| Message ID | 20211216133456.21002-1-xueshuai@linux.alibaba.com (mailing list archive) |
|---|---|
| State | Changes Requested, archived |
| Headers | show |
| Series | [RESEND,v4] ACPI: Move sdei_init and ghes_init ahead to handle platform errors earlier | expand |
On Thu, Dec 16, 2021 at 2:35 PM Shuai Xue <xueshuai@linux.alibaba.com> wrote: > > On an ACPI system, ACPI is initialised very early from a subsys_initcall(), > while SDEI is not ready until a subsys_initcall_sync(). > > The SDEI driver provides functions (e.g. apei_sdei_register_ghes, > apei_sdei_unregister_ghes) to register or unregister event callback for > dispatcher in firmware. When the GHES driver probing, it registers the > corresponding callback according to the notification type specified by > GHES. If the GHES notification type is SDEI, the GHES driver will call > apei_sdei_register_ghes to register event call. > > When the firmware emits an event, it migrates the handling of the event > into the kernel at the registered entry-point __sdei_asm_handler. And > finally, the kernel will call the registered event callback and return > status_code to indicate the status of event handling. SDEI_EV_FAILED > indicates that the kernel failed to handle the event. > > Consequently, when an error occurs during kernel booting, the kernel is > unable to handle and report errors until the GHES driver is initialized by > device_initcall(), in which the event callback is registered. All errors > that occurred before GHES initialization are missed and there is no chance > to report and find them again. > > From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus > memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage > the estatus memory pool. On the other hand, ghes_init() relies on > sdei_init() to detect the SDEI version and the framework for registering > and unregistering events. By the way, I don't figure out why acpi_hest_init > is called in acpi_pci_root_init, it don't rely on any other thing. May it > could be moved further, following acpi_iort_init in acpi_init. > > sdei_init() relies on ACPI table which is initialized subsys_initcall(): > acpi_init(), acpi_bus_init(), acpi_load_tables(), acpi_tb_laod_namespace(). > May it should be also moved further, after acpi_load_tables. > > In this patch, move sdei_init and ghes_init as far ahead as possible, right > after acpi_hest_init(). > > Signed-off-by: Shuai Xue <xueshuai@linux.alibaba.com> This needs ACKs from Bjorn and Boris/Tony/James/. > --- > drivers/acpi/apei/ghes.c | 18 ++++++++---------- > drivers/acpi/pci_root.c | 5 ++++- > drivers/firmware/arm_sdei.c | 13 ++----------- > include/acpi/apei.h | 2 ++ > include/linux/arm_sdei.h | 2 ++ > 5 files changed, 18 insertions(+), 22 deletions(-) > > diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c > index 0c8330ed1ffd..b11e46fb4b3d 100644 > --- a/drivers/acpi/apei/ghes.c > +++ b/drivers/acpi/apei/ghes.c > @@ -1457,27 +1457,26 @@ static struct platform_driver ghes_platform_driver = { > .remove = ghes_remove, > }; > > -static int __init ghes_init(void) > +void __init ghes_init(void) > { > int rc; > > if (acpi_disabled) > - return -ENODEV; > + return; > > switch (hest_disable) { > case HEST_NOT_FOUND: > - return -ENODEV; > + pr_info(GHES_PFX "HEST is not found!\n"); > + return; > case HEST_DISABLED: > pr_info(GHES_PFX "HEST is not enabled!\n"); > - return -EINVAL; > + return; > default: > break; > } > > - if (ghes_disable) { > + if (ghes_disable) > pr_info(GHES_PFX "GHES is not enabled!\n"); > - return -EINVAL; > - } > > ghes_nmi_init_cxt(); > > @@ -1495,8 +1494,7 @@ static int __init ghes_init(void) > else > pr_info(GHES_PFX "Failed to enable APEI firmware first mode.\n"); > > - return 0; > + return; > err: > - return rc; > + ghes_disable = 1; > } > -device_initcall(ghes_init); > diff --git a/drivers/acpi/pci_root.c b/drivers/acpi/pci_root.c > index ab2f7dfb0c44..1260bb556184 100644 > --- a/drivers/acpi/pci_root.c > +++ b/drivers/acpi/pci_root.c > @@ -23,7 +23,7 @@ > #include <linux/dmi.h> > #include <linux/platform_data/x86/apple.h> > #include <acpi/apei.h> /* for acpi_hest_init() */ > - > +#include <linux/arm_sdei.h> /* for sdei_init() */ > #include "internal.h" > > #define ACPI_PCI_ROOT_CLASS "pci_bridge" > @@ -946,6 +946,9 @@ struct pci_bus *acpi_pci_root_create(struct acpi_pci_root *root, > void __init acpi_pci_root_init(void) > { > acpi_hest_init(); > + sdei_init(); > + ghes_init(); > + > if (acpi_pci_disabled) > return; > > diff --git a/drivers/firmware/arm_sdei.c b/drivers/firmware/arm_sdei.c > index a7e762c352f9..1e1a51510e83 100644 > --- a/drivers/firmware/arm_sdei.c > +++ b/drivers/firmware/arm_sdei.c > @@ -1059,14 +1059,14 @@ static bool __init sdei_present_acpi(void) > return true; > } > > -static int __init sdei_init(void) > +void __init sdei_init(void) > { > struct platform_device *pdev; > int ret; > > ret = platform_driver_register(&sdei_driver); > if (ret || !sdei_present_acpi()) > - return ret; > + return; > > pdev = platform_device_register_simple(sdei_driver.driver.name, > 0, NULL, 0); > @@ -1076,17 +1076,8 @@ static int __init sdei_init(void) > pr_info("Failed to register ACPI:SDEI platform device %d\n", > ret); > } > - > - return ret; > } > > -/* > - * On an ACPI system SDEI needs to be ready before HEST:GHES tries to register > - * its events. ACPI is initialised from a subsys_initcall(), GHES is initialised > - * by device_initcall(). We want to be called in the middle. > - */ > -subsys_initcall_sync(sdei_init); > - > int sdei_event_handler(struct pt_regs *regs, > struct sdei_registered_event *arg) > { > diff --git a/include/acpi/apei.h b/include/acpi/apei.h > index ece0a8af2bae..7dbd6363fda7 100644 > --- a/include/acpi/apei.h > +++ b/include/acpi/apei.h > @@ -27,8 +27,10 @@ extern int hest_disable; > extern int erst_disable; > #ifdef CONFIG_ACPI_APEI_GHES > extern bool ghes_disable; > +void __init ghes_init(void); > #else > #define ghes_disable 1 > +static inline void ghes_init(void) { return; } > #endif > > #ifdef CONFIG_ACPI_APEI > diff --git a/include/linux/arm_sdei.h b/include/linux/arm_sdei.h > index 0a241c5c911d..9c987188b692 100644 > --- a/include/linux/arm_sdei.h > +++ b/include/linux/arm_sdei.h > @@ -46,9 +46,11 @@ int sdei_unregister_ghes(struct ghes *ghes); > /* For use by arch code when CPU hotplug notifiers are not appropriate. */ > int sdei_mask_local_cpu(void); > int sdei_unmask_local_cpu(void); > +void __init sdei_init(void); > #else > static inline int sdei_mask_local_cpu(void) { return 0; } > static inline int sdei_unmask_local_cpu(void) { return 0; } > +static inline void sdei_init(void) { return ; } > #endif /* CONFIG_ARM_SDE_INTERFACE */ > > > -- > 2.20.1.12.g72788fdb >
Hi Rafeal, Thank you for your reply. On 2021/12/18 AM2:17, Rafael J. Wysocki wrote: > On Thu, Dec 16, 2021 at 2:35 PM Shuai Xue <xueshuai@linux.alibaba.com> wrote: >> >> On an ACPI system, ACPI is initialised very early from a subsys_initcall(), >> while SDEI is not ready until a subsys_initcall_sync(). >> >> The SDEI driver provides functions (e.g. apei_sdei_register_ghes, >> apei_sdei_unregister_ghes) to register or unregister event callback for >> dispatcher in firmware. When the GHES driver probing, it registers the >> corresponding callback according to the notification type specified by >> GHES. If the GHES notification type is SDEI, the GHES driver will call >> apei_sdei_register_ghes to register event call. >> >> When the firmware emits an event, it migrates the handling of the event >> into the kernel at the registered entry-point __sdei_asm_handler. And >> finally, the kernel will call the registered event callback and return >> status_code to indicate the status of event handling. SDEI_EV_FAILED >> indicates that the kernel failed to handle the event. >> >> Consequently, when an error occurs during kernel booting, the kernel is >> unable to handle and report errors until the GHES driver is initialized by >> device_initcall(), in which the event callback is registered. All errors >> that occurred before GHES initialization are missed and there is no chance >> to report and find them again. >> >> From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus >> memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage >> the estatus memory pool. On the other hand, ghes_init() relies on >> sdei_init() to detect the SDEI version and the framework for registering >> and unregistering events. By the way, I don't figure out why acpi_hest_init >> is called in acpi_pci_root_init, it don't rely on any other thing. May it >> could be moved further, following acpi_iort_init in acpi_init. >> >> sdei_init() relies on ACPI table which is initialized subsys_initcall(): >> acpi_init(), acpi_bus_init(), acpi_load_tables(), acpi_tb_laod_namespace(). >> May it should be also moved further, after acpi_load_tables. >> >> In this patch, move sdei_init and ghes_init as far ahead as possible, right >> after acpi_hest_init(). >> >> Signed-off-by: Shuai Xue <xueshuai@linux.alibaba.com> > > This needs ACKs from Bjorn and Boris/Tony/James/. OK, I will keep an eye on their response. I am wondering that do you have any comments for revision? Thank you. Best Regrads, Shuai >> --- >> drivers/acpi/apei/ghes.c | 18 ++++++++---------- >> drivers/acpi/pci_root.c | 5 ++++- >> drivers/firmware/arm_sdei.c | 13 ++----------- >> include/acpi/apei.h | 2 ++ >> include/linux/arm_sdei.h | 2 ++ >> 5 files changed, 18 insertions(+), 22 deletions(-) >> >> diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c >> index 0c8330ed1ffd..b11e46fb4b3d 100644 >> --- a/drivers/acpi/apei/ghes.c >> +++ b/drivers/acpi/apei/ghes.c >> @@ -1457,27 +1457,26 @@ static struct platform_driver ghes_platform_driver = { >> .remove = ghes_remove, >> }; >> >> -static int __init ghes_init(void) >> +void __init ghes_init(void) >> { >> int rc; >> >> if (acpi_disabled) >> - return -ENODEV; >> + return; >> >> switch (hest_disable) { >> case HEST_NOT_FOUND: >> - return -ENODEV; >> + pr_info(GHES_PFX "HEST is not found!\n"); >> + return; >> case HEST_DISABLED: >> pr_info(GHES_PFX "HEST is not enabled!\n"); >> - return -EINVAL; >> + return; >> default: >> break; >> } >> >> - if (ghes_disable) { >> + if (ghes_disable) >> pr_info(GHES_PFX "GHES is not enabled!\n"); >> - return -EINVAL; >> - } >> >> ghes_nmi_init_cxt(); >> >> @@ -1495,8 +1494,7 @@ static int __init ghes_init(void) >> else >> pr_info(GHES_PFX "Failed to enable APEI firmware first mode.\n"); >> >> - return 0; >> + return; >> err: >> - return rc; >> + ghes_disable = 1; >> } >> -device_initcall(ghes_init); >> diff --git a/drivers/acpi/pci_root.c b/drivers/acpi/pci_root.c >> index ab2f7dfb0c44..1260bb556184 100644 >> --- a/drivers/acpi/pci_root.c >> +++ b/drivers/acpi/pci_root.c >> @@ -23,7 +23,7 @@ >> #include <linux/dmi.h> >> #include <linux/platform_data/x86/apple.h> >> #include <acpi/apei.h> /* for acpi_hest_init() */ >> - >> +#include <linux/arm_sdei.h> /* for sdei_init() */ >> #include "internal.h" >> >> #define ACPI_PCI_ROOT_CLASS "pci_bridge" >> @@ -946,6 +946,9 @@ struct pci_bus *acpi_pci_root_create(struct acpi_pci_root *root, >> void __init acpi_pci_root_init(void) >> { >> acpi_hest_init(); >> + sdei_init(); >> + ghes_init(); >> + >> if (acpi_pci_disabled) >> return; >> >> diff --git a/drivers/firmware/arm_sdei.c b/drivers/firmware/arm_sdei.c >> index a7e762c352f9..1e1a51510e83 100644 >> --- a/drivers/firmware/arm_sdei.c >> +++ b/drivers/firmware/arm_sdei.c >> @@ -1059,14 +1059,14 @@ static bool __init sdei_present_acpi(void) >> return true; >> } >> >> -static int __init sdei_init(void) >> +void __init sdei_init(void) >> { >> struct platform_device *pdev; >> int ret; >> >> ret = platform_driver_register(&sdei_driver); >> if (ret || !sdei_present_acpi()) >> - return ret; >> + return; >> >> pdev = platform_device_register_simple(sdei_driver.driver.name, >> 0, NULL, 0); >> @@ -1076,17 +1076,8 @@ static int __init sdei_init(void) >> pr_info("Failed to register ACPI:SDEI platform device %d\n", >> ret); >> } >> - >> - return ret; >> } >> >> -/* >> - * On an ACPI system SDEI needs to be ready before HEST:GHES tries to register >> - * its events. ACPI is initialised from a subsys_initcall(), GHES is initialised >> - * by device_initcall(). We want to be called in the middle. >> - */ >> -subsys_initcall_sync(sdei_init); >> - >> int sdei_event_handler(struct pt_regs *regs, >> struct sdei_registered_event *arg) >> { >> diff --git a/include/acpi/apei.h b/include/acpi/apei.h >> index ece0a8af2bae..7dbd6363fda7 100644 >> --- a/include/acpi/apei.h >> +++ b/include/acpi/apei.h >> @@ -27,8 +27,10 @@ extern int hest_disable; >> extern int erst_disable; >> #ifdef CONFIG_ACPI_APEI_GHES >> extern bool ghes_disable; >> +void __init ghes_init(void); >> #else >> #define ghes_disable 1 >> +static inline void ghes_init(void) { return; } >> #endif >> >> #ifdef CONFIG_ACPI_APEI >> diff --git a/include/linux/arm_sdei.h b/include/linux/arm_sdei.h >> index 0a241c5c911d..9c987188b692 100644 >> --- a/include/linux/arm_sdei.h >> +++ b/include/linux/arm_sdei.h >> @@ -46,9 +46,11 @@ int sdei_unregister_ghes(struct ghes *ghes); >> /* For use by arch code when CPU hotplug notifiers are not appropriate. */ >> int sdei_mask_local_cpu(void); >> int sdei_unmask_local_cpu(void); >> +void __init sdei_init(void); >> #else >> static inline int sdei_mask_local_cpu(void) { return 0; } >> static inline int sdei_unmask_local_cpu(void) { return 0; } >> +static inline void sdei_init(void) { return ; } >> #endif /* CONFIG_ARM_SDE_INTERFACE */ >> >> >> -- >> 2.20.1.12.g72788fdb >>
On Thu, Dec 16, 2021 at 09:34:56PM +0800, Shuai Xue wrote: > On an ACPI system, ACPI is initialised very early from a subsys_initcall(), > while SDEI is not ready until a subsys_initcall_sync(). > > The SDEI driver provides functions (e.g. apei_sdei_register_ghes, > apei_sdei_unregister_ghes) to register or unregister event callback for > dispatcher in firmware. When the GHES driver probing, it registers the > corresponding callback according to the notification type specified by > GHES. If the GHES notification type is SDEI, the GHES driver will call > apei_sdei_register_ghes to register event call. > > When the firmware emits an event, it migrates the handling of the event > into the kernel at the registered entry-point __sdei_asm_handler. And > finally, the kernel will call the registered event callback and return > status_code to indicate the status of event handling. SDEI_EV_FAILED > indicates that the kernel failed to handle the event. > > Consequently, when an error occurs during kernel booting, the kernel is > unable to handle and report errors until the GHES driver is initialized by > device_initcall(), in which the event callback is registered. All errors > that occurred before GHES initialization are missed and there is no chance > to report and find them again. > > From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus > memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage > the estatus memory pool. On the other hand, ghes_init() relies on > sdei_init() to detect the SDEI version and the framework for registering > and unregistering events. > By the way, I don't figure out why acpi_hest_init is called in > acpi_pci_root_init, it don't rely on any other thing. May it could > be moved further, following acpi_iort_init in acpi_init. I think you should drop the "By the way ..." text or move it after the "---" at the bottom of your commit log. It doesn't help understand this patch. > sdei_init() relies on ACPI table which is initialized > subsys_initcall(): acpi_init(), acpi_bus_init(), acpi_load_tables(), > acpi_tb_laod_namespace(). May it should be also moved further, > after acpi_load_tables. This text also doesn't seem relevant to this patch. > In this patch, move sdei_init and ghes_init as far ahead as > possible, right after acpi_hest_init(). I'm having a hard time figuring out the reason for this patch. Apparently the relevant parts are sdei_init() and ghes_init(). Today they are executed in that order: subsys_initcall_sync(sdei_init); device_initcall(ghes_init); After this patch, they would be executed in the same order, but called explicitly instead of as initcalls: acpi_pci_root_init() { acpi_hest_init(); sdei_init(); ghes_init(); ... Explicit calls are certainly better than initcalls, but that doesn't seem to be the reason for this patch. Does this patch fix a bug? If so, what is the bug? You say that currently "errors that occur before GHES initialization are missed". Isn't that still true after this patch? Does this patch merely reduce the time before GHES initialization? If so, I'm dubious, because we have to tolerate an arbitrary amount of time there. s/acpi_tb_laod_namespace/acpi_tb_load_namespace/ You use "()" after function names sometimes, but not always. Please do it consistently. > -device_initcall(ghes_init); > void __init acpi_pci_root_init(void) > { > acpi_hest_init(); > + sdei_init(); > + ghes_init(); What's the connection between PCI, SDEI, and GHES? As far as I can tell, SDEI and GHES are not PCI-specific, so it doesn't seem like they should be initialized here in acpi_pci_root_init(). > -subsys_initcall_sync(sdei_init);
Hi, Bjorn, Thank you for your comments. 在 2021/12/22 AM7:17, Bjorn Helgaas 写道: > On Thu, Dec 16, 2021 at 09:34:56PM +0800, Shuai Xue wrote: >> On an ACPI system, ACPI is initialised very early from a subsys_initcall(), >> while SDEI is not ready until a subsys_initcall_sync(). >> >> The SDEI driver provides functions (e.g. apei_sdei_register_ghes, >> apei_sdei_unregister_ghes) to register or unregister event callback for >> dispatcher in firmware. When the GHES driver probing, it registers the >> corresponding callback according to the notification type specified by >> GHES. If the GHES notification type is SDEI, the GHES driver will call >> apei_sdei_register_ghes to register event call. >> >> When the firmware emits an event, it migrates the handling of the event >> into the kernel at the registered entry-point __sdei_asm_handler. And >> finally, the kernel will call the registered event callback and return >> status_code to indicate the status of event handling. SDEI_EV_FAILED >> indicates that the kernel failed to handle the event. >> >> Consequently, when an error occurs during kernel booting, the kernel is >> unable to handle and report errors until the GHES driver is initialized by >> device_initcall(), in which the event callback is registered. All errors >> that occurred before GHES initialization are missed and there is no chance >> to report and find them again. >> >> From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus >> memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage >> the estatus memory pool. On the other hand, ghes_init() relies on >> sdei_init() to detect the SDEI version and the framework for registering >> and unregistering events. > >> By the way, I don't figure out why acpi_hest_init is called in >> acpi_pci_root_init, it don't rely on any other thing. May it could >> be moved further, following acpi_iort_init in acpi_init. > > I think you should drop the "By the way ..." text or move it after the > "---" at the bottom of your commit log. It doesn't help understand > this patch. I will fix it in next version. >> sdei_init() relies on ACPI table which is initialized >> subsys_initcall(): acpi_init(), acpi_bus_init(), acpi_load_tables(), >> acpi_tb_laod_namespace(). May it should be also moved further, >> after acpi_load_tables. > > This text also doesn't seem relevant to this patch. I will delete it in next version. >> In this patch, move sdei_init and ghes_init as far ahead as >> possible, right after acpi_hest_init(). > > I'm having a hard time figuring out the reason for this patch. > > Apparently the relevant parts are sdei_init() and ghes_init(). > Today they are executed in that order: > > subsys_initcall_sync(sdei_init); > device_initcall(ghes_init); > > After this patch, they would be executed in the same order, but called > explicitly instead of as initcalls: > > acpi_pci_root_init() > { > acpi_hest_init(); > sdei_init(); > ghes_init(); > ... > > Explicit calls are certainly better than initcalls, but that doesn't > seem to be the reason for this patch. > > Does this patch fix a bug? If so, what is the bug? Yes. When the kernel booting, the console logs many times from firmware before GHES drivers init: Trip in MM PCIe RAS handle(Intr:910) Clean PE[1.1.1] ERR_STS:0x4000100 -> 0 INT_STS:F0000000 Find RP(98:1.0) --Walk dev(98:1.0) CE:0 UCE:4000 ... ERROR: sdei_dispatch_event(32a) ret:-1 --handler(910) end This is because the callback function has not been registered yet. Previously reported errors will be overwritten by new ones. Therefore, all errors that occurred before GHES initialization are missed and there is no chance to report and find them again. > You say that currently "errors that occur before GHES initialization > are missed". Isn't that still true after this patch? Does this patch > merely reduce the time before GHES initialization? If so, I'm > dubious, because we have to tolerate an arbitrary amount of time > there. After this patch, there are still errors missing. As you mentioned, we have to tolerate it until the software reporting capability is built. Yes, this patch merely reduce the time before GHES initialization. The boot dmesg before this patch: [ 3.688586] HEST: Table parsing has been initialized. ... [ 33.204340] calling sdei_init+0x0/0x120 @ 1 [ 33.208645] sdei: SDEIv1.0 (0x0) detected in firmware. ... [ 36.005390] calling ghes_init+0x0/0x11c @ 1 [ 36.190021] GHES: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. After this patch, the boot dmesg like bellow: [ 3.688664] HEST: Table parsing has been initialized. [ 3.688691] sdei: SDEIv1.0 (0x0) detected in firmware. [ 3.694557] GHES: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. As we can see, the initialization of GHES is advanced by 33 seconds. So, in my opinion, this patch is necessary, right? (It should be noted that the effect of optimization varies with the platform.) > s/acpi_tb_laod_namespace/acpi_tb_load_namespace/ > You use "()" after function names sometimes, but not always. Please > do it consistently. Thank you for pointing this out. I will fix it in next version. >> -device_initcall(ghes_init); > >> void __init acpi_pci_root_init(void) >> { >> acpi_hest_init(); >> + sdei_init(); >> + ghes_init(); > > What's the connection between PCI, SDEI, and GHES? As far as I can > tell, SDEI and GHES are not PCI-specific, so it doesn't seem like they > should be initialized here in acpi_pci_root_init(). The only reason is that acpi_hest_init() is initialized here. From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage the estatus memory pool. On the other hand, ghes_init() relies on sdei_init() to detect the SDEI version and the framework for registering and unregistering events. The dependencies are as follows ghes_init() => acpi_hest_init() ghes_init() => sdei_init() I don't figure out why acpi_hest_init() is called in acpi_pci_root_init(), it don't rely on any other thing. I am wondering that should we moved all of them further? e.g. following acpi_iort_init() in acpi_init(). Best Regards, Shuai
[+to Rafael, question about HEST/GHES/SDEI init] On Thu, Dec 23, 2021 at 04:11:11PM +0800, Shuai Xue wrote: > 在 2021/12/22 AM7:17, Bjorn Helgaas 写道: > > On Thu, Dec 16, 2021 at 09:34:56PM +0800, Shuai Xue wrote: > >> On an ACPI system, ACPI is initialised very early from a subsys_initcall(), > >> while SDEI is not ready until a subsys_initcall_sync(). > >> > >> The SDEI driver provides functions (e.g. apei_sdei_register_ghes, > >> apei_sdei_unregister_ghes) to register or unregister event callback for > >> dispatcher in firmware. When the GHES driver probing, it registers the > >> corresponding callback according to the notification type specified by > >> GHES. If the GHES notification type is SDEI, the GHES driver will call > >> apei_sdei_register_ghes to register event call. > >> > >> When the firmware emits an event, it migrates the handling of the event > >> into the kernel at the registered entry-point __sdei_asm_handler. And > >> finally, the kernel will call the registered event callback and return > >> status_code to indicate the status of event handling. SDEI_EV_FAILED > >> indicates that the kernel failed to handle the event. > >> > >> Consequently, when an error occurs during kernel booting, the kernel is > >> unable to handle and report errors until the GHES driver is initialized by > >> device_initcall(), in which the event callback is registered. All errors > >> that occurred before GHES initialization are missed and there is no chance > >> to report and find them again. > >> > >> From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus > >> memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage > >> the estatus memory pool. On the other hand, ghes_init() relies on > >> sdei_init() to detect the SDEI version and the framework for registering > >> and unregistering events. > > > >> By the way, I don't figure out why acpi_hest_init is called in > >> acpi_pci_root_init, it don't rely on any other thing. May it could > >> be moved further, following acpi_iort_init in acpi_init. > >> sdei_init() relies on ACPI table which is initialized > >> subsys_initcall(): acpi_init(), acpi_bus_init(), acpi_load_tables(), > >> acpi_tb_laod_namespace(). May it should be also moved further, > >> after acpi_load_tables. > >> In this patch, move sdei_init and ghes_init as far ahead as > >> possible, right after acpi_hest_init(). > > > > I'm having a hard time figuring out the reason for this patch. > > > > Apparently the relevant parts are sdei_init() and ghes_init(). > > Today they are executed in that order: > > > > subsys_initcall_sync(sdei_init); > > device_initcall(ghes_init); > > > > After this patch, they would be executed in the same order, but called > > explicitly instead of as initcalls: > > > > acpi_pci_root_init() > > { > > acpi_hest_init(); > > sdei_init(); > > ghes_init(); > > ... > > > > Explicit calls are certainly better than initcalls, but that doesn't > > seem to be the reason for this patch. > > > > Does this patch fix a bug? If so, what is the bug? > > Yes. When the kernel booting, the console logs many times from firmware > before GHES drivers init: > > Trip in MM PCIe RAS handle(Intr:910) > Clean PE[1.1.1] ERR_STS:0x4000100 -> 0 INT_STS:F0000000 > Find RP(98:1.0) > --Walk dev(98:1.0) CE:0 UCE:4000 > ... > ERROR: sdei_dispatch_event(32a) ret:-1 > --handler(910) end > > This is because the callback function has not been registered yet. > Previously reported errors will be overwritten by new ones. Therefore, > all errors that occurred before GHES initialization are missed > and there is no chance to report and find them again. > > > You say that currently "errors that occur before GHES initialization > > are missed". Isn't that still true after this patch? Does this patch > > merely reduce the time before GHES initialization? If so, I'm > > dubious, because we have to tolerate an arbitrary amount of time > > there. > > After this patch, there are still errors missing. As you mentioned, > we have to tolerate it until the software reporting capability is built. > > Yes, this patch merely reduce the time before GHES initialization. It's not a bug that errors that happen before the callback are lost. At least, it's not a *Linux* bug. It might be a poor design of the firmware error reporting interface. If the only point of this patch is to reduce the time before GHES initialization, the commit log should clearly say that. > The boot dmesg before this patch: > > [ 3.688586] HEST: Table parsing has been initialized. > ... > [ 33.204340] calling sdei_init+0x0/0x120 @ 1 > [ 33.208645] sdei: SDEIv1.0 (0x0) detected in firmware. > ... > [ 36.005390] calling ghes_init+0x0/0x11c @ 1 > [ 36.190021] GHES: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. > > > After this patch, the boot dmesg like bellow: > > [ 3.688664] HEST: Table parsing has been initialized. > [ 3.688691] sdei: SDEIv1.0 (0x0) detected in firmware. > [ 3.694557] GHES: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. [Tangent: I think this GHES message is confusing. What "APEI bit" does this refer to? The only bits I remember are the Flags bits in HEST Error Source Descriptor Entries, e.g., ACPI v6.3, sec 18.3.2. "WHEA _OSC" means nothing to me, and I didn't find anything useful with grep, other than that "WHEA" might be an obsolete name for what we now call "APEI". I don't think there's anything in _OSC that mentions "firmware first." I don't remember anything in the spec about a way to *enable* Firmware First Error Handling needing (I'm looking at ACPI v6.3, sec 18.4). I think the "firmware first" information is useless to the OS -- as far as I can tell, the spec says nothing about anything the OS should do based on the FIRMWARE_FIRST bits.] > As we can see, the initialization of GHES is advanced by 33 seconds. > So, in my opinion, this patch is necessary, right? > (It should be noted that the effect of optimization varies with the platform.) > >> -device_initcall(ghes_init); > > > >> void __init acpi_pci_root_init(void) > >> { > >> acpi_hest_init(); > >> + sdei_init(); > >> + ghes_init(); > > > > What's the connection between PCI, SDEI, and GHES? As far as I can > > tell, SDEI and GHES are not PCI-specific, so it doesn't seem like they > > should be initialized here in acpi_pci_root_init(). > > The only reason is that acpi_hest_init() is initialized here. > > From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus > memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage > the estatus memory pool. On the other hand, ghes_init() relies on > sdei_init() to detect the SDEI version and the framework for registering > and unregistering events. The dependencies are as follows > > ghes_init() => acpi_hest_init() > ghes_init() => sdei_init() > > I don't figure out why acpi_hest_init() is called in > acpi_pci_root_init(), it don't rely on any other thing. > I am wondering that should we moved all of them further? e.g. > following acpi_iort_init() in acpi_init(). I don't know why acpi_hest_init() is called from acpi_pci_root_init(). It looks like HEST can support error sources other than PCI (IA-32 Machine Checks, NMIs, GHES, etc.) It was added by 415e12b23792 ("PCI/ACPI: Request _OSC control once for each root bridge (v3)"); maybe Rafael remembers why. Seem like acpi_hest_init(), sdei_init(), and ghes_init() should all go somewhere else, but I don't know where. Maybe somewhere in acpi_init()? Bjorn
Hi, Bjorn, Thank you for your comments. 在 2021/12/24 AM8:17, Bjorn Helgaas 写道: > [+to Rafael, question about HEST/GHES/SDEI init] > > On Thu, Dec 23, 2021 at 04:11:11PM +0800, Shuai Xue wrote: >> 在 2021/12/22 AM7:17, Bjorn Helgaas 写道: >>> On Thu, Dec 16, 2021 at 09:34:56PM +0800, Shuai Xue wrote: >>>> On an ACPI system, ACPI is initialised very early from a subsys_initcall(), >>>> while SDEI is not ready until a subsys_initcall_sync(). >>>> >>>> The SDEI driver provides functions (e.g. apei_sdei_register_ghes, >>>> apei_sdei_unregister_ghes) to register or unregister event callback for >>>> dispatcher in firmware. When the GHES driver probing, it registers the >>>> corresponding callback according to the notification type specified by >>>> GHES. If the GHES notification type is SDEI, the GHES driver will call >>>> apei_sdei_register_ghes to register event call. >>>> >>>> When the firmware emits an event, it migrates the handling of the event >>>> into the kernel at the registered entry-point __sdei_asm_handler. And >>>> finally, the kernel will call the registered event callback and return >>>> status_code to indicate the status of event handling. SDEI_EV_FAILED >>>> indicates that the kernel failed to handle the event. >>>> >>>> Consequently, when an error occurs during kernel booting, the kernel is >>>> unable to handle and report errors until the GHES driver is initialized by >>>> device_initcall(), in which the event callback is registered. All errors >>>> that occurred before GHES initialization are missed and there is no chance >>>> to report and find them again. >>>> >>>> From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus >>>> memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage >>>> the estatus memory pool. On the other hand, ghes_init() relies on >>>> sdei_init() to detect the SDEI version and the framework for registering >>>> and unregistering events. >>> >>>> By the way, I don't figure out why acpi_hest_init is called in >>>> acpi_pci_root_init, it don't rely on any other thing. May it could >>>> be moved further, following acpi_iort_init in acpi_init. > >>>> sdei_init() relies on ACPI table which is initialized >>>> subsys_initcall(): acpi_init(), acpi_bus_init(), acpi_load_tables(), >>>> acpi_tb_laod_namespace(). May it should be also moved further, >>>> after acpi_load_tables. > >>>> In this patch, move sdei_init and ghes_init as far ahead as >>>> possible, right after acpi_hest_init(). >>> >>> I'm having a hard time figuring out the reason for this patch. >>> >>> Apparently the relevant parts are sdei_init() and ghes_init(). >>> Today they are executed in that order: >>> >>> subsys_initcall_sync(sdei_init); >>> device_initcall(ghes_init); >>> >>> After this patch, they would be executed in the same order, but called >>> explicitly instead of as initcalls: >>> >>> acpi_pci_root_init() >>> { >>> acpi_hest_init(); >>> sdei_init(); >>> ghes_init(); >>> ... >>> >>> Explicit calls are certainly better than initcalls, but that doesn't >>> seem to be the reason for this patch. >>> >>> Does this patch fix a bug? If so, what is the bug? >> >> Yes. When the kernel booting, the console logs many times from firmware >> before GHES drivers init: >> >> Trip in MM PCIe RAS handle(Intr:910) >> Clean PE[1.1.1] ERR_STS:0x4000100 -> 0 INT_STS:F0000000 >> Find RP(98:1.0) >> --Walk dev(98:1.0) CE:0 UCE:4000 >> ... >> ERROR: sdei_dispatch_event(32a) ret:-1 >> --handler(910) end >> >> This is because the callback function has not been registered yet. >> Previously reported errors will be overwritten by new ones. Therefore, >> all errors that occurred before GHES initialization are missed >> and there is no chance to report and find them again. >> >>> You say that currently "errors that occur before GHES initialization >>> are missed". Isn't that still true after this patch? Does this patch >>> merely reduce the time before GHES initialization? If so, I'm >>> dubious, because we have to tolerate an arbitrary amount of time >>> there. >> >> After this patch, there are still errors missing. As you mentioned, >> we have to tolerate it until the software reporting capability is built. >> >> Yes, this patch merely reduce the time before GHES initialization. > > It's not a bug that errors that happen before the callback are lost. > At least, it's not a *Linux* bug. It might be a poor design of the > firmware error reporting interface. > > If the only point of this patch is to reduce the time before GHES > initialization, the commit log should clearly say that. Yep, it is a design defect of firmware. I will explicitly document the purpose of this patch in next version. >> The boot dmesg before this patch: >> >> [ 3.688586] HEST: Table parsing has been initialized. >> ... >> [ 33.204340] calling sdei_init+0x0/0x120 @ 1 >> [ 33.208645] sdei: SDEIv1.0 (0x0) detected in firmware. >> ... >> [ 36.005390] calling ghes_init+0x0/0x11c @ 1 >> [ 36.190021] GHES: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. >> >> >> After this patch, the boot dmesg like bellow: >> >> [ 3.688664] HEST: Table parsing has been initialized. >> [ 3.688691] sdei: SDEIv1.0 (0x0) detected in firmware. >> [ 3.694557] GHES: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. After this patch, we use explicit call to init GHES rather than initcall. The GHES message here is just to prove that GHES initialization is advanced to 3.694557 seconds. > [Tangent: I think this GHES message is confusing. What "APEI bit" > does this refer to? The only bits I remember are the Flags bits in > HEST Error Source Descriptor Entries, e.g., ACPI v6.3, sec 18.3.2. From the log code behavior, the APEI bit refers to OSC_SB_APEI_SUPPORT. (BIT 4, ACPI v6.4, sec 6.2.11.2, table Table 6.13 Platform-Wide _OSC Capabilities DWORD 2. [1]) /* code in drivers/acpi/apei/ghes.c */ if (rc == 0 && osc_sb_apei_support_acked) pr_info(GHES_PFX "APEI firmware first mode is enabled by APEI bit and WHEA _OSC.\n"); /* code in drivers/acpi/bus.c */ osc_sb_apei_support_acked = capbuf_ret[OSC_SUPPORT_DWORD] & OSC_SB_APEI_SUPPORT; > "WHEA _OSC" means nothing to me, and I didn't find anything useful > with grep, other than that "WHEA" might be an obsolete name for what > we now call "APEI". Yep, WHEA is an obsolete name. I think apei_osc_setup() is added for compatibility in commit eccddd32ced0 ("ACPI, APEI, Add APEI bit support in generic _OSC call")' The ACPI spec Platform-Wide OSPM Capabilities defines UUID "0811B06E-4A27-44F9-8D60-3CBBC22E7B48" is used in acpi_bus_osc_support(). The Microsoft WHEA defines UUID "ed855e0c-6c90-47bf-a62a-26de0fc5ad5c" that is used in apei_osc_setup(). [2][3] From the discussion[4], the GHES message is to distinguish between the two specs. To avoid confusion, can we change the message as follow. - generic _OSC succeeded, APEI _OSC failed: APEI firmware first mode is enabled by APEI bit. - generic _OSC failed, APEI _OSC succeeded: APEI firmware first mode is enabled by WHEA _OSC. - both succeeded: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. - both failed: Failed to enable APEI firmware first mode! > I don't think there's anything in _OSC that mentions "firmware first." > > I don't remember anything in the spec about a way to *enable* Firmware > First Error Handling needing (I'm looking at ACPI v6.3, sec 18.4). > > I think the "firmware first" information is useless to the OS -- as > far as I can tell, the spec says nothing about anything the OS should > do based on the FIRMWARE_FIRST bits.] GHES works in "Firmware First" mode to report platform hardware error and it depends on APEI interface. (drivers/acpi/apei/ghes.c) * Generic Hardware Error Source provides a way to report platform * hardware errors (such as that from chipset). It works in so called * "Firmware First" mode, that is, hardware errors are reported to * firmware firstly, then reported to Linux by firmware. Based on above, I think the GHES message just means that this booting kernel supports APEI or WHEA, and GHES report Firmware first errors on this machine. Is it fine to keep the message? >> As we can see, the initialization of GHES is advanced by 33 seconds. >> So, in my opinion, this patch is necessary, right? >> (It should be noted that the effect of optimization varies with the platform.) > >>>> -device_initcall(ghes_init); >>> >>>> void __init acpi_pci_root_init(void) >>>> { >>>> acpi_hest_init(); >>>> + sdei_init(); >>>> + ghes_init(); >>> >>> What's the connection between PCI, SDEI, and GHES? As far as I can >>> tell, SDEI and GHES are not PCI-specific, so it doesn't seem like they >>> should be initialized here in acpi_pci_root_init(). >> >> The only reason is that acpi_hest_init() is initialized here. >> >> From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus >> memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage >> the estatus memory pool. On the other hand, ghes_init() relies on >> sdei_init() to detect the SDEI version and the framework for registering >> and unregistering events. The dependencies are as follows >> >> ghes_init() => acpi_hest_init() >> ghes_init() => sdei_init() >> >> I don't figure out why acpi_hest_init() is called in >> acpi_pci_root_init(), it don't rely on any other thing. >> I am wondering that should we moved all of them further? e.g. >> following acpi_iort_init() in acpi_init(). > > I don't know why acpi_hest_init() is called from acpi_pci_root_init(). > It looks like HEST can support error sources other than PCI (IA-32 > Machine Checks, NMIs, GHES, etc.) It was added by 415e12b23792 > ("PCI/ACPI: Request _OSC control once for each root bridge (v3)"); > maybe Rafael remembers why. > > Seem like acpi_hest_init(), sdei_init(), and ghes_init() should all go > somewhere else, but I don't know where. Maybe somewhere in > acpi_init()? I tried to move them into acpi_init(), and it works well. @@ -1251,6 +1251,9 @@ static int __init acpi_init(void) pci_mmcfg_late_init(); acpi_iort_init(); + acpi_hest_init(); + sdei_init(); + ghes_init(); acpi_scan_init(); acpi_ec_init(); What's your opinion? Best Regards, Shuai [1] https://uefi.org/specs/ACPI/6.4/06_Device_Configuration/Device_Configuration.html#platform-wide-osc-capabilities-dword-2 [2] https://www.mail-archive.com/ubuntu-bugs@lists.ubuntu.com/msg4718503.html [3] http://www.guyreams.com/PDF/whea.pdf [4] https://patchwork.kernel.org/project/linux-acpi/patch/1308640587-24502-5-git-send-email-ying.huang@intel.com/#1978922
Hi, Bjorn, Rafael, Happy New Year :) I am wondering that do you have any comment to this implementation？ Any feedback are welcomed. Thank you. Best regrads, Shuai 在 2021/12/24 PM4:55, Shuai Xue 写道: > Hi, Bjorn, > > Thank you for your comments. > > 在 2021/12/24 AM8:17, Bjorn Helgaas 写道: >> [+to Rafael, question about HEST/GHES/SDEI init] >> >> On Thu, Dec 23, 2021 at 04:11:11PM +0800, Shuai Xue wrote: >>> 在 2021/12/22 AM7:17, Bjorn Helgaas 写道: >>>> On Thu, Dec 16, 2021 at 09:34:56PM +0800, Shuai Xue wrote: >>>>> On an ACPI system, ACPI is initialised very early from a subsys_initcall(), >>>>> while SDEI is not ready until a subsys_initcall_sync(). >>>>> >>>>> The SDEI driver provides functions (e.g. apei_sdei_register_ghes, >>>>> apei_sdei_unregister_ghes) to register or unregister event callback for >>>>> dispatcher in firmware. When the GHES driver probing, it registers the >>>>> corresponding callback according to the notification type specified by >>>>> GHES. If the GHES notification type is SDEI, the GHES driver will call >>>>> apei_sdei_register_ghes to register event call. >>>>> >>>>> When the firmware emits an event, it migrates the handling of the event >>>>> into the kernel at the registered entry-point __sdei_asm_handler. And >>>>> finally, the kernel will call the registered event callback and return >>>>> status_code to indicate the status of event handling. SDEI_EV_FAILED >>>>> indicates that the kernel failed to handle the event. >>>>> >>>>> Consequently, when an error occurs during kernel booting, the kernel is >>>>> unable to handle and report errors until the GHES driver is initialized by >>>>> device_initcall(), in which the event callback is registered. All errors >>>>> that occurred before GHES initialization are missed and there is no chance >>>>> to report and find them again. >>>>> >>>>> From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus >>>>> memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage >>>>> the estatus memory pool. On the other hand, ghes_init() relies on >>>>> sdei_init() to detect the SDEI version and the framework for registering >>>>> and unregistering events. >>>> >>>>> By the way, I don't figure out why acpi_hest_init is called in >>>>> acpi_pci_root_init, it don't rely on any other thing. May it could >>>>> be moved further, following acpi_iort_init in acpi_init. >> >>>>> sdei_init() relies on ACPI table which is initialized >>>>> subsys_initcall(): acpi_init(), acpi_bus_init(), acpi_load_tables(), >>>>> acpi_tb_laod_namespace(). May it should be also moved further, >>>>> after acpi_load_tables. >> >>>>> In this patch, move sdei_init and ghes_init as far ahead as >>>>> possible, right after acpi_hest_init(). >>>> >>>> I'm having a hard time figuring out the reason for this patch. >>>> >>>> Apparently the relevant parts are sdei_init() and ghes_init(). >>>> Today they are executed in that order: >>>> >>>> subsys_initcall_sync(sdei_init); >>>> device_initcall(ghes_init); >>>> >>>> After this patch, they would be executed in the same order, but called >>>> explicitly instead of as initcalls: >>>> >>>> acpi_pci_root_init() >>>> { >>>> acpi_hest_init(); >>>> sdei_init(); >>>> ghes_init(); >>>> ... >>>> >>>> Explicit calls are certainly better than initcalls, but that doesn't >>>> seem to be the reason for this patch. >>>> >>>> Does this patch fix a bug? If so, what is the bug? >>> >>> Yes. When the kernel booting, the console logs many times from firmware >>> before GHES drivers init: >>> >>> Trip in MM PCIe RAS handle(Intr:910) >>> Clean PE[1.1.1] ERR_STS:0x4000100 -> 0 INT_STS:F0000000 >>> Find RP(98:1.0) >>> --Walk dev(98:1.0) CE:0 UCE:4000 >>> ... >>> ERROR: sdei_dispatch_event(32a) ret:-1 >>> --handler(910) end >>> >>> This is because the callback function has not been registered yet. >>> Previously reported errors will be overwritten by new ones. Therefore, >>> all errors that occurred before GHES initialization are missed >>> and there is no chance to report and find them again. >>> >>>> You say that currently "errors that occur before GHES initialization >>>> are missed". Isn't that still true after this patch? Does this patch >>>> merely reduce the time before GHES initialization? If so, I'm >>>> dubious, because we have to tolerate an arbitrary amount of time >>>> there. >>> >>> After this patch, there are still errors missing. As you mentioned, >>> we have to tolerate it until the software reporting capability is built. >>> >>> Yes, this patch merely reduce the time before GHES initialization. >> >> It's not a bug that errors that happen before the callback are lost. >> At least, it's not a *Linux* bug. It might be a poor design of the >> firmware error reporting interface. >> >> If the only point of this patch is to reduce the time before GHES >> initialization, the commit log should clearly say that. > > Yep, it is a design defect of firmware. I will explicitly document > the purpose of this patch in next version. > >>> The boot dmesg before this patch: >>> >>> [ 3.688586] HEST: Table parsing has been initialized. >>> ... >>> [ 33.204340] calling sdei_init+0x0/0x120 @ 1 >>> [ 33.208645] sdei: SDEIv1.0 (0x0) detected in firmware. >>> ... >>> [ 36.005390] calling ghes_init+0x0/0x11c @ 1 >>> [ 36.190021] GHES: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. >>> >>> >>> After this patch, the boot dmesg like bellow: >>> >>> [ 3.688664] HEST: Table parsing has been initialized. >>> [ 3.688691] sdei: SDEIv1.0 (0x0) detected in firmware. >>> [ 3.694557] GHES: APEI firmware first mode is enabled by APEI bit and WHEA _OSC. > > After this patch, we use explicit call to init GHES rather than initcall. > The GHES message here is just to prove that GHES initialization is > advanced to 3.694557 seconds. > > >> [Tangent: I think this GHES message is confusing. What "APEI bit" >> does this refer to? The only bits I remember are the Flags bits in >> HEST Error Source Descriptor Entries, e.g., ACPI v6.3, sec 18.3.2. > > From the log code behavior, the APEI bit refers to OSC_SB_APEI_SUPPORT. > (BIT 4, ACPI v6.4, sec 6.2.11.2, table Table 6.13 Platform-Wide _OSC > Capabilities DWORD 2. [1]) > > /* code in drivers/acpi/apei/ghes.c */ > if (rc == 0 && osc_sb_apei_support_acked) > pr_info(GHES_PFX "APEI firmware first mode is enabled by APEI bit and WHEA _OSC.\n"); > > /* code in drivers/acpi/bus.c */ > osc_sb_apei_support_acked = > capbuf_ret[OSC_SUPPORT_DWORD] & OSC_SB_APEI_SUPPORT; > > >> "WHEA _OSC" means nothing to me, and I didn't find anything useful >> with grep, other than that "WHEA" might be an obsolete name for what >> we now call "APEI". > > Yep, WHEA is an obsolete name. I think apei_osc_setup() is added for compatibility > in commit eccddd32ced0 ("ACPI, APEI, Add APEI bit support in generic _OSC call")' > > The ACPI spec Platform-Wide OSPM Capabilities defines > UUID "0811B06E-4A27-44F9-8D60-3CBBC22E7B48" is used in > acpi_bus_osc_support(). The Microsoft WHEA defines > UUID "ed855e0c-6c90-47bf-a62a-26de0fc5ad5c" that is used > in apei_osc_setup(). [2][3] > > From the discussion[4], the GHES message is to distinguish between > the two specs. > > To avoid confusion, can we change the message as follow. > > - generic _OSC succeeded, APEI _OSC failed: APEI firmware first mode is > enabled by APEI bit. > - generic _OSC failed, APEI _OSC succeeded: APEI firmware first mode is > enabled by WHEA _OSC. > - both succeeded: APEI firmware first mode is enabled by APEI bit and > WHEA _OSC. > - both failed: Failed to enable APEI firmware first mode! > > >> I don't think there's anything in _OSC that mentions "firmware first." >> >> I don't remember anything in the spec about a way to *enable* Firmware >> First Error Handling needing (I'm looking at ACPI v6.3, sec 18.4). >> >> I think the "firmware first" information is useless to the OS -- as >> far as I can tell, the spec says nothing about anything the OS should >> do based on the FIRMWARE_FIRST bits.] > > GHES works in "Firmware First" mode to report platform hardware error > and it depends on APEI interface. (drivers/acpi/apei/ghes.c) > > * Generic Hardware Error Source provides a way to report platform > * hardware errors (such as that from chipset). It works in so called > * "Firmware First" mode, that is, hardware errors are reported to > * firmware firstly, then reported to Linux by firmware. > > Based on above, I think the GHES message just means that this booting > kernel supports APEI or WHEA, and GHES report Firmware first errors on > this machine. Is it fine to keep the message? > > >>> As we can see, the initialization of GHES is advanced by 33 seconds. >>> So, in my opinion, this patch is necessary, right? >>> (It should be noted that the effect of optimization varies with the platform.) >> >>>>> -device_initcall(ghes_init); >>>> >>>>> void __init acpi_pci_root_init(void) >>>>> { >>>>> acpi_hest_init(); >>>>> + sdei_init(); >>>>> + ghes_init(); >>>> >>>> What's the connection between PCI, SDEI, and GHES? As far as I can >>>> tell, SDEI and GHES are not PCI-specific, so it doesn't seem like they >>>> should be initialized here in acpi_pci_root_init(). >>> >>> The only reason is that acpi_hest_init() is initialized here. >>> >>> From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus >>> memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage >>> the estatus memory pool. On the other hand, ghes_init() relies on >>> sdei_init() to detect the SDEI version and the framework for registering >>> and unregistering events. The dependencies are as follows >>> >>> ghes_init() => acpi_hest_init() >>> ghes_init() => sdei_init() >>> >>> I don't figure out why acpi_hest_init() is called in >>> acpi_pci_root_init(), it don't rely on any other thing. >>> I am wondering that should we moved all of them further? e.g. >>> following acpi_iort_init() in acpi_init(). >> >> I don't know why acpi_hest_init() is called from acpi_pci_root_init(). >> It looks like HEST can support error sources other than PCI (IA-32 >> Machine Checks, NMIs, GHES, etc.) It was added by 415e12b23792 >> ("PCI/ACPI: Request _OSC control once for each root bridge (v3)"); >> maybe Rafael remembers why. >> >> Seem like acpi_hest_init(), sdei_init(), and ghes_init() should all go >> somewhere else, but I don't know where. Maybe somewhere in >> acpi_init()? > > I tried to move them into acpi_init(), and it works well. > > @@ -1251,6 +1251,9 @@ static int __init acpi_init(void) > > pci_mmcfg_late_init(); > acpi_iort_init(); > + acpi_hest_init(); > + sdei_init(); > + ghes_init(); > acpi_scan_init(); > acpi_ec_init(); > > What's your opinion? > > Best Regards, > > Shuai > > > [1] https://uefi.org/specs/ACPI/6.4/06_Device_Configuration/Device_Configuration.html#platform-wide-osc-capabilities-dword-2 > [2] https://www.mail-archive.com/ubuntu-bugs@lists.ubuntu.com/msg4718503.html > [3] http://www.guyreams.com/PDF/whea.pdf > [4] https://patchwork.kernel.org/project/linux-acpi/patch/1308640587-24502-5-git-send-email-ying.huang@intel.com/#1978922
diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c index 0c8330ed1ffd..b11e46fb4b3d 100644 --- a/drivers/acpi/apei/ghes.c +++ b/drivers/acpi/apei/ghes.c @@ -1457,27 +1457,26 @@ static struct platform_driver ghes_platform_driver = { .remove = ghes_remove, }; -static int __init ghes_init(void) +void __init ghes_init(void) { int rc; if (acpi_disabled) - return -ENODEV; + return; switch (hest_disable) { case HEST_NOT_FOUND: - return -ENODEV; + pr_info(GHES_PFX "HEST is not found!\n"); + return; case HEST_DISABLED: pr_info(GHES_PFX "HEST is not enabled!\n"); - return -EINVAL; + return; default: break; } - if (ghes_disable) { + if (ghes_disable) pr_info(GHES_PFX "GHES is not enabled!\n"); - return -EINVAL; - } ghes_nmi_init_cxt(); @@ -1495,8 +1494,7 @@ static int __init ghes_init(void) else pr_info(GHES_PFX "Failed to enable APEI firmware first mode.\n"); - return 0; + return; err: - return rc; + ghes_disable = 1; } -device_initcall(ghes_init); diff --git a/drivers/acpi/pci_root.c b/drivers/acpi/pci_root.c index ab2f7dfb0c44..1260bb556184 100644 --- a/drivers/acpi/pci_root.c +++ b/drivers/acpi/pci_root.c @@ -23,7 +23,7 @@ #include <linux/dmi.h> #include <linux/platform_data/x86/apple.h> #include <acpi/apei.h> /* for acpi_hest_init() */ - +#include <linux/arm_sdei.h> /* for sdei_init() */ #include "internal.h" #define ACPI_PCI_ROOT_CLASS "pci_bridge" @@ -946,6 +946,9 @@ struct pci_bus *acpi_pci_root_create(struct acpi_pci_root *root, void __init acpi_pci_root_init(void) { acpi_hest_init(); + sdei_init(); + ghes_init(); + if (acpi_pci_disabled) return; diff --git a/drivers/firmware/arm_sdei.c b/drivers/firmware/arm_sdei.c index a7e762c352f9..1e1a51510e83 100644 --- a/drivers/firmware/arm_sdei.c +++ b/drivers/firmware/arm_sdei.c @@ -1059,14 +1059,14 @@ static bool __init sdei_present_acpi(void) return true; } -static int __init sdei_init(void) +void __init sdei_init(void) { struct platform_device *pdev; int ret; ret = platform_driver_register(&sdei_driver); if (ret || !sdei_present_acpi()) - return ret; + return; pdev = platform_device_register_simple(sdei_driver.driver.name, 0, NULL, 0); @@ -1076,17 +1076,8 @@ static int __init sdei_init(void) pr_info("Failed to register ACPI:SDEI platform device %d\n", ret); } - - return ret; } -/* - * On an ACPI system SDEI needs to be ready before HEST:GHES tries to register - * its events. ACPI is initialised from a subsys_initcall(), GHES is initialised - * by device_initcall(). We want to be called in the middle. - */ -subsys_initcall_sync(sdei_init); - int sdei_event_handler(struct pt_regs *regs, struct sdei_registered_event *arg) { diff --git a/include/acpi/apei.h b/include/acpi/apei.h index ece0a8af2bae..7dbd6363fda7 100644 --- a/include/acpi/apei.h +++ b/include/acpi/apei.h @@ -27,8 +27,10 @@ extern int hest_disable; extern int erst_disable; #ifdef CONFIG_ACPI_APEI_GHES extern bool ghes_disable; +void __init ghes_init(void); #else #define ghes_disable 1 +static inline void ghes_init(void) { return; } #endif #ifdef CONFIG_ACPI_APEI diff --git a/include/linux/arm_sdei.h b/include/linux/arm_sdei.h index 0a241c5c911d..9c987188b692 100644 --- a/include/linux/arm_sdei.h +++ b/include/linux/arm_sdei.h @@ -46,9 +46,11 @@ int sdei_unregister_ghes(struct ghes *ghes); /* For use by arch code when CPU hotplug notifiers are not appropriate. */ int sdei_mask_local_cpu(void); int sdei_unmask_local_cpu(void); +void __init sdei_init(void); #else static inline int sdei_mask_local_cpu(void) { return 0; } static inline int sdei_unmask_local_cpu(void) { return 0; } +static inline void sdei_init(void) { return ; } #endif /* CONFIG_ARM_SDE_INTERFACE */
On an ACPI system, ACPI is initialised very early from a subsys_initcall(), while SDEI is not ready until a subsys_initcall_sync(). The SDEI driver provides functions (e.g. apei_sdei_register_ghes, apei_sdei_unregister_ghes) to register or unregister event callback for dispatcher in firmware. When the GHES driver probing, it registers the corresponding callback according to the notification type specified by GHES. If the GHES notification type is SDEI, the GHES driver will call apei_sdei_register_ghes to register event call. When the firmware emits an event, it migrates the handling of the event into the kernel at the registered entry-point __sdei_asm_handler. And finally, the kernel will call the registered event callback and return status_code to indicate the status of event handling. SDEI_EV_FAILED indicates that the kernel failed to handle the event. Consequently, when an error occurs during kernel booting, the kernel is unable to handle and report errors until the GHES driver is initialized by device_initcall(), in which the event callback is registered. All errors that occurred before GHES initialization are missed and there is no chance to report and find them again. From commit e147133a42cb ("ACPI / APEI: Make hest.c manage the estatus memory pool") was merged, ghes_init() relies on acpi_hest_init() to manage the estatus memory pool. On the other hand, ghes_init() relies on sdei_init() to detect the SDEI version and the framework for registering and unregistering events. By the way, I don't figure out why acpi_hest_init is called in acpi_pci_root_init, it don't rely on any other thing. May it could be moved further, following acpi_iort_init in acpi_init. sdei_init() relies on ACPI table which is initialized subsys_initcall(): acpi_init(), acpi_bus_init(), acpi_load_tables(), acpi_tb_laod_namespace(). May it should be also moved further, after acpi_load_tables. In this patch, move sdei_init and ghes_init as far ahead as possible, right after acpi_hest_init(). Signed-off-by: Shuai Xue <xueshuai@linux.alibaba.com> --- drivers/acpi/apei/ghes.c | 18 ++++++++---------- drivers/acpi/pci_root.c | 5 ++++- drivers/firmware/arm_sdei.c | 13 ++----------- include/acpi/apei.h | 2 ++ include/linux/arm_sdei.h | 2 ++ 5 files changed, 18 insertions(+), 22 deletions(-)