Re: [PATCH v5 07/10] perf record: implement -z,--compression_level=n option and compression

[Alexey Budankov <alexey.budankov@linux.intel.com>]

On 08.03.2019 13:46, Jiri Olsa wrote:
> On Thu, Mar 07, 2019 at 06:26:47PM +0300, Alexey Budankov wrote:
>>
>> On 07.03.2019 15:14, Jiri Olsa wrote:
>>> On Thu, Mar 07, 2019 at 11:39:46AM +0300, Alexey Budankov wrote:
>>>>
>>>> On 05.03.2019 15:25, Jiri Olsa wrote:
>>>>> On Fri, Mar 01, 2019 at 06:58:32PM +0300, Alexey Budankov wrote:
>>>>>
>>>>> SNIP
>>>>>
>>>>>>  
>>>>>>  	/*
>>>>>>  	 * Increment md->refcount to guard md->data[idx] buffer
>>>>>> @@ -350,7 +357,7 @@ int perf_mmap__aio_push(struct perf_mmap *md, void *to, int idx,
>>>>>>  	md->prev = head;
>>>>>>  	perf_mmap__consume(md);
>>>>>>  
>>>>>> -	rc = push(to, &md->aio.cblocks[idx], md->aio.data[idx], size0 + size, *off);
>>>>>> +	rc = push(to, md->aio.data[idx], size0 + size, *off, &md->aio.cblocks[idx]);
>>>>>>  	if (!rc) {
>>>>>>  		*off += size0 + size;
>>>>>>  	} else {
>>>>>> @@ -556,13 +563,15 @@ int perf_mmap__read_init(struct perf_mmap *map)
>>>>>>  }
>>>>>>  
>>>>>>  int perf_mmap__push(struct perf_mmap *md, void *to,
>>>>>> -		    int push(struct perf_mmap *map, void *to, void *buf, size_t size))
>>>>>> +		    int push(struct perf_mmap *map, void *to, void *buf, size_t size),
>>>>>> +		    perf_mmap__compress_fn_t compress, void *comp_data)
>>>>>>  {
>>>>>>  	u64 head = perf_mmap__read_head(md);
>>>>>>  	unsigned char *data = md->base + page_size;
>>>>>>  	unsigned long size;
>>>>>>  	void *buf;
>>>>>>  	int rc = 0;
>>>>>> +	size_t mmap_len = perf_mmap__mmap_len(md);
>>>>>>  
>>>>>>  	rc = perf_mmap__read_init(md);
>>>>>>  	if (rc < 0)
>>>>>> @@ -574,7 +583,10 @@ int perf_mmap__push(struct perf_mmap *md, void *to,
>>>>>>  		buf = &data[md->start & md->mask];
>>>>>>  		size = md->mask + 1 - (md->start & md->mask);
>>>>>>  		md->start += size;
>>>>>> -
>>>>>> +		if (compress) {
>>>>>> +			size = compress(comp_data, md->data, mmap_len, buf, size);
>>>>>> +			buf = md->data;
>>>>>> +		}
>>>>>>  		if (push(md, to, buf, size) < 0) {
>>>>>>  			rc = -1;
>>>>>>  			goto out;
>>>>>
>>>>> when we discussed the compress callback should be another layer
>>>>> in perf_mmap__push I was thinking more of the layered/fifo design,
>>>>> like:
>>>>>
>>>>> normaly we call:
>>>>>
>>>>> 	perf_mmap__push(... push = record__pushfn ...)
>>>>> 		-> reads mmap data and calls push(data), which translates as:
>>>>>
>>>>> 		record__pushfn(data);
>>>>> 			- which stores the data
>>>>>
>>>>>
>>>>> for compressed it'd be:
>>>>>
>>>>> 	perf_mmap__push(... push = compressed_push ...)
>>>>>
>>>>> 		-> reads mmap data and calls push(data), which translates as:
>>>>>
>>>>> 		compressed_push(data)
>>>>> 			-> reads data, compresses them and calls, next push callback in line:
>>>>>
>>>>> 			record__pushfn(data)
>>>>> 				- which stores the data
>>>>>
>>>>>
>>>>> there'd need to be the logic for compressed_push to
>>>>> remember the 'next push' function
>>>>
>>>> That is suboptimal for AIO. Also compression is an independent operation that 
>>>> could be applied on any of push stages you mean.
>>>
>>> not sure what you mean by suboptimal, but I think
>>> that it can still happen in subsequent push callback
>>>
>>>>
>>>>>
>>>>> but I think this was the original idea behind the
>>>>> perf_mmap__push -> it gets the data and pushes them for
>>>>> the next processing.. it should stay as simple as that
>>>>
>>>> Agree on keeping simplicity and, at the moment, there is no any push to the next 
>>>> processing in the code so provided implementation fits as for serial as for AIO
>>>> at the same time sticking to simplicity as much as possibly. If you see something 
>>>> that would fit better please speak up and share.
>>>
>>> I have to insist that perf_mmap__push stays untouched
>>> and we do other processing in the push callbacks
>>
>> What is about perf_mmap__aio_push()?
>>
>> Without compression it does 
>> 	memcpy(), memcpy(), aio_push()
>>
>> With compression its does
>> 	memcpy_with_compression(), memcpy_with_compression(), aio_push()
> 
> so to be on the same page.. normal processing without compression is:
> 
>   perf_mmap__push does:
> 	push(mmap buf)
> 	  record__pushfn
> 	    record__write
> 	      write(buf)
> 
>   perf_mmap__aio_push does:
> 	memcpy(aio buf, mmap buf)
> 	  push(aio buf)
> 	    record__aio_pushfn
> 	      record__aio_write
> 	        aio_write(aio buf)
> 
> 
> and for compression it would be:
> 
>   perf_mmap__push does:
> 	push(mmap buf)
> 	  compress_push
> 	    memcpy(compress buffer, mmapbuf)  EXTRA copy
> 	      record__pushfn
> 	        record__write
> 		  write(buf)
> 
>   perf_mmap__aio_push does:
> 	memcpy(aio buf, mmap buf)
> 	  memcpy(compress buffer, mmapbuf)    EXTRA copy
> 	    push(aio buf)
> 	      record__aio_pushfn
> 	        record__aio_write
> 	          aio_write(aio buf)
> 
> 
> side note: that actualy makes me think why do we even have perf_mmap__aio_push,
> it looks like we could copy the buf in the callback push function with no harm?

Well, yes, perf_mmap__aio_push() can be avoided and perf_mmap__push() can be used
as for serial as for AIO, moving all the specifics to record code from mmap.c,
like this:

Serial
	perf_mmap__push(, record__pushfn)
 	   push(), possibly two times
	      record__pushfn()
                 if (-z) zstd_compress(map->base => map->data) <-- compressing memcpy()
                 record__write(-z ? map->data, map->base)
AIO
	record__aio_push()
	   perf_mmap__push(, record__aio_pushfn())
              push(), possibly two times
                 record__aio_pushfn()
                    if (-z) zstd_compress(map->base => map->aio.data[i]) <--- compressing memcpy()
                    else memcpy(map->base => map->aio.data[i]) <--- plain memcpy()
           record__aio_write(map->aio.data[i]) 

So now it looks optimal as from performance and data loss reduction 
perspective as from design perspective. What do you think?

~Alexey

> 
> so.. there's one extra memcpy for compression, is it right?
> I might miss some part which makes this scheme unusable..
> 
> thanks,
> jirka
>