STABILITY: SMP improvements

.gitignore

@@ -30,3 +30,4 @@ limine
 .DS_Store
 /build/
 *.o
+disk.img

src/arch/interrupts.asm

@@ -45,7 +45,11 @@ isr%2_wrapper:
     push r14
     push r15
     
-    ; Save SSE/FPU state (fxsave requires 16-byte alignment)
+    test qword [rsp + 144], 3
+    jz %%skip_swap
+    swapgs
+%%skip_swap:
+    
     sub rsp, 512
     fxsave [rsp]
 
@@ -76,6 +80,12 @@ isr%2_wrapper:
     pop rcx
     pop rbx
     pop rax
+
+    test qword [rsp + 24], 3
+    jz %%skip_swap_back
+    swapgs
+%%skip_swap_back:
+
     add rsp, 16 ; drop dummy vector and error code
     iretq
 %endmacro
@@ -164,7 +174,11 @@ exception_common:
     push r14
     push r15
 
-    ; Save SSE/FPU state (fxsave requires 16-byte alignment)
+    test qword [rsp + 144], 3
+    jz .skip_swap_exc
+    swapgs
+.skip_swap_exc:
+    
     sub rsp, 512
     fxsave [rsp]
 
@@ -196,6 +210,12 @@ exception_common:
     pop rcx
     pop rbx
     pop rax
+
+    test qword [rsp + 24], 3
+    jz .skip_swap_back_exc
+    swapgs
+.skip_swap_back_exc:
+
     add rsp, 16 ; drop vector and error code
     iretq
 

src/arch/syscalls.asm

@@ -15,15 +15,14 @@ section .text
 ; R9  = arg5
 
 syscall_entry:
-    ; 1. Switch to Kernel Stack safely
+    swapgs 
-    ; Note: For true SMP safety, we need per-CPU storage (via swapgs).
-    ; For now, we use a global scratch which is only safe because we mask interrupts on entry.
-    mov [rel user_rsp_scratch], rsp
-    mov rsp, [rel kernel_syscall_stack]
     
-    ; 2. Build iretq frame (compatible with registers_t)
+    mov [gs:40], rsp
+    mov rsp, [gs:48]
+
+    ; 2. Build iretq frame 
     push 0x1B           ; SS (User Data)
-    push qword [rel user_rsp_scratch] ; RSP
+    push qword [gs:40]  ; RSP
     push r11            ; RFLAGS (captured by syscall)
     push 0x23           ; CS (User Code)
     push rcx            ; RIP (return address from syscall)
@@ -81,14 +80,7 @@ syscall_entry:
     pop rax
     add rsp, 16 ; drop int_no/err_code
     
-    ; Debug: check RIP before iretq
+    swapgs 
-    ; We can't easily print from here without destroying registers, 
-    ; but we can at least check if it's canonical.
-    
     iretq
 
 section .bss
-global kernel_syscall_stack
-global user_rsp_scratch
-kernel_syscall_stack: resq 1
-user_rsp_scratch: resq 1

src/core/main.c

@@ -158,6 +158,8 @@ void kmain(void) {
         // The memory manager will now scan the memory map and manage all usable regions.
         memory_manager_init_from_memmap(memmap_request.response);
         serial_write("[DEBUG] memory_manager_init OK\n");
+        smp_init_bsp();
+        serial_write("[DEBUG] smp_init_bsp OK\n");
     } else {
         serial_write("[DEBUG] ERROR: No usable memory for heap! Check Limine memmap.\n");
         hcf();

src/fs/fat32.c

@@ -1561,15 +1561,11 @@ static int vfs_realfs_read(void *fs_private, void *file_handle, void *buf, int s
     uint8_t *cluster_buf = (uint8_t*)kmalloc(cluster_size);
     if (!cluster_buf) return -1;
 
-    extern void serial_write(const char*);
-    serial_write("[VFS] vfs_realfs_read enter\n");
-
     uint64_t rflags = spinlock_acquire_irqsave(&vol->lock);
     int ret = realfs_read_file(handle, buf, size, cluster_buf);
     spinlock_release_irqrestore(&vol->lock, rflags);
 
     kfree(cluster_buf);
-    serial_write("[VFS] vfs_realfs_read exit\n");
     return ret;
 }
 
@@ -1583,15 +1579,11 @@ static int vfs_realfs_write(void *fs_private, void *file_handle, const void *buf
     uint8_t *cluster_buf = (uint8_t*)kmalloc(cluster_size);
     if (!cluster_buf) return -1;
     
-    extern void serial_write(const char*);
-    serial_write("[VFS] vfs_realfs_write enter\n");
-
     uint64_t rflags = spinlock_acquire_irqsave(&vol->lock);
     int ret = realfs_write_file(handle, buf, size, cluster_buf);
     spinlock_release_irqrestore(&vol->lock, rflags);
 
     kfree(cluster_buf);
-    serial_write("[VFS] vfs_realfs_write exit\n");
     return ret;
 }
 

src/sys/process.c

@@ -23,10 +23,10 @@ process_t processes[MAX_PROCESSES] __attribute__((aligned(16)));
 int process_count = 0;
 static process_t* current_process[MAX_CPUS_SCHED] = {0}; // Per-CPU
 static uint32_t next_pid = 0;
-static void *free_kernel_stack_later = NULL;
+static void *free_kernel_stack_later[MAX_CPUS_SCHED] = {0};
-static uint64_t free_pml4_later = 0;
+static uint64_t free_pml4_later[MAX_CPUS_SCHED] = {0};
 static spinlock_t runqueue_lock = SPINLOCK_INIT;
-static uint32_t next_cpu_assign = 1; // Round-robin CPU assignment (start from CPU 1)
+static uint32_t next_cpu_assign = 1; 
 
 void process_init(void) {
     for (int i = 0; i < MAX_PROCESSES; i++) {
@@ -376,21 +376,35 @@ process_t* process_get_current(void) {
 }
 
 uint64_t process_schedule(uint64_t current_rsp) {
-    if (free_kernel_stack_later) {
+    uint32_t my_cpu = smp_this_cpu_id();
-        kfree(free_kernel_stack_later);
+    uint64_t rflags = spinlock_acquire_irqsave(&runqueue_lock);
-        free_kernel_stack_later = NULL;
+
+    void *cleanup_stack = NULL;
+    uint64_t cleanup_pml4 = 0;
+
+    if (free_kernel_stack_later[my_cpu]) {
+        cleanup_stack = free_kernel_stack_later[my_cpu];
+        free_kernel_stack_later[my_cpu] = NULL;
     }
-    if (free_pml4_later) {
+    if (free_pml4_later[my_cpu]) {
-        extern void paging_destroy_user_pml4_phys(uint64_t pml4_phys);
+        cleanup_pml4 = free_pml4_later[my_cpu];
-        paging_destroy_user_pml4_phys(free_pml4_later);
+        free_pml4_later[my_cpu] = 0;
-        free_pml4_later = 0;
     }
 
-    uint32_t my_cpu = smp_this_cpu_id();
     process_t *cur = current_process[my_cpu];
     
-    if (!cur || !cur->next || cur == cur->next) 
+    if (!cur || !cur->next || cur == cur->next) {
+        spinlock_release_irqrestore(&runqueue_lock, rflags);
+
+        // Perform cleanup outside the lock
+        if (cleanup_stack) kfree(cleanup_stack);
+        if (cleanup_pml4) {
+            extern void paging_destroy_user_pml4_phys(uint64_t pml4_phys);
+            paging_destroy_user_pml4_phys(cleanup_pml4);
+        }
+
         return current_rsp;
+    }
         
     // Save context
     cur->rsp = current_rsp;
@@ -412,11 +426,8 @@ uint64_t process_schedule(uint64_t current_rsp) {
         next_proc = next_proc->next;
     }
     
-    // If we didn't find a ready process for our CPU, stay on current (unless we are terminated)
     if (next_proc->cpu_affinity != my_cpu || next_proc->pid == 0xFFFFFFFF) {
-        // Fallback to idle if current is terminated
         if (cur && cur->pid == 0xFFFFFFFF) {
-            // Find the idle process for this CPU
             for (int i = 0; i < MAX_PROCESSES; i++) {
                 if (processes[i].pid == 0 || (processes[i].cpu_affinity == my_cpu && processes[i].is_user == false)) {
                     next_proc = &processes[i];
@@ -424,18 +435,25 @@ uint64_t process_schedule(uint64_t current_rsp) {
                 }
             }
         } else {
+            spinlock_release_irqrestore(&runqueue_lock, rflags);
+
+            if (cleanup_stack) kfree(cleanup_stack);
+            if (cleanup_pml4) {
+                extern void paging_destroy_user_pml4_phys(uint64_t pml4_phys);
+                paging_destroy_user_pml4_phys(cleanup_pml4);
+            }
+
             return current_rsp;
         }
     }
     
     current_process[my_cpu] = next_proc;
     
-    // Update Kernel Stack for User Mode interrupts and System Calls
     if (current_process[my_cpu]->is_user && current_process[my_cpu]->kernel_stack) {
         tss_set_stack_cpu(my_cpu, current_process[my_cpu]->kernel_stack);
-        if (my_cpu == 0) {
+        cpu_state_t *cpu_state = smp_get_cpu(my_cpu);
-            extern uint64_t kernel_syscall_stack;
+        if (cpu_state) {
-            kernel_syscall_stack = current_process[my_cpu]->kernel_stack;
+            cpu_state->kernel_syscall_stack = current_process[my_cpu]->kernel_stack;
         }
     }
     
@@ -443,8 +461,16 @@ uint64_t process_schedule(uint64_t current_rsp) {
     paging_switch_directory(current_process[my_cpu]->pml4_phys);
     
     current_process[my_cpu]->ticks++;
+    uint64_t next_rsp = current_process[my_cpu]->rsp;
 
-    return current_process[my_cpu]->rsp;
+    spinlock_release_irqrestore(&runqueue_lock, rflags);
+    if (cleanup_stack) kfree(cleanup_stack);
+    if (cleanup_pml4) {
+        extern void paging_destroy_user_pml4_phys(uint64_t pml4_phys);
+        paging_destroy_user_pml4_phys(cleanup_pml4);
+    }
+
+    return next_rsp;
 }
 
 process_t* process_get_by_pid(uint32_t pid) {
@@ -600,25 +626,22 @@ uint64_t process_terminate_current(void) {
     // 4. Load context for the NEXT process
     if (current_process[my_cpu]->is_user && current_process[my_cpu]->kernel_stack) {
         tss_set_stack_cpu(my_cpu, current_process[my_cpu]->kernel_stack);
-        if (my_cpu == 0) {
+        cpu_state_t *cpu_state = smp_get_cpu(my_cpu);
-            extern uint64_t kernel_syscall_stack;
+        if (cpu_state) {
-            kernel_syscall_stack = current_process[my_cpu]->kernel_stack;
+            cpu_state->kernel_syscall_stack = current_process[my_cpu]->kernel_stack;
         }
     }
     
     paging_switch_directory(current_process[my_cpu]->pml4_phys);
 
-    // 5. Free memory
-    if (to_delete->user_stack_alloc) kfree(to_delete->user_stack_alloc);
    
-    extern void paging_destroy_user_pml4_phys(uint64_t pml4_phys);
+        kfree(to_delete->user_stack_alloc);
-    if (to_delete->pml4_phys && to_delete->is_user) {
+        to_delete->user_stack_alloc = NULL;
-        paging_destroy_user_pml4_phys(to_delete->pml4_phys);
     }
 
-    to_delete->user_stack_alloc = NULL;
+    free_kernel_stack_later[my_cpu] = to_delete->kernel_stack_alloc;
-    free_kernel_stack_later = to_delete->kernel_stack_alloc;
     to_delete->kernel_stack_alloc = NULL;
+    free_pml4_later[my_cpu] = to_delete->pml4_phys;
     to_delete->pml4_phys = 0;
     
     uint64_t next_rsp = current_process[my_cpu]->rsp;
@@ -666,4 +689,3 @@ process_t* process_get_by_ui_window(void *win) {
     }
     return NULL;
 }
-

src/sys/smp.c

@@ -18,6 +18,16 @@ extern void serial_write_hex(uint64_t n);
 static cpu_state_t *cpu_states = NULL;
 static uint32_t total_cpus = 0;
 static uint32_t bsp_lapic_id = 0;
+static cpu_state_t *bsp_cpu_state = NULL;
+
+#define MSR_GS_BASE         0xC0000101
+#define MSR_KERNEL_GS_BASE  0xC0000102
+
+static inline void wrmsr(uint32_t msr, uint64_t value) {
+    uint32_t low = (uint32_t)value;
+    uint32_t high = (uint32_t)(value >> 32);
+    asm volatile("wrmsr" : : "c"(msr), "a"(low), "d"(high));
+}
 
 static uint32_t read_lapic_id(void) {
     uint32_t eax, ebx, ecx, edx;
@@ -27,6 +37,12 @@ static uint32_t read_lapic_id(void) {
 
 uint32_t smp_this_cpu_id(void) {
     if (total_cpus <= 1) return 0;
+    
+    // Use GS-based self-pointer to get the structure first
+    cpu_state_t *state;
+    asm volatile("movq %%gs:0, %0" : "=r"(state) : : "memory");
+    if (state) return state->cpu_id;
+
     uint32_t lapic = read_lapic_id();
     for (uint32_t i = 0; i < total_cpus; i++) {
         if (cpu_states[i].lapic_id == lapic) return i;
@@ -68,13 +84,21 @@ static void ap_entry(struct limine_smp_info *info) {
     extern void idt_load(void);
     idt_load();
 
+    extern void syscall_init(void);
+    syscall_init();
+
     uint64_t kernel_cr3 = paging_get_pml4_phys();
     asm volatile("mov %0, %%cr3" : : "r"(kernel_cr3));
 
     extern void lapic_enable(void);
     lapic_enable();
 
+    cpu_states[my_id].self = &cpu_states[my_id];
     cpu_states[my_id].online = true;
+    cpu_states[my_id].kernel_syscall_stack = cpu_states[my_id].kernel_stack;
+
+    wrmsr(MSR_GS_BASE, (uint64_t)&cpu_states[my_id]);
+    wrmsr(MSR_KERNEL_GS_BASE, (uint64_t)&cpu_states[my_id]);
 
     serial_write("[SMP] AP ");
     serial_write_num(my_id);
@@ -90,6 +114,19 @@ static void ap_entry(struct limine_smp_info *info) {
     work_queue_drain_loop();
 }
 
+void smp_init_bsp(void) {
+    static cpu_state_t bsp_state_static = {0};
+    bsp_state_static.cpu_id = 0;
+    bsp_state_static.lapic_id = read_lapic_id();
+    bsp_state_static.self = &bsp_state_static;
+    bsp_state_static.online = true;
+    
+    wrmsr(MSR_GS_BASE, (uint64_t)&bsp_state_static);
+    wrmsr(MSR_KERNEL_GS_BASE, (uint64_t)&bsp_state_static);
+    
+    bsp_cpu_state = &bsp_state_static;
+}
+
 // --- SMP Initialization ---
 uint32_t smp_init(struct limine_smp_response *smp_resp) {
     if (!smp_resp || smp_resp->cpu_count <= 1) {
@@ -132,8 +169,15 @@ uint32_t smp_init(struct limine_smp_response *smp_resp) {
         cpu_states[i].lapic_id = cpu->lapic_id;
 
         if (cpu->lapic_id == bsp_lapic_id) {
-            cpu_states[i].online = true;
+            cpu_states[i] = *bsp_cpu_state; // Copy early BSP state
+            cpu_states[i].self = &cpu_states[i];
+            
+            cpu_states[i].kernel_stack = 0; // Limine stack for now
+            cpu_states[i].kernel_syscall_stack = 0; 
             bsp_index = i;
+            wrmsr(MSR_GS_BASE, (uint64_t)&cpu_states[i]);
+            wrmsr(MSR_KERNEL_GS_BASE, (uint64_t)&cpu_states[i]);
+            
             serial_write("[SMP] BSP CPU ");
             serial_write_num(i);
             serial_write(" (LAPIC ");

src/sys/smp.h

@@ -8,29 +8,27 @@
 #include <stdbool.h>
 #include "spinlock.h"
 
-// Per-CPU state. Dynamically allocated at boot based on actual CPU count.
 typedef struct cpu_state {
-    uint32_t cpu_id;           // Logical CPU index (0 = BSP)
+    struct cpu_state *self;    
-    uint32_t lapic_id;         // Local APIC ID from Limine
+    uint32_t cpu_id;           
-    uint64_t kernel_stack;     // Top of kernel stack for this CPU
+    uint32_t lapic_id;         
-    void    *kernel_stack_alloc; // Base allocation for kfree
+    uint64_t kernel_stack;     
-    volatile bool online;      // True once AP is fully initialized
+    void    *kernel_stack_alloc; 
+    volatile bool online;      
+    uint64_t user_rsp_scratch;  
+    uint64_t kernel_syscall_stack; 
 } cpu_state_t;
  
-// Initialize SMP — call after GDT/IDT/memory init but before wm_init.
+ void smp_init_bsp(void);
-// Pass the Limine SMP response. APs will be started and will enter their
+
-// idle loops. Returns the number of CPUs brought online.
+
 struct limine_smp_response;
 uint32_t smp_init(struct limine_smp_response *smp_resp);
 
-// Get the current CPU index (0 = BSP). Uses CPUID to read LAPIC ID,
-// then looks up in the cpu table.
 uint32_t smp_this_cpu_id(void);
 
-// Total number of CPUs online.
 uint32_t smp_cpu_count(void);
 
-// Get per-CPU state by index.
 cpu_state_t *smp_get_cpu(uint32_t cpu_id);
 
 #endif

src/sys/syscall.c

@@ -77,6 +77,14 @@ static void smp_user_wrapper(void *arg) {
 }
 
 void syscall_init(void) {
+    uint64_t efer = rdmsr(MSR_EFER);
+    efer |= 1; 
+    wrmsr(MSR_EFER, efer);
+    uint64_t star = ((uint64_t)0x001B << 48) | ((uint64_t)0x0008 << 32);
+    wrmsr(MSR_STAR, star);
+    extern void syscall_entry(void);
+    wrmsr(MSR_LSTAR, (uint64_t)syscall_entry);
+    wrmsr(MSR_FMASK, 0x200); 
 }
 
 static void user_window_close(Window *win) {

src/wm/explorer.c

@@ -710,12 +710,7 @@ static void explorer_load_directory(Window *win, const char *path) {
 
     int count = vfs_list_directory(path, entries, capacity);
     
-    // Trace string to see if we reached here
-    extern void serial_write(const char*); extern void serial_write_num(uint32_t);
-    serial_write("[EXPLORER] load_directory: "); serial_write(path); serial_write(" | loop start. count: "); serial_write_num(count); serial_write("\n");
-        
     while (count == capacity) {
-        serial_write("[EXPLORER] Doubling capacity to: "); serial_write_num(capacity * 2); serial_write("\n");
         capacity *= 2;
         vfs_dirent_t *new_entries = (vfs_dirent_t*)krealloc(entries, capacity * sizeof(vfs_dirent_t));
         if (!new_entries) { kfree(entries); return; }
@@ -723,8 +718,6 @@ static void explorer_load_directory(Window *win, const char *path) {
         count = vfs_list_directory(path, entries, capacity);
     }
     
-    serial_write("[EXPLORER] load_directory loop complete.\n");
-    
     if (state->items_capacity < count) {
         int new_cap = count < EXPLORER_INITIAL_CAPACITY ? EXPLORER_INITIAL_CAPACITY : count;
         ExplorerItem *new_items = (ExplorerItem*)krealloc(state->items, new_cap * sizeof(ExplorerItem));