for cxl memory metadata for software coherency.

Author: vickiegpt

include/cxl_backend.h

@@ -193,18 +193,53 @@ const char* cxl_backend_type_name(cxl_backend_type_t type);
 #define CXL_SHM_CACHELINE_SIZE 64
 
 /* Request types */
-#define CXL_SHM_REQ_NONE      0
-#define CXL_SHM_REQ_READ      1
-#define CXL_SHM_REQ_WRITE     2
-#define CXL_SHM_REQ_ATOMIC_FAA 3
-#define CXL_SHM_REQ_ATOMIC_CAS 4
-#define CXL_SHM_REQ_FENCE     5
+#define CXL_SHM_REQ_NONE          0
+#define CXL_SHM_REQ_READ          1
+#define CXL_SHM_REQ_WRITE         2
+#define CXL_SHM_REQ_ATOMIC_FAA    3
+#define CXL_SHM_REQ_ATOMIC_CAS    4
+#define CXL_SHM_REQ_FENCE         5
+#define CXL_SHM_REQ_READ_META     6   /* Read with metadata */
+#define CXL_SHM_REQ_WRITE_META    7   /* Write with metadata */
+#define CXL_SHM_REQ_GET_META      8   /* Get metadata only */
+#define CXL_SHM_REQ_SET_META      9   /* Set metadata only */
 
 /* Response status */
 #define CXL_SHM_RESP_NONE     0
 #define CXL_SHM_RESP_OK       1
 #define CXL_SHM_RESP_ERROR    2
 
+/* MESI Cache States */
+#define CXL_CACHE_INVALID     0
+#define CXL_CACHE_SHARED      1
+#define CXL_CACHE_EXCLUSIVE   2
+#define CXL_CACHE_MODIFIED    3
+
+/* Metadata flags */
+#define CXL_META_FLAG_DIRTY   0x01
+#define CXL_META_FLAG_LOCKED  0x02
+#define CXL_META_FLAG_PINNED  0x04
+
+/* Cacheline metadata structure (64 bytes) - compatible with Splash libpgas */
+typedef struct {
+    uint8_t cache_state;        /* MESI state */
+    uint8_t owner_id;           /* Current owner host/thread ID */
+    uint16_t sharers_bitmap;    /* Bitmap of hosts/threads sharing this line */
+    uint32_t access_count;      /* Number of accesses */
+    uint64_t last_access_time;  /* Timestamp of last access */
+    uint64_t virtual_addr;      /* Virtual address mapping */
+    uint64_t physical_addr;     /* Physical address */
+    uint32_t version;           /* Version number for coherency */
+    uint8_t flags;              /* Various flags (dirty, locked, etc.) */
+    uint8_t reserved[23];       /* Reserved for future use */
+} __attribute__((packed)) cxl_cacheline_metadata_t;
+
+/* PGAS Memory Entry: 128 bytes (64 data + 64 metadata) per cacheline */
+typedef struct {
+    uint8_t data[CXL_SHM_CACHELINE_SIZE];  /* 64 bytes of data */
+    cxl_cacheline_metadata_t metadata;      /* 64 bytes of metadata */
+} __attribute__((packed, aligned(128))) cxl_pgas_entry_t;
+
 /* Shared memory slot for request/response */
 typedef struct {
     volatile uint32_t req_type;      /* Request type */
@@ -215,23 +250,29 @@ typedef struct {
     volatile uint64_t expected;      /* Expected value for CAS */
     volatile uint64_t latency_ns;    /* Simulated latency */
     volatile uint64_t timestamp;     /* Request timestamp */
-    uint8_t data[CXL_SHM_CACHELINE_SIZE];  /* Data buffer */
-    uint8_t padding[64 - 8];         /* Align to 128 bytes */
-} __attribute__((aligned(128))) cxl_shm_slot_t;
+    uint8_t data[CXL_SHM_CACHELINE_SIZE];  /* Data buffer (64 bytes) */
+    cxl_cacheline_metadata_t metadata;     /* Metadata buffer (64 bytes) */
+} __attribute__((aligned(256))) cxl_shm_slot_t;
 
 /* Shared memory header */
 typedef struct {
     uint64_t magic;                  /* Magic number for validation */
     uint32_t version;                /* Protocol version */
     uint32_t num_slots;              /* Number of request slots */
     volatile uint32_t server_ready;  /* Server is ready flag */
-    uint32_t reserved;
+    uint32_t flags;                  /* Header flags */
     uint64_t memory_base;            /* Base address of simulated memory */
     uint64_t memory_size;            /* Size of simulated memory */
-    uint8_t padding[64 - 40];        /* Pad header to 64 bytes */
+    uint64_t num_cachelines;         /* Number of cachelines (memory_size/64) */
+    uint32_t metadata_enabled;       /* 1 if metadata transfer is enabled */
+    uint32_t entry_size;             /* Size of each entry (64 or 128 bytes) */
+    uint8_t padding[64 - 56];        /* Pad header to 64 bytes */
     cxl_shm_slot_t slots[];          /* Request/response slots */
 } __attribute__((aligned(64))) cxl_shm_header_t;
 
+/* Header flags */
+#define CXL_SHM_FLAG_METADATA_ENABLED  0x01
+
 /* Size calculation */
 #define CXL_SHM_HEADER_SIZE(nslots) \
     (sizeof(cxl_shm_header_t) + (nslots) * sizeof(cxl_shm_slot_t))

include/shm_communication.h

@@ -18,18 +18,29 @@
 #include <atomic>
 #include <cstring>
 
+// Operation type constants for SHM mode
+constexpr uint8_t SHM_OP_READ = 0;
+constexpr uint8_t SHM_OP_WRITE = 1;
+constexpr uint8_t SHM_OP_GET_SHM_INFO = 2;
+constexpr uint8_t SHM_OP_ATOMIC_FAA = 3;   // Fetch-and-Add
+constexpr uint8_t SHM_OP_ATOMIC_CAS = 4;   // Compare-and-Swap
+constexpr uint8_t SHM_OP_FENCE = 5;        // Memory fence
+
 // Request/Response structures matching TCP version
 struct ShmRequest {
-    uint8_t op_type;      // 0=READ, 1=WRITE, 2=GET_SHM_INFO
+    uint8_t op_type;      // 0=READ, 1=WRITE, 2=GET_SHM_INFO, 3=ATOMIC_FAA, 4=ATOMIC_CAS, 5=FENCE
     uint64_t addr;
     uint64_t size;
     uint64_t timestamp;
+    uint64_t value;       // Value for FAA (add value) or CAS (desired value)
+    uint64_t expected;    // Expected value for CAS operation
     uint8_t data[64];     // Cacheline data
 };
 
 struct ShmResponse {
     uint8_t status;
     uint64_t latency_ns;
+    uint64_t old_value;   // Previous value returned by atomic operations
     uint8_t data[64];
 };