Date:   Wed, 5 Dec 2018 14:25:50 +0800
From:   leo.yan@...aro.org
To:     Mathieu Poirier <mathieu.poirier@...aro.org>
Cc:     Arnaldo Carvalho de Melo <acme@...nel.org>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Jiri Olsa <jolsa@...hat.com>,
        Namhyung Kim <namhyung@...nel.org>,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
        Mike Leach <mike.leach@...aro.org>,
        Robert Walker <robert.walker@....com>,
        Al Grant <Al.Grant@....com>,
        Coresight ML <coresight@...ts.linaro.org>,
        Andi Kleen <andi@...stfloor.org>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Arnaldo Carvalho de Melo <acme@...hat.com>
Subject: Re: [PATCH v2 1/2] perf cs-etm: Set branch instruction flags in
 packet

On Mon, Nov 19, 2018 at 03:26:17PM -0700, Mathieu Poirier wrote:
> On Sun, Nov 11, 2018 at 01:07:55PM +0800, Leo Yan wrote:
> > The perf sample data contains flags to indicate the hardware trace data
> > is belonging to which type branch instruction, thus this can be used to
> > print out the human readable string.  Arm CoreSight ETM sample data is
> > missed to set flags and it is always set to zeros, this results in perf
> > tool skips to print string for instruction types.
> > 
> > Arm CoreSight ETM supports different kinds instruction of A64, A32 and
> > T32; this patch is to set branch instruction flags in packet for these
> > ISAs.
> > 
> > The brief idea for patch implementation is describe as below:
> > 
> > - For element with OCSD_GEN_TRC_ELEM_TRACE_ON type, it is taken as trace
> >   beginning packet; for element with OCSD_GEN_TRC_ELEM_NO_SYNC or
> >   OCSD_GEN_TRC_ELEM_EO_TRACE, these two kinds elements are used to set
> >   for trace end;
> > 
> >   As Mike suggested the packet stream might have more than one two
> >   TRACE_ON packets, the first one TRACE_ON packet indicates trace end
> >   and the second one is taken as trace restarting.  We will handle this
> >   special case in the upper layer with packet queue handling, which has
> >   more context so it's more suitable fix up for it.  This will be
> >   accomplished in the sequential patch.
> > 
> > - For instruction range packet, mainly base on three factors to decide
> >   the branch instruction types:
> > 
> >   elem->last_i_type
> >   elem->last_i_subtype
> >   elem->last_instr_cond
> > 
> >   If the instruction is immediate branch but without link and return
> >   flag, we consider it as function internal branch;  in fact the
> >   immediate branch also can be used to invoke the function entry,
> >   usually this is only used in assembly code to directly call a symbol
> >   and don't expect to return back; after reviewing kernel normal
> >   functions and user space programs, both of them are very seldom to use
> >   immediate branch for function call.  On the other hand, if we want to
> >   decide the immediate branch is for function branch jumping or for
> >   function calling, we need to rely on the start address of next packet
> >   and check the symbol offset for the start address,  this will
> >   introduce much complexity in the implementation.  So for this version
> >   we simply consider immediate branch as function internal branch.
> >   Moreover, we rely on 'elem->last_instr_cond' to decide if the branch
> >   instruction is a conditional branch or not.
> > 
> >   If the instruction is immediate branch with link, it's instruction
> >   'BL' and which is used for function call.
> > 
> >   If the instruction is indirect branch and with subtype
> >   OCSD_S_INSTR_V7_IMPLIED_RET, the decoders gives the hint the function
> >   return for below cases related with A32/T32 instruction; set this
> >   branch flag as function return (Thanks for Al's suggestion).
> > 
> >     BX R14
> >     MOV PC, LR
> >     POP {…, PC}
> >     LDR PC, [SP], #offset
> > 
> >   If the instruction is indirect branch without link, this is
> >   corresponding to instruction 'BR', this instruction usually is used
> >   for dynamic link lib with below usage; so we think it's a return
> >   instruction.
> > 
> >     0000000000000680 <.plt>:
> >      680:   a9bf7bf0        stp     x16, x30, [sp, #-16]!
> >      684:   90000090        adrp    x16, 10000 <__FRAME_END__+0xf630>
> >      688:   f947fe11        ldr     x17, [x16, #4088]
> >      68c:   913fe210        add     x16, x16, #0xff8
> >      690:   d61f0220        br      x17
> > 
> >   If the instruction is indirect branch with link, e.g BLR, we think
> >   it's a function call.
> > 
> >   For function return, ARMv8 introduces a dedicated instruction 'ret',
> >   which has flag of OCSD_S_INSTR_V8_RET.
> > 
> > - For exception packets, this patch divides into three types:
> > 
> >   The first type of exception is caused by external logics like bus,
> >   interrupt controller, debug module or PE reset or halt; this is
> >   corresponding to flags "bcyi" which defined in doc perf-script.txt;
> > 
> >   The second type is for system call, this is set as "bcs" by following
> >   definition in the doc;
> > 
> >   The third type is for CPU trap, data and instruction prefetch abort,
> >   alignment abort; usually these exceptions are synchronous for CPU, so
> >   set them as "bci" type.
> 
> This is too long and needs to be broken down into pieces.  I would split this
> patch in 3 heat, one for NO_SYNC and TRACE_ON, one for INSTR_RANGE and one for
> ELEM_EXCEPTION/ELEM_EXCEPTION_RET. 

This makes sense, will split to 3 patches.

> > 
> > Cc: Mathieu Poirier <mathieu.poirier@...aro.org>
> > Cc: Mike Leach <mike.leach@...aro.org>
> > Cc: Robert Walker <robert.walker@....com>
> > Cc: Al Grant <Al.Grant@....com>
> > Cc: Andi Kleen <andi@...stfloor.org>
> > Cc: Adrian Hunter <adrian.hunter@...el.com>
> > Cc: Arnaldo Carvalho de Melo <acme@...hat.com>
> > Signed-off-by: Leo Yan <leo.yan@...aro.org>
> > ---
> >  tools/perf/util/cs-etm-decoder/cs-etm-decoder.c | 168 ++++++++++++++++++++++++
> >  tools/perf/util/cs-etm-decoder/cs-etm-decoder.h |   1 +
> >  2 files changed, 169 insertions(+)
> > 
> > diff --git a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c
> > index d1a6cbc..0e50c52 100644
> > --- a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c
> > +++ b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c
> > @@ -303,6 +303,7 @@ cs_etm_decoder__buffer_packet(struct cs_etm_decoder *decoder,
> >  	decoder->packet_buffer[et].instr_count = 0;
> >  	decoder->packet_buffer[et].last_instr_taken_branch = false;
> >  	decoder->packet_buffer[et].last_instr_size = 0;
> > +	decoder->packet_buffer[et].flags = 0;
> 
> Since PERF_IP_FLAG_BRANCH is '0', I would set this to UNINT32_MAX.

PERF_IP_FLAG_BRANCH is bit 0 is set (so it's 1) but not 0.  If
initialize value to UNINT32_MAX (0xFFFF,FFFF) that means all flags has
been set and this will introduce confusion.  So will keep to init
flags to 0.

> >  
> >  	if (decoder->packet_count == MAX_BUFFER - 1)
> >  		return OCSD_RESP_WAIT;
> > @@ -437,6 +438,171 @@ cs_etm_decoder__buffer_exception_ret(struct cs_etm_decoder *decoder,
> >  					     CS_ETM_EXCEPTION_RET);
> >  }
> >  
> > +static void cs_etm_decoder__set_sample_flags(
> > +				const void *context,
> > +				const ocsd_generic_trace_elem *elem)
> > +{
> > +	struct cs_etm_decoder *decoder = (struct cs_etm_decoder *) context;
> > +	struct cs_etm_packet *packet;
> > +	u32 exc_num;
> > +
> > +	packet = &decoder->packet_buffer[decoder->tail];
> > +
> > +	switch (elem->elem_type) {
> > +	case OCSD_GEN_TRC_ELEM_TRACE_ON:
> > +		packet->flags = PERF_IP_FLAG_BRANCH |
> > +				PERF_IP_FLAG_TRACE_BEGIN;
> > +		break;
> > +
> > +	case OCSD_GEN_TRC_ELEM_NO_SYNC:
> > +	case OCSD_GEN_TRC_ELEM_EO_TRACE:
> > +		packet->flags = PERF_IP_FLAG_BRANCH |
> > +				PERF_IP_FLAG_TRACE_END;
> > +		break;
> > +
> > +	case OCSD_GEN_TRC_ELEM_INSTR_RANGE:
> > +		/*
> > +		 * Immediate branch instruction without neither link nor
> > +		 * return flag, it's normal branch instruction within
> > +		 * the function.
> > +		 */
> > +		if (elem->last_i_type == OCSD_INSTR_BR &&
> > +		    elem->last_i_subtype == OCSD_S_INSTR_NONE) {
> > +			packet->flags = PERF_IP_FLAG_BRANCH;
> > +
> > +			if (elem->last_instr_cond)
> > +				packet->flags |= PERF_IP_FLAG_CONDITIONAL;
> > +		}
> > +
> > +		/*
> > +		 * Immediate branch instruction with link (e.g. BL), this is
> > +		 * branch instruction for function call.
> > +		 */
> > +		if (elem->last_i_type == OCSD_INSTR_BR &&
> > +		    elem->last_i_subtype == OCSD_S_INSTR_BR_LINK)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_CALL;
> > +
> > +		/*
> > +		 * Indirect branch instruction with subtype of
> > +		 * OCSD_S_INSTR_V7_IMPLIED_RET, this is explicit hint for
> > +		 * function return for A32/T32.
> > +		 */
> > +		if (elem->last_i_type == OCSD_INSTR_BR_INDIRECT &&
> > +		    elem->last_i_subtype == OCSD_S_INSTR_V7_IMPLIED_RET)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_RETURN;
> > +
> > +		/*
> > +		 * Indirect branch instruction without link (e.g. BR), usually
> > +		 * this is used for function return, especially for functions
> > +		 * within dynamic link lib.
> > +		 */
> > +		if (elem->last_i_type == OCSD_INSTR_BR_INDIRECT &&
> > +		    elem->last_i_subtype == OCSD_S_INSTR_NONE)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_RETURN;
> > +
> > +		/*
> > +		 * Indirect branch instruction with link (e.g. BLR), this is
> > +		 * branch instruction for function call.
> > +		 */
> > +		if (elem->last_i_type == OCSD_INSTR_BR_INDIRECT &&
> > +		    elem->last_i_subtype == OCSD_S_INSTR_BR_LINK)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_CALL;
> > +
> > +		/* Return instruction for function return. */
> > +		if (elem->last_i_type == OCSD_INSTR_BR_INDIRECT &&
> > +		    elem->last_i_subtype == OCSD_S_INSTR_V8_RET)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_RETURN;
> 
> I would swap the last to if() condition so that the (BRANCH | RETURN) flags
> are all at the same place.

Will follow this in new patch.

> > +
> > +		break;
> > +
> > +	case OCSD_GEN_TRC_ELEM_EXCEPTION:
> > +
> > +#define OCSD_EXC_RESET			0
> > +#define OCSD_EXC_DEBUG_HALT		1
> > +#define OCSD_EXC_CALL			2
> > +#define OCSD_EXC_TRAP			3
> > +#define OCSD_EXC_SYSTEM_ERROR		4
> > +#define OCSD_EXC_INST_DEBUG		6
> > +#define OCSD_EXC_DATA_DEBUG		7
> > +#define OCSD_EXC_ALIGNMENT		10
> > +#define OCSD_EXC_INST_FAULT		11
> > +#define OCSD_EXC_DATA_FAULT		12
> > +#define OCSD_EXC_IRQ			14
> > +#define OCSD_EXC_FIQ			15
> 
> Where did you get the above?  To me this is something that should come from the
> library.

The concept is coming from OpenCSD lib but OpenCSD doesn't define
macros for them.

Will polish this and add these macros into OpenCSD header file.

> > +
> > +		exc_num = decoder->exc_num[packet->cpu];
> > +
> > +		/*
> > +		 * The exceptions are triggered by external signals
> > +		 * from bus, interrupt controller, debug module,
> > +		 * PE reset or halt.
> > +		 */
> > +		if (exc_num == OCSD_EXC_RESET ||
> > +		    exc_num == OCSD_EXC_DEBUG_HALT ||
> > +		    exc_num == OCSD_EXC_SYSTEM_ERROR ||
> > +		    exc_num == OCSD_EXC_INST_DEBUG ||
> > +		    exc_num == OCSD_EXC_DATA_DEBUG ||
> > +		    exc_num == OCSD_EXC_IRQ ||
> > +		    exc_num == OCSD_EXC_FIQ)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_CALL |
> > +					PERF_IP_FLAG_ASYNC |
> > +					PERF_IP_FLAG_INTERRUPT;
> > +
> > +		/* The exception is for system call. */
> > +		if (exc_num == OCSD_EXC_CALL)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_CALL |
> > +					PERF_IP_FLAG_SYSCALLRET;
> > +
> > +		/*
> > +		 * The exception is introduced by trap, instruction &
> > +		 * data fault or alignment errors.
> > +		 */
> > +		if (exc_num == OCSD_EXC_TRAP ||
> > +		    exc_num == OCSD_EXC_ALIGNMENT ||
> > +		    exc_num == OCSD_EXC_INST_FAULT ||
> > +		    exc_num == OCSD_EXC_DATA_FAULT)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_CALL |
> > +					PERF_IP_FLAG_INTERRUPT;
> > +
> > +		break;
> > +
> > +	case OCSD_GEN_TRC_ELEM_EXCEPTION_RET:
> > +
> > +		exc_num = decoder->exc_num[packet->cpu];
> > +
> > +		if (exc_num == OCSD_EXC_CALL)
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_RETURN |
> > +					PERF_IP_FLAG_SYSCALLRET;
> > +		else
> > +			packet->flags = PERF_IP_FLAG_BRANCH |
> > +					PERF_IP_FLAG_RETURN |
> > +					PERF_IP_FLAG_INTERRUPT;
> > +
> > +		break;
> > +
> > +	case OCSD_GEN_TRC_ELEM_UNKNOWN:
> > +	case OCSD_GEN_TRC_ELEM_PE_CONTEXT:
> > +	case OCSD_GEN_TRC_ELEM_ADDR_NACC:
> > +	case OCSD_GEN_TRC_ELEM_TIMESTAMP:
> > +	case OCSD_GEN_TRC_ELEM_CYCLE_COUNT:
> > +	case OCSD_GEN_TRC_ELEM_ADDR_UNKNOWN:
> > +	case OCSD_GEN_TRC_ELEM_EVENT:
> > +	case OCSD_GEN_TRC_ELEM_SWTRACE:
> > +	case OCSD_GEN_TRC_ELEM_CUSTOM:
> > +	default:
> > +		break;
> > +	}
> > +}
> > +
> >  static ocsd_datapath_resp_t cs_etm_decoder__gen_trace_elem_printer(
> >  				const void *context,
> >  				const ocsd_trc_index_t indx __maybe_unused,
> > @@ -484,6 +650,8 @@ static ocsd_datapath_resp_t cs_etm_decoder__gen_trace_elem_printer(
> >  		break;
> >  	}
> >  
> > +	cs_etm_decoder__set_sample_flags(context, elem);
> > +
> 
> I was toying with the idea of setting the flags in each of the case statement
> found in cs_etm_decoder__gen_trace_elem_printer().  But that would move more
> code around and the end result would be the same so let's keep it that way until
> we have a good reason to split it.

Do you sugguest to keep current implementation rather than to
split flags setting in each of the case statement in
cs_etm_decoder__gen_trace_elem_printer()?

I am not 100% sure if I understand correctly for "split it" (split flags
setting vs split functions).  So please correct me if I misunderstand
this.

Thanks,
Leo Yan

> >  	return resp;
> >  }
> >  
> > diff --git a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h
> > index 0d1c18d..71df908 100644
> > --- a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h
> > +++ b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h
> > @@ -47,6 +47,7 @@ struct cs_etm_packet {
> >  	u8 last_instr_taken_branch;
> >  	u8 last_instr_size;
> >  	int cpu;
> > +	u32 flags;
> >  };
> >  
> >  struct cs_etm_queue;
> > -- 
> > 2.7.4
> > 

Powered by blists - more mailing lists