7 files changed, 321 insertions, 56 deletions

diff --git a/src/kernel/Makefile b/src/kernel/Makefile
index f7cee23..1b01f13 100644
--- a/src/kernel/Makefile
+++ b/src/kernel/Makefile
@@ -3,7 +3,7 @@ OBJ = core/loader.o core/dbglog.o \
 	core/gdt.o core/idt.o core/interrupt.o core/context_switch.o core/thread.o \
 	core/frame.o core/paging.o core/freemem.o core/region.o core/kmalloc.o \
 	core/worker.o core/prng.o \
-	user/vfs.o user/nullfs.o user/process.o user/elf.o user/syscall.o user/ipc.o \
+	user/vfs.o user/nullfs.o user/process.o user/elf.o user/syscall.o user/ipc.o user/pager.o \
 	dev/pci.o dev/pciide.o \
 	fs/iso9660.o
 
diff --git a/src/kernel/include/pager.h b/src/kernel/include/pager.h
new file mode 100644
index 0000000..5491d10
--- /dev/null
+++ b/src/kernel/include/pager.h
@@ -0,0 +1,73 @@
+#pragma once
+
+#include <hashtbl.h>
+#include <mutex.h>
+
+#include <paging.h>
+
+#define PAGE_PINNED		0x01
+#define PAGE_ACCESSED	0x02
+#define PAGE_DIRTY		0x04
+
+typedef struct pager pager_t;
+typedef struct user_region user_region_t;
+typedef struct fs_node fs_node_t;
+
+typedef struct {
+	size_t (*read)(fs_node_t* data, size_t offset, size_t len, char* ptr);
+	size_t (*write)(fs_node_t* data, size_t offset, size_t len, const char* ptr);
+	bool (*resize)(fs_node_t* data, size_t new_size);
+} vfs_pager_ops_t;
+
+typedef struct {
+	void (*page_in)(pager_t *p, size_t offset, size_t len);			// allocates the pages
+	void (*page_commit)(pager_t *p, size_t offset, size_t len);		// writes pages back to storage (if applicable)
+	void (*page_out)(pager_t *p, size_t offset, size_t len);		// frees the pages (no need to write back)
+	void (*page_release)(pager_t *p, size_t offset, size_t len);	// called on delete/resize
+} pager_ops_t;
+
+typedef struct pager {
+	pager_ops_t *ops;
+
+	union {
+		struct {
+			fs_node_t* vfs_data;
+			vfs_pager_ops_t *vfs_ops;
+		} vfs_pager;
+		struct {
+			size_t phys_offset;
+		} device_pager;
+	};
+
+	size_t size;
+
+	hashtbl_t *pages;
+	mutex_t lock;
+
+	user_region_t *maps;
+} pager_t;
+
+pager_t* new_swap_pager(size_t size);
+pager_t* new_vfs_pager(size_t size, fs_node_t* vfs_node, vfs_pager_ops_t *vfs_ops);
+pager_t* new_dev_pager(size_t size, size_t phys_offset);
+
+void delete_pager(pager_t *p);
+
+bool pager_resize(pager_t *p, size_t newsize);
+
+// void pager_pin_region(pager_t *pager, size_t offset, size_t len);	// implement later
+// void pager_unpin_region(pager_t *pager, size_t offset, size_t len);
+
+void pager_access(pager_t *p, size_t offset, size_t len, bool accessed, bool dirty);
+void pager_commit_dirty(pager_t *p);
+
+int pager_get_frame(pager_t *p, size_t offset);
+
+size_t pager_read(pager_t *p, size_t offset, size_t len, char* buf);
+size_t pager_write(pager_t *p, size_t offset, size_t len, const char* buf);
+
+void pager_add_map(pager_t *p, user_region_t *reg);
+void pager_rm_map(pager_t *p, user_region_t *reg);
+
+
+/* vim: set ts=4 sw=4 tw=0 noet :*/
diff --git a/src/kernel/include/process.h b/src/kernel/include/process.h
index 8dfa274..8cdfe85 100644
--- a/src/kernel/include/process.h
+++ b/src/kernel/include/process.h
@@ -15,9 +15,9 @@
 
 #include <thread.h>
 #include <vfs.h>
+#include <pager.h>
 
-#include <mmap.h>
-
+#include <mmap.h>	// common header for mmaps
 #include <proc.h> 	// common header defining process statuses
 
 #define USERSTACK_ADDR	0xB8000000
@@ -31,12 +31,15 @@ typedef struct user_region {
 	void* addr;
 	size_t size;
 
-	int mode;
+	fs_handle_t *file;
+	pager_t *pager;
+	size_t offset;
+	bool own_pager;
 
-	fs_handle_t *file;		// null if not mmaped-file
-	size_t file_offset;
+	int16_t mode;
 
-	struct user_region *next;
+	struct user_region *next_in_proc;
+	struct user_region *next_for_pager;
 } user_region_t;
 
 typedef struct process {
diff --git a/src/kernel/include/sys.h b/src/kernel/include/sys.h
index 944df96..61bfc59 100644
--- a/src/kernel/include/sys.h
+++ b/src/kernel/include/sys.h
@@ -75,9 +75,10 @@ static inline void invlpg(void* addr) {
 
 #define PAGE_SIZE		0x1000
 #define PAGE_MASK		0xFFFFF000
-#define PAGE_ALIGN_DOWN(x)	(((size_t)x) & PAGE_MASK) 
-#define PAGE_ALIGN_UP(x)  	((((size_t)x)&(~PAGE_MASK)) == 0 ? ((size_t)x) : (((size_t)x) & PAGE_MASK) + PAGE_SIZE)
-#define PAGE_ID(x)			(((size_t)x) / PAGE_SIZE)
+#define PAGE_ALIGNED(x)     ((((size_t)(x)) & (~PAGE_MASK)) == 0)
+#define PAGE_ALIGN_DOWN(x)	(((size_t)(x)) & PAGE_MASK) 
+#define PAGE_ALIGN_UP(x)  	((((size_t)(x))&(~PAGE_MASK)) == 0 ? ((size_t)x) : (((size_t)x) & PAGE_MASK) + PAGE_SIZE)
+#define PAGE_ID(x)			(((size_t)(x)) / PAGE_SIZE)
 #define PAGE_SHIFT		12
 #define PT_SHIFT		10
 // PAGE_SHIFT + PT_SHIFT + PT_SHIFT = 32
diff --git a/src/kernel/include/vfs.h b/src/kernel/include/vfs.h
index cc6d11c..cdd9e8d 100644
--- a/src/kernel/include/vfs.h
+++ b/src/kernel/include/vfs.h
@@ -8,6 +8,8 @@
 
 #include <fs.h> 		// common header
 
+#include <pager.h>
+
 // How to use :
 // - When using a filesystem : never call the operations in fs_*_ops_t directly, use
 // 		the functions defined bellow in section "public functions";
@@ -108,6 +110,8 @@ typedef struct fs_node {
 	// These fields are filled by the FS's specific walk() code
 	fs_node_ops_t *ops;
 	fs_node_ptr data;
+
+	pager_t *pager;
 } fs_node_t;
 
 // -------------------------------------------
diff --git a/src/kernel/user/pager.c b/src/kernel/user/pager.c
new file mode 100644
index 0000000..c892599
--- /dev/null
+++ b/src/kernel/user/pager.c
@@ -0,0 +1,183 @@
+#include <string.h>
+
+#include <pager.h>
+#include <vfs.h>
+#include <process.h>
+#include <frame.h>
+
+#define ENT_FRAME_SHIFT 12
+
+
+// ========== //
+// SWAP PAGER //
+// ========== //
+
+static void swap_page_in(pager_t *p, size_t offset, size_t len);
+static void swap_page_out(pager_t *p, size_t offset, size_t len);
+static void swap_page_release(pager_t *p, size_t offset, size_t len);
+
+static pager_ops_t swap_pager_ops = {
+	.page_in = swap_page_in,
+	.page_commit = 0,
+	.page_out = swap_page_out,
+	.page_release = swap_page_release,
+};
+
+pager_t *new_swap_pager(size_t size) {
+	pager_t *p = (pager_t*)malloc(sizeof(pager_t));
+	if (p == 0) return 0;
+
+	p->pages = create_hashtbl(id_key_eq_fun, id_hash_fun, 0);
+	if (p->pages == 0) goto error;
+
+	p->ops = &swap_pager_ops;
+	p->size = size;
+	p->lock = MUTEX_UNLOCKED;
+	p->maps = 0;
+
+	return p;
+
+error:
+	if (p) free(p);
+	return 0;
+}
+
+void swap_page_in(pager_t *p, size_t offset, size_t len) {
+	ASSERT(PAGE_ALIGNED(offset));
+	size_t npages = PAGE_ALIGN_UP(len) / PAGE_SIZE;
+
+	void *region = region_alloc(PAGE_SIZE, "Page zero area", 0);
+	if (region == 0) return;
+
+	for (size_t i = 0; i < npages; i++) {
+		size_t page = offset + (i * PAGE_SIZE);
+		if (hashtbl_find(p->pages, (void*)page) == 0) {
+			uint32_t frame = frame_alloc(1);
+			if (frame != 0) {
+				pd_map_page(region, frame, true);
+				memset(region, 0, PAGE_SIZE);
+				pd_unmap_page(region);
+			}
+			if (!hashtbl_add(p->pages, (void*)page, (void*)(frame << ENT_FRAME_SHIFT))) {
+				frame_free(frame, 1);
+			}
+		}
+	}
+
+	region_free(region);
+}
+
+void swap_page_out(pager_t *p, size_t offset, size_t len) {
+	// later : save to swap file...
+}
+
+void swap_page_release(pager_t *p, size_t offset, size_t len) {
+	ASSERT(PAGE_ALIGNED(offset));
+	size_t npages = PAGE_ALIGN_UP(len) / PAGE_SIZE;
+
+	for (size_t i = 0; i < npages; i++) {
+		size_t page = offset + (i * PAGE_SIZE);
+		uint32_t frame = (uint32_t)hashtbl_find(p->pages, (void*)page) >> ENT_FRAME_SHIFT;
+		if (frame != 0) {
+			hashtbl_remove(p->pages, (void*)page);
+			frame_free(frame, 1);
+		}
+	}
+}
+
+// ======================= //
+// GENERIC PAGER FUNCTIONS //
+// ======================= //
+
+void delete_pager(pager_t *p) {
+	ASSERT(p->maps == 0);
+
+	pager_commit_dirty(p);
+
+	mutex_lock(&p->lock);
+
+	p->ops->page_release(p, 0, p->size);
+	delete_hashtbl(p->pages);
+	free(p);
+}
+
+void pager_access(pager_t *p, size_t offset, size_t len, bool accessed, bool dirty) {
+	mutex_lock(&p->lock);
+
+	ASSERT(PAGE_ALIGNED(offset));
+	size_t npages = PAGE_ALIGN_UP(len) / PAGE_SIZE;
+
+	for (size_t i = 0; i < npages; i++) {
+		size_t page = offset + (i * PAGE_SIZE);
+		uint32_t v = (uint32_t)hashtbl_find(p->pages, (void*)page);
+		if (v != 0) {
+			hashtbl_change(p->pages, (void*)page,
+				(void*)(v | (accessed ? PAGE_ACCESSED : 0) | (dirty ? PAGE_DIRTY : 0)));
+		}
+	}
+
+	mutex_unlock(&p->lock);
+}
+
+void pager_commit_dirty(pager_t *p) {
+	if (!p->ops->page_commit) return;
+
+	mutex_lock(&p->lock);
+
+	for (size_t i = 0; i < p->size; i += PAGE_SIZE) {
+		uint32_t v = (uint32_t)hashtbl_find(p->pages, (void*)i);
+		if (v & PAGE_DIRTY) {
+			p->ops->page_commit(p, i, PAGE_SIZE);
+			hashtbl_change(p->pages, (void*)i, (void*)(v & ~PAGE_DIRTY));
+		}
+
+	}
+
+	mutex_unlock(&p->lock);
+}
+
+int pager_get_frame(pager_t *p, size_t offset) {
+	mutex_lock(&p->lock);
+
+	ASSERT(PAGE_ALIGNED(offset));
+
+	uint32_t v = (uint32_t)hashtbl_find(p->pages, (void*)offset);
+	if (v == 0) {
+		p->ops->page_in(p, offset, PAGE_SIZE);
+	}
+
+	uint32_t ret = (uint32_t)hashtbl_find(p->pages, (void*)offset) >> ENT_FRAME_SHIFT;
+
+	mutex_unlock(&p->lock);
+
+	return ret;
+}
+
+void pager_add_map(pager_t *p, user_region_t *reg) {
+	mutex_lock(&p->lock);
+
+	reg->next_for_pager = p->maps;
+	p->maps = reg;
+
+	mutex_unlock(&p->lock);
+}
+
+void pager_rm_map(pager_t *p, user_region_t *reg) {
+	mutex_lock(&p->lock);
+
+	if (p->maps == reg) {
+		p->maps = reg->next_for_pager;
+	} else {
+		for (user_region_t *it = p->maps; it->next_for_pager != 0; it = it->next_for_pager) {
+			if (it->next_for_pager == reg) {
+				it->next_for_pager = reg->next_for_pager;
+				break;
+			}
+		}
+	}
+	reg->next_for_pager = 0;
+
+	mutex_unlock(&p->lock);
+}
+
+/* vim: set ts=4 sw=4 tw=0 noet :*/
diff --git a/src/kernel/user/process.c b/src/kernel/user/process.c
index 2812fc6..1b3013b 100644
--- a/src/kernel/user/process.c
+++ b/src/kernel/user/process.c
@@ -489,32 +489,38 @@ bool mmap(process_t *proc, void* addr, size_t size, int mode) {
 
 	if ((uint32_t)addr & (~PAGE_MASK)) return false;
 
+	if (addr >= (void*)K_HIGHHALF_ADDR || addr + size > (void*)K_HIGHHALF_ADDR
+		|| addr + size <= addr || size == 0) return false;
+
 	user_region_t *r = (user_region_t*)malloc(sizeof(user_region_t));
 	if (r == 0) return false;
 
+	r->proc = proc;
 	r->addr = addr;
-	r->size = PAGE_ALIGN_UP(size);
+	r->size = size;
 
-	if (r->addr >= (void*)K_HIGHHALF_ADDR || r->addr + r->size > (void*)K_HIGHHALF_ADDR
-		|| r->addr + r->size <= r->addr || r->size == 0) {
-		free(r);
-		return false;
-	}
+	r->own_pager = true;
+	r->pager = new_swap_pager(size);
+	if (r->pager == 0) goto error;
+	r->offset = 0;
 
 	bool add_ok = btree_add(proc->regions_idx, r->addr, r);
-	if (!add_ok) {
-		free(r);
-		return false;
-	}
+	if (!add_ok) goto error;
 
 	r->mode = mode;
 	r->file = 0;
-	r->file_offset = 0;
 
-	r->next = proc->regions;
+	r->next_in_proc = proc->regions;
 	proc->regions = r;
 
+	pager_add_map(r->pager, r);
+
 	return true;
+
+error:
+	if (r && r->pager) delete_pager(r->pager);
+	if (r) free(r);
+	return false;
 }
 
 bool mmap_file(process_t *proc, fs_handle_t *h, size_t offset, void* addr, size_t size, int mode) {
@@ -527,34 +533,38 @@ bool mmap_file(process_t *proc, fs_handle_t *h, size_t offset, void* addr, size_
 	if (!(fmode & FM_MMAP) || !(fmode & FM_READ)) return false;
 	if ((mode & MM_WRITE) && !(fmode & FM_WRITE)) return false;
 
+	ASSERT(h->node->pager != 0);
+
+	if (addr >= (void*)K_HIGHHALF_ADDR || addr + size > (void*)K_HIGHHALF_ADDR
+		|| addr + size <= addr || size == 0) return false;
+
 	user_region_t *r = (user_region_t*)malloc(sizeof(user_region_t));
 	if (r == 0) return false;
 
 	r->addr = addr;
-	r->size = PAGE_ALIGN_UP(size);
-
-	if (r->addr >= (void*)K_HIGHHALF_ADDR || r->addr + r->size > (void*)K_HIGHHALF_ADDR
-		|| r->addr + r->size <= r->addr || r->size == 0) {
-		free(r);
-		return false;
-	}
+	r->size = size;
 
 	bool add_ok = btree_add(proc->regions_idx, r->addr, r);
-	if (!add_ok) {
-		free(r);
-		return false;
-	}
+	if (!add_ok) goto error;
 
 	ref_file(h);
+	r->file = h;
 
 	r->mode = mode;
-	r->file = h;
-	r->file_offset = offset;
+	r->offset = offset;
+	r->pager = h->node->pager;
+	r->own_pager = false;
 
-	r->next = proc->regions;
+	pager_add_map(r->pager, r);
+
+	r->next_in_proc = proc->regions;
 	proc->regions = r;
 
 	return true;
+
+error:
+	if (r) free(r);
+	return false;
 }
 
 bool mchmap(process_t *proc, void* addr, int mode) {
@@ -592,11 +602,11 @@ bool munmap(process_t *proc, void* addr) {
 	if (r == 0) return false;
 
 	if (proc->regions == r) {
-		proc->regions = r->next;
+		proc->regions = r->next_in_proc;
 	} else {
-		for (user_region_t *it = proc->regions; it != 0; it = it->next) {
-			if (it->next == r) {
-				it->next = r->next;
+		for (user_region_t *it = proc->regions; it != 0; it = it->next_in_proc) {
+			if (it->next_in_proc == r) {
+				it->next_in_proc = r->next_in_proc;
 				break;
 			}
 		}
@@ -609,22 +619,20 @@ bool munmap(process_t *proc, void* addr) {
 	switch_pagedir(proc->pd);
 	for (void* it = r->addr; it < r->addr + r->size; it += PAGE_SIZE) {
 		uint32_t ent = pd_get_entry(it);
-		uint32_t frame = pd_get_frame(it);
 
 		if (ent & PTE_PRESENT) {
-			if ((ent & PTE_DIRTY) && (r->mode & MM_WRITE) && r->file != 0) {
-				// TODO COMMIT PAGE!!
-			}
+			pager_access(r->pager,
+				it - r->addr + r->offset, PAGE_SIZE,
+				(ent & PTE_ACCESSED), (ent & PTE_DIRTY));
 			pd_unmap_page(it);
-			if (r->file != 0) {
-				// frame is owned by process
-				frame_free(frame, 1);
-			}
 		}
 	}
 	switch_pagedir(save_pd);
 
+	pager_rm_map(r->pager, r);
+
 	if (r->file != 0) unref_file(r->file);
+	if (r->own_pager) delete_pager(r->pager);
 
 	free(r);
 
@@ -663,12 +671,7 @@ static void proc_usermem_pf(void* p, registers_t *regs, void* addr) {
 
 	uint32_t frame;
 	do {
-		if (r->file == 0) {
-			frame = frame_alloc(1);
-		} else {
-			PANIC("Not implemented mmap (AWFUL TODO)");
-			// Here we must get the page from the cache
-		}
+		frame = pager_get_frame(r->pager, addr - r->addr + r->offset);
 		if (frame == 0) {
 			free_some_memory();
 		}
@@ -677,8 +680,6 @@ static void proc_usermem_pf(void* p, registers_t *regs, void* addr) {
 	while(!pd_map_page(addr, frame, (r->mode & MM_WRITE) != 0)) {
 		free_some_memory();
 	}
-
-	if (r->file == 0) memset(addr, 0, PAGE_SIZE);		// zero out
 }
 
 void probe_for_read(const void* addr, size_t len) {