talos-obmc-linux/drivers/fsi, branch dev-4.13

fsi/fsi-master-gpio: Delay sampling of FSI data input

2018-05-17T02:26:51+00:00

Most SoC GPIO implementations, including the Aspeed one, have synchronizers on the GPIO inputs. This means that the value read from a GPIO is a couple of clocks old, from whatever clock source feeds those synchronizers. In practice, this means that in no-delay mode, we are using a value that can potentially be a bit too old and too close to the clock edge establishing the data on the other side of the link. The voltage converters we use on some systems make this worse and sensitive to things like voltage fluctuations etc... This is, we believe, the cause of occasional CRC errors encountered during heavy activity on the LPC bus. This is fixed by introducing a dummy GPIO read before the actual data read. It slows down SBEFIFO by about 15% (less than any delay primitive) and the end result is so far solid. Signed-off-by: Benjamin Herrenschmidt Reviewed-by: Christopher Bostic Signed-off-by: Joel Stanley

fsi/fsi-master-gpio: Reduce dpoll clocks

2018-05-11T10:49:14+00:00

FSI_GPIO_DPOLL_CLOCKS is the number of clocks before sending a DPOLL command after receiving a BUSY status. It should be at least tSendDelay (16 clocks). According to comments in the code, it needs to also be at least 21 clocks due to HW issues. It's currently 100 clocks which impacts performances negatively in some cases. Reduces it in half to 50 clocks which seems to still be solid. Signed-off-by: Benjamin Herrenschmidt Reviewed-by: Christopher Bostic Signed-off-by: Joel Stanley

fsi/fsi-master-gpio: Reduce turnaround clocks

2018-05-11T10:48:18+00:00

FSI_GPIO_PRIME_SLAVE_CLOCKS is the number of clocks if the "idle" phase between the end of a response and the beginning of the next one. It corresponds to tSendDelay in the FSI specification. The default value in the slave is 16 clocks. 100 is way overkill and significantly reduces the driver performance. This changes it to 20 (which gives the HW a bit of margin still just in case). Signed-off-by: Benjamin Herrenschmidt Reviewed-by: Christopher Bostic Signed-off-by: Joel Stanley

fsi/fsi-master-gpio: Add "no-gpio-delays" option

2018-05-11T10:48:13+00:00

This adds support for an optional device-tree property that makes the driver skip all the delays around clocking the GPIOs and set it in the device-tree of common POWER9 based OpenPower platforms. This useful on chips like the AST2500 where the GPIO block is running at a fairly low clock frequency (25Mhz typically). In this case, the delays are unnecessary and due to the low precision of the timers, actually quite harmful in terms of performance. Signed-off-by: Benjamin Herrenschmidt Reviewed-by: Christopher Bostic Signed-off-by: Joel Stanley

fsi/fsi-master-gpio: Sample input data on different clock phase

2018-05-11T10:48:09+00:00

We currently sample the input data right after we toggle the clock low, then high. The slave establishes the data on the rising edge, so this is not ideal. We should sample it on the low phase instead. This currently works because we have an extra delay, but subsequent patches will remove it. Signed-off-by: Benjamin Herrenschmidt Reviewed-by: Christopher Bostic Signed-off-by: Joel Stanley

fsi: occ: Add check for OCC response checksum

2018-05-07T11:34:09+00:00

The OCC specification indicates that it is an error scenario if the response checksum doesn't match the sum of the bytes of the response. The driver needs to perform this calculation and check, and return an error if it's a mismatch. OpenBMC-Staging-Count: 1 Signed-off-by: Eddie James Acked-by: Andrew Jeffery Signed-off-by: Joel Stanley

fsi: occ: Use PutOCCSRAM in circular mode for OCC attention

2018-05-01T03:28:43+00:00

We can't do SCOMs from the BMC when the host is booted in secure mode, and without triggering the attention of the OCC our commands won't be processed. Further, the SCOM operation fails, which causes the OCC driver's probe to fail, which leads the BMC to think that the OCC has failed. We have an alternative to the SCOM though: We can trigger the OCC by writing the correct attention magic in circular buffer mode. The PutOCCSRAM operation returns the written data length in the response payload before the 0xCODE word, so the condition testing the success of the operation is rearranged. OpenBMC-Staging-Count: 1 Signed-off-by: Andrew Jeffery Reviewed-by: Matt Spinler mspinler@linux.vnet.ibm.com Reviewed-by: Eddie James Signed-off-by: Joel Stanley

fsi: gpio: Use a mutex to protect transfers

2018-03-07T04:46:58+00:00

Reduce time spent with interrupts disabled by limiting the critical sections to bitbanging FSI symbols. We only need to ensure exclusive use of the bus for an entire transfer, not that the transfer be performed in atomic context. OpenBMC-Staging-Count: 1 Signed-off-by: Jeremy Kerr Signed-off-by: Andrew Jeffery Signed-off-by: Joel Stanley

fsi: gpio: Remove unused 'id' variable

2018-03-07T04:46:57+00:00

OpenBMC-Staging-Count: 1 Signed-off-by: Andrew Jeffery Signed-off-by: Joel Stanley

fsi: sbefifo: Switch list_lock from spinlock to mutex

2018-03-07T04:46:57+00:00

We allocate GFP_KERNEL memory for a list element in sbefifo_enq_xfr() (elided from the stack trace presumably by inlining) but do so under list_lock, which was previously a spin lock. Enabling lock debugging gives the following warning: BUG: sleeping function called from invalid context at mm/slab.h:408 in_atomic(): 1, irqs_disabled(): 128, pid: 110, name: kworker/u2:9 INFO: lockdep is turned off. irq event stamp: 9002 hardirqs last enabled at (9001): [<801105bc>] kmem_cache_alloc_trace+0x8c/0x284 hardirqs last disabled at (9002): [<8049458c>] _raw_spin_lock_irqsave+0x2c/0xa0 softirqs last enabled at (5370): [<8000987c>] __do_softirq+0x38c/0x510 softirqs last disabled at (5363): [<8001f154>] irq_exit+0x120/0x16c CPU: 0 PID: 110 Comm: kworker/u2:9 Tainted: G W 4.10.17-00534-gb846201e7a2d #2277 Hardware name: ASpeed SoC Workqueue: occ occ_worker [<800108f8>] (unwind_backtrace) from [<8000e02c>] (show_stack+0x20/0x24) [<8000e02c>] (show_stack) from [<8022d1c0>] (dump_stack+0x20/0x28) [<8022d1c0>] (dump_stack) from [<80048bf8>] (___might_sleep+0x174/0x2ac) [<80048bf8>] (___might_sleep) from [<80048da0>] (__might_sleep+0x70/0xb0) [<80048da0>] (__might_sleep) from [<801106f8>] (kmem_cache_alloc_trace+0x1c8/0x284) [<801106f8>] (kmem_cache_alloc_trace) from [<803330e8>] (sbefifo_write_common+0x228/0x4b4) [<803330e8>] (sbefifo_write_common) from [<80333740>] (sbefifo_drv_write+0x24/0x28) [<80333740>] (sbefifo_drv_write) from [<80333b58>] (occ_write_sbefifo+0x44/0x60) [<80333b58>] (occ_write_sbefifo) from [<80334530>] (occ_worker+0x38/0x4a0) [<80334530>] (occ_worker) from [<80034a2c>] (process_one_work+0x23c/0x6d4) [<80034a2c>] (process_one_work) from [<80034f08>] (worker_thread+0x44/0x528) [<80034f08>] (worker_thread) from [<8003c8c0>] (kthread+0x160/0x17c) [<8003c8c0>] (kthread) from [<8000a888>] (ret_from_fork+0x14/0x2c) Avoid the warning (or using GFP_ATOMIC memory) by switching to a mutex, as we only require mutual exclusion, not pre-emption be disabled (occ_worker() is executed on a workqueue). This should also reduce the latency impact of calling into the OCC. OpenBMC-Staging-Count: 1 Signed-off-by: Andrew Jeffery Acked-by: Eddie James Signed-off-by: Joel Stanley