[PATCH v2 0/2] arch_topology: Pre-allocate cacheinfo from primary CPU

[PATCH v2 0/2] arch_topology: Pre-allocate cacheinfo from primary CPU

[Radu Rendec]

Commit 5944ce092b97 ("arch_topology: Build cacheinfo from primary CPU")
tries to build the cacheinfo from the primary CPU prior to secondary
CPUs boot, if the DT/ACPI description contains cache information.
However, if such information is not present, it still reverts to the old
behavior, which allocates the cacheinfo memory on each secondary CPU. On
RT kernels, this triggers a "BUG: sleeping function called from invalid
context" because the allocation is done before preemption is first
enabled on the secondary CPU.

The solution is to add cache information to DT/ACPI, but at least on
arm64 systems this can be avoided by leveraging automatic detection
(through the CLIDR_EL1 register), which is already implemented but
currently doesn't work on RT kernels for the reason described above.

This patch series attempts to enable automatic detection for RT kernels
when no DT/ACPI cache information is available, by pre-allocating
cacheinfo memory on the primary CPU.

The first patch adds an architecture independent infrastructure that
allows architecture specific code to take an early guess at the number
of cache leaves of the secodary CPUs, while it runs in preemptible
context on the primary CPU. At the same time, it gives architecture
specific code the opportunity to go back later, while it runs on the
secondary CPU, and reallocate the cacheinfo memory if the initial guess
proves to be wrong.

The second patch leverages the infrastructure implemented in the first
patch and enables early cache depth detection for arm64.

The patch series is based on an RFC patch that was posted to the
linux-arm-kernel mailing list and discussed with a smaller audience:
https://lore.kernel.org/all/20230323224242.31142-1-rrendec@redhat.com/

Radu Rendec (2):
  cacheinfo: Add arch specific early level initializer
  cacheinfo: Add arm64 early level initializer implementation

 arch/arm64/kernel/cacheinfo.c | 32 +++++++++++++++-----
 drivers/base/cacheinfo.c      | 57 +++++++++++++++++++++++------------
 include/linux/cacheinfo.h     |  2 ++
 3 files changed, 64 insertions(+), 27 deletions(-)

-- 
2.39.2

[PATCH v2 1/2] cacheinfo: Add arch specific early level initializer

[Radu Rendec]

This patch gives of architecture specific code the ability to initialize
the cache level and allocate cacheinfo memory early, when cache level
initialization runs on the primary CPU for all possible CPUs.

This is part of a patch series that attempts to further the work in
commit 5944ce092b97 ("arch_topology: Build cacheinfo from primary CPU").
Previously, in the absence of any DT/ACPI cache info, architecture
specific cache detection and info allocation for secondary CPUs would
happen in non-preemptible context during early CPU initialization and
trigger a "BUG: sleeping function called from invalid context" splat on
an RT kernel.

More specifically, this patch adds the early_cache_level() function,
which is called by fetch_cache_info() as a fallback when the number of
cache leaves cannot be extracted from DT/ACPI. In the default generic
(weak) implementation, this new function returns -ENOENT, which
preserves the original behavior for architectures that do not implement
the function.

Since early detection can get the number of cache leaves wrong in some
cases*, additional logic is added to still call init_cache_level() later
on the secondary CPU, therefore giving the architecture specific code an
opportunity to go back and fix the initial guess. Again, the original
behavior is preserved for architectures that do not implement the new
function.

* For example, on arm64, CLIDR_EL1 detection works only when it runs on
  the current CPU. In other words, a CPU cannot detect the cache depth
  for any other CPU than itself.

Signed-off-by: Radu Rendec <rrendec@redhat.com>
---
 drivers/base/cacheinfo.c  | 57 ++++++++++++++++++++++++++-------------
 include/linux/cacheinfo.h |  2 ++
 2 files changed, 40 insertions(+), 19 deletions(-)

diff --git a/drivers/base/cacheinfo.c b/drivers/base/cacheinfo.c
index f6573c335f4c..7f8ac0cb549f 100644
--- a/drivers/base/cacheinfo.c
+++ b/drivers/base/cacheinfo.c
@@ -398,6 +398,11 @@ static void free_cache_attributes(unsigned int cpu)
 	cache_shared_cpu_map_remove(cpu);
 }
 
+int __weak early_cache_level(unsigned int cpu)
+{
+	return -ENOENT;
+}
+
 int __weak init_cache_level(unsigned int cpu)
 {
 	return -ENOENT;
@@ -423,51 +428,65 @@ int allocate_cache_info(int cpu)
 
 int fetch_cache_info(unsigned int cpu)
 {
-	struct cpu_cacheinfo *this_cpu_ci;
+	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
 	unsigned int levels = 0, split_levels = 0;
 	int ret;
 
-	if (acpi_disabled) {
+	if (acpi_disabled)
 		ret = init_of_cache_level(cpu);
-		if (ret < 0)
-			return ret;
-	} else {
+	else {
 		ret = acpi_get_cache_info(cpu, &levels, &split_levels);
-		if (ret < 0)
+		if (!ret) {
+			this_cpu_ci->num_levels = levels;
+			/*
+			 * This assumes that:
+			 * - there cannot be any split caches (data/instruction)
+			 *   above a unified cache
+			 * - data/instruction caches come by pair
+			 */
+			this_cpu_ci->num_leaves = levels + split_levels;
+		}
+	}
+
+	if (ret || !cache_leaves(cpu)) {
+		ret = early_cache_level(cpu);
+		if (ret)
 			return ret;
 
-		this_cpu_ci = get_cpu_cacheinfo(cpu);
-		this_cpu_ci->num_levels = levels;
-		/*
-		 * This assumes that:
-		 * - there cannot be any split caches (data/instruction)
-		 *   above a unified cache
-		 * - data/instruction caches come by pair
-		 */
-		this_cpu_ci->num_leaves = levels + split_levels;
+		if (!cache_leaves(cpu))
+			return -ENOENT;
+
+		this_cpu_ci->early_arch_info = true;
 	}
-	if (!cache_leaves(cpu))
-		return -ENOENT;
 
 	return allocate_cache_info(cpu);
 }
 
 int detect_cache_attributes(unsigned int cpu)
 {
+	unsigned int early_leaves = cache_leaves(cpu);
 	int ret;
 
 	/* Since early initialization/allocation of the cacheinfo is allowed
 	 * via fetch_cache_info() and this also gets called as CPU hotplug
 	 * callbacks via cacheinfo_cpu_online, the init/alloc can be skipped
 	 * as it will happen only once (the cacheinfo memory is never freed).
-	 * Just populate the cacheinfo.
+	 * Just populate the cacheinfo. However, if the cacheinfo has been
+	 * allocated early through the arch-specific early_cache_level() call,
+	 * there is a chance the info is wrong (this can happen on arm64). In
+	 * that case, call init_cache_level() anyway to give the arch-specific
+	 * code a chance to make things right.
 	 */
-	if (per_cpu_cacheinfo(cpu))
+	if (per_cpu_cacheinfo(cpu) && !ci_cacheinfo(cpu)->early_arch_info)
 		goto populate_leaves;
 
 	if (init_cache_level(cpu) || !cache_leaves(cpu))
 		return -ENOENT;
 
+	if (cache_leaves(cpu) <= early_leaves)
+		goto populate_leaves;
+
+	kfree(per_cpu_cacheinfo(cpu));
 	ret = allocate_cache_info(cpu);
 	if (ret)
 		return ret;
diff --git a/include/linux/cacheinfo.h b/include/linux/cacheinfo.h
index 908e19d17f49..c9d44308fc42 100644
--- a/include/linux/cacheinfo.h
+++ b/include/linux/cacheinfo.h
@@ -76,9 +76,11 @@ struct cpu_cacheinfo {
 	unsigned int num_levels;
 	unsigned int num_leaves;
 	bool cpu_map_populated;
+	bool early_arch_info;
 };
 
 struct cpu_cacheinfo *get_cpu_cacheinfo(unsigned int cpu);
+int early_cache_level(unsigned int cpu);
 int init_cache_level(unsigned int cpu);
 int init_of_cache_level(unsigned int cpu);
 int populate_cache_leaves(unsigned int cpu);
-- 
2.39.2

[PATCH v2 2/2] cacheinfo: Add arm64 early level initializer implementation

[Radu Rendec]

This patch adds an architecture specific early cache level detection
handler for arm64. This is basically the CLIDR_EL1 based detection that
was previously done (only) in init_cache_level().

This is part of a patch series that attempts to further the work in
commit 5944ce092b97 ("arch_topology: Build cacheinfo from primary CPU").
Previously, in the absence of any DT/ACPI cache info, architecture
specific cache detection and info allocation for secondary CPUs would
happen in non-preemptible context during early CPU initialization and
trigger a "BUG: sleeping function called from invalid context" splat on
an RT kernel.

This patch does not solve the problem completely for RT kernels. It
relies on the assumption that on most systems, the CPUs are symmetrical
and therefore have the same number of cache leaves. The cacheinfo memory
is allocated early (on the primary CPU), relying on the new handler. If
later (when CLIDR_EL1 based detection runs again on the secondary CPU)
the initial assumption proves to be wrong and the CPU has in fact more
leaves, the cacheinfo memory is reallocated, and that still triggers a
splat on an RT kernel.

In other words, asymmetrical CPU systems *must* still provide cacheinfo
data in DT/ACPI to avoid the splat on RT kernels (unless secondary CPUs
happen to have less leaves than the primary CPU). But symmetrical CPU
systems (the majority) can now get away without the additional DT/ACPI
data and rely on CLIDR_EL1 based detection.

Signed-off-by: Radu Rendec <rrendec@redhat.com>
---
 arch/arm64/kernel/cacheinfo.c | 32 ++++++++++++++++++++++++--------
 1 file changed, 24 insertions(+), 8 deletions(-)

diff --git a/arch/arm64/kernel/cacheinfo.c b/arch/arm64/kernel/cacheinfo.c
index c307f69e9b55..520d17e4ebe9 100644
--- a/arch/arm64/kernel/cacheinfo.c
+++ b/arch/arm64/kernel/cacheinfo.c
@@ -38,21 +38,37 @@ static void ci_leaf_init(struct cacheinfo *this_leaf,
 	this_leaf->type = type;
 }
 
-int init_cache_level(unsigned int cpu)
+static void detect_cache_level(unsigned int *level, unsigned int *leaves)
 {
-	unsigned int ctype, level, leaves;
-	int fw_level, ret;
-	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
+	unsigned int ctype;
 
-	for (level = 1, leaves = 0; level <= MAX_CACHE_LEVEL; level++) {
-		ctype = get_cache_type(level);
+	for (*level = 1, *leaves = 0; *level <= MAX_CACHE_LEVEL; (*level)++) {
+		ctype = get_cache_type(*level);
 		if (ctype == CACHE_TYPE_NOCACHE) {
-			level--;
+			(*level)--;
 			break;
 		}
 		/* Separate instruction and data caches */
-		leaves += (ctype == CACHE_TYPE_SEPARATE) ? 2 : 1;
+		*leaves += (ctype == CACHE_TYPE_SEPARATE) ? 2 : 1;
 	}
+}
+
+int early_cache_level(unsigned int cpu)
+{
+	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
+
+	detect_cache_level(&this_cpu_ci->num_levels, &this_cpu_ci->num_leaves);
+
+	return 0;
+}
+
+int init_cache_level(unsigned int cpu)
+{
+	unsigned int level, leaves;
+	int fw_level, ret;
+	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
+
+	detect_cache_level(&level, &leaves);
 
 	if (acpi_disabled) {
 		fw_level = of_find_last_cache_level(cpu);
-- 
2.39.2

Re: [PATCH v2 1/2] cacheinfo: Add arch specific early level initializer

[Pierre Gondois]

Hello Radu,

On 4/4/23 01:15, Radu Rendec wrote:
> This patch gives of architecture specific code the ability to initialize
> the cache level and allocate cacheinfo memory early, when cache level
> initialization runs on the primary CPU for all possible CPUs.
> 
> This is part of a patch series that attempts to further the work in
> commit 5944ce092b97 ("arch_topology: Build cacheinfo from primary CPU").
> Previously, in the absence of any DT/ACPI cache info, architecture
> specific cache detection and info allocation for secondary CPUs would
> happen in non-preemptible context during early CPU initialization and
> trigger a "BUG: sleeping function called from invalid context" splat on
> an RT kernel.
> 
> More specifically, this patch adds the early_cache_level() function,
> which is called by fetch_cache_info() as a fallback when the number of
> cache leaves cannot be extracted from DT/ACPI. In the default generic
> (weak) implementation, this new function returns -ENOENT, which
> preserves the original behavior for architectures that do not implement
> the function.
> 
> Since early detection can get the number of cache leaves wrong in some
> cases*, additional logic is added to still call init_cache_level() later
> on the secondary CPU, therefore giving the architecture specific code an
> opportunity to go back and fix the initial guess. Again, the original
> behavior is preserved for architectures that do not implement the new
> function.
> 
> * For example, on arm64, CLIDR_EL1 detection works only when it runs on
>    the current CPU. In other words, a CPU cannot detect the cache depth
>    for any other CPU than itself.
> 
> Signed-off-by: Radu Rendec <rrendec@redhat.com>
> ---
>   drivers/base/cacheinfo.c  | 57 ++++++++++++++++++++++++++-------------
>   include/linux/cacheinfo.h |  2 ++
>   2 files changed, 40 insertions(+), 19 deletions(-)
> 
> diff --git a/drivers/base/cacheinfo.c b/drivers/base/cacheinfo.c
> index f6573c335f4c..7f8ac0cb549f 100644
> --- a/drivers/base/cacheinfo.c
> +++ b/drivers/base/cacheinfo.c
> @@ -398,6 +398,11 @@ static void free_cache_attributes(unsigned int cpu)
>   	cache_shared_cpu_map_remove(cpu);
>   }
>   
> +int __weak early_cache_level(unsigned int cpu)
> +{
> +	return -ENOENT;
> +}
> +
>   int __weak init_cache_level(unsigned int cpu)
>   {
>   	return -ENOENT;
> @@ -423,51 +428,65 @@ int allocate_cache_info(int cpu)
>   
>   int fetch_cache_info(unsigned int cpu)
>   {
> -	struct cpu_cacheinfo *this_cpu_ci;
> +	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
>   	unsigned int levels = 0, split_levels = 0;
>   	int ret;
>   
> -	if (acpi_disabled) {
> +	if (acpi_disabled)
>   		ret = init_of_cache_level(cpu);
> -		if (ret < 0)
> -			return ret;
> -	} else {
> +	else {

NIT: I think braces should be used in the first branch if they
are used in the second branch, cf. './scripts/checkpatch.pl --strict'

>   		ret = acpi_get_cache_info(cpu, &levels, &split_levels);
> -		if (ret < 0)
> +		if (!ret) {
> +			this_cpu_ci->num_levels = levels;
> +			/*
> +			 * This assumes that:
> +			 * - there cannot be any split caches (data/instruction)
> +			 *   above a unified cache
> +			 * - data/instruction caches come by pair
> +			 */
> +			this_cpu_ci->num_leaves = levels + split_levels;
> +		}
> +	}
> +
> +	if (ret || !cache_leaves(cpu)) {
> +		ret = early_cache_level(cpu);
> +		if (ret)
>   			return ret;
>   
> -		this_cpu_ci = get_cpu_cacheinfo(cpu);
> -		this_cpu_ci->num_levels = levels;
> -		/*
> -		 * This assumes that:
> -		 * - there cannot be any split caches (data/instruction)
> -		 *   above a unified cache
> -		 * - data/instruction caches come by pair
> -		 */
> -		this_cpu_ci->num_leaves = levels + split_levels;
> +		if (!cache_leaves(cpu))
> +			return -ENOENT;
> +
> +		this_cpu_ci->early_arch_info = true;
>   	}
> -	if (!cache_leaves(cpu))
> -		return -ENOENT;
>   
>   	return allocate_cache_info(cpu);
>   }
>   
>   int detect_cache_attributes(unsigned int cpu)
>   {
[start]
> +	unsigned int early_leaves = cache_leaves(cpu);
>   	int ret;
>   
>   	/* Since early initialization/allocation of the cacheinfo is allowed
>   	 * via fetch_cache_info() and this also gets called as CPU hotplug
>   	 * callbacks via cacheinfo_cpu_online, the init/alloc can be skipped
>   	 * as it will happen only once (the cacheinfo memory is never freed).
> -	 * Just populate the cacheinfo.
> +	 * Just populate the cacheinfo. However, if the cacheinfo has been
> +	 * allocated early through the arch-specific early_cache_level() call,
> +	 * there is a chance the info is wrong (this can happen on arm64). In
> +	 * that case, call init_cache_level() anyway to give the arch-specific
> +	 * code a chance to make things right.
>   	 */
> -	if (per_cpu_cacheinfo(cpu))
> +	if (per_cpu_cacheinfo(cpu) && !ci_cacheinfo(cpu)->early_arch_info)
>   		goto populate_leaves;
>   
>   	if (init_cache_level(cpu) || !cache_leaves(cpu))
>   		return -ENOENT;
>   
> +	if (cache_leaves(cpu) <= early_leaves)
> +		goto populate_leaves;
> +
> +	kfree(per_cpu_cacheinfo(cpu))>   	ret = allocate_cache_info(cpu);
>   	if (ret)
>   		return ret;
[stop]

Maybe this would be the occasion to put the code between the start/stop in a separate
function and remove the 'populate_leaves' label. The code seems correct, but it was
already a bit complex to read before the patch.


> diff --git a/include/linux/cacheinfo.h b/include/linux/cacheinfo.h
> index 908e19d17f49..c9d44308fc42 100644
> --- a/include/linux/cacheinfo.h
> +++ b/include/linux/cacheinfo.h
> @@ -76,9 +76,11 @@ struct cpu_cacheinfo {
>   	unsigned int num_levels;
>   	unsigned int num_leaves;
>   	bool cpu_map_populated;
> +	bool early_arch_info;
>   };
>   
>   struct cpu_cacheinfo *get_cpu_cacheinfo(unsigned int cpu);
> +int early_cache_level(unsigned int cpu);
>   int init_cache_level(unsigned int cpu);
>   int init_of_cache_level(unsigned int cpu);
>   int populate_cache_leaves(unsigned int cpu);

Re: [PATCH v2 1/2] cacheinfo: Add arch specific early level initializer

[Radu Rendec]

Hello Pierre,

On Thu, 2023-04-06 at 10:17 +0200, Pierre Gondois wrote:
> [start]
> > +       unsigned int early_leaves = cache_leaves(cpu);
> >         int ret;
> >   
> >         /* Since early initialization/allocation of the cacheinfo is allowed
> >          * via fetch_cache_info() and this also gets called as CPU hotplug
> >          * callbacks via cacheinfo_cpu_online, the init/alloc can be skipped
> >          * as it will happen only once (the cacheinfo memory is never freed).
> > -        * Just populate the cacheinfo.
> > +        * Just populate the cacheinfo. However, if the cacheinfo has been
> > +        * allocated early through the arch-specific early_cache_level() call,
> > +        * there is a chance the info is wrong (this can happen on arm64). In
> > +        * that case, call init_cache_level() anyway to give the arch-specific
> > +        * code a chance to make things right.
> >          */
> > -       if (per_cpu_cacheinfo(cpu))
> > +       if (per_cpu_cacheinfo(cpu) && !ci_cacheinfo(cpu)->early_arch_info)
> >                 goto populate_leaves;
> >   
> >         if (init_cache_level(cpu) || !cache_leaves(cpu))
> >                 return -ENOENT;
> >   
> > +       if (cache_leaves(cpu) <= early_leaves)
> > +               goto populate_leaves;
> > +
> > +       kfree(per_cpu_cacheinfo(cpu))
> > 
> >         ret = allocate_cache_info(cpu);
> >         if (ret)
> >                 return ret;
> [stop]
> 
> Maybe this would be the occasion to put the code between the start/stop in a separate
> function and remove the 'populate_leaves' label. The code seems correct, but it was
> already a bit complex to read before the patch.

Yes, that makes sense. I will address this (and your other comment
about the braces) and post v3 shortly. Thanks for reviewing the patch
and for the feedback!

After I had sent v2, I realized there was something missing from that
code between start/stop. I think we should also set the early_arch_info
flag back to false to prevent another detection/reallocation in case
detect_cache_attributes() is called again (I'm thinking CPU hotplug).
I will address this as well in v3.

Best regards,
Radu