Print the error information when the kernel RNIC driver did not load
properly, and fix the cleanup logic for the exceptional exit.
Change-Id: I97a45e73d830280b994818f3defc491bc2b6b020
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
If the RDMA transport failed to initialize, g_rdma.event_channel may be
NULL.
Change-Id: I4510ee5893389f244f0fbaa1cd4a182868939b25
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
For iWARP devices, buffers that are intended to be the
target of an RDMA read initiated by the target must additionally
have IBV_ACCESS_REMOTE_WRITE permission. This is because iWARP's
RDMA read path essentially requests the remote side to do
an RDMA write.
This is unfortunate because there is no way to differentiate between
memory that the remote side can do an RDMA write to and memory
that will only be the target of RDMA reads initiated by the
target. There is nothing we can do about this serious deficiency in
the specification, however, so we have to live with it.
Change-Id: I3d2f2814ce0cb1df4e5347296ef371db4d16be21
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
I believe this is required for NICs to report, but handle
the case where it isn't reported.
Change-Id: I38d10c3590d1df8bb902ab312af0f9e01b9e5032
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
This makes it consistent with the way connections and
requests work.
Change-Id: Ifb97499ba72f7dfd02ac54ba1b622726d266262c
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
The shared memory pool for a session is associated with
a particular RNIC via the protection domain. New connections
attempting to join a session that came in on a different RNIC
can't use that memory, so must be rejected.
Change-Id: Ibd79fe90566a231f76b7472e5e9b484c3e528454
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Rearrange the functions in rdma.c to match the order
of the function pointers in the transport. No other
code changes.
Change-Id: I9dbc68912ecd5dfdf53f20b4807d4116933a3c3a
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Use the lower level registration functions. The RDMA-CM
examples use the ibv_* versions, so who knows if the
rdma_reg_* wrappers are even well tested.
Change-Id: I8e8250ab09a1401e636aebe2fc04a60806f7a827
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Previously, we mixed use free and spdk_nvmf_rdma_conn_destroy to
free allocated spdk_nvmf_rdma_conn structure, which sounds not
exactly free all the resources.
Change-Id: I2917b442c34d63ba5c014add58f429ae4b831595
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
The RDMA API doesn't say whether the wr is copied, so be
safe and allocate it on the heap.
Change-Id: I091af50aa031e1861333f19d864eb52335d6b756
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
This changes as little code as possible while still creating
a single public API header. This enables future clean up
of the public API and clarification of the exposed
concepts.
Change-Id: I780e7a5a9afd27acf0276516bd71b896ad301c50
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Remove #includes for all DPDK headers that weren't
necessary.
Change-Id: Ib02522e0f04e64a1c98afceb7508cc0e8d931a9d
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Use the env library to perform all memory allocations
that previously called DPDK directly.
Change-Id: I6d33e85bde99796e0c85277d6d4880521c34f10d
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
All completion queues for the same listen address
now share a common completion queue channel.
Change-Id: I42c149fe7e221951e8a3826b1713482c37a265b8
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
These 4 callbacks can be condensed into two callbacks, which
simplifies the API.
Change-Id: I069da00de34b252753cdc8961439e13a75d1cc68
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
We already require the assert header from the C standard library,
so use that instead of RTE_VERIFY to further isolate DPDK
dependencies.
Change-Id: I4a718af858c88aff6080e33e6c3dd533c077b8f4
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Instead of polling for only 1 completion at at time,
poll for batches of 32.
Change-Id: I5ef99a270489e7b3d2a58cb765915f187775a93e
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
ibv_poll_cq is actually an expensive call to make, so take
steps to begin to minimize the number of times it is called.
Change-Id: I6fc64979604220eb8cacd612b46e3a3b1bca0924
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Fix copy-and-paste errors - when polling the recv CQ, we should print
"Recv" instead of "Send" in log messages.
Signed-off-by: Roland Dreier <roland@purestorage.com>
This function always shows up as one of the hottest functions when
profiling. I believe it is the memset that is expensive, so instead
use default initialization when the wr is declared on the stack
and just set the members that need to be updated in the function.
Also make the function inline for good measure.
Change-Id: I29e24cdd375311fa033b5a6df772ff4f73e35302
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
The application is now entirely responsible for scheduling subsystem
pollers and sending events between threads.
Change-Id: I88da1f53b5e8852c7c4acd6f0a7a1e2219fbed41
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
The "+" is not correct, should be "-". Currently,
the issue doest not happen since the offset is 0,
then both + and - is OK. But if we adjust the location
of spdk_nvmf_conn or spdk_nvmf_request, we can find
this bug.
Change-Id: Ib358dc729da901a69442d0402a6089989f49b05c
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
These are specified as "ASCII string", which means they should be
left-aligned and padded with spaces, according to the NVMe base
specification.
Change-Id: I25babe0ca417c2e16137b0bfc41fc7834277114e
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Currently the NVMf target listens for new connections on any address.
Instead, listen only on the addresses specified by the user.
Change-Id: Idb6d37c422e442fc70a8673bd3fcfb9c27b57828
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
The NVMe over Fabrics target library now exposes a simple function call
that polls the acceptor once, and the application handles registration
of the poller.
Also rename the transport function pointers related to the acceptor so
they better reflect their purpose.
Change-Id: I5fa0d516586bf17e73afeb88ff3c2d5b0d46794d
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This will become more important when other transports are added.
For now, it is also useful to be able to start nvmf_tgt on systems
without RDMA hardware.
Change-Id: I6b9002cc7711f928c4e6b73adcd9b677349ebdd6
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Users can specify the core for each subsystem and the acceptor listen routine
to run on different cores for performance consideration.
Change-Id: I4bd1a96f39194c870863b4b778e6ea7cf8fc1a2d
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
This leaves more flexibility for future changes to the poller
representation without requiring API changes (after this one).
It also prevents the user from accidentally using poller fields in a
non-thread-safe way, since they can't be accessed directly anymore.
Change-Id: I7677d5b93668665d29ae39c5e0ba74333ad3f878
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
The NVMe submission queue head wraparound point can be determined in the
generic NVMe over Fabrics layer; it should not be using the RDMA
connection queue depth.
Change-Id: I9da8f09e4f057f8fdc1ff4c6cc5f48cea7123e11
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Use the max_queue_depth parameter rather than rdma_conn->max_queue_depth
so that we can start to eliminate rdma_conn->max_queue_depth.
Change-Id: I1670c634e6d12aa004fb5a10338b7624850fbc4a
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
The large buffer pool allocation was using the per-connection queue
depth, whereas the RDMA memory region registration was using the global
RDMA max queue depth. These sizes need to match, so use the global RDMA
max queue depth for both calls.
Change-Id: Iae161b719e09e19ca3e81df6593b68a4a2e86614
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Use the new timer-based poller functionality to replace rte_timer.
Change-Id: Ic40653306cc73b40139fe18e06bab29b35721a43
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Allow higher queue depths by allowing many more send/recv
operations than read/write.
Change-Id: I66c424a6463e5e09be6d5463667241ce9271404b
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
The target can only provide updates to sq_head inside
of completions. Therefore, we must update sq_head prior
to sending the completion or we'll incorrectly get into
queue full scenarios.
Change-Id: If2925d39570bbc247801219f352e690d33132a2d
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
This allows the target to poll for internal completions
at higher priority.
Change-Id: I895c33a594a7d7c0545aa3a8405a296be3c106fb
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
This ensures that the data buffers are not in use
when we go to send the completion.
Change-Id: I30467b3e3964001150f81b21e5b695dcd0974b0c
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
This is useful for holding session-wide buffer pools.
Change-Id: I7024da24b210a2205bf1e159d5935e0093b81120
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
For small SGLs, even if they are keyed and not inline, use the
buffer we allocated for inline data.
Change-Id: I5051c43aabacb20a4247b2feaf2af801dba5f5a9
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Read/Write depth is much lower than Send/Recv depth.
Calculate them separately to prepare for supporting
a larger number of receives than read/writes.
Currently, the target still only exposes a queue depth
equal to the read/write depth.
Change-Id: I08a7434d4ace8d696ae7e1eee241047004de7cc5
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
These don't actually work quite yet, but pipe the
configuration file data through to where it will
be needed.
Change-Id: I95512d718d45b936fa85c03c0b80689ce3c866bc
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
