| Message ID | 20210819054542.608745-2-kai.heng.feng@canonical.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | r8169: Implement dynamic ASPM mechanism for recent 1.0/2.5Gbps Realtek NICs | expand |
| Context | Check | Description |
|---|---|---|
| netdev/cover_letter | success | Link |
| netdev/fixes_present | success | Link |
| netdev/patch_count | success | Link |
| netdev/tree_selection | success | Clearly marked for net-next |
| netdev/subject_prefix | success | Link |
| netdev/cc_maintainers | success | CCed 5 of 5 maintainers |
| netdev/source_inline | success | Was 0 now: 0 |
| netdev/verify_signedoff | success | Link |
| netdev/module_param | success | Was 0 now: 0 |
| netdev/build_32bit | success | Errors and warnings before: 0 this patch: 0 |
| netdev/kdoc | success | Errors and warnings before: 0 this patch: 0 |
| netdev/verify_fixes | success | Link |
| netdev/checkpatch | warning | WARNING: line length of 88 exceeds 80 columns |
| netdev/build_allmodconfig_warn | success | Errors and warnings before: 0 this patch: 0 |
| netdev/header_inline | success | Link |
On Thu, Aug 19, 2021 at 01:45:40PM +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. On some platforms but not on others? Maybe the PCIe topology is a factor? Do you have bug reports with data, e.g., "lspci -vv" output? > 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, and vice versa. Presumably there's a time interval related to the 10 packets? For example, do you want to disable ASPM if 10 packets are received (or sent?) in a certain amount of time? > The possible reason for this is > likely because the buffer on the chip is too small for its ASPM exit > latency. Maybe this means the chip advertises incorrect exit latencies? If so, maybe a quirk could override that? > 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. What exactly is "dynamic ASPM"? I see Heiner's comment about this being intended only for a downstream kernel. But why? > Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com> > --- > 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 | 44 ++++++++++++++++++++++- > 1 file changed, 43 insertions(+), 1 deletion(-) > > diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c > index 7a69b468584a2..3359509c1c351 100644 > --- a/drivers/net/ethernet/realtek/r8169_main.c > +++ b/drivers/net/ethernet/realtek/r8169_main.c > @@ -624,6 +624,10 @@ struct rtl8169_private { > > unsigned supports_gmii:1; > unsigned aspm_manageable:1; > + unsigned rtl_aspm_enabled:1; > + struct delayed_work aspm_toggle; > + atomic_t aspm_packet_count; > + > dma_addr_t counters_phys_addr; > struct rtl8169_counters *counters; > struct rtl8169_tc_offsets tc_offset; > @@ -2665,8 +2669,13 @@ static void rtl_pcie_state_l2l3_disable(struct rtl8169_private *tp) > > static void rtl_hw_aspm_clkreq_enable(struct rtl8169_private *tp, bool enable) > { > + if (!tp->aspm_manageable && enable) > + return; > + > + tp->rtl_aspm_enabled = enable; > + > /* Don't enable ASPM in the chip if OS can't control ASPM */ > - if (enable && tp->aspm_manageable) { > + if (enable) { > RTL_W8(tp, Config5, RTL_R8(tp, Config5) | ASPM_en); > RTL_W8(tp, Config2, RTL_R8(tp, Config2) | ClkReqEn); > } else { > @@ -4415,6 +4424,7 @@ 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; > @@ -4559,6 +4569,8 @@ 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; > } > > @@ -4666,8 +4678,32 @@ 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); > + 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); > + } > + > + 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); > > @@ -4694,6 +4730,8 @@ static void rtl8169_up(struct rtl8169_private *tp) > rtl_reset_work(tp); > > phy_start(tp->phydev); > + > + schedule_delayed_work(&tp->aspm_toggle, msecs_to_jiffies(ASPM_TOGGLE_INTERVAL)); > } > > static int rtl8169_close(struct net_device *dev) > @@ -5354,6 +5392,10 @@ 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 >
On 19.08.2021 13:42, Bjorn Helgaas wrote: > On Thu, Aug 19, 2021 at 01:45:40PM +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. > > On some platforms but not on others? Maybe the PCIe topology is a > factor? Do you have bug reports with data, e.g., "lspci -vv" output? > >> 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, and vice versa. > > Presumably there's a time interval related to the 10 packets? For > example, do you want to disable ASPM if 10 packets are received (or > sent?) in a certain amount of time? > >> The possible reason for this is >> likely because the buffer on the chip is too small for its ASPM exit >> latency. > > Maybe this means the chip advertises incorrect exit latencies? If so, > maybe a quirk could override that? > >> 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. > > What exactly is "dynamic ASPM"? > > I see Heiner's comment about this being intended only for a downstream > kernel. But why? > We've seen various more or less obvious symptoms caused by the broken ASPM support on Realtek network chips. Unfortunately Realtek releases neither datasheets nor errata information. Last time we attempted to re-enable ASPM numerous problem reports came in. These Realtek chips are used on basically every consumer mainboard. The proposed workaround has potential side effects: In case of a congestion in the chip it may take up to a second until ASPM gets disabled, what may affect performance, especially in case of alternating traffic patterns. Also we can't expect support from Realtek. Having said that my decision was that it's too risky to re-enable ASPM in mainline even with this workaround in place. Kai-Heng weights the power saving higher and wants to take the risk in his downstream kernel. If there are no problems downstream after few months, then this workaround may make it to mainline. >> Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com> >> --- >> 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 | 44 ++++++++++++++++++++++- >> 1 file changed, 43 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c >> index 7a69b468584a2..3359509c1c351 100644 >> --- a/drivers/net/ethernet/realtek/r8169_main.c >> +++ b/drivers/net/ethernet/realtek/r8169_main.c >> @@ -624,6 +624,10 @@ struct rtl8169_private { >> >> unsigned supports_gmii:1; >> unsigned aspm_manageable:1; >> + unsigned rtl_aspm_enabled:1; >> + struct delayed_work aspm_toggle; >> + atomic_t aspm_packet_count; >> + >> dma_addr_t counters_phys_addr; >> struct rtl8169_counters *counters; >> struct rtl8169_tc_offsets tc_offset; >> @@ -2665,8 +2669,13 @@ static void rtl_pcie_state_l2l3_disable(struct rtl8169_private *tp) >> >> static void rtl_hw_aspm_clkreq_enable(struct rtl8169_private *tp, bool enable) >> { >> + if (!tp->aspm_manageable && enable) >> + return; >> + >> + tp->rtl_aspm_enabled = enable; >> + >> /* Don't enable ASPM in the chip if OS can't control ASPM */ >> - if (enable && tp->aspm_manageable) { >> + if (enable) { >> RTL_W8(tp, Config5, RTL_R8(tp, Config5) | ASPM_en); >> RTL_W8(tp, Config2, RTL_R8(tp, Config2) | ClkReqEn); >> } else { >> @@ -4415,6 +4424,7 @@ 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; >> @@ -4559,6 +4569,8 @@ 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; >> } >> >> @@ -4666,8 +4678,32 @@ 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); >> + 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); >> + } >> + >> + 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); >> >> @@ -4694,6 +4730,8 @@ static void rtl8169_up(struct rtl8169_private *tp) >> rtl_reset_work(tp); >> >> phy_start(tp->phydev); >> + >> + schedule_delayed_work(&tp->aspm_toggle, msecs_to_jiffies(ASPM_TOGGLE_INTERVAL)); >> } >> >> static int rtl8169_close(struct net_device *dev) >> @@ -5354,6 +5392,10 @@ 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 >>
On Thu, Aug 19, 2021 at 05:45:22PM +0200, Heiner Kallweit wrote: > On 19.08.2021 13:42, Bjorn Helgaas wrote: > > On Thu, Aug 19, 2021 at 01:45:40PM +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. > > > > On some platforms but not on others? Maybe the PCIe topology is a > > factor? Do you have bug reports with data, e.g., "lspci -vv" output? > > > >> 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, and vice versa. > > > > Presumably there's a time interval related to the 10 packets? For > > example, do you want to disable ASPM if 10 packets are received (or > > sent?) in a certain amount of time? > > > >> The possible reason for this is > >> likely because the buffer on the chip is too small for its ASPM exit > >> latency. > > > > Maybe this means the chip advertises incorrect exit latencies? If so, > > maybe a quirk could override that? > > > >> 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. > > > > What exactly is "dynamic ASPM"? > > > > I see Heiner's comment about this being intended only for a downstream > > kernel. But why? > > > We've seen various more or less obvious symptoms caused by the broken > ASPM support on Realtek network chips. Unfortunately Realtek releases > neither datasheets nor errata information. > Last time we attempted to re-enable ASPM numerous problem reports came > in. These Realtek chips are used on basically every consumer mainboard. > The proposed workaround has potential side effects: In case of a > congestion in the chip it may take up to a second until ASPM gets > disabled, what may affect performance, especially in case of alternating > traffic patterns. Also we can't expect support from Realtek. > Having said that my decision was that it's too risky to re-enable ASPM > in mainline even with this workaround in place. Kai-Heng weights the > power saving higher and wants to take the risk in his downstream kernel. > If there are no problems downstream after few months, then this > workaround may make it to mainline. Since ASPM apparently works well on some platforms but not others, I'd suspect some incorrect exit latencies. Ideally we'd have some launchpad/bugzilla links, and a better understanding of the problem, and maybe a quirk that makes this work on all platforms without mucking up the driver with ASPM tweaks. But I'm a little out of turn here because the only direct impact to the PCI core is the pcie_aspm_supported() interface. It *looks* like these patches don't actually touch the PCIe architected ASPM controls in Link Control; all I see is mucking with Realtek-specific registers. I think this is more work than it should be and likely to be not as reliable as it should be. But I guess that's up to you guys. Bjorn
On Sat, Aug 21, 2021 at 5:03 AM Bjorn Helgaas <helgaas@kernel.org> wrote: > > On Thu, Aug 19, 2021 at 05:45:22PM +0200, Heiner Kallweit wrote: > > On 19.08.2021 13:42, Bjorn Helgaas wrote: > > > On Thu, Aug 19, 2021 at 01:45:40PM +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. > > > > > > On some platforms but not on others? Maybe the PCIe topology is a > > > factor? Do you have bug reports with data, e.g., "lspci -vv" output? > > > > > >> 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, and vice versa. > > > > > > Presumably there's a time interval related to the 10 packets? For > > > example, do you want to disable ASPM if 10 packets are received (or > > > sent?) in a certain amount of time? > > > > > >> The possible reason for this is > > >> likely because the buffer on the chip is too small for its ASPM exit > > >> latency. > > > > > > Maybe this means the chip advertises incorrect exit latencies? If so, > > > maybe a quirk could override that? > > > > > >> 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. > > > > > > What exactly is "dynamic ASPM"? > > > > > > I see Heiner's comment about this being intended only for a downstream > > > kernel. But why? > > > > > We've seen various more or less obvious symptoms caused by the broken > > ASPM support on Realtek network chips. Unfortunately Realtek releases > > neither datasheets nor errata information. > > Last time we attempted to re-enable ASPM numerous problem reports came > > in. These Realtek chips are used on basically every consumer mainboard. > > The proposed workaround has potential side effects: In case of a > > congestion in the chip it may take up to a second until ASPM gets > > disabled, what may affect performance, especially in case of alternating > > traffic patterns. Also we can't expect support from Realtek. > > Having said that my decision was that it's too risky to re-enable ASPM > > in mainline even with this workaround in place. Kai-Heng weights the > > power saving higher and wants to take the risk in his downstream kernel. > > If there are no problems downstream after few months, then this > > workaround may make it to mainline. > > Since ASPM apparently works well on some platforms but not others, I'd > suspect some incorrect exit latencies. Can be, but if their dynamic ASPM mechanism can workaround the issue, maybe their hardware is just designed that way? > > Ideally we'd have some launchpad/bugzilla links, and a better > understanding of the problem, and maybe a quirk that makes this work > on all platforms without mucking up the driver with ASPM tweaks. The tweaks is OS-agnostic and is also implemented in Windows. > > But I'm a little out of turn here because the only direct impact to > the PCI core is the pcie_aspm_supported() interface. It *looks* like > these patches don't actually touch the PCIe architected ASPM controls > in Link Control; all I see is mucking with Realtek-specific registers. AFAICT, Realtek ethernet NIC and wireless NIC both have two layers of ASPM, one is the regular PCIe ASPM, and a Realtek specific internal ASPM. Both have to be enabled to really make ASPM work for them. Kai-Heng > > I think this is more work than it should be and likely to be not as > reliable as it should be. But I guess that's up to you guys. > > Bjorn
On Tue, Aug 24, 2021 at 03:39:35PM +0800, Kai-Heng Feng wrote: > On Sat, Aug 21, 2021 at 5:03 AM Bjorn Helgaas <helgaas@kernel.org> wrote: > > > > On Thu, Aug 19, 2021 at 05:45:22PM +0200, Heiner Kallweit wrote: > > > On 19.08.2021 13:42, Bjorn Helgaas wrote: > > > > On Thu, Aug 19, 2021 at 01:45:40PM +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. > > > > > > > > On some platforms but not on others? Maybe the PCIe topology is a > > > > factor? Do you have bug reports with data, e.g., "lspci -vv" output? > > > > > > > >> 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, and vice versa. > > > > > > > > Presumably there's a time interval related to the 10 packets? For > > > > example, do you want to disable ASPM if 10 packets are received (or > > > > sent?) in a certain amount of time? > > > > > > > >> The possible reason for this is > > > >> likely because the buffer on the chip is too small for its ASPM exit > > > >> latency. > > > > > > > > Maybe this means the chip advertises incorrect exit latencies? If so, > > > > maybe a quirk could override that? > > > > > > > >> 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. > > > > > > > > What exactly is "dynamic ASPM"? > > > > > > > > I see Heiner's comment about this being intended only for a downstream > > > > kernel. But why? > > > > > > > We've seen various more or less obvious symptoms caused by the broken > > > ASPM support on Realtek network chips. Unfortunately Realtek releases > > > neither datasheets nor errata information. > > > Last time we attempted to re-enable ASPM numerous problem reports came > > > in. These Realtek chips are used on basically every consumer mainboard. > > > The proposed workaround has potential side effects: In case of a > > > congestion in the chip it may take up to a second until ASPM gets > > > disabled, what may affect performance, especially in case of alternating > > > traffic patterns. Also we can't expect support from Realtek. > > > Having said that my decision was that it's too risky to re-enable ASPM > > > in mainline even with this workaround in place. Kai-Heng weights the > > > power saving higher and wants to take the risk in his downstream kernel. > > > If there are no problems downstream after few months, then this > > > workaround may make it to mainline. > > > > Since ASPM apparently works well on some platforms but not others, I'd > > suspect some incorrect exit latencies. > > Can be, but if their dynamic ASPM mechanism can workaround the issue, > maybe their hardware is just designed that way? Designed what way? You mean the hardware uses the architected ASPM control bits in the PCIe capability to control some ASPM functionality that doesn't work like the spec says it should work? > > Ideally we'd have some launchpad/bugzilla links, and a better > > understanding of the problem, and maybe a quirk that makes this work > > on all platforms without mucking up the driver with ASPM tweaks. > > The tweaks is OS-agnostic and is also implemented in Windows. I assume you mean these tweaks are also implemented in the Windows *driver* from Realtek. That's not a very convincing argument that this is the way it should work. If ASPM works well on some platforms, we should be able to make it work well on other platforms, too. The actual data ("lspci -vvxxx") from working and problematic platforms might have hints. > > But I'm a little out of turn here because the only direct impact to > > the PCI core is the pcie_aspm_supported() interface. It *looks* like > > these patches don't actually touch the PCIe architected ASPM controls > > in Link Control; all I see is mucking with Realtek-specific registers. > > AFAICT, Realtek ethernet NIC and wireless NIC both have two layers of > ASPM, one is the regular PCIe ASPM, and a Realtek specific internal > ASPM. > Both have to be enabled to really make ASPM work for them. It's common for devices to have chicken bits. But when a feature is enabled, it should work as defined by the PCIe spec so it will work with other spec-compliant devices. Bjorn
On Tue, Aug 24, 2021 at 10:53 PM Bjorn Helgaas <helgaas@kernel.org> wrote: > > On Tue, Aug 24, 2021 at 03:39:35PM +0800, Kai-Heng Feng wrote: > > On Sat, Aug 21, 2021 at 5:03 AM Bjorn Helgaas <helgaas@kernel.org> wrote: > > > > > > On Thu, Aug 19, 2021 at 05:45:22PM +0200, Heiner Kallweit wrote: > > > > On 19.08.2021 13:42, Bjorn Helgaas wrote: > > > > > On Thu, Aug 19, 2021 at 01:45:40PM +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. > > > > > > > > > > On some platforms but not on others? Maybe the PCIe topology is a > > > > > factor? Do you have bug reports with data, e.g., "lspci -vv" output? > > > > > > > > > >> 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, and vice versa. > > > > > > > > > > Presumably there's a time interval related to the 10 packets? For > > > > > example, do you want to disable ASPM if 10 packets are received (or > > > > > sent?) in a certain amount of time? > > > > > > > > > >> The possible reason for this is > > > > >> likely because the buffer on the chip is too small for its ASPM exit > > > > >> latency. > > > > > > > > > > Maybe this means the chip advertises incorrect exit latencies? If so, > > > > > maybe a quirk could override that? > > > > > > > > > >> 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. > > > > > > > > > > What exactly is "dynamic ASPM"? > > > > > > > > > > I see Heiner's comment about this being intended only for a downstream > > > > > kernel. But why? > > > > > > > > > We've seen various more or less obvious symptoms caused by the broken > > > > ASPM support on Realtek network chips. Unfortunately Realtek releases > > > > neither datasheets nor errata information. > > > > Last time we attempted to re-enable ASPM numerous problem reports came > > > > in. These Realtek chips are used on basically every consumer mainboard. > > > > The proposed workaround has potential side effects: In case of a > > > > congestion in the chip it may take up to a second until ASPM gets > > > > disabled, what may affect performance, especially in case of alternating > > > > traffic patterns. Also we can't expect support from Realtek. > > > > Having said that my decision was that it's too risky to re-enable ASPM > > > > in mainline even with this workaround in place. Kai-Heng weights the > > > > power saving higher and wants to take the risk in his downstream kernel. > > > > If there are no problems downstream after few months, then this > > > > workaround may make it to mainline. > > > > > > Since ASPM apparently works well on some platforms but not others, I'd > > > suspect some incorrect exit latencies. > > > > Can be, but if their dynamic ASPM mechanism can workaround the issue, > > maybe their hardware is just designed that way? > > Designed what way? You mean the hardware uses the architected ASPM > control bits in the PCIe capability to control some ASPM functionality > that doesn't work like the spec says it should work? Yes, it requires both standard PCIe ASPM control bits and Realtek specific register bits to make ASPM really work. Does PCI spec mandates PCIe config space to be the only way to enable ASPM? > > > > Ideally we'd have some launchpad/bugzilla links, and a better > > > understanding of the problem, and maybe a quirk that makes this work > > > on all platforms without mucking up the driver with ASPM tweaks. > > > > The tweaks is OS-agnostic and is also implemented in Windows. > > I assume you mean these tweaks are also implemented in the Windows > *driver* from Realtek. That's not a very convincing argument that > this is the way it should work. Since Realtek doesn't publish any erratum so following the driver tweaks is the most practical way to improve the situation under Linux. The same tweaks (i.e. dynamically enable/disable ASPM) can also be found in another driver, drivers/infiniband/hw/hfi1/aspm.c. > > If ASPM works well on some platforms, we should be able to make it > work well on other platforms, too. The actual data ("lspci -vvxxx") > from working and problematic platforms might have hints. OK, I'll ask affected users' lspci data. > > > > > But I'm a little out of turn here because the only direct impact to > > > the PCI core is the pcie_aspm_supported() interface. It *looks* like > > > these patches don't actually touch the PCIe architected ASPM controls > > > in Link Control; all I see is mucking with Realtek-specific registers. > > > > AFAICT, Realtek ethernet NIC and wireless NIC both have two layers of > > ASPM, one is the regular PCIe ASPM, and a Realtek specific internal > > ASPM. > > Both have to be enabled to really make ASPM work for them. > > It's common for devices to have chicken bits. But when a feature is > enabled, it should work as defined by the PCIe spec so it will work > with other spec-compliant devices. I have no idea why they designed ASPM in two layers. Only Realtek knows the reason... > > Bjorn
diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c index 7a69b468584a2..3359509c1c351 100644 --- a/drivers/net/ethernet/realtek/r8169_main.c +++ b/drivers/net/ethernet/realtek/r8169_main.c @@ -624,6 +624,10 @@ struct rtl8169_private { unsigned supports_gmii:1; unsigned aspm_manageable:1; + unsigned rtl_aspm_enabled:1; + struct delayed_work aspm_toggle; + atomic_t aspm_packet_count; + dma_addr_t counters_phys_addr; struct rtl8169_counters *counters; struct rtl8169_tc_offsets tc_offset; @@ -2665,8 +2669,13 @@ static void rtl_pcie_state_l2l3_disable(struct rtl8169_private *tp) static void rtl_hw_aspm_clkreq_enable(struct rtl8169_private *tp, bool enable) { + if (!tp->aspm_manageable && enable) + return; + + tp->rtl_aspm_enabled = enable; + /* Don't enable ASPM in the chip if OS can't control ASPM */ - if (enable && tp->aspm_manageable) { + if (enable) { RTL_W8(tp, Config5, RTL_R8(tp, Config5) | ASPM_en); RTL_W8(tp, Config2, RTL_R8(tp, Config2) | ClkReqEn); } else { @@ -4415,6 +4424,7 @@ 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; @@ -4559,6 +4569,8 @@ 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; } @@ -4666,8 +4678,32 @@ 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); + 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); + } + + 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); @@ -4694,6 +4730,8 @@ static void rtl8169_up(struct rtl8169_private *tp) rtl_reset_work(tp); phy_start(tp->phydev); + + schedule_delayed_work(&tp->aspm_toggle, msecs_to_jiffies(ASPM_TOGGLE_INTERVAL)); } static int rtl8169_close(struct net_device *dev) @@ -5354,6 +5392,10 @@ 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;
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, 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. Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com> --- 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 | 44 ++++++++++++++++++++++- 1 file changed, 43 insertions(+), 1 deletion(-)