[PATCH 5/7] platform/x86/amd/hsmp: Add SMN read IOCTL support
Muralidhara M K <[email protected]> Wed, 29 Jul 2026 22:10:32 +0530
| Newsgroups | org.kernel.vger.platform-driver-x86,org.kernel.vger.linux-kernel |
|---|---|
| Message-ID | <[email protected]> |
Bringing up a client platform means reading SMN registers that the HSMP message set does not cover, and today that means either a debugfs interface that is not present on production kernels or an out-of-tree module. Add HSMP_IOCTL_SMN_CMD so a privileged tool can issue a 32-bit SMN read for a given socket through /dev/hsmp, routed through the same amd_smn_hsmp_rdwr() helper the driver already uses for its mailbox. The ioctl is read-only. struct hsmp_smn_message carries a @write flag so write support could be added later without changing the layout, but a request that sets it is rejected with -EPERM regardless of the caller's privilege or open mode: an SMN write can have side effects anywhere in the SoC, and nothing in this driver can bound them. Reads still require CAP_SYS_RAWIO, because even reading can touch registers with read side effects. Every field of the request falls on its natural alignment under the surrounding #pragma pack(4), so the struct is a tight 12 bytes with the same wire layout for 32-bit and 64-bit callers, and the reserved byte is rejected when non-zero so a future kernel can repurpose it without breaking deployed userspace. The handler takes hsmp_sock_rwsem for read across the bounds check and the access. /dev/hsmp is a singleton character device that outlives an individual socket unbind, so an ioctl on an already-open fd can run concurrently with teardown, and remove takes the same lock for write while it tears the socket array down; without the lock the check against num_sockets could be made against a count that is already being torn down. The copy_to_user() is deliberately outside the lock, because faulting in a userfaultfd-backed destination can block indefinitely and would otherwise leave a socket unbind waiting for the write lock. This mirrors HSMP_IOCTL_GET_TELEMETRY_DATA. The user-controlled socket index is clamped with array_index_nospec() before it is used to reach per-node data, mitigating Spectre v1 (CVE-2017-5753). Signed-off-by: Muralidhara M K <[email protected]> --- arch/x86/include/uapi/asm/amd_hsmp.h | 34 ++++++++++++++ drivers/platform/x86/amd/hsmp/hsmp.c | 69 ++++++++++++++++++++++++++++ 2 files changed, 103 insertions(+) diff --git a/arch/x86/include/uapi/asm/amd_hsmp.h b/arch/x86/include/uapi/asm/amd_hsmp.h index b4c3ddb9d1c1..9c4ad22e47ae 100644 --- a/arch/x86/include/uapi/asm/amd_hsmp.h +++ b/arch/x86/include/uapi/asm/amd_hsmp.h @@ -79,6 +79,31 @@ struct hsmp_message { __u16 sock_ind; /* socket number */ }; +/** + * struct hsmp_smn_message - Request descriptor for the HSMP SMN read IOCTL + * @smn_address: Input. SMN address to read. + * @value: Output. Populated by the kernel with the value read. + * @sock_ind: Input. Socket index the read is issued on. + * @write: Must be zero. This IOCTL is read-only, so a request with + * @write set is rejected with -EPERM whatever the caller's + * privilege or open mode. The field is kept so that write + * support could be added later without changing the layout. + * @reserved: Reserved for future use. Callers must set this to zero; a + * non-zero value is rejected with -EINVAL so future kernels can + * repurpose the field without breaking deployed userspace. + * + * Every field falls on its natural alignment under the surrounding + * #pragma pack(4), so the struct is a tight 12 bytes with the same wire + * layout on 32-bit and 64-bit userspace. + */ +struct hsmp_smn_message { + __u32 smn_address; + __u32 value; + __u16 sock_ind; + __u8 write; + __u8 reserved; +}; + enum hsmp_msg_type { HSMP_RSVD = -1, HSMP_SET = 0, @@ -664,6 +689,15 @@ struct hsmp_telemetry_data { #define HSMP_IOCTL_GET_TELEMETRY_DATA \ _IOW(HSMP_BASE_IOCTL_NR, 1, struct hsmp_telemetry_data) +/* + * Read a 32-bit SMN register on a given socket. This reaches registers + * outside the HSMP message ABI, so it requires CAP_SYS_RAWIO. + * + * The direction is _IOWR because the kernel reads the request struct and + * writes the value it read back into the same struct. + */ +#define HSMP_IOCTL_SMN_CMD _IOWR(HSMP_BASE_IOCTL_NR, 2, struct hsmp_smn_message) + /* * Client HSMP messages supported on the Family 1Ah client platforms: * Models 80h-87h (Medusa1), Models 88h-8Fh (Olympic Ridge) and diff --git a/drivers/platform/x86/amd/hsmp/hsmp.c b/drivers/platform/x86/amd/hsmp/hsmp.c index 2326506929a4..ec11c5840b87 100644 --- a/drivers/platform/x86/amd/hsmp/hsmp.c +++ b/drivers/platform/x86/amd/hsmp/hsmp.c @@ -8,8 +8,10 @@ */ #include <asm/amd/hsmp.h> +#include <asm/amd/node.h> #include <linux/acpi.h> +#include <linux/capability.h> #include <linux/cleanup.h> #include <linux/delay.h> #include <linux/device.h> @@ -430,6 +432,71 @@ static bool is_get_msg(struct hsmp_message *msg) return false; } +/* + * Read a 32-bit SMN register on the requested socket. + * + * Raw SMN access reaches registers that are not part of the HSMP message ABI + * and whose side effects the driver cannot reason about, so it is gated on + * CAP_SYS_RAWIO. Writes are not offered at all: the request carries a @write + * flag so support could be added later, and a request that sets it today is + * refused whatever the caller's privilege or open mode. + */ +static long hsmp_ioctl_smn(struct file *fp, unsigned long arg) +{ + void __user *arguser = (void __user *)arg; + struct hsmp_smn_message smn; + unsigned int sock_ind; + int ret; + + if (!capable(CAP_SYS_RAWIO)) + return -EPERM; + + /* The value read travels back in the request struct. */ + if (!(fp->f_mode & FMODE_READ)) + return -EPERM; + + if (copy_from_user(&smn, arguser, sizeof(smn))) + return -EFAULT; + + if (smn.write) + return -EPERM; + + if (smn.reserved) + return -EINVAL; + + /* + * /dev/hsmp is a singleton character device that outlives an individual + * socket unbind, so an ioctl on an already-open fd can run concurrently + * with socket teardown. Hold hsmp_sock_rwsem for read across the bounds + * check and the access, since remove takes the same lock for write + * while it tears the socket array down. + */ + scoped_guard(rwsem_read, &hsmp_sock_rwsem) { + if (!hsmp_pdev.sock || smn.sock_ind >= hsmp_pdev.num_sockets) + return -ENODEV; + + /* + * Sanitize the user-controlled socket index against speculative + * execution. The bounds check above retires the out-of-range + * case with -ENODEV, but a mispredicted branch can still let + * the CPU speculatively use the index to reach per-node data + * and pull arbitrary kernel memory into the cache (Spectre v1, + * CVE-2017-5753). + */ + sock_ind = array_index_nospec(smn.sock_ind, hsmp_pdev.num_sockets); + + ret = amd_smn_hsmp_rdwr(sock_ind, smn.smn_address, &smn.value, false); + } + + if (ret) + return ret; + + if (copy_to_user(arguser, &smn, sizeof(smn))) + return -EFAULT; + + return 0; +} + static long hsmp_ioctl_msg(struct file *fp, unsigned long arg) { int __user *arguser = (int __user *)arg; @@ -620,6 +687,8 @@ long hsmp_ioctl(struct file *fp, unsigned int cmd, unsigned long arg) return hsmp_ioctl_msg(fp, arg); case HSMP_IOCTL_GET_TELEMETRY_DATA: return hsmp_ioctl_get_telemetry(fp, arg); + case HSMP_IOCTL_SMN_CMD: + return hsmp_ioctl_smn(fp, arg); default: return -ENOTTY; } -- 2.34.1