This patch adds context relative numa affinity conditioned on the
module parameter numa_aware. The qib_ctxtdata has an additional
node_id member and qib_create_ctxtdata() has an addition node_id
parameter.
The allocations within the hdr queue and eager queue setup routines
now take this additional member and adjust allocations as necesary.
PSM will pass the either current numa node or the node closest to the
HCA depending on numa_aware. Verbs will always use the node closest to
the HCA.
Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com>
Signed-off-by: Vinit Agnihotri <vinit.abhay.agnihotri@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
This patch adds DCA cache warming for systems that support DCA.
The code uses cpu affinity notification to react to an affinity change
from a user mode program like irqbalance and (re-)program the chip
accordingly. This notification avoids reading the current cpu on every
interrupt.
Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
[ Add Kconfig dependency on SMP && GENERIC_HARDIRQS to avoid failure to
build due to undefined struct irq_affinity_notify. - Roland ]
Signed-off-by: Roland Dreier <roland@purestorage.com>
Covers the rest of the uses of pci error handler.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Commit 36a8f01cd2 ("IB/qib: Add congestion control agent
implementation") tries to store the value 1984 in a u8, which leads to
truncation. Fix this by making the member big enough.
This bug was detected by a smatch warning.
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Add a congestion control agent in the driver that handles gets and
sets from the congestion control manager in the fabric for the
Performance Scale Messaging (PSM) library.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Profiling has shown that sdma_lock is proving a bottleneck for
performance. The situations include:
- RDMA reads when krcvqs > 1
- post sends from multiple threads
For RDMA read the current global qib_wq mechanism runs on all CPUs
and contends for the sdma_lock when multiple RMDA read requests are
fielded on differenct CPUs. For post sends, the direct call to
qib_do_send() from multiple threads causes the contention.
Since the sdma mechanism is per port, this fix converts the existing
workqueue to a per port single thread workqueue to reduce the lock
contention in the RDMA read case, and for any other case where the QP
is scheduled via the workqueue mechanism from more than 1 CPU.
For the post send case, This patch modifies the post send code to test
for a non empty sdma engine. If the sdma is not idle the (now single
thread) workqueue will be used to trigger the send engine instead of
the direct call to qib_do_send().
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
This patch reorganizes the QP and devdata files to be more cache line aware.
qib_qp fields in particular are split into read-mostly, send, and receive fields.
qib_devdata fields are split into read-mostly and read/write fields
Testing has show that bidirectional tests improve by as much as 100%
with this patch.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
This patch optimizes pio buffer allocation in the kernel.
For qib, kernel pio buffers are used for sending acks. The code to
allocate the buffer would always start at 0 until it found a buffer.
This means that an average of 64 comparisions were done on each
allocate, since the busy bit won't be cleared until the bits are
refreshed when buffers are exhausted.
This patch adds two new fields in the devdata struct, last_pio and
min_kernel_pio. last_pio is the last buffer that was allocated.
min_kernel_pio is the lowest potential available buffer.
min_kernel_pio is modifed as contexts are allocated and deallocted.
Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Call irq_set_affinity_hint() to give userspace programs such as
irqbalance the information to be able to distribute qib interrupts
appropriately.
The logic allocates all non-receive interrupts to the first CPU local
to the HCA. Receive interrupts are allocated round robin starting
with the second CPU local to the HCA with potential wrap back to the
second CPU.
This patch also adds a refinement to the name registered for MSI-X
interrupts so that user level scripts can determine the device
associated with the IRQs when there are multiple HCAs with a
potentially different set of local CPUs.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The heavy weight spinlock in qib_lookup_qpn() is replaced with RCU.
The hash list itself is now accessed via jhash functions instead of mod.
The changes should benefit multiple receive contexts in different
processors by not contending for the lock just to read the hash
structures.
The patch also adds a lookaside_qp (pointer) and a lookaside_qpn in
the context. The interrupt handler will test the current packet's qpn
against lookaside_qpn if the lookaside_qp pointer is non-NULL. The
pointer is NULL'ed when the interrupt handler exits.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The context init now saves a shift from rcvegrbufs_perchunk
rcvegrbufs_perchunk_shift using ilog2. A BUG_ON() protects the
power of 2 assumption.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The code that was recently introduced to report the number
of free contexts is flawed for multiple HCAs:
/* Return the number of free user ports (contexts) available. */
return scnprintf(buf, PAGE_SIZE, "%u\n", dd->cfgctxts -
dd->first_user_ctxt - (u32)qib_stats.sps_ctxts);
The qib_stats is global to the module, not per HCA, so the code is broken
for multiple HCAs.
This patch adds a qib_devdata field, freectxts, that reflects the free
contexts for this HCA.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Reviewed-by: Ram Vepa <ram.vepa@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
With ib_qib options:
options ib_qib krcvqs=1 pcie_caps=0x51 rcvhdrcnt=4096 singleport=1 ibmtu=4
a run of ib_write_bw -a yields the following:
------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec]
1048576 5000 2910.64 229.80
------------------------------------------------------------------
The top cpu use in a profile is:
CPU: Intel Architectural Perfmon, speed 2400.15 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit mask
of 0x00 (No unit mask) count 1002300
Counted LLC_MISSES events (Last level cache demand requests from this core that
missed the LLC) with a unit mask of 0x41 (No unit mask) count 10000
samples % samples % app name symbol name
15237 29.2642 964 17.1195 ib_qib.ko qib_7322intr
12320 23.6618 1040 18.4692 ib_qib.ko handle_7322_errors
4106 7.8860 0 0 vmlinux vsnprintf
Analysis of the stats, profile, the code, and the annotated profile indicate:
- All of the overflow interrupts (one per packet overflow) are
serviced on CPU0 with no mitigation on the frequency.
- All of the receive interrupts are being serviced by CPU0. (That is
the way truescale.cmds statically allocates the kctx IRQs to CPU)
- The code is spending all of its time servicing QIB_I_C_ERROR
RcvEgrFullErr interrupts on CPU0, starving the packet receive
processing.
- The decode_err routine is very inefficient, using a printf variant
to format a "%s" and continues to loop when the errs mask has been
cleared.
- Both qib_7322intr and handle_7322_errors read pci registers, which
is very inefficient.
The fix does the following:
- Adds a tasklet to service QIB_I_C_ERROR
- Replaces the very inefficient scnprintf() with a memcpy(). A field
is added to qib_hwerror_msgs to save the sizeof("string") at
compile time so that a strlen is not needed during err_decode().
- The most frequent errors (Overflows) are serviced first to exit the
loop as early as possible.
- The loop now exits as soon as the errs mask is clear rather than
fruitlessly looping through the msp array.
With this fix the performance changes to:
------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec]
1048576 5000 2990.64 2941.35
------------------------------------------------------------------
During testing of the error handling overflow patch, it was determined
that some CPU's were slower when servicing both overflow and receive
interrupts on CPU0 with different MSI interrupt vectors.
This patch adds an option (krcvq01_no_msi) to not use a dedicated MSI
interrupt for kctx's < 2 and to service them on the default interrupt.
For some CPUs, the cost of the interrupt enter/exit is more costly
than then the additional PCI read in the default handler.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
For SusieQ we need to write to the interrupt timer register before
updating the header queue head with interrupt count. This is to
ensure that the timer is enabled properly and a receive available
interrupt is delivered. Otherwise this interrupt can be lost if the
receiver header/eager queues are full before the timer is enabled.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Noticed this odd looking thing in dmesg:
ib_qib 0000:02:00.0: <3>ib_qib: Unable to enable pcie error reporting: -5
which is due to a bad use of dev_info.
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Acked-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix build failure on sparc64 which is missing the include of
<linux/slab.h> via <asm/pci.h> that x86, powerpc, ia64, etc. have.
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Rather than use a variable size array allocation on the stack,
define a constant for the maximum array size possible.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Don't set write combining via PAT on the VL15 buffers to avoid a rare
problem with unaligned writes from interrupt-flushed store buffers.
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add a low-level IB driver for QLogic PCIe adapters.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
