- 27 Oct, 2015 40 commits
-
-
Mike Marciniszyn authored
Add additional rc traces to aid in debugging rc retry logic. Reviewed-by:
Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by:
Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by:
Ira Weiny <ira.weiny@intel.com> Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
Jubin John authored
The name used to create the verbs txreq cache was not qualified with the unit number. This causes a panic when destroying the cache on a dual HFI systems. Create a unique name with the unit number with this patch Reviewed-by:
Jubin John <jubin.john@intel.com> Signed-off-by:
-
Easwar Hariharan authored
Using fw_sbus_load to control SBus firmware load doesn't scale across multiple HFI1 cards in a single system. This patch ensures that the SBus firmware is loaded once per ASIC. Reviewed-by:
Dean Luick <dean.luick@intel.com> Reviewed-by:
Easwar Hariharan <easwar.hariharan@intel.com> Signed-off-by:
-
Dean Luick authored
When under heavy load, the receive interrupt handler can run too long with IRQs disabled. Add a mixed-mode threading scheme. Initially process packets in the handler for quick responses (latency). If there are too many packets to process move to a thread to continue (bandwidth). Reviewed-by:
-
Dean Luick authored
In preparation for threading the receive interrupt, add irqsaves in the packet processing path. When the receive interrupt is threaded, the packet processing path is no longer guaranteed to have IRQs disabled. Add irqsaves where needed on several locks in the packet processing path. Anything that did not have an obvious, "close" irqsave in its caller is a candidate. Reviewed-by:
-
Ignacio Hernandez authored
A value of 2048 increased PSM performance while not impacting verbs performance. Reviewed-by:
Mitko Haralanov <mitko.haralanov@intel.com> Signed-off-by:
Ignacio Hernandez <ignacio.hernandez@intel.com> Signed-off-by:
-
Mitko Haralanov authored
The SDMA engines were configured to generate progress interrupts every time they processed N/2 descriptors (where N is the size of the descriptor queue). This interval was too infrequent, leading to degraded performance. This commit adds a module parameter, as well as a recommended default, which allows for the tuning of the interrupt frequency. Reviewed-by:
-
Mitko Haralanov authored
The address being mapped into a process's memory for notification events was improperly calculated due to cast that was happening too early. dd->events is a pointer and should have been casted to unsigned long after the pointer arithmetic was done, not before. As a result, processes were looking at the wrong place and not seeing their notification events. Signed-off-by:
-
Mitko Haralanov authored
User SDMA keeps track of progress into the submitted IO vectors by tracking an offset into the vectors when packets are submitted. This offset is updated after a successful submission of a txreq to the SDMA engine. The same offset was used when determining whether an IO vector should be 'freed' (pages unpinned) in the SDMA callback functions. This was causing a silent data corruption in big jobs (> 2 nodes, 120 ranks each) on the receive side because the send side was mistakenly unpinning the vector pages before the HW has processed all descriptors referencing the vector. Reviewed-by:
-
Easwar Hariharan authored
The DC firmware overwrites the enable_lane_tx register and does not update it on a host request to go to Poll. This causes an infinite loop through the LNI state machine if a link width downgrade occurs. This patch re-sets the enable_lane_tx register to all 4 lanes. Reviewed-by:
-
Dean Luick authored
When under heavy load, the send handler can run too long without allowing other tasks to run. Add a conditional resched to break this up. Reviewed-by:
-
Caz Yokoyama authored
Reset firmware instead of reloading Sbus firmware if it's already done for this ASIC. To work around thermal polling problem in firmware, don't reload Sbus firmware, instead, reset the firmware on the initialization of the second HFI. Reviewed-by:
Caz Yokoyama <caz.yokoyama@intel.com> Signed-off-by:
-
Jareer Abdel-Qader authored
Driver does not verify userid for shared context assignments, allowing malicious user access. Reviewed by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by:
Jareer H Abdel-Qader <jareer.h.abdel-qader@intel.com> Signed-off-by:
-
Niranjana Vishwanathapura authored
This fixes transmit errors when the number of scatter gather elements in the request is more that the number of per packet descriptors supported by the hardware, allocate and coalesce the extra scatter gather elements into a single buffer. The last descriptor is reserved and used for this coalesced buffer. Verbs potentially need this support when transferring small data chunks involving different memory regions. Reviewed-by:
Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com> Signed-off-by:
-
Easwar Hariharan authored
The QSFP interface code has been running without issues and the flag is never set to off. This patch removes the QSFP_ENABLED bit from HFI1_CAP. Reviewed-by:
-
Vennila Megavannan authored
If packets stop egressing the hardware link, software can lock up. Implement a timeout for send context halt recovery. This patch increases the timeout for packet egress to 500 us and timer resets to zero if the packet occupancy changes. Also we bounce the link on time out. Reviewed-by:
Vennila Megavannan <vennila.megavannan@intel.com> Signed-off-by:
-
Dean Luick authored
The latest version of the 8051 firmware will wait longer when bringing the link down. Extend the driver's timeout to go with that. Reviewed-by:
-
Easwar Hariharan authored
The ASIC registers were not reset on FLR, and the code to protect the ASIC block against multiple initializations by peer HFIs did not extend to multiple ASICs in a system. This patch addresses this gap. Reviewed-by:
-
Mike Marciniszyn authored
The clear_ahg call is new in hfi1 vs qib. For small messages the progress routine always builds one and clears out the ahg state when the queue has gone to empty which is the predominant case for small messages. Inline the routine and avoid the call to sdma_ahg_free to mitigate the performance impact. Finally, move the routine to qp.h for scope reasons. Reviewed-by:
-
Muhammad Falak R Wani authored
Use the timer API function setup_timer instead of init_timer, removing the structure field assignments. Signed-off-by:
Muhammad Falak R Wani <falakreyaz@gmail.com> Signed-off-by:
-
Muhammad Falak R Wani authored
Use the timer API function setup_timer instead of init_timer, removing the structure field assignments. <smpl> @timer@ expression e1,e2,e3,fn_ptr; @@ -init_timer(&e1); +setup_timer(&e1, fn_ptr, e2); ... when != fn_ptr = e3 -e1.function = fn_ptr; -e1.data = e2; </smpl> Signed-off-by:
-
Mike Rapoport authored
The {x,y}LCD fields of struct sm750_dev are not used and can be removed. Signed-off-by:Mike Rapoport <mike.rapoport@gmail.com> Signed-off-by:
-
Mike Rapoport authored
There are several occurencies of '#ifdef CAP_EXP<something>', spelled differently each time. None of these is ever defined and therefore they enclose dead code that can be removed. Signed-off-by:
-
Mike Rapoport authored
Both struct lynx_share and struct sm750_dev reprsent some parts of the SM750 graphics adapter. There is no point to keep these parts in different structures. Signed-off-by:
-
Mike Rapoport authored
The lynx_share field in lynx_cursor structure is never used and can be removed. Signed-off-by:
-
Mike Rapoport authored
Fix wrong indentation of a line in lynxfb_set_fbinfo Signed-off-by:
-
Mike Rapoport authored
Use sm750_dev as private driver data for the PCI device Signed-off-by:
-
Mike Rapoport authored
Use proper 'struct sm750_dev *' type instead of 'void *' Signed-off-by:
-
Mike Rapoport authored
Use sm750_dev rather than lynx_share as parameter for hw_sm750_map, hw_sm750_inithw and hw_sm750_initAccel functions. Signed-off-by:
-
Mike Rapoport authored
The hw_sm750_getVMSize wraps ddk750_getVMSize without any added functionality. Call ddk750_getVMSize directly and remove hw_sm750_getVMSize. Signed-off-by:
-
Mike Rapoport authored
Use sm750_dev instead of lynx_share in the private data of the frambuffer device. Signed-off-by:
-
Mike Rapoport authored
The PROC_SPEC* typedefs are not used and can be removed Signed-off-by:
-
Mike Rapoport authored
The struct sm750_state is never used on its own. Merge it into the struct sm750_dev. Signed-off-by:
-
Mike Rapoport authored
Use consistent naming for 'struct sm750_dev' variables Signed-off-by:
-
Mike Rapoport authored
The struct sm750_share represents the device state and name sm750_dev seems more appropriate. Signed-off-by:
-
Mike Rapoport authored
There is no actual need to split type definitions and global function declarations between lots of small headers. Signed-off-by:
-
Mike Rapoport authored
The smi_indent variable is only declared, but never used. Remove it. Signed-off-by:
-
Mike Rapoport authored
The custom macro PADDING differs from standard kernel ALIGN only in parameters order. Replace PADDING with ALIGN and remove it's definition. Signed-off-by:
-
Mike Rapoport authored
The SPC_SM750 define is not used and can be removed Signed-off-by:
-
Mike Rapoport authored
The methods in lynx_cursor always use the same implementation and there is no point to use redirection rather than direct function call. Signed-off-by:
-
