd9253f3eab9d47c0593767
This test already links nvme.c via OTHER_FILES, which provides the
g_spdk_nvme_driver definition.
Also make g_request static while we're here, since it is only used
within this test file.
Change-Id: I9f1a2a7b3d40d2ea0989f6f781a4d220afbd3f0e
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Since nvme_ns_cmd.c now walks the SGL, some of the test code
needs to also be updated to initialize and return correct values
such as ctrlr->flags and sge_length.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I521213695def35d0897aabf57a0638a6c347632e
Now that the hotplug code is isolated in nvme_pcie.c, it can call the
PCIe transport attach function directly.
Change-Id: I2df3b9168473b537cc9b13367e06d3d3b6fa22be
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
The other simplifications to probe_info and trid made the
trtype argument redundant.
Change-Id: Ie7bea4e2204e690dc4909eeacd065e0722b53272
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
The probe_info was reduced to just containing a
transport_id, so remove probe_info entirely.
Change-Id: Ica9a22d126cd14e282decd3eea1a0afe0460f099
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Scanning the transport may result in both new
devices and removed devices, so pass the callback
for both operations.
Change-Id: I6f73dbe6fd7cf61575c354b43f8ae3e2a01e2965
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Simplify the arguments to nvme_transport_ctrlr_scan to take
a transport id that identifies the discovery service (or
NULL to scan PCIe).
Further, separate scan into two functions - scan and attach.
Scan is for scanning an entire bus, attach is for a specific
device.
Change-Id: I464f351a02a04bc5a45096dcf5dc8fc5ac489041
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
This is a small step toward making discovery more like
scanning a local PCI bus.
Change-Id: Ie7149ad060f2eeb56939b1241187bdf09681f2aa
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
It's not the whole transport - it's just an enum for the
type of transport.
Change-Id: Ia435a21792f221ddf50ddf4f0923c6152622eccb
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Change the PCI enumeration API to individual functions per device type
so that only the drivers that are actually in use get linked into the
final executable. All of the common code is still shared internally in
the env_dpdk library.
Change-Id: I2ba83afe59202a510f999a0674e23e60b6581221
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This version of multi-process support needs to have DPDK 16.11 builtin.
Change-Id: I3352944516f327800b4bd640347afc6127d82ed4
Signed-off-by: GangCao <gang.cao@intel.com>
Let the transport access the controller options during
ctrlr_construct().
Change-Id: I83590c111e75c843685dd9315f0f08416168356d
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Use the NVMe over Fabrics spec definitions for TRTYPE rather than the
internal library transport type.
Change-Id: Idead559a8f8d95274fc580d10e82033822e6eda8
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Considering the process can be terminated in the cases like ctrl+c,
kill command or memory fault, the ref is tracked in the per process
structure spdk_nvme_controller_process and whenever there is other
process attaches or detaches the controller, a scan will be issued
to cleanup those unexpectedly exited processes.
Change-Id: Ib4f974f567a865748d42da4ead49edd383dfc752
Signed-off-by: GangCao <gang.cao@intel.com>
Instead of the next_sge callback returning the physical address
directly, make it return the virtual address and convert to physical
address inside the NVMe library.
This is necessary for NVMe over Fabrics host support, since the RDMA
userspace API requires virtual addresses rather than physical addresses.
It is also more consistent with the normal non-SGL NVMe functions that
already take virtual addresses.
Change-Id: I79a7af64ead987535f6bf3057b2b22aef3171c5b
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Function pointers will not work for the DPDK multi-process model (they
can have different addresses in different processes), so define a
transport enum and dispatch functions that switch on the transport type
instead.
Change-Id: Ic16866786eba5e523ce533e56e7a5c92672eb2a5
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
spdk_nvme_probe() will now provide a struct spdk_nvme_probe_info to the
probe and attach callbacks in place of the PCI device pointer.
This struct contains the useful information that could be retrieved from
the PCI device during probe.
The goal of this change is to allow expansion of the probe information
in the future when other transports (specifically, NVMe over Fabrics)
are added that do not necessarily use PCI addressing or device IDs.
Change-Id: I59a2a9e874e248ce5fa1d7f4b57c8056962ff3cd
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
- Split the part that gets a PCI device's address into its own function,
spdk_pci_device_get_addr(). This is useful outside of the comparison
function and is orthogonal to comparing addresses.
- Make the comparison function take two addresses instead of a device
and an address. The more general form will be useful with addresses
that are not directly associated with a device. Because of this, also
rename the function from spdk_pci_device_compare_addr() to
spdk_pci_addr_compare().
- Return a signed value similar to strcmp() so that addresses can be
ordered, not just compared for equality.
Change-Id: Idf304454af09ea57f1e1d5dc3a39b077378cecad
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Make the transport ctrlr_construct callback responsible for allocating
its own controller.
Change-Id: I5102ee233df23e27349410ed063cde8bfdce4c67
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This will allow factoring out PCIe-specific code into a swappable
transport so that NVMe over Fabrics host support can be added.
Change-Id: I4df74dd268d655e3b36e8d6114ebe7d79a24844d
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Enforce exactly one trailing \n, and fix all of the existing cases.
Change-Id: I6218e4700e90aeb647eaee78089530c79993c8c8
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This patch also drops support for automatically unbinding
devices from the kernel - run scripts/setup.sh first.
Our generic pci interface is now hidden behind include/spdk/pci.h
and implemented in lib/util/pci.c. We no longer wrap the calls
in nvme_impl.h or ioat_impl.h. The implementation now only uses
DPDK and the libpciaccess dependency has been removed. If using
a version of DPDK earlier than 16.07, enumerating devices
by class code isn't available and only Intel SSDs will be
discovered. DPDK 16.07 adds enumeration by class code and all
NVMe devices will be correctly discovered.
Change-Id: I0e8bac36b5ca57df604a2b310c47342c67dc9f3c
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Provide a convenience wrapper for general purpose dataset
management commands. The previous wrapper for deallocate
was difficult to use correctly and only for deallocate.
Note that the name is "dataset_management" as opposed to
"data_set_management" to match the NVMe specification.
It's questionable whether "dataset" is valid English, but
it is best to match the specification.
Change-Id: Ifc03d66dbabeabe8146968cf8a09f7ac3446ad68
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Rather than forcing the NVMe library user to pass a specially-allocated
block of memory (e.g. rte_malloc() in the case of the default
nvme_impl.h), just make the NVMe library allocate a suitable buffer
itself and copy to/from the user buffer as needed.
The fast path I/O functions still require special rte_malloc()
allocations, since we don't want to add an allocation and copy to the
I/O critical path.
Change-Id: I7fe88c0ba60c859a33bbe95b7713f423c6bf1ea8
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Use SPDK_CU_ASSERT_FATAL instead of CU_ASSERT_FATAL so static analyzers
recognize that g_request cannot be NULL in the following lines.
Change-Id: Ie7ab3bd34a177bea0d565441014e8db12be8bb01
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Resolve relative paths before using them to clean up command lines.
This should also help shorten the overall command line length that gets
embedded in the binary and used when locating the executable from a
coredump.
Change-Id: Ibff9849ede198bb04313496c8b7131485ffaf14f
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Rename nvme_remove_child_request() to nvme_request_remove_child() and
move it next to nvme_request_add_child() to make the symmetry clear.
Change-Id: I78747c44ab3db1a656b33555a45f634dc5a55b31
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This patch is used to add a nvme_request remove child
helpler function
Change-Id: I1e5bb228d53333ca3601f4ae30fcd801ea39e532
Signed-off-by: Ziye Yang <ziye.yang@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>
Provide a new structure, spdk_nvme_ctrlr_opts, to let the user modify
the default controller initialization options during probe/attach.
Currently, only the number of queue pairs can be modified in this way;
other options will be added later.
Change-Id: Ie27b9429291d93a9353c0d820f0ad467d3b0e7cb
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>
Add CUnit test case to verify payload_offset value in split_test4
Change-Id: I4a9a33854295ed802709bbe4f11746d284ed8cbd
Signed-off-by: Liang Yan <liangx.yan@intel.com>
When multiple NVMe controllers are being initialized during
spdk_nvme_probe(), we can overlap the hardware resets of all controllers
to improve startup time.
Rewrite the initialization sequence as a polling function,
nvme_ctrlr_process_init(), that maintains a per-controller state machine
to determine which initialization step is underway. Each step also has
a timeout to ensure the process will terminate if the hardware is hung.
Currently, only the hardware reset (toggling of CC.EN and waiting for
CSTS.RDY) is done in parallel; the rest of initialization is done
sequentially in nvme_ctrlr_start() as before. These steps could also be
parallelized in a similar framework if measurements indicate that they
take a significant amount of time.
Change-Id: I02ce5863f1b5c13ad65ccd8be571085528d98bd5
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Two test cases should be added.
1. ns cmd don't have child requests;
2. Assert that the correct number of child requests are created
and verify that each one has the correct payload_offset.
Change-Id: I1182a1a6673ceaf2ba35be268f80d8668af82848
Signed-off-by: Liang Yan <liangx.yan@intel.com>
