RE: [RFC PATCH 4/6] arm_mpam: Introduce flexible CLOSID/RMID translation

From
Shaopeng Tan (Fujitsu) <hidden>
Date
2025-01-08 06:46:16
Also in
lkml
Reply to
Dave Martin, 2025-01-02
Hello Dave,
Hi,

On Fri, Dec 20, 2024 at 05:01:30AM +0000, Shaopeng Tan (Fujitsu) wrote:
quoted
Hello Dave,
[...]
quoted
quoted
diff --git a/drivers/platform/arm64/mpam/mpam_resctrl.c
b/drivers/platform/arm64/mpam/mpam_resctrl.c
index 30f2caec11d7..0473286ec65a 100644
--- a/drivers/platform/arm64/mpam/mpam_resctrl.c
+++ b/drivers/platform/arm64/mpam/mpam_resctrl.c
[...]
quoted
quoted
@@ -163,35 +175,71 @@ static bool mpam_resctrl_hide_cdp(enum
resctrl_res_level rid)
  */
 u32 resctrl_arch_get_num_closid(struct rdt_resource *ignored)  {
-	return mpam_partid_max + 1;
+	u32 res = (mpam_partid_max + 1) / partid_per_closid;
If there is a remainder, are you going to ignore the remainder PARTIDs?
It might be better to notify the user with a warning message.
You are right: the remainder is just thrown away, because the remaining
PARTIDs would not be sufficient for a whole resctrl control group.

It would be a good idea to print a warning somewhere though; I will try to add
that.

[...]
quoted
I think it is better to check whether PARTID narrowing feature is
supported or not before using this function.
If it is checked here, since this function is called multiple times,
the message should be printed only once by WARN_ON_ONCE().

Also, if the partid_per_closid is greater than (mpam_partid_max + 1),
returning "0" will not work properly.
To switch to default behavior (PARTID narrowing disabled), should it
return mapm_patid_max+ 1)?
The idea is that partid_per_closid is set by some configuration logic that is not
implemented yet.  This will be a bit more complicated than just exposing this
variable directly as a kernel parameter.

Currently I just implemented a basic range check to prevent the kernel
parameter being set to stupid values, but I didn't try to make it completely
robust.


In any case, since partid_per_closid defaults to 1, the behavior of this function
should just match the non-PARTID-Narrowing behaviour unless the user
requested a different value.

Bearing that in mind, I think that a "bad" value of partid_per_closid should not
be seen here, when the rest of the implementation is complete.

Does this sound sensible?

It could make sense to have a WARN_ON_ONCE() here anyway, though, since
this code is not on a fast path.
It sounds sensible.
quoted
quoted
-u32 resctrl_arch_system_num_rmid_idx(void)
+static void mpam_resctrl_partid_range(u32 closid, enum
+resctrl_conf_type
type,
+				      const struct rdt_resource *r,
+				      u16 *min_partid, u16 *max_partid)
 {
-	u8 closid_shift = fls(mpam_pmg_max);
-	u32 num_partid = resctrl_arch_get_num_closid(NULL);
+	u16 base_partid = closid;
+	u16 span = 1;
+
+	if (cdp_enabled) {
+		base_partid *= 2;
+		if (mpam_resctrl_hide_cdp(r->rid) ||
+		    type == CDP_NONE)
+			span *= 2;
+	}

-	return num_partid << closid_shift;
+	*min_partid = base_partid * partid_per_closid;
+	if (max_partid)
+		*max_partid = *min_partid + (span * partid_per_closid - 1);
 }
I think the min_partid/max_partid is different depending on the type
(CDP_DATA or CDP_CODE).
You're right!  Looking at my fixup patches from the end of December, it looks
like I already encountered this problem and wrote a workaround, but I didn't
post it at the time.

[...]
quoted
quoted
@@ -1190,22 +1219,13 @@ int resctrl_arch_update_one(struct
rdt_resource *r, struct rdt_ctrl_domain *d,
 		return -EINVAL;
 	}

-	/*
-	 * When CDP is enabled, but the resource doesn't support it, we need
to
-	 * apply the same configuration to the other partid.
-	 */
-	if (mpam_resctrl_hide_cdp(r->rid)) {
-		partid = resctrl_get_config_index(closid, CDP_CODE);
+	for (partid = min_partid; partid <= max_partid; partid++) {
 		err = mpam_apply_config(dom->comp, partid, &cfg);
 		if (err)
-			return err;
-
-		partid = resctrl_get_config_index(closid, CDP_DATA);
-		return mpam_apply_config(dom->comp, partid, &cfg);
-
-	} else {
-		return mpam_apply_config(dom->comp, partid, &cfg);
+			break;
 	}
+
+	return err;
 }
For a mixture of MSCs system, MSCs that do not support PARTID
narrowing and support Maximum Partition feature may not work properly.
This is true.

For this series, we just trust that the user understands what they are doing, but
the complete implementation will need additional checks.

My plan was to do the appropriate checks when the driver starts up, to decide
whether it is safe to set partid_per_closid to a value different from 1.

Once we get to this code, the decision has already been taken and so the value
of partid_per_closid (and the resulting mapping of resctrl groups onto PARTIDs)
can be trusted as being appropriate for the hardware.

Can you see any problems with this approach?
This approach sounds sensible.
[...]
quoted
I know this is not a complete implementation.
Mapping relationship between intPARTIDs and reqPARTIDs need to be
written to register MPAMCFG_INTPARTID/MPAMCFG_PART_SEL.

Best regards,
Shaopeng TAN
Agreed.  I have some prototype code for that, but I need to split up the patches
and clean it up before posting it.

Thanks for the review.
I will review the rest of the prototype code when it is posted.
 
Best regards,
Shaopeng TAN
Thread (11 messages) 11 messages, 2 authors, last reply 2025-01-08
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help