The hfi1_<set|clear>_ctxt_<j|p>key functions take a context index and
look up the context based on that index.
Since the context index is being retrieved from the context, this
doesn't seem optimal.
Pass the context pointer for use, rather than the context index.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@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 array index for the rcd array is sized several different ways
throughout the code.
Use the user interface size (u16) as the standard size and update the
necessary code to reflect this.
u16 is large enough for the largest amount of supported contexts.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@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>
In the error path for context allocation, the file descriptor pointer
should not point to a context when an error occurs.
Clean up the appropriate references on error.
Fixes: Commit 62239fc6e5 ("IB/hfi1: Clean up on context initialization failure")
Reviewed-by: Ira Weiny <ira.weiny@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>
When the debugpat kernel boot flag is turned on the following
traces are printed:
[ 1884.793168] x86/PAT: Overlap at 0x90000000-0x92000000
[ 1884.803510] x86/PAT: reserve_memtype added [mem 0x91200000-0x9127ffff],
track uncached-minus, req write-combining, ret uncached-minus
[ 1884.818167] hfi1 0000:05:00.0: hfi1_0: WC Remapped RcvArray:
ffffc9000a980000
The ioremap_wc() clearly is not returning a write combining mapping due
to an overlap where the RcvArray is mapped in a uncached mapping prior
to creating the proposed write combining mapping.
The patch replaces the single base register for uncached CSRs that
used to overlap the RcvArray with two mappings. One, kregbase1, from the
bar0 up to the RcvArray and another, kregbase2, from the end of the
RcvArray to the pio send buffer space. A new dd field, base2_start,
is used to convert the zero-based offset in the CSR routines to the
correct kregbase1/kregbase2 mapping. A single direct write of the
RcvArray CSRs is replaced with hfi1_put_tid() to insure correct access
using the new disjoint mapping.
Additionally, the kregend field is deleted since it is only ever written.
patdebug now shows the RcvArray as write combining:
[ 35.688990] x86/PAT: reserve_memtype added [mem 0x91200000-0x9127ffff],
track write-combining, req write-combining, ret write-combining
To insulate from any potential issues with write combining, all
writeq are now flushed in hfi1_put_tid() and rcv_array_wc_fill().
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@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>
The caller to the driver marks GFP_NOIO allocations with help
of memalloc_noio-* calls now. This makes redundant to pass down
to the driver gfp flags, which can be GFP_KERNEL only.
The patch removes the gfp flags argument and updates all driver paths.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Acked-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
As the code stands today the array access in remap_intr() is OK. To
future proof the code though we should explicitly check to ensure the
index value is not outside of the valid range. This is not a straight
forward calculation so err on the side of caution.
Reviewed-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>
IPOIB is calling free_rdma_netdev even though alloc_rdma_netdev has
returned -EOPNOTSUPP.
Move free_rdma_netdev from ib_device structure to rdma_netdev structure
thus ensuring proper cleanup function is called for the rdma net device.
Fix the following trace:
ib0: Failed to modify QP to ERROR state
BUG: unable to handle kernel paging request at 0000000000001d20
IP: hfi1_vnic_free_rn+0x26/0xb0 [hfi1]
Call Trace:
ipoib_remove_one+0xbe/0x160 [ib_ipoib]
ib_unregister_device+0xd0/0x170 [ib_core]
rvt_unregister_device+0x29/0x90 [rdmavt]
hfi1_unregister_ib_device+0x1a/0x100 [hfi1]
remove_one+0x4b/0x220 [hfi1]
pci_device_remove+0x39/0xc0
device_release_driver_internal+0x141/0x200
driver_detach+0x3f/0x80
bus_remove_driver+0x55/0xd0
driver_unregister+0x2c/0x50
pci_unregister_driver+0x2a/0xa0
hfi1_mod_cleanup+0x10/0xf65 [hfi1]
SyS_delete_module+0x171/0x250
do_syscall_64+0x67/0x150
entry_SYSCALL64_slow_path+0x25/0x25
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Driver does not check whether proper configuration file exist in EPROM,
and treats empty partition as possible valid configuration, preventing
fallback to default firmware. Change EPROM read function to treat
missing magic number as read error.
Reviewed-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Base receive contexts can be used by sub contexts. Because of this,
resources for the context cannot be completely freed until all sub
contexts are done using the base context.
Introduce a reference count so that the base receive context can be
freed only when all sub contexts are done with it.
Use the provided function call for setting default send context
integrity rather than the manual method.
The cleanup path does not set all variables back to NULL after freeing
resources. Since the clean up code can get called more than once,
(e.g. during context close and on the error path), it is necessary to
make sure that all the variables are NULLed.
Possible crash are:
BUG: unable to handle kernel paging request at 0000000001908900
IP: read_csr+0x24/0x30 [hfi1]
RIP: 0010:read_csr+0x24/0x30 [hfi1]
Call Trace:
sc_disable+0x40/0x110 [hfi1]
hfi1_file_close+0x16f/0x360 [hfi1]
__fput+0xe7/0x210
____fput+0xe/0x10
or
kernel BUG at mm/slub.c:3877!
RIP: 0010:kfree+0x14f/0x170
Call Trace:
hfi1_free_ctxtdata+0x19a/0x2b0 [hfi1]
? hfi1_user_exp_rcv_grp_free+0x73/0x80 [hfi1]
hfi1_file_close+0x20f/0x360 [hfi1]
__fput+0xe7/0x210
____fput+0xe/0x10
Fixes: Commit 62239fc6e5 ("IB/hfi1: Clean up on context initialization failure")
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
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>
The expected receive lists (tid_xxx_list) are not initialized until
late in the receive context initialization. If an error happens
before the initialization, a NULL pointer access will occur during
cleanup.
Initialized the lists sooner rather than later to avoid this Oops:
IP: unlock_exp_tids.isra.11+0x26/0xd0 [hfi1]
RIP: 0010:unlock_exp_tids.isra.11+0x26/0xd0 [hfi1]
Call Trace:
hfi1_user_exp_rcv_free+0x79/0xb0 [hfi1]
hfi1_file_close+0x87/0x360 [hfi1]
__fput+0xe7/0x210
____fput+0xe/0x10
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
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>
This patch adds a trace for putting a TID and
for writing the RcvArray CSR.
The CSR access template can be easily extended for additional
CSR readq/writeq calls.
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@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>
Currently, platform configuration can be read from EFI variable
for discrete cards. It will happen when reading from EPROM fails.
EFI variables should not be queried for platform configuration
in any scenario.
Reviewed-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
According to IBTA spec a QKey violation should not result in a bad qkey
trap being triggered for UD queue pairs. Also since it is a silent error
we do not increment the q_key violation or the dropped packet counters.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
In process_receive_packet the packet header field is used to calculate
the length of the packet. However this is not necessarily setup. In fact
only if the ECN prescan is enabled will the packet header be valid at
this point.
The code works as is because we do not do anything with the packet
length at this point in the packet processing. The length and header are
setup correctly in hfi1_setup_ib_header which is called by the following
sequence:
process_receive_packet()
-> rhf_receieve_function_map[]()
--> process_receive_ib()
---> hfi1_setup_9B_packet()
----> hfi1_setup_ib_header()
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
As the code stands today the array access in remap_intr() is OK. To
future proof the code though we should explicitly check to ensure the
index value is not outside of the valid range. This is not a straight
forward calculation so err on the side of caution.
Reviewed-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>
sdma_init does not take a number of sdma engine parameters,
rather it initializes all of the sdma engines.
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Reclassify messages printed out to /var/log/messages into
warnings and errors to facilitate debugging in the future
for issues related to the platform config logic.
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>
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>
SGEs that are contiguous needlessly consume driver dependent TX resources.
The lkey validation logic is enhanced to compress the SGE that ends
up in the send wqe when consecutive addresses are detected.
The lkey validation API used to return 1 (success) or 0 (fail).
The return value is now an -errno, 0 (compressed), or 1 (uncompressed). A
additional argument is added to pass the last SQE for the compression.
Loopback callers always pass a NULL to last_sge since the optimization is
of little benefit in that situation.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Brian Welty <brian.welty@intel.com>
Signed-off-by: Venkata Sandeep Dhanalakota <venkata.s.dhanalakota@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>
Calls to trace incoming packets will now receive the packet
context as parameter. This enables trace support for future
packet types.
Header trace output is in the format <field>:<value>
which makes parsing easier.
input_ibhdr trace before change:
<idle>-0 [001] d.h. 5904.250925: input_ibhdr: [0000:05:00.0] vl 0
lver 0 sl 0 lnh 2,LRH_BTH dlid 0002 len 18 slid 0001 op
0x64,UD_SEND_ONLY se 0 m 0 pad 0 tver 0 pkey 0xffff f 0 b 0 qpn 0x000001
a 0 psn 0x000001b2 deth qkey 0x80010000 sqpn 0x000001
input_ibhdr trace after change:
<idle>-0 [001] d.h. 6655.714488: input_ibhdr: [0000:05:00.0] (IB)
len:124 sc:0 dlid:0x0001 slid:0x0002 lnh:2,LRH_BTH lver:0 sl:0 age:0
becn:0 fecn:0 l4:0 rc:0 entropy:0 op:0x64,UD_SEND_ONLY se:0 m:0 pad:0
tver:0 pkey:0x7fff f:0 b:0 qpn:0x000001 a:0 psn:0x00000036 hlen:8 deth
qkey:0x80010000 sqpn:0x000001
Reviewed-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@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 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 handling of IB_RDMA_WRITE_ONLY_WITH_IMMEDIATE will leak a memory
reference when a buffer cannot be allocated for returning the immediate
data.
The issue is that the rkey validation has already occurred and the RNR
nak fails to release the reference that was fruitlessly gotten. The
the peer will send the identical single packet request when its RNR
timer pops.
The fix is to release the held reference prior to the rnr nak exit.
This is the only sequence the requires both rkey validation and the
buffer allocation on the same packet.
Cc: Stable <stable@vger.kernel.org> # 4.7+
Tested-by: Tadeusz Struk <tadeusz.struk@intel.com>
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>
Keep VL15 credits at 0 during LNI, before link-up. Store
VL15 credits value during verify cap interrupt and set
in after link-up. This addresses an issue where VL15 MAD
packets could be sent by one side of the link before
the other side is ready to receive them.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The Omni-Path adapter driver fails to load on the ppc64le platform
due to invalid PCI setup.
This patch makes the PCI configuration more robust and will
fix 64 bit addressing for ppc64le.
Signed-off-by: Steven L Roberts <robers97@gmail.com>
Acked-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Pull more rdma updates from Doug Ledford:
"As mentioned in my first pull request, this is the subsequent pull
requests I had. This is all I have, and in fact this cleans out the
RDMA subsystem's entire patchworks queue of kernel changes that are
ready to go (well, it did for the weekend anyway, a few new patches
are in, but they'll be coming during the -rc cycle).
The first tag contains a single patch that would have conflicted if
taken from my tree or DaveM's tree as it needed our trees merged to
come cleanly.
The second tag contains the patch series from Intel plus three other
stragllers that came in late last week. I took them because it allowed
me to legitimately claim that the RDMA patchworks queue was, for a
short time, 100% cleared of all waiting kernel patches, woohoo! :-).
I have it under my for-next tag, so it did get 0day and linux- next
over the end of last week, and linux-next did show one minor conflict.
Summary:
'for-linus' tag:
- mlx5/IPoIB fixup patch
'for-next' tag:
- the hfi1 15 patch set that landed late
- IPoIB get_link_ksettings which landed late because I asked for a
respin
- one late rxe change
- one -rc worthy fix that's in early"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dledford/rdma:
IB/mlx5: Enable IPoIB acceleration
* tag 'for-next' of git://git.kernel.org/pub/scm/linux/kernel/git/dledford/rdma:
rxe: expose num_possible_cpus() cnum_comp_vectors
IB/rxe: Update caller's CRC for RXE_MEM_TYPE_DMA memory type
IB/hfi1: Clean up on context initialization failure
IB/hfi1: Fix an assign/ordering issue with shared context IDs
IB/hfi1: Clean up context initialization
IB/hfi1: Correctly clear the pkey
IB/hfi1: Search shared contexts on the opened device, not all devices
IB/hfi1: Remove atomic operations for SDMA_REQ_HAVE_AHG bit
IB/hfi1: Use filedata rather than filepointer
IB/hfi1: Name function prototype parameters
IB/hfi1: Fix a subcontext memory leak
IB/hfi1: Return an error on memory allocation failure
IB/hfi1: Adjust default eager_buffer_size to 8MB
IB/hfi1: Get rid of divide when setting the tx request header
IB/hfi1: Fix yield logic in send engine
IB/hfi1, IB/rdmavt: Move r_adefered to r_lock cache line
IB/hfi1: Fix checks for Offline transient state
IB/ipoib: add get_link_ksettings in ethtool
