[RFC] [PATCH net-next v6 0/3] r8169: Implement dynamic ASPM mechanism for recent 1.0/2.5Gbps Realtek NICs

[RFC] [PATCH net-next v6 0/3] r8169: Implement dynamic ASPM mechanism for recent 1.0/2.5Gbps Realtek NICs

[Kai-Heng Feng]

The purpose of the series is to get comments and reviews so we can merge
and test the series in downstream kernel.

The latest Realtek vendor driver and its Windows driver implements a
feature called "dynamic ASPM" which can improve performance on it's
ethernet NICs.

Heiner Kallweit pointed out the potential root cause can be that the
buffer is to small for its ASPM exit latency.

So bring the dynamic ASPM to r8169 so we can have both nice performance
and powersaving at the same time.

For the slow/fast alternating traffic pattern, we'll need some real
world test to know if we need to lower the dynamic ASPM interval.

v5:
https://lore.kernel.org/netdev/20210916154417.664323-1-kai.heng.feng@canonical.com/

v4:
https://lore.kernel.org/netdev/20210827171452.217123-1-kai.heng.feng@canonical.com/

v3:
https://lore.kernel.org/netdev/20210819054542.608745-1-kai.heng.feng@canonical.com/

v2:
https://lore.kernel.org/netdev/20210812155341.817031-1-kai.heng.feng@canonical.com/

v1:
https://lore.kernel.org/netdev/20210803152823.515849-1-kai.heng.feng@canonical.com/

Kai-Heng Feng (3):
  PCI/ASPM: Introduce a new helper to report ASPM capability
  r8169: Enable chip-specific ASPM regardless of PCIe ASPM status
  r8169: Implement dynamic ASPM mechanism

 drivers/net/ethernet/realtek/r8169_main.c | 69 ++++++++++++++++++++---
 drivers/pci/pcie/aspm.c                   | 11 ++++
 include/linux/pci.h                       |  2 +
 3 files changed, 73 insertions(+), 9 deletions(-)

-- 
2.32.0

[RFC] [PATCH net-next v5 1/3] PCI/ASPM: Introduce a new helper to report ASPM capability

[Kai-Heng Feng]

Introduce a new helper, pcie_aspm_capable(), to report ASPM capability.

The user will be introduced by next patch.

Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
---
v6:
v5:
 - No change.

v4:
 - Report aspm_capable instead.

v3:
 - This is a new patch

 drivers/pci/pcie/aspm.c | 11 +++++++++++
 include/linux/pci.h     |  2 ++
 2 files changed, 13 insertions(+)

diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
index 013a47f587cea..788e7496f33b1 100644
--- a/drivers/pci/pcie/aspm.c
+++ b/drivers/pci/pcie/aspm.c
@@ -1201,6 +1201,17 @@ bool pcie_aspm_enabled(struct pci_dev *pdev)
 }
 EXPORT_SYMBOL_GPL(pcie_aspm_enabled);
 
+bool pcie_aspm_capable(struct pci_dev *pdev)
+{
+	struct pcie_link_state *link = pcie_aspm_get_link(pdev);
+
+	if (!link)
+		return false;
+
+	return link->aspm_capable;
+}
+EXPORT_SYMBOL_GPL(pcie_aspm_capable);
+
 static ssize_t aspm_attr_show_common(struct device *dev,
 				     struct device_attribute *attr,
 				     char *buf, u8 state)
diff --git a/include/linux/pci.h b/include/linux/pci.h
index cd8aa6fce2041..a17baa39141f4 100644
--- a/include/linux/pci.h
+++ b/include/linux/pci.h
@@ -1639,6 +1639,7 @@ int pci_disable_link_state_locked(struct pci_dev *pdev, int state);
 void pcie_no_aspm(void);
 bool pcie_aspm_support_enabled(void);
 bool pcie_aspm_enabled(struct pci_dev *pdev);
+bool pcie_aspm_capable(struct pci_dev *pdev);
 #else
 static inline int pci_disable_link_state(struct pci_dev *pdev, int state)
 { return 0; }
@@ -1647,6 +1648,7 @@ static inline int pci_disable_link_state_locked(struct pci_dev *pdev, int state)
 static inline void pcie_no_aspm(void) { }
 static inline bool pcie_aspm_support_enabled(void) { return false; }
 static inline bool pcie_aspm_enabled(struct pci_dev *pdev) { return false; }
+static inline bool pcie_aspm_capable(struct pci_dev *pdev) { return false; }
 #endif
 
 #ifdef CONFIG_PCIEAER
-- 
2.32.0

[RFC] [PATCH net-next v6 2/3] r8169: Enable chip-specific ASPM regardless of PCIe ASPM status

[Kai-Heng Feng]

To really enable ASPM on r8169 NICs, both standard PCIe ASPM and
chip-specific ASPM have to be enabled at the same time.

Since PCIe ASPM can be enabled or disabled vis sysfs and there's no
mechanism to notify driver about ASPM change, unconditionally enable
chip-specific ASPM to make ASPM really take into effect.

Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
---
v6:
 - Unconditionally enable chip-specific ASPM.

v5:
 - New patch.

 drivers/net/ethernet/realtek/r8169_main.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
index 0199914440abc..53936ebb3b3a6 100644
--- a/drivers/net/ethernet/realtek/r8169_main.c
+++ b/drivers/net/ethernet/realtek/r8169_main.c
@@ -622,7 +622,6 @@ struct rtl8169_private {
 	} wk;
 
 	unsigned supports_gmii:1;
-	unsigned aspm_manageable:1;
 	dma_addr_t counters_phys_addr;
 	struct rtl8169_counters *counters;
 	struct rtl8169_tc_offsets tc_offset;
@@ -2664,8 +2663,13 @@ static void rtl_enable_exit_l1(struct rtl8169_private *tp)
 
 static void rtl_hw_aspm_clkreq_enable(struct rtl8169_private *tp, bool enable)
 {
-	/* Don't enable ASPM in the chip if OS can't control ASPM */
-	if (enable && tp->aspm_manageable) {
+	struct pci_dev *pdev = tp->pci_dev;
+
+	/* Skip if PCIe ASPM isn't possible */
+	if (!pcie_aspm_support_enabled() || !pcie_aspm_capable(pdev))
+		return;
+
+	if (enable) {
 		RTL_W8(tp, Config5, RTL_R8(tp, Config5) | ASPM_en);
 		RTL_W8(tp, Config2, RTL_R8(tp, Config2) | ClkReqEn);
 	} else {
@@ -5272,8 +5276,7 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
 	/* Disable ASPM L1 as that cause random device stop working
 	 * problems as well as full system hangs for some PCIe devices users.
 	 */
-	rc = pci_disable_link_state(pdev, PCIE_LINK_STATE_L1);
-	tp->aspm_manageable = !rc;
+	pci_disable_link_state(pdev, PCIE_LINK_STATE_L1);
 
 	/* enable device (incl. PCI PM wakeup and hotplug setup) */
 	rc = pcim_enable_device(pdev);
-- 
2.32.0

[RFC] [PATCH net-next v6 3/3] r8169: Implement dynamic ASPM mechanism

[Kai-Heng Feng]

r8169 NICs on some platforms have abysmal speed when ASPM is enabled.
Same issue can be observed with older vendor drivers.

The issue is however solved by the latest vendor driver. There's a new
mechanism, which disables r8169's internal ASPM when the NIC traffic has
more than 10 packets per second, and vice versa. The possible reason for
this is likely because the buffer on the chip is too small for its ASPM
exit latency.

Realtek confirmed that all their PCIe LAN NICs, r8106, r8168 and r8125
use dynamic ASPM under Windows. So implement the same mechanism here to
resolve the issue.

Also introduce a lock to prevent race on accessing config registers.

Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=214307
Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
---
v6:
 - Wording change.
 - Add bugzilla link.

v5:
 - Split out aspm_manageable replacement as another patch.
 - Introduce a lock for lock_config_regs() and unlock_config_regs().

v4:
 - Squash two patches
 - Remove aspm_manageable and use pcie_aspm_capable()
   pcie_aspm_enabled() accordingly

v3:
 - Use msecs_to_jiffies() for delay time
 - Use atomic_t instead of mutex for bh
 - Mention the buffer size and ASPM exit latency in commit message

v2: 
 - Use delayed_work instead of timer_list to avoid interrupt context
 - Use mutex to serialize packet counter read/write
 - Wording change
 drivers/net/ethernet/realtek/r8169_main.c | 58 +++++++++++++++++++++--
 1 file changed, 53 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
index 53936ebb3b3a6..9c10a908c08fb 100644
--- a/drivers/net/ethernet/realtek/r8169_main.c
+++ b/drivers/net/ethernet/realtek/r8169_main.c
@@ -622,6 +622,11 @@ struct rtl8169_private {
 	} wk;
 
 	unsigned supports_gmii:1;
+	unsigned rtl_aspm_enabled:1;
+	struct delayed_work aspm_toggle;
+	atomic_t aspm_packet_count;
+	struct mutex config_lock;
+
 	dma_addr_t counters_phys_addr;
 	struct rtl8169_counters *counters;
 	struct rtl8169_tc_offsets tc_offset;
@@ -670,12 +675,14 @@ static inline struct device *tp_to_dev(struct rtl8169_private *tp)
 
 static void rtl_lock_config_regs(struct rtl8169_private *tp)
 {
+	mutex_lock(&tp->config_lock);
 	RTL_W8(tp, Cfg9346, Cfg9346_Lock);
 }
 
 static void rtl_unlock_config_regs(struct rtl8169_private *tp)
 {
 	RTL_W8(tp, Cfg9346, Cfg9346_Unlock);
+	mutex_unlock(&tp->config_lock);
 }
 
 static void rtl_pci_commit(struct rtl8169_private *tp)
@@ -2669,6 +2676,8 @@ static void rtl_hw_aspm_clkreq_enable(struct rtl8169_private *tp, bool enable)
 	if (!pcie_aspm_support_enabled() || !pcie_aspm_capable(pdev))
 		return;
 
+	tp->rtl_aspm_enabled = enable;
+
 	if (enable) {
 		RTL_W8(tp, Config5, RTL_R8(tp, Config5) | ASPM_en);
 		RTL_W8(tp, Config2, RTL_R8(tp, Config2) | ClkReqEn);
@@ -4407,6 +4416,7 @@ static void rtl_tx(struct net_device *dev, struct rtl8169_private *tp,
 
 	dirty_tx = tp->dirty_tx;
 
+	atomic_add(tp->cur_tx - dirty_tx, &tp->aspm_packet_count);
 	while (READ_ONCE(tp->cur_tx) != dirty_tx) {
 		unsigned int entry = dirty_tx % NUM_TX_DESC;
 		u32 status;
@@ -4551,6 +4561,8 @@ static int rtl_rx(struct net_device *dev, struct rtl8169_private *tp, int budget
 		rtl8169_mark_to_asic(desc);
 	}
 
+	atomic_add(count, &tp->aspm_packet_count);
+
 	return count;
 }
 
@@ -4658,8 +4670,39 @@ static int r8169_phy_connect(struct rtl8169_private *tp)
 	return 0;
 }
 
+#define ASPM_PACKET_THRESHOLD 10
+#define ASPM_TOGGLE_INTERVAL 1000
+
+static void rtl8169_aspm_toggle(struct work_struct *work)
+{
+	struct rtl8169_private *tp = container_of(work, struct rtl8169_private,
+						  aspm_toggle.work);
+	int packet_count;
+	bool enable;
+
+	packet_count = atomic_xchg(&tp->aspm_packet_count, 0);
+
+	if (pcie_aspm_enabled(tp->pci_dev)) {
+		enable = packet_count <= ASPM_PACKET_THRESHOLD;
+
+		if (tp->rtl_aspm_enabled != enable) {
+			rtl_unlock_config_regs(tp);
+			rtl_hw_aspm_clkreq_enable(tp, enable);
+			rtl_lock_config_regs(tp);
+		}
+	} else if (tp->rtl_aspm_enabled) {
+		rtl_unlock_config_regs(tp);
+		rtl_hw_aspm_clkreq_enable(tp, false);
+		rtl_lock_config_regs(tp);
+	}
+
+	schedule_delayed_work(&tp->aspm_toggle, msecs_to_jiffies(ASPM_TOGGLE_INTERVAL));
+}
+
 static void rtl8169_down(struct rtl8169_private *tp)
 {
+	cancel_delayed_work_sync(&tp->aspm_toggle);
+
 	/* Clear all task flags */
 	bitmap_zero(tp->wk.flags, RTL_FLAG_MAX);
 
@@ -4686,6 +4729,10 @@ static void rtl8169_up(struct rtl8169_private *tp)
 	rtl_reset_work(tp);
 
 	phy_start(tp->phydev);
+
+	/* pcie_aspm_capable may change after system resume */
+	if (pcie_aspm_support_enabled() && pcie_aspm_capable(tp->pci_dev))
+		schedule_delayed_work(&tp->aspm_toggle, 0);
 }
 
 static int rtl8169_close(struct net_device *dev)
@@ -5273,11 +5320,6 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
 	if (rc)
 		return rc;
 
-	/* Disable ASPM L1 as that cause random device stop working
-	 * problems as well as full system hangs for some PCIe devices users.
-	 */
-	pci_disable_link_state(pdev, PCIE_LINK_STATE_L1);
-
 	/* enable device (incl. PCI PM wakeup and hotplug setup) */
 	rc = pcim_enable_device(pdev);
 	if (rc < 0) {
@@ -5307,6 +5349,8 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
 		return rc;
 	}
 
+	mutex_init(&tp->config_lock);
+
 	tp->mmio_addr = pcim_iomap_table(pdev)[region];
 
 	xid = (RTL_R32(tp, TxConfig) >> 20) & 0xfcf;
@@ -5344,6 +5388,10 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
 
 	INIT_WORK(&tp->wk.work, rtl_task);
 
+	INIT_DELAYED_WORK(&tp->aspm_toggle, rtl8169_aspm_toggle);
+
+	atomic_set(&tp->aspm_packet_count, 0);
+
 	rtl_init_mac_address(tp);
 
 	dev->ethtool_ops = &rtl8169_ethtool_ops;
-- 
2.32.0

Re: [RFC] [PATCH net-next v6 3/3] r8169: Implement dynamic ASPM mechanism

[Bjorn Helgaas]

On Fri, Oct 08, 2021 at 12:15:52AM +0800, Kai-Heng Feng wrote:
> r8169 NICs on some platforms have abysmal speed when ASPM is enabled.
> Same issue can be observed with older vendor drivers.
> 
> The issue is however solved by the latest vendor driver. There's a new
> mechanism, which disables r8169's internal ASPM when the NIC traffic has
> more than 10 packets per second, and vice versa. The possible reason for
> this is likely because the buffer on the chip is too small for its ASPM
> exit latency.

Because the NIC works fine on some platforms with ASPM fully enabled,
I would describe this as a "workaround" for a bug where we don't know
the root cause, not a "solution".

> Realtek confirmed that all their PCIe LAN NICs, r8106, r8168 and r8125
> use dynamic ASPM under Windows. So implement the same mechanism here to
> resolve the issue.
> 
> Also introduce a lock to prevent race on accessing config registers.

Strictly speaking, the addition of the lock should be a separate patch
since it's not directly related to the ASPM change.

A little more below...

> Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=214307
> Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
> ---
> v6:
>  - Wording change.
>  - Add bugzilla link.
> 
> v5:
>  - Split out aspm_manageable replacement as another patch.
>  - Introduce a lock for lock_config_regs() and unlock_config_regs().
> 
> v4:
>  - Squash two patches
>  - Remove aspm_manageable and use pcie_aspm_capable()
>    pcie_aspm_enabled() accordingly
> 
> v3:
>  - Use msecs_to_jiffies() for delay time
>  - Use atomic_t instead of mutex for bh
>  - Mention the buffer size and ASPM exit latency in commit message
> 
> v2: 
>  - Use delayed_work instead of timer_list to avoid interrupt context
>  - Use mutex to serialize packet counter read/write
>  - Wording change
>  drivers/net/ethernet/realtek/r8169_main.c | 58 +++++++++++++++++++++--
>  1 file changed, 53 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
> index 53936ebb3b3a6..9c10a908c08fb 100644
> --- a/drivers/net/ethernet/realtek/r8169_main.c
> +++ b/drivers/net/ethernet/realtek/r8169_main.c
> @@ -622,6 +622,11 @@ struct rtl8169_private {
>  	} wk;
>  
>  	unsigned supports_gmii:1;
> +	unsigned rtl_aspm_enabled:1;
> +	struct delayed_work aspm_toggle;
> +	atomic_t aspm_packet_count;
> +	struct mutex config_lock;
> +
>  	dma_addr_t counters_phys_addr;
>  	struct rtl8169_counters *counters;
>  	struct rtl8169_tc_offsets tc_offset;
> @@ -670,12 +675,14 @@ static inline struct device *tp_to_dev(struct rtl8169_private *tp)
>  
>  static void rtl_lock_config_regs(struct rtl8169_private *tp)
>  {
> +	mutex_lock(&tp->config_lock);
>  	RTL_W8(tp, Cfg9346, Cfg9346_Lock);
>  }
>  
>  static void rtl_unlock_config_regs(struct rtl8169_private *tp)
>  {
>  	RTL_W8(tp, Cfg9346, Cfg9346_Unlock);
> +	mutex_unlock(&tp->config_lock);
>  }
>  
>  static void rtl_pci_commit(struct rtl8169_private *tp)
> @@ -2669,6 +2676,8 @@ static void rtl_hw_aspm_clkreq_enable(struct rtl8169_private *tp, bool enable)
>  	if (!pcie_aspm_support_enabled() || !pcie_aspm_capable(pdev))
>  		return;
>  
> +	tp->rtl_aspm_enabled = enable;
> +
>  	if (enable) {
>  		RTL_W8(tp, Config5, RTL_R8(tp, Config5) | ASPM_en);
>  		RTL_W8(tp, Config2, RTL_R8(tp, Config2) | ClkReqEn);
> @@ -4407,6 +4416,7 @@ static void rtl_tx(struct net_device *dev, struct rtl8169_private *tp,
>  
>  	dirty_tx = tp->dirty_tx;
>  
> +	atomic_add(tp->cur_tx - dirty_tx, &tp->aspm_packet_count);
>  	while (READ_ONCE(tp->cur_tx) != dirty_tx) {
>  		unsigned int entry = dirty_tx % NUM_TX_DESC;
>  		u32 status;
> @@ -4551,6 +4561,8 @@ static int rtl_rx(struct net_device *dev, struct rtl8169_private *tp, int budget
>  		rtl8169_mark_to_asic(desc);
>  	}
>  
> +	atomic_add(count, &tp->aspm_packet_count);
> +
>  	return count;
>  }
>  
> @@ -4658,8 +4670,39 @@ static int r8169_phy_connect(struct rtl8169_private *tp)
>  	return 0;
>  }
>  
> +#define ASPM_PACKET_THRESHOLD 10
> +#define ASPM_TOGGLE_INTERVAL 1000
> +
> +static void rtl8169_aspm_toggle(struct work_struct *work)
> +{
> +	struct rtl8169_private *tp = container_of(work, struct rtl8169_private,
> +						  aspm_toggle.work);
> +	int packet_count;
> +	bool enable;
> +
> +	packet_count = atomic_xchg(&tp->aspm_packet_count, 0);
> +
> +	if (pcie_aspm_enabled(tp->pci_dev)) {
> +		enable = packet_count <= ASPM_PACKET_THRESHOLD;
> +
> +		if (tp->rtl_aspm_enabled != enable) {
> +			rtl_unlock_config_regs(tp);
> +			rtl_hw_aspm_clkreq_enable(tp, enable);
> +			rtl_lock_config_regs(tp);
> +		}
> +	} else if (tp->rtl_aspm_enabled) {
> +		rtl_unlock_config_regs(tp);
> +		rtl_hw_aspm_clkreq_enable(tp, false);
> +		rtl_lock_config_regs(tp);
> +	}

IIUC the way the "dynamic ASPM" works is that rtl8169_aspm_toggle()
runs every second (1000ms).  If the NIC has sent or received fewer
than 10 packets in the last second, you make sure ASPM is enabled.  If
it has sent or received more than 10 packets, you disable ASPM.

Since the disable is done in rtl_hw_aspm_clkreq_enable() with
chip-specific registers, I suppose lspci and the like still show ASPM
as being enabled.  Not really a problem, I guess.

It looks like this disables ASPM completely, even though the NIC
apparently works correctly with L0s and L1.1 enabled, right?

I suppose that on the Intel system, if we enable ASPM, the link goes
to L1.2, and the NIC immediately receives 1000 packets in that second
before we can disable ASPM again, we probably drop a few packets?

Whereas on the AMD system, we probably *never* drop any packets even
with L1.2 enabled all the time?

And if we actually knew the root cause and could set the correct LTR
values or whatever is wrong on the Intel system, we probably wouldn't
need this dynamic scheme?

> +	schedule_delayed_work(&tp->aspm_toggle, msecs_to_jiffies(ASPM_TOGGLE_INTERVAL));
> +}
> +
>  static void rtl8169_down(struct rtl8169_private *tp)
>  {
> +	cancel_delayed_work_sync(&tp->aspm_toggle);
> +
>  	/* Clear all task flags */
>  	bitmap_zero(tp->wk.flags, RTL_FLAG_MAX);
>  
> @@ -4686,6 +4729,10 @@ static void rtl8169_up(struct rtl8169_private *tp)
>  	rtl_reset_work(tp);
>  
>  	phy_start(tp->phydev);
> +
> +	/* pcie_aspm_capable may change after system resume */
> +	if (pcie_aspm_support_enabled() && pcie_aspm_capable(tp->pci_dev))
> +		schedule_delayed_work(&tp->aspm_toggle, 0);
>  }
>  
>  static int rtl8169_close(struct net_device *dev)
> @@ -5273,11 +5320,6 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
>  	if (rc)
>  		return rc;
>  
> -	/* Disable ASPM L1 as that cause random device stop working
> -	 * problems as well as full system hangs for some PCIe devices users.
> -	 */
> -	pci_disable_link_state(pdev, PCIE_LINK_STATE_L1);
> -
>  	/* enable device (incl. PCI PM wakeup and hotplug setup) */
>  	rc = pcim_enable_device(pdev);
>  	if (rc < 0) {
> @@ -5307,6 +5349,8 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
>  		return rc;
>  	}
>  
> +	mutex_init(&tp->config_lock);
> +
>  	tp->mmio_addr = pcim_iomap_table(pdev)[region];
>  
>  	xid = (RTL_R32(tp, TxConfig) >> 20) & 0xfcf;
> @@ -5344,6 +5388,10 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
>  
>  	INIT_WORK(&tp->wk.work, rtl_task);
>  
> +	INIT_DELAYED_WORK(&tp->aspm_toggle, rtl8169_aspm_toggle);
> +
> +	atomic_set(&tp->aspm_packet_count, 0);
> +
>  	rtl_init_mac_address(tp);
>  
>  	dev->ethtool_ops = &rtl8169_ethtool_ops;
> -- 
> 2.32.0
> 

Re: [RFC] [PATCH net-next v6 3/3] r8169: Implement dynamic ASPM mechanism

[Kai-Heng Feng]

On Fri, Oct 8, 2021 at 3:11 AM Bjorn Helgaas <helgaas@kernel.org> wrote:
>
> On Fri, Oct 08, 2021 at 12:15:52AM +0800, Kai-Heng Feng wrote:
> > r8169 NICs on some platforms have abysmal speed when ASPM is enabled.
> > Same issue can be observed with older vendor drivers.
> >
> > The issue is however solved by the latest vendor driver. There's a new
> > mechanism, which disables r8169's internal ASPM when the NIC traffic has
> > more than 10 packets per second, and vice versa. The possible reason for
> > this is likely because the buffer on the chip is too small for its ASPM
> > exit latency.
>
> Because the NIC works fine on some platforms with ASPM fully enabled,
> I would describe this as a "workaround" for a bug where we don't know
> the root cause, not a "solution".

OK, will change the wording.

>
> > Realtek confirmed that all their PCIe LAN NICs, r8106, r8168 and r8125
> > use dynamic ASPM under Windows. So implement the same mechanism here to
> > resolve the issue.
> >
> > Also introduce a lock to prevent race on accessing config registers.
>
> Strictly speaking, the addition of the lock should be a separate patch
> since it's not directly related to the ASPM change.

Will separate it to another patch.

>
> A little more below...
>
> > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=214307
> > Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
> > ---
> > v6:
> >  - Wording change.
> >  - Add bugzilla link.
> >
> > v5:
> >  - Split out aspm_manageable replacement as another patch.
> >  - Introduce a lock for lock_config_regs() and unlock_config_regs().
> >
> > v4:
> >  - Squash two patches
> >  - Remove aspm_manageable and use pcie_aspm_capable()
> >    pcie_aspm_enabled() accordingly
> >
> > v3:
> >  - Use msecs_to_jiffies() for delay time
> >  - Use atomic_t instead of mutex for bh
> >  - Mention the buffer size and ASPM exit latency in commit message
> >
> > v2:
> >  - Use delayed_work instead of timer_list to avoid interrupt context
> >  - Use mutex to serialize packet counter read/write
> >  - Wording change
> >  drivers/net/ethernet/realtek/r8169_main.c | 58 +++++++++++++++++++++--
> >  1 file changed, 53 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
> > index 53936ebb3b3a6..9c10a908c08fb 100644
> > --- a/drivers/net/ethernet/realtek/r8169_main.c
> > +++ b/drivers/net/ethernet/realtek/r8169_main.c
> > @@ -622,6 +622,11 @@ struct rtl8169_private {
> >       } wk;
> >
> >       unsigned supports_gmii:1;
> > +     unsigned rtl_aspm_enabled:1;
> > +     struct delayed_work aspm_toggle;
> > +     atomic_t aspm_packet_count;
> > +     struct mutex config_lock;
> > +
> >       dma_addr_t counters_phys_addr;
> >       struct rtl8169_counters *counters;
> >       struct rtl8169_tc_offsets tc_offset;
> > @@ -670,12 +675,14 @@ static inline struct device *tp_to_dev(struct rtl8169_private *tp)
> >
> >  static void rtl_lock_config_regs(struct rtl8169_private *tp)
> >  {
> > +     mutex_lock(&tp->config_lock);
> >       RTL_W8(tp, Cfg9346, Cfg9346_Lock);
> >  }
> >
> >  static void rtl_unlock_config_regs(struct rtl8169_private *tp)
> >  {
> >       RTL_W8(tp, Cfg9346, Cfg9346_Unlock);
> > +     mutex_unlock(&tp->config_lock);
> >  }
> >
> >  static void rtl_pci_commit(struct rtl8169_private *tp)
> > @@ -2669,6 +2676,8 @@ static void rtl_hw_aspm_clkreq_enable(struct rtl8169_private *tp, bool enable)
> >       if (!pcie_aspm_support_enabled() || !pcie_aspm_capable(pdev))
> >               return;
> >
> > +     tp->rtl_aspm_enabled = enable;
> > +
> >       if (enable) {
> >               RTL_W8(tp, Config5, RTL_R8(tp, Config5) | ASPM_en);
> >               RTL_W8(tp, Config2, RTL_R8(tp, Config2) | ClkReqEn);
> > @@ -4407,6 +4416,7 @@ static void rtl_tx(struct net_device *dev, struct rtl8169_private *tp,
> >
> >       dirty_tx = tp->dirty_tx;
> >
> > +     atomic_add(tp->cur_tx - dirty_tx, &tp->aspm_packet_count);
> >       while (READ_ONCE(tp->cur_tx) != dirty_tx) {
> >               unsigned int entry = dirty_tx % NUM_TX_DESC;
> >               u32 status;
> > @@ -4551,6 +4561,8 @@ static int rtl_rx(struct net_device *dev, struct rtl8169_private *tp, int budget
> >               rtl8169_mark_to_asic(desc);
> >       }
> >
> > +     atomic_add(count, &tp->aspm_packet_count);
> > +
> >       return count;
> >  }
> >
> > @@ -4658,8 +4670,39 @@ static int r8169_phy_connect(struct rtl8169_private *tp)
> >       return 0;
> >  }
> >
> > +#define ASPM_PACKET_THRESHOLD 10
> > +#define ASPM_TOGGLE_INTERVAL 1000
> > +
> > +static void rtl8169_aspm_toggle(struct work_struct *work)
> > +{
> > +     struct rtl8169_private *tp = container_of(work, struct rtl8169_private,
> > +                                               aspm_toggle.work);
> > +     int packet_count;
> > +     bool enable;
> > +
> > +     packet_count = atomic_xchg(&tp->aspm_packet_count, 0);
> > +
> > +     if (pcie_aspm_enabled(tp->pci_dev)) {
> > +             enable = packet_count <= ASPM_PACKET_THRESHOLD;
> > +
> > +             if (tp->rtl_aspm_enabled != enable) {
> > +                     rtl_unlock_config_regs(tp);
> > +                     rtl_hw_aspm_clkreq_enable(tp, enable);
> > +                     rtl_lock_config_regs(tp);
> > +             }
> > +     } else if (tp->rtl_aspm_enabled) {
> > +             rtl_unlock_config_regs(tp);
> > +             rtl_hw_aspm_clkreq_enable(tp, false);
> > +             rtl_lock_config_regs(tp);
> > +     }
>
> IIUC the way the "dynamic ASPM" works is that rtl8169_aspm_toggle()
> runs every second (1000ms).  If the NIC has sent or received fewer
> than 10 packets in the last second, you make sure ASPM is enabled.  If
> it has sent or received more than 10 packets, you disable ASPM.

Yes, this is what this patch does.

>
> Since the disable is done in rtl_hw_aspm_clkreq_enable() with
> chip-specific registers, I suppose lspci and the like still show ASPM
> as being enabled.  Not really a problem, I guess.
>
> It looks like this disables ASPM completely, even though the NIC
> apparently works correctly with L0s and L1.1 enabled, right?

I've seen bug reports that ASPM L0s and L1.1 caused the NIC stops to working.
So dynamic ASPM strikes the right

>
> I suppose that on the Intel system, if we enable ASPM, the link goes
> to L1.2, and the NIC immediately receives 1000 packets in that second
> before we can disable ASPM again, we probably drop a few packets?
>
> Whereas on the AMD system, we probably *never* drop any packets even
> with L1.2 enabled all the time?

Yes and yes.

>
> And if we actually knew the root cause and could set the correct LTR
> values or whatever is wrong on the Intel system, we probably wouldn't
> need this dynamic scheme?

Because Realtek already implemented the dynamic ASPM workaround in
their Windows and Linux driver, they never bother to find the root
cause.
So we'll never know what really happens here.

Kai-Heng

>
> > +     schedule_delayed_work(&tp->aspm_toggle, msecs_to_jiffies(ASPM_TOGGLE_INTERVAL));
> > +}
> > +
> >  static void rtl8169_down(struct rtl8169_private *tp)
> >  {
> > +     cancel_delayed_work_sync(&tp->aspm_toggle);
> > +
> >       /* Clear all task flags */
> >       bitmap_zero(tp->wk.flags, RTL_FLAG_MAX);
> >
> > @@ -4686,6 +4729,10 @@ static void rtl8169_up(struct rtl8169_private *tp)
> >       rtl_reset_work(tp);
> >
> >       phy_start(tp->phydev);
> > +
> > +     /* pcie_aspm_capable may change after system resume */
> > +     if (pcie_aspm_support_enabled() && pcie_aspm_capable(tp->pci_dev))
> > +             schedule_delayed_work(&tp->aspm_toggle, 0);
> >  }
> >
> >  static int rtl8169_close(struct net_device *dev)
> > @@ -5273,11 +5320,6 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
> >       if (rc)
> >               return rc;
> >
> > -     /* Disable ASPM L1 as that cause random device stop working
> > -      * problems as well as full system hangs for some PCIe devices users.
> > -      */
> > -     pci_disable_link_state(pdev, PCIE_LINK_STATE_L1);
> > -
> >       /* enable device (incl. PCI PM wakeup and hotplug setup) */
> >       rc = pcim_enable_device(pdev);
> >       if (rc < 0) {
> > @@ -5307,6 +5349,8 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
> >               return rc;
> >       }
> >
> > +     mutex_init(&tp->config_lock);
> > +
> >       tp->mmio_addr = pcim_iomap_table(pdev)[region];
> >
> >       xid = (RTL_R32(tp, TxConfig) >> 20) & 0xfcf;
> > @@ -5344,6 +5388,10 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
> >
> >       INIT_WORK(&tp->wk.work, rtl_task);
> >
> > +     INIT_DELAYED_WORK(&tp->aspm_toggle, rtl8169_aspm_toggle);
> > +
> > +     atomic_set(&tp->aspm_packet_count, 0);
> > +
> >       rtl_init_mac_address(tp);
> >
> >       dev->ethtool_ops = &rtl8169_ethtool_ops;
> > --
> > 2.32.0
> >

Re: [RFC] [PATCH net-next v6 3/3] r8169: Implement dynamic ASPM mechanism

[Bjorn Helgaas]

On Fri, Oct 08, 2021 at 02:18:55PM +0800, Kai-Heng Feng wrote:
> On Fri, Oct 8, 2021 at 3:11 AM Bjorn Helgaas <helgaas@kernel.org> wrote:
> > On Fri, Oct 08, 2021 at 12:15:52AM +0800, Kai-Heng Feng wrote:
> > > r8169 NICs on some platforms have abysmal speed when ASPM is enabled.
> > > Same issue can be observed with older vendor drivers.
> > >
> > > The issue is however solved by the latest vendor driver. There's a new
> > > mechanism, which disables r8169's internal ASPM when the NIC traffic has
> > > more than 10 packets per second, and vice versa. The possible reason for
> > > this is likely because the buffer on the chip is too small for its ASPM
> > > exit latency.
> > > ...

> > I suppose that on the Intel system, if we enable ASPM, the link goes
> > to L1.2, and the NIC immediately receives 1000 packets in that second
> > before we can disable ASPM again, we probably drop a few packets?
> >
> > Whereas on the AMD system, we probably *never* drop any packets even
> > with L1.2 enabled all the time?
> 
> Yes and yes.

The fact that we drop some packets with dynamic ASPM on the Intel
system means we must be giving up some performance.

And I guess that on the AMD system, we should get full performance but
we must be using a little more power (probably unmeasurable) because
ASPM *could* be always enabled but dynamic ASPM disables it some of
the time.

> > And if we actually knew the root cause and could set the correct LTR
> > values or whatever is wrong on the Intel system, we probably wouldn't
> > need this dynamic scheme?
> 
> Because Realtek already implemented the dynamic ASPM workaround in
> their Windows and Linux driver, they never bother to find the root
> cause.
> So we'll never know what really happens here.

Looks like it.  Somebody with a PCIe analyzer could probably make
progress, but I agree, that doesn't seem likely.

Realtek no doubt has the equipment to do this, but apparently they
don't think it's worthwhile.  In their defense, the Linux ASPM code is
pretty impenetrable and there could be a problem there that causes or
contributes to this.

Bjorn

Re: [RFC] [PATCH net-next v5 1/3] PCI/ASPM: Introduce a new helper to report ASPM capability

[Bjorn Helgaas]

On Fri, Oct 08, 2021 at 12:15:50AM +0800, Kai-Heng Feng wrote:
> Introduce a new helper, pcie_aspm_capable(), to report ASPM capability.
> 
> The user will be introduced by next patch.
> 
> Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com>

Change subject to:

  PCI/ASPM: Add pcie_aspm_capable()

Acked-by: Bjorn Helgaas <bhelgaas@google.com>

> ---
> v6:
> v5:
>  - No change.
> 
> v4:
>  - Report aspm_capable instead.
> 
> v3:
>  - This is a new patch
> 
>  drivers/pci/pcie/aspm.c | 11 +++++++++++
>  include/linux/pci.h     |  2 ++
>  2 files changed, 13 insertions(+)
> 
> diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> index 013a47f587cea..788e7496f33b1 100644
> --- a/drivers/pci/pcie/aspm.c
> +++ b/drivers/pci/pcie/aspm.c
> @@ -1201,6 +1201,17 @@ bool pcie_aspm_enabled(struct pci_dev *pdev)
>  }
>  EXPORT_SYMBOL_GPL(pcie_aspm_enabled);
>  
> +bool pcie_aspm_capable(struct pci_dev *pdev)
> +{
> +	struct pcie_link_state *link = pcie_aspm_get_link(pdev);
> +
> +	if (!link)
> +		return false;
> +
> +	return link->aspm_capable;
> +}
> +EXPORT_SYMBOL_GPL(pcie_aspm_capable);
> +
>  static ssize_t aspm_attr_show_common(struct device *dev,
>  				     struct device_attribute *attr,
>  				     char *buf, u8 state)
> diff --git a/include/linux/pci.h b/include/linux/pci.h
> index cd8aa6fce2041..a17baa39141f4 100644
> --- a/include/linux/pci.h
> +++ b/include/linux/pci.h
> @@ -1639,6 +1639,7 @@ int pci_disable_link_state_locked(struct pci_dev *pdev, int state);
>  void pcie_no_aspm(void);
>  bool pcie_aspm_support_enabled(void);
>  bool pcie_aspm_enabled(struct pci_dev *pdev);
> +bool pcie_aspm_capable(struct pci_dev *pdev);
>  #else
>  static inline int pci_disable_link_state(struct pci_dev *pdev, int state)
>  { return 0; }
> @@ -1647,6 +1648,7 @@ static inline int pci_disable_link_state_locked(struct pci_dev *pdev, int state)
>  static inline void pcie_no_aspm(void) { }
>  static inline bool pcie_aspm_support_enabled(void) { return false; }
>  static inline bool pcie_aspm_enabled(struct pci_dev *pdev) { return false; }
> +static inline bool pcie_aspm_capable(struct pci_dev *pdev) { return false; }
>  #endif
>  
>  #ifdef CONFIG_PCIEAER
> -- 
> 2.32.0
> 

Re: [RFC] [PATCH net-next v6 3/3] r8169: Implement dynamic ASPM mechanism

[Kai-Heng Feng]

On Fri, Oct 8, 2021 at 9:58 PM Bjorn Helgaas <helgaas@kernel.org> wrote:
>
> On Fri, Oct 08, 2021 at 02:18:55PM +0800, Kai-Heng Feng wrote:
> > On Fri, Oct 8, 2021 at 3:11 AM Bjorn Helgaas <helgaas@kernel.org> wrote:
> > > On Fri, Oct 08, 2021 at 12:15:52AM +0800, Kai-Heng Feng wrote:
> > > > r8169 NICs on some platforms have abysmal speed when ASPM is enabled.
> > > > Same issue can be observed with older vendor drivers.
> > > >
> > > > The issue is however solved by the latest vendor driver. There's a new
> > > > mechanism, which disables r8169's internal ASPM when the NIC traffic has
> > > > more than 10 packets per second, and vice versa. The possible reason for
> > > > this is likely because the buffer on the chip is too small for its ASPM
> > > > exit latency.
> > > > ...
>
> > > I suppose that on the Intel system, if we enable ASPM, the link goes
> > > to L1.2, and the NIC immediately receives 1000 packets in that second
> > > before we can disable ASPM again, we probably drop a few packets?
> > >
> > > Whereas on the AMD system, we probably *never* drop any packets even
> > > with L1.2 enabled all the time?
> >
> > Yes and yes.
>
> The fact that we drop some packets with dynamic ASPM on the Intel
> system means we must be giving up some performance.
>
> And I guess that on the AMD system, we should get full performance but
> we must be using a little more power (probably unmeasurable) because
> ASPM *could* be always enabled but dynamic ASPM disables it some of
> the time.

Yes that's the case here.

>
> > > And if we actually knew the root cause and could set the correct LTR
> > > values or whatever is wrong on the Intel system, we probably wouldn't
> > > need this dynamic scheme?
> >
> > Because Realtek already implemented the dynamic ASPM workaround in
> > their Windows and Linux driver, they never bother to find the root
> > cause.
> > So we'll never know what really happens here.
>
> Looks like it.  Somebody with a PCIe analyzer could probably make
> progress, but I agree, that doesn't seem likely.
>
> Realtek no doubt has the equipment to do this, but apparently they
> don't think it's worthwhile.  In their defense, the Linux ASPM code is
> pretty impenetrable and there could be a problem there that causes or
> contributes to this.

I do hope they can put more effort on their ethernet driver like what
they do on their wireless drivers.

Kai-Heng

>
> Bjorn