x86/amd_nb: Enhance SMN access error checking

AMD Zen-based systems use a System Management Network (SMN) that
provides access to implementation-specific registers.

SMN accesses are done indirectly through an index/data pair in PCI
config space. The accesses can fail for a variety of reasons.

Include code comments to describe some possible scenarios.

Require error checking for callers of amd_smn_read() and amd_smn_write().
This is needed because many error conditions cannot be checked by these
functions.

  [ bp: Touchup comment. ]

Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
Link: https://lore.kernel.org/r/20240606-fix-smn-bad-read-v4-4-ffde21931c3f@amd.com

Author: yghannam

arch/x86/include/asm/amd_nb.h

@@ -21,8 +21,8 @@ extern int amd_numa_init(void);
 extern int amd_get_subcaches(int);
 extern int amd_set_subcaches(int, unsigned long);
 
-extern int amd_smn_read(u16 node, u32 address, u32 *value);
-extern int amd_smn_write(u16 node, u32 address, u32 value);
+int __must_check amd_smn_read(u16 node, u32 address, u32 *value);
+int __must_check amd_smn_write(u16 node, u32 address, u32 value);
 
 struct amd_l3_cache {
 	unsigned indices;

arch/x86/kernel/amd_nb.c

@@ -180,6 +180,43 @@ static struct pci_dev *next_northbridge(struct pci_dev *dev,
 	return dev;
 }
 
+/*
+ * SMN accesses may fail in ways that are difficult to detect here in the called
+ * functions amd_smn_read() and amd_smn_write(). Therefore, callers must do
+ * their own checking based on what behavior they expect.
+ *
+ * For SMN reads, the returned value may be zero if the register is Read-as-Zero.
+ * Or it may be a "PCI Error Response", e.g. all 0xFFs. The "PCI Error Response"
+ * can be checked here, and a proper error code can be returned.
+ *
+ * But the Read-as-Zero response cannot be verified here. A value of 0 may be
+ * correct in some cases, so callers must check that this correct is for the
+ * register/fields they need.
+ *
+ * For SMN writes, success can be determined through a "write and read back"
+ * However, this is not robust when done here.
+ *
+ * Possible issues:
+ *
+ * 1) Bits that are "Write-1-to-Clear". In this case, the read value should
+ *    *not* match the write value.
+ *
+ * 2) Bits that are "Read-as-Zero"/"Writes-Ignored". This information cannot be
+ *    known here.
+ *
+ * 3) Bits that are "Reserved / Set to 1". Ditto above.
+ *
+ * Callers of amd_smn_write() should do the "write and read back" check
+ * themselves, if needed.
+ *
+ * For #1, they can see if their target bits got cleared.
+ *
+ * For #2 and #3, they can check if their target bits got set as intended.
+ *
+ * This matches what is done for RDMSR/WRMSR. As long as there's no #GP, then
+ * the operation is considered a success, and the caller does their own
+ * checking.
+ */
 static int __amd_smn_rw(u16 node, u32 address, u32 *value, bool write)
 {
 	struct pci_dev *root;
@@ -202,9 +239,6 @@ static int __amd_smn_rw(u16 node, u32 address, u32 *value, bool write)
 
 	err = (write ? pci_write_config_dword(root, 0x64, *value)
 		     : pci_read_config_dword(root, 0x64, value));
-	if (err)
-		pr_warn("Error %s SMN address 0x%x.\n",
-			(write ? "writing to" : "reading from"), address);
 
 out_unlock:
 	mutex_unlock(&smn_mutex);
@@ -213,7 +247,7 @@ static int __amd_smn_rw(u16 node, u32 address, u32 *value, bool write)
 	return err;
 }
 
-int amd_smn_read(u16 node, u32 address, u32 *value)
+int __must_check amd_smn_read(u16 node, u32 address, u32 *value)
 {
 	int err = __amd_smn_rw(node, address, value, false);
 
@@ -226,7 +260,7 @@ int amd_smn_read(u16 node, u32 address, u32 *value)
 }
 EXPORT_SYMBOL_GPL(amd_smn_read);
 
-int amd_smn_write(u16 node, u32 address, u32 value)
+int __must_check amd_smn_write(u16 node, u32 address, u32 value)
 {
 	return __amd_smn_rw(node, address, &value, true);
 }