[PATCH v3 00/17] AMD MCA interrupts rework

[PATCH v3 00/17] AMD MCA interrupts rework

[Yazen Ghannam]

Hi all,

This set unifies the AMD MCA interrupt handlers with common MCA code.
The goal is to avoid duplicating functionality like reading and clearing
MCA banks.

Based on feedback, this revision also include changes to the MCA init
flow.

Patches 1-4:
General fixes and cleanups.

Patches 5-10:
Add BSP-only init flow and related changes.

Patches 11-15:
Updates from v1 set.

Patch 16:
Interrupt storm handling rebased on current set.

Patch 17:
Fix how poll settings are restored after a CMCI storm.

Thanks,
Yazen

---
Changes in v3:
- Rebased on tip/x86/merge rather than tip/master.
- Updated MSR access helpers (*msrl -> *msrq).
- Add patch to fix polling after a storm.
- Link to v2: https://lore.kernel.org/r/20250213-wip-mca-updates-v2-0-3636547fe05f@amd.com

Changes in v2:
- Add general cleanup pre-patches.
- Add changes for BSP-only init.
- Add interrupt storm handling for AMD.
- Link to v1: https://lore.kernel.org/r/20240523155641.2805411-1-yazen.ghannam@amd.com

---
Borislav Petkov (1):
      x86/mce: Cleanup bank processing on init

Smita Koralahalli (1):
      x86/mce: Handle AMD threshold interrupt storms

Yazen Ghannam (15):
      x86/mce: Don't remove sysfs if thresholding sysfs init fails
      x86/mce/amd: Remove return value for mce_threshold_{create,remove}_device()
      x86/mce/amd: Remove smca_banks_map
      x86/mce/amd: Put list_head in threshold_bank
      x86/mce: Remove __mcheck_cpu_init_early()
      x86/mce: Define BSP-only init
      x86/mce: Define BSP-only SMCA init
      x86/mce: Do 'UNKNOWN' vendor check early
      x86/mce: Separate global and per-CPU quirks
      x86/mce: Move machine_check_poll() status checks to helper functions
      x86/mce: Unify AMD THR handler with MCA Polling
      x86/mce: Unify AMD DFR handler with MCA Polling
      x86/mce/amd: Enable interrupt vectors once per-CPU on SMCA systems
      x86/mce/amd: Support SMCA Corrected Error Interrupt
      x86/mce: Restore poll settings after storm subsides

 arch/x86/include/asm/mce.h          |  11 +-
 arch/x86/kernel/cpu/common.c        |   1 +
 arch/x86/kernel/cpu/mce/amd.c       | 397 +++++++++++++-----------------------
 arch/x86/kernel/cpu/mce/core.c      | 331 +++++++++++++++---------------
 arch/x86/kernel/cpu/mce/intel.c     |  18 ++
 arch/x86/kernel/cpu/mce/internal.h  |  10 +
 arch/x86/kernel/cpu/mce/threshold.c |   3 +
 7 files changed, 347 insertions(+), 424 deletions(-)
---
base-commit: e6090e017e4b1e2a1e461750b7281a05f4e07a76
change-id: 20250210-wip-mca-updates-bed2a67c9c57

[PATCH v3 01/17] x86/mce: Don't remove sysfs if thresholding sysfs init fails

[Yazen Ghannam]

Currently, the MCE subsystem sysfs interface will be removed if the
thresholding sysfs interface fails to be created. A common failure is
due to new MCA bank types that are not recognized and don't have a short
name set.

The MCA thresholding feature is optional and should not break the common
MCE sysfs interface. Also, new MCA bank types are occasionally
introduced, and updates will be needed to recognize them. But likewise,
this should not break the common sysfs interface.

Keep the MCE sysfs interface regardless of the status of the
thresholding sysfs interface.

Cc: stable@vger.kernel.org
Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-1-3636547fe05f@amd.com
    
    v2->v3:
    * Added tags from Qiuxu and Tony.
    
    v1->v2:
    * New in v2.
    * Included stable tag but there's no specific commit for Fixes.

 arch/x86/kernel/cpu/mce/core.c | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index 255927f0284e..72c2aa0809c0 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -2801,15 +2801,9 @@ static int mce_cpu_dead(unsigned int cpu)
 static int mce_cpu_online(unsigned int cpu)
 {
 	struct timer_list *t = this_cpu_ptr(&mce_timer);
-	int ret;
 
 	mce_device_create(cpu);
-
-	ret = mce_threshold_create_device(cpu);
-	if (ret) {
-		mce_device_remove(cpu);
-		return ret;
-	}
+	mce_threshold_create_device(cpu);
 	mce_reenable_cpu();
 	mce_start_timer(t);
 	return 0;

-- 
2.49.0

[PATCH v3 02/17] x86/mce/amd: Remove return value for mce_threshold_{create,remove}_device()

[Yazen Ghannam]

The return values are not checked, so set return type to 'void'.

Also, move function declarations to internal.h, since these functions are
only used within the MCE subsystem.

Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-2-3636547fe05f@amd.com
    
    v2->v3:
    * Include mce_threshold_remove_device().
    
    v1->v2:
    * New in v2.

 arch/x86/include/asm/mce.h         |  6 ------
 arch/x86/kernel/cpu/mce/amd.c      | 22 ++++++++++------------
 arch/x86/kernel/cpu/mce/internal.h |  4 ++++
 3 files changed, 14 insertions(+), 18 deletions(-)

diff --git a/arch/x86/include/asm/mce.h b/arch/x86/include/asm/mce.h
index 6c77c03139f7..752802bf966b 100644
--- a/arch/x86/include/asm/mce.h
+++ b/arch/x86/include/asm/mce.h
@@ -371,15 +371,9 @@ enum smca_bank_types {
 
 extern bool amd_mce_is_memory_error(struct mce *m);
 
-extern int mce_threshold_create_device(unsigned int cpu);
-extern int mce_threshold_remove_device(unsigned int cpu);
-
 void mce_amd_feature_init(struct cpuinfo_x86 *c);
 enum smca_bank_types smca_get_bank_type(unsigned int cpu, unsigned int bank);
 #else
-
-static inline int mce_threshold_create_device(unsigned int cpu)		{ return 0; };
-static inline int mce_threshold_remove_device(unsigned int cpu)		{ return 0; };
 static inline bool amd_mce_is_memory_error(struct mce *m)		{ return false; };
 static inline void mce_amd_feature_init(struct cpuinfo_x86 *c)		{ }
 #endif
diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 9d852c3b2cb5..7ff479c679fb 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -1290,12 +1290,12 @@ static void __threshold_remove_device(struct threshold_bank **bp)
 	kfree(bp);
 }
 
-int mce_threshold_remove_device(unsigned int cpu)
+void mce_threshold_remove_device(unsigned int cpu)
 {
 	struct threshold_bank **bp = this_cpu_read(threshold_banks);
 
 	if (!bp)
-		return 0;
+		return;
 
 	/*
 	 * Clear the pointer before cleaning up, so that the interrupt won't
@@ -1304,7 +1304,7 @@ int mce_threshold_remove_device(unsigned int cpu)
 	this_cpu_write(threshold_banks, NULL);
 
 	__threshold_remove_device(bp);
-	return 0;
+	return;
 }
 
 /**
@@ -1318,36 +1318,34 @@ int mce_threshold_remove_device(unsigned int cpu)
  * thread running on @cpu.  The callback is invoked on all CPUs which are
  * online when the callback is installed or during a real hotplug event.
  */
-int mce_threshold_create_device(unsigned int cpu)
+void mce_threshold_create_device(unsigned int cpu)
 {
 	unsigned int numbanks, bank;
 	struct threshold_bank **bp;
-	int err;
 
 	if (!mce_flags.amd_threshold)
-		return 0;
+		return;
 
 	bp = this_cpu_read(threshold_banks);
 	if (bp)
-		return 0;
+		return;
 
 	numbanks = this_cpu_read(mce_num_banks);
 	bp = kcalloc(numbanks, sizeof(*bp), GFP_KERNEL);
 	if (!bp)
-		return -ENOMEM;
+		return;
 
 	for (bank = 0; bank < numbanks; ++bank) {
 		if (!(this_cpu_read(bank_map) & BIT_ULL(bank)))
 			continue;
-		err = threshold_create_bank(bp, cpu, bank);
-		if (err) {
+		if (threshold_create_bank(bp, cpu, bank)) {
 			__threshold_remove_device(bp);
-			return err;
+			return;
 		}
 	}
 	this_cpu_write(threshold_banks, bp);
 
 	if (thresholding_irq_en)
 		mce_threshold_vector = amd_threshold_interrupt;
-	return 0;
+	return;
 }
diff --git a/arch/x86/kernel/cpu/mce/internal.h b/arch/x86/kernel/cpu/mce/internal.h
index b5ba598e54cb..64ac25b95360 100644
--- a/arch/x86/kernel/cpu/mce/internal.h
+++ b/arch/x86/kernel/cpu/mce/internal.h
@@ -265,6 +265,8 @@ void mce_prep_record_common(struct mce *m);
 void mce_prep_record_per_cpu(unsigned int cpu, struct mce *m);
 
 #ifdef CONFIG_X86_MCE_AMD
+void mce_threshold_create_device(unsigned int cpu);
+void mce_threshold_remove_device(unsigned int cpu);
 extern bool amd_filter_mce(struct mce *m);
 bool amd_mce_usable_address(struct mce *m);
 
@@ -293,6 +295,8 @@ static __always_inline void smca_extract_err_addr(struct mce *m)
 }
 
 #else
+static inline void mce_threshold_create_device(unsigned int cpu)	{ }
+static inline void mce_threshold_remove_device(unsigned int cpu)	{ }
 static inline bool amd_filter_mce(struct mce *m) { return false; }
 static inline bool amd_mce_usable_address(struct mce *m) { return false; }
 static inline void smca_extract_err_addr(struct mce *m) { }

-- 
2.49.0

[PATCH v3 03/17] x86/mce/amd: Remove smca_banks_map

[Yazen Ghannam]

The MCx_MISC0[BlkPtr] field was used on legacy systems to hold a
register offset for the next MCx_MISC* register. In this way, an
implementation-specific number of registers can be discovered at
runtime.

The MCAX/SMCA register space simplifies this by always including
the MCx_MISC[1-4] registers. The MCx_MISC0[BlkPtr] field is used to
indicate (true/false) whether any MCx_MISC[1-4] registers are present.
But it indicates neither which ones nor how many. Therefore, all the
registers are accessed and their bits are checked.

AMD systems generally enforce a Read-as-Zero/Writes-Ignored policy for
unused registers. Therefore, there is no harm to read an unused
register. This is already done in practice for most of the MCx_MISC
registers.

Remove the smca_banks_map variable as it is effectively redundant.

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-3-3636547fe05f@amd.com
    
    v2->v3:
    * Minor edit in commit message.
    * Added tags from Qiuxu and Tony.
    
    v1->v2:
    * New in v2.

 arch/x86/kernel/cpu/mce/amd.c | 30 ------------------------------
 1 file changed, 30 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 7ff479c679fb..46ff41c1b50d 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -252,9 +252,6 @@ static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
  */
 static DEFINE_PER_CPU(u64, bank_map);
 
-/* Map of banks that have more than MCA_MISC0 available. */
-static DEFINE_PER_CPU(u64, smca_misc_banks_map);
-
 static void amd_threshold_interrupt(void);
 static void amd_deferred_error_interrupt(void);
 
@@ -264,28 +261,6 @@ static void default_deferred_error_interrupt(void)
 }
 void (*deferred_error_int_vector)(void) = default_deferred_error_interrupt;
 
-static void smca_set_misc_banks_map(unsigned int bank, unsigned int cpu)
-{
-	u32 low, high;
-
-	/*
-	 * For SMCA enabled processors, BLKPTR field of the first MISC register
-	 * (MCx_MISC0) indicates presence of additional MISC regs set (MISC1-4).
-	 */
-	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_CONFIG(bank), &low, &high))
-		return;
-
-	if (!(low & MCI_CONFIG_MCAX))
-		return;
-
-	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_MISC(bank), &low, &high))
-		return;
-
-	if (low & MASK_BLKPTR_LO)
-		per_cpu(smca_misc_banks_map, cpu) |= BIT_ULL(bank);
-
-}
-
 static void smca_configure(unsigned int bank, unsigned int cpu)
 {
 	u8 *bank_counts = this_cpu_ptr(smca_bank_counts);
@@ -326,8 +301,6 @@ static void smca_configure(unsigned int bank, unsigned int cpu)
 		wrmsr(smca_config, low, high);
 	}
 
-	smca_set_misc_banks_map(bank, cpu);
-
 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_IPID(bank), &low, &high)) {
 		pr_warn("Failed to read MCA_IPID for bank %d\n", bank);
 		return;
@@ -532,9 +505,6 @@ static u32 smca_get_block_address(unsigned int bank, unsigned int block,
 	if (!block)
 		return MSR_AMD64_SMCA_MCx_MISC(bank);
 
-	if (!(per_cpu(smca_misc_banks_map, cpu) & BIT_ULL(bank)))
-		return 0;
-
 	return MSR_AMD64_SMCA_MCx_MISCy(bank, block - 1);
 }
 

-- 
2.49.0

[PATCH v3 04/17] x86/mce/amd: Put list_head in threshold_bank

[Yazen Ghannam]

The threshold_bank structure is a container for one or more
threshold_block structures. Currently, the container has a single
pointer to the 'first' threshold_block structure which then has a linked
list of the remaining threshold_block structures.

This results in an extra level of indirection where the 'first' block is
checked before iterating over the remaining blocks.

Remove the indirection by including the head of the block list in the
threshold_bank structure which already acts as a container for all the
bank's thresholding blocks.

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-4-3636547fe05f@amd.com
    
    v2->v3:
    * Added tags from Qiuxu and Tony.
    
    v1->v2:
    * New in v2.

 arch/x86/kernel/cpu/mce/amd.c | 43 ++++++++++++-------------------------------
 1 file changed, 12 insertions(+), 31 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 46ff41c1b50d..8e5a07f78346 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -241,7 +241,8 @@ struct threshold_block {
 
 struct threshold_bank {
 	struct kobject		*kobj;
-	struct threshold_block	*blocks;
+	/* List of threshold blocks within this MCA bank. */
+	struct list_head	miscj;
 };
 
 static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
@@ -901,9 +902,9 @@ static void log_and_reset_block(struct threshold_block *block)
  */
 static void amd_threshold_interrupt(void)
 {
-	struct threshold_block *first_block = NULL, *block = NULL, *tmp = NULL;
-	struct threshold_bank **bp = this_cpu_read(threshold_banks);
+	struct threshold_bank **bp = this_cpu_read(threshold_banks), *thr_bank;
 	unsigned int bank, cpu = smp_processor_id();
+	struct threshold_block *block, *tmp;
 
 	/*
 	 * Validate that the threshold bank has been initialized already. The
@@ -917,16 +918,11 @@ static void amd_threshold_interrupt(void)
 		if (!(per_cpu(bank_map, cpu) & BIT_ULL(bank)))
 			continue;
 
-		first_block = bp[bank]->blocks;
-		if (!first_block)
+		thr_bank = bp[bank];
+		if (!thr_bank)
 			continue;
 
-		/*
-		 * The first block is also the head of the list. Check it first
-		 * before iterating over the rest.
-		 */
-		log_and_reset_block(first_block);
-		list_for_each_entry_safe(block, tmp, &first_block->miscj, miscj)
+		list_for_each_entry_safe(block, tmp, &thr_bank->miscj, miscj)
 			log_and_reset_block(block);
 	}
 }
@@ -1145,13 +1141,7 @@ static int allocate_threshold_blocks(unsigned int cpu, struct threshold_bank *tb
 		default_attrs[2] = NULL;
 	}
 
-	INIT_LIST_HEAD(&b->miscj);
-
-	/* This is safe as @tb is not visible yet */
-	if (tb->blocks)
-		list_add(&b->miscj, &tb->blocks->miscj);
-	else
-		tb->blocks = b;
+	list_add(&b->miscj, &tb->miscj);
 
 	err = kobject_init_and_add(&b->kobj, &threshold_ktype, tb->kobj, get_name(cpu, bank, b));
 	if (err)
@@ -1202,6 +1192,8 @@ static int threshold_create_bank(struct threshold_bank **bp, unsigned int cpu,
 		goto out_free;
 	}
 
+	INIT_LIST_HEAD(&b->miscj);
+
 	err = allocate_threshold_blocks(cpu, b, bank, 0, mca_msr_reg(bank, MCA_MISC));
 	if (err)
 		goto out_kobj;
@@ -1222,26 +1214,15 @@ static void threshold_block_release(struct kobject *kobj)
 	kfree(to_block(kobj));
 }
 
-static void deallocate_threshold_blocks(struct threshold_bank *bank)
+static void threshold_remove_bank(struct threshold_bank *bank)
 {
 	struct threshold_block *pos, *tmp;
 
-	list_for_each_entry_safe(pos, tmp, &bank->blocks->miscj, miscj) {
+	list_for_each_entry_safe(pos, tmp, &bank->miscj, miscj) {
 		list_del(&pos->miscj);
 		kobject_put(&pos->kobj);
 	}
 
-	kobject_put(&bank->blocks->kobj);
-}
-
-static void threshold_remove_bank(struct threshold_bank *bank)
-{
-	if (!bank->blocks)
-		goto out_free;
-
-	deallocate_threshold_blocks(bank);
-
-out_free:
 	kobject_put(bank->kobj);
 	kfree(bank);
 }

-- 
2.49.0

[PATCH v3 05/17] x86/mce: Cleanup bank processing on init

[Yazen Ghannam]

From: Borislav Petkov <bp@suse.de>

Unify the bank preparation into __mcheck_cpu_init_clear_banks(), rename
that function to what it does now - prepares banks. Do this so that
generic and vendor banks init goes first so that settings done during
that init can take effect before the first bank polling takes place.

Move __mcheck_cpu_check_banks() into __mcheck_cpu_init_prepare_banks()
as it already loops over the banks.

The MCP_DONTLOG flag is no longer needed, since the MCA polling function
is now called only if boot-time logging should be done.

Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Yazen Ghannam <yazen.ghannam@amd.com>
Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-5-3636547fe05f@amd.com
    
    v2->v3:
    * Update commit message.
    * Add tags from Qiuxu and Tony.
    
    v1->v2:
    * New in v2, but based on old patch (see link).
    * Kept old tags for reference.

 arch/x86/include/asm/mce.h     |  3 +-
 arch/x86/kernel/cpu/mce/core.c | 63 ++++++++++++------------------------------
 2 files changed, 19 insertions(+), 47 deletions(-)

diff --git a/arch/x86/include/asm/mce.h b/arch/x86/include/asm/mce.h
index 752802bf966b..3224f3862dc8 100644
--- a/arch/x86/include/asm/mce.h
+++ b/arch/x86/include/asm/mce.h
@@ -290,8 +290,7 @@ DECLARE_PER_CPU(mce_banks_t, mce_poll_banks);
 enum mcp_flags {
 	MCP_TIMESTAMP	= BIT(0),	/* log time stamp */
 	MCP_UC		= BIT(1),	/* log uncorrected errors */
-	MCP_DONTLOG	= BIT(2),	/* only clear, don't log */
-	MCP_QUEUE_LOG	= BIT(3),	/* only queue to genpool */
+	MCP_QUEUE_LOG	= BIT(2),	/* only queue to genpool */
 };
 
 void machine_check_poll(enum mcp_flags flags, mce_banks_t *b);
diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index 72c2aa0809c0..ee801f8862d8 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -807,9 +807,6 @@ void machine_check_poll(enum mcp_flags flags, mce_banks_t *b)
 		continue;
 
 log_it:
-		if (flags & MCP_DONTLOG)
-			goto clear_it;
-
 		mce_read_aux(&err, i);
 		m->severity = mce_severity(m, NULL, NULL, false);
 		/*
@@ -1807,7 +1804,7 @@ static void __mcheck_cpu_mce_banks_init(void)
 		/*
 		 * Init them all, __mcheck_cpu_apply_quirks() is going to apply
 		 * the required vendor quirks before
-		 * __mcheck_cpu_init_clear_banks() does the final bank setup.
+		 * __mcheck_cpu_init_prepare_banks() does the final bank setup.
 		 */
 		b->ctl = -1ULL;
 		b->init = true;
@@ -1846,21 +1843,8 @@ static void __mcheck_cpu_cap_init(void)
 
 static void __mcheck_cpu_init_generic(void)
 {
-	enum mcp_flags m_fl = 0;
-	mce_banks_t all_banks;
 	u64 cap;
 
-	if (!mca_cfg.bootlog)
-		m_fl = MCP_DONTLOG;
-
-	/*
-	 * Log the machine checks left over from the previous reset. Log them
-	 * only, do not start processing them. That will happen in mcheck_late_init()
-	 * when all consumers have been registered on the notifier chain.
-	 */
-	bitmap_fill(all_banks, MAX_NR_BANKS);
-	machine_check_poll(MCP_UC | MCP_QUEUE_LOG | m_fl, &all_banks);
-
 	cr4_set_bits(X86_CR4_MCE);
 
 	rdmsrq(MSR_IA32_MCG_CAP, cap);
@@ -1868,36 +1852,23 @@ static void __mcheck_cpu_init_generic(void)
 		wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
 }
 
-static void __mcheck_cpu_init_clear_banks(void)
+static void __mcheck_cpu_init_prepare_banks(void)
 {
 	struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
+	u64 msrval;
 	int i;
 
-	for (i = 0; i < this_cpu_read(mce_num_banks); i++) {
-		struct mce_bank *b = &mce_banks[i];
+	/*
+	 * Log the machine checks left over from the previous reset. Log them
+	 * only, do not start processing them. That will happen in mcheck_late_init()
+	 * when all consumers have been registered on the notifier chain.
+	 */
+	if (mca_cfg.bootlog) {
+		mce_banks_t all_banks;
 
-		if (!b->init)
-			continue;
-		wrmsrq(mca_msr_reg(i, MCA_CTL), b->ctl);
-		wrmsrq(mca_msr_reg(i, MCA_STATUS), 0);
+		bitmap_fill(all_banks, MAX_NR_BANKS);
+		machine_check_poll(MCP_UC | MCP_QUEUE_LOG, &all_banks);
 	}
-}
-
-/*
- * Do a final check to see if there are any unused/RAZ banks.
- *
- * This must be done after the banks have been initialized and any quirks have
- * been applied.
- *
- * Do not call this from any user-initiated flows, e.g. CPU hotplug or sysfs.
- * Otherwise, a user who disables a bank will not be able to re-enable it
- * without a system reboot.
- */
-static void __mcheck_cpu_check_banks(void)
-{
-	struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
-	u64 msrval;
-	int i;
 
 	for (i = 0; i < this_cpu_read(mce_num_banks); i++) {
 		struct mce_bank *b = &mce_banks[i];
@@ -1905,6 +1876,9 @@ static void __mcheck_cpu_check_banks(void)
 		if (!b->init)
 			continue;
 
+		wrmsrq(mca_msr_reg(i, MCA_CTL), b->ctl);
+		wrmsrq(mca_msr_reg(i, MCA_STATUS), 0);
+
 		rdmsrq(mca_msr_reg(i, MCA_CTL), msrval);
 		b->init = !!msrval;
 	}
@@ -2310,8 +2284,7 @@ void mcheck_cpu_init(struct cpuinfo_x86 *c)
 	__mcheck_cpu_init_early(c);
 	__mcheck_cpu_init_generic();
 	__mcheck_cpu_init_vendor(c);
-	__mcheck_cpu_init_clear_banks();
-	__mcheck_cpu_check_banks();
+	__mcheck_cpu_init_prepare_banks();
 	__mcheck_cpu_setup_timer();
 }
 
@@ -2479,7 +2452,7 @@ static void mce_syscore_resume(void)
 {
 	__mcheck_cpu_init_generic();
 	__mcheck_cpu_init_vendor(raw_cpu_ptr(&cpu_info));
-	__mcheck_cpu_init_clear_banks();
+	__mcheck_cpu_init_prepare_banks();
 }
 
 static struct syscore_ops mce_syscore_ops = {
@@ -2497,7 +2470,7 @@ static void mce_cpu_restart(void *data)
 	if (!mce_available(raw_cpu_ptr(&cpu_info)))
 		return;
 	__mcheck_cpu_init_generic();
-	__mcheck_cpu_init_clear_banks();
+	__mcheck_cpu_init_prepare_banks();
 	__mcheck_cpu_init_timer();
 }
 

-- 
2.49.0

[PATCH v3 06/17] x86/mce: Remove __mcheck_cpu_init_early()

[Yazen Ghannam]

The __mcheck_cpu_init_early() function was introduced so that some
vendor-specific features are detected before the first MCA polling event
done in __mcheck_cpu_init_generic().

Currently, __mcheck_cpu_init_early() is only used on AMD-based systems and
additional code will be needed to support various system configurations.

However, the current and future vendor-specific code should be done during
vendor init. This keeps all the vendor code in a common location and
simplifies the generic init flow.

Move all the __mcheck_cpu_init_early() code into mce_amd_feature_init().

Also, move __mcheck_cpu_init_generic() after
__mcheck_cpu_init_prepare_banks() so that MCA is enabled after the first
MCA polling event.

Additionally, this brings the MCA init flow closer to what is described
in the x86 docs.

The AMD PPRs say
  "The operating system must initialize the MCA_CONFIG registers prior to
  initialization of the MCA_CTL registers.

  The MCA_CTL registers must be initialized prior to enabling the error
  reporting banks in MCG_CTL".

However, the Intel SDM "Machine-Check Initialization Pseudocode" says
MCG_CTL first then MCi_CTL.

But both agree that CR4.MCE should be set last.

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-6-3636547fe05f@amd.com
    
    v2->v3:
    * Update commit message.
    * Add tags from Qiuxu and Tony.
    
    v1->v2:
    * New in v2, but based on old patch (see link).
    * Changed cpu_has() to cpu_feature_enabled().

 arch/x86/kernel/cpu/mce/amd.c  |  4 ++++
 arch/x86/kernel/cpu/mce/core.c | 20 +++-----------------
 2 files changed, 7 insertions(+), 17 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 8e5a07f78346..aa23139a3092 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -656,6 +656,10 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 	u32 low = 0, high = 0, address = 0;
 	int offset = -1;
 
+	mce_flags.overflow_recov = cpu_feature_enabled(X86_FEATURE_OVERFLOW_RECOV);
+	mce_flags.succor	 = cpu_feature_enabled(X86_FEATURE_SUCCOR);
+	mce_flags.smca		 = cpu_feature_enabled(X86_FEATURE_SMCA);
+	mce_flags.amd_threshold	 = 1;
 
 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
 		if (mce_flags.smca)
diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index ee801f8862d8..331cd8984395 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -2029,19 +2029,6 @@ static bool __mcheck_cpu_ancient_init(struct cpuinfo_x86 *c)
 	return false;
 }
 
-/*
- * Init basic CPU features needed for early decoding of MCEs.
- */
-static void __mcheck_cpu_init_early(struct cpuinfo_x86 *c)
-{
-	if (c->x86_vendor == X86_VENDOR_AMD || c->x86_vendor == X86_VENDOR_HYGON) {
-		mce_flags.overflow_recov = !!cpu_has(c, X86_FEATURE_OVERFLOW_RECOV);
-		mce_flags.succor	 = !!cpu_has(c, X86_FEATURE_SUCCOR);
-		mce_flags.smca		 = !!cpu_has(c, X86_FEATURE_SMCA);
-		mce_flags.amd_threshold	 = 1;
-	}
-}
-
 static void mce_centaur_feature_init(struct cpuinfo_x86 *c)
 {
 	struct mca_config *cfg = &mca_cfg;
@@ -2281,10 +2268,9 @@ void mcheck_cpu_init(struct cpuinfo_x86 *c)
 
 	mca_cfg.initialized = 1;
 
-	__mcheck_cpu_init_early(c);
-	__mcheck_cpu_init_generic();
 	__mcheck_cpu_init_vendor(c);
 	__mcheck_cpu_init_prepare_banks();
+	__mcheck_cpu_init_generic();
 	__mcheck_cpu_setup_timer();
 }
 
@@ -2450,9 +2436,9 @@ static void mce_syscore_shutdown(void)
  */
 static void mce_syscore_resume(void)
 {
-	__mcheck_cpu_init_generic();
 	__mcheck_cpu_init_vendor(raw_cpu_ptr(&cpu_info));
 	__mcheck_cpu_init_prepare_banks();
+	__mcheck_cpu_init_generic();
 }
 
 static struct syscore_ops mce_syscore_ops = {
@@ -2469,8 +2455,8 @@ static void mce_cpu_restart(void *data)
 {
 	if (!mce_available(raw_cpu_ptr(&cpu_info)))
 		return;
-	__mcheck_cpu_init_generic();
 	__mcheck_cpu_init_prepare_banks();
+	__mcheck_cpu_init_generic();
 	__mcheck_cpu_init_timer();
 }
 

-- 
2.49.0

[PATCH v3 07/17] x86/mce: Define BSP-only init

[Yazen Ghannam]

Currently, MCA initialization is executed identically on each CPU as
they are brought online. However, a number of MCA initialization tasks
only need to be done once.

Define a function to collect all 'global' init tasks and call this from
the BSP only. Start with CPU features.

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-7-3636547fe05f@amd.com
    
    v2->v3:
    * Add tags from Qiuxu and Tony.
    
    v1->v2:
    * New in v2.

 arch/x86/include/asm/mce.h     |  2 ++
 arch/x86/kernel/cpu/common.c   |  1 +
 arch/x86/kernel/cpu/mce/amd.c  |  3 ---
 arch/x86/kernel/cpu/mce/core.c | 29 ++++++++++++++++++++++-------
 4 files changed, 25 insertions(+), 10 deletions(-)

diff --git a/arch/x86/include/asm/mce.h b/arch/x86/include/asm/mce.h
index 3224f3862dc8..0108f69ec46a 100644
--- a/arch/x86/include/asm/mce.h
+++ b/arch/x86/include/asm/mce.h
@@ -241,12 +241,14 @@ struct cper_ia_proc_ctx;
 
 #ifdef CONFIG_X86_MCE
 int mcheck_init(void);
+void cpu_mca_init(struct cpuinfo_x86 *c);
 void mcheck_cpu_init(struct cpuinfo_x86 *c);
 void mcheck_cpu_clear(struct cpuinfo_x86 *c);
 int apei_smca_report_x86_error(struct cper_ia_proc_ctx *ctx_info,
 			       u64 lapic_id);
 #else
 static inline int mcheck_init(void) { return 0; }
+static inline void cpu_mca_init(struct cpuinfo_x86 *c) {}
 static inline void mcheck_cpu_init(struct cpuinfo_x86 *c) {}
 static inline void mcheck_cpu_clear(struct cpuinfo_x86 *c) {}
 static inline int apei_smca_report_x86_error(struct cper_ia_proc_ctx *ctx_info,
diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index 079ded4eeb86..8e3e51281f12 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -1690,6 +1690,7 @@ static void __init early_identify_cpu(struct cpuinfo_x86 *c)
 		setup_clear_cpu_cap(X86_FEATURE_LA57);
 
 	detect_nopl();
+	cpu_mca_init(c);
 }
 
 void __init init_cpu_devs(void)
diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index aa23139a3092..206973d7dbcc 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -656,9 +656,6 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 	u32 low = 0, high = 0, address = 0;
 	int offset = -1;
 
-	mce_flags.overflow_recov = cpu_feature_enabled(X86_FEATURE_OVERFLOW_RECOV);
-	mce_flags.succor	 = cpu_feature_enabled(X86_FEATURE_SUCCOR);
-	mce_flags.smca		 = cpu_feature_enabled(X86_FEATURE_SMCA);
 	mce_flags.amd_threshold	 = 1;
 
 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index 331cd8984395..d0a29e22cab0 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -1832,13 +1832,6 @@ static void __mcheck_cpu_cap_init(void)
 	this_cpu_write(mce_num_banks, b);
 
 	__mcheck_cpu_mce_banks_init();
-
-	/* Use accurate RIP reporting if available. */
-	if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9)
-		mca_cfg.rip_msr = MSR_IA32_MCG_EIP;
-
-	if (cap & MCG_SER_P)
-		mca_cfg.ser = 1;
 }
 
 static void __mcheck_cpu_init_generic(void)
@@ -2238,6 +2231,28 @@ DEFINE_IDTENTRY_RAW(exc_machine_check)
 }
 #endif
 
+/* Called only on the boot CPU. */
+void cpu_mca_init(struct cpuinfo_x86 *c)
+{
+	u64 cap;
+
+	if (!mce_available(c))
+		return;
+
+	mce_flags.overflow_recov = cpu_feature_enabled(X86_FEATURE_OVERFLOW_RECOV);
+	mce_flags.succor	 = cpu_feature_enabled(X86_FEATURE_SUCCOR);
+	mce_flags.smca		 = cpu_feature_enabled(X86_FEATURE_SMCA);
+
+	rdmsrl(MSR_IA32_MCG_CAP, cap);
+
+	/* Use accurate RIP reporting if available. */
+	if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9)
+		mca_cfg.rip_msr = MSR_IA32_MCG_EIP;
+
+	if (cap & MCG_SER_P)
+		mca_cfg.ser = 1;
+}
+
 /*
  * Called for each booted CPU to set up machine checks.
  * Must be called with preempt off:

-- 
2.49.0

Re: [PATCH v3 07/17] x86/mce: Define BSP-only init

[Borislav Petkov]

On Tue, Apr 15, 2025 at 02:55:02PM +0000, Yazen Ghannam wrote:
> +/* Called only on the boot CPU. */
> +void cpu_mca_init(struct cpuinfo_x86 *c)

I guess mca_bsp_init() or mca_init() or so. Probably former as the name makes
it perfectly clear and obviates the need for the comment above.

Because there's also mcheck_cpu_init() which is per-CPU and that's confusion
waiting to happen.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 07/17] x86/mce: Define BSP-only init

[Yazen Ghannam]

On Thu, Apr 17, 2025 at 04:18:35AM +0200, Borislav Petkov wrote:
> On Tue, Apr 15, 2025 at 02:55:02PM +0000, Yazen Ghannam wrote:
> > +/* Called only on the boot CPU. */
> > +void cpu_mca_init(struct cpuinfo_x86 *c)
> 
> I guess mca_bsp_init() or mca_init() or so. Probably former as the name makes
> it perfectly clear and obviates the need for the comment above.
> 
> Because there's also mcheck_cpu_init() which is per-CPU and that's confusion
> waiting to happen.
> 

Okay, will change.

Thanks,
Yazen

[PATCH v3 08/17] x86/mce: Define BSP-only SMCA init

[Yazen Ghannam]

Currently on AMD systems, MCA interrupt handler functions are set during
CPU init. However, the functions only need to be set once for the whole
system.

Assign the handlers only during BSP init. Do so only for SMCA systems to
maintain the old behavior for legacy systems.

Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-8-3636547fe05f@amd.com
    
    v2->v3:
    * No change.
    
    v1->v2:
    * New in v2.

 arch/x86/kernel/cpu/mce/amd.c      | 6 ++++++
 arch/x86/kernel/cpu/mce/core.c     | 3 +++
 arch/x86/kernel/cpu/mce/internal.h | 2 ++
 3 files changed, 11 insertions(+)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 206973d7dbcc..0d84b171b851 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -687,6 +687,12 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 		deferred_error_interrupt_enable(c);
 }
 
+void mce_smca_cpu_init(void)
+{
+	mce_threshold_vector	  = amd_threshold_interrupt;
+	deferred_error_int_vector = amd_deferred_error_interrupt;
+}
+
 /*
  * DRAM ECC errors are reported in the Northbridge (bank 4) with
  * Extended Error Code 8.
diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index d0a29e22cab0..444d006366fd 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -2243,6 +2243,9 @@ void cpu_mca_init(struct cpuinfo_x86 *c)
 	mce_flags.succor	 = cpu_feature_enabled(X86_FEATURE_SUCCOR);
 	mce_flags.smca		 = cpu_feature_enabled(X86_FEATURE_SMCA);
 
+	if (mce_flags.smca)
+		mce_smca_cpu_init();
+
 	rdmsrl(MSR_IA32_MCG_CAP, cap);
 
 	/* Use accurate RIP reporting if available. */
diff --git a/arch/x86/kernel/cpu/mce/internal.h b/arch/x86/kernel/cpu/mce/internal.h
index 64ac25b95360..87b69935d57d 100644
--- a/arch/x86/kernel/cpu/mce/internal.h
+++ b/arch/x86/kernel/cpu/mce/internal.h
@@ -294,12 +294,14 @@ static __always_inline void smca_extract_err_addr(struct mce *m)
 	m->addr &= GENMASK_ULL(55, lsb);
 }
 
+void mce_smca_cpu_init(void);
 #else
 static inline void mce_threshold_create_device(unsigned int cpu)	{ }
 static inline void mce_threshold_remove_device(unsigned int cpu)	{ }
 static inline bool amd_filter_mce(struct mce *m) { return false; }
 static inline bool amd_mce_usable_address(struct mce *m) { return false; }
 static inline void smca_extract_err_addr(struct mce *m) { }
+static inline void mce_smca_cpu_init(void) {}
 #endif
 
 #ifdef CONFIG_X86_ANCIENT_MCE

-- 
2.49.0

Re: [PATCH v3 08/17] x86/mce: Define BSP-only SMCA init

[Borislav Petkov]

On Tue, Apr 15, 2025 at 02:55:03PM +0000, Yazen Ghannam wrote:
> diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
> index 206973d7dbcc..0d84b171b851 100644
> --- a/arch/x86/kernel/cpu/mce/amd.c
> +++ b/arch/x86/kernel/cpu/mce/amd.c
> @@ -687,6 +687,12 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
>  		deferred_error_interrupt_enable(c);
>  }
>  
> +void mce_smca_cpu_init(void)

Same note as about the previous patch.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 08/17] x86/mce: Define BSP-only SMCA init

[Yazen Ghannam]

On Thu, Apr 17, 2025 at 11:52:24AM +0200, Borislav Petkov wrote:
> On Tue, Apr 15, 2025 at 02:55:03PM +0000, Yazen Ghannam wrote:
> > diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
> > index 206973d7dbcc..0d84b171b851 100644
> > --- a/arch/x86/kernel/cpu/mce/amd.c
> > +++ b/arch/x86/kernel/cpu/mce/amd.c
> > @@ -687,6 +687,12 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
> >  		deferred_error_interrupt_enable(c);
> >  }
> >  
> > +void mce_smca_cpu_init(void)
> 
> Same note as about the previous patch.
> 

Ack. Will change.

Thanks,
Yazen

[PATCH v3 09/17] x86/mce: Do 'UNKNOWN' vendor check early

[Yazen Ghannam]

The 'UNKNOWN' vendor check is handled as a quirk that is run on each
online CPU. However, all CPUs are expected to have the same vendor.

Move the 'UNKNOWN' vendor check to the BSP-only init so it is done early
and once. Remove the unnecessary return value from the quirks check.

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-9-3636547fe05f@amd.com
    
    v2->v3:
    * Add tags from Qiuxu and Tony.
    
    v1->v2:
    * New in v2.

 arch/x86/kernel/cpu/mce/core.c | 18 ++++++++----------
 1 file changed, 8 insertions(+), 10 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index 444d006366fd..9dc9d672a7d1 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -1974,14 +1974,11 @@ static void apply_quirks_zhaoxin(struct cpuinfo_x86 *c)
 }
 
 /* Add per CPU specific workarounds here */
-static bool __mcheck_cpu_apply_quirks(struct cpuinfo_x86 *c)
+static void __mcheck_cpu_apply_quirks(struct cpuinfo_x86 *c)
 {
 	struct mca_config *cfg = &mca_cfg;
 
 	switch (c->x86_vendor) {
-	case X86_VENDOR_UNKNOWN:
-		pr_info("unknown CPU type - not enabling MCE support\n");
-		return false;
 	case X86_VENDOR_AMD:
 		apply_quirks_amd(c);
 		break;
@@ -1997,8 +1994,6 @@ static bool __mcheck_cpu_apply_quirks(struct cpuinfo_x86 *c)
 		cfg->monarch_timeout = 0;
 	if (cfg->bootlog != 0)
 		cfg->panic_timeout = 30;
-
-	return true;
 }
 
 static bool __mcheck_cpu_ancient_init(struct cpuinfo_x86 *c)
@@ -2239,6 +2234,12 @@ void cpu_mca_init(struct cpuinfo_x86 *c)
 	if (!mce_available(c))
 		return;
 
+	if (c->x86_vendor == X86_VENDOR_UNKNOWN) {
+		mca_cfg.disabled = 1;
+		pr_info("unknown CPU type - not enabling MCE support\n");
+		return;
+	}
+
 	mce_flags.overflow_recov = cpu_feature_enabled(X86_FEATURE_OVERFLOW_RECOV);
 	mce_flags.succor	 = cpu_feature_enabled(X86_FEATURE_SUCCOR);
 	mce_flags.smca		 = cpu_feature_enabled(X86_FEATURE_SMCA);
@@ -2273,10 +2274,7 @@ void mcheck_cpu_init(struct cpuinfo_x86 *c)
 
 	__mcheck_cpu_cap_init();
 
-	if (!__mcheck_cpu_apply_quirks(c)) {
-		mca_cfg.disabled = 1;
-		return;
-	}
+	__mcheck_cpu_apply_quirks(c);
 
 	if (!mce_gen_pool_init()) {
 		mca_cfg.disabled = 1;

-- 
2.49.0

[PATCH v3 10/17] x86/mce: Separate global and per-CPU quirks

[Yazen Ghannam]

Many quirks are global configuration settings and a handful apply to
each CPU.

Move the per-CPU quirks to vendor init to execute them on each online
CPU. Set the global quirks during BSP-only init so they're only executed
once and early.

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-10-3636547fe05f@amd.com
    
    v2->v3:
    * Update code comment.
    * Add tags from Qiuxu and Tony.
    
    v1->v2:
    * New in v2.

 arch/x86/kernel/cpu/mce/amd.c   | 23 +++++++++++++++++++++++
 arch/x86/kernel/cpu/mce/core.c  | 36 ++----------------------------------
 arch/x86/kernel/cpu/mce/intel.c | 18 ++++++++++++++++++
 3 files changed, 43 insertions(+), 34 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 0d84b171b851..6a69cac36c18 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -649,6 +649,28 @@ static void disable_err_thresholding(struct cpuinfo_x86 *c, unsigned int bank)
 		wrmsrq(MSR_K7_HWCR, hwcr);
 }
 
+static void amd_apply_quirks(struct cpuinfo_x86 *c)
+{
+	struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
+
+	/* This should be disabled by the BIOS, but isn't always */
+	if (c->x86 == 15 && this_cpu_read(mce_num_banks) > 4) {
+		/*
+		 * disable GART TBL walk error reporting, which
+		 * trips off incorrectly with the IOMMU & 3ware
+		 * & Cerberus:
+		 */
+		clear_bit(10, (unsigned long *)&mce_banks[4].ctl);
+	}
+
+	/*
+	 * Various K7s with broken bank 0 around. Always disable
+	 * by default.
+	 */
+	if (c->x86 == 6 && this_cpu_read(mce_num_banks))
+		mce_banks[0].ctl = 0;
+}
+
 /* cpu init entry point, called from mce.c with preempt off */
 void mce_amd_feature_init(struct cpuinfo_x86 *c)
 {
@@ -656,6 +678,7 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 	u32 low = 0, high = 0, address = 0;
 	int offset = -1;
 
+	amd_apply_quirks(c);
 	mce_flags.amd_threshold	 = 1;
 
 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index 9dc9d672a7d1..413c68f18084 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -1879,18 +1879,6 @@ static void __mcheck_cpu_init_prepare_banks(void)
 
 static void apply_quirks_amd(struct cpuinfo_x86 *c)
 {
-	struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
-
-	/* This should be disabled by the BIOS, but isn't always */
-	if (c->x86 == 15 && this_cpu_read(mce_num_banks) > 4) {
-		/*
-		 * disable GART TBL walk error reporting, which
-		 * trips off incorrectly with the IOMMU & 3ware
-		 * & Cerberus:
-		 */
-		clear_bit(10, (unsigned long *)&mce_banks[4].ctl);
-	}
-
 	if (c->x86 < 0x11 && mca_cfg.bootlog < 0) {
 		/*
 		 * Lots of broken BIOS around that don't clear them
@@ -1899,13 +1887,6 @@ static void apply_quirks_amd(struct cpuinfo_x86 *c)
 		mca_cfg.bootlog = 0;
 	}
 
-	/*
-	 * Various K7s with broken bank 0 around. Always disable
-	 * by default.
-	 */
-	if (c->x86 == 6 && this_cpu_read(mce_num_banks))
-		mce_banks[0].ctl = 0;
-
 	/*
 	 * overflow_recov is supported for F15h Models 00h-0fh
 	 * even though we don't have a CPUID bit for it.
@@ -1919,23 +1900,10 @@ static void apply_quirks_amd(struct cpuinfo_x86 *c)
 
 static void apply_quirks_intel(struct cpuinfo_x86 *c)
 {
-	struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
-
 	/* Older CPUs (prior to family 6) don't need quirks. */
 	if (c->x86_vfm < INTEL_PENTIUM_PRO)
 		return;
 
-	/*
-	 * SDM documents that on family 6 bank 0 should not be written
-	 * because it aliases to another special BIOS controlled
-	 * register.
-	 * But it's not aliased anymore on model 0x1a+
-	 * Don't ignore bank 0 completely because there could be a
-	 * valid event later, merely don't write CTL0.
-	 */
-	if (c->x86_vfm < INTEL_NEHALEM_EP && this_cpu_read(mce_num_banks))
-		mce_banks[0].init = false;
-
 	/*
 	 * All newer Intel systems support MCE broadcasting. Enable
 	 * synchronization with a one second timeout.
@@ -2255,6 +2223,8 @@ void cpu_mca_init(struct cpuinfo_x86 *c)
 
 	if (cap & MCG_SER_P)
 		mca_cfg.ser = 1;
+
+	__mcheck_cpu_apply_quirks(c);
 }
 
 /*
@@ -2274,8 +2244,6 @@ void mcheck_cpu_init(struct cpuinfo_x86 *c)
 
 	__mcheck_cpu_cap_init();
 
-	__mcheck_cpu_apply_quirks(c);
-
 	if (!mce_gen_pool_init()) {
 		mca_cfg.disabled = 1;
 		pr_emerg("Couldn't allocate MCE records pool!\n");
diff --git a/arch/x86/kernel/cpu/mce/intel.c b/arch/x86/kernel/cpu/mce/intel.c
index efcf21e9552e..ae9417d634ac 100644
--- a/arch/x86/kernel/cpu/mce/intel.c
+++ b/arch/x86/kernel/cpu/mce/intel.c
@@ -468,8 +468,26 @@ static void intel_imc_init(struct cpuinfo_x86 *c)
 	}
 }
 
+static void intel_apply_quirks(struct cpuinfo_x86 *c)
+{
+	/*
+	 * SDM documents that on family 6 bank 0 should not be written
+	 * because it aliases to another special BIOS controlled
+	 * register.
+	 * But it's not aliased anymore on model 0x1a+
+	 * Don't ignore bank 0 completely because there could be a
+	 * valid event later, merely don't write CTL0.
+	 *
+	 * Older CPUs (prior to family 6) can't reach this point and already
+	 * return early due to the check of __mcheck_cpu_ancient_init().
+	 */
+	if (c->x86_vfm < INTEL_NEHALEM_EP && this_cpu_read(mce_num_banks))
+		this_cpu_ptr(mce_banks_array)[0].init = false;
+}
+
 void mce_intel_feature_init(struct cpuinfo_x86 *c)
 {
+	intel_apply_quirks(c);
 	intel_init_cmci();
 	intel_init_lmce();
 	intel_imc_init(c);

-- 
2.49.0

Re: [PATCH v3 10/17] x86/mce: Separate global and per-CPU quirks

[Borislav Petkov]

On Tue, Apr 15, 2025 at 02:55:05PM +0000, Yazen Ghannam wrote:
>  /* cpu init entry point, called from mce.c with preempt off */
>  void mce_amd_feature_init(struct cpuinfo_x86 *c)
>  {
> @@ -656,6 +678,7 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
>  	u32 low = 0, high = 0, address = 0;
>  	int offset = -1;
>  
> +	amd_apply_quirks(c);

<---- newline here.

>  	mce_flags.amd_threshold	 = 1;
>  
>  	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {

...

> @@ -2255,6 +2223,8 @@ void cpu_mca_init(struct cpuinfo_x86 *c)
>  
>  	if (cap & MCG_SER_P)
>  		mca_cfg.ser = 1;
> +
> +	__mcheck_cpu_apply_quirks(c);

Right, this is not a per-CPU init function anymore but a one-time thing and
called only here so put its contents here and whack it.

Thx.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 10/17] x86/mce: Separate global and per-CPU quirks

[Yazen Ghannam]

On Thu, Apr 17, 2025 at 02:16:31PM +0200, Borislav Petkov wrote:
> On Tue, Apr 15, 2025 at 02:55:05PM +0000, Yazen Ghannam wrote:
> >  /* cpu init entry point, called from mce.c with preempt off */
> >  void mce_amd_feature_init(struct cpuinfo_x86 *c)
> >  {
> > @@ -656,6 +678,7 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
> >  	u32 low = 0, high = 0, address = 0;
> >  	int offset = -1;
> >  
> > +	amd_apply_quirks(c);
> 
> <---- newline here.
> 

Ack.

> >  	mce_flags.amd_threshold	 = 1;
> >  
> >  	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
> 
> ...
> 
> > @@ -2255,6 +2223,8 @@ void cpu_mca_init(struct cpuinfo_x86 *c)
> >  
> >  	if (cap & MCG_SER_P)
> >  		mca_cfg.ser = 1;
> > +
> > +	__mcheck_cpu_apply_quirks(c);
> 
> Right, this is not a per-CPU init function anymore but a one-time thing and
> called only here so put its contents here and whack it.
> 

Will do.

Thanks,
Yazen

[PATCH v3 11/17] x86/mce: Move machine_check_poll() status checks to helper functions

[Yazen Ghannam]

There are a number of generic and vendor-specific status checks in
machine_check_poll(). These are used to determine if an error should be
skipped.

Move these into helper functions. Future vendor-specific checks will be
added to the helpers.

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-11-3636547fe05f@amd.com
    
    v2->v3:
    * Add tags from Qiuxu and Tony.
    
    v1->v2:
    * Change log_poll_error() to should_log_poll_error().
    * Keep code comment.

 arch/x86/kernel/cpu/mce/core.c | 88 +++++++++++++++++++++++-------------------
 1 file changed, 48 insertions(+), 40 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index 413c68f18084..c82c9e435066 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -714,6 +714,52 @@ static noinstr void mce_read_aux(struct mce_hw_err *err, int i)
 
 DEFINE_PER_CPU(unsigned, mce_poll_count);
 
+/*
+ * Newer Intel systems that support software error
+ * recovery need to make additional checks. Other
+ * CPUs should skip over uncorrected errors, but log
+ * everything else.
+ */
+static bool ser_should_log_poll_error(struct mce *m)
+{
+	/* Log "not enabled" (speculative) errors */
+	if (!(m->status & MCI_STATUS_EN))
+		return true;
+
+	/*
+	 * Log UCNA (SDM: 15.6.3 "UCR Error Classification")
+	 * UC == 1 && PCC == 0 && S == 0
+	 */
+	if (!(m->status & MCI_STATUS_PCC) && !(m->status & MCI_STATUS_S))
+		return true;
+
+	return false;
+}
+
+static bool should_log_poll_error(enum mcp_flags flags, struct mce_hw_err *err)
+{
+	struct mce *m = &err->m;
+
+	/* If this entry is not valid, ignore it. */
+	if (!(m->status & MCI_STATUS_VAL))
+		return false;
+
+	/*
+	 * If we are logging everything (at CPU online) or this
+	 * is a corrected error, then we must log it.
+	 */
+	if ((flags & MCP_UC) || !(m->status & MCI_STATUS_UC))
+		return true;
+
+	if (mca_cfg.ser)
+		return ser_should_log_poll_error(m);
+
+	if (m->status & MCI_STATUS_UC)
+		return false;
+
+	return true;
+}
+
 /*
  * Poll for corrected events or events that happened before reset.
  * Those are just logged through /dev/mcelog.
@@ -765,48 +811,10 @@ void machine_check_poll(enum mcp_flags flags, mce_banks_t *b)
 		if (!mca_cfg.cmci_disabled)
 			mce_track_storm(m);
 
-		/* If this entry is not valid, ignore it */
-		if (!(m->status & MCI_STATUS_VAL))
+		/* Verify that the error should be logged based on hardware conditions. */
+		if (!should_log_poll_error(flags, &err))
 			continue;
 
-		/*
-		 * If we are logging everything (at CPU online) or this
-		 * is a corrected error, then we must log it.
-		 */
-		if ((flags & MCP_UC) || !(m->status & MCI_STATUS_UC))
-			goto log_it;
-
-		/*
-		 * Newer Intel systems that support software error
-		 * recovery need to make additional checks. Other
-		 * CPUs should skip over uncorrected errors, but log
-		 * everything else.
-		 */
-		if (!mca_cfg.ser) {
-			if (m->status & MCI_STATUS_UC)
-				continue;
-			goto log_it;
-		}
-
-		/* Log "not enabled" (speculative) errors */
-		if (!(m->status & MCI_STATUS_EN))
-			goto log_it;
-
-		/*
-		 * Log UCNA (SDM: 15.6.3 "UCR Error Classification")
-		 * UC == 1 && PCC == 0 && S == 0
-		 */
-		if (!(m->status & MCI_STATUS_PCC) && !(m->status & MCI_STATUS_S))
-			goto log_it;
-
-		/*
-		 * Skip anything else. Presumption is that our read of this
-		 * bank is racing with a machine check. Leave the log alone
-		 * for do_machine_check() to deal with it.
-		 */
-		continue;
-
-log_it:
 		mce_read_aux(&err, i);
 		m->severity = mce_severity(m, NULL, NULL, false);
 		/*

-- 
2.49.0

[PATCH v3 12/17] x86/mce: Unify AMD THR handler with MCA Polling

[Yazen Ghannam]

AMD systems optionally support an MCA thresholding interrupt. The
interrupt should be used as another signal to trigger MCA polling. This
is similar to how the Intel Corrected Machine Check interrupt (CMCI) is
handled.

AMD MCA thresholding is managed using the MCA_MISC registers within an
MCA bank. The OS will need to modify the hardware error count field in
order to reset the threshold limit and rearm the interrupt. Management
of the MCA_MISC register should be done as a follow up to the basic MCA
polling flow. It should not be the main focus of the interrupt handler.

Furthermore, future systems will have the ability to send an MCA
thresholding interrupt to the OS even when the OS does not manage the
feature, i.e. MCA_MISC registers are Read-as-Zero/Locked.

Call the common MCA polling function when handling the MCA thresholding
interrupt. This will allow the OS to find any valid errors whether or
not the MCA thresholding feature is OS-managed. Also, this allows the
common MCA polling options and kernel parameters to apply to AMD
systems.

Add a callback to the MCA polling function to check and reset any
threshold blocks that have reached their threshold limit.

Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-12-3636547fe05f@amd.com
    
    v2->v3:
    * Add tags from Qiuxu and Tony.
    
    v1->v2:
    * Start collecting per-CPU items in a struct.
    * Keep and use mce_flags.amd_threshold.

 arch/x86/kernel/cpu/mce/amd.c      | 49 ++++++++++++++++----------------------
 arch/x86/kernel/cpu/mce/core.c     |  3 +++
 arch/x86/kernel/cpu/mce/internal.h |  2 ++
 3 files changed, 26 insertions(+), 28 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 6a69cac36c18..f8755a21fd48 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -54,6 +54,12 @@
 
 static bool thresholding_irq_en;
 
+struct mce_amd_cpu_data {
+	mce_banks_t     thr_intr_banks;
+};
+
+static DEFINE_PER_CPU_READ_MOSTLY(struct mce_amd_cpu_data, mce_amd_data);
+
 static const char * const th_names[] = {
 	"load_store",
 	"insn_fetch",
@@ -559,6 +565,7 @@ prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
 	if (!b.interrupt_capable)
 		goto done;
 
+	__set_bit(bank, this_cpu_ptr(&mce_amd_data)->thr_intr_banks);
 	b.interrupt_enable = 1;
 
 	if (!mce_flags.smca) {
@@ -898,12 +905,7 @@ static void amd_deferred_error_interrupt(void)
 		log_error_deferred(bank);
 }
 
-static void log_error_thresholding(unsigned int bank, u64 misc)
-{
-	_log_error_deferred(bank, misc);
-}
-
-static void log_and_reset_block(struct threshold_block *block)
+static void reset_block(struct threshold_block *block)
 {
 	struct thresh_restart tr;
 	u32 low = 0, high = 0;
@@ -917,23 +919,14 @@ static void log_and_reset_block(struct threshold_block *block)
 	if (!(high & MASK_OVERFLOW_HI))
 		return;
 
-	/* Log the MCE which caused the threshold event. */
-	log_error_thresholding(block->bank, ((u64)high << 32) | low);
-
-	/* Reset threshold block after logging error. */
 	memset(&tr, 0, sizeof(tr));
 	tr.b = block;
 	threshold_restart_bank(&tr);
 }
 
-/*
- * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
- * goes off when error_count reaches threshold_limit.
- */
-static void amd_threshold_interrupt(void)
+void amd_reset_thr_limit(unsigned int bank)
 {
-	struct threshold_bank **bp = this_cpu_read(threshold_banks), *thr_bank;
-	unsigned int bank, cpu = smp_processor_id();
+	struct threshold_bank **bp = this_cpu_read(threshold_banks);
 	struct threshold_block *block, *tmp;
 
 	/*
@@ -941,20 +934,20 @@ static void amd_threshold_interrupt(void)
 	 * handler is installed at boot time, but on a hotplug event the
 	 * interrupt might fire before the data has been initialized.
 	 */
-	if (!bp)
+	if (!bp || !bp[bank])
 		return;
 
-	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
-		if (!(per_cpu(bank_map, cpu) & BIT_ULL(bank)))
-			continue;
-
-		thr_bank = bp[bank];
-		if (!thr_bank)
-			continue;
+	list_for_each_entry_safe(block, tmp, &bp[bank]->miscj, miscj)
+		reset_block(block);
+}
 
-		list_for_each_entry_safe(block, tmp, &thr_bank->miscj, miscj)
-			log_and_reset_block(block);
-	}
+/*
+ * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
+ * goes off when error_count reaches threshold_limit.
+ */
+static void amd_threshold_interrupt(void)
+{
+	machine_check_poll(MCP_TIMESTAMP, &this_cpu_ptr(&mce_amd_data)->thr_intr_banks);
 }
 
 /*
diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index c82c9e435066..de85b014653f 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -831,6 +831,9 @@ void machine_check_poll(enum mcp_flags flags, mce_banks_t *b)
 			mce_log(&err);
 
 clear_it:
+		if (mce_flags.amd_threshold)
+			amd_reset_thr_limit(i);
+
 		/*
 		 * Clear state for this bank.
 		 */
diff --git a/arch/x86/kernel/cpu/mce/internal.h b/arch/x86/kernel/cpu/mce/internal.h
index 87b69935d57d..aeb0a998f553 100644
--- a/arch/x86/kernel/cpu/mce/internal.h
+++ b/arch/x86/kernel/cpu/mce/internal.h
@@ -269,6 +269,7 @@ void mce_threshold_create_device(unsigned int cpu);
 void mce_threshold_remove_device(unsigned int cpu);
 extern bool amd_filter_mce(struct mce *m);
 bool amd_mce_usable_address(struct mce *m);
+void amd_reset_thr_limit(unsigned int bank);
 
 /*
  * If MCA_CONFIG[McaLsbInStatusSupported] is set, extract ErrAddr in bits
@@ -300,6 +301,7 @@ static inline void mce_threshold_create_device(unsigned int cpu)	{ }
 static inline void mce_threshold_remove_device(unsigned int cpu)	{ }
 static inline bool amd_filter_mce(struct mce *m) { return false; }
 static inline bool amd_mce_usable_address(struct mce *m) { return false; }
+static inline void amd_reset_thr_limit(unsigned int bank) { }
 static inline void smca_extract_err_addr(struct mce *m) { }
 static inline void mce_smca_cpu_init(void) {}
 #endif

-- 
2.49.0

[PATCH v3 13/17] x86/mce: Unify AMD DFR handler with MCA Polling

[Yazen Ghannam]

AMD systems optionally support a deferred error interrupt. The interrupt
should be used as another signal to trigger MCA polling. This is similar
to how other MCA interrupts are handled.

Deferred errors do not require any special handling related to the
interrupt, e.g. resetting or rearming the interrupt, etc.

However, Scalable MCA systems include a pair of registers, MCA_DESTAT
and MCA_DEADDR, that should be checked for valid errors. This check
should be done whenever MCA registers are polled. Currently, the
deferred error interrupt does this check, but the MCA polling function
does not.

Call the MCA polling function when handling the deferred error
interrupt. This keeps all "polling" cases in a common function.

Call the polling function only for banks that have the deferred error
interrupt enabled.

Add an SMCA status check helper. This will do the same status check and
register clearing that the interrupt handler has done. And it extends
the common polling flow to find AMD deferred errors.

Remove old code whose functionality is already covered in the common MCA
code.

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-13-3636547fe05f@amd.com
    
    v2->v3:
    * Add tags from Qiuxu and Tony.
    
    v1->v2:
    * Keep code comment.
    * Log directly from helper function rather than pass values.

 arch/x86/kernel/cpu/mce/amd.c  | 103 ++---------------------------------------
 arch/x86/kernel/cpu/mce/core.c |  60 +++++++++++++++++++++++-
 2 files changed, 64 insertions(+), 99 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index f8755a21fd48..62c4fe98d02a 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -56,6 +56,7 @@ static bool thresholding_irq_en;
 
 struct mce_amd_cpu_data {
 	mce_banks_t     thr_intr_banks;
+	mce_banks_t     dfr_intr_banks;
 };
 
 static DEFINE_PER_CPU_READ_MOSTLY(struct mce_amd_cpu_data, mce_amd_data);
@@ -300,8 +301,10 @@ static void smca_configure(unsigned int bank, unsigned int cpu)
 		 * APIC based interrupt. First, check that no interrupt has been
 		 * set.
 		 */
-		if ((low & BIT(5)) && !((high >> 5) & 0x3))
+		if ((low & BIT(5)) && !((high >> 5) & 0x3)) {
+			__set_bit(bank, this_cpu_ptr(&mce_amd_data)->dfr_intr_banks);
 			high |= BIT(5);
+		}
 
 		this_cpu_ptr(mce_banks_array)[bank].lsb_in_status = !!(low & BIT(8));
 
@@ -794,37 +797,6 @@ bool amd_mce_usable_address(struct mce *m)
 	return false;
 }
 
-static void __log_error(unsigned int bank, u64 status, u64 addr, u64 misc)
-{
-	struct mce_hw_err err;
-	struct mce *m = &err.m;
-
-	mce_prep_record(&err);
-
-	m->status = status;
-	m->misc   = misc;
-	m->bank   = bank;
-	m->tsc	 = rdtsc();
-
-	if (m->status & MCI_STATUS_ADDRV) {
-		m->addr = addr;
-
-		smca_extract_err_addr(m);
-	}
-
-	if (mce_flags.smca) {
-		rdmsrq(MSR_AMD64_SMCA_MCx_IPID(bank), m->ipid);
-
-		if (m->status & MCI_STATUS_SYNDV) {
-			rdmsrq(MSR_AMD64_SMCA_MCx_SYND(bank), m->synd);
-			rdmsrq(MSR_AMD64_SMCA_MCx_SYND1(bank), err.vendor.amd.synd1);
-			rdmsrq(MSR_AMD64_SMCA_MCx_SYND2(bank), err.vendor.amd.synd2);
-		}
-	}
-
-	mce_log(&err);
-}
-
 DEFINE_IDTENTRY_SYSVEC(sysvec_deferred_error)
 {
 	trace_deferred_error_apic_entry(DEFERRED_ERROR_VECTOR);
@@ -834,75 +806,10 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_deferred_error)
 	apic_eoi();
 }
 
-/*
- * Returns true if the logged error is deferred. False, otherwise.
- */
-static inline bool
-_log_error_bank(unsigned int bank, u32 msr_stat, u32 msr_addr, u64 misc)
-{
-	u64 status, addr = 0;
-
-	rdmsrq(msr_stat, status);
-	if (!(status & MCI_STATUS_VAL))
-		return false;
-
-	if (status & MCI_STATUS_ADDRV)
-		rdmsrq(msr_addr, addr);
-
-	__log_error(bank, status, addr, misc);
-
-	wrmsrq(msr_stat, 0);
-
-	return status & MCI_STATUS_DEFERRED;
-}
-
-static bool _log_error_deferred(unsigned int bank, u32 misc)
-{
-	if (!_log_error_bank(bank, mca_msr_reg(bank, MCA_STATUS),
-			     mca_msr_reg(bank, MCA_ADDR), misc))
-		return false;
-
-	/*
-	 * Non-SMCA systems don't have MCA_DESTAT/MCA_DEADDR registers.
-	 * Return true here to avoid accessing these registers.
-	 */
-	if (!mce_flags.smca)
-		return true;
-
-	/* Clear MCA_DESTAT if the deferred error was logged from MCA_STATUS. */
-	wrmsrq(MSR_AMD64_SMCA_MCx_DESTAT(bank), 0);
-	return true;
-}
-
-/*
- * We have three scenarios for checking for Deferred errors:
- *
- * 1) Non-SMCA systems check MCA_STATUS and log error if found.
- * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
- *    clear MCA_DESTAT.
- * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
- *    log it.
- */
-static void log_error_deferred(unsigned int bank)
-{
-	if (_log_error_deferred(bank, 0))
-		return;
-
-	/*
-	 * Only deferred errors are logged in MCA_DE{STAT,ADDR} so just check
-	 * for a valid error.
-	 */
-	_log_error_bank(bank, MSR_AMD64_SMCA_MCx_DESTAT(bank),
-			      MSR_AMD64_SMCA_MCx_DEADDR(bank), 0);
-}
-
 /* APIC interrupt handler for deferred errors */
 static void amd_deferred_error_interrupt(void)
 {
-	unsigned int bank;
-
-	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank)
-		log_error_deferred(bank);
+	machine_check_poll(MCP_TIMESTAMP, &this_cpu_ptr(&mce_amd_data)->dfr_intr_banks);
 }
 
 static void reset_block(struct threshold_block *block)
diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index de85b014653f..0a2a97681266 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -714,6 +714,61 @@ static noinstr void mce_read_aux(struct mce_hw_err *err, int i)
 
 DEFINE_PER_CPU(unsigned, mce_poll_count);
 
+/*
+ * We have three scenarios for checking for Deferred errors:
+ *
+ * 1) Non-SMCA systems check MCA_STATUS and log error if found.
+ * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
+ *    clear MCA_DESTAT.
+ * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
+ *    log it.
+ */
+static bool smca_should_log_poll_error(enum mcp_flags flags, struct mce_hw_err *err)
+{
+	struct mce *m = &err->m;
+
+	/*
+	 * If this is a deferred error found in MCA_STATUS, then clear
+	 * the redundant data from the MCA_DESTAT register.
+	 */
+	if (m->status & MCI_STATUS_VAL) {
+		if (m->status & MCI_STATUS_DEFERRED)
+			mce_wrmsrq(MSR_AMD64_SMCA_MCx_DESTAT(m->bank), 0);
+
+		return true;
+	}
+
+	/*
+	 * If the MCA_DESTAT register has valid data, then use
+	 * it as the status register.
+	 */
+	m->status = mce_rdmsrq(MSR_AMD64_SMCA_MCx_DESTAT(m->bank));
+
+	if (!(m->status & MCI_STATUS_VAL))
+		return false;
+
+	/*
+	 * Gather all relevant data now and log the record before clearing
+	 * the deferred status register. This avoids needing to go back to
+	 * the polling function for these actions.
+	 */
+	mce_read_aux(err, m->bank);
+
+	if (m->status & MCI_STATUS_ADDRV)
+		m->addr = mce_rdmsrq(MSR_AMD64_SMCA_MCx_DEADDR(m->bank));
+
+	smca_extract_err_addr(m);
+	m->severity = mce_severity(m, NULL, NULL, false);
+
+	if (flags & MCP_QUEUE_LOG)
+		mce_gen_pool_add(err);
+	else
+		mce_log(err);
+
+	mce_wrmsrq(MSR_AMD64_SMCA_MCx_DESTAT(m->bank), 0);
+	return false;
+}
+
 /*
  * Newer Intel systems that support software error
  * recovery need to make additional checks. Other
@@ -740,6 +795,9 @@ static bool should_log_poll_error(enum mcp_flags flags, struct mce_hw_err *err)
 {
 	struct mce *m = &err->m;
 
+	if (mce_flags.smca)
+		return smca_should_log_poll_error(flags, err);
+
 	/* If this entry is not valid, ignore it. */
 	if (!(m->status & MCI_STATUS_VAL))
 		return false;
@@ -2226,7 +2284,7 @@ void cpu_mca_init(struct cpuinfo_x86 *c)
 	if (mce_flags.smca)
 		mce_smca_cpu_init();
 
-	rdmsrl(MSR_IA32_MCG_CAP, cap);
+	rdmsrq(MSR_IA32_MCG_CAP, cap);
 
 	/* Use accurate RIP reporting if available. */
 	if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9)

-- 
2.49.0

Re: [PATCH v3 13/17] x86/mce: Unify AMD DFR handler with MCA Polling

[Borislav Petkov]

On Tue, Apr 15, 2025 at 02:55:08PM +0000, Yazen Ghannam wrote:
> +static bool smca_should_log_poll_error(enum mcp_flags flags, struct mce_hw_err *err)
> +{
> +	struct mce *m = &err->m;
> +
> +	/*
> +	 * If this is a deferred error found in MCA_STATUS, then clear
> +	 * the redundant data from the MCA_DESTAT register.
> +	 */
> +	if (m->status & MCI_STATUS_VAL) {
> +		if (m->status & MCI_STATUS_DEFERRED)
> +			mce_wrmsrq(MSR_AMD64_SMCA_MCx_DESTAT(m->bank), 0);
> +
> +		return true;
> +	}
> +
> +	/*
> +	 * If the MCA_DESTAT register has valid data, then use
> +	 * it as the status register.
> +	 */
> +	m->status = mce_rdmsrq(MSR_AMD64_SMCA_MCx_DESTAT(m->bank));
> +

Superfluous newline.

> +	if (!(m->status & MCI_STATUS_VAL))
> +		return false;
> +
> +	/*
> +	 * Gather all relevant data now and log the record before clearing
> +	 * the deferred status register. This avoids needing to go back to
> +	 * the polling function for these actions.
> +	 */
> +	mce_read_aux(err, m->bank);
> +
> +	if (m->status & MCI_STATUS_ADDRV)
> +		m->addr = mce_rdmsrq(MSR_AMD64_SMCA_MCx_DEADDR(m->bank));
> +
> +	smca_extract_err_addr(m);
> +	m->severity = mce_severity(m, NULL, NULL, false);
> +
> +	if (flags & MCP_QUEUE_LOG)
> +		mce_gen_pool_add(err);
> +	else
> +		mce_log(err);

Except you have a function which is called "should log" which also does the
logging.

And you have that same logging being done in machine_check_poll().

This code needs more designing.

Thx.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 13/17] x86/mce: Unify AMD DFR handler with MCA Polling

[Yazen Ghannam]

On Wed, May 07, 2025 at 11:20:42AM +0200, Borislav Petkov wrote:
> On Tue, Apr 15, 2025 at 02:55:08PM +0000, Yazen Ghannam wrote:
> > +static bool smca_should_log_poll_error(enum mcp_flags flags, struct mce_hw_err *err)
> > +{
> > +	struct mce *m = &err->m;
> > +
> > +	/*
> > +	 * If this is a deferred error found in MCA_STATUS, then clear
> > +	 * the redundant data from the MCA_DESTAT register.
> > +	 */
> > +	if (m->status & MCI_STATUS_VAL) {
> > +		if (m->status & MCI_STATUS_DEFERRED)
> > +			mce_wrmsrq(MSR_AMD64_SMCA_MCx_DESTAT(m->bank), 0);
> > +
> > +		return true;
> > +	}
> > +
> > +	/*
> > +	 * If the MCA_DESTAT register has valid data, then use
> > +	 * it as the status register.
> > +	 */
> > +	m->status = mce_rdmsrq(MSR_AMD64_SMCA_MCx_DESTAT(m->bank));
> > +
> 
> Superfluous newline.
> 

Ack.

> > +	if (!(m->status & MCI_STATUS_VAL))
> > +		return false;
> > +
> > +	/*
> > +	 * Gather all relevant data now and log the record before clearing
> > +	 * the deferred status register. This avoids needing to go back to
> > +	 * the polling function for these actions.
> > +	 */
> > +	mce_read_aux(err, m->bank);
> > +
> > +	if (m->status & MCI_STATUS_ADDRV)
> > +		m->addr = mce_rdmsrq(MSR_AMD64_SMCA_MCx_DEADDR(m->bank));
> > +
> > +	smca_extract_err_addr(m);
> > +	m->severity = mce_severity(m, NULL, NULL, false);
> > +
> > +	if (flags & MCP_QUEUE_LOG)
> > +		mce_gen_pool_add(err);
> > +	else
> > +		mce_log(err);
> 
> Except you have a function which is called "should log" which also does the
> logging.
> 
> And you have that same logging being done in machine_check_poll().
> 
> This code needs more designing.
> 

Okay, will do.

Thanks,
Yazen

[PATCH v3 14/17] x86/mce/amd: Enable interrupt vectors once per-CPU on SMCA systems

[Yazen Ghannam]

Scalable MCA systems have a per-CPU register that gives the APIC LVT
offset for the thresholding and deferred error interrupts.

Currently, this register is read once to set up the deferred error
interrupt and then read again for each thresholding block. Furthermore,
the APIC LVT registers are configured each time, but they only need to
be configured once per-CPU.

Move the APIC LVT setup to the early part of CPU init, so that the
registers are set up once. Also, this ensures that the kernel is ready
to service the interrupts before the individual error sources (each MCA
bank) are enabled.

Apply this change only to SMCA systems to avoid breaking any legacy
behavior. The deferred error interrupt is technically advertised by the
SUCCOR feature. However, this was first made available on SMCA systems.
Therefore, only set up the deferred error interrupt on SMCA systems and
simplify the code.

Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-14-3636547fe05f@amd.com
    
    v2->v3:
    * Add tags from Tony.
    
    v1->v2:
    * Use new per-CPU struct.
    * Don't set up interrupt vectors.

 arch/x86/kernel/cpu/mce/amd.c | 113 ++++++++++++++++++------------------------
 1 file changed, 48 insertions(+), 65 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 62c4fe98d02a..9e226bdbdc40 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -43,9 +43,6 @@
 /* Deferred error settings */
 #define MSR_CU_DEF_ERR		0xC0000410
 #define MASK_DEF_LVTOFF		0x000000F0
-#define MASK_DEF_INT_TYPE	0x00000006
-#define DEF_LVT_OFF		0x2
-#define DEF_INT_TYPE_APIC	0x2
 
 /* Scalable MCA: */
 
@@ -57,6 +54,8 @@ static bool thresholding_irq_en;
 struct mce_amd_cpu_data {
 	mce_banks_t     thr_intr_banks;
 	mce_banks_t     dfr_intr_banks;
+	bool		thr_intr_en;
+	bool		dfr_intr_en;
 };
 
 static DEFINE_PER_CPU_READ_MOSTLY(struct mce_amd_cpu_data, mce_amd_data);
@@ -271,6 +270,7 @@ void (*deferred_error_int_vector)(void) = default_deferred_error_interrupt;
 
 static void smca_configure(unsigned int bank, unsigned int cpu)
 {
+	struct mce_amd_cpu_data *data = this_cpu_ptr(&mce_amd_data);
 	u8 *bank_counts = this_cpu_ptr(smca_bank_counts);
 	const struct smca_hwid *s_hwid;
 	unsigned int i, hwid_mcatype;
@@ -301,8 +301,8 @@ static void smca_configure(unsigned int bank, unsigned int cpu)
 		 * APIC based interrupt. First, check that no interrupt has been
 		 * set.
 		 */
-		if ((low & BIT(5)) && !((high >> 5) & 0x3)) {
-			__set_bit(bank, this_cpu_ptr(&mce_amd_data)->dfr_intr_banks);
+		if ((low & BIT(5)) && !((high >> 5) & 0x3) && data->dfr_intr_en) {
+			__set_bit(bank, data->dfr_intr_banks);
 			high |= BIT(5);
 		}
 
@@ -378,6 +378,14 @@ static bool lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
 {
 	int msr = (hi & MASK_LVTOFF_HI) >> 20;
 
+	/*
+	 * On SMCA CPUs, LVT offset is programmed at a different MSR, and
+	 * the BIOS provides the value. The original field where LVT offset
+	 * was set is reserved. Return early here:
+	 */
+	if (mce_flags.smca)
+		return false;
+
 	if (apic < 0) {
 		pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
@@ -386,14 +394,6 @@ static bool lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
 	}
 
 	if (apic != msr) {
-		/*
-		 * On SMCA CPUs, LVT offset is programmed at a different MSR, and
-		 * the BIOS provides the value. The original field where LVT offset
-		 * was set is reserved. Return early here:
-		 */
-		if (mce_flags.smca)
-			return false;
-
 		pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
 		       b->cpu, apic, b->bank, b->block, b->address, hi, lo);
@@ -474,41 +474,6 @@ static int setup_APIC_mce_threshold(int reserved, int new)
 	return reserved;
 }
 
-static int setup_APIC_deferred_error(int reserved, int new)
-{
-	if (reserved < 0 && !setup_APIC_eilvt(new, DEFERRED_ERROR_VECTOR,
-					      APIC_EILVT_MSG_FIX, 0))
-		return new;
-
-	return reserved;
-}
-
-static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
-{
-	u32 low = 0, high = 0;
-	int def_offset = -1, def_new;
-
-	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
-		return;
-
-	def_new = (low & MASK_DEF_LVTOFF) >> 4;
-	if (!(low & MASK_DEF_LVTOFF)) {
-		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
-		def_new = DEF_LVT_OFF;
-		low = (low & ~MASK_DEF_LVTOFF) | (DEF_LVT_OFF << 4);
-	}
-
-	def_offset = setup_APIC_deferred_error(def_offset, def_new);
-	if ((def_offset == def_new) &&
-	    (deferred_error_int_vector != amd_deferred_error_interrupt))
-		deferred_error_int_vector = amd_deferred_error_interrupt;
-
-	if (!mce_flags.smca)
-		low = (low & ~MASK_DEF_INT_TYPE) | DEF_INT_TYPE_APIC;
-
-	wrmsr(MSR_CU_DEF_ERR, low, high);
-}
-
 static u32 smca_get_block_address(unsigned int bank, unsigned int block,
 				  unsigned int cpu)
 {
@@ -551,7 +516,6 @@ prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
 			int offset, u32 misc_high)
 {
 	unsigned int cpu = smp_processor_id();
-	u32 smca_low, smca_high;
 	struct threshold_block b;
 	int new;
 
@@ -571,18 +535,10 @@ prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
 	__set_bit(bank, this_cpu_ptr(&mce_amd_data)->thr_intr_banks);
 	b.interrupt_enable = 1;
 
-	if (!mce_flags.smca) {
-		new = (misc_high & MASK_LVTOFF_HI) >> 20;
-		goto set_offset;
-	}
-
-	/* Gather LVT offset for thresholding: */
-	if (rdmsr_safe(MSR_CU_DEF_ERR, &smca_low, &smca_high))
-		goto out;
-
-	new = (smca_low & SMCA_THR_LVT_OFF) >> 12;
+	if (mce_flags.smca)
+		goto done;
 
-set_offset:
+	new = (misc_high & MASK_LVTOFF_HI) >> 20;
 	offset = setup_APIC_mce_threshold(offset, new);
 	if (offset == new)
 		thresholding_irq_en = true;
@@ -590,7 +546,6 @@ prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
 done:
 	mce_threshold_block_init(&b, offset);
 
-out:
 	return offset;
 }
 
@@ -659,6 +614,32 @@ static void disable_err_thresholding(struct cpuinfo_x86 *c, unsigned int bank)
 		wrmsrq(MSR_K7_HWCR, hwcr);
 }
 
+/*
+ * Enable the APIC LVT interrupt vectors once per-CPU. This should be done before hardware is
+ * ready to send interrupts.
+ *
+ * Individual error sources are enabled later during per-bank init.
+ */
+static void smca_enable_interrupt_vectors(void)
+{
+	struct mce_amd_cpu_data *data = this_cpu_ptr(&mce_amd_data);
+	u64 mca_intr_cfg, offset;
+
+	if (!mce_flags.smca || !mce_flags.succor)
+		return;
+
+	if (rdmsrq_safe(MSR_CU_DEF_ERR, &mca_intr_cfg))
+		return;
+
+	offset = (mca_intr_cfg & SMCA_THR_LVT_OFF) >> 12;
+	if (!setup_APIC_eilvt(offset, THRESHOLD_APIC_VECTOR, APIC_EILVT_MSG_FIX, 0))
+		data->thr_intr_en = true;
+
+	offset = (mca_intr_cfg & MASK_DEF_LVTOFF) >> 4;
+	if (!setup_APIC_eilvt(offset, DEFERRED_ERROR_VECTOR, APIC_EILVT_MSG_FIX, 0))
+		data->dfr_intr_en = true;
+}
+
 static void amd_apply_quirks(struct cpuinfo_x86 *c)
 {
 	struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
@@ -690,11 +671,16 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 
 	amd_apply_quirks(c);
 	mce_flags.amd_threshold	 = 1;
+	smca_enable_interrupt_vectors();
 
 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
-		if (mce_flags.smca)
+		if (mce_flags.smca) {
 			smca_configure(bank, cpu);
 
+			if (!this_cpu_ptr(&mce_amd_data)->thr_intr_en)
+				continue;
+		}
+
 		disable_err_thresholding(c, bank);
 
 		for (block = 0; block < NR_BLOCKS; ++block) {
@@ -715,9 +701,6 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 			offset = prepare_threshold_block(bank, block, address, offset, high);
 		}
 	}
-
-	if (mce_flags.succor)
-		deferred_error_interrupt_enable(c);
 }
 
 void mce_smca_cpu_init(void)

-- 
2.49.0

Re: [PATCH v3 14/17] x86/mce/amd: Enable interrupt vectors once per-CPU on SMCA systems

[Borislav Petkov]

On Tue, Apr 15, 2025 at 02:55:09PM +0000, Yazen Ghannam wrote:
> -static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
> -{
> -	u32 low = 0, high = 0;
> -	int def_offset = -1, def_new;
> -
> -	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
> -		return;
> -
> -	def_new = (low & MASK_DEF_LVTOFF) >> 4;
> -	if (!(low & MASK_DEF_LVTOFF)) {
> -		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");

I'm not sure why it is ok to remove that one.

/me goes and digs into lore...

Here's why we did it back then:

https://lore.kernel.org/all/5547906E.3060701@amd.com/

and apparently that was for some bulldozer BIOS.

How can we trust Zen BIOS all of a sudden?

;-)

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 14/17] x86/mce/amd: Enable interrupt vectors once per-CPU on SMCA systems

[Yazen Ghannam]

On Wed, May 07, 2025 at 09:35:39PM +0200, Borislav Petkov wrote:
> On Tue, Apr 15, 2025 at 02:55:09PM +0000, Yazen Ghannam wrote:
> > -static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
> > -{
> > -	u32 low = 0, high = 0;
> > -	int def_offset = -1, def_new;
> > -
> > -	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
> > -		return;
> > -
> > -	def_new = (low & MASK_DEF_LVTOFF) >> 4;
> > -	if (!(low & MASK_DEF_LVTOFF)) {
> > -		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
> 
> I'm not sure why it is ok to remove that one.
> 
> /me goes and digs into lore...
> 
> Here's why we did it back then:
> 
> https://lore.kernel.org/all/5547906E.3060701@amd.com/
> 
> and apparently that was for some bulldozer BIOS.
> 
> How can we trust Zen BIOS all of a sudden?
> 
> ;-)
> 

Let me flip it around. Why is this check needed at all? Was there ever a
real issue to resolve? It seems to me the deferred error updates are
just following what other code did.

I figure the reason to have the platform give the offset to the OS is so
the OS doesn't hard code it (in case it needs to change). These offsets
were hard coded in the past (conflict between IBS/THR), and it caused
problems when the offsets switched in the hardware. The registers that
give the offsets were introduced soon after, I think.

So the checks we do are defeating the purpose. The OS is still hard
coding the offsets. The goal of this change is to follow the intent of
the design. Sometimes we need to let go and trust [the BIOS]. ;)

Now we could update the checks to verify that an offset is not used for
multiple interrupt sources.

Let's follow up with the design folks to be sure.

Thanks,
Yazen

Re: [PATCH v3 14/17] x86/mce/amd: Enable interrupt vectors once per-CPU on SMCA systems

[Borislav Petkov]

On Thu, May 08, 2025 at 11:53:00AM -0400, Yazen Ghannam wrote:
> Let me flip it around. Why is this check needed at all?

As I said above, some BIOS f*ckup.

> Was there ever a real issue to resolve?

Not that I remember...

> It seems to me the deferred error updates are just following what other code
> did.

Let's search the web for it:

* https://bbs.archlinux.org/viewtopic.php?id=299379

- silly guests, who cares

* https://gitlab.com/qemu-project/qemu/-/issues/2571

- another misguided qemu...

Aha:

https://lore.kernel.org/lkml/20241219124426.325747-1-pbonzini@redhat.com

the usual virt silly stuff.

> I figure the reason to have the platform give the offset to the OS is so
> the OS doesn't hard code it (in case it needs to change). These offsets
> were hard coded in the past (conflict between IBS/THR), and it caused
> problems when the offsets switched in the hardware. The registers that
> give the offsets were introduced soon after, I think.

Right.

> So the checks we do are defeating the purpose. The OS is still hard
> coding the offsets. The goal of this change is to follow the intent of
> the design. Sometimes we need to let go and trust [the BIOS]. ;)

Look at you being silly :-P

> Now we could update the checks to verify that an offset is not used for
> multiple interrupt sources.

... or, we won't do anything until someone in BIOS f*cks up again.
 
> Let's follow up with the design folks to be sure.

Yah, sounds like we will have to verify them after all. You can see how
universally widespread the trust in BIOS is...

:-P

In any case, whatever you do, when you axe off stuff, write in the commit
message why you do so. Silently removing it is making me want to know why it
is ok now.

Thx.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 14/17] x86/mce/amd: Enable interrupt vectors once per-CPU on SMCA systems

[Yazen Ghannam]

On Fri, May 09, 2025 at 04:08:21PM +0200, Borislav Petkov wrote:
> On Thu, May 08, 2025 at 11:53:00AM -0400, Yazen Ghannam wrote:
> > Let me flip it around. Why is this check needed at all?
> 
> As I said above, some BIOS f*ckup.
> 
> > Was there ever a real issue to resolve?
> 
> Not that I remember...
> 
> > It seems to me the deferred error updates are just following what other code
> > did.
> 
> Let's search the web for it:
> 
> * https://bbs.archlinux.org/viewtopic.php?id=299379
> 
> - silly guests, who cares
> 
> * https://gitlab.com/qemu-project/qemu/-/issues/2571
> 
> - another misguided qemu...
> 
> Aha:
> 
> https://lore.kernel.org/lkml/20241219124426.325747-1-pbonzini@redhat.com
> 
> the usual virt silly stuff.
> 
> > I figure the reason to have the platform give the offset to the OS is so
> > the OS doesn't hard code it (in case it needs to change). These offsets
> > were hard coded in the past (conflict between IBS/THR), and it caused
> > problems when the offsets switched in the hardware. The registers that
> > give the offsets were introduced soon after, I think.
> 
> Right.
> 
> > So the checks we do are defeating the purpose. The OS is still hard
> > coding the offsets. The goal of this change is to follow the intent of
> > the design. Sometimes we need to let go and trust [the BIOS]. ;)
> 
> Look at you being silly :-P
> 
> > Now we could update the checks to verify that an offset is not used for
> > multiple interrupt sources.
> 
> ... or, we won't do anything until someone in BIOS f*cks up again.
>  
> > Let's follow up with the design folks to be sure.
> 
> Yah, sounds like we will have to verify them after all. You can see how
> universally widespread the trust in BIOS is...
> 
> :-P
> 
> In any case, whatever you do, when you axe off stuff, write in the commit
> message why you do so. Silently removing it is making me want to know why it
> is ok now.
> 

Right, it sounds like we should take the values from the platform and
just make sure they aren't used for multiple sources. In other words, we
don't hard code the offsets, and we verify that each source has a unique
offset.

I agree we can leave this for now. So I'll drop this part from the patch.

I think this topic can be a separate set, and it should cover all APIC
LVT sources including IBS. I'll add it to the to-do list. :)

Thanks,
Yazen

[PATCH v3 15/17] x86/mce/amd: Support SMCA Corrected Error Interrupt

[Yazen Ghannam]

AMD systems optionally support MCA thresholding which provides the
ability for hardware to send an interrupt when a set error threshold is
reached. This feature counts errors of all severities, but it is
commonly used to report correctable errors with an interrupt rather than
polling.

Scalable MCA systems allow the Platform to take control of this feature.
In this case, the OS will not see the feature configuration and control
bits in the MCA_MISC* registers. The OS will not receive the MCA
thresholding interrupt, and it will need to poll for correctable errors.

A "corrected error interrupt" will be available on Scalable MCA systems.
This will be used in the same configuration where the Platform controls
MCA thresholding. However, the Platform will now be able to send the
MCA thresholding interrupt to the OS.

Check for the feature bit in the MCA_CONFIG register and confirm that
the MCA thresholding interrupt handler is already enabled. If successful,
set the feature enable bit in the MCA_CONFIG register to indicate to the
Platform that the OS is ready for the interrupt.

Tested-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-15-3636547fe05f@amd.com
    
    v2->v3:
    * Add tags from Tony.
    
    v1->v2:
    * Use new per-CPU struct.

 arch/x86/kernel/cpu/mce/amd.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index 9e226bdbdc40..d76a64c47a6d 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -306,6 +306,11 @@ static void smca_configure(unsigned int bank, unsigned int cpu)
 			high |= BIT(5);
 		}
 
+		if ((low & BIT(10)) && data->thr_intr_en) {
+			__set_bit(bank, data->thr_intr_banks);
+			high |= BIT(8);
+		}
+
 		this_cpu_ptr(mce_banks_array)[bank].lsb_in_status = !!(low & BIT(8));
 
 		wrmsr(smca_config, low, high);

-- 
2.49.0

Re: [PATCH v3 15/17] x86/mce/amd: Support SMCA Corrected Error Interrupt

[Borislav Petkov]

On Tue, Apr 15, 2025 at 02:55:10PM +0000, Yazen Ghannam wrote:
> @@ -306,6 +306,11 @@ static void smca_configure(unsigned int bank, unsigned int cpu)
>  			high |= BIT(5);
>  		}

Yeah, the above statements explain in comments what they do so that we don't
have to define the bits but use them straight "naked" with the BIT macro.
I think you'd need to put something along the lines of that text...

> Check for the feature bit in the MCA_CONFIG register and confirm that
> the MCA thresholding interrupt handler is already enabled. If successful,
> set the feature enable bit in the MCA_CONFIG register to indicate to the
> Platform that the OS is ready for the interrupt.

... here.

<---

> +		if ((low & BIT(10)) && data->thr_intr_en) {
> +			__set_bit(bank, data->thr_intr_banks);
> +			high |= BIT(8);
> +		}
> +
>  		this_cpu_ptr(mce_banks_array)[bank].lsb_in_status = !!(low & BIT(8));
>  
>  		wrmsr(smca_config, low, high);
> 
> -- 
> 2.49.0
> 

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 15/17] x86/mce/amd: Support SMCA Corrected Error Interrupt

[Yazen Ghannam]

On Fri, May 09, 2025 at 09:37:21PM +0200, Borislav Petkov wrote:
> On Tue, Apr 15, 2025 at 02:55:10PM +0000, Yazen Ghannam wrote:
> > @@ -306,6 +306,11 @@ static void smca_configure(unsigned int bank, unsigned int cpu)
> >  			high |= BIT(5);
> >  		}
> 
> Yeah, the above statements explain in comments what they do so that we don't
> have to define the bits but use them straight "naked" with the BIT macro.
> I think you'd need to put something along the lines of that text...
> 
> > Check for the feature bit in the MCA_CONFIG register and confirm that
> > the MCA thresholding interrupt handler is already enabled. If successful,
> > set the feature enable bit in the MCA_CONFIG register to indicate to the
> > Platform that the OS is ready for the interrupt.
> 
> ... here.
> 
> <---
> 

Okay, will do.

Thanks,
Yazen

[PATCH v3 16/17] x86/mce: Handle AMD threshold interrupt storms

[Yazen Ghannam]

From: Smita Koralahalli <Smita.KoralahalliChannabasappa@amd.com>

Extend the logic of handling CMCI storms to AMD threshold interrupts.

Rely on the similar approach as of Intel's CMCI to mitigate storms per
CPU and per bank. But, unlike CMCI, do not set thresholds and reduce
interrupt rate on a storm. Rather, disable the interrupt on the
corresponding CPU and bank. Re-enable back the interrupts if enough
consecutive polls of the bank show no corrected errors (30, as
programmed by Intel).

Turning off the threshold interrupts would be a better solution on AMD
systems as other error severities will still be handled even if the
threshold interrupts are disabled.

[Tony: Small tweak because mce_handle_storm() isn't a pointer now]
[Yazen: Rebase and simplify]

Reviewed-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
Signed-off-by: Smita Koralahalli <Smita.KoralahalliChannabasappa@amd.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---

Notes:
    Link:
    https://lore.kernel.org/r/20250213-wip-mca-updates-v2-16-3636547fe05f@amd.com
    
    v2->v3:
    * Add tag from Qiuxu.
    
    v1->v2:
    * New in v2, but based on older patch.
    * Rebased on current set and simplified.
    * Kept old tags.

 arch/x86/kernel/cpu/mce/amd.c       | 18 ++++++++++++++++++
 arch/x86/kernel/cpu/mce/internal.h  |  2 ++
 arch/x86/kernel/cpu/mce/threshold.c |  3 +++
 3 files changed, 23 insertions(+)

diff --git a/arch/x86/kernel/cpu/mce/amd.c b/arch/x86/kernel/cpu/mce/amd.c
index d76a64c47a6d..93f6cececad4 100644
--- a/arch/x86/kernel/cpu/mce/amd.c
+++ b/arch/x86/kernel/cpu/mce/amd.c
@@ -1218,3 +1218,21 @@ void mce_threshold_create_device(unsigned int cpu)
 		mce_threshold_vector = amd_threshold_interrupt;
 	return;
 }
+
+void mce_amd_handle_storm(unsigned int bank, bool on)
+{
+	struct threshold_bank **thr_banks = this_cpu_read(threshold_banks);
+	struct threshold_block *block, *tmp;
+	struct thresh_restart tr;
+
+	if (!thr_banks || !thr_banks[bank])
+		return;
+
+	memset(&tr, 0, sizeof(tr));
+
+	list_for_each_entry_safe(block, tmp, &thr_banks[bank]->miscj, miscj) {
+		tr.b = block;
+		tr.b->interrupt_enable = on;
+		threshold_restart_bank(&tr);
+	}
+}
diff --git a/arch/x86/kernel/cpu/mce/internal.h b/arch/x86/kernel/cpu/mce/internal.h
index aeb0a998f553..0dd77fa18d06 100644
--- a/arch/x86/kernel/cpu/mce/internal.h
+++ b/arch/x86/kernel/cpu/mce/internal.h
@@ -267,6 +267,7 @@ void mce_prep_record_per_cpu(unsigned int cpu, struct mce *m);
 #ifdef CONFIG_X86_MCE_AMD
 void mce_threshold_create_device(unsigned int cpu);
 void mce_threshold_remove_device(unsigned int cpu);
+void mce_amd_handle_storm(unsigned int bank, bool on);
 extern bool amd_filter_mce(struct mce *m);
 bool amd_mce_usable_address(struct mce *m);
 void amd_reset_thr_limit(unsigned int bank);
@@ -299,6 +300,7 @@ void mce_smca_cpu_init(void);
 #else
 static inline void mce_threshold_create_device(unsigned int cpu)	{ }
 static inline void mce_threshold_remove_device(unsigned int cpu)	{ }
+static inline void mce_amd_handle_storm(unsigned int bank, bool on)	{ }
 static inline bool amd_filter_mce(struct mce *m) { return false; }
 static inline bool amd_mce_usable_address(struct mce *m) { return false; }
 static inline void amd_reset_thr_limit(unsigned int bank) { }
diff --git a/arch/x86/kernel/cpu/mce/threshold.c b/arch/x86/kernel/cpu/mce/threshold.c
index f4a007616468..45144598ec74 100644
--- a/arch/x86/kernel/cpu/mce/threshold.c
+++ b/arch/x86/kernel/cpu/mce/threshold.c
@@ -63,6 +63,9 @@ static void mce_handle_storm(unsigned int bank, bool on)
 	case X86_VENDOR_INTEL:
 		mce_intel_handle_storm(bank, on);
 		break;
+	case X86_VENDOR_AMD:
+		mce_amd_handle_storm(bank, on);
+		break;
 	}
 }
 

-- 
2.49.0

[PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Yazen Ghannam]

Users can disable MCA polling by setting the "ignore_ce" parameter or by
setting "check_interval=0". This tells the kernel to *not* start the MCE
timer on a CPU.

During a CMCI storm, the MCE timer will be started with a fixed
interval. After the storm subsides, the timer's next interval is set to
check_interval.

This disregards the user's input through "ignore_ce" and
"check_interval". Furthermore, if "check_interval=0", then the new timer
will run faster than expected.

Create a new helper to check these conditions and use it when a CMCI
storm ends.

Fixes: 7eae17c4add5 ("x86/mce: Add per-bank CMCI storm mitigation")
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
Cc: stable@vger.kernel.org
---

Notes:
    v2->v3:
    * New in v3.

 arch/x86/kernel/cpu/mce/core.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index 0a2a97681266..131015f5eadc 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -1806,6 +1806,11 @@ static void mc_poll_banks_default(void)
 
 void (*mc_poll_banks)(void) = mc_poll_banks_default;
 
+static bool should_enable_timer(unsigned long iv)
+{
+	return !mca_cfg.ignore_ce && iv;
+}
+
 static void mce_timer_fn(struct timer_list *t)
 {
 	struct timer_list *cpu_t = this_cpu_ptr(&mce_timer);
@@ -1829,7 +1834,7 @@ static void mce_timer_fn(struct timer_list *t)
 
 	if (mce_get_storm_mode()) {
 		__start_timer(t, HZ);
-	} else {
+	} else if (should_enable_timer(iv)) {
 		__this_cpu_write(mce_next_interval, iv);
 		__start_timer(t, iv);
 	}
@@ -2142,7 +2147,7 @@ static void mce_start_timer(struct timer_list *t)
 {
 	unsigned long iv = check_interval * HZ;
 
-	if (mca_cfg.ignore_ce || !iv)
+	if (!should_enable_timer(iv))
 		return;
 
 	this_cpu_write(mce_next_interval, iv);

-- 
2.49.0

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Borislav Petkov]

On Tue, Apr 15, 2025 at 02:55:12PM +0000, Yazen Ghannam wrote:
> Users can disable MCA polling by setting the "ignore_ce" parameter or by
> setting "check_interval=0". This tells the kernel to *not* start the MCE
> timer on a CPU.
> 
> During a CMCI storm, the MCE timer will be started with a fixed
> interval.

Why?

If a user doesn't want CEs, why are we even bothering with CMCI storms?

Might as well disable the storm handling code altogether...

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Yazen Ghannam]

On Mon, May 12, 2025 at 09:46:16AM +0200, Borislav Petkov wrote:
> On Tue, Apr 15, 2025 at 02:55:12PM +0000, Yazen Ghannam wrote:
> > Users can disable MCA polling by setting the "ignore_ce" parameter or by
> > setting "check_interval=0". This tells the kernel to *not* start the MCE
> > timer on a CPU.
> > 
> > During a CMCI storm, the MCE timer will be started with a fixed
> > interval.
> 
> Why?
> 
> If a user doesn't want CEs, why are we even bothering with CMCI storms?
> 
> Might as well disable the storm handling code altogether...
> 

The use case is "disable MCA polling". I just gave two examples of how
this can be done.

We can focus on "check_interval=0". The user wants to disable MCA
polling and rely only on interrupts. They still want to see the CEs.

If an interrupt storm happens, the kernel will switch to polling with
our MCE timer.

When the storm ends, the kernel should go back to how things were before
the storm. If there was a timer before, then go back to the old
interval. If there was *not* a timer before, then delete/remove the
timer.

Thanks,
Yazen

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Luck, Tony]

On Mon, May 12, 2025 at 11:43:15AM -0400, Yazen Ghannam wrote:
> On Mon, May 12, 2025 at 09:46:16AM +0200, Borislav Petkov wrote:
> > On Tue, Apr 15, 2025 at 02:55:12PM +0000, Yazen Ghannam wrote:
> > > Users can disable MCA polling by setting the "ignore_ce" parameter or by
> > > setting "check_interval=0". This tells the kernel to *not* start the MCE
> > > timer on a CPU.
> > > 
> > > During a CMCI storm, the MCE timer will be started with a fixed
> > > interval.

I think you just need some more words at the start of this second
paragraph to avoid confusion when reading together with the previous
one.

Perhaps:

If the user did not disable CMCI, then storms can occur. When these
happen the MCE timer will be started ...

-Tony

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Yazen Ghannam]

On Mon, May 12, 2025 at 08:53:08AM -0700, Luck, Tony wrote:
> On Mon, May 12, 2025 at 11:43:15AM -0400, Yazen Ghannam wrote:
> > On Mon, May 12, 2025 at 09:46:16AM +0200, Borislav Petkov wrote:
> > > On Tue, Apr 15, 2025 at 02:55:12PM +0000, Yazen Ghannam wrote:
> > > > Users can disable MCA polling by setting the "ignore_ce" parameter or by
> > > > setting "check_interval=0". This tells the kernel to *not* start the MCE
> > > > timer on a CPU.
> > > > 
> > > > During a CMCI storm, the MCE timer will be started with a fixed
> > > > interval.
> 
> I think you just need some more words at the start of this second
> paragraph to avoid confusion when reading together with the previous
> one.
> 
> Perhaps:
> 
> If the user did not disable CMCI, then storms can occur. When these
> happen the MCE timer will be started ...
> 

Yes, good point. I'll update the commit message.

Thanks,
Yazen

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Borislav Petkov]

On Mon, May 12, 2025 at 11:43:15AM -0400, Yazen Ghannam wrote:
> The use case is "disable MCA polling". I just gave two examples of how
> this can be done.

Our documentation says:

                ignore_ce
                        disable features for corrected errors, e.g.
                        polling timer and CMCI.  All events reported as 
                        corrected are not cleared by OS and remained in its
                        error banks.

                        Usually this disablement is not recommended, however
                        if there is an agent checking/clearing corrected
                        errors (e.g. BIOS or hardware monitoring 
                        applications), conflicting with OS's error handling,
                        and you cannot deactivate the agent, then this option
                        will be a help.

it basically disables all: polling *and* CMCI.

So why do we even bother with storms?

> We can focus on "check_interval=0". The user wants to disable MCA
> polling and rely only on interrupts. They still want to see the CEs.

Is that a use case we support?

Where is that documented?

I can see why someone would want to avoid the recurrent polling but I'm not
sure we explicitly say that somewhere in the text...

> When the storm ends, the kernel should go back to how things were before
> the storm. If there was a timer before, then go back to the old
> interval. If there was *not* a timer before, then delete/remove the
> timer.

That I agree with.

Thx.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Yazen Ghannam]

On Tue, May 13, 2025 at 07:55:43PM +0200, Borislav Petkov wrote:
> On Mon, May 12, 2025 at 11:43:15AM -0400, Yazen Ghannam wrote:
> > The use case is "disable MCA polling". I just gave two examples of how
> > this can be done.
> 
> Our documentation says:
> 
>                 ignore_ce
>                         disable features for corrected errors, e.g.
>                         polling timer and CMCI.  All events reported as 
>                         corrected are not cleared by OS and remained in its
>                         error banks.
> 
>                         Usually this disablement is not recommended, however
>                         if there is an agent checking/clearing corrected
>                         errors (e.g. BIOS or hardware monitoring 
>                         applications), conflicting with OS's error handling,
>                         and you cannot deactivate the agent, then this option
>                         will be a help.
> 
> it basically disables all: polling *and* CMCI.
> 
> So why do we even bother with storms?
> 

I think I see your point. The AMD MCA Thresholding init flow doesn't
check "ignore_ce". We should fix that to have more feature parity
between vendors.

Another item for the to do list. :)

> > We can focus on "check_interval=0". The user wants to disable MCA
> > polling and rely only on interrupts. They still want to see the CEs.
> 
> Is that a use case we support?
> 
> Where is that documented?
> 
> I can see why someone would want to avoid the recurrent polling but I'm not
> sure we explicitly say that somewhere in the text...
> 

Right, good point. This may come down to another vendor difference.

On Intel, the set of banks for polling and for CMCI are mutually
exclusive. So polling can be effectively disabled if all banks support
CMCI.

On AMD, polling and interrupt are independent. We still poll all banks
even if they are interrupt-capable. I think we discussed this in a
previous revision of this set.

So I think we should document this use case. And also consider changing
the polling/interrupt behavior on AMD. Even if this use case is
documented, it still requires user intervention which we can avoid if we
change the kernel behavior.

> > When the storm ends, the kernel should go back to how things were before
> > the storm. If there was a timer before, then go back to the old
> > interval. If there was *not* a timer before, then delete/remove the
> > timer.
> 
> That I agree with.
> 

Okay.

Thanks,
Yazen

RE: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Luck, Tony]

> On AMD, polling and interrupt are independent. We still poll all banks
> even if they are interrupt-capable. I think we discussed this in a
> previous revision of this set.

Can you race and double report the same error if a polling interval
and interrupt happen together?

Disabling polling for interrupt capable banks happened before I
started looking at this code. But I assumed it was to avoid double
report.

-Tony

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Yazen Ghannam]

On Tue, May 13, 2025 at 10:07:13PM +0000, Luck, Tony wrote:
> > On AMD, polling and interrupt are independent. We still poll all banks
> > even if they are interrupt-capable. I think we discussed this in a
> > previous revision of this set.
> 
> Can you race and double report the same error if a polling interval
> and interrupt happen together?
> 

Maybe, but probably very unlikely.

On AMD, MCA bank management is always 'local', i.e. per-CPU.

If a CPU is in the polling function, can it be preempted by an interrupt
(not MCE)?

> Disabling polling for interrupt capable banks happened before I
> started looking at this code. But I assumed it was to avoid double
> report.
> 

Ah okay. I assumed it was a performance thing too. But maybe that's just
a nice side effect.

Thanks,
Yazen

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Borislav Petkov]

On Wed, May 14, 2025 at 10:34:16AM -0400, Yazen Ghannam wrote:
> On AMD, MCA bank management is always 'local', i.e. per-CPU.
> 
> If a CPU is in the polling function, can it be preempted by an interrupt
> (not MCE)?

Well, ofc. We're polling with interrupts enabled.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v3 17/17] x86/mce: Restore poll settings after storm subsides

[Yazen Ghannam]

On Thu, May 15, 2025 at 02:37:08PM +0200, Borislav Petkov wrote:
> On Wed, May 14, 2025 at 10:34:16AM -0400, Yazen Ghannam wrote:
> > On AMD, MCA bank management is always 'local', i.e. per-CPU.
> > 
> > If a CPU is in the polling function, can it be preempted by an interrupt
> > (not MCE)?
> 
> Well, ofc. We're polling with interrupts enabled.
> 

Right.

The polling function is called from a timer. I expect the timer is
checked during a timer tick or during rescheduling.

Even though these events are interrupt-driven, it doesn't make sense to
stay in interrupt context the whole time. I think this is where my
thoughts were.

So there's a slight change of double counting errors if the polling
function is interrupted between reading MCA_STATUS in a bank and
clearing it.

Thanks,
Yazen