The hfi1_cdbg() macro can be instantiated in the hot path even when it
is not in use. This shows up on perf profiles.
Rework the macros (for SDMA and MMU), to use the trace interface directly
to eliminate this performance hit.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Clean up user_exp_rcv.c file by moving structure definitions into header
file user_exp_rcv.h. Since these structure definitions depend on the
structure definitions in mmu_rb.h, move #include "mmu_rb.h" above
the include "user_exp_rcv.h" or include of header files that include
user_exp_rcv.h
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
num_user_pages() function has been defined in both user_exp_rcv.c file
and user_sdma.c file. Move the function definition to a header file so
there is only one definition in the source repo.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
In pin_vector_pages() function, if there is any error while pinning the
pages or while adding a pinned buffer to the cache, the bail out code
needs to unpin any pinned pages that are not in the cache and adjust the
n_locked counter that counts the total pages pinned. The current bail
out code doesn't seem to be doing it right in two cases:
1. Before pinning required pages for a buffer, the SDMA pinned buffer
cache is searched to see if the virtual address range that needs to be
pinned is already pinned. If there isn't a hit in the cache, a new node
is created for the buffer and is added to the cache after the buffer is
pinned. If adding the new node to the cache fails, the n_locked count is
decremented properly but the pinned pages are not freed. This commit
fixes this issue.
2. If there is a hit in the SDMA cache, but the cached buffer doesn't
have enough pages to cover the entire address range that needs to be
pinned, the node for the cached buffer is extracted from the cache,
remaining pages needed are pinned and added to the node. The node is
finally added back into the cache. If there is an error pinning the
extra pages, the bail out code frees all the pages in the node but the
n_locked count is not being decremented by the no of pages in the node
that are freed. This commit fixes this issue.
This commit fixes the above two issues by creating a new function that
frees the pages in a node and decrements the n_locked count by the
number of pages freed.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Performance analysis shows that the cache callback function
sdma_kmem_cache_ctor contributes to 1/2 of the kmem_cache_allocs
time.
Since all of the fields in the allocated data structure are initialized
in the code path, remove the _ctor function.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
PIO/SDMA send logic now uses the hdr_type field to determine
the type of packet that has been constructed. Based on the hdr_type,
certain things such as PBC flags, padding count and the LT extra
trailing bytes are determined.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The allocate_ctxt() function adds the context to the fd data structure.
Since the context is not completely initialized, this can cause confusion
as to whether the context is valid or not.
Move the fd reference from allocate_ctxt() to setup_base_ctxt().
Update the necessary functions to be aware of this move.
Reviewed-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
When an egress resource(SDMA descriptors, pio credits) is not available,
a sending thread will be put on the resource's wait queue. When the
resource becomes available again, up to a fixed number of sending threads
can be awakened sequentially and removed from the wait queue, depending
on the number of waiting threads and the number of free resources. Since
each awakened sending thread will send as many packets as possible, it
is highly likely that the first sending thread will consume all the
egress resources. Subsequently, it will be put back to the end of the wait
queue. Depending on the timing when the later sending threads wake up,
they may not be able to send any packet and be again put back to the end
of the wait queue sequentially, right behind the first sending thread.
This starvation cycle continues until some sending threads exceed their
retry limit and consequently fail.
This patch fixes the issue by two simple approaches:
(1) Any starved sending thread will be put to the head of the wait queue
while a served sending thread will be put to the tail;
(2) The most starved sending thread will be served first.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Atomic bit tests are used to single errors and the
completion of request submissions. These operations
don't need to be atomic and show to be expensive on
the profile.
Replace each atomic bit operation with a bool type
and a READ_ONCE/WRITE_ONCE pairing.
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>
The atomic SDMA_REQ_SEND_DONE bit is set by the
process-level code, and then the same process-level
code uses the bit to test that all packets have been
submitted incurring a costly atomic read.
Use a bool type with a READ_ONCE/WRITE_ONCE
pairing for this bit, and use the same condition that
is used to set the bit to test that all packets have
been submitted.
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>
The current user SDMA request structure layout has holes.
The cachelines can be reduced to improve cacheline trading.
Separate fields in the following categories: mostly read,
writable and shared with interrupt.
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>
An RB tree is used for the SDMA pinning cache. Cache
entries are extracted and reinserted from the tree
in case the address range for it changes. However,
if the address range for the entry doesn't change,
deleting the entry from the RB tree is not necessary.
This affects performance since the tree needs to be
rebalanced for each insertion, and this happens in
the hot path. Optimize RB search by not removing
entries when it's not needed.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Mitko Haralanov <mitko.haralanov@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>
The tx request is unnecessarily initialized in the hot
code path with memset(), however, there's no need to do
this as most fields are initialized later on. this
initialization shows to be costly in the profile.
Remove unnecessary initialization from tx request and make
sure all variables are initialized properly.
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>
The error path for context initialization is not consistent. Cleanup all
resources on failure.
Removed unused variable user_event_mask.
Add the _BASE_FAILED bit to the event flags so that a base context can
notify waiting sub contexts that they cannot continue.
Running out of sub contexts is an EBUSY result, not EINVAL.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The AHG index is only accessed in the request call
from user space, so there's no need for atomic semantics.
Replace atomic operations for SDMA_REQ_HAVE_AHG bit
with a test of the AHG index.
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>
Div instructions show costly in profiles when
the tx request header is set. Using right shift
instead of a divide operation reduces the cycles
spent in the function that sets the tx request
header as shown in the profile. Use right shift
operation instead.
Profile before change:
43.24% 009
|
|--23.41%-- user_sdma_send_pkts
| |
| |--99.90%-- hfi1_user_sdma_process_requestAfter:
Profile after change:
45.75% 009
|
|--14.81%-- user_sdma_send_pkts
| |
| |--99.95%-- hfi1_user_sdma_process_request
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>
Replace the specification of a data structure by a reference to
the desired member as the parameter for the operator "sizeof" to make
the corresponding size determination a bit safer according to
the Linux coding style convention.
Signed-off-by: Markus Elfring <elfring@users.sourceforge.net>
Signed-off-by: Doug Ledford <dledford@redhat.com>
* Pass a product for a call of the function "vmalloc_user" without storing
it in an intermediate variable.
* Delete the local variable "memsize" which became unnecessary with
this refactoring.
Signed-off-by: Markus Elfring <elfring@users.sourceforge.net>
Signed-off-by: Doug Ledford <dledford@redhat.com>
* Multiplications for the size determination of memory allocations
indicated that array data structures should be processed.
Thus reuse the corresponding function "kcalloc".
This issue was detected by using the Coccinelle software.
* Replace the specification of a data type by a pointer dereference
to make the corresponding size determination a bit safer according to
the Linux coding style convention.
Signed-off-by: Markus Elfring <elfring@users.sourceforge.net>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Pull rdma updates from Doug Ledford:
"This is the complete update for the rdma stack for this release cycle.
Most of it is typical driver and core updates, but there is the
entirely new VMWare pvrdma driver. You may have noticed that there
were changes in DaveM's pull request to the bnxt Ethernet driver to
support a RoCE RDMA driver. The bnxt_re driver was tentatively set to
be pulled in this release cycle, but it simply wasn't ready in time
and was dropped (a few review comments still to address, and some
multi-arch build issues like prefetch() not working across all
arches).
Summary:
- shared mlx5 updates with net stack (will drop out on merge if
Dave's tree has already been merged)
- driver updates: cxgb4, hfi1, hns-roce, i40iw, mlx4, mlx5, qedr, rxe
- debug cleanups
- new connection rejection helpers
- SRP updates
- various misc fixes
- new paravirt driver from vmware"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dledford/rdma: (210 commits)
IB: Add vmw_pvrdma driver
IB/mlx4: fix improper return value
IB/ocrdma: fix bad initialization
infiniband: nes: return value of skb_linearize should be handled
MAINTAINERS: Update Intel RDMA RNIC driver maintainers
MAINTAINERS: Remove Mitesh Ahuja from emulex maintainers
IB/core: fix unmap_sg argument
qede: fix general protection fault may occur on probe
IB/mthca: Replace pci_pool_alloc by pci_pool_zalloc
mlx5, calc_sq_size(): Make a debug message more informative
mlx5: Remove a set-but-not-used variable
mlx5: Use { } instead of { 0 } to init struct
IB/srp: Make writing the add_target sysfs attr interruptible
IB/srp: Make mapping failures easier to debug
IB/srp: Make login failures easier to debug
IB/srp: Introduce a local variable in srp_add_one()
IB/srp: Fix CONFIG_DYNAMIC_DEBUG=n build
IB/multicast: Check ib_find_pkey() return value
IPoIB: Avoid reading an uninitialized member variable
IB/mad: Fix an array index check
...
For the received packets with payload less or equal 8DWS
RxDmaDataFifoRdUncErr is not reported. There is set RHF.EccErr
if the header is not suppressed. When such packet is detected
on the send side the header suppression mechanism is disabled
by clearing SH bit in the packet header.
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jakub Pawlak <jakub.pawlak@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Some users want more control over which cpu cores are being used by the
driver. For example, users might want to restrict the driver to some
specified subset of the cores so that they can appropriately partition
processes, irq handlers, and work threads.
To allow the user to fine tune system affinity settings new sysfs
attributes are introduced per sdma engine. This patch adds a new
attribute type for sdma engine and a new cpu_list attribute.
When the user writes a cpu range to the cpu_list attribute the driver
will create an internal cpu->sdma map, which will be used later as a
look-up table to choose an optimal engine for a user requests.
Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Reviewed-by: Jianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: Tadeusz Struk <tadeusz.struk@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Each user SDMA request coming into the driver may contain multiple packets.
Each user packet may use multiple SDMA descriptors to fill the send buffer.
The field seqsubmitted in struct user_sdma_request counts the number of
user packets submitted to an SDMA engine. Sometimes, the intermediate count
may not be updated properly. However, once all the packets' descriptors
are successfully submitted to the SDMA engine, the final count is updated
correctly. But, if only some of the packets are submitted to the engine due
to an error, the intermediate count doesn't reflect the partial number of
packets submitted to the SDMA engine. This can cause a hang later in the
code as the count of packets submitted to the SDMA engine doesn't match the
the count of packets processed by the SDMA engine.
Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
In the set_txreq_header_ahg(), The KDETH Intr bit is obtained from the
header in the user sdma request using a KDETH_GET shift and mask macro.
This value is then futher right shifted by 16 causing us to lose the
value i.e it is shifted to zero, leading to the following
smatch warning:
drivers/infiniband/hw/hfi1/user_sdma.c:1482 set_txreq_header_ahg()
warn: mask and shift to zero
The Intr bit should be left shifted into its correct position in the
KDETH header before the AHG update.
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The ops->remove() callback was called by hfi1_mmu_unregister() with a
NULL mm argument while holding a spinlock. In the case of sdma_rb_remove()
this caused it to pass current->mm to hfi1_release_user_pages()
This had 2 problems. First this would attempt to acquire the mmap_sem
under a spin lock. Second the use of current->mm is not always guaranteed
to be the proper mm when the fd is being closed.
Rather than depend on this implicit behavior we move all calls to
ops->remove outside of the spinlock. This also allows the correct
mm to be used in the remove callback without fear of deadlock.
Because the MMU notifier is not guaranteed to hold mm->mmap_sem, but
usually does, we must delay all remove callbacks until out of the notifier,
when the callbacks can take the mmap_sem if they need to.
Code comments were added to clarify what the expectations are for the
users of the mmu rb tree.
Suggested-by: Jim Foraker <foraker1@llnl.gov>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Use the new cache evict operation in the SDMA code. This allows the cache
to properly coordinate evicts and removes, preventing any race. With this
change, the separate list, lock, and race flag are not needed.
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The objects which use cache handling should reference their own handler
object not the internal data structure it uses to track the nodes.
Have the "users" of the mmu notifier code pass opaque objects which can
then be properly used in the mmu callbacks depending on the owners needs.
This patch has the additional benefit that operations no longer require a
look up in a list to find the handlers.
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The hfi1 driver registers a mmu_notifier callback when /dev/hfi1_* is
opened, and unregisters it when the device is closed. The driver
incorrectly assumes that the close will always happen from the same
context as the open. In particular, closes due to SIGKILL or OOM killer
activity may happen from a different context. In these cases, the wrong
mm is passed to mmu_notifier_unregister(), which causes improper reference
counting for the victim mm, and eventual memory corruption.
Preserve the mm for all open file descriptors and use this mm rather than
current->mm for memory operations for the lifetime of that fd. Note: this
patch leaves 1 use of current->mm in place. This use is removed in a
follow on patch because other functional changes were required prior to
that use being removed.
If registration fails, there is no reason to keep the handler object
around. Free the handler object rather than add it to the list to
prevent any mmu_notifier operations, including unregister, when
registration fails.
Suggested-by: Jim Foraker <foraker1@llnl.gov>
Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The user SDMA in-use claim bit is in the structure that gets zeroed out
once the claim is made. Move the request in-use flag into its own bit
array and use that for atomic claims. This cleans up the claim code and
removes any race possibility.
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
If unable to insert node into the RB tree cache, node will be freed
before returning from the function. Null out iovec's pointer to node
so iovec does not try to free it later.
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
If a context has not been assigned or assignment failed, pq may be NULL.
Move the unregister within the protection of the null check.
Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The driver pads non-double word multiple message sizes but it doesn't
account for this padding when the packet length is calculated. Also, the
data length is miscalculated for message sizes less than 4 bytes due to
the bit representation in LRH. And there's a check for non-double word
multiple message sizes that prevents these messages from being sent.
This patch fixes length miscalculations and enables the functionality to
send non-double word multiple message sizes.
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Currently each user context is assigned a single SDMA engine
based on the VL, context id, and subcontext id. That means for
MPI applications, each rank can only use one SDMA engine for
all messages. This may create unwanted backup for independent
messages going to different destinations upon congestion at one
destination.
This patch adds the packet "dlid" to the formula of SDMA engine
selection for user SDMA requests. A simple hash table is used
to maintain even distribution among the available SDMA engines
regardless how the "dlid" values are distributed.
Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Tadeusz Struk <tadeusz.struk@intel.com>
Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
