Currently the driver doesn't support completion vectors. These
are used to indicate which sets of CQs should be grouped together
into the same vector. A vector is a CQ processing thread that
runs on a specific CPU.
If an application has several CQs bound to different completion
vectors, and each completion vector runs on different CPUs, then
the completion queue workload is balanced. This helps scale as more
nodes are used.
Implement CQ completion vector support using a global workqueue
where a CQ entry is queued to the CPU corresponding to the CQ's
completion vector. Since the workqueue is global, it's guaranteed
to always be there when queueing CQ entries; Therefore, the RCU
locking for cq->rdi->worker in the hot path is superfluous.
Each completion vector is assigned to a different CPU. The number of
completion vectors available is computed by taking the number of
online, physical CPUs from the local NUMA node and subtracting the
CPUs used for kernel receive queues and the general interrupt.
Special use cases:
* If there are no CPUs left for completion vectors, the same CPU
for the general interrupt is used; Therefore, there would only
be one completion vector available.
* For multi-HFI systems, the number of completion vectors available
for each device is the total number of completion vectors in
the local NUMA node divided by the number of devices in the same
NUMA node. If there's a division remainder, the first device to
get initialized gets an extra completion vector.
Upon a CQ creation, an invalid completion vector could be specified.
Handle it as follows:
* If the completion vector is less than 0, set it to 0.
* Set the completion vector to the result of the passed completion
vector moded with the number of device completion vectors
available.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
All threads queuing CQ entries on different CQs are unnecessarily
synchronized by a spin lock to check if the CQ kthread worker hasn't
been destroyed before queuing an CQ entry.
The lock used in 6efaf10f16 ("IB/rdmavt: Avoid queuing work into a
destroyed cq kthread worker") is a device global lock and will have
poor performance at scale as completions are entered from a large
number of CPUs.
Convert to use RCU where the read side of RCU is rvt_cq_enter() to
determine that the worker is alive prior to triggering the
completion event.
Apply write side RCU semantics in rvt_driver_cq_init() and
rvt_cq_exit().
Fixes: 6efaf10f16 ("IB/rdmavt: Avoid queuing work into a destroyed cq kthread worker")
Cc: <stable@vger.kernel.org> # 4.14.x
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This matches what the userspace copy of this header has been doing
for a while. imm_data is an opaque 4 byte array carried over the network,
and invalidate_rkey is in CPU byte order.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Use correct parameter names and formatting in function kernel-doc notation
to eliminate warnings from scripts/kernel-doc.
../drivers/infiniband/sw/rdmavt/mr.c:784: warning: Excess function parameter 'ibmfr' description in 'rvt_map_phys_fmr'
../drivers/infiniband/sw/rdmavt/vt.c:234: warning: Excess function parameter 'intex' description in 'rvt_query_pkey'
../drivers/infiniband/sw/rdmavt/vt.c:266: warning: Excess function parameter 'index' description in 'rvt_query_gid'
../drivers/infiniband/sw/rdmavt/vt.c:306: warning: Excess function parameter 'data' description in 'rvt_alloc_ucontext'
../drivers/infiniband/sw/rdmavt/cq.c:65: warning: Excess function parameter 'sig' description in 'rvt_cq_enter'
../drivers/infiniband/sw/rdmavt/qp.c:279: warning: Excess function parameter 'qpt' description in 'rvt_free_all_qps'
../drivers/infiniband/sw/rdmavt/mcast.c:282: warning: Excess function parameter 'igd' description in 'rvt_attach_mcast'
../drivers/infiniband/sw/rdmavt/mcast.c:345: warning: Excess function parameter 'igd' description in 'rvt_detach_mcast'
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Doug Ledford <dledford@redhat.com>
Cc: Jason Gunthorpe <jgg@mellanox.com>
Cc: linux-doc@vger.kernel.org
Acked-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
CQ allocation does not ensure that completion queue entries
and the completion queue structure are allocated on the correct
numa node.
Fix by allocating the rvt_cq and kernel CQ entries on the device node,
leaving the user CQ entries on the default local node. Also ensure
CQ resizes use the correct allocator when extending a CQ.
Reviewed-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This patch series primarily increases sizes of variables that hold
lid values from 16 to 32 bits. Additionally, it adds a check in
the IB mad stack to verify a properly formatted MAD when OPA
extended LIDs are used.
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
slid field in struct ib_wc is increased to 32 bits.
This enables core components to use larger LIDs if needed.
The user ABI is unchanged and return 16 bit values when queried.
Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The following fields are defined for filtering and triggering:
- wr_id
- status
- opcode
- qpn
- length
- idx
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Get rid of this warning:
drivers/infiniband/sw/rdmavt/cq.c: In function ‘rvt_cq_exit’:
drivers/infiniband/sw/rdmavt/cq.c:542:2: warning: ‘worker’ may be used uninitialized in this function [-Wmaybe-uninitialized]
kthread_destroy_worker(worker);
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
by fixing the function to actually work.
Fixes: 6efaf10f16 ("IB/rdmavt: Avoid queuing work into a destroyed cq kthread worker")
Cc: Petr Mladek <pmladek@suse.com>
Cc: Doug Ledford <dledford@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Use the new API to create and destroy the cq kthread worker.
The API hides some implementation details.
In particular, kthread_create_worker() allocates and initializes
struct kthread_worker. It runs the kthread the right way and stores
task_struct into the worker structure. In addition, the *on_cpu()
variant binds the kthread to the given cpu and the related memory
node.
kthread_destroy_worker() flushes all pending works, stops
the kthread and frees the structure.
This patch does not change the existing behavior. Note that we must
use the on_cpu() variant because the function starts the kthread
and it must bind it to the right CPU before waking. The numa node
is associated for given CPU as well.
Signed-off-by: Petr Mladek <pmladek@suse.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The memory barrier is not enough to protect queuing works into
a destroyed cq kthread. Just imagine the following situation:
CPU1 CPU2
rvt_cq_enter()
worker = cq->rdi->worker;
rvt_cq_exit()
rdi->worker = NULL;
smp_wmb();
kthread_flush_worker(worker);
kthread_stop(worker->task);
kfree(worker);
// nothing queued yet =>
// nothing flushed and
// happily stopped and freed
if (likely(worker)) {
// true => read before CPU2 acted
cq->notify = RVT_CQ_NONE;
cq->triggered++;
kthread_queue_work(worker, &cq->comptask);
BANG: worker has been flushed/stopped/freed in the meantime.
This patch solves this by protecting the critical sections by
rdi->n_cqs_lock. It seems that this lock is not much contended
and looks reasonable for this purpose.
One catch is that rvt_cq_enter() might be called from IRQ context.
Therefore we must always take the lock with IRQs disabled to avoid
a possible deadlock.
Signed-off-by: Petr Mladek <pmladek@suse.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
A good practice is to prefix the names of functions by the name
of the subsystem.
The kthread worker API is a mix of classic kthreads and workqueues. Each
worker has a dedicated kthread. It runs a generic function that process
queued works. It is implemented as part of the kthread subsystem.
This patch renames the existing kthread worker API to use
the corresponding name from the workqueues API prefixed by
kthread_:
__init_kthread_worker() -> __kthread_init_worker()
init_kthread_worker() -> kthread_init_worker()
init_kthread_work() -> kthread_init_work()
insert_kthread_work() -> kthread_insert_work()
queue_kthread_work() -> kthread_queue_work()
flush_kthread_work() -> kthread_flush_work()
flush_kthread_worker() -> kthread_flush_worker()
Note that the names of DEFINE_KTHREAD_WORK*() macros stay
as they are. It is common that the "DEFINE_" prefix has
precedence over the subsystem names.
Note that INIT() macros and init() functions use different
naming scheme. There is no good solution. There are several
reasons for this solution:
+ "init" in the function names stands for the verb "initialize"
aka "initialize worker". While "INIT" in the macro names
stands for the noun "INITIALIZER" aka "worker initializer".
+ INIT() macros are used only in DEFINE() macros
+ init() functions are used close to the other kthread()
functions. It looks much better if all the functions
use the same scheme.
+ There will be also kthread_destroy_worker() that will
be used close to kthread_cancel_work(). It is related
to the init() function. Again it looks better if all
functions use the same naming scheme.
+ there are several precedents for such init() function
names, e.g. amd_iommu_init_device(), free_area_init_node(),
jump_label_init_type(), regmap_init_mmio_clk(),
+ It is not an argument but it was inconsistent even before.
[arnd@arndb.de: fix linux-next merge conflict]
Link: http://lkml.kernel.org/r/20160908135724.1311726-1-arnd@arndb.de
Link: http://lkml.kernel.org/r/1470754545-17632-3-git-send-email-pmladek@suse.com
Suggested-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Petr Mladek <pmladek@suse.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Jiri Kosina <jkosina@suse.cz>
Cc: Borislav Petkov <bp@suse.de>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The priority of the send engines is higher than the CQ completion
thread potentially causing completions to be starved for very
fast interfaces.
Change the CQ kthread to match the send engine threads to minimize
this delay for ULP completion processing.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
While hfi1 and qib were still supporting bits and pieces of core verbs
components there needed to be a way to convey if rdmavt should handle
allocation and initialize of resources like the queue pair table. Now
that all of this is moved into rdmavt there is no need for these flags.
They are no longer used in the drivers.
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Add, remove, and otherwise clean up existing comments that are leftover
from the initial code postings of rdmavt. Many of the comments were added
to provide an idea on the direction we were thinking of going. Now that the
design is solidified make a pass over and clean everything up. Also add
details where lacking.
Ensure all non static functions have nano comments.
Reviewed-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Return directly from rvt_resize_cq rather than use a goto/label.
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Update all files added by rdmavt which do not yet have 2016 as the
copyright year.
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Brings in completion queue functionality. A kthread worker is added to
the rvt_dev_info to serve as a worker for completion queues.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Create stubs for completion queue creation, polling,
resizing, calling for notification, and destroying.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Sebastian Sanchez