[BUG] error trace debugfs entry (etrace) is always empty

[keyonjie]

The error trace debugfs entry (/sys/kernel/debug/sof/etrace) is always empty which lead to we can't get any error trace with sof-logger (sof-logger -l *.ldc), this is bad for debugging when DMA trace doesn't work because of FW hanged or crashed.

To Reproduce
Do anything can lead to FW reporting tr_err() or comp_err(), and check the "sof-logger -l *.ldc".

Reproduction Rate
100%

Expected behavior
We should see the error trace logs in sof-logger output.

Impact
This will make it difficult for FW debugging when FW is hang or crashed.

Environment
Any Intel platforms

[paulstelian97]

Either this is a regression (I have used it on i.MX8 initially before adding the DMA trace support -- heck, before adding the host DMA in the first place) or something is different between Intel and NXP platforms.

Is there any known working commit a few months ago so you can start bisecting?

[plbossart]

We should really have a first message to check that the trace works. An empty trace is not-testable, it can mean the trace is not working or there is no error message. An 'Hello SOF" seems required for CI/daily tests to check the trace functionality, no?

@mmaka1 @lgirdwood @xiulipan @marc-hb FYI

[marc-hb]

@keyonjie does revert #3245 help?

We should really have a first message to check that the trace works. An empty trace is not-testable, it can mean the trace is not working or there is no error message. An 'Hello SOF" seems required for CI/daily tests to check the trace functionality, no?

thesofproject/sof-test#297 thesofproject/sof-test#298

[slawblauciak]

Issue not apparent when testing with the slim driver, maybe a kernel regression?

[dbaluta]

Cc: @IulianOlaru249 @crazoes

I also noticed this last week but I thought it's just me being unable to use sof-logger properly.

[lyakh]

I hardly remember seeing error trace data in "etrace" ever. Looking through both kernel and firmware sources it seems like the kernel expects two memory window regions - SOF_IPC_REGION_DEBUG for debug and SOF_IPC_REGION_TRACE for errors, in the firmware e.g. on CAVS platforms they are initialised by EXT_MAN_PORT() in src/platform/intel/cavs/platform.c - so far so good. But I don't see the DEBUG region being used. Both _trace_error_with_ids() and _trace_event_with_ids() seem to translate to the same code in the end. I also checked one of daily tests on ICL from yesterday, where there is an "ERROR" entry in debug trace from hda_dma_link_check_xrun() where tr_err() is called, but "etrace" is empty. So, it seems like a global error to me.

[lgirdwood]

We should really have a first message to check that the trace works. An empty trace is not-testable, it can mean the trace is not working or there is no error message. An 'Hello SOF" seems required for CI/daily tests to check the trace functionality, no?

@mmaka1 @lgirdwood @xiulipan @marc-hb FYI

Ack - "Hello SOF". @marc-hb do you fancy making this change, it should be a one liner for each platform in the FW ready function to add a trace_error call with the above string..

[xiulipan]

@lgirdwood I am OK with this!
@marc-hb And I think you can also make patches to sof-test scripts to make a valid test for error logger now.

[lyakh]

@xiulipan @akloniex so how is the error trace supposed to work and when did it work last time? I looked at v1.0 and v1.1 and even there I couldn't so far find that out.

[xiulipan]

@lyakh To understand how the trace work, we need to know we have 2 kinds of trace:

1. error trace: a simple shared memory window between DSP and host (same as the IPC mailbox), which means this kind of trace can only be directly access when DSP in D0. To make it work with PM, we have some method to cache this filed into debugfs when enter suspend, but I think it is not prefect at now and will lose data in some case (tplg load fail message will lost becaise of unfinished multiple core support feature)
2. DMA trace: something like a capture stream. We have set up buffer in debugfs to store data copied by DMA from DSP. As the buffer is limited size, so it can only hold limit size logs in the past. But if you open the sof-logger before the test, it can record all logs.

To understand more about error trace, we need first to know where is the buffer. Following code take APL as example:
We have about 8KB buffer for error log:

sof/src/platform/apollolake/include/platform/lib/memory.h

Lines 175 to 181 in cf74566

	/* window 3 */
	#define SRAM_TRACE_BASE SRAM_WND_BASE
	#if CONFIG_TRACE
	#define SRAM_TRACE_SIZE 0x2000
	#else
	#define SRAM_TRACE_SIZE 0
	#endif

sof/src/platform/intel/cavs/include/cavs/lib/mailbox.h

Lines 38 to 40 in cf74566

	/* window 3 - trace */
	#define MAILBOX_TRACE_SIZE SRAM_TRACE_SIZE
	#define MAILBOX_TRACE_BASE SRAM_TRACE_BASE

Function for trace written to mailbox:

sof/src/trace/trace.c

Lines 68 to 97 in cf74566

	#if CONFIG_TRACEM
	static inline void mtrace_event(const char *data, uint32_t length)
	{
	struct trace *trace = trace_get();
	volatile char *t;
	uint32_t i, available;
	available = MAILBOX_TRACE_SIZE - trace->pos;
	t = (volatile char *)(MAILBOX_TRACE_BASE);
	/* write until we run out of space */
	for (i = 0; i < available && i < length; i++)
	t[trace->pos + i] = data[i];
	dcache_writeback_region((void *)&t[trace->pos], i);
	trace->pos += length;
	/* if there was more data than space available, wrap back */
	if (length > available) {
	for (i = 0; i < length - available; i++)
	t[i] = data[available + i];
	dcache_writeback_region((void *)t, i);
	trace->pos = i;
	}
	platform_shared_commit(trace, sizeof(*trace));
	}
	#endif /* CONFIG_TRACEM */

I am also look into this issue, it seems we have some error in trace implement.

[lyakh]

@lyakh To understand how the trace work, we need to know we have 2 kinds of trace:
[snip]
I am also look into this issue, it seems we have some error in trace implement.

@xiulipan thanks for the explanation, that's also my impression, that the current implementation isn't right and I don't understand whether and when and how it ever worked in the past.