Re: [PATCH v3 1/1] regmap: Synchronize cache for the page selector

[Andy Shevchenko <andriy.shevchenko@linux.intel.com>]

On Tue, Jan 28, 2025 at 06:43:26PM +0200, Andy Shevchenko wrote:
> On Tue, Jan 28, 2025 at 05:08:08PM +0100, Marek Szyprowski wrote:
> > On 21.01.2025 14:29, Andy Shevchenko wrote:
> > > On Tue, Jan 21, 2025 at 08:33:09AM +0100, Marek Szyprowski wrote:
> > >> On 17.01.2025 18:28, Andy Shevchenko wrote:
> > >>> On Fri, Jan 17, 2025 at 05:05:42PM +0100, Marek Szyprowski wrote:
> > >>>
> > >>> Does it fail in the same way?
> > >> Yes, the hw revision is reported as zero in this case: LT9611 revision:
> > >> 0x00.00.00
> > > Hmm... This is very interesting! It means that the page selector is a bit
> > > magical there. Dmitry, can you chime in and perhaps shed some light on this?
> > >
> > >>>> Does it mean that there is really a bug in the driver?
> > >>> Without looking at the datasheet it's hard to say. At least what I found so far
> > >>> is one page of the I²C traffic dump on Windows as an example how to use their
> > >>> evaluation board and software, but it doesn't unveil the bigger picture. At
> > >>> least what I think is going on here is that the programming is not so easy as
> > >>> just paging. Something is more complicated there.
> > >>>
> > >>> But at least (and as Mark said) the most of the regmap based drivers got
> > >>> the ranges wrong (so, at least there is one bug in the driver).
> > >> I can do more experiments if this helps. Do you need a dump of all
> > >> regmap accesses or i2c traffic from this driver?
> > > It would be helpful! Traces from the failed and non-failed cases
> > > till the firmware revision and chip ID reading would be enough to
> > > start with.
> > 
> > I'm sorry for the delay, I was a bit busy with other stuff.
> 
> No problem and thanks for sharing.
> 
> > Here are logs (all values are in hex):
> > 
> > next-20250128 (probe broken):
> > root@target:~# dmesg | grep regmap
> > [   14.817604] regmap_write reg 80ee <- 1
> > [   14.823036] regmap_read reg 8100 -> 0
> > [   14.827631] regmap_read reg 8101 -> 0
> > [   14.832130] regmap_read reg 8102 -> 0
> 
> 
> 
> > next-20250128 + 1fd60ed1700c reverted (probe okay):
> > root@target:~# dmesg | grep regmap
> > [   13.565920] regmap_write reg 80ee <- 1
> > [   13.567509] regmap_read reg 8100 -> 17
> > [   13.568219] regmap_read reg 8101 -> 4
> > [   13.568909] regmap_read reg 8102 -> 93
> 
> Something is missing here. Like we have an identical start and an immediate
> failure. If you did it via printk() approach, it's probably wrong as my patch
> uses internal regmap function. Most likely you need to enable trace events
> for regmap and collect those for let's say 2 seconds:
> 
> 	echo 1 > ...trace events...
> 	modprobe ...
> 	sleep 2
> 	echo 0 > ...trace events...
> 
> and dump the buffer to a file. It might have though more than needed
> as some other devices might also use regmap at the same time. I don't remember
> if the trace events for regmap have a device instance name field which can be
> used as a filter.
> 
> Alternatively you may also try to add a printk() into regmap core, but I don't
> think it's more practical than trace events.

Meanwhile, can you test this patch (on top of non-working case)?

diff --git a/drivers/base/regmap/regmap.c b/drivers/base/regmap/regmap.c
index 2314744201b4..f799a7a80231 100644
--- a/drivers/base/regmap/regmap.c
+++ b/drivers/base/regmap/regmap.c
@@ -1553,8 +1553,19 @@ static int _regmap_select_page(struct regmap *map, unsigned int *reg,
 		 * virtual copy as well.
 		 */
 		if (page_chg &&
-		    in_range(range->selector_reg, range->window_start, range->window_len))
+		    in_range(range->selector_reg, range->window_start, range->window_len)) {
+			bool bypass, cache_only;
+
+			bypass = map->cache_bypass;
+			cache_only = map->cache_only;
+			map->cache_bypass = false;
+			map->cache_only = true;
+
 			_regmap_update_bits(map, sel_register, mask, val, NULL, false);
+
+			map->cache_bypass = bypass;
+			map->cache_only = cache_only;
+		}
 	}
 
 	*reg = range->window_start + win_offset;

If I understood the case, the affected driver doesn't use case and we actually
write to the selector register twice which most likely messes up the things.
But this is only a theory (since we don't have the traces yet).

-- 
With Best Regards,
Andy Shevchenko