The era of hand-rolled HIWORD_UPDATE macros is over, at least for those
drivers that use constant masks.
Like many other Rockchip drivers, rockchip-dfi brings with it its own
HIWORD_UPDATE macro. This variant doesn't shift the value (and like the
others, doesn't do any checking).
Remove it, and replace instances of it with hw_bitfield.h's
FIELD_PREP_WM16. Since FIELD_PREP_WM16 requires contiguous masks and
shifts the value for us, some reshuffling of definitions needs to
happen.
This gives us better compile-time error checking, and in my opinion,
nicer code.
Tested on an RK3568 ODROID-M1 board (LPDDR4X at 1560 MHz, an RK3588
Radxa ROCK 5B board (LPDDR4X at 2112 MHz) and an RK3588 Radxa ROCK 5T
board (LPDDR5 at 2400 MHz). perf measurements were consistent with the
measurements of stress-ng --stream in all cases.
Signed-off-by: Nicolas Frattaroli <nicolas.frattaroli@collabora.com>
Signed-off-by: Yury Norov (NVIDIA) <yury.norov@gmail.com>
The Rockchip RK3588 SoC can also support LPDDR5 memory. This type of
memory needs some special case handling in the rockchip-dfi driver.
Add support for it in rockchip-dfi, as well as the needed GRF register
definitions.
This has been tested as returning both the right cycle count and
bandwidth on a LPDDR5 board where the CKR bit is 1. I couldn't test
whether the values are correct on a system where CKR is 0, as I'm not
savvy enough with the Rockchip tooling to know whether this can be set
in the DDR init blob.
Downstream has some special case handling for a hardware version where
not just the control bits differ, but also the register. Since I don't
know whether that hardware version is in any production silicon, it's
left unimplemented for now, with an error message urging users to report
if they have such a system.
There is a slight change of behaviour for non-LPDDR5 systems: instead of
writing 0 as the control flags to the control register and pretending
everything is alright if the memory type is unknown, we now explicitly
return an error.
Signed-off-by: Nicolas Frattaroli <nicolas.frattaroli@collabora.com>
Reviewed-by: Sascha Hauer <s.hauer@pengutronix.de>
Acked-by: Heiko Stuebner <heiko@sntech.de>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
Link: https://patchwork.kernel.org/project/linux-pm/patch/20250530-rk3588-dfi-improvements-v1-2-6e077c243a95@collabora.com/
On RK3588 with LPDDR4X memory, the cycle count as returned by
perf stat -a -e rockchip_ddr/cycles/ sleep 1
consistently reads half as much as what the actual DDR frequency is at.
For a LPDDR4X module running at 2112MHz, I get more like 1056059916
cycles per second, which is almost bang-on half what it should be. No,
I'm not mixing up megatransfers and megahertz.
Consulting the downstream driver, this appears to be because the RK3588
hardware specifically (and RK3528 as well, for future reference) needs a
multiplier of 2 to get to the correct frequency with everything but
LPDDR5.
The RK3588's actual memory bandwidth measurements in MB/s are correct
however, as confirmed with stress-ng --stream. This makes me think the
access counters are not affected in the same way. This tracks with the
vendor kernel not multiplying the access counts either.
Solve this by adding a new member to the dfi struct, which each SoC can
set to whatever they want, but defaults to 1 if left unset by the SoC
init functions. The event_get_count op can then use this multiplier if
the cycle count is requested.
The cycle multiplier is not used in rockchip_dfi_get_event because the
vendor driver doesn't use it there either, and we don't do other actual
bandwidth unit conversion stuff in there anyway.
Fixes: 481d97ba61 ("PM / devfreq: rockchip-dfi: add support for RK3588")
Signed-off-by: Nicolas Frattaroli <nicolas.frattaroli@collabora.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
Link: https://lore.kernel.org/lkml/20250530-rk3588-dfi-improvements-v1-1-6e077c243a95@collabora.com/
The continual trickle of small conversion patches is grating on me, and
is really not helping. Just get rid of the 'remove_new' member
function, which is just an alias for the plain 'remove', and had a
comment to that effect:
/*
* .remove_new() is a relic from a prototype conversion of .remove().
* New drivers are supposed to implement .remove(). Once all drivers are
* converted to not use .remove_new any more, it will be dropped.
*/
This was just a tree-wide 'sed' script that replaced '.remove_new' with
'.remove', with some care taken to turn a subsequent tab into two tabs
to make things line up.
I did do some minimal manual whitespace adjustment for places that used
spaces to line things up.
Then I just removed the old (sic) .remove_new member function, and this
is the end result. No more unnecessary conversion noise.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The .remove() callback for a platform driver returns an int which makes
many driver authors wrongly assume it's possible to do error handling by
returning an error code. However the value returned is ignored (apart
from emitting a warning) and this typically results in resource leaks.
To improve here there is a quest to make the remove callback return
void. In the first step of this quest all drivers are converted to
.remove_new(), which already returns void. Eventually after all drivers
are converted, .remove_new() will be renamed to .remove().
Trivially convert this driver from always returning zero in the remove
callback to the void returning variant.
Signed-off-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The .remove() callback for a platform driver returns an int which makes
many driver authors wrongly assume it's possible to do error handling by
returning an error code. However the value returned is ignored (apart
from emitting a warning) and this typically results in resource leaks.
To improve here there is a quest to make the remove callback return
void. In the first step of this quest all drivers are converted to
.remove_new(), which already returns void. Eventually after all drivers
are converted, .remove_new() will be renamed to .remove().
Trivially convert this driver from always returning zero in the remove
callback to the void returning variant.
Signed-off-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The currently supported RK3399 has a set of registers per channel, but
it has only a single DDRMON_CTRL register. With upcoming RK3588 this
will be different, the RK3588 has a DDRMON_CTRL register per channel.
Instead of expecting a single DDRMON_CTRL register, loop over the
channels and write the channel specific DDRMON_CTRL register. Break
out early out of the loop when there is only a single DDRMON_CTRL
register like on the RK3399.
Link: https://lore.kernel.org/all/20231018061714.3553817-19-s.hauer@pengutronix.de/
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Sebastian Reichel <sebastian.reichel@collabora.com>
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The DFI is a unit which is suitable for measuring DDR utilization, but
so far it could only be used as an event driver for the DDR frequency
scaling driver. This adds perf support to the DFI driver.
Usage with the 'perf' tool can look like:
perf stat -a -e rockchip_ddr/cycles/,\
rockchip_ddr/read-bytes/,\
rockchip_ddr/write-bytes/,\
rockchip_ddr/bytes/ sleep 1
Performance counter stats for 'system wide':
1582524826 rockchip_ddr/cycles/
1802.25 MB rockchip_ddr/read-bytes/
1793.72 MB rockchip_ddr/write-bytes/
3595.90 MB rockchip_ddr/bytes/
1.014369709 seconds time elapsed
perf support has been tested on a RK3568 and a RK3399, the latter with
dual channel DDR.
Link: https://lore.kernel.org/all/20231019064819.3496740-1-s.hauer@pengutronix.de/
Reviewed-by: Sebastian Reichel <sebastian.reichel@collabora.com>
Acked-by: Chanwoo Choi <cw00.choi@samsung.com>
Acked-by: Heiko Stuebner <heiko@sntech.de>
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
[cw00.choi: Fix typo from 'write_acccess' to 'write_access']
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
Use the HIWORD_UPDATE() define known from other rockchip drivers to
make the defines look less odd to the readers who've seen other
rockchip drivers.
The HIWORD registers have their functional bits in the lower 16 bits
whereas the upper 16 bits contain a mask. Only the functional bits that
have the corresponding mask bit set are modified during a write. Although
the register writes look different, the end result should be the same,
at least there's no functional change intended with this patch.
Link: https://lore.kernel.org/all/20231018061714.3553817-10-s.hauer@pengutronix.de/
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
Reviewed-by: Sebastian Reichel <sebastian.reichel@collabora.com>
Acked-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The DDRTYPE defines are named to be RK3399 specific, but they can be
used for other Rockchip SoCs as well, so replace the RK3399_PMUGRF_
prefix with ROCKCHIP_. They are defined in a SoC specific header
file, so when generalizing the prefix also move the new defines to
a SoC agnostic header file. While at it use GENMASK to define the
DDRTYPE bitfield and give it a name including the full register name.
Link: https://lore.kernel.org/all/20231018061714.3553817-9-s.hauer@pengutronix.de/
Reviewed-by: Sebastian Reichel <sebastian.reichel@collabora.com>
Acked-by: Chanwoo Choi <cw00.choi@samsung.com>
Acked-by: Heiko Stuebner <heiko@sntech.de>
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The DDR_MON counters are free running counters. These are resetted to 0
when starting them over like currently done when reading the current
counter values.
Resetting the counters becomes a problem with perf support we want to
add later, because perf needs counters that are not modified elsewhere.
This patch removes resetting the counters and keeps them running
instead. That means we no longer use the absolute counter values but
instead compare them with the counter values we read last time. Not
stopping the counters also has the impact that they are running while
we are reading them. We cannot read multiple timers atomically, so
the values do not exactly fit together. The effect should be negligible
though as the time between two measurements is some orders of magnitude
bigger than the time we need to read multiple registers.
Link: https://lore.kernel.org/all/20231018061714.3553817-7-s.hauer@pengutronix.de/
Reviewed-by: Sebastian Reichel <sebastian.reichel@collabora.com>
Acked-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
