diff options
Diffstat (limited to 'manuals/volta/gv100/dev_fifo.ref.txt')
| -rw-r--r-- | manuals/volta/gv100/dev_fifo.ref.txt | 97 |
1 files changed, 97 insertions, 0 deletions
diff --git a/manuals/volta/gv100/dev_fifo.ref.txt b/manuals/volta/gv100/dev_fifo.ref.txt index 8b590cb..fe0bea2 100644 --- a/manuals/volta/gv100/dev_fifo.ref.txt +++ b/manuals/volta/gv100/dev_fifo.ref.txt @@ -640,6 +640,103 @@ DEALINGS IN THE SOFTWARE. #define NV_PFIFO_PBDMA_STATUS_INST_VALID_FALSE 0x00000000 /* R-E-V */ #define NV_PFIFO_PBDMA_STATUS_INST_VALID_TRUE 0x00000001 /* R---V */ +Channel Teardown Sequence +=============================================================================== + + This section describes the sequence software (specifically RM) can use to +tear down a channel for robust channels (RC) recovery or in the case of a fault. + + In the case of a fault, Host does not guarantee that a PBDMA has saved out +prior to RM receiving notification of the fault. RM must determine which +context has faulted by processing the fault buffer as described in the +NV_PFB_PRI_MMU_FAULT_BUFFER_* register documentation in pri_mmu_hub.ref and in +the fault buffer NV_MMU_FAULT_BUF_ENTRY documentation in dev_mmu_fault.ref. +This context can then be torn down using the following procedure. + Note when a PBDMA fault or CE fault occurs, the PBDMA will save out +automatically. The TSG related to the context in which the fault occurred will +not be scheduled again until the fault is handled. + In the case of some other issue requiring the engine to be reset, the TSG +will need to be manually preempted. + In all cases, a PBDMA interrupt may occur prior to the PBDMA being able to +switch out. SW must handle these interrupts according to the relevant handling +procedure before the PBDMA preempt can complete. + +Context TSG tear-down procedure: + + 1. Disable scheduling for the engine's runlist via NV_PFIFO_SCHED_DISABLE. + This enables SW to determine whether a context has hung later in the + process: otherwise, ongoing work on the runlist may keep ENG_STATUS from + reaching a steady state. + + 2. Disable all channels in the TSG being torn down or submit a new runlist + that does not contain the TSG. This is to prevent the TSG from being + rescheduled once scheduling is re-enabled in step 6. + + 3. Initiate a preempt of the engine by writing the bit associated with its + runlist to NV_PFIFO_RUNLIST_PREEMPT. This allows us to begin the preempt + process prior to doing the slow register reads needed to determine whether + the context has hit any interrupts or is hung. Do not poll + NV_PFIFO_RUNLIST_PREEMPT for the preempt to complete. + + 4. Check for interrupts or hangs while waiting for the preempt to complete. + During the below polling, any stalling interrupts relating to the runlist + must be detected and handled in order for the preemption to complete. SW + may opt to simply reset the engine immediately, or perform the following + sub-steps to more cleanly tear down the context: + + a. Wait for PBDMA preempt completion: For each PBDMA which serves the + runlist, poll NV_PFIFO_PBDMA_STATUS(pbdma_id) to reach CHAN_STATUS + INVALID, indicating the no further work will run on the PBDMA during the + tear-down sequence. Interleaved with the polling, PBDMA interrupts must + be serviced as they arise: such an interrupt can prevent the PBDMA from + completing its channel save. + + b. Wait for engine context preempt completion: For each engine served by the + runlist, read NV_PFIFO_ENGINE_STATUS(engine_id) to verify the channel/TSG + has saved off the engine, or tell if the CTXSW is hung, via the + CTX_STATUS, ID, and NEXT_ID fields. Take action based on the following + values for the CTX_STATUS field: + + i. CTX_STATUS_SWITCH: Engine save hasn't started yet, continue to poll + (repeat step 4b). + + ii. CTX_STATUS_INVALID: The engine context has switched off. The + preemption step for this engine is complete. + + iii. CTX_STATUS_VALID or CTX_STATUS_CTXSW_SAVE: check the ID field: + * If ID matches the TSG for the context being torn down, the engine + reset procedure can be performed (see step 5), or SW can continue + waiting by repeating step 4b. + * If ID does NOT match, then skip engine reset (skip step 5) for this + engine. The context isn't running on the engine. + + iv. CTX_STATUS_LOAD: check the NEXT_ID field: + * If NEXT_ID matches the TSG of the context being torn down, the engine + is loading the context and reset (see step 5) can be performed + immediately or after a delay to allow the context a chance to load and + be saved off. + * If NEXT_ID does not match the TSG ID or CHID then the context is no + longer on the engine. Skip engine reset (skip step 5) for this + engine. + + SW may alternatively wait for the CTX_STATUS to reach INVALID, but this + may take longer if an unrelated context is currently on the engine or + being switched to. + + 5. If a reset is needed as determined by step 4: + + a. Halt the memory interface for the engine (as per the relevant engine + procedure). + + b. Reset the engine via NV_PMC_ENABLE. + + c. Take the engine out of reset and re-init the engine (as per the relevant + engine procedure) + + 6. Re-enable scheduling for the engine's runlist via NV_PFIFO_SCHED_ENABLE. + +After this sequence, resources for the channels in the TSG may be reclaimed. + -------------------------------------------------------------------------------- KEY LEGEND -------------------------------------------------------------------------------- |