| Message ID | 20240717205511.2541693-4-wei.huang2@amd.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Bjorn Helgaas |
| Headers | show |
| Series | PCIe TPH and cache direct injection support | expand |
On Wed, Jul 17, 2024 at 03:55:04PM -0500, Wei Huang wrote: > TLP headers with incorrect steering tags (e.g. caused by buggy driver) > can potentially cause issues when the system hardware consumes the tags. Hmm. What kind of issues? Crash? Data corruption? Poor performance? > Provide a kernel option, with related helper functions, to completely > prevent TPH from being enabled. Also would be nice to have a hint about the difference between "notph" and "nostmode". Maybe that goes in the "nostmode" patch? I'm not super clear on all the differences here. > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -4655,6 +4655,7 @@ > nomio [S390] Do not use MIO instructions. > norid [S390] ignore the RID field and force use of > one PCI domain per PCI function > + notph [PCIE] Do not use PCIe TPH Expand acronym here since there's no helpful context. Can also include "(TPH)" if that's useful. > @@ -322,8 +323,12 @@ static long local_pci_probe(void *_ddi) > pm_runtime_get_sync(dev); > pci_dev->driver = pci_drv; > rc = pci_drv->probe(pci_dev, ddi->id); > - if (!rc) > + if (!rc) { > + if (pci_tph_disabled()) > + pcie_tph_disable(pci_dev); I'm not really a fan of cluttering probe() like this. Can't we disable it in pcie_tph_init() so all devices start off with TPH disabled, and then check pci_tph_disabled() in whatever interface drivers use to enable TPH? > +bool pci_tph_disabled(void) > +{ > + return pcie_tph_disabled; > +} > +EXPORT_SYMBOL_GPL(pci_tph_disabled); Other related interfaces use "pcie" prefix; I think this should match. Do drivers need this? Would be nice not to export it unless they do. Bjorn
On 7/17/24 21:55, Wei Huang wrote: > TLP headers with incorrect steering tags (e.g. caused by buggy driver) > can potentially cause issues when the system hardware consumes the tags. > Provide a kernel option, with related helper functions, to completely > prevent TPH from being enabled. Maybe rephrase it for including a potential buggy device, including the cpu. Also, what about handling this with a no-tph-allow device list instead of a generic binary option for the whole system? Foreseeing some buggy or poor-performance implementations, or specific use cases where it could be counterproductive, maybe supporting both options. > Co-developed-by: Eric Van Tassell <Eric.VanTassell@amd.com> > Signed-off-by: Eric Van Tassell <Eric.VanTassell@amd.com> > Signed-off-by: Wei Huang <wei.huang2@amd.com> > Reviewed-by: Ajit Khaparde <ajit.khaparde@broadcom.com> > Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com> > Reviewed-by: Andy Gospodarek <andrew.gospodarek@broadcom.com> > --- > .../admin-guide/kernel-parameters.txt | 1 + > drivers/pci/pci-driver.c | 7 +++++- > drivers/pci/pci.c | 12 +++++++++ > drivers/pci/pcie/tph.c | 25 +++++++++++++++++++ > include/linux/pci-tph.h | 18 +++++++++++++ > include/linux/pci.h | 1 + > 6 files changed, 63 insertions(+), 1 deletion(-) > create mode 100644 include/linux/pci-tph.h > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > index b2057241ea6c..65581ebd9b50 100644 > --- a/Documentation/admin-guide/kernel-parameters.txt > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -4655,6 +4655,7 @@ > nomio [S390] Do not use MIO instructions. > norid [S390] ignore the RID field and force use of > one PCI domain per PCI function > + notph [PCIE] Do not use PCIe TPH > > pcie_aspm= [PCIE] Forcibly enable or ignore PCIe Active State Power > Management. > diff --git a/drivers/pci/pci-driver.c b/drivers/pci/pci-driver.c > index af2996d0d17f..9722d070c0ca 100644 > --- a/drivers/pci/pci-driver.c > +++ b/drivers/pci/pci-driver.c > @@ -21,6 +21,7 @@ > #include <linux/acpi.h> > #include <linux/dma-map-ops.h> > #include <linux/iommu.h> > +#include <linux/pci-tph.h> > #include "pci.h" > #include "pcie/portdrv.h" > > @@ -322,8 +323,12 @@ static long local_pci_probe(void *_ddi) > pm_runtime_get_sync(dev); > pci_dev->driver = pci_drv; > rc = pci_drv->probe(pci_dev, ddi->id); > - if (!rc) > + if (!rc) { > + if (pci_tph_disabled()) > + pcie_tph_disable(pci_dev); > + > return rc; > + } > if (rc < 0) { > pci_dev->driver = NULL; > pm_runtime_put_sync(dev); > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > index 02b1d81b1419..4cbfd5b53be8 100644 > --- a/drivers/pci/pci.c > +++ b/drivers/pci/pci.c > @@ -157,6 +157,9 @@ static bool pcie_ari_disabled; > /* If set, the PCIe ATS capability will not be used. */ > static bool pcie_ats_disabled; > > +/* If set, the PCIe TPH capability will not be used. */ > +static bool pcie_tph_disabled; > + > /* If set, the PCI config space of each device is printed during boot. */ > bool pci_early_dump; > > @@ -166,6 +169,12 @@ bool pci_ats_disabled(void) > } > EXPORT_SYMBOL_GPL(pci_ats_disabled); > > +bool pci_tph_disabled(void) > +{ > + return pcie_tph_disabled; > +} > +EXPORT_SYMBOL_GPL(pci_tph_disabled); > + > /* Disable bridge_d3 for all PCIe ports */ > static bool pci_bridge_d3_disable; > /* Force bridge_d3 for all PCIe ports */ > @@ -6869,6 +6878,9 @@ static int __init pci_setup(char *str) > pci_no_domains(); > } else if (!strncmp(str, "noari", 5)) { > pcie_ari_disabled = true; > + } else if (!strcmp(str, "notph")) { > + pr_info("PCIe: TPH is disabled\n"); > + pcie_tph_disabled = true; > } else if (!strncmp(str, "cbiosize=", 9)) { > pci_cardbus_io_size = memparse(str + 9, &str); > } else if (!strncmp(str, "cbmemsize=", 10)) { > diff --git a/drivers/pci/pcie/tph.c b/drivers/pci/pcie/tph.c > index e385b871333e..ad58a892792c 100644 > --- a/drivers/pci/pcie/tph.c > +++ b/drivers/pci/pcie/tph.c > @@ -7,8 +7,33 @@ > * Wei Huang <wei.huang2@amd.com> > */ > > +#include <linux/pci.h> > +#include <linux/bitfield.h> > +#include <linux/pci-tph.h> > + > #include "../pci.h" > > +/* Update the TPH Requester Enable field of TPH Control Register */ > +static void set_ctrl_reg_req_en(struct pci_dev *pdev, u8 req_type) > +{ > + u32 reg_val; > + > + pci_read_config_dword(pdev, pdev->tph_cap + PCI_TPH_CTRL, ®_val); > + > + reg_val &= ~PCI_TPH_CTRL_REQ_EN_MASK; > + reg_val |= FIELD_PREP(PCI_TPH_CTRL_REQ_EN_MASK, req_type); > + > + pci_write_config_dword(pdev, pdev->tph_cap + PCI_TPH_CTRL, reg_val); > +} > + > +void pcie_tph_disable(struct pci_dev *pdev) > +{ > + if (!pdev->tph_cap) > + return; > + > + set_ctrl_reg_req_en(pdev, PCI_TPH_REQ_DISABLE); > +} > + > void pcie_tph_init(struct pci_dev *pdev) > { > pdev->tph_cap = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_TPH); > diff --git a/include/linux/pci-tph.h b/include/linux/pci-tph.h > new file mode 100644 > index 000000000000..e0b782bda929 > --- /dev/null > +++ b/include/linux/pci-tph.h > @@ -0,0 +1,18 @@ > +/* SPDX-License-Identifier: GPL-2.0 */ > +/* > + * TPH (TLP Processing Hints) > + * > + * Copyright (C) 2024 Advanced Micro Devices, Inc. > + * Eric Van Tassell <Eric.VanTassell@amd.com> > + * Wei Huang <wei.huang2@amd.com> > + */ > +#ifndef LINUX_PCI_TPH_H > +#define LINUX_PCI_TPH_H > + > +#ifdef CONFIG_PCIE_TPH > +void pcie_tph_disable(struct pci_dev *dev); > +#else > +static inline void pcie_tph_disable(struct pci_dev *dev) {} > +#endif > + > +#endif /* LINUX_PCI_TPH_H */ > diff --git a/include/linux/pci.h b/include/linux/pci.h > index 6631ebe80ca9..05fbbd9ad6b4 100644 > --- a/include/linux/pci.h > +++ b/include/linux/pci.h > @@ -1840,6 +1840,7 @@ static inline bool pci_aer_available(void) { return false; } > #endif > > bool pci_ats_disabled(void); > +bool pci_tph_disabled(void); > > #ifdef CONFIG_PCIE_PTM > int pci_enable_ptm(struct pci_dev *dev, u8 *granularity);
On Wed, Jul 24, 2024 at 03:45:34PM +0100, Alejandro Lucero Palau wrote: > On 7/17/24 21:55, Wei Huang wrote: > > TLP headers with incorrect steering tags (e.g. caused by buggy driver) > > can potentially cause issues when the system hardware consumes the tags. > > Provide a kernel option, with related helper functions, to completely > > prevent TPH from being enabled. > > Maybe rephrase it for including a potential buggy device, including the cpu. > > Also, what about handling this with a no-tph-allow device list instead of a > generic binary option for the whole system? > > Foreseeing some buggy or poor-performance implementations, or specific use > cases where it could be counterproductive, maybe supporting both options. Makes sense if/when we need it. IMO no point in adding an empty list of known-broken devices.
On 7/23/24 17:41, Bjorn Helgaas wrote: > On Wed, Jul 17, 2024 at 03:55:04PM -0500, Wei Huang wrote: >> TLP headers with incorrect steering tags (e.g. caused by buggy driver) >> can potentially cause issues when the system hardware consumes the tags. > > Hmm. What kind of issues? Crash? Data corruption? Poor > performance? Not crash or functionality errors. Usually it is QoS related because of resource competition. AMD has > >> Provide a kernel option, with related helper functions, to completely >> prevent TPH from being enabled. > > Also would be nice to have a hint about the difference between "notph" > and "nostmode". Maybe that goes in the "nostmode" patch? I'm not > super clear on all the differences here. I can combine them. Here is the combination and it meaning based on TPH Control Register values: Requestor Enable | ST Mode | Meaning --------------------------------------------------------------- 00 | xx | TPH disabled (i.e. notph) 01 | 00 | TPH enabled, NO ST Mode (i.e. nostmode) 01 or 11 | 01 | Interrupt Vector mode 01 or 11 | 10 | Device specific mode If you have any other thoughts on how to approach these modes, please let me know. > >> +++ b/Documentation/admin-guide/kernel-parameters.txt >> @@ -4655,6 +4655,7 @@ >> nomio [S390] Do not use MIO instructions. >> norid [S390] ignore the RID field and force use of >> one PCI domain per PCI function >> + notph [PCIE] Do not use PCIe TPH > > Expand acronym here since there's no helpful context. Can also > include "(TPH)" if that's useful. > >> @@ -322,8 +323,12 @@ static long local_pci_probe(void *_ddi) >> pm_runtime_get_sync(dev); >> pci_dev->driver = pci_drv; >> rc = pci_drv->probe(pci_dev, ddi->id); >> - if (!rc) >> + if (!rc) { >> + if (pci_tph_disabled()) >> + pcie_tph_disable(pci_dev); > > I'm not really a fan of cluttering probe() like this. Can't we > disable it in pcie_tph_init() so all devices start off with TPH > disabled, and then check pci_tph_disabled() in whatever interface > drivers use to enable TPH? > >> +bool pci_tph_disabled(void) >> +{ >> + return pcie_tph_disabled; >> +} >> +EXPORT_SYMBOL_GPL(pci_tph_disabled); > > Other related interfaces use "pcie" prefix; I think this should match. > > Do drivers need this? Would be nice not to export it unless they do. > > Bjorn
On 7/24/24 10:36, Bjorn Helgaas wrote: > On Wed, Jul 24, 2024 at 03:45:34PM +0100, Alejandro Lucero Palau wrote: >> On 7/17/24 21:55, Wei Huang wrote: >>> TLP headers with incorrect steering tags (e.g. caused by buggy driver) >>> can potentially cause issues when the system hardware consumes the tags. >>> Provide a kernel option, with related helper functions, to completely >>> prevent TPH from being enabled. >> >> Maybe rephrase it for including a potential buggy device, including the cpu. >> >> Also, what about handling this with a no-tph-allow device list instead of a >> generic binary option for the whole system? >> >> Foreseeing some buggy or poor-performance implementations, or specific use >> cases where it could be counterproductive, maybe supporting both options. > > Makes sense if/when we need it. IMO no point in adding an empty list > of known-broken devices. We can add quirks.c, or something similar later, after TPH support is enabled.
On Wed, Jul 24, 2024 at 03:05:59PM -0500, Wei Huang wrote: > > > On 7/23/24 17:41, Bjorn Helgaas wrote: > > On Wed, Jul 17, 2024 at 03:55:04PM -0500, Wei Huang wrote: > >> TLP headers with incorrect steering tags (e.g. caused by buggy driver) > >> can potentially cause issues when the system hardware consumes the tags. > > > > Hmm. What kind of issues? Crash? Data corruption? Poor > > performance? > > Not crash or functionality errors. Usually it is QoS related because of > resource competition. AMD has Looks like you had more to say here? I *assume* that both the PH hint and the Steering Tags are only *hints* and there's no excuse for hardware to corrupt anything (e.g., by omitting cache maintenance) even if the hint turns out to be wrong. If that's the case, I assume "can potentially cause issues" really just means "might lead to lower performance". That's what I want to clarify and confirm. > >> Provide a kernel option, with related helper functions, to completely > >> prevent TPH from being enabled. > > > > Also would be nice to have a hint about the difference between "notph" > > and "nostmode". Maybe that goes in the "nostmode" patch? I'm not > > super clear on all the differences here. > > I can combine them. Here is the combination and it meaning based on TPH > Control Register values: > > Requestor Enable | ST Mode | Meaning > --------------------------------------------------------------- > 00 | xx | TPH disabled (i.e. notph) > 01 | 00 | TPH enabled, NO ST Mode (i.e. nostmode) > 01 or 11 | 01 | Interrupt Vector mode > 01 or 11 | 10 | Device specific mode > > If you have any other thoughts on how to approach these modes, please > let me know. IIRC, there's no interface in this series that reall does anything with TPH per se; drivers would only use the ST-related things. If that's the case, maybe "pci=notph" isn't needed yet. Bjorn
On 7/25/24 16:29, Bjorn Helgaas wrote: > On Wed, Jul 24, 2024 at 03:05:59PM -0500, Wei Huang wrote: >> >> >> On 7/23/24 17:41, Bjorn Helgaas wrote: >>> On Wed, Jul 17, 2024 at 03:55:04PM -0500, Wei Huang wrote: >>>> TLP headers with incorrect steering tags (e.g. caused by buggy driver) >>>> can potentially cause issues when the system hardware consumes the tags. >>> >>> Hmm. What kind of issues? Crash? Data corruption? Poor >>> performance? >> >> Not crash or functionality errors. Usually it is QoS related because of >> resource competition. AMD has > > Looks like you had more to say here? I hit the send button too fast. What I wanted to say was there will be AMD QoS patches to control TPH. Note that they will be hooked up under x86/resctrl. Since they are AMD specific, it will be independent from PCIe subsystem code. > > I *assume* that both the PH hint and the Steering Tags are only > *hints* and there's no excuse for hardware to corrupt anything (e.g., > by omitting cache maintenance) even if the hint turns out to be wrong. > If that's the case, I assume "can potentially cause issues" really > just means "might lead to lower performance". That's what I want to > clarify and confirm. Corrrect, only QoS-related concerns. There won't be any correctness concerns. > >>>> Provide a kernel option, with related helper functions, to completely >>>> prevent TPH from being enabled. >>> >>> Also would be nice to have a hint about the difference between "notph" >>> and "nostmode". Maybe that goes in the "nostmode" patch? I'm not >>> super clear on all the differences here. >> >> I can combine them. Here is the combination and it meaning based on TPH >> Control Register values: >> >> Requestor Enable | ST Mode | Meaning >> --------------------------------------------------------------- >> 00 | xx | TPH disabled (i.e. notph) >> 01 | 00 | TPH enabled, NO ST Mode (i.e. nostmode) >> 01 or 11 | 01 | Interrupt Vector mode >> 01 or 11 | 10 | Device specific mode >> >> If you have any other thoughts on how to approach these modes, please >> let me know. > > IIRC, there's no interface in this series that reall does anything > with TPH per se; drivers would only use the ST-related things. > > If that's the case, maybe "pci=notph" isn't needed yet. I can go with it. There will be a BIOS option to turn it off on AMD platform. I would expect similar options on other vendors' platforms. So I am not overly concerned about dropping pci=notph. > > Bjorn
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt index b2057241ea6c..65581ebd9b50 100644 --- a/Documentation/admin-guide/kernel-parameters.txt +++ b/Documentation/admin-guide/kernel-parameters.txt @@ -4655,6 +4655,7 @@ nomio [S390] Do not use MIO instructions. norid [S390] ignore the RID field and force use of one PCI domain per PCI function + notph [PCIE] Do not use PCIe TPH pcie_aspm= [PCIE] Forcibly enable or ignore PCIe Active State Power Management. diff --git a/drivers/pci/pci-driver.c b/drivers/pci/pci-driver.c index af2996d0d17f..9722d070c0ca 100644 --- a/drivers/pci/pci-driver.c +++ b/drivers/pci/pci-driver.c @@ -21,6 +21,7 @@ #include <linux/acpi.h> #include <linux/dma-map-ops.h> #include <linux/iommu.h> +#include <linux/pci-tph.h> #include "pci.h" #include "pcie/portdrv.h" @@ -322,8 +323,12 @@ static long local_pci_probe(void *_ddi) pm_runtime_get_sync(dev); pci_dev->driver = pci_drv; rc = pci_drv->probe(pci_dev, ddi->id); - if (!rc) + if (!rc) { + if (pci_tph_disabled()) + pcie_tph_disable(pci_dev); + return rc; + } if (rc < 0) { pci_dev->driver = NULL; pm_runtime_put_sync(dev); diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c index 02b1d81b1419..4cbfd5b53be8 100644 --- a/drivers/pci/pci.c +++ b/drivers/pci/pci.c @@ -157,6 +157,9 @@ static bool pcie_ari_disabled; /* If set, the PCIe ATS capability will not be used. */ static bool pcie_ats_disabled; +/* If set, the PCIe TPH capability will not be used. */ +static bool pcie_tph_disabled; + /* If set, the PCI config space of each device is printed during boot. */ bool pci_early_dump; @@ -166,6 +169,12 @@ bool pci_ats_disabled(void) } EXPORT_SYMBOL_GPL(pci_ats_disabled); +bool pci_tph_disabled(void) +{ + return pcie_tph_disabled; +} +EXPORT_SYMBOL_GPL(pci_tph_disabled); + /* Disable bridge_d3 for all PCIe ports */ static bool pci_bridge_d3_disable; /* Force bridge_d3 for all PCIe ports */ @@ -6869,6 +6878,9 @@ static int __init pci_setup(char *str) pci_no_domains(); } else if (!strncmp(str, "noari", 5)) { pcie_ari_disabled = true; + } else if (!strcmp(str, "notph")) { + pr_info("PCIe: TPH is disabled\n"); + pcie_tph_disabled = true; } else if (!strncmp(str, "cbiosize=", 9)) { pci_cardbus_io_size = memparse(str + 9, &str); } else if (!strncmp(str, "cbmemsize=", 10)) { diff --git a/drivers/pci/pcie/tph.c b/drivers/pci/pcie/tph.c index e385b871333e..ad58a892792c 100644 --- a/drivers/pci/pcie/tph.c +++ b/drivers/pci/pcie/tph.c @@ -7,8 +7,33 @@ * Wei Huang <wei.huang2@amd.com> */ +#include <linux/pci.h> +#include <linux/bitfield.h> +#include <linux/pci-tph.h> + #include "../pci.h" +/* Update the TPH Requester Enable field of TPH Control Register */ +static void set_ctrl_reg_req_en(struct pci_dev *pdev, u8 req_type) +{ + u32 reg_val; + + pci_read_config_dword(pdev, pdev->tph_cap + PCI_TPH_CTRL, ®_val); + + reg_val &= ~PCI_TPH_CTRL_REQ_EN_MASK; + reg_val |= FIELD_PREP(PCI_TPH_CTRL_REQ_EN_MASK, req_type); + + pci_write_config_dword(pdev, pdev->tph_cap + PCI_TPH_CTRL, reg_val); +} + +void pcie_tph_disable(struct pci_dev *pdev) +{ + if (!pdev->tph_cap) + return; + + set_ctrl_reg_req_en(pdev, PCI_TPH_REQ_DISABLE); +} + void pcie_tph_init(struct pci_dev *pdev) { pdev->tph_cap = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_TPH); diff --git a/include/linux/pci-tph.h b/include/linux/pci-tph.h new file mode 100644 index 000000000000..e0b782bda929 --- /dev/null +++ b/include/linux/pci-tph.h @@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +/* + * TPH (TLP Processing Hints) + * + * Copyright (C) 2024 Advanced Micro Devices, Inc. + * Eric Van Tassell <Eric.VanTassell@amd.com> + * Wei Huang <wei.huang2@amd.com> + */ +#ifndef LINUX_PCI_TPH_H +#define LINUX_PCI_TPH_H + +#ifdef CONFIG_PCIE_TPH +void pcie_tph_disable(struct pci_dev *dev); +#else +static inline void pcie_tph_disable(struct pci_dev *dev) {} +#endif + +#endif /* LINUX_PCI_TPH_H */ diff --git a/include/linux/pci.h b/include/linux/pci.h index 6631ebe80ca9..05fbbd9ad6b4 100644 --- a/include/linux/pci.h +++ b/include/linux/pci.h @@ -1840,6 +1840,7 @@ static inline bool pci_aer_available(void) { return false; } #endif bool pci_ats_disabled(void); +bool pci_tph_disabled(void); #ifdef CONFIG_PCIE_PTM int pci_enable_ptm(struct pci_dev *dev, u8 *granularity);