lib: fs: implement memory-mapped file i/o

lib/libfs/context.c

@@ -1,6 +1,7 @@
 #include "btree.h"
 #include "file.h"
 #include "interface.h"
+#include "mapping.h"
 
 #include <fs/allocator.h>
 #include <fs/context.h>
@@ -8,9 +9,16 @@
 #include <fs/inode.h>
 #include <fs/status.h>
 #include <fs/superblock.h>
+#include <mango/handle.h>
 #include <mango/log.h>
+#include <mango/object.h>
+#include <mango/signal.h>
+#include <mango/task.h>
+#include <mango/vm.h>
 #include <stdio.h>
 
+#define TEMP_OBJECT_SIZE 0x10000
+
 BTREE_DEFINE_SIMPLE_GET(struct fs_file, unsigned long, f_node, f_id, get_file);
 BTREE_DEFINE_SIMPLE_INSERT(struct fs_file, f_node, f_id, put_file);
 
@@ -18,12 +26,21 @@ struct fs_context {
 	struct fs_superblock *ctx_sb;
 	struct fs_allocator *ctx_alloc;
 	struct btree ctx_filelist;
+	kern_handle_t ctx_vm_controller;
+	kern_handle_t ctx_channel;
+	kern_handle_t ctx_temp_object;
+	void *ctx_temp_object_buf;
 
 	struct fs_vtable ctx_vtable;
 };
 
 struct fs_context *fs_context_create(struct fs_allocator *alloc)
 {
+	kern_handle_t self, address_space;
+	task_self(&self);
+	task_get_address_space(self, &address_space);
+	kern_handle_close(self);
+
 	struct fs_context *ctx = fs_alloc(alloc, sizeof *ctx);
 	if (!ctx) {
 		return NULL;
@@ -31,10 +48,48 @@ struct fs_context *fs_context_create(struct fs_allocator *alloc)
 
 	memset(ctx, 0x0, sizeof *ctx);
 
+	kern_status_t status = vm_controller_create(&ctx->ctx_vm_controller);
+	if (status != KERN_OK) {
+		fs_free(alloc, ctx);
+		return NULL;
+	}
+
+	status = vm_object_create(
+		NULL,
+		0,
+		TEMP_OBJECT_SIZE,
+		VM_PROT_READ | VM_PROT_WRITE | VM_PROT_USER,
+		&ctx->ctx_temp_object);
+	if (status != KERN_OK) {
+		kern_handle_close(ctx->ctx_vm_controller);
+		fs_free(alloc, ctx);
+		return NULL;
+	}
+
+	virt_addr_t temp_buffer;
+	status = address_space_map(
+		address_space,
+		MAP_ADDRESS_ANY,
+		ctx->ctx_temp_object,
+		0,
+		TEMP_OBJECT_SIZE,
+		VM_PROT_READ | VM_PROT_WRITE | VM_PROT_USER,
+		&temp_buffer);
+	if (status != KERN_OK) {
+		kern_handle_close(ctx->ctx_temp_object);
+		kern_handle_close(ctx->ctx_vm_controller);
+		fs_free(alloc, ctx);
+		return NULL;
+	}
+
+	ctx->ctx_temp_object_buf = (void *)temp_buffer;
 	ctx->ctx_alloc = alloc;
 
 	ctx->ctx_vtable.open = fs_msg_open;
+	ctx->ctx_vtable.close = fs_msg_close;
 	ctx->ctx_vtable.read = fs_msg_read;
+	ctx->ctx_vtable.write = fs_msg_write;
+	ctx->ctx_vtable.map = fs_msg_map;
 
 	return ctx;
 }
@@ -73,6 +128,112 @@ enum fs_status fs_context_unmount_filesystem(struct fs_context *ctx)
 	return FS_ERR_NOT_IMPLEMENTED;
 }
 
+void fs_context_set_channel(struct fs_context *ctx, kern_handle_t channel)
+{
+	ctx->ctx_channel = channel;
+}
+
+kern_handle_t fs_context_get_vm_controller(const struct fs_context *ctx)
+{
+	return ctx->ctx_vm_controller;
+}
+
+static enum fs_status handle_msg(struct fs_context *ctx)
+{
+	xpc_msg_t msg;
+	kern_status_t status = xpc_msg_recv_nowait(ctx->ctx_channel, &msg);
+	if (status == KERN_NO_ENTRY) {
+		return FS_SUCCESS;
+	}
+
+	if (status != KERN_OK) {
+		kern_logf("message recv error %d", status);
+		return FS_ERR_INTERNAL_FAILURE;
+	}
+
+	switch (msg.msg_header.hdr_interface) {
+	case INTERFACE_FS:
+		status = fs_context_dispatch_msg(ctx, &msg);
+		break;
+	default:
+		kern_logf(
+			"unknown message protocol %u",
+			msg.msg_header.hdr_interface);
+		xpc_msg_reply_error(&msg, KERN_UNSUPPORTED);
+		break;
+	}
+
+	return FS_SUCCESS;
+}
+
+static enum fs_status handle_page_request(struct fs_context *ctx)
+{
+	equeue_packet_page_request_t packet;
+	vm_controller_recv(ctx->ctx_vm_controller, &packet);
+	struct file_mapping *mapping = (struct file_mapping *)packet.req_vmo;
+	kern_logf(
+		"received page request [%zx-%zx] for file %s",
+		packet.req_offset,
+		packet.req_offset + packet.req_length,
+		mapping->m_file->f_dent->d_name);
+
+	size_t length = packet.req_length;
+	if (length > TEMP_OBJECT_SIZE) {
+		length = TEMP_OBJECT_SIZE;
+	}
+
+	char *dst = ctx->ctx_temp_object_buf;
+	xpc_buffer_t buf = XPC_LOCAL_BUFFER_OUT(dst, TEMP_OBJECT_SIZE);
+	enum fs_status status = fs_file_read_at(
+		mapping->m_file,
+		&buf,
+		mapping->m_file_offset + packet.req_offset,
+		length);
+	if (status != FS_SUCCESS) {
+		kern_logf("map-read failed with code %d", status);
+		return status;
+	}
+
+	vm_controller_supply_pages(
+		ctx->ctx_vm_controller,
+		mapping->m_vmo,
+		packet.req_offset,
+		ctx->ctx_temp_object,
+		0,
+		packet.req_length);
+	return FS_SUCCESS;
+}
+
+enum fs_status fs_context_handle_request(struct fs_context *ctx)
+{
+	kern_wait_item_t waiters[] = {
+		{
+			.w_handle = ctx->ctx_channel,
+			.w_waitfor = CHANNEL_SIGNAL_MSG_RECEIVED,
+		},
+		{
+			.w_handle = ctx->ctx_vm_controller,
+			.w_waitfor = VM_CONTROLLER_SIGNAL_REQUEST_RECEIVED,
+		},
+	};
+	const size_t nr_waiters = sizeof waiters / sizeof waiters[0];
+
+	kern_status_t kstatus = kern_object_wait(waiters, nr_waiters);
+	if (kstatus != KERN_OK) {
+		return FS_ERR_INTERNAL_FAILURE;
+	}
+
+	if (waiters[0].w_observed & CHANNEL_SIGNAL_MSG_RECEIVED) {
+		return handle_msg(ctx);
+	}
+
+	if (waiters[1].w_observed & VM_CONTROLLER_SIGNAL_REQUEST_RECEIVED) {
+		return handle_page_request(ctx);
+	}
+
+	return FS_SUCCESS;
+}
+
 struct fs_file *fs_context_open_file(struct fs_context *ctx, unsigned long id)
 {
 	struct fs_file *f = get_file(&ctx->ctx_filelist, id);

lib/libfs/file.c

@@ -25,9 +25,22 @@ enum fs_status fs_file_read(
 	return status;
 }
 
-enum fs_status fs_file_write(
+enum fs_status fs_file_read_at(
 	struct fs_file *f,
 	struct xpc_buffer *buf,
+	off_t offset,
+	size_t count)
+{
+	if (!f->f_ops || !f->f_ops->f_read) {
+		return FS_ERR_NOT_IMPLEMENTED;
+	}
+
+	return f->f_ops->f_read(f, buf, count, &offset);
+}
+
+enum fs_status fs_file_write(
+	struct fs_file *f,
+	const struct xpc_buffer *buf,
 	size_t count)
 {
 	if (!f->f_ops || !f->f_ops->f_write) {

lib/libfs/file.h

@@ -2,6 +2,7 @@
 #define _FS_FILE_H_
 
 #include "btree.h"
+#include "queue.h"
 
 #include <fs/dentry.h>
 #include <fs/file.h>
@@ -18,6 +19,8 @@ struct fs_file {
 	off_t f_seek;
 	struct fs_inode *f_inode;
 	struct fs_dentry *f_dent;
+
+	struct queue f_mappings;
 };
 
 #endif

lib/libfs/include/fs/context.h

@@ -30,6 +30,12 @@ extern enum fs_status fs_context_mount_filesystem(
 	enum fs_mount_flags flags);
 extern enum fs_status fs_context_unmount_filesystem(struct fs_context *ctx);
 
+extern void fs_context_set_channel(
+	struct fs_context *ctx,
+	kern_handle_t channel);
+extern kern_handle_t fs_context_get_vm_controller(const struct fs_context *ctx);
+extern enum fs_status fs_context_handle_request(struct fs_context *ctx);
+
 extern struct fs_file *fs_context_open_file(
 	struct fs_context *ctx,
 	unsigned long id);

lib/libfs/include/fs/file.h

@@ -27,9 +27,14 @@ extern enum fs_status fs_file_read(
 	struct fs_file *f,
 	struct xpc_buffer *buf,
 	size_t count);
-extern enum fs_status fs_file_write(
+extern enum fs_status fs_file_read_at(
 	struct fs_file *f,
 	struct xpc_buffer *buf,
+	off_t offset,
+	size_t count);
+extern enum fs_status fs_file_write(
+	struct fs_file *f,
+	const struct xpc_buffer *buf,
 	size_t count);
 
 #endif

lib/libfs/interface.h

@@ -17,9 +17,8 @@ extern kern_status_t fs_msg_open(
 	int *out_err,
 	void *arg);
 extern kern_status_t fs_msg_close(
-	const struct msg_endpoint *sender,
+	xpc_context_t *ctx,
-	const char *path,
+	const xpc_endpoint_t *sender,
-	int flags,
 	int *out_err,
 	void *arg);
 
@@ -31,5 +30,21 @@ extern kern_status_t fs_msg_read(
 	size_t *out_nr_read,
 	xpc_buffer_t *out_data,
 	void *arg);
+extern kern_status_t fs_msg_write(
+	xpc_context_t *ctx,
+	const xpc_endpoint_t *sender,
+	const xpc_buffer_t *data,
+	int *out_err,
+	size_t *out_nr_written,
+	void *arg);
+
+extern kern_status_t fs_msg_map(
+	xpc_context_t *ctx,
+	const xpc_endpoint_t *sender,
+	int prot,
+	int flags,
+	int *out_err,
+	kern_handle_t *out_vmo,
+	void *arg);
 
 #endif

lib/libfs/interface/close.c

@@ -0,0 +1,14 @@
+#include "../interface.h"
+
+#include <errno.h>
+#include <mango/status.h>
+
+extern kern_status_t fs_msg_close(
+	xpc_context_t *ctx,
+	const xpc_endpoint_t *sender,
+	int *out_err,
+	void *arg)
+{
+	*out_err = ENOSYS;
+	return KERN_OK;
+}

lib/libfs/interface/map.c

@@ -0,0 +1,74 @@
+#include "../file.h"
+#include "../mapping.h"
+
+#include <errno.h>
+#include <fs/context.h>
+#include <fs/file.h>
+#include <fs/status.h>
+#include <mango/handle.h>
+#include <mango/log.h>
+#include <mango/vm.h>
+#include <stdio.h>
+#include <sys/mman.h>
+
+extern kern_status_t fs_msg_map(
+	xpc_context_t *xpc,
+	const xpc_endpoint_t *sender,
+	int prot,
+	int flags,
+	int *out_err,
+	kern_handle_t *out_vmo,
+	void *arg)
+{
+	struct fs_context *ctx = arg;
+	struct fs_file *f = fs_context_get_file(ctx, sender->e_port);
+	if (!f) {
+		*out_err = EBADF;
+		return KERN_OK;
+	}
+
+	struct file_mapping *mapping = fs_context_alloc(ctx, sizeof *mapping);
+	if (!mapping) {
+		*out_err = ENOMEM;
+		return KERN_OK;
+	}
+
+	kern_logf("mapping file %s", f->f_dent->d_name);
+	vm_prot_t vm_prot = VM_PROT_USER;
+	if (prot & PROT_READ) {
+		vm_prot |= VM_PROT_READ;
+	}
+
+	if (prot & PROT_WRITE) {
+		vm_prot |= VM_PROT_WRITE;
+	}
+
+	if (prot & PROT_EXEC) {
+		vm_prot |= VM_PROT_EXEC;
+	}
+
+	kern_handle_t vmo = KERN_HANDLE_INVALID;
+	kern_status_t status = vm_controller_create_object(
+		fs_context_get_vm_controller(ctx),
+		f->f_dent->d_name,
+		strlen(f->f_dent->d_name),
+		(equeue_key_t)mapping,
+		f->f_inode->i_size,
+		vm_prot,
+		&vmo);
+
+	if (status != KERN_OK) {
+		fs_context_free(ctx, mapping);
+		*out_err = __errno_from_kern_status(status);
+		return KERN_OK;
+	}
+
+	mapping->m_file = f;
+	mapping->m_file_offset = 0;
+	kern_handle_duplicate(vmo, &mapping->m_vmo);
+	queue_push_back(&f->f_mappings, &mapping->m_entry);
+
+	*out_err = SUCCESS;
+	*out_vmo = vmo;
+	return KERN_OK;
+}

lib/libfs/interface/write.c

@@ -0,0 +1,29 @@
+#include <errno.h>
+#include <fs/context.h>
+#include <fs/file.h>
+#include <fs/status.h>
+
+extern kern_status_t fs_msg_write(
+	xpc_context_t *xpc,
+	const xpc_endpoint_t *sender,
+	const xpc_buffer_t *data,
+	int *out_err,
+	size_t *out_nr_written,
+	void *arg)
+{
+	struct fs_context *ctx = arg;
+	struct fs_file *f = fs_context_get_file(ctx, sender->e_port);
+	if (!f) {
+		*out_err = EBADF;
+		return KERN_OK;
+	}
+
+	size_t start = fs_file_get_cursor(f);
+	enum fs_status status = fs_file_write(f, data, data->buf_len);
+	size_t end = fs_file_get_cursor(f);
+
+	*out_err = fs_status_to_errno(status);
+	*out_nr_written = end - start;
+
+	return KERN_OK;
+}

lib/libfs/mapping.h

@@ -0,0 +1,17 @@
+#ifndef MAPPING_H_
+#define MAPPING_H_
+
+#include "queue.h"
+
+#include <mango/types.h>
+
+struct fs_file;
+
+struct file_mapping {
+	struct fs_file *m_file;
+	kern_handle_t m_vmo;
+	off_t m_file_offset;
+	struct queue_entry m_entry;
+};
+
+#endif

lib/libfs/queue.c

@@ -0,0 +1,138 @@
+#include "queue.h"
+
+size_t queue_length(struct queue *q)
+{
+	size_t i = 0;
+	struct queue_entry *x = q->q_first;
+	while (x) {
+		i++;
+		x = x->qe_next;
+	}
+
+	return i;
+}
+
+void queue_insert_before(
+	struct queue *q,
+	struct queue_entry *entry,
+	struct queue_entry *before)
+{
+	struct queue_entry *x = before->qe_prev;
+	if (x) {
+		x->qe_next = entry;
+	} else {
+		q->q_first = entry;
+	}
+
+	entry->qe_prev = x;
+
+	before->qe_prev = entry;
+	entry->qe_next = before;
+}
+
+void queue_insert_after(
+	struct queue *q,
+	struct queue_entry *entry,
+	struct queue_entry *after)
+{
+	struct queue_entry *x = after->qe_next;
+	if (x) {
+		x->qe_prev = entry;
+	} else {
+		q->q_last = entry;
+	}
+
+	entry->qe_prev = x;
+
+	after->qe_next = entry;
+	entry->qe_prev = after;
+}
+
+void queue_push_front(struct queue *q, struct queue_entry *entry)
+{
+	if (q->q_first) {
+		q->q_first->qe_prev = entry;
+	}
+
+	entry->qe_next = q->q_first;
+	entry->qe_prev = NULL;
+
+	q->q_first = entry;
+
+	if (!q->q_last) {
+		q->q_last = entry;
+	}
+}
+
+void queue_push_back(struct queue *q, struct queue_entry *entry)
+{
+	if (q->q_last) {
+		q->q_last->qe_next = entry;
+	}
+
+	entry->qe_prev = q->q_last;
+	entry->qe_next = NULL;
+
+	q->q_last = entry;
+
+	if (!q->q_first) {
+		q->q_first = entry;
+	}
+}
+
+struct queue_entry *queue_pop_front(struct queue *q)
+{
+	struct queue_entry *x = q->q_first;
+	if (x) {
+		queue_delete(q, x);
+	}
+
+	return x;
+}
+
+struct queue_entry *queue_pop_back(struct queue *q)
+{
+	struct queue_entry *x = q->q_last;
+	if (x) {
+		queue_delete(q, x);
+	}
+
+	return x;
+}
+
+void queue_delete(struct queue *q, struct queue_entry *entry)
+{
+	if (!entry) {
+		return;
+	}
+
+	if (entry == q->q_first) {
+		q->q_first = q->q_first->qe_next;
+	}
+
+	if (entry == q->q_last) {
+		q->q_last = q->q_last->qe_prev;
+	}
+
+	if (entry->qe_next) {
+		entry->qe_next->qe_prev = entry->qe_prev;
+	}
+
+	if (entry->qe_prev) {
+		entry->qe_prev->qe_next = entry->qe_next;
+	}
+
+	entry->qe_next = entry->qe_prev = NULL;
+}
+
+void queue_delete_all(struct queue *q)
+{
+	struct queue_entry *x = q->q_first;
+	while (x) {
+		struct queue_entry *next = x->qe_next;
+		x->qe_next = x->qe_prev = NULL;
+		x = next;
+	}
+
+	q->q_first = q->q_last = NULL;
+}

lib/libfs/queue.h

@@ -0,0 +1,100 @@
+#ifndef QUEUE_H_
+#define QUEUE_H_
+
+#include <stdbool.h>
+#include <string.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define QUEUE_CONTAINER(t, m, v)                                               \
+	((void *)((v) ? (uintptr_t)(v) - (offsetof(t, m)) : 0))
+
+#define QUEUE_INIT ((struct queue) {.q_first = NULL, .q_last = NULL})
+#define QUEUE_ENTRY_INIT                                                       \
+	((struct queue_entry) {.qe_next = NULL, .qe_prev = NULL})
+
+#define queue_foreach(iter_type, iter_name, queue_name, node_member)           \
+	for (iter_type *iter_name = (iter_type *)QUEUE_CONTAINER(              \
+		     iter_type,                                                \
+		     node_member,                                              \
+		     queue_first(queue_name));                                 \
+	     iter_name;                                                        \
+	     iter_name = (iter_type *)QUEUE_CONTAINER(                         \
+		     iter_type,                                                \
+		     node_member,                                              \
+		     queue_next(&((iter_name)->node_member))))
+
+#define queue_foreach_r(iter_type, iter_name, queue_name, node_member)         \
+	for (iter_type *iter_name = (iter_type *)QUEUE_CONTAINER(              \
+		     iter_type,                                                \
+		     node_member,                                              \
+		     queue_last(queue_name));                                  \
+	     iter_name;                                                        \
+	     iter_name = (iter_type *)QUEUE_CONTAINER(                         \
+		     iter_type,                                                \
+		     node_member,                                              \
+		     queue_prev(&((iter_name)->node_member))))
+
+struct queue_entry {
+	struct queue_entry *qe_next;
+	struct queue_entry *qe_prev;
+};
+
+struct queue {
+	struct queue_entry *q_first;
+	struct queue_entry *q_last;
+};
+
+static inline void queue_init(struct queue *q)
+{
+	memset(q, 0x00, sizeof *q);
+}
+static inline bool queue_empty(struct queue *q)
+{
+	return q->q_first == NULL;
+}
+
+static inline struct queue_entry *queue_first(struct queue *q)
+{
+	return q->q_first;
+}
+static inline struct queue_entry *queue_last(struct queue *q)
+{
+	return q->q_last;
+}
+static inline struct queue_entry *queue_next(struct queue_entry *entry)
+{
+	return entry->qe_next;
+}
+static inline struct queue_entry *queue_prev(struct queue_entry *entry)
+{
+	return entry->qe_prev;
+}
+
+extern size_t queue_length(struct queue *q);
+
+extern void queue_insert_before(
+	struct queue *q,
+	struct queue_entry *entry,
+	struct queue_entry *before);
+extern void queue_insert_after(
+	struct queue *q,
+	struct queue_entry *entry,
+	struct queue_entry *after);
+
+extern void queue_push_front(struct queue *q, struct queue_entry *entry);
+extern void queue_push_back(struct queue *q, struct queue_entry *entry);
+
+extern struct queue_entry *queue_pop_front(struct queue *q);
+extern struct queue_entry *queue_pop_back(struct queue *q);
+
+extern void queue_delete(struct queue *q, struct queue_entry *entry);
+extern void queue_delete_all(struct queue *q);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif