Make ghes_edac a proper module
Subsystems (15)
v8 cover Make ghes_edac a proper module
Commit dc4e8c07e9e2 ("ACPI: APEI: explicit init of HEST and GHES in apci_init()") introduced a bug that ghes_edac_register() would be invoked before edac_init(). Because at that time, the bus "edac" hasn't been even registered, this created sysfs /devices/mc0 instead of /sys/devices/system/edac/mc/mc0 on an Ampere eMag server.
The solution is to make ghes_edac a proper module.
Changelog:
v8:
- merge v7 two force_enable and ghes_get_devices() patches into one
- make force_enable static
v7:https://lore.kernel.org/lkml/[email protected]/
- remove the ghes_edac_preferred and ghes_present (suggested by Borislav)
- adjust the patch splitting, no major functional changes
- remove the r-b tag in those changed patches
v6:https://www.spinics.net/lists/kernel/msg4511453.html
- no code changes from v5 patches
- add the reviewed and acked by from Toshi
- describe the removal of ghes_edac_force_enable checking in Patch 05
v5: https://www.spinics.net/lists/kernel/msg4502787.html
- add the review-by from Toshi for patch 04 and 06
- refine the commit msg
- remove the unconditional set of ghes_edac_force_enable on Arm
v4: https://lore.kernel.org/lkml/[email protected]/
- move the kernel boot option to ghes module parameter
- collapse th ghes_present and ghes_edac_preferred into one patch
v3: https://lore.kernel.org/lkml/[email protected]/
- refine the commit logs
- introduce ghes preferred and present flag (by Toshi)
- move force_load to setup parameter
- add the ghes_edac_preferred() check for x86/Arm edac drivers
v2: https://lore.kernel.org/lkml/[email protected]/
- add acked-by tag of Patch 1 from Ard
- split the notifier patch
- add 2 patch to get regular drivers selected when ghes edac is not loaded
- fix an errno in igen6 driver
- add a patch to fix the sparse warning of ghes
- refine the commit logs
v1: https://lore.kernel.org/lkml/[email protected]/
Jia He (7):
efi/cper: export several helpers for ghes_edac to use EDAC/ghes: Add a notifier for reporting memory errors EDAC/ghes: Prepare to make ghes_edac a proper module EDAC/ghes: Make ghes_edac a proper module to remove the dependency on
ghes
EDAC: Add the ghes_get_devices() check for chipset-specific edac
drivers
apei/ghes: Use unrcu_pointer for cmpxchg
EDAC/igen6: Return consistent errno when another edac driver is
enabled
drivers/acpi/apei/ghes.c | 69 +++++++++++++++++++++++--- drivers/edac/Kconfig | 4 +- drivers/edac/amd64_edac.c | 3 ++ drivers/edac/armada_xp_edac.c | 3 ++ drivers/edac/edac_module.h | 1 + drivers/edac/ghes_edac.c | 90 +++++++++++++++++++++------------- drivers/edac/i10nm_base.c | 3 ++ drivers/edac/igen6_edac.c | 5 +- drivers/edac/layerscape_edac.c | 3 ++ drivers/edac/pnd2_edac.c | 3 ++ drivers/edac/sb_edac.c | 3 ++ drivers/edac/skx_base.c | 3 ++ drivers/edac/thunderx_edac.c | 3 ++ drivers/edac/xgene_edac.c | 3 ++ drivers/firmware/efi/cper.c | 3 ++ include/acpi/ghes.h | 34 ++++--------- 16 files changed, 164 insertions(+), 69 deletions(-)
Signature · 1 line
2.25.1
PATCH v8 1/7 efi/cper: export several helpers for ghes_edac to use
Before ghes_edac can be turned back into a proper module again, export several helpers which are going to be used by it.
--- drivers/firmware/efi/cper.c | 3 +++ 1 file changed, 3 insertions(+)
drivers/firmware/efi/cper.c +3 -0
@@ -290,6 +290,7 @@ int cper_mem_err_location(struct cper_mem_err_compact *mem, char *msg)return n; } +EXPORT_SYMBOL_GPL (cper_mem_err_location);int cper_dimm_err_location (struct cper_mem_err_compact *mem,char *msg) {
@@ -310,6 +311,7 @@ int cper_dimm_err_location(struct cper_mem_err_compact *mem, char *msg)return n; } +EXPORT_SYMBOL_GPL (cper_dimm_err_location);void cper_mem_err_pack (const struct cper_sec_mem_err *mem,struct cper_mem_err_compact *cmem)
@@ -331,6 +333,7 @@ void cper_mem_err_pack(const struct cper_sec_mem_err *mem, cmem->mem_array_handle = mem->mem_array_handle; cmem->mem_dev_handle = mem->mem_dev_handle; } +EXPORT_SYMBOL_GPL (cper_mem_err_pack);const char *cper_mem_err_unpack (struct trace_seq *p,struct cper_mem_err_compact *cmem)
Signature · 1 line
2.25.1
PATCH v8 2/7 EDAC/ghes: Add a notifier for reporting memory errors
In reply to Jia HeTo make a proper module, add a notifier for reporting memory errors. Use an atomic notifier because calls sites like ghes_proc_in_irq() run in interrupt context.
--- drivers/acpi/apei/ghes.c | 16 +++++++++++++++- drivers/edac/ghes_edac.c | 19 +++++++++++++++++-- include/acpi/ghes.h | 10 +++------- 3 files changed, 35 insertions(+), 10 deletions(-)
drivers/acpi/apei/ghes.c +15 -1
@@ -94,6 +94,8 @@#define FIX_APEI_GHES_SDEI_CRITICAL __end_of_fixed_addresses#endif +static ATOMIC_NOTIFIER_HEAD (ghes_report_chain); +static inline bool is_hest_type_generic_v2 (struct ghes *ghes) {return ghes->generic->header.type ==ACPI_HEST_TYPE_GENERIC_ERROR_V2 ;
@@ -645,7 +647,7 @@ static bool ghes_do_proc(struct ghes *ghes,if (guid_equal (sec_type, &CPER_SEC_PLATFORM_MEM )) {struct cper_sec_mem_err *mem_err =acpi_hest_get_payload (gdata); -ghes_edac_report_mem_error (sev, mem_err); +atomic_notifier_call_chain (&ghes_report_chain, sev, mem_err);arch_apei_report_mem_error (sev, mem_err); queued =ghes_handle_memory_failure (gdata, sev);
@@ -1497,3 +1499,15 @@ void __init acpi_ghes_init(void)else pr_info (GHES_PFX "Failed to enable APEI firmware first mode.\n" ); } + +void ghes_register_report_chain (struct notifier_block *nb) +{ +atomic_notifier_chain_register (&ghes_report_chain, nb); +} +EXPORT_SYMBOL_GPL (ghes_register_report_chain); + +void ghes_unregister_report_chain (struct notifier_block *nb) +{ +atomic_notifier_chain_unregister (&ghes_report_chain, nb); +} +EXPORT_SYMBOL_GPL (ghes_unregister_report_chain);
drivers/edac/ghes_edac.c +17 -2
@@ -14,6 +14,7 @@#include <linux/dmi.h> #include "edac_module.h" #include <ras/ras_event.h> +#include <linux/notifier.h> #define OTHER_DETAIL_LEN 400
@@ -267,11 +268,14 @@ static int print_mem_error_other_detail(const struct cper_sec_mem_err *mem, charreturn n; } -void ghes_edac_report_mem_error (int sev,struct cper_sec_mem_err *mem_err) +static int ghes_edac_report_mem_error (struct notifier_block *nb, +unsigned long val,void *data) { +struct cper_sec_mem_err *mem_err = (struct cper_sec_mem_err *)data;struct cper_mem_err_compact cmem;struct edac_raw_error_desc *e;struct mem_ctl_info *mci; +unsigned long sev = val;struct ghes_pvt *pvt;unsigned long flags;char *p;
@@ -282,7 +286,7 @@ void ghes_edac_report_mem_error(int sev, struct cper_sec_mem_err *mem_err)* know. */ if (WARN_ON_ONCE (in_nmi ())) -return ; +return NOTIFY_OK ;spin_lock_irqsave (&ghes_lock, flags);
@@ -374,8 +378,15 @@ void ghes_edac_report_mem_error(int sev, struct cper_sec_mem_err *mem_err) unlock:spin_unlock_irqrestore (&ghes_lock, flags); + +return NOTIFY_OK ; } +static struct notifier_block ghes_edac_mem_err_nb = { + .notifier_call = ghes_edac_report_mem_error, + .priority =0 , +}; +/* * Known systems that are safe to enable this module. */
@@ -503,6 +514,8 @@ int ghes_edac_register(struct ghes *ghes, struct device *dev) ghes_pvt = pvt;spin_unlock_irqrestore (&ghes_lock, flags); +ghes_register_report_chain (&ghes_edac_mem_err_nb); +/* only set on success */ refcount_set (&ghes_refcount,1 );
@@ -548,6 +561,8 @@ void ghes_edac_unregister(struct ghes *ghes)if (mci)edac_mc_free (mci); +ghes_unregister_report_chain (&ghes_edac_mem_err_nb); + unlock:mutex_unlock (&ghes_reg_mutex); }
include/acpi/ghes.h +3 -7
@@ -76,18 +76,11 @@ int ghes_estatus_pool_init(int num_ghes);/* From drivers/edac/ghes_edac.c */ #ifdef CONFIG_EDAC_GHES -void ghes_edac_report_mem_error (int sev,struct cper_sec_mem_err *mem_err); -int ghes_edac_register (struct ghes *ghes,struct device *dev);void ghes_edac_unregister (struct ghes *ghes);#else -static inline void ghes_edac_report_mem_error (int sev, -struct cper_sec_mem_err *mem_err) -{ -} -static inline int ghes_edac_register (struct ghes *ghes,struct device *dev) {return -ENODEV ;
@@ -145,4 +138,7 @@ int ghes_notify_sea(void);static inline int ghes_notify_sea (void ) {return -ENOENT ; }#endif +struct notifier_block ; +extern void ghes_register_report_chain (struct notifier_block *nb); +extern void ghes_unregister_report_chain (struct notifier_block *nb);#endif /* GHES_H */
Signature · 1 line
2.25.1
PATCH v8 3/7 EDAC/ghes: Prepare to make ghes_edac a proper module
In reply to Jia HeTo make ghes_edac a proper module, prepare to decouple its dependencies on ghes.
ghes_edac_register() is too late to set the module flag ghes_edac.force_load. Also, other edac drivers should not be able to control this flag. Move this flag to the module parameter in ghes and make it static.
Introduce a helper ghes_get_devices(), which returns the dev list GHES probed when the platform-check passes on the system.
The previous force_load check is not needed in ghes_edac_unregister() since it will be checked in module_init of ghes_edac later.
--- drivers/acpi/apei/ghes.c | 46 ++++++++++++++++++++++++++++++++++++++++ drivers/edac/ghes_edac.c | 35 ++---------------------------- include/acpi/ghes.h | 6 ++++++ 3 files changed, 54 insertions(+), 33 deletions(-)
drivers/acpi/apei/ghes.c +46 -0
@@ -109,6 +109,13 @@ static inline bool is_hest_type_generic_v2(struct ghes *ghes)bool ghes_disable;module_param_named (disable, ghes_disable, bool,0 ); +/* +* "ghes.edac_force_enable" forcibly enables ghes_edac and skips the platform +* check. +*/ +static bool ghes_edac_force_enable; +module_param_named (edac_force_enable, ghes_edac_force_enable, bool,0 ); +/* * All error sources notified with HED (Hardware Error Device) share a * single notifier callback, so they need to be linked and checked one
@@ -120,6 +127,9 @@ module_param_named(disable, ghes_disable, bool, 0);static LIST_HEAD (ghes_hed);static DEFINE_MUTEX (ghes_list_mutex); +static LIST_HEAD (ghes_devs); +static DEFINE_MUTEX (ghes_devs_mutex); +/* * Because the memory area used to transfer hardware error information * from BIOS to Linux can be determined only in NMI, IRQ or timer
@@ -1380,6 +1390,12 @@ static int ghes_probe(struct platform_device *ghes_dev)ghes_edac_register (ghes, &ghes_dev->dev); + ghes->dev = &ghes_dev->dev; + +mutex_lock (&ghes_devs_mutex); +list_add_tail (&ghes->elist, &ghes_devs); +mutex_unlock (&ghes_devs_mutex); +/* Handle any pending errors right away */ spin_lock_irqsave (&ghes_notify_lock_irq, flags);ghes_proc (ghes);
@@ -1444,6 +1460,10 @@ static int ghes_remove(struct platform_device *ghes_dev)ghes_edac_unregister (ghes); +mutex_lock (&ghes_devs_mutex); +list_del (&ghes->elist); +mutex_unlock (&ghes_devs_mutex); +kfree (ghes);platform_set_drvdata (ghes_dev,NULL );
@@ -1500,6 +1520,32 @@ void __init acpi_ghes_init(void)pr_info (GHES_PFX "Failed to enable APEI firmware first mode.\n" ); } +/* +* Known x86 systems that prefer GHES error reporting: +*/ +static struct acpi_platform_list plat_list[] = { + {"HPE " ,"Server " ,0 ,ACPI_SIG_FADT , all_versions}, + { }/* End */ +}; + +struct list_head *ghes_get_devices (void ) +{ +int idx =-1 ; + +if (IS_ENABLED (CONFIG_X86 )) { + idx =acpi_match_platform_list (plat_list); +if (idx <0 ) { +if (!ghes_edac_force_enable) +return NULL ; + +pr_warn_once ("Force-loading ghes_edac on an unsupported platform. You're on your own!\n" ); + } + } + +return &ghes_devs; +} +EXPORT_SYMBOL_GPL (ghes_get_devices); +void ghes_register_report_chain (struct notifier_block *nb) {atomic_notifier_chain_register (&ghes_report_chain, nb);
drivers/edac/ghes_edac.c +2 -33
@@ -54,10 +54,6 @@ static DEFINE_MUTEX(ghes_reg_mutex);*/ static DEFINE_SPINLOCK (ghes_lock); -/* "ghes_edac.force_load=1" skips the platform check */ -static bool __read_mostly force_load; -module_param (force_load, bool,0 ); -static bool system_scanned;/* Memory Device - Type 17 of SMBIOS spec */
@@ -387,14 +383,6 @@ static struct notifier_block ghes_edac_mem_err_nb = { .priority = 0, }; -/* -* Known systems that are safe to enable this module. -*/ -static struct acpi_platform_list plat_list[] = { - {"HPE " ,"Server " ,0 ,ACPI_SIG_FADT , all_versions}, - { }/* End */ -}; -int ghes_edac_register (struct ghes *ghes,struct device *dev) {bool fake = false;
@@ -402,19 +390,8 @@ int ghes_edac_register(struct ghes *ghes, struct device *dev)struct ghes_pvt *pvt;struct edac_mc_layer layers[1 ];unsigned long flags; -int idx =-1 ;int rc =0 ; -if (IS_ENABLED (CONFIG_X86 )) { -/* Check if safe to enable on this system */ - idx =acpi_match_platform_list (plat_list); -if (!force_load && idx <0 ) -return -ENODEV ; - }else { - force_load = true; - idx =0 ; - } -/* finish another registration/unregistration instance first */ mutex_lock (&ghes_reg_mutex);
@@ -458,15 +435,10 @@ int ghes_edac_register(struct ghes *ghes, struct device *dev)pr_info ("This system has a very crappy BIOS: It doesn't even list the DIMMS.\n" );pr_info ("Its SMBIOS info is wrong. It is doubtful that the error report would\n" );pr_info ("work on such system. Use this driver with caution\n" ); - }else if (idx <0 ) { -pr_info ("This EDAC driver relies on BIOS to enumerate memory and get error reports.\n" ); -pr_info ("Unfortunately, not all BIOSes reflect the memory layout correctly.\n" ); -pr_info ("So, the end result of using this driver varies from vendor to vendor.\n" ); -pr_info ("If you find incorrect reports, please contact your hardware vendor\n" ); -pr_info ("to correct its BIOS.\n" ); -pr_info ("This system has %d DIMM sockets.\n" , ghes_hw.num_dimms); } +pr_info ("This system has %d DIMM sockets.\n" , ghes_hw.num_dimms); +if (!fake) {struct dimm_info *src, *dst;int i =0 ;
@@ -535,9 +507,6 @@ void ghes_edac_unregister(struct ghes *ghes)struct mem_ctl_info *mci;unsigned long flags; -if (!force_load) -return ; -mutex_lock (&ghes_reg_mutex); system_scanned = false;
include/acpi/ghes.h +6 -0
@@ -27,6 +27,8 @@ struct ghes {struct timer_list timer;unsigned int irq; }; +struct device *dev; +struct list_head elist; };struct ghes_estatus_node {
@@ -80,6 +82,8 @@ int ghes_edac_register(struct ghes *ghes, struct device *dev);void ghes_edac_unregister (struct ghes *ghes); +struct list_head *ghes_get_devices (void ); +#else static inline int ghes_edac_register (struct ghes *ghes,struct device *dev) {
@@ -89,6 +93,8 @@ static inline int ghes_edac_register(struct ghes *ghes, struct device *dev)static inline void ghes_edac_unregister (struct ghes *ghes) { } + +static inline struct list_head *ghes_get_devices (void ) {return NULL ; }#endif static inline int acpi_hest_get_version (struct acpi_hest_generic_data *gdata)
Signature · 1 line
2.25.1
PATCH v8 4/7 EDAC/ghes: Make ghes_edac a proper module to remove the dependency on ghes
In reply to Jia HeCommit dc4e8c07e9e2 ("ACPI: APEI: explicit init of HEST and GHES in apci_init()") introduced a bug that ghes_edac_register() would be invoked before edac_init(). Because at that time, the bus "edac" hadn't been even registered, this created sysfs /devices/mc0 instead of /sys/devices/system/edac/mc/mc0 on an Ampere eMag server.
To remove the dependency of ghes_edac on ghes, make it a proper module. Use a list to save the probing devices in ghes_probe(), and defer the ghes_edac_register() to module_init() of the new ghes_edac module by iterating over the devices list.
--- drivers/acpi/apei/ghes.c | 4 ---- drivers/edac/Kconfig | 4 ++-- drivers/edac/ghes_edac.c | 40 ++++++++++++++++++++++++++++++++++++++-- include/acpi/ghes.h | 22 ++-------------------- 4 files changed, 42 insertions(+), 28 deletions(-)
drivers/acpi/apei/ghes.c +0 -4
@@ -1388,8 +1388,6 @@ static int ghes_probe(struct platform_device *ghes_dev)platform_set_drvdata (ghes_dev, ghes); -ghes_edac_register (ghes, &ghes_dev->dev); - ghes->dev = &ghes_dev->dev;mutex_lock (&ghes_devs_mutex);
@@ -1458,8 +1456,6 @@ static int ghes_remove(struct platform_device *ghes_dev)ghes_fini (ghes); -ghes_edac_unregister (ghes); -mutex_lock (&ghes_devs_mutex);list_del (&ghes->elist);mutex_unlock (&ghes_devs_mutex);
drivers/edac/Kconfig +2 -2
@@ -53,8 +53,8 @@ config EDAC_DECODE_MCE has been initialized. config EDAC_GHES - bool "Output ACPI APEI/GHES BIOS detected errors via EDAC" - depends on ACPI_APEI_GHES && (EDAC=y) + tristate "Output ACPI APEI/GHES BIOS detected errors via EDAC" + depends on ACPI_APEI_GHES select UEFI_CPER help Not all machines support hardware-driven error report. Some of those
drivers/edac/ghes_edac.c +38 -2
@@ -56,6 +56,8 @@ static DEFINE_SPINLOCK(ghes_lock);static bool system_scanned; +static struct list_head *ghes_devs; +/* Memory Device - Type 17 of SMBIOS spec */ struct memdev_dmi_entry {u8 type;
@@ -383,7 +385,7 @@ static struct notifier_block ghes_edac_mem_err_nb = { .priority = 0, }; -int ghes_edac_register (struct ghes *ghes,struct device *dev) +static int ghes_edac_register (struct device *dev) {bool fake = false;struct mem_ctl_info *mci;
@@ -502,7 +504,7 @@ int ghes_edac_register(struct ghes *ghes, struct device *dev)return rc; } -void ghes_edac_unregister (struct ghes *ghes) +static void ghes_edac_unregister (struct ghes *ghes) {struct mem_ctl_info *mci;unsigned long flags;
@@ -535,3 +537,37 @@ void ghes_edac_unregister(struct ghes *ghes) unlock:mutex_unlock (&ghes_reg_mutex); } + +static int __init ghes_edac_init (void ) +{ +struct ghes *g, *g_tmp; + + ghes_devs =ghes_get_devices (); +if (!ghes_devs) +return -ENODEV ; + +if (list_empty (ghes_devs)) { +pr_info ("GHES probing device list is empty" ); +return -ENODEV ; + } + +list_for_each_entry_safe (g, g_tmp, ghes_devs, elist) { +ghes_edac_register (g->dev); + } + +return 0 ; +} +module_init (ghes_edac_init); + +static void __exit ghes_edac_exit(void) +{ +struct ghes *g, *g_tmp; + +list_for_each_entry_safe (g, g_tmp, ghes_devs, elist) { +ghes_edac_unregister (g); + } +} +module_exit (ghes_edac_exit); + +MODULE_LICENSE ("GPL" ); +MODULE_DESCRIPTION ("Output ACPI APEI/GHES BIOS detected errors module via EDAC" );
include/acpi/ghes.h +2 -20
@@ -71,32 +71,14 @@ int ghes_register_vendor_record_notifier(struct notifier_block *nb);* @nb: pointer to the notifier_block structure of the vendor record handler. */ void ghes_unregister_vendor_record_notifier (struct notifier_block *nb); -#endif - -int ghes_estatus_pool_init (int num_ghes); - -/* From drivers/edac/ghes_edac.c */ - -#ifdef CONFIG_EDAC_GHES -int ghes_edac_register (struct ghes *ghes,struct device *dev); - -void ghes_edac_unregister (struct ghes *ghes);struct list_head *ghes_get_devices (void ); -#else -static inline int ghes_edac_register (struct ghes *ghes,struct device *dev) -{ -return -ENODEV ; -} - -static inline void ghes_edac_unregister (struct ghes *ghes) -{ -} -static inline struct list_head *ghes_get_devices (void ) {return NULL ; }#endif +int ghes_estatus_pool_init (int num_ghes); +static inline int acpi_hest_get_version (struct acpi_hest_generic_data *gdata) {return gdata->revision >>8 ;
Signature · 1 line
2.25.1
PATCH v8 5/7 EDAC: Add the ghes_get_devices() check for chipset-specific edac drivers
In reply to Jia HeAdd ghes_get_devices() check for chipset-specific edac drivers to ensure that ghes_edac is used on the platform where ghes_edac is preferred over chipset-specific edac driver.
Unlike the existing edac_get_owner() check, the ghes_get_devices() check works independent to the module_init ordering.
--- drivers/edac/amd64_edac.c | 3 +++ drivers/edac/armada_xp_edac.c | 3 +++ drivers/edac/edac_module.h | 1 + drivers/edac/i10nm_base.c | 3 +++ drivers/edac/igen6_edac.c | 3 +++ drivers/edac/layerscape_edac.c | 3 +++ drivers/edac/pnd2_edac.c | 3 +++ drivers/edac/sb_edac.c | 3 +++ drivers/edac/skx_base.c | 3 +++ drivers/edac/thunderx_edac.c | 3 +++ drivers/edac/xgene_edac.c | 3 +++ 11 files changed, 31 insertions(+)
drivers/edac/amd64_edac.c +3 -0
@@ -4329,6 +4329,9 @@ static int __init amd64_edac_init(void)int err = -ENODEV ;int i; +if (ghes_get_devices ()) +return -EBUSY ; + owner =edac_get_owner ();if (owner &&strncmp (owner,EDAC_MOD_STR ,sizeof (EDAC_MOD_STR )))return -EBUSY ;
drivers/edac/armada_xp_edac.c +3 -0
@@ -599,6 +599,9 @@ static int __init armada_xp_edac_init(void) {int res; +if (ghes_get_devices ()) +return -EBUSY ; +/* only polling is supported */ edac_op_state =EDAC_OPSTATE_POLL ;
drivers/edac/edac_module.h +1 -0
@@ -11,6 +11,7 @@#ifndef __EDAC_MODULE_H__#define __EDAC_MODULE_H__ +#include <acpi/ghes.h> #include "edac_mc.h" #include "edac_pci.h" #include "edac_device.h"
drivers/edac/i10nm_base.c +3 -0
@@ -548,6 +548,9 @@ static int __init i10nm_init(void)edac_dbg (2 ,"\n" ); +if (ghes_get_devices ()) +return -EBUSY ; + owner =edac_get_owner ();if (owner &&strncmp (owner,EDAC_MOD_STR ,sizeof (EDAC_MOD_STR )))return -EBUSY ;
drivers/edac/igen6_edac.c +3 -0
@@ -1271,6 +1271,9 @@ static int __init igen6_init(void)edac_dbg (2 ,"\n" ); +if (ghes_get_devices ()) +return -EBUSY ; + owner =edac_get_owner ();if (owner &&strncmp (owner,EDAC_MOD_STR ,sizeof (EDAC_MOD_STR )))return -ENODEV ;
drivers/edac/layerscape_edac.c +3 -0
@@ -38,6 +38,9 @@ static int __init fsl_ddr_mc_init(void) {int res; +if (ghes_get_devices ()) +return -EBUSY ; +/* make sure error reporting method is sane */ switch (edac_op_state) {case EDAC_OPSTATE_POLL :
drivers/edac/pnd2_edac.c +3 -0
@@ -1528,6 +1528,9 @@ static int __init pnd2_init(void)edac_dbg (2 ,"\n" ); +if (ghes_get_devices ()) +return -EBUSY ; + owner =edac_get_owner ();if (owner &&strncmp (owner,EDAC_MOD_STR ,sizeof (EDAC_MOD_STR )))return -EBUSY ;
drivers/edac/sb_edac.c +3 -0
@@ -3506,6 +3506,9 @@ static int __init sbridge_init(void)edac_dbg (2 ,"\n" ); +if (ghes_get_devices ()) +return -EBUSY ; + owner =edac_get_owner ();if (owner &&strncmp (owner,EDAC_MOD_STR ,sizeof (EDAC_MOD_STR )))return -EBUSY ;
drivers/edac/skx_base.c +3 -0
@@ -653,6 +653,9 @@ static int __init skx_init(void)edac_dbg (2 ,"\n" ); +if (ghes_get_devices ()) +return -EBUSY ; + owner =edac_get_owner ();if (owner &&strncmp (owner,EDAC_MOD_STR ,sizeof (EDAC_MOD_STR )))return -EBUSY ;
drivers/edac/thunderx_edac.c +3 -0
@@ -2114,6 +2114,9 @@ static int __init thunderx_edac_init(void) {int rc =0 ; +if (ghes_get_devices ()) +return -EBUSY ; + rc =pci_register_driver (&thunderx_lmc_driver);if (rc)return rc;
drivers/edac/xgene_edac.c +3 -0
@@ -2004,6 +2004,9 @@ static int __init xgene_edac_init(void) {int rc; +if (ghes_get_devices ()) +return -EBUSY ; +/* Make sure error reporting method is sane */ switch (edac_op_state) {case EDAC_OPSTATE_POLL :
Signature · 1 line
2.25.1
PATCH v8 6/7 apei/ghes: Use unrcu_pointer for cmpxchg
In reply to Jia Heghes_estatus_caches should be add rcu annotation to avoid sparse warnings.
drivers/acpi/apei/ghes.c:733:25: sparse: sparse: incompatible types in comparison expression (different address spaces):
drivers/acpi/apei/ghes.c:733:25: sparse: struct ghes_estatus_cache [noderef] __rcu *
drivers/acpi/apei/ghes.c:733:25: sparse: struct ghes_estatus_cache *
drivers/acpi/apei/ghes.c:813:25: sparse: sparse: incompatible types in comparison expression (different address spaces):
drivers/acpi/apei/ghes.c:813:25: sparse: struct ghes_estatus_cache [noderef] __rcu *
drivers/acpi/apei/ghes.c:813:25: sparse: struct ghes_estatus_cache *
unrcu_pointer is to strip the __rcu in cmpxchg.
--- drivers/acpi/apei/ghes.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/acpi/apei/ghes.c +4 -3
@@ -150,7 +150,7 @@ struct ghes_vendor_record_entry {static struct gen_pool *ghes_estatus_pool;static unsigned long ghes_estatus_pool_size_request; -static struct ghes_estatus_cache *ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ]; +static struct ghes_estatus_cache __rcu *ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ];static atomic_t ghes_estatus_cache_alloced;static int ghes_panic_timeout__read_mostly =30 ;
@@ -840,8 +840,9 @@ static void ghes_estatus_cache_add( }/* new_cache must be put into array after its contents are written */ smp_wmb (); -if (slot !=-1 &&cmpxchg (ghes_estatus_caches + slot, - slot_cache, new_cache) == slot_cache) { +if (slot !=-1 &&unrcu_pointer (cmpxchg (ghes_estatus_caches + slot, +RCU_INITIALIZER (slot_cache), +RCU_INITIALIZER (new_cache))) == slot_cache) {if (slot_cache)call_rcu (&slot_cache->rcu, ghes_estatus_cache_rcu_free); }else
Signature · 1 line
2.25.1
Borislav Petkov
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 9 lines · Jia He
ghes_estatus_cachesshould be add rcu annotation to avoid sparse warnings.
drivers/acpi/apei/ghes.c:733:25: sparse: sparse: incompatible types in comparison expression (different address spaces):
drivers/acpi/apei/ghes.c:733:25: sparse: structghes_estatus_cache[noderef]__rcu*
drivers/acpi/apei/ghes.c:733:25: sparse: structghes_estatus_cache*
drivers/acpi/apei/ghes.c:813:25: sparse: sparse: incompatible types in comparison expression (different address spaces):
drivers/acpi/apei/ghes.c:813:25: sparse: structghes_estatus_cache[noderef]__rcu*
drivers/acpi/apei/ghes.c:813:25: sparse: structghes_estatus_cache*
unrcu_pointeris to strip the__rcuin cmpxchg.
Is this only to shut up sparse or actually fixing anything?
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteRE: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Hi Borislav
Quoted 26 lines
-----Original Message-----
From: Borislav Petkov <[email protected]>
Sent: Tuesday, October 11, 2022 6:34 PM
To: Justin He <[email protected]>
Cc: Len Brown <[email protected]>; James Morse <[email protected]>; Tony Luck <[email protected]>; Mauro Carvalho Chehab <[email protected]>; Robert Richter <[email protected]>; Robert Moore <[email protected]>; Qiuxu Zhuo <[email protected]>; Yazen Ghannam <[email protected]>; Jan Luebbe <[email protected]>; Khuong Dinh <[email protected]>; Kani Toshi <[email protected]>; Ard Biesheuvel <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Rafael J . Wysocki <[email protected]>; Shuai Xue <[email protected]>; Jarkko Sakkinen <[email protected]>; [email protected]; nd <[email protected]>; kernel test robot <[email protected]>
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchgQuoted 2 lines · Jia He
ghes_estatus_cachesshould be add rcu annotation to avoid sparse warnings.
drivers/acpi/apei/ghes.c:733:25: sparse: sparse: incompatible types incomparison expression (different address spaces):
Quoted 1 line
drivers/acpi/apei/ghes.c:733:25: sparse: structghes_estatus_cache[noderef]
__rcu*Quoted 2 lines
drivers/acpi/apei/ghes.c:733:25: sparse: structghes_estatus_cache*
drivers/acpi/apei/ghes.c:813:25: sparse: sparse: incompatible types in
Quoted 8 lines
comparison expression (different address spaces):
Quoted 1 line
drivers/acpi/apei/ghes.c:813:25: sparse: structghes_estatus_cache[noderef]
__rcu*Quoted 3 lines
drivers/acpi/apei/ghes.c:813:25: sparse: structghes_estatus_cache*
unrcu_pointeris to strip the__rcuin cmpxchg.Is this only to shut up sparse or actually fixing anything?
My original purpose is to make it pass the sparse checking.
--
Cheers,
Justin (Jia He)
Borislav Petkov
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 1 line · Justin He
My original purpose is to make it pass the sparse checking.
Then do this pls.
This is a combined diff - do a second patch which does only remove the smp_wmb(). The smp_wmb() there is not needed as the cmpxchg() already implies a smp_mb() so there's no need for that separate, explicit one.
But it would be prudent to have it in a separate patch just in case.
Thx.
---
drivers/acpi/apei/ghes.c +2 -4
@@ -154,7 +154,7 @@ struct ghes_vendor_record_entry {static struct gen_pool *ghes_estatus_pool;static unsigned long ghes_estatus_pool_size_request; -static struct ghes_estatus_cache *ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ]; +static struct ghes_estatus_cache __rcu *ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ];static atomic_t ghes_estatus_cache_alloced;static int ghes_panic_timeout__read_mostly =30 ;
@@ -842,9 +842,7 @@ static void ghes_estatus_cache_add( slot_cache = cache; } } -/* new_cache must be put into array after its contents are written */ -smp_wmb (); -if (slot !=-1 &&cmpxchg (ghes_estatus_caches + slot, +if (slot !=-1 &&cmpxchg ((struct ghes_estatus_cache **__force )(ghes_estatus_caches + slot), slot_cache, new_cache) == slot_cache) {if (slot_cache)call_rcu (&slot_cache->rcu, ghes_estatus_cache_rcu_free);
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteRE: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 23 lines
-----Original Message-----
From: Borislav Petkov <[email protected]>
Sent: Tuesday, October 11, 2022 10:46 PM
To: Justin He <[email protected]>
Cc: Len Brown <[email protected]>; James Morse <[email protected]>; Tony Luck <[email protected]>; Mauro Carvalho Chehab <[email protected]>; Robert Richter <[email protected]>; Robert Moore <[email protected]>; Qiuxu Zhuo <[email protected]>; Yazen Ghannam <[email protected]>; Jan Luebbe <[email protected]>; Khuong Dinh <[email protected]>; Kani Toshi <[email protected]>; Ard Biesheuvel <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Rafael J . Wysocki <[email protected]>; Shuai Xue <[email protected]>; Jarkko Sakkinen <[email protected]>; [email protected]; nd <[email protected]>; kernel test robot <[email protected]>
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchgQuoted 1 line · Justin He
My original purpose is to make it pass the sparse checking.
Quoted 8 lines
Then do this pls.
This is a combined diff - do a second patch which does only remove the
smp_wmb(). Thesmp_wmb()there is not needed as thecmpxchg()already implies asmp_mb()so there's no need for that separate, explicit one.But it would be prudent to have it in a separate patch just in case.
Sure, I will insert/append your patch in my v9 series if there are no more comments.
Thanks for the
RE: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
In reply to Borislav PetkovHi Borislav
Quoted 18 lines
-----Original Message-----
From: Borislav Petkov <[email protected]>
Sent: Tuesday, October 11, 2022 10:46 PM
To: Justin He <[email protected]>
Cc: Len Brown <[email protected]>; James Morse <[email protected]>; Tony Luck <[email protected]>; Mauro Carvalho Chehab <[email protected]>; Robert Richter <[email protected]>; Robert Moore <[email protected]>; Qiuxu Zhuo <[email protected]>; Yazen Ghannam <[email protected]>; Jan Luebbe <[email protected]>; Khuong Dinh <[email protected]>; Kani Toshi <[email protected]>; Ard Biesheuvel <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Rafael J . Wysocki <[email protected]>; Shuai Xue <[email protected]>; Jarkko Sakkinen <[email protected]>; [email protected]; nd <[email protected]>; kernel test robot <[email protected]>
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchg
Quoted 9 lines
Quoted 1 line · Justin He
My original purpose is to make it pass the sparse checking.
Then do this pls.
This is a combined diff - do a second patch which does only remove the
smp_wmb(). Thesmp_wmb()there is not needed as thecmpxchg()already implies asmp_mb()so there's no need for that separate, explicit one.
I have a concern about what if cmpxchg failed? Do we have to still guarantee the ordering since cmpxchg will not imply a smp_mb if it failed.
Besides, I didn't find the paired smp_mb or smp_rmb for this smp_wmb. Do you have any ideas?
--
Cheers,
Justin (Jia He)
Borislav Petkov
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 3 lines · Justin He
I have a concern about what if cmpxchg failed? Do we have to still guarantee the ordering since cmpxchg will not imply a
smp_mbif it failed.
Of course it will imply that. At least on x86 it does. smp_wmb() is a compiler barrier there and cmpxchg() already has that barrier semantics by clobbering "memory". I'm pretty sure you should have the same thing on ARM.
And even if that weren't the case, the write barrier is, as the comment says, "new_cache must be put into array after its contents are written".
Are we writing anything into the cache if cmpxchg fails?
Quoted 1 line
Besides, I didn't find the paired
smp_mborsmp_rmbfor thissmp_wmb.
Why would there be pairs? I don't understand that statement here.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteArd Biesheuvel
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 11 lines · Borislav Petkov
Quoted 3 lines · Justin He
I have a concern about what if cmpxchg failed? Do we have to still guarantee the ordering since cmpxchg will not imply a
smp_mbif it failed.Of course it will imply that. At least on x86 it does.
smp_wmb()is a compiler barrier there andcmpxchg()already has that barrier semantics by clobbering "memory". I'm pretty sure you should have the same thing on ARM.
No it definitely does not imply that. A memory clobber is a codegen construct, and the hardware could still complete the writes in a way that could result in another observer seeing a mix of old and new values that is inconsistent with the ordering of the stores as issued by the compiler.
Quoted 4 lines
says, "
new_cachemust be put into array after its contents are written".Are we writing anything into the cache if cmpxchg fails?
The cache fields get updated but the pointer to the struct is never shared globally if the cmpxchg() fails so not having the barrier on failure should not be an issue here.
Quoted 4 lines
Quoted 1 line
Besides, I didn't find the paired
smp_mborsmp_rmbfor thissmp_wmb.Why would there be pairs? I don't understand that statement here.
Typically, the other observer pairs the write barrier with a read barrier.
In this case, the other observer appears to be ghes_estatus_cached(), and the reads of the cache struct fields must be ordered after the read of the cache struct's address. However, there is an implicit ordering there through an address dependency (you cannot dereference a struct without knowing its address) so the ordering is implied (and rcu_dereference() has a READ_ONCE() inside so we are guaranteed to always dereference the same struct, even if the array slot gets updated concurrently.)
If you want to get rid of the barrier, you could drop it and change the cmpxchg() to cmpxchg_release().
Justin: so why are the RCU_INITIALIZER()s needed here?
Borislav Petkov
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 5 lines · Ard Biesheuvel
No it definitely does not imply that. A memory clobber is a codegen construct, and the hardware could still complete the writes in a way that could result in another observer seeing a mix of old and new values that is inconsistent with the ordering of the stores as issued by the compiler.
Yes, but look at the code. There's a:
smp_wmb();
which on x86 is
#define smp_wmb() barrier()
which is
#define barrier() __asm__ __volatile__("": : :"memory")so there wasn't a hardware memory barrier there in the first place.
Unless ARM does something else in those primitives...
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
In reply to Ard BiesheuvelQuoted 13 lines · Ard Biesheuvel
Quoted 11 lines · Borislav Petkov
Quoted 3 lines · Justin He
I have a concern about what if cmpxchg failed? Do we have to still guarantee the ordering since cmpxchg will not imply a
smp_mbif it failed.Of course it will imply that. At least on x86 it does.
smp_wmb()is a compiler barrier there andcmpxchg()already has that barrier semantics by clobbering "memory". I'm pretty sure you should have the same thing on ARM.
Quoted 5 lines
No it definitely does not imply that. A memory clobber is a codegen construct, and the hardware could still complete the writes in a way that could result in another observer seeing a mix of old and new values that is inconsistent with the ordering of the stores as issued by the compiler.
Borislav is thinking too much x86. Failed cmpxchg() does indeed not imply any memory ordering for all architectures -- and while the memory clobber (aka. barrier()) is equivalent to an smp_wmb() on x86, that most certainly doesn't hold for non x86 code.
Quoted 8 lines
Quoted 4 lines
says, "
new_cachemust be put into array after its contents are written".Are we writing anything into the cache if cmpxchg fails?
The cache fields get updated but the pointer to the struct is never shared globally if the
cmpxchg()fails so not having the barrier on failure should not be an issue here.
That is how I read the code too; so if the cmpxchg fails the object is not published and nobody cares about the ordering.
Quoted 17 lines
Quoted 5 lines
Quoted 2 lines
Besides, I didn't find the paired
smp_mborsmp_rmbfor thissmp_wmb.Why would there be pairs? I don't understand that statement here.
Typically, the other observer pairs the write barrier with a read barrier.
In this case, the other observer appears to be
ghes_estatus_cached(), and the reads of the cache struct fields must be ordered after the read of the cache struct's address. However, there is an implicit ordering there through an address dependency (you cannot dereference a struct without knowing its address) so the ordering is implied (andrcu_dereference()has aREAD_ONCE()inside so we are guaranteed to always dereference the same struct, even if the array slot gets updated concurrently.)
Quoted 2 lines
If you want to get rid of the barrier, you could drop it and change the
cmpxchg()tocmpxchg_release().
cmpxchg_release() is strictly weaker than cmpxchg(); so I don't see the point there other than optimizing for weak architectures. It can't fundamentally fix anything.
Borislav Petkov
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 4 lines · Peter Zijlstra
Borislav is thinking too much x86. Failed
cmpxchg()does indeed not imply any memory ordering for all architectures -- and while the memory clobber (aka.barrier()) is equivalent to ansmp_wmb()on x86, that most certainly doesn't hold for non x86 code.
Right, but the patch was addied by an Intel person, CCed:
152cef40a808 ("ACPI, APEI, GHES, Error records content based throttle")a
So I don't think he was thinking about ARM when doing that.
And that commit message doesn't say one whit why that memory barrier is needed there.
Reading that comment, it sounds like he wanted a hw memory barrier - MFENCE - but I don't see how normal data dependency wouldn't enforce the proper order already...
So that barrier looks out of place there.
Btw, this is the next perfect example why I'm asking people to write proper commit messages so that when we do git archeology later, we can figure out why something was done the way it has been.
And in this case, we can't. ;-\
Because writing proper commit messages is for losers. Yeah, right.</sarcasm>
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteArd Biesheuvel
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 13 lines · Borislav Petkov
Quoted 4 lines · Peter Zijlstra
Borislav is thinking too much x86. Failed
cmpxchg()does indeed not imply any memory ordering for all architectures -- and while the memory clobber (aka.barrier()) is equivalent to ansmp_wmb()on x86, that most certainly doesn't hold for non x86 code.Right, but the patch was addied by an Intel person, CCed:
152cef40a808("ACPI, APEI, GHES, Error records content based throttle")aSo I don't think he was thinking about ARM when doing that.
Quoted 9 lines
And that commit message doesn't say one whit why that memory barrier is needed there.
Reading that comment, it sounds like he wanted a hw memory barrier - MFENCE - but I don't see how normal data dependency wouldn't enforce the proper order already...
So that barrier looks out of place there.
The cache struct pointer should not be published until after the struct itself is fully populated. So on the producer side, some kind of hardware barrier is definitely needed, or the struct may appear half-baked to other cores that can read the updated pointer.
But as Peter points out, cmpxchg() itself has the required barrier semantics already, so the separate smp_wmb() is likely unnecessary. And as I suggested earlier, a full barrier is not necessary so we could relax this to cmpxchg_release() if desired.
OTOH the code seems to be working fine as is, so why modify it at all? (apart from the purely cosmetic changes)
Quoted 8 lines
Btw, this is the next perfect example why I'm asking people to write proper commit messages so that when we do git archeology later, we can figure out why something was done the way it has been.
And in this case, we can't. ;-\
Because writing proper commit messages is for losers. Yeah, right.</sarcasm>
Yeah, agree there.
Borislav Petkov
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 4 lines · Ard Biesheuvel
The cache struct pointer should not be published until after the struct itself is fully populated. So on the producer side, some kind of hardware barrier is definitely needed, or the struct may appear half-baked to other cores that can read the updated pointer.
Ah, right you are, ghes_estatus_cached() is called in all kinds of contexts and by other cores, sure.
Quoted 2 lines
OTOH the code seems to be working fine as is, so why modify it at all? (apart from the purely cosmetic changes)
peterz questioned that smp_wmb() there and then we started digging. And frankly, even if removing that barrier won't make any difference, I'd still prefer it gone and have the code simpler and cleaner.
Thx.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteRE: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
In reply to Ard BiesheuvelHi Ard
Quoted 60 lines
-----Original Message-----
From: Ard Biesheuvel <[email protected]>
Sent: Thursday, October 13, 2022 11:41 PM
To: Borislav Petkov <[email protected]>
Cc: Justin He <[email protected]>; Len Brown <[email protected]>; James Morse <[email protected]>; Tony Luck <[email protected]>; Mauro Carvalho Chehab <[email protected]>; Robert Richter <[email protected]>; Robert Moore <[email protected]>; Qiuxu Zhuo <[email protected]>; Yazen Ghannam <[email protected]>; Jan Luebbe <[email protected]>; Khuong Dinh <[email protected]>; Kani Toshi <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Rafael J . Wysocki <[email protected]>; Shuai Xue <[email protected]>; Jarkko Sakkinen <[email protected]>; [email protected]; nd <[email protected]>; kernel test robot <[email protected]>
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchgQuoted 11 lines · Borislav Petkov
Quoted 3 lines · Justin He
I have a concern about what if cmpxchg failed? Do we have to still guarantee the ordering since cmpxchg will not imply a
smp_mbif it failed.Of course it will imply that. At least on x86 it does.
smp_wmb()is a compiler barrier there andcmpxchg()already has that barrier semantics by clobbering "memory". I'm pretty sure you should have the same thing on ARM.No it definitely does not imply that. A memory clobber is a codegen construct, and the hardware could still complete the writes in a way that could result in another observer seeing a mix of old and new values that is inconsistent with the ordering of the stores as issued by the compiler.
Quoted 4 lines
says, "
new_cachemust be put into array after its contents are written".Are we writing anything into the cache if cmpxchg fails?
The cache fields get updated but the pointer to the struct is never shared globally if the
cmpxchg()fails so not having the barrier on failure should not be an issue here.Quoted 4 lines
Quoted 1 line
Besides, I didn't find the paired
smp_mborsmp_rmbfor thissmp_wmb.Why would there be pairs? I don't understand that statement here.
Typically, the other observer pairs the write barrier with a read barrier.
In this case, the other observer appears to be
ghes_estatus_cached(), and the reads of the cache struct fields must be ordered after the read of the cache struct's address. However, there is an implicit ordering there through an address dependency (you cannot dereference a struct without knowing its address) so the ordering is implied (andrcu_dereference()has aREAD_ONCE()inside so we are guaranteed to always dereference the same struct, even if the array slot gets updated concurrently.)
Quoted 4 lines
If you want to get rid of the barrier, you could drop it and change the
cmpxchg()tocmpxchg_release().Justin: so why are the RCU_INITIALIZER()s needed here?
In my this patch, I add the "__rcu" to the definition of ghes_estatus_caches. Hence Sparse will still have the warning on X86 with this RCU_INITIALIZER cast. drivers/acpi/apei/ghes.c:843:27: sparse: warning: incorrect type in initializer (different address spaces)drivers/acpi/apei/ghes.c:843:27: sparse: expected struct ghes_estatus_cache [noderef] __rcu *__olddrivers/acpi/apei/ghes.c:843:27: sparse: got struct ghes_estatus_cache *[assigned] slot_cachedrivers/acpi/apei/ghes.c:843:27: sparse: warning: incorrect type in initializer (different address spaces)drivers/acpi/apei/ghes.c:843:27: sparse: expected struct ghes_estatus_cache [noderef] __rcu *__newdrivers/acpi/apei/ghes.c:843:27: sparse: got struct ghes_estatus_cache *[assigned] new_cache
On Arm, IMO the macro cmpxchg doesn't care about it, that is, sparse will not report warnings with or without RCU_INITIALIZER cast.
I tend to remain this cast, what do you think of it.
--
Cheers,
Justin (Jia He)
Ard Biesheuvel
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 77 lines · Justin He
Hi Ard
Quoted 60 lines
-----Original Message-----
From: Ard Biesheuvel <[email protected]>
Sent: Thursday, October 13, 2022 11:41 PM
To: Borislav Petkov <[email protected]>
Cc: Justin He <[email protected]>; Len Brown <[email protected]>; James Morse <[email protected]>; Tony Luck <[email protected]>; Mauro Carvalho Chehab <[email protected]>; Robert Richter <[email protected]>; Robert Moore <[email protected]>; Qiuxu Zhuo <[email protected]>; Yazen Ghannam <[email protected]>; Jan Luebbe <[email protected]>; Khuong Dinh <[email protected]>; Kani Toshi <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Rafael J . Wysocki <[email protected]>; Shuai Xue <[email protected]>; Jarkko Sakkinen <[email protected]>; [email protected]; nd <[email protected]>; kernel test robot <[email protected]>
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchgQuoted 11 lines · Borislav Petkov
Quoted 3 lines · Justin He
I have a concern about what if cmpxchg failed? Do we have to still guarantee the ordering since cmpxchg will not imply a
smp_mbif it failed.Of course it will imply that. At least on x86 it does.
smp_wmb()is a compiler barrier there andcmpxchg()already has that barrier semantics by clobbering "memory". I'm pretty sure you should have the same thing on ARM.No it definitely does not imply that. A memory clobber is a codegen construct, and the hardware could still complete the writes in a way that could result in another observer seeing a mix of old and new values that is inconsistent with the ordering of the stores as issued by the compiler.
Quoted 4 lines
says, "
new_cachemust be put into array after its contents are written".Are we writing anything into the cache if cmpxchg fails?
The cache fields get updated but the pointer to the struct is never shared globally if the
cmpxchg()fails so not having the barrier on failure should not be an issue here.Quoted 4 lines
Quoted 1 line
Besides, I didn't find the paired
smp_mborsmp_rmbfor thissmp_wmb.Why would there be pairs? I don't understand that statement here.
Typically, the other observer pairs the write barrier with a read barrier.
In this case, the other observer appears to be
ghes_estatus_cached(), and the reads of the cache struct fields must be ordered after the read of the cache struct's address. However, there is an implicit ordering there through an address dependency (you cannot dereference a struct without knowing its address) so the ordering is implied (andrcu_dereference()has aREAD_ONCE()inside so we are guaranteed to always dereference the same struct, even if the array slot gets updated concurrently.)Quoted 4 lines
If you want to get rid of the barrier, you could drop it and change the
cmpxchg()tocmpxchg_release().Justin: so why are the RCU_INITIALIZER()s needed here?
In my this patch, I add the "
__rcu" to the definition ofghes_estatus_caches. Hence Sparse will still have the warning on X86 with thisRCU_INITIALIZERcast.drivers/acpi/apei/ghes.c:843:27: sparse: warning: incorrect type in initializer (different address spaces)drivers/acpi/apei/ghes.c:843:27: sparse: expected structghes_estatus_cache[noderef]__rcu*__olddrivers/acpi/apei/ghes.c:843:27: sparse: got structghes_estatus_cache*[assigned]slot_cachedrivers/acpi/apei/ghes.c:843:27: sparse: warning: incorrect type in initializer (different address spaces)drivers/acpi/apei/ghes.c:843:27: sparse: expected structghes_estatus_cache[noderef]__rcu*__newdrivers/acpi/apei/ghes.c:843:27: sparse: got structghes_estatus_cache*[assigned]new_cache
Quoted 5 lines
On Arm, IMO the macro cmpxchg doesn't care about it, that is, sparse will not report warnings with or without
RCU_INITIALIZERcast.I tend to remain this cast, what do you think of it.
OK, fair enough, I had only tested the arm64 build myself.
But just putting unrcu_pointer() and RCU_INITIALIZER() arbitrarily to shut up sparse is a bit sloppy, imho. Passing around pointers like this code does makes that necessary, unfortunately, so it would be nice if we could clean that up, by getting rid of the slot_cache variable.
And now that I have spent some time looking at this code, I wonder what the point of the cmpxchg() is in the first place. ghes_estatus_cache_add() selects a slot, and either succeeds in replacing its contents with a pointer to a new cached item, or it just gives up and frees the new item again, without attempting to select another slot even if one might be available.
Since we only insert new items, the race can only cause a failure if the selected slot was updated with another new item concurrently, which means that it is arbitrary which of those two items gets dropped. This means we don't need the cmpxchg() and the special case, and we can just drop the existing item unconditionally. Note that this does not result in loss of error events, it simply means we might cause a false cache miss, and report the same event one additional time in quick succession even if the cache should have prevented that.
------------------------8<------------------------------
drivers/acpi/apei/ghes.c +0 -1
@@ -138,7 +138,7 @@ struct ghes_vendor_record_entry {static struct gen_pool *ghes_estatus_pool;static unsigned long ghes_estatus_pool_size_request; -static struct ghes_estatus_cache
*ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE];
+static struct ghes_estatus_cache __rcu
*ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE];
static atomic_t ghes_estatus_cache_alloced;
static int ghes_panic_timeout __read_mostly = 30;
@@ -773,31 +773,26 @@ static struct ghes_estatus_cache *ghes_estatus_cache_alloc(
return cache;
}
-static void ghes_estatus_cache_free(struct ghes_estatus_cache *cache) +static void ghes_estatus_cache_rcu_free(struct rcu_head *head)
{
+ struct ghes_estatus_cache *cache;
u32 len;+ cache = container_of(head, struct ghes_estatus_cache, rcu);
len = cper_estatus_len(GHES_ESTATUS_FROM_CACHE(cache));
len = GHES_ESTATUS_CACHE_LEN(len);
gen_pool_free(ghes_estatus_pool, (unsigned long)cache, len);
atomic_dec(&ghes_estatus_cache_alloced);}
-static void ghes_estatus_cache_rcu_free(struct rcu_head *head) -{
- struct ghes_estatus_cache *cache;
-
- cache = container_of(head, struct ghes_estatus_cache, rcu);
- ghes_estatus_cache_free(cache);
-}
-
static void ghes_estatus_cache_add(
struct acpi_hest_generic *generic,
struct acpi_hest_generic_status *estatus)
{
int i, slot = -1, count;
unsigned long long now, duration, period, max_period = 0;
- struct ghes_estatus_cache *cache, *slot_cache = NULL, *new_cache;
+ struct ghes_estatus_cache *cache, *new_cache;
+ struct ghes_estatus_cache __rcu *victim; new_cache = ghes_estatus_cache_alloc(generic, estatus);
if (new_cache == NULL)
@@ -808,13 +803,11 @@ static void ghes_estatus_cache_add(
cache = rcu_dereference(ghes_estatus_caches[i]);
if (cache == NULL) {
slot = i;
- slot_cache = NULL;
break;
}
duration = now - cache->time_in;
if (duration >= GHES_ESTATUS_IN_CACHE_MAX_NSEC) {
slot = i;
- slot_cache = cache;
break;
}
count = atomic_read(&cache->count);
@@ -823,17 +816,28 @@ static void ghes_estatus_cache_add(
if (period > max_period) {
max_period = period;
slot = i;
- slot_cache = cache;
}
}
- /* new_cache must be put into array after its contents are written */
- smp_wmb();
- if (slot != -1 && cmpxchg(ghes_estatus_caches + slot,- slot_cache, new_cache) == slot_cache) {
- if (slot_cache)
- call_rcu(&slot_cache->rcu, ghes_estatus_cache_rcu_free);
- } else
- ghes_estatus_cache_free(new_cache);
+ if (slot != -1) {
+ /*
+ * Use release semantics to ensure that ghes_estatus_cached() + * running on another CPU will see the updated cache fields if + * it can see the new value of the pointer. + */
+ victim = xchg_release(ghes_estatus_caches + slot,
+ RCU_INITIALIZER(new_cache));
+
+ /*
+ * At this point, victim may point to a cached item different + * from the one based on which we selected the slot. Instead of + * going to the loop again to pick another slot, let's just + * drop the other item anyway: this may cause a false cache + * miss later on, but that won't cause any problems. + */
+ if (victim)
+ call_rcu(&rcu_dereference(victim)->rcu,
+ ghes_estatus_cache_rcu_free);
+ }
rcu_read_unlock();
}
Re: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 9 lines · Ard Biesheuvel
(quoted diff) +9
-0+ if (slot != -1) { + /* + * Use release semantics to ensure that ghes_estatus_cached() + * running on another CPU will see the updated cache fields if + * it can see the new value of the pointer. + */ + victim = xchg_release(ghes_estatus_caches + slot, + RCU_INITIALIZER(new_cache)); +
Review
Quoted 10 lines
(quoted diff) +10
-0+ /* + * At this point, victim may point to a cached item different + * from the one based on which we selected the slot. Instead of + * going to the loop again to pick another slot, let's just + * drop the other item anyway: this may cause a false cache + * miss later on, but that won't cause any problems. + */ + if (victim) { + call_rcu(&rcu_dereference(victim)->rcu, + ghes_estatus_cache_rcu_free);
}
I think you can use unrcu_pointer() here instead, there should not be a data dependency since the ->rcu member itself should be otherwise unused (and if it were, we wouldn't care about its previous content anyway).
But only Alpha cares about that distinction anyway, so *shrug*.
While I much like the xchg() variant; I still don't really fancy the verbage the sparse nonsense makes us do.
victim = xchg_release(&ghes_estatus_caches[slot], new_cache); if (victim) call_rcu(&victim->rcu, ghes_estatus_cache_rcu_free);
is much nicer code.
Over all; I'd simply ignore sparse (I often do).
Quoted 3 lines
+ }
rcu_read_unlock();}
Ard Biesheuvel
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Quoted 21 lines · Peter Zijlstra
Quoted 9 lines · Ard Biesheuvel
(quoted diff) +9
-0+ if (slot != -1) { + /* + * Use release semantics to ensure that ghes_estatus_cached() + * running on another CPU will see the updated cache fields if + * it can see the new value of the pointer. + */ + victim = xchg_release(ghes_estatus_caches + slot, + RCU_INITIALIZER(new_cache)); +Review
Quoted 10 lines
(quoted diff) +10
-0+ /* + * At this point, victim may point to a cached item different + * from the one based on which we selected the slot. Instead of + * going to the loop again to pick another slot, let's just + * drop the other item anyway: this may cause a false cache + * miss later on, but that won't cause any problems. + */ + if (victim) { + call_rcu(&rcu_dereference(victim)->rcu, + ghes_estatus_cache_rcu_free);
Quoted 8 lines
}I think you can use
unrcu_pointer()here instead, there should not be a data dependency since the ->rcu member itself should be otherwise unused (and if it were, we wouldn't care about its previous content anyway).But only Alpha cares about that distinction anyway, so *shrug*.
Ah yeah good point - and we are not actually dereferencing the pointer at all here, just adding an offset to get at the address of the rcu member.
So we can take this block out of the rcu_read_lock() section as well.
Quoted 11 lines
While I much like the
xchg()variant; I still don't really fancy the verbage the sparse nonsense makes us do.victim = xchg_release(&ghes_estatus_caches[slot], new_cache); if (victim) call_rcu(&victim->rcu, ghes_estatus_cache_rcu_free);is much nicer code.
Over all; I'd simply ignore sparse (I often do).
No disagreement there.
RE: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Hi Ard
Quoted 56 lines
-----Original Message-----
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchgQuoted 5 lines · Peter Zijlstra
Review
Quoted 3 lines · Ard Biesheuvel
(quoted diff) +3
-0+ if (slot != -1) { + /* + * Use release semantics to ensure that
ghes_estatus_cached()Quoted 1 line
Quoted 1 line
+ * running on another CPU will see the updated cache
fields if
Quoted 4 lines
Review
Quoted 4 lines
(quoted diff) +4
-0+ * it can see the new value of the pointer. + */ + victim = xchg_release(ghes_estatus_caches + slot, +
RCU_INITIALIZER(new_cache));Quoted 3 lines
Review
Quoted 3 lines
(quoted diff) +3
-0+ + /* + * At this point, victim may point to a cached itemdifferent
Quoted 1 line
Quoted 1 line
+ * from the one based on which we selected the slot.
Instead of
Quoted 1 line
Quoted 1 line
+ * going to the loop again to pick another slot, let's
just
Quoted 1 line
Quoted 1 line
+ * drop the other item anyway: this may cause a false
cache
Quoted 5 lines
Review
Quoted 5 lines
(quoted diff) +5
-0+ * miss later on, but that won't cause any problems. + */ + if (victim) { + call_rcu(&rcu_dereference(victim)->rcu, + ghes_estatus_cache_rcu_free);Quoted 8 lines
}I think you can use
unrcu_pointer()here instead, there should not be a data dependency since the ->rcu member itself should be otherwise unused (and if it were, we wouldn't care about its previous content anyway).But only Alpha cares about that distinction anyway, so *shrug*.
Ah yeah good point - and we are not actually dereferencing the pointer at all here, just adding an offset to get at the address of the rcu member.
So we can take this block out of the
rcu_read_lock()section as well.Quoted 4 lines
While I much like the
xchg()variant; I still don't really fancy the verbage the sparse nonsense makes us do.victim = xchg_release(&ghes_estatus_caches[slot],
new_cache);Quoted 4 lines
if (victim) call_rcu(&victim->rcu,
ghes_estatus_cache_rcu_free);
Quoted 6 lines
Quoted 4 lines
is much nicer code.
Over all; I'd simply ignore sparse (I often do).
No disagreement there.
What do you think of the updated patch:
apei/ghes: Use xchg() for updating new cache slot instead of
cmpxchg()
From: Ard Biesheuvel <[email protected]>
ghes_estatus_cache_add() selects a slot, and either succeeds in replacing its contents with a pointer to a new cached item, or it just gives up and frees the new item again, without attempting to select another slot even if one might be available.
Since only inserting new items is needed, the race can only cause a failure if the selected slot was updated with another new item concurrently, which means that it is arbitrary which of those two items gets dropped. This means the cmpxchg() and the special case are not necessary, and hence just drop the existing item unconditionally. Note that this does not result in loss of error events, it simply means we might cause a false cache miss, and report the same event one additional time in quick succession even if the cache should have prevented that.
---
[Justin]: I removed __rcu annotation of victim, removed the RCU_INITIALIZER cast and added the unptr for xchg.
drivers/acpi/apei/ghes.c | 44 ++++++++++++++++++++--------------------
1 file changed, 22 insertions(+), 22 deletions(-)
drivers/acpi/apei/ghes.c +22 -22
@@ -150,7 +150,7 @@ struct ghes_vendor_record_entry {static struct gen_pool *ghes_estatus_pool;static unsigned long ghes_estatus_pool_size_request; -static struct ghes_estatus_cache *ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ]; +static struct ghes_estatus_cache __rcu *ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ];static atomic_t ghes_estatus_cache_alloced;static int ghes_panic_timeout__read_mostly =30 ;
@@ -785,31 +785,26 @@ static struct ghes_estatus_cache *ghes_estatus_cache_alloc(return cache; } -static void ghes_estatus_cache_free (struct ghes_estatus_cache *cache) +static void ghes_estatus_cache_rcu_free (struct rcu_head *head) { +struct ghes_estatus_cache *cache;u32 len; + cache = container_of(head, struct ghes_estatus_cache, rcu); len =cper_estatus_len (GHES_ESTATUS_FROM_CACHE (cache)); len =GHES_ESTATUS_CACHE_LEN (len);gen_pool_free (ghes_estatus_pool, (unsigned long )cache, len);atomic_dec (&ghes_estatus_cache_alloced); } -static void ghes_estatus_cache_rcu_free (struct rcu_head *head) -{ -struct ghes_estatus_cache *cache; - - cache = container_of(head, struct ghes_estatus_cache, rcu); -ghes_estatus_cache_free (cache); -} -static void ghes_estatus_cache_add (struct acpi_hest_generic *generic,struct acpi_hest_generic_status *estatus) {int i, slot =-1 , count;unsigned long long now, duration, period, max_period =0 ; -struct ghes_estatus_cache *cache, *slot_cache =NULL , *new_cache; +struct ghes_estatus_cache *cache, *new_cache; +struct ghes_estatus_cache *victim; new_cache =ghes_estatus_cache_alloc (generic, estatus);if (new_cache ==NULL )
@@ -820,13 +815,11 @@ static void ghes_estatus_cache_add( cache =rcu_dereference (ghes_estatus_caches[i]);if (cache ==NULL ) { slot = i; - slot_cache =NULL ;break ; } duration = now - cache->time_in;if (duration >=GHES_ESTATUS_IN_CACHE_MAX_NSEC ) { slot = i; - slot_cache = cache;break ; } count =atomic_read (&cache->count);
@@ -835,17 +828,24 @@ static void ghes_estatus_cache_add(if (period > max_period) { max_period = period; slot = i; - slot_cache = cache; } } -/* new_cache must be put into array after its contents are written */ -smp_wmb (); -if (slot !=-1 &&cmpxchg (ghes_estatus_caches + slot, - slot_cache, new_cache) == slot_cache) { -if (slot_cache) -call_rcu (&slot_cache->rcu, ghes_estatus_cache_rcu_free); - }else -ghes_estatus_cache_free (new_cache); +if (slot !=-1 ) { +/* +* Use release semantics to ensure that ghes_estatus_cached() +* running on another CPU will see the updated cache fields if +* it can see the new value of the pointer. +* At this point, victim may point to a cached item different +* from the one based on which we selected the slot. Instead of +* going to the loop again to pick another slot, let's just +* drop the other item anyway: this may cause a false cache +* miss later on, but that won't cause any problems. +*/ + victim =unrcu_pointer (xchg_release (&ghes_estatus_caches[slot], + new_cache)); +if (victim) +call_rcu (&victim->rcu, ghes_estatus_cache_rcu_free); + }rcu_read_unlock (); }
IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.
Ard Biesheuvel
maintainerRe: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Hi Justin,
Quoted 78 lines · Justin He
Hi Ard
Quoted 56 lines
-----Original Message-----
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchgQuoted 5 lines · Peter Zijlstra
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add(
Quoted 3 lines · Ard Biesheuvel
drivers/acpi/apei/ghes.c +3
-0+if (slot !=-1 ) { +/* +* Use release semantics to ensure that
ghes_estatus_cached()Quoted 1 line
Quoted 1 line
+ * running on another CPU will see the updated cache
fields if
Quoted 4 lines
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add(
Quoted 4 lines
drivers/acpi/apei/ghes.c +4
-0+* it can see the new value of the pointer. +*/ + victim = xchg_release(ghes_estatus_caches + slot, +
RCU_INITIALIZER(new_cache));Quoted 3 lines
Review
Quoted 3 lines
(quoted diff) +3
-0+ + /* + * At this point, victim may point to a cached itemdifferent
Quoted 1 line
Quoted 1 line
+ * from the one based on which we selected the slot.
Instead of
Quoted 1 line
Quoted 1 line
+ * going to the loop again to pick another slot, let's
just
Quoted 1 line
Quoted 1 line
+ * drop the other item anyway: this may cause a false
cache
Quoted 5 lines
Review
Quoted 5 lines
(quoted diff) +5
-0+ * miss later on, but that won't cause any problems. + */ + if (victim) { + call_rcu(&rcu_dereference(victim)->rcu, + ghes_estatus_cache_rcu_free);Quoted 8 lines
}I think you can use
unrcu_pointer()here instead, there should not be a data dependency since the ->rcu member itself should be otherwise unused (and if it were, we wouldn't care about its previous content anyway).But only Alpha cares about that distinction anyway, so *shrug*.
Ah yeah good point - and we are not actually dereferencing the pointer at all here, just adding an offset to get at the address of the rcu member.
So we can take this block out of the
rcu_read_lock()section as well.Quoted 4 lines
While I much like the
xchg()variant; I still don't really fancy the verbage the sparse nonsense makes us do.victim = xchg_release(&ghes_estatus_caches[slot],
new_cache);Quoted 4 lines
if (victim) call_rcu(&victim->rcu,
ghes_estatus_cache_rcu_free);Quoted 6 lines
Quoted 4 lines
is much nicer code.
Over all; I'd simply ignore sparse (I often do).
No disagreement there.
What do you think of the updated patch:
apei/ghes: Use
xchg()for updating new cache slot instead of
cmpxchg()From: Ard Biesheuvel <[email protected]>
ghes_estatus_cache_add()selects a slot, and either succeeds in replacing its contents with a pointer to a new cached item, or it just gives up and frees the new item again, without attempting to select another slot even if one might be available.
Quoted 9 lines
Since only inserting new items is needed, the race can only cause a failure if the selected slot was updated with another new item concurrently, which means that it is arbitrary which of those two items gets dropped. This means the
cmpxchg()and the special case are not necessary, and hence just drop the existing item unconditionally. Note that this does not result in loss of error events, it simply means we might cause a false cache miss, and report the same event one additional time in quick succession even if the cache should have prevented that.
Please add a line here
Quoted 83 lines
Signed-off-by Jia He <[email protected]>Signed-off-by Ard Biesheuvel <[email protected]>---
[Justin]: I removed__rcuannotation of victim, removed theRCU_INITIALIZERcast and added the unptr for xchg.
drivers/acpi/apei/ghes.c| 44 ++++++++++++++++++++--------------------1 file changed, 22 insertions(+), 22 deletions(-)drivers/acpi/apei/ghes.c +6
-17@@ -150,7 +150,7 @@ struct ghes_vendor_record_entry {static struct gen_pool *ghes_estatus_pool;static unsigned long ghes_estatus_pool_size_request; -static struct ghes_estatus_cache *ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ]; +static struct ghes_estatus_cache __rcu *ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ];static atomic_t ghes_estatus_cache_alloced;static int ghes_panic_timeout__read_mostly =30 ;@@ -785,31 +785,26 @@ static struct ghes_estatus_cache *ghes_estatus_cache_alloc(return cache; } -static void ghes_estatus_cache_free (struct ghes_estatus_cache *cache) +static void ghes_estatus_cache_rcu_free (struct rcu_head *head) { +struct ghes_estatus_cache *cache;u32 len; + cache = container_of(head, struct ghes_estatus_cache, rcu); len =cper_estatus_len (GHES_ESTATUS_FROM_CACHE (cache)); len =GHES_ESTATUS_CACHE_LEN (len);gen_pool_free (ghes_estatus_pool, (unsigned long )cache, len);atomic_dec (&ghes_estatus_cache_alloced); } -static void ghes_estatus_cache_rcu_free (struct rcu_head *head) -{ -struct ghes_estatus_cache *cache; - - cache = container_of(head, struct ghes_estatus_cache, rcu); -ghes_estatus_cache_free (cache); -} -static void ghes_estatus_cache_add (struct acpi_hest_generic *generic,struct acpi_hest_generic_status *estatus) {int i, slot =-1 , count;unsigned long long now, duration, period, max_period =0 ; -struct ghes_estatus_cache *cache, *slot_cache =NULL , *new_cache; +struct ghes_estatus_cache *cache, *new_cache; +struct ghes_estatus_cache *victim; new_cache =ghes_estatus_cache_alloc (generic, estatus);if (new_cache ==NULL )@@ -820,13 +815,11 @@ static void ghes_estatus_cache_add( cache =rcu_dereference (ghes_estatus_caches[i]);if (cache ==NULL ) { slot = i; - slot_cache =NULL ;break ; } duration = now - cache->time_in;if (duration >=GHES_ESTATUS_IN_CACHE_MAX_NSEC ) { slot = i; - slot_cache = cache;break ; } count =atomic_read (&cache->count);@@ -835,17 +828,24 @@ static void ghes_estatus_cache_add(if (period > max_period) { max_period = period; slot = i; - slot_cache = cache; } } -/* new_cache must be put into array after its contents are written */ -smp_wmb (); - if (slot != -1 && cmpxchg(ghes_estatus_caches + slot,
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add( · line 837
Quoted 10 lines
drivers/acpi/apei/ghes.c +5
-5- slot_cache, new_cache) == slot_cache) { -if (slot_cache) -call_rcu (&slot_cache->rcu, ghes_estatus_cache_rcu_free); - } else -ghes_estatus_cache_free (new_cache ); +if (slot !=-1 ) { +/* +* Use release semantics to ensure that ghes_estatus_cached() +* running on another CPU will see the updated cache fields if +* it can see the new value of the pointer.
Please move the comment back where it was. 'At this point' is now ambiguous because victim has not been assigned yet.
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add( · line 845
Quoted 8 lines
drivers/acpi/apei/ghes.c +8
-0+* At this point, victim may point to a cached item different +* from the one based on which we selected the slot. Instead of +* going to the loop again to pick another slot, let's just +* drop the other item anyway: this may cause a false cache +* miss later on, but that won't cause any problems. +*/ + victim =unrcu_pointer (xchg_release (&ghes_estatus_caches[slot], + new_cache));
Doesn't this still trigger the sparse warning on x86?
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add( · line 847
Quoted 2 lines
drivers/acpi/apei/ghes.c +2
-0+if (victim) +call_rcu (&victim->rcu, ghes_estatus_cache_rcu_free);
I think it is better to add back the __rcu annotation to 'victim', and change this line to
call_rcu(&unrcu_pointer(victim)->rcu, ghes_estatus_cache_rcu_free);
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add( · line 849
Quoted 2 lines
drivers/acpi/apei/ghes.c +1
-0+ }rcu_read_unlock ();
This can now be moved before the if()
Quoted 3 lines
drivers/acpi/apei/ghes.c +0
-0}IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.
Please get rid of this footer.
RE: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
Hi Ard
Quoted 231 lines
-----Original Message-----
From: Ard Biesheuvel <[email protected]>
Sent: Monday, October 17, 2022 5:27 PM
To: Justin He <[email protected]>
Cc: Borislav Petkov <[email protected]>; Len Brown <[email protected]>; James Morse <[email protected]>; Tony Luck <[email protected]>; Mauro Carvalho Chehab <[email protected]>; Peter Zijlstra <[email protected]>; Robert Richter <[email protected]>; Robert Moore <[email protected]>; Qiuxu Zhuo <[email protected]>; Yazen Ghannam <[email protected]>; Jan Luebbe <[email protected]>; Khuong Dinh <[email protected]>; Kani Toshi <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Rafael J . Wysocki <[email protected]>; Shuai Xue <[email protected]>; Jarkko Sakkinen <[email protected]>; [email protected]; kernel test robot <[email protected]>
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchgHi Justin,
Quoted 45 lines · Justin He
Hi Ard
Quoted 42 lines
-----Original Message-----
Subject: Re: [PATCH v8 6/7] apei/ghes: Useunrcu_pointerfor cmpxchgQuoted 5 lines · Peter Zijlstra
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add(
Quoted 3 lines · Ard Biesheuvel
drivers/acpi/apei/ghes.c +3
-0+if (slot !=-1 ) { +/* +* Use release semantics to ensure that
ghes_estatus_cached()Quoted 2 lines
Quoted 2 lines
+ * running on another CPU will see the updated
+ cachefields if
Quoted 5 lines
Review
Quoted 5 lines
(quoted diff) +5
-0+ * it can see the new value of the pointer. + */ + victim = xchg_release(ghes_estatus_caches + + slot, +
RCU_INITIALIZER(new_cache));Quoted 4 lines
Review
Quoted 4 lines
(quoted diff) +4
-0+ + /* + * At this point, victim may point to a cached + itemdifferent
Quoted 1 line
Quoted 1 line
+ * from the one based on which we selected the slot.
Instead of
Quoted 2 lines
Quoted 2 lines
+ * going to the loop again to pick another slot,
+ let'sjust
Quoted 2 lines
Quoted 2 lines
+ * drop the other item anyway: this may cause a
+ falsecache
Quoted 10 lines
Review
Quoted 5 lines
(quoted diff) +5
-0+ * miss later on, but that won't cause any problems. + */ + if (victim) { + call_rcu(&rcu_dereference(victim)->rcu, + ghes_estatus_cache_rcu_free);}I think you can use
unrcu_pointer()here instead, there should not be a data dependency since the ->rcu member itself should be otherwise unused (and if it were, we wouldn't care about its previouscontent anyway).
Quoted 6 lines
Quoted 6 lines
Quoted 3 lines
But only Alpha cares about that distinction anyway, so *shrug*.
Ah yeah good point - and we are not actually dereferencing the pointer at all here, just adding an offset to get at the address of the rcu
member.
Quoted 32 lines
Quoted 13 lines
So we can take this block out of the
rcu_read_lock()section as well.Quoted 4 lines
While I much like the
xchg()variant; I still don't really fancy the verbage the sparse nonsense makes us do.victim = xchg_release(&ghes_estatus_caches[slot],
new_cache);Quoted 4 lines
if (victim) call_rcu(&victim->rcu,
ghes_estatus_cache_rcu_free);Quoted 6 lines
Quoted 4 lines
is much nicer code.
Over all; I'd simply ignore sparse (I often do).
No disagreement there.
What do you think of the updated patch:
apei/ghes: Use
xchg()for updating new cache slot instead of
cmpxchg()From: Ard Biesheuvel <[email protected]>
ghes_estatus_cache_add()selects a slot, and either succeeds in replacing its contents with a pointer to a new cached item, or it just gives up and frees the new item again, without attempting to select another slot even if one might be available.Quoted 8 lines
Since only inserting new items is needed, the race can only cause a failure if the selected slot was updated with another new item concurrently, which means that it is arbitrary which of those two items gets dropped. This means the
cmpxchg()and the special case are not necessary, and hence just drop the existing item unconditionally. Note that this does not result in loss of error events, it simply means we might cause a false cache miss, and report the same event one additional time in quick succession even if the cache should have preventedthat.
Quoted 1 line
Please add a line here
Co-developed-by Jia He <[email protected]>Quoted 44 lines
Signed-off-by Jia He <[email protected]>Signed-off-by Ard Biesheuvel <[email protected]>---
[Justin]: I removed__rcuannotation of victim, removed theRCU_INITIALIZERcast and added the unptr for xchg.
drivers/acpi/apei/ghes.c| 44 ++++++++++++++++++++--------------------1 file changed, 22 insertions(+), 22 deletions(-)drivers/acpi/apei/ghes.c index renamed from drivers/acpi/apei/ghes.c +0
-0
27c72b175e4b..5fc8a135450b100644drivers/acpi/apei/ghes.c +0
-0@@ -150,7 +150,7 @@ struct ghes_vendor_record_entry { static struct
gen_pool*ghes_estatus_pool; static unsigned longghes_estatus_pool_size_request;-static struct
ghes_estatus_cache
*ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE];drivers/acpi/apei/ghes.c +2
-0+static struct ghes_estatus_cache __rcu +*ghes_estatus_caches[GHES_ESTATUS_CACHES_SIZE ];static atomic_t ghes_estatus_cache_alloced; static int ghes_panic_timeout __read_mostly = 30; @@ -785,31 +785,26@@ static struct
ghes_estatus_cache*ghes_estatus_cache_alloc(return cache;}
drivers/acpi/apei/ghes.c +3
-4-static void ghes_estatus_cache_free (struct ghes_estatus_cache *cache) +static void ghes_estatus_cache_rcu_free (struct rcu_head *head) { +struct ghes_estatus_cache *cache;u32 len; + cache = container_of(head, struct ghes_estatus_cache, rcu); len =cper_estatus_len (GHES_ESTATUS_FROM_CACHE (cache)); len =GHES_ESTATUS_CACHE_LEN (len);gen_pool_free (ghes_estatus_pool, (unsigned long )cache, len);atomic_dec (&ghes_estatus_cache_alloced); } -static void ghes_estatus_cache_rcu_free(struct rcu_head *head) -{ -struct ghes_estatus_cache *cache; -drivers/acpi/apei/ghes.c static struct ghes_estatus_cache *ghes_estatus_cache_alloc(
Quoted 10 lines
drivers/acpi/apei/ghes.c +0
-5- cache = container_of(head, struct ghes_estatus_cache, rcu); -ghes_estatus_cache_free (cache); -} -static void ghes_estatus_cache_add (struct acpi_hest_generic *generic,struct acpi_hest_generic_status *estatus) {int i, slot =-1 , count;unsigned long long now, duration, period, max_period =0 ; - struct ghes_estatus_cache *cache, *slot_cache = NULL,
*new_cache;Quoted 9 lines
drivers/acpi/apei/ghes.c +2
-0+struct ghes_estatus_cache *cache, *new_cache; +struct ghes_estatus_cache *victim; new_cache =ghes_estatus_cache_alloc (generic, estatus);if (new_cache ==NULL )@@ -820,13 +815,11 @@ static void ghes_estatus_cache_add( cache =rcu_dereference (ghes_estatus_caches[i]);if (cache ==NULL ) { slot = i;drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add( · line 825
Quoted 10 lines
drivers/acpi/apei/ghes.c +0
-2- slot_cache =NULL ;break ; } duration = now - cache->time_in;if (duration >=GHES_ESTATUS_IN_CACHE_MAX_NSEC ) { slot = i; - slot_cache = cache;break ; } count = atomic_read(&cache->count); @@ -835,17+828,24
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add(
Quoted 8 lines
@@ static void ghes_estatus_cache_add(
if (period > max_period) { max_period = period; slot = i;drivers/acpi/apei/ghes.c +0
-2- slot_cache = cache; } } -/* new_cache must be put into array after its contents are written */
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add(
Quoted 3 lines
drivers/acpi/apei/ghes.c +0
-3-smp_wmb (); -if (slot !=-1 &&cmpxchg (ghes_estatus_caches + slot, - slot_cache, new_cache) == slot_cache){
drivers/acpi/apei/ghes.c
Quoted 2 lines
drivers/acpi/apei/ghes.c +0
-2-if (slot_cache) - call_rcu(&slot_cache->rcu,
ghes_estatus_cache_rcu_free);drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add(
Quoted 5 lines
drivers/acpi/apei/ghes.c +3
-2- }else -ghes_estatus_cache_free (new_cache); +if (slot !=-1 ) { +/* +* Use release semantics to ensure that
ghes_estatus_cached()Quoted 1 line
+ * running on another CPU will see the updated cache
fields if
Quoted 1 line
+ * it can see the new value of the pointer.
Please move the comment back where it was. 'At this point' is now ambiguous because victim has not been assigned yet.
Quoted 1 line
+ * At this point, victim may point to a cached item
different
Quoted 1 line
+ * from the one based on which we selected the slot.
Instead of
drivers/acpi/apei/ghes.c
Quoted 2 lines
drivers/acpi/apei/ghes.c +2
-0+* going to the loop again to pick another slot, let's just +* drop the other item anyway: this may cause a false
Quoted 8 lines
cache
drivers/acpi/apei/ghes.c static void ghes_estatus_cache_add(
Quoted 3 lines
drivers/acpi/apei/ghes.c +3
-0+* miss later on, but that won't cause any problems. +*/ + victim =unrcu_pointer(xchg_release(&ghes_estatus_caches[slot],
Quoted 1 line
+
new_cache));Doesn't this still trigger the sparse warning on x86?
Yes,
I will add the RCU_INITIALIZER back if Peter doesn't object.
--
Cheers,
Justin (Jia He)
Re: [PATCH v8 6/7] apei/ghes: Use unrcu_pointer for cmpxchg
In reply to Justin HeQuoted 10 lines · Justin He
Quoted 4 lines
This is a combined diff - do a second patch which does only remove the
smp_wmb(). Thesmp_wmb()there is not needed as thecmpxchg()already implies asmp_mb()so there's no need for that separate, explicit one.I have a concern about what if cmpxchg failed? Do we have to still guarantee the ordering since cmpxchg will not imply a
smp_mbif it failed.Besides, I didn't find the paired
smp_mborsmp_rmbfor thissmp_wmb. Do you have any ideas?
failed cmpxchg does indeed not imply smp_mb; but in that case I can't find a store it orders against; and the comment is utterly shite since it doesn't spell out the ordering in any case.
The way I read that code is that the cmpxchg effectively publishes the data and all it wants to ensure is that if the pointer is publised the object is complete -- in which case the cpmxchg() is sufficient, on success the object is published and you get the ordering, on failure the object isn't published and nobody cares about the ordering anyway.
If there's anything else, the comment is in dire need of fixing anyway.
PATCH v8 7/7 EDAC/igen6: Return consistent errno when another edac driver is enabled
In reply to Jia HeOnly a single edac driver can be enabled for EDAC MC. The igen6_init() should be returned with EBUSY instead of ENODEV, which is consistent with other edac drivers.
--- drivers/edac/igen6_edac.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/edac/igen6_edac.c +1 -1
@@ -1276,7 +1276,7 @@ static int __init igen6_init(void) owner =edac_get_owner ();if (owner &&strncmp (owner,EDAC_MOD_STR ,sizeof (EDAC_MOD_STR ))) -return -ENODEV ; +return -EBUSY ; edac_op_state =EDAC_OPSTATE_NMI ;
Signature · 1 line
2.25.1