[dpdk-dev] [PATCH v5] sched: make RED scaling configurable

Dewar, Alan ad759e at intl.att.com
Fri Jan 12 10:38:04 CET 2018


Hi Cristian,

Responses inline below.

Regards
Alan

> -----Original Message-----
> From: Dumitrescu, Cristian [mailto:cristian.dumitrescu at intel.com] 
> Sent: Thursday, January 11, 2018 1:11 PM
> To: alangordondewar at gmail.com
> Cc: dev at dpdk.org; Alan Dewar <alan.dewar at att.com>; Kantecki, Tomasz <tomasz.kantecki at intel.com>; Singh, Jasvinder <jasvinder.singh at intel.com>
> Subject: RE: [PATCH v5] sched: make RED scaling configurable
>
> Hi Alan,
>
> More issues and questions below:
>
>...<snip>
>
> > diff --git a/lib/librte_sched/rte_red.c b/lib/librte_sched/rte_red.h
> I suggest we use uint32_t for rte_red_scaling and rte_red_max_threshold, even though their values can fit into uint8_t and uint16_t respectively.

Will do.

> > diff --git a/lib/librte_sched/rte_red.c b/lib/librte_sched/rte_red.c
> > +int
> > +rte_red_set_scaling(uint16_t max_red_queue_length) {
> > +	int8_t count;
> > +
> > +	if (rte_red_init_done)
> > +		/**
> > +		 * Can't change the scaling once the red table has been
> > +		 * computed.
> > +		 */
> > +		return -1;
>
> Is there a reason why we cannot simply reset the scaling here?

Actually we could, but I was originally thinking that you might be happier keeping with a one-time RED initialization function, but then had to introduce the rte_reset_red_scaling function for the unit-tests. I'm happy to do RED reinitialization here, if you are.

> > +
> > +	if (max_red_queue_length < RTE_RED_MIN_QUEUE_LENGTH)
> > +		return -2;
> > +
> > +	if (max_red_queue_length > RTE_RED_MAX_QUEUE_LENGTH)
> > +		return -3;
> > +
> > +	if (!rte_is_power_of_2(max_red_queue_length))
> > +		return -4;
> > +
> > +	count = 0;
> > +	while (max_red_queue_length != 0) {
> > +		max_red_queue_length >>= 1;
> > +		count++;
> > +	}
>
> This does not look right to me. I think you want to compute the log2 of max_red_queue_length here, but your result (count) is bigger by 1 than it should be, right?
> When max_red_queue_length = RTE_RED_DEFAULT_QUEUE_LENGTH = 1024, the result should be: count = RTE_RED_SCALING = 10, not 11.
>

Well spotted.

> I suggest you use rte_bsf32() function for this purpose.

I wasn't aware of this function, it appears to do the necessary so will do.

> > +
> > +	rte_red_scaling -= count - RTE_RED_SCALING;
>
> Why not simply: rte_red_scaling = count?

Stupidity on my part! - Will do.

> > +	rte_red_max_threshold = max_red_queue_length - 1;
> > +	return 0;
> > +}
> > +
>
> > diff --git a/lib/librte_sched/rte_sched_version.map
> > b/lib/librte_sched/rte_sched_version.map
> > index 3aa159a..262bece 100644
> > --- a/lib/librte_sched/rte_sched_version.map
> > +++ b/lib/librte_sched/rte_sched_version.map
> > @@ -29,3 +29,10 @@ DPDK_2.1 {
> >  	rte_sched_port_pkt_read_color;
> > 
> >  } DPDK_2.0;
> > +
> > +DPDK_17.08 {
> > +	global;
> > +
> > +	__rte_red_reset;
> > +	rte_red_set_scaling;
> > +} DPDK_2.1;
>
> You need to put the correct DPDK release number here.

Will do.

> You also need to do more work to make sure the share library support  is not broken:
> -need to increase LIBABIVER in Makefile
> -need to update the library .so number in release notes -need to check that build does not fail for shared libraries
>
> (You can use this patch as example: https://urldefense.proofpoint.com/v2/url?u=https-3A__dpdk.org_dev_patchwork_patch_33109_&d=DwIFAg&c=LFYZ-o9_HUMeMTSQicvjIg&r=MSobi4bsl2plY_3b1jvY6kXt-SHzSxMj5YhuyxfTjJE&m=D7OTgDyX27pcg-xs8N7FQzX4lAkkcPqbPAhJp_IyrNI&s=mCvsqitq8Y6vD3UC-r1WEGtrRp17bsmJaEHGby97zYw&e=  )

Okay, I need to look into this stuff - I'll come back with some questions if anything isn't clear.

> > diff --git a/test/test/test_red.c b/test/test/test_red.c
> The RED  test code fails to link, please fix:

Yes - will do.

> export RTE_TARGET=x86_64-native-linuxapp-gcc
>
> [ ]$ ~/git_dpdk_red/dpdk$ make -C test/test
> make: Entering directory 'dpdk/test/test'
>...<snip>
>    LD test
> test_red.o: In function `test_rte_red_init':
> test_red.c:(.text+0x2cd): undefined reference to `__rte_red_reset'
> test_red.c:(.text+0x2da): undefined reference to `rte_red_set_scaling'
> test_red.c:(.text+0x2ed): undefined reference to `rte_red_max_threshold'
> test_red.o: In function `increase_actual_qsize':
> test_red.c:(.text+0x48c): undefined reference to `rte_red_scaling'
> test_red.c:(.text+0x533): undefined reference to `rte_red_scaling'
> test_red.o: In function `enqueue_dequeue_func':
> test_red.c:(.text+0x70c): undefined reference to `rte_red_scaling'
> test_red.c:(.text+0x7a5): undefined reference to `rte_red_scaling'
> test_red.o: In function `increase_average_qsize':
> test_red.c:(.text+0x95c): undefined reference to `rte_red_scaling'
> test_red.o:test_red.c:(.text+0x9f1): more undefined references to `rte_red_scaling' follow
> collect2: error: ld returned 1 exit status
> dpdk/mk/rte.app.mk:306: recipe for target 'test' failed
> make: *** [test] Error 1
> make: Leaving directory 'dpdk/test/test'
>
> Regards,
> Cristian



More information about the dev mailing list