[RFC PATCH v4 0/4] PCI/ASPM: Remove struct aspm_latency

[RFC PATCH v4 0/4] PCI/ASPM: Remove struct aspm_latency

[Saheed O. Bolarinwa]

From: "Bolarinwa O. Saheed" <refactormyself@gmail.com>

To validate and set link latency capability, `struct aspm_latency` and
related members defined within `struct pcie_link_state` are used.
However, since there are not many access to theses values, it is
possible to directly access and compute these values.

Doing this will also reduce the dependency on `struct pcie_link_state`.

The series removes `struct aspm_latency` and related members within
`struct pcie_link_state`. All latencies are now calculated when needed.


VERSION CHANGES:
- v2:
»       - directly access downstream by calling `pci_function_0()`
»         instead of using the `struct pcie_link_state`

- v3:
»       - rebase on Linux 5.15-rc2

- v4 (this version)
	- Create a seprate path to move pci_function_0() upward

MERGE NOTICE:
These series are based on
»       'commit e4e737bb5c17 ("Linux 5.15-rc2")'

Saheed O. Bolarinwa (4):
  PCI/ASPM: Move pci_function_0() upward
  PCI/ASPM: Do not cache link latencies
  PCI/ASPM: Remove struct pcie_link_state.acceptable
  PCI/ASPM: Remove struct aspm_latency

 drivers/pci/pcie/aspm.c | 88 +++++++++++++++++------------------------
 1 file changed, 37 insertions(+), 51 deletions(-)

-- 
2.20.1

[RFC PATCH v4 1/4] PCI/ASPM: Move pci_function_0() upward

[Saheed O. Bolarinwa]

From: "Bolarinwa O. Saheed" <refactormyself@gmail.com>

To call pci_function_0() directly from other functions,
move its definition upward to a more accessible location.

Signed-off-by: Bolarinwa O. Saheed <refactormyself@gmail.com>
---
 drivers/pci/pcie/aspm.c | 28 ++++++++++++++--------------
 1 file changed, 14 insertions(+), 14 deletions(-)

diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
index 013a47f587ce..a6d89c2c5b60 100644
--- a/drivers/pci/pcie/aspm.c
+++ b/drivers/pci/pcie/aspm.c
@@ -105,6 +105,20 @@ static const char *policy_str[] = {
 
 #define LINK_RETRAIN_TIMEOUT HZ
 
+/*
+ * The L1 PM substate capability is only implemented in function 0 in a
+ * multi function device.
+ */
+static struct pci_dev *pci_function_0(struct pci_bus *linkbus)
+{
+	struct pci_dev *child;
+
+	list_for_each_entry(child, &linkbus->devices, bus_list)
+		if (PCI_FUNC(child->devfn) == 0)
+			return child;
+	return NULL;
+}
+
 static int policy_to_aspm_state(struct pcie_link_state *link)
 {
 	switch (aspm_policy) {
@@ -423,20 +437,6 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 	}
 }
 
-/*
- * The L1 PM substate capability is only implemented in function 0 in a
- * multi function device.
- */
-static struct pci_dev *pci_function_0(struct pci_bus *linkbus)
-{
-	struct pci_dev *child;
-
-	list_for_each_entry(child, &linkbus->devices, bus_list)
-		if (PCI_FUNC(child->devfn) == 0)
-			return child;
-	return NULL;
-}
-
 static void pci_clear_and_set_dword(struct pci_dev *pdev, int pos,
 				    u32 clear, u32 set)
 {
-- 
2.20.1

[RFC PATCH v4 2/4] PCI/ASPM: Do not cache link latencies

[Saheed O. Bolarinwa]

The latencies of the upstream and downstream are calculated within
pcie_aspm_cap_init() and cached in struct pcie_link_state.latency_*
These values are only used in pcie_aspm_check_latency() where they are
compared with the acceptable latencies on the link.

- remove `latency_*` entries from struct pcie_link_state.
- calculate the latencies directly where they are needed.

Signed-off-by: Saheed O. Bolarinwa <refactormyself@gmail.com>
---
 drivers/pci/pcie/aspm.c | 25 ++++++++++++++-----------
 1 file changed, 14 insertions(+), 11 deletions(-)

diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
index a6d89c2c5b60..9e74df7b9dc0 100644
--- a/drivers/pci/pcie/aspm.c
+++ b/drivers/pci/pcie/aspm.c
@@ -66,9 +66,6 @@ struct pcie_link_state {
 	u32 clkpm_default:1;		/* Default Clock PM state by BIOS */
 	u32 clkpm_disable:1;		/* Clock PM disabled */
 
-	/* Exit latencies */
-	struct aspm_latency latency_up;	/* Upstream direction exit latency */
-	struct aspm_latency latency_dw;	/* Downstream direction exit latency */
 	/*
 	 * Endpoint acceptable latencies. A pcie downstream port only
 	 * has one slot under it, so at most there are 8 functions.
@@ -392,7 +389,8 @@ static void encode_l12_threshold(u32 threshold_us, u32 *scale, u32 *value)
 
 static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 {
-	u32 latency, l1_switch_latency = 0;
+	u32 latency, lnkcap_up, lnkcap_dw, l1_switch_latency = 0;
+	struct aspm_latency latency_up, latency_dw;
 	struct aspm_latency *acceptable;
 	struct pcie_link_state *link;
 
@@ -405,14 +403,23 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 	acceptable = &link->acceptable[PCI_FUNC(endpoint->devfn)];
 
 	while (link) {
+		/* Read direction exit latencies */
+		pcie_capability_read_dword(link->pdev, PCI_EXP_LNKCAP, &lnkcap_up);
+		pcie_capability_read_dword(pci_function_0(link->pdev->subordinate),
+					   PCI_EXP_LNKCAP, &lnkcap_dw);
+		latency_up.l0s = calc_l0s_latency(lnkcap_up);
+		latency_up.l1 = calc_l1_latency(lnkcap_up);
+		latency_dw.l0s = calc_l0s_latency(lnkcap_dw);
+		latency_dw.l1 = calc_l1_latency(lnkcap_dw);
+
 		/* Check upstream direction L0s latency */
 		if ((link->aspm_capable & ASPM_STATE_L0S_UP) &&
-		    (link->latency_up.l0s > acceptable->l0s))
+		    (latency_up.l0s > acceptable->l0s))
 			link->aspm_capable &= ~ASPM_STATE_L0S_UP;
 
 		/* Check downstream direction L0s latency */
 		if ((link->aspm_capable & ASPM_STATE_L0S_DW) &&
-		    (link->latency_dw.l0s > acceptable->l0s))
+		    (latency_dw.l0s > acceptable->l0s))
 			link->aspm_capable &= ~ASPM_STATE_L0S_DW;
 		/*
 		 * Check L1 latency.
@@ -427,7 +434,7 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 		 * L1 exit latencies advertised by a device include L1
 		 * substate latencies (and hence do not do any check).
 		 */
-		latency = max_t(u32, link->latency_up.l1, link->latency_dw.l1);
+		latency = max_t(u32, latency_up.l1, latency_dw.l1);
 		if ((link->aspm_capable & ASPM_STATE_L1) &&
 		    (latency + l1_switch_latency > acceptable->l1))
 			link->aspm_capable &= ~ASPM_STATE_L1;
@@ -593,8 +600,6 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)
 		link->aspm_enabled |= ASPM_STATE_L0S_UP;
 	if (parent_lnkctl & PCI_EXP_LNKCTL_ASPM_L0S)
 		link->aspm_enabled |= ASPM_STATE_L0S_DW;
-	link->latency_up.l0s = calc_l0s_latency(parent_lnkcap);
-	link->latency_dw.l0s = calc_l0s_latency(child_lnkcap);
 
 	/* Setup L1 state */
 	if (parent_lnkcap & child_lnkcap & PCI_EXP_LNKCAP_ASPM_L1)
@@ -602,8 +607,6 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)
 
 	if (parent_lnkctl & child_lnkctl & PCI_EXP_LNKCTL_ASPM_L1)
 		link->aspm_enabled |= ASPM_STATE_L1;
-	link->latency_up.l1 = calc_l1_latency(parent_lnkcap);
-	link->latency_dw.l1 = calc_l1_latency(child_lnkcap);
 
 	/* Setup L1 substate */
 	pci_read_config_dword(parent, parent->l1ss + PCI_L1SS_CAP,
-- 
2.20.1

[RFC PATCH v4 3/4] PCI/ASPM: Remove struct pcie_link_state.acceptable

[Saheed O. Bolarinwa]

The acceptable latencies for each device on the bus are calculated within
pcie_aspm_cap_init() and cached in struct pcie_link_state.acceptable.
They are only used within pcie_aspm_check_latency() to validate actual
latencies. Thus, it is possible to avoid caching these values.

 - remove `acceptable` from struct pcie_link_state
 - calculate the acceptable latency for individual device directly
 - remove the calculations done within pcie_aspm_cap_init()

Signed-off-by: Saheed O. Bolarinwa <refactormyself@gmail.com>
---
 drivers/pci/pcie/aspm.c | 27 ++++++++-------------------
 1 file changed, 8 insertions(+), 19 deletions(-)

diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
index 9e74df7b9dc0..6afbb86d07b8 100644
--- a/drivers/pci/pcie/aspm.c
+++ b/drivers/pci/pcie/aspm.c
@@ -65,12 +65,6 @@ struct pcie_link_state {
 	u32 clkpm_enabled:1;		/* Current Clock PM state */
 	u32 clkpm_default:1;		/* Default Clock PM state by BIOS */
 	u32 clkpm_disable:1;		/* Clock PM disabled */
-
-	/*
-	 * Endpoint acceptable latencies. A pcie downstream port only
-	 * has one slot under it, so at most there are 8 functions.
-	 */
-	struct aspm_latency acceptable[8];
 };
 
 static int aspm_disabled, aspm_force;
@@ -389,7 +383,7 @@ static void encode_l12_threshold(u32 threshold_us, u32 *scale, u32 *value)
 
 static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 {
-	u32 latency, lnkcap_up, lnkcap_dw, l1_switch_latency = 0;
+	u32 reg32, latency, encoding, lnkcap_up, lnkcap_dw, l1_switch_latency = 0;
 	struct aspm_latency latency_up, latency_dw;
 	struct aspm_latency *acceptable;
 	struct pcie_link_state *link;
@@ -400,7 +394,13 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 		return;
 
 	link = endpoint->bus->self->link_state;
-	acceptable = &link->acceptable[PCI_FUNC(endpoint->devfn)];
+	pcie_capability_read_dword(endpoint, PCI_EXP_DEVCAP, &reg32);
+	/* Calculate endpoint L0s acceptable latency */
+	encoding = (reg32 & PCI_EXP_DEVCAP_L0S) >> 6;
+	acceptable->l0s = calc_l0s_acceptable(encoding);
+	/* Calculate endpoint L1 acceptable latency */
+	encoding = (reg32 & PCI_EXP_DEVCAP_L1) >> 9;
+	acceptable->l1 = calc_l1_acceptable(encoding);
 
 	while (link) {
 		/* Read direction exit latencies */
@@ -663,22 +663,11 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)
 
 	/* Get and check endpoint acceptable latencies */
 	list_for_each_entry(child, &linkbus->devices, bus_list) {
-		u32 reg32, encoding;
-		struct aspm_latency *acceptable =
-			&link->acceptable[PCI_FUNC(child->devfn)];
 
 		if (pci_pcie_type(child) != PCI_EXP_TYPE_ENDPOINT &&
 		    pci_pcie_type(child) != PCI_EXP_TYPE_LEG_END)
 			continue;
 
-		pcie_capability_read_dword(child, PCI_EXP_DEVCAP, &reg32);
-		/* Calculate endpoint L0s acceptable latency */
-		encoding = (reg32 & PCI_EXP_DEVCAP_L0S) >> 6;
-		acceptable->l0s = calc_l0s_acceptable(encoding);
-		/* Calculate endpoint L1 acceptable latency */
-		encoding = (reg32 & PCI_EXP_DEVCAP_L1) >> 9;
-		acceptable->l1 = calc_l1_acceptable(encoding);
-
 		pcie_aspm_check_latency(child);
 	}
 }
-- 
2.20.1

[RFC PATCH v4 4/4] PCI/ASPM: Remove struct aspm_latency

[Saheed O. Bolarinwa]

The struct aspm_latency is now used only inside pcie_aspm_check_latency().

  - replace struct aspm_latency variables with u32 variables
  - remove struct aspm_latency

Signed-off-by: Saheed O. Bolarinwa <refactormyself@gmail.com>
---
 drivers/pci/pcie/aspm.c | 30 ++++++++++++------------------
 1 file changed, 12 insertions(+), 18 deletions(-)

diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
index 6afbb86d07b8..19459cb0af34 100644
--- a/drivers/pci/pcie/aspm.c
+++ b/drivers/pci/pcie/aspm.c
@@ -41,11 +41,6 @@
 #define ASPM_STATE_ALL		(ASPM_STATE_L0S | ASPM_STATE_L1 |	\
 				 ASPM_STATE_L1SS)
 
-struct aspm_latency {
-	u32 l0s;			/* L0s latency (nsec) */
-	u32 l1;				/* L1 latency (nsec) */
-};
-
 struct pcie_link_state {
 	struct pci_dev *pdev;		/* Upstream component of the Link */
 	struct pci_dev *downstream;	/* Downstream component, function 0 */
@@ -384,8 +379,8 @@ static void encode_l12_threshold(u32 threshold_us, u32 *scale, u32 *value)
 static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 {
 	u32 reg32, latency, encoding, lnkcap_up, lnkcap_dw, l1_switch_latency = 0;
-	struct aspm_latency latency_up, latency_dw;
-	struct aspm_latency *acceptable;
+	u32 latency_up_l0s, latency_up_l1, latency_dw_l0s, latency_dw_l1;
+	u32 acceptable_l0s, acceptable_l1;
 	struct pcie_link_state *link;
 
 	/* Device not in D0 doesn't need latency check */
@@ -397,29 +392,29 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 	pcie_capability_read_dword(endpoint, PCI_EXP_DEVCAP, &reg32);
 	/* Calculate endpoint L0s acceptable latency */
 	encoding = (reg32 & PCI_EXP_DEVCAP_L0S) >> 6;
-	acceptable->l0s = calc_l0s_acceptable(encoding);
+	acceptable_l0s = calc_l0s_acceptable(encoding);
 	/* Calculate endpoint L1 acceptable latency */
 	encoding = (reg32 & PCI_EXP_DEVCAP_L1) >> 9;
-	acceptable->l1 = calc_l1_acceptable(encoding);
+	acceptable_l1 = calc_l1_acceptable(encoding);
 
 	while (link) {
 		/* Read direction exit latencies */
 		pcie_capability_read_dword(link->pdev, PCI_EXP_LNKCAP, &lnkcap_up);
 		pcie_capability_read_dword(pci_function_0(link->pdev->subordinate),
 					   PCI_EXP_LNKCAP, &lnkcap_dw);
-		latency_up.l0s = calc_l0s_latency(lnkcap_up);
-		latency_up.l1 = calc_l1_latency(lnkcap_up);
-		latency_dw.l0s = calc_l0s_latency(lnkcap_dw);
-		latency_dw.l1 = calc_l1_latency(lnkcap_dw);
+		latency_up_l0s = calc_l0s_latency(lnkcap_up);
+		latency_up_l1 = calc_l1_latency(lnkcap_up);
+		latency_dw_l0s = calc_l0s_latency(lnkcap_dw);
+		latency_dw_l1 = calc_l1_latency(lnkcap_dw);
 
 		/* Check upstream direction L0s latency */
 		if ((link->aspm_capable & ASPM_STATE_L0S_UP) &&
-		    (latency_up.l0s > acceptable->l0s))
+		    (latency_up_l0s > acceptable_l0s))
 			link->aspm_capable &= ~ASPM_STATE_L0S_UP;
 
 		/* Check downstream direction L0s latency */
 		if ((link->aspm_capable & ASPM_STATE_L0S_DW) &&
-		    (latency_dw.l0s > acceptable->l0s))
+		    (latency_dw_l0s > acceptable_l0s))
 			link->aspm_capable &= ~ASPM_STATE_L0S_DW;
 		/*
 		 * Check L1 latency.
@@ -434,9 +429,9 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
 		 * L1 exit latencies advertised by a device include L1
 		 * substate latencies (and hence do not do any check).
 		 */
-		latency = max_t(u32, latency_up.l1, latency_dw.l1);
+		latency = max_t(u32, latency_up_l1, latency_dw_l1);
 		if ((link->aspm_capable & ASPM_STATE_L1) &&
-		    (latency + l1_switch_latency > acceptable->l1))
+		    (latency + l1_switch_latency > acceptable_l1))
 			link->aspm_capable &= ~ASPM_STATE_L1;
 		l1_switch_latency += 1000;
 
@@ -663,7 +658,6 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)
 
 	/* Get and check endpoint acceptable latencies */
 	list_for_each_entry(child, &linkbus->devices, bus_list) {
-
 		if (pci_pcie_type(child) != PCI_EXP_TYPE_ENDPOINT &&
 		    pci_pcie_type(child) != PCI_EXP_TYPE_LEG_END)
 			continue;
-- 
2.20.1

Re: [RFC PATCH v4 1/4] PCI/ASPM: Move pci_function_0() upward

[Christoph Hellwig]

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [RFC PATCH v4 2/4] PCI/ASPM: Do not cache link latencies

[Christoph Hellwig]

On Sat, Nov 06, 2021 at 06:53:03PM +0100, Saheed O. Bolarinwa wrote:
> The latencies of the upstream and downstream are calculated within
> pcie_aspm_cap_init() and cached in struct pcie_link_state.latency_*
> These values are only used in pcie_aspm_check_latency() where they are
> compared with the acceptable latencies on the link.
> 
> - remove `latency_*` entries from struct pcie_link_state.
> - calculate the latencies directly where they are needed.
> 
> Signed-off-by: Saheed O. Bolarinwa <refactormyself@gmail.com>
> ---
>  drivers/pci/pcie/aspm.c | 25 ++++++++++++++-----------
>  1 file changed, 14 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> index a6d89c2c5b60..9e74df7b9dc0 100644
> --- a/drivers/pci/pcie/aspm.c
> +++ b/drivers/pci/pcie/aspm.c
> @@ -66,9 +66,6 @@ struct pcie_link_state {
>  	u32 clkpm_default:1;		/* Default Clock PM state by BIOS */
>  	u32 clkpm_disable:1;		/* Clock PM disabled */
>  
> -	/* Exit latencies */
> -	struct aspm_latency latency_up;	/* Upstream direction exit latency */
> -	struct aspm_latency latency_dw;	/* Downstream direction exit latency */
>  	/*
>  	 * Endpoint acceptable latencies. A pcie downstream port only
>  	 * has one slot under it, so at most there are 8 functions.
> @@ -392,7 +389,8 @@ static void encode_l12_threshold(u32 threshold_us, u32 *scale, u32 *value)
>  
>  static void pcie_aspm_check_latency(struct pci_dev *endpoint)
>  {
> -	u32 latency, l1_switch_latency = 0;
> +	u32 latency, lnkcap_up, lnkcap_dw, l1_switch_latency = 0;
> +	struct aspm_latency latency_up, latency_dw;
>  	struct aspm_latency *acceptable;
>  	struct pcie_link_state *link;
>  
> @@ -405,14 +403,23 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
>  	acceptable = &link->acceptable[PCI_FUNC(endpoint->devfn)];
>  
>  	while (link) {
> +		/* Read direction exit latencies */
> +		pcie_capability_read_dword(link->pdev, PCI_EXP_LNKCAP, &lnkcap_up);
> +		pcie_capability_read_dword(pci_function_0(link->pdev->subordinate),

Please avoid the overly long lines.

Re: [RFC PATCH v4 3/4] PCI/ASPM: Remove struct pcie_link_state.acceptable

[Christoph Hellwig]

On Sat, Nov 06, 2021 at 06:53:04PM +0100, Saheed O. Bolarinwa wrote:
> +	u32 reg32, latency, encoding, lnkcap_up, lnkcap_dw, l1_switch_latency = 0;

Another unreadably long line, just split this up.

Re: [RFC PATCH v4 4/4] PCI/ASPM: Remove struct aspm_latency

[Christoph Hellwig]

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>