- 11 Mar, 2016 26 commits
-
-
Dennis Dalessandro authored
This patch adds the MR datastructures based on hfi1 into rvt. For now the data structures are defined in include/rdma/rdma_vt.h but once all MR functionality has been moved from the drivers into rvt these should move to rdmavt/mr.h Reviewed-by:
Ira Weiny <ira.weiny@intel.com> Reviewed-by:
Dean Luick <dean.luick@intel.com> 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>
-
Dennis Dalessandro authored
This adds the post sned and recv function stubs. Reviewed-by:
-
Dennis Dalessandro authored
Create stubs for completion queue creation, polling, resizing, calling for notification, and destroying. Reviewed-by:
-
Dennis Dalessandro authored
This adds the get port immutable verbs call. Reviewed-by:
-
Dennis Dalessandro authored
Adds the stub for the mmap verbs call. Reviewed-by:
-
Dennis Dalessandro authored
This adds the stub for process mad. More study is needed to determine the final MAD interaction between the driver and rvt. Reviewed-by:
-
Dennis Dalessandro authored
Adds the function stubs for attach and detach multicast. Reviewed-by:
-
Dennis Dalessandro authored
Adds the stubs for create, modify, query, and destory for shared request queues. Reviewed-by:
-
Dennis Dalessandro authored
Adds the function stubs for allocating, and registering memory regions, as well as deregistering them. Reviewed-by:
-
Dennis Dalessandro authored
Adds the stubs for create, destroy, modify, and query of the address handle. Reviewed-by:
-
Dennis Dalessandro authored
Adds the stubs for create, modify, destroy and query functions for queue pairs. Reviewed-by:
-
Dennis Dalessandro authored
Add the stubs to allocate and deallocate user contexts. This will be handled completely by rvt. Reviewed-by:
-
Dennis Dalessandro authored
The handler for query gid operates along the same lines as the query pkey handler. The driver will take care to keep the guid table updated. Reviewed-by:
-
Dennis Dalessandro authored
The pkey table will reside in the rvt structure but it will be modified only when the driver requests then rvt will simply read the value to return in the query. Reviewed-by:
-
Dennis Dalessandro authored
This patch adds the query and modify port stubs. The query will mostly entail the driver returning everything in the ib_port_attr which will get handed back to the verbs layer. The modify will need some API helpers in the driver. The send_trap and post_mad_send are still issues to address. Reviewed-by:
-
Dennis Dalessandro authored
Adds the stubs which will handle the query and modify device functions. At this time the only intention is to support changing the node desc and the guid via these calls. Reviewed-by:
-
Dennis Dalessandro authored
Add a macro to cut down on writing the same lines over and over again for what will be a large number of functions that will be supported. Reviewed-by:
-
Dennis Dalessandro authored
Instead of trying to handle each parameter separately, add ib_device_attr to rvt_driver_params. This means drivers will fill this in and pass to the rvt registration function. Reviewed-by:
-
Dennis Dalessandro authored
Add datastructure for and allocation/deallocation of protection domains for RDMAVT. Reviewed-by:
-
Dennis Dalessandro authored
This patch adds dma functions to rdmavt. The source is hfi1's version of dma.c which will be removed by a subsequent hfi1 patch. Reviewed-by:
-
Dennis Dalessandro authored
This patch introduces the basics for a new module called rdma_vt. This new driver is a software implementation of the InfiniBand verbs and aims to replace the multiple implementations that exist and duplicate each others' code. While the call to actually register the device with the IB core happens in rdma_vt, most of the work is still done in the drivers themselves. This will be changing in a follow on patch this is just laying the groundwork for this infrastructure. Reviewed-by:
-
Jim Snow authored
The link state will transition from ARMED to ACTIVE when a non-SC15 packet arrives, but the driver might not notice the change. With this fix, if the slowpath receive interrupt handler sees a non-SC15 packet while in the ARMED state, we queue work to call linkstate_active_work from process context to promote it to ACTIVE. Reviewed-by:
Jim Snow <jim.m.snow@intel.com> Signed-off-by:
Brendan Cunningham <brendan.cunningham@intel.com> Signed-off-by:
Jubin John <jubin.john@intel.com> Signed-off-by:
-
Ira Weiny authored
Total XMIT wait needs to sum the xmit wait values of all the VLs not just those requested in the query. Also, make the algorithm used for both PortStatus and PortDataCounters the same. Reviewed-by:
Arthur Kepner <arthur.kepner@intel.com> Reviewed-by:
Breyer, Scott J <scott.j.breyer@intel.com> Signed-off-by:
Ira Weiny <iweiny@gmail.com> Signed-off-by:
-
Edward Mascarenhas authored
Clean up comments by deleting numbering and terms internal to Intel. The information on the actual bugs is not deleted. Reviewed-by:
Edward Mascarenhas <edward.mascarenhas@intel.com> Signed-off-by:
-
Dean Luick authored
The variable "index" increments the same as dd->ndevcntrs. Just use the later. Remove uneeded usage of "index" in the fill loop - it is not used there or later in the function. Reviewed-by:
-
Vennila Megavannan authored
Added the following per sdma engine stats: - SendDmaDescFetchedCnt - software maintained count of SDMA interrupts (SDmaInt, SDmaIdleInt, SDmaProgressInt) - software maintained counts of SDMA error cases Reviewed-by:
Vennila Megavannan <vennila.megavannan@intel.com> Signed-off-by:
-
- 01 Mar, 2016 6 commits
-
-
jubin.john@intel.com authored
Change the default number of krcvqs to number of numa nodes + 1 based on the performance data collected. Reviewed-by:
-
Mark F. Brown authored
The krcvqs parameter is displayed incorrectly in sysfs. The workaround is to set the param type as uint. Reviewed-by:
Mitko Haralanov <mitko.haralanov@intel.com> Signed-off-by:
Mark F. Brown <mark.f.brown@intel.com> Signed-off-by:
-
Harish Chegondi authored
This would reduce L2 cache misses on s_sde in the _hfi1_schedule_send function when invoked from post_send thereby improving performance of post_send. Reviewed-by:
Harish Chegondi <harish.chegondi@intel.com> Signed-off-by:
-
jubin.john@intel.com authored
This patch fixes the checkpatch issue: CHECK: Prefer using the BIT macro Use of BIT macro for HDRQ_INCREMENT in chip.h causes a change in format specifier for error message in init.c in order to avoid a build warning. Reviewed-by:
-
Dean Luick authored
Correctly set half-swing for integrated devices. A0 needs all fields set for CcePcieCtrl. B0 and later only need a few fields set. Reviewed-by:
Stuart Summers <john.s.summers@intel.com> Signed-off-by:
-
Ira Weiny authored
To be used in future patches add dd_dev_dbg. dd_* functions properly decode the hfi1_devdata structure used throughout the driver Signed-off-by:
-
- 29 Feb, 2016 8 commits
-
-
Mitko Haralanov authored
This commit "flips the switch" on the TID caching feature implemented in this patch series. As well as enabling the new feature by tying the new function with the PSM API, it also cleans up the old unneeded code, data structure members, and variables. Due to difference in operation and information, the tracing functions related to expected receives had to be changed. This patch include these changes. The tracing function changes could not be split into a separate commit without including both tracing variants at the same time. This would have caused other complications and ugliness. Signed-off-by:
-
Mitko Haralanov authored
The previous patch in the series added the free/invalidate function bodies. Now, it's time for the programming side. This large function takes the user's buffer, breaks it up into manageable chunks, allocates enough RcvArray groups and programs the chunks into the RcvArray entries in the hardware. With this function, the TID caching functionality is implemented. However, it is still unused. The switch will come in a later patch in the series, which will remove the old functionality and switch the driver over to TID caching. Signed-off-by:
-
Mitko Haralanov authored
Up to now, the functions which cleared the programmed TID entries and gave PSM the list of invalidated TID entries were just stubs. With this commit, the bodies of these functions are added. This commit is a bit asymmetric as it only contains the free code path. This is done on purpose to help with patch reviews as the programming code path is much longer. Signed-off-by:
-
Mitko Haralanov authored
TID caching will rely on the MMU notifier to be told when memory is being invalidated. When the callback is called, the driver will find all RcvArray entries that span the invalidated buffer and "schedule" them to be freed by the PSM library. This function is currently unused and is being added in preparation for the TID caching feature. Signed-off-by:
-
Mitko Haralanov authored
The upcoming TID caching feature requires different data structures and, by extension, different initialization for each of the MPI processes. The two new functions (currently unused) perform the required initialization and freeing of required resources and structures. Signed-off-by:
-
Mitko Haralanov authored
The exp_lock lock does not need to be a spinlock as all its uses are in process context and allowing the process to sleep when the mutex is contended might be beneficial. Signed-off-by:
-
Mitko Haralanov authored
Functions added by this patch are building blocks for the upcoming TID caching functionality. The functions added are currently unsed (and marked as such.) The functions' purposes are to find physically contigous pages in the user's virtual buffer, program the RcvArray group entries with these physical chunks, and unprogram the RcvArray groups. Reviewed-by:
-
Mitko Haralanov authored
Definitions and functions use to manage sets of TID/RcvArray groups. These will be used by the TID cacheline functionality coming with later patches. TID groups (or RcvArray groups) are groups of TID/RcvArray entries organized in sets of 8 and aligned on cacheline boundaries. The TID/RcvArray entries are managed in this way to make taking advantage of write-combining easier - each group is a entire cacheline. rcv_array_wc_fill() is provided to allow of generating writes to TIDs which are not currently being used in order to cause the flush of the write-combining buffer. Signed-off-by:
-
