#include <slab.h>
#include <kmalloc.h>
#include <kfs.h>
+#include <ext2fs.h>
#include <pmap.h>
#include <umem.h>
#include <smp.h>
spinlock_t super_blocks_lock = SPINLOCK_INITIALIZER;
struct fs_type_tailq file_systems = TAILQ_HEAD_INITIALIZER(file_systems);
struct namespace default_ns;
-// TODO: temp dcache, holds all dentries ever for now
-struct dentry_slist dcache = SLIST_HEAD_INITIALIZER(dcache);
-spinlock_t dcache_lock = SPINLOCK_INITIALIZER;
struct kmem_cache *dentry_kcache; // not to be confused with the dcache
struct kmem_cache *inode_kcache;
* with multiple namespaces on the same FS yet. Note if you mount the same FS
* multiple times, you only have one FS still (and one SB). If we ever support
* that... */
-struct vfsmount *mount_fs(struct fs_type *fs, char *dev_name,
- struct dentry *mnt_pt, int flags,
- struct namespace *ns)
+struct vfsmount *__mount_fs(struct fs_type *fs, char *dev_name,
+ struct dentry *mnt_pt, int flags,
+ struct namespace *ns)
{
struct super_block *sb;
struct vfsmount *vmnt = kmalloc(sizeof(struct vfsmount), 0);
/* this first ref is stored in the NS tailq below */
kref_init(&vmnt->mnt_kref, fake_release, 1);
- /* Build the vfsmount, if there is no mnt_pt, mnt is the root vfsmount (for now).
- * fields related to the actual FS, like the sb and the mnt_root are set in
- * the fs-specific get_sb() call. */
+ /* Build the vfsmount, if there is no mnt_pt, mnt is the root vfsmount (for
+ * now). fields related to the actual FS, like the sb and the mnt_root are
+ * set in the fs-specific get_sb() call. */
if (!mnt_pt) {
vmnt->mnt_parent = NULL;
vmnt->mnt_mountpoint = NULL;
/* build list of all FS's in the system. put yours here. if this is ever
* done on the fly, we'll need to lock. */
TAILQ_INSERT_TAIL(&file_systems, &kfs_fs_type, list);
+#ifdef CONFIG_EXT2FS
+ TAILQ_INSERT_TAIL(&file_systems, &ext2_fs_type, list);
+#endif
TAILQ_FOREACH(fs, &file_systems, list)
printk("Supports the %s Filesystem\n", fs->name);
/* mounting KFS at the root (/), pending root= parameters */
// TODO: linux creates a temp root_fs, then mounts the real root onto that
- default_ns.root = mount_fs(&kfs_fs_type, "RAM", NULL, 0, &default_ns);
+ default_ns.root = __mount_fs(&kfs_fs_type, "RAM", NULL, 0, &default_ns);
printk("vfs_init() completed\n");
}
+/* FS's can provide another, if they want */
+int generic_dentry_hash(struct dentry *dentry, struct qstr *qstr)
+{
+ unsigned long hash = 5381;
+
+ for (int i = 0; i < qstr->len; i++) {
+ /* hash * 33 + c, djb2's technique */
+ hash = ((hash << 5) + hash) + qstr->name[i];
+ }
+ return hash;
+}
+
/* Builds / populates the qstr of a dentry based on its d_iname. If there is an
* l_name, (long), it will use that instead of the inline name. This will
* probably change a bit. */
void qstr_builder(struct dentry *dentry, char *l_name)
{
dentry->d_name.name = l_name ? l_name : dentry->d_iname;
- // TODO: pending what we actually do in d_hash
- //dentry->d_name.hash = dentry->d_op->d_hash(dentry, &dentry->d_name);
- dentry->d_name.hash = 0xcafebabe;
dentry->d_name.len = strnlen(dentry->d_name.name, MAX_FILENAME_SZ);
+ dentry->d_name.hash = dentry->d_op->d_hash(dentry, &dentry->d_name);
}
/* Useful little helper - return the string ptr for a given file */
return file->f_dentry->d_name.name;
}
-/* Some issues with this, coupled closely to fs_lookup. This assumes that
- * negative dentries are not returned (might differ from linux) */
+/* Some issues with this, coupled closely to fs_lookup.
+ *
+ * Note the use of __dentry_free, instead of kref_put. In those cases, we don't
+ * want to treat it like a kref and we have the only reference to it, so it is
+ * okay to do this. It makes dentry_release() easier too. */
static struct dentry *do_lookup(struct dentry *parent, char *name)
{
- struct dentry *dentry;
- /* TODO: look up in the dentry cache first */
- dentry = get_dentry(parent->d_sb, parent, name);
- dentry = parent->d_inode->i_op->lookup(parent->d_inode, dentry, 0);
- /* insert in dentry cache */
+ struct dentry *result, *query;
+ query = get_dentry(parent->d_sb, parent, name);
+ if (!query) {
+ warn("OOM in do_lookup(), probably wasn't expected\n");
+ return 0;
+ }
+ result = dcache_get(parent->d_sb, query);
+ if (result) {
+ __dentry_free(query);
+ return result;
+ }
+ /* No result, check for negative */
+ if (query->d_flags & DENTRY_NEGATIVE) {
+ __dentry_free(query);
+ return 0;
+ }
+ /* not in the dcache at all, need to consult the FS */
+ result = parent->d_inode->i_op->lookup(parent->d_inode, query, 0);
+ if (!result) {
+ /* Note the USED flag will get turned off when this gets added to the
+ * LRU in dentry_release(). There's a slight race here that we'll panic
+ * on, but I want to catch it (in dcache_put()) for now. */
+ query->d_flags |= DENTRY_NEGATIVE;
+ dcache_put(parent->d_sb, query);
+ kref_put(&query->d_kref);
+ return 0;
+ }
+ dcache_put(parent->d_sb, result);
+ /* This is because KFS doesn't return the same dentry, but ext2 does. this
+ * is ugly and needs to be fixed. (TODO) */
+ if (result != query)
+ __dentry_free(query);
+
/* TODO: if the following are done by us, how do we know the i_ino?
* also need to handle inodes that are already read in! For now, we're
* going to have the FS handle it in it's lookup() method:
* - get a new inode
* - read in the inode
* - put in the inode cache */
- return dentry;
+ return result;
}
/* Update ND such that it represents having followed dentry. IAW the nd
printd("CLIMB_UP, from %s\n", nd->dentry->d_name.name);
/* Top of the world, just return. Should also check for being at the top of
* the current process's namespace (TODO) */
- if (!nd->dentry->d_parent)
+ if (!nd->dentry->d_parent || (nd->dentry->d_parent == nd->dentry))
return -1;
/* Check if we are at the top of a mount, if so, we need to follow
* backwards, and then climb_up from that one. We might need to climb
return 0;
}
+/* nd->dentry might be on a mount point, so we need to move on to the child
+ * mount's root. */
static int follow_mount(struct nameidata *nd)
{
- /* Detect mount, follow, etc... (TODO!) */
+ if (!nd->dentry->d_mount_point)
+ return 0;
+ next_link(nd->dentry->d_mounted_fs->mnt_root, nd);
return 0;
}
return 0;
if (nd->depth > MAX_SYMLINK_DEPTH)
return -ELOOP;
- printd("Following symlink for dentry %08p %s\n", nd->dentry,
+ printd("Following symlink for dentry %p %s\n", nd->dentry,
nd->dentry->d_name.name);
nd->depth++;
symname = nd->dentry->d_inode->i_op->readlink(nd->dentry);
int retval;
printd("Path lookup for %s\n", path);
/* we allow absolute lookups with no process context */
+ /* TODO: RCU read lock on pwd or kref_not_zero in a loop. concurrent chdir
+ * could decref nd->dentry before we get to incref it below. */
if (path[0] == '/') { /* absolute lookup */
if (!current)
nd->dentry = default_ns.root->mnt_root;
}
}
+/* External version of mount, only call this after having a / mount */
+int mount_fs(struct fs_type *fs, char *dev_name, char *path, int flags)
+{
+ struct nameidata nd_r = {0}, *nd = &nd_r;
+ int retval = 0;
+ retval = path_lookup(path, LOOKUP_DIRECTORY, nd);
+ if (retval)
+ goto out;
+ /* taking the namespace of the vfsmount of path */
+ if (!__mount_fs(fs, dev_name, nd->dentry, flags, nd->mnt->mnt_namespace))
+ retval = -EINVAL;
+out:
+ path_release(nd);
+ return retval;
+}
+
/* Superblock functions */
+/* Dentry "hash" function for the hash table to use. Since we already have the
+ * hash in the qstr, we don't need to rehash. Also, note we'll be using the
+ * dentry in question as both the key and the value. */
+static size_t __dcache_hash(void *k)
+{
+ return (size_t)((struct dentry*)k)->d_name.hash;
+}
+
+/* Dentry cache hashtable equality function. This means we need to pass in some
+ * minimal dentry when doing a lookup. */
+static ssize_t __dcache_eq(void *k1, void *k2)
+{
+ if (((struct dentry*)k1)->d_parent != ((struct dentry*)k2)->d_parent)
+ return 0;
+ /* TODO: use the FS-specific string comparison */
+ return !strcmp(((struct dentry*)k1)->d_name.name,
+ ((struct dentry*)k2)->d_name.name);
+}
+
/* Helper to alloc and initialize a generic superblock. This handles all the
* VFS related things, like lists. Each FS will need to handle its own things
* in it's *_get_sb(), usually involving reading off the disc. */
TAILQ_INIT(&sb->s_inodes);
TAILQ_INIT(&sb->s_dirty_i);
TAILQ_INIT(&sb->s_io_wb);
- SLIST_INIT(&sb->s_anon_d);
+ TAILQ_INIT(&sb->s_lru_d);
TAILQ_INIT(&sb->s_files);
+ sb->s_dcache = create_hashtable(100, __dcache_hash, __dcache_eq);
+ sb->s_icache = create_hashtable(100, __generic_hash, __generic_eq);
+ spinlock_init(&sb->s_lru_lock);
+ spinlock_init(&sb->s_dcache_lock);
+ spinlock_init(&sb->s_icache_lock);
sb->s_fs_info = 0; // can override somewhere else
return sb;
}
{
/* Build and init the first dentry / inode. The dentry ref is stored later
* by vfsmount's mnt_root. The parent is dealt with later. */
- struct dentry *d_root = get_dentry(sb, 0, "/"); /* probably right */
+ struct dentry *d_root = get_dentry_with_ops(sb, 0, "/", d_op);
+ if (!d_root)
+ panic("OOM! init_sb() can't fail yet!");
/* a lot of here on down is normally done in lookup() or create, since
* get_dentry isn't a fully usable dentry. The two FS-specific settings are
* normally inherited from a parent within the same FS in get_dentry, but we
/* TODO: do we need to read in the inode? can we do this on demand? */
/* if this FS is already mounted, we'll need to do something different. */
sb->s_op->read_inode(inode);
+ icache_put(sb, inode);
/* Link the dentry and SB to the VFS mount */
vmnt->mnt_root = d_root; /* ref comes from get_dentry */
vmnt->mnt_sb = sb;
if (vmnt->mnt_mountpoint) {
kref_get(&vmnt->mnt_mountpoint->d_kref, 1); /* held by d_root */
d_root->d_parent = vmnt->mnt_mountpoint; /* dentry of the root */
+ } else {
+ d_root->d_parent = d_root; /* set root as its own parent */
}
/* insert the dentry into the dentry cache. when's the earliest we can?
* when's the earliest we should? what about concurrent accesses to the
* same dentry? should be locking the dentry... */
- dcache_put(d_root); // TODO: should set a d_flag too
+ dcache_put(sb, d_root);
kref_put(&inode->i_kref); /* give up the ref from get_inode() */
}
/* Dentry Functions */
-/* Helper to alloc and initialize a generic dentry. The following needs to be
- * set still: d_op (if no parent), d_fs_info (opt), d_inode, connect the inode
- * to the dentry (and up the d_kref again), maybe dcache_put(). The inode
- * stitching is done in get_inode() or lookup (depending on the FS).
- * The setting of the d_op might be problematic when dealing with mounts. Just
- * overwrite it.
- *
- * If the name is longer than the inline name, it will kmalloc a buffer, so
- * don't worry about the storage for *name after calling this. */
-struct dentry *get_dentry(struct super_block *sb, struct dentry *parent,
- char *name)
+static void dentry_set_name(struct dentry *dentry, char *name)
{
- assert(name);
size_t name_len = strnlen(name, MAX_FILENAME_SZ); /* not including \0! */
- struct dentry *dentry = kmem_cache_alloc(dentry_kcache, 0);
char *l_name = 0;
+ if (name_len < DNAME_INLINE_LEN) {
+ strncpy(dentry->d_iname, name, name_len);
+ dentry->d_iname[name_len] = '\0';
+ qstr_builder(dentry, 0);
+ } else {
+ l_name = kmalloc(name_len + 1, 0);
+ assert(l_name);
+ strncpy(l_name, name, name_len);
+ l_name[name_len] = '\0';
+ qstr_builder(dentry, l_name);
+ }
+}
- if (!dentry)
+/* Gets a dentry. If there is no parent, use d_op. Only called directly by
+ * superblock init code. */
+struct dentry *get_dentry_with_ops(struct super_block *sb,
+ struct dentry *parent, char *name,
+ struct dentry_operations *d_op)
+{
+ assert(name);
+ struct dentry *dentry = kmem_cache_alloc(dentry_kcache, 0);
+
+ if (!dentry) {
+ set_errno(ENOMEM);
return 0;
+ }
//memset(dentry, 0, sizeof(struct dentry));
kref_init(&dentry->d_kref, dentry_release, 1); /* this ref is returned */
spinlock_init(&dentry->d_lock);
if (parent) { /* no parent for rootfs mount */
kref_get(&parent->d_kref, 1);
dentry->d_op = parent->d_op; /* d_op set in init_sb for parentless */
+ } else {
+ dentry->d_op = d_op;
}
dentry->d_parent = parent;
- dentry->d_flags = 0; /* related to its dcache state */
+ dentry->d_flags = DENTRY_USED;
dentry->d_fs_info = 0;
- SLIST_INIT(&dentry->d_bucket);
- if (name_len < DNAME_INLINE_LEN) {
- strncpy(dentry->d_iname, name, name_len);
- dentry->d_iname[name_len] = '\0';
- qstr_builder(dentry, 0);
- } else {
- l_name = kmalloc(name_len + 1, 0);
- assert(l_name);
- strncpy(l_name, name, name_len);
- l_name[name_len] = '\0';
- qstr_builder(dentry, l_name);
- }
+ dentry_set_name(dentry, name);
/* Catch bugs by aggressively zeroing this (o/w we use old stuff) */
dentry->d_inode = 0;
return dentry;
}
-/* Adds a dentry to the dcache. */
-void dcache_put(struct dentry *dentry)
+/* Helper to alloc and initialize a generic dentry. The following needs to be
+ * set still: d_op (if no parent), d_fs_info (opt), d_inode, connect the inode
+ * to the dentry (and up the d_kref again), maybe dcache_put(). The inode
+ * stitching is done in get_inode() or lookup (depending on the FS).
+ * The setting of the d_op might be problematic when dealing with mounts. Just
+ * overwrite it.
+ *
+ * If the name is longer than the inline name, it will kmalloc a buffer, so
+ * don't worry about the storage for *name after calling this. */
+struct dentry *get_dentry(struct super_block *sb, struct dentry *parent,
+ char *name)
{
-#if 0 /* pending a more thorough review of the dcache */
- /* TODO: should set a d_flag too */
- spin_lock(&dcache_lock);
- SLIST_INSERT_HEAD(&dcache, dentry, d_hash);
- spin_unlock(&dcache_lock);
-#endif
+ return get_dentry_with_ops(sb, parent, name, 0);
+}
+
+/* Called when the dentry is unreferenced (after kref == 0). This works closely
+ * with the resurrection in dcache_get().
+ *
+ * The dentry is still in the dcache, but needs to be un-USED and added to the
+ * LRU dentry list. Even dentries that were used in a failed lookup need to be
+ * cached - they ought to be the negative dentries. Note that all dentries have
+ * parents, even negative ones (it is needed to find it in the dcache). */
+void dentry_release(struct kref *kref)
+{
+ struct dentry *dentry = container_of(kref, struct dentry, d_kref);
+
+ printd("'Releasing' dentry %p: %s\n", dentry, dentry->d_name.name);
+ /* DYING dentries (recently unlinked / rmdir'd) just get freed */
+ if (dentry->d_flags & DENTRY_DYING) {
+ __dentry_free(dentry);
+ return;
+ }
+ /* This lock ensures the USED state and the TAILQ membership is in sync.
+ * Also used to check the refcnt, though that might not be necessary. */
+ spin_lock(&dentry->d_lock);
+ /* While locked, we need to double check the kref, in case someone already
+ * reup'd it. Re-up? you're crazy! Reee-up, you're outta yo mind! */
+ if (!kref_refcnt(&dentry->d_kref)) {
+ /* Note this is where negative dentries get set UNUSED */
+ if (dentry->d_flags & DENTRY_USED) {
+ dentry->d_flags &= ~DENTRY_USED;
+ spin_lock(&dentry->d_sb->s_lru_lock);
+ TAILQ_INSERT_TAIL(&dentry->d_sb->s_lru_d, dentry, d_lru);
+ spin_unlock(&dentry->d_sb->s_lru_lock);
+ } else {
+ /* and make sure it wasn't USED, then UNUSED again */
+ /* TODO: think about issues with this */
+ warn("This should be rare. Tell brho this happened.");
+ }
+ }
+ spin_unlock(&dentry->d_lock);
}
-/* Cleans up the dentry (after ref == 0). We still may want it, and this is
- * where we should add it to the dentry cache. (TODO). For now, we do nothing,
- * since we don't have a dcache. Also, if i_nlink == 0, never cache it.
- *
+/* Called when we really dealloc and get rid of a dentry (like when it is
+ * removed from the dcache, either for memory or correctness reasons)
+ *
* This has to handle two types of dentries: full ones (ones that had been used)
* and ones that had been just for lookups - hence the check for d_inode.
*
* Note that dentries pin and kref their inodes. When all the dentries are
* gone, we want the inode to be released via kref. The inode has internal /
* weak references to the dentry, which are not refcounted. */
-void dentry_release(struct kref *kref)
+void __dentry_free(struct dentry *dentry)
{
- struct dentry *dentry = container_of(kref, struct dentry, d_kref);
- printd("Freeing dentry %08p: %s\n", dentry, dentry->d_name.name);
+ if (dentry->d_inode)
+ printd("Freeing dentry %p: %s\n", dentry, dentry->d_name.name);
assert(dentry->d_op); /* catch bugs. a while back, some lacked d_op */
dentry->d_op->d_release(dentry);
/* TODO: check/test the boundaries on this. */
return dentry;
}
+/* Get a dentry from the dcache. At a minimum, we need the name hash and parent
+ * in what_i_want, though most uses will probably be from a get_dentry() call.
+ * We pass in the SB in the off chance that we don't want to use a get'd dentry.
+ *
+ * The unusual variable name (instead of just "key" or something) is named after
+ * ex-SPC Castro's porn folder. Caller deals with the memory for what_i_want.
+ *
+ * If the dentry is negative, we don't return the actual result - instead, we
+ * set the negative flag in 'what i want'. The reason is we don't want to
+ * kref_get() and then immediately put (causing dentry_release()). This also
+ * means that dentry_release() should never get someone who wasn't USED (barring
+ * the race, which it handles). And we don't need to ever have a dentry set as
+ * USED and NEGATIVE (which is always wrong, but would be needed for a cleaner
+ * dentry_release()).
+ *
+ * This is where we do the "kref resurrection" - we are returning a kref'd
+ * object, even if it wasn't kref'd before. This means the dcache does NOT hold
+ * krefs (it is a weak/internal ref), but it is a source of kref generation. We
+ * sync up with the possible freeing of the dentry by locking the table. See
+ * Doc/kref for more info. */
+struct dentry *dcache_get(struct super_block *sb, struct dentry *what_i_want)
+{
+ struct dentry *found;
+ /* This lock protects the hash, as well as ensures the returned object
+ * doesn't get deleted/freed out from under us */
+ spin_lock(&sb->s_dcache_lock);
+ found = hashtable_search(sb->s_dcache, what_i_want);
+ if (found) {
+ if (found->d_flags & DENTRY_NEGATIVE) {
+ what_i_want->d_flags |= DENTRY_NEGATIVE;
+ spin_unlock(&sb->s_dcache_lock);
+ return 0;
+ }
+ spin_lock(&found->d_lock);
+ __kref_get(&found->d_kref, 1); /* prob could be done outside the lock*/
+ /* If we're here (after kreffing) and it is not USED, we are the one who
+ * should resurrect */
+ if (!(found->d_flags & DENTRY_USED)) {
+ found->d_flags |= DENTRY_USED;
+ spin_lock(&sb->s_lru_lock);
+ TAILQ_REMOVE(&sb->s_lru_d, found, d_lru);
+ spin_unlock(&sb->s_lru_lock);
+ }
+ spin_unlock(&found->d_lock);
+ }
+ spin_unlock(&sb->s_dcache_lock);
+ return found;
+}
+
+/* Adds a dentry to the dcache. Note the *dentry is both the key and the value.
+ * If the value was already in there (which can happen iff it was negative), for
+ * now we'll remove it and put the new one in there. */
+void dcache_put(struct super_block *sb, struct dentry *key_val)
+{
+ struct dentry *old;
+ int retval;
+ spin_lock(&sb->s_dcache_lock);
+ old = hashtable_remove(sb->s_dcache, key_val);
+ /* if it is old and non-negative, our caller lost a race with someone else
+ * adding the dentry. but since we yanked it out, like a bunch of idiots,
+ * we still have to put it back. should be fairly rare. */
+ if (old && (old->d_flags & DENTRY_NEGATIVE)) {
+ /* This is possible, but rare for now (about to be put on the LRU) */
+ assert(!(old->d_flags & DENTRY_USED));
+ assert(!kref_refcnt(&old->d_kref));
+ spin_lock(&sb->s_lru_lock);
+ TAILQ_REMOVE(&sb->s_lru_d, old, d_lru);
+ spin_unlock(&sb->s_lru_lock);
+ /* TODO: this seems suspect. isn't this the same memory as key_val?
+ * in which case, we just adjust the flags (remove NEG) and reinsert? */
+ assert(old != key_val); // checking TODO comment
+ __dentry_free(old);
+ }
+ /* this returns 0 on failure (TODO: Fix this ghetto shit) */
+ retval = hashtable_insert(sb->s_dcache, key_val, key_val);
+ assert(retval);
+ spin_unlock(&sb->s_dcache_lock);
+}
+
+/* Will remove and return the dentry. Caller deallocs the key, but the retval
+ * won't have a reference. * Returns 0 if it wasn't found. Callers can't
+ * assume much - they should not use the reference they *get back*, (if they
+ * already had one for key, they can use that). There may be other users out
+ * there. */
+struct dentry *dcache_remove(struct super_block *sb, struct dentry *key)
+{
+ struct dentry *retval;
+ spin_lock(&sb->s_dcache_lock);
+ retval = hashtable_remove(sb->s_dcache, key);
+ spin_unlock(&sb->s_dcache_lock);
+ return retval;
+}
+
+/* This will clean out the LRU list, which are the unused dentries of the dentry
+ * cache. This will optionally only free the negative ones. Note that we grab
+ * the hash lock for the time we traverse the LRU list - this prevents someone
+ * from getting a kref from the dcache, which could cause us trouble (we rip
+ * someone off the list, who isn't unused, and they try to rip them off the
+ * list). */
+void dcache_prune(struct super_block *sb, bool negative_only)
+{
+ struct dentry *d_i, *temp;
+ struct dentry_tailq victims = TAILQ_HEAD_INITIALIZER(victims);
+
+ spin_lock(&sb->s_dcache_lock);
+ spin_lock(&sb->s_lru_lock);
+ TAILQ_FOREACH_SAFE(d_i, &sb->s_lru_d, d_lru, temp) {
+ if (!(d_i->d_flags & DENTRY_USED)) {
+ if (negative_only && !(d_i->d_flags & DENTRY_NEGATIVE))
+ continue;
+ /* another place where we'd be better off with tools, not sol'ns */
+ hashtable_remove(sb->s_dcache, d_i);
+ TAILQ_REMOVE(&sb->s_lru_d, d_i, d_lru);
+ TAILQ_INSERT_HEAD(&victims, d_i, d_lru);
+ }
+ }
+ spin_unlock(&sb->s_lru_lock);
+ spin_unlock(&sb->s_dcache_lock);
+ /* Now do the actual freeing, outside of the hash/LRU list locks. This is
+ * necessary since __dentry_free() will decref its parent, which may get
+ * released and try to add itself to the LRU. */
+ TAILQ_FOREACH_SAFE(d_i, &victims, d_lru, temp) {
+ TAILQ_REMOVE(&victims, d_i, d_lru);
+ assert(!kref_refcnt(&d_i->d_kref));
+ __dentry_free(d_i);
+ }
+ /* It is possible at this point that there are new items on the LRU. We
+ * could loop back until that list is empty, if we care about this. */
+}
+
/* Inode Functions */
/* Creates and initializes a new inode. Generic fields are filled in.
inode->i_ino = 0; /* set by caller later */
inode->i_blksize = sb->s_blocksize;
spinlock_init(&inode->i_lock);
+ kref_get(&sb->s_kref, 1); /* could allow the dentry to pin it */
inode->i_sb = sb;
+ inode->i_rdev = 0; /* this has no real meaning yet */
+ inode->i_bdev = sb->s_bdev; /* storing an uncounted ref */
inode->i_state = 0; /* need real states, like I_NEW */
inode->dirtied_when = 0;
inode->i_flags = 0;
atomic_set(&inode->i_writecount, 0);
/* Set up the page_map structures. Default is to use the embedded one.
* Might push some of this back into specific FSs. For now, the FS tells us
- * what pm_op they want via i_pm.pm_op, which we use when we point i_mapping
- * to i_pm. */
+ * what pm_op they want via i_pm.pm_op, which we set again in pm_init() */
inode->i_mapping = &inode->i_pm;
- inode->i_mapping->pm_host = inode;
- radix_tree_init(&inode->i_mapping->pm_tree);
- spinlock_init(&inode->i_mapping->pm_tree_lock);
- inode->i_mapping->pm_flags = 0;
+ pm_init(inode->i_mapping, inode->i_pm.pm_op, inode);
return inode;
}
+/* Helper: loads/ reads in the inode numbered ino and attaches it to dentry */
+void load_inode(struct dentry *dentry, unsigned long ino)
+{
+ struct inode *inode;
+
+ /* look it up in the inode cache first */
+ inode = icache_get(dentry->d_sb, ino);
+ if (inode) {
+ /* connect the dentry to its inode */
+ TAILQ_INSERT_TAIL(&inode->i_dentry, dentry, d_alias);
+ dentry->d_inode = inode; /* storing the ref we got from icache_get */
+ return;
+ }
+ /* otherwise, we need to do it manually */
+ inode = get_inode(dentry);
+ inode->i_ino = ino;
+ dentry->d_sb->s_op->read_inode(inode);
+ /* TODO: race here, two creators could miss in the cache, and then get here.
+ * need a way to sync across a blocking call. needs to be either at this
+ * point in the code or per the ino (dentries could be different) */
+ icache_put(dentry->d_sb, inode);
+ kref_put(&inode->i_kref);
+}
+
/* Helper op, used when creating regular files, directories, symlinks, etc.
* Note we make a distinction between the mode and the file type (for now).
* After calling this, call the FS specific version (create or mkdir), which
* note we don't pass this an nd, like Linux does... */
static struct inode *create_inode(struct dentry *dentry, int mode)
{
+ uint64_t now = epoch_seconds();
/* note it is the i_ino that uniquely identifies a file in the specific
* filesystem. there's a diff between creating an inode (even for an in-use
* ino) and then filling it in, and vs creating a brand new one.
inode->i_nlink = 1;
inode->i_size = 0;
inode->i_blocks = 0;
- inode->i_atime.tv_sec = 0; /* TODO: now! */
- inode->i_ctime.tv_sec = 0;
- inode->i_mtime.tv_sec = 0;
- inode->i_atime.tv_nsec = 0; /* are these supposed to be the extra ns? */
+ inode->i_atime.tv_sec = now;
+ inode->i_ctime.tv_sec = now;
+ inode->i_mtime.tv_sec = now;
+ inode->i_atime.tv_nsec = 0;
inode->i_ctime.tv_nsec = 0;
inode->i_mtime.tv_nsec = 0;
inode->i_bdev = inode->i_sb->s_bdev;
/* Create a new disk inode in dir associated with dentry, with the given mode.
* called when creating a regular file. dir is the directory/parent. dentry is
- * the dentry of the inode we are creating. Note the lack of the nd...
- * Also, we do the nlink++ in here, since we want to give the FS's a chance to
- * fail. */
+ * the dentry of the inode we are creating. Note the lack of the nd... */
int create_file(struct inode *dir, struct dentry *dentry, int mode)
{
struct inode *new_file = create_inode(dentry, mode);
if (!new_file)
return -1;
dir->i_op->create(dir, dentry, mode, 0);
- dir->i_nlink++;
+ icache_put(new_file->i_sb, new_file);
kref_put(&new_file->i_kref);
return 0;
}
if (!new_dir)
return -1;
dir->i_op->mkdir(dir, dentry, mode);
- dir->i_nlink++;
+ dir->i_nlink++; /* Directories get a hardlink for every child dir */
/* Make sure my parent tracks me. This is okay, since no directory (dir)
* can have more than one dentry */
struct dentry *parent = TAILQ_FIRST(&dir->i_dentry);
assert(parent && parent == TAILQ_LAST(&dir->i_dentry, dentry_tailq));
/* parent dentry tracks dentry as a subdir, weak reference */
TAILQ_INSERT_TAIL(&parent->d_subdirs, dentry, d_subdirs_link);
+ icache_put(new_dir->i_sb, new_dir);
kref_put(&new_dir->i_kref);
return 0;
}
if (!new_sym)
return -1;
dir->i_op->symlink(dir, dentry, symname);
- dir->i_nlink++; /* TODO: race with this, among other things */
+ icache_put(new_sym->i_sb, new_sym);
kref_put(&new_sym->i_kref);
return 0;
}
{
struct inode *inode = container_of(kref, struct inode, i_kref);
TAILQ_REMOVE(&inode->i_sb->s_inodes, inode, i_sb_list);
- /* If we still have links, just dealloc the in-memory inode. if we have no
- * links, we need to delete it too (which calls destroy). */
- if (inode->i_nlink)
- inode->i_sb->s_op->dealloc_inode(inode);
- else
+ icache_remove(inode->i_sb, inode->i_ino);
+ /* Might need to write back or delete the file/inode */
+ if (inode->i_nlink) {
+ if (inode->i_state & I_STATE_DIRTY)
+ inode->i_sb->s_op->write_inode(inode, TRUE);
+ } else {
inode->i_sb->s_op->delete_inode(inode);
+ }
+ if (S_ISFIFO(inode->i_mode)) {
+ page_decref(kva2page(inode->i_pipe->p_buf));
+ kfree(inode->i_pipe);
+ }
+ /* TODO: (BDEV) */
+ // kref_put(inode->i_bdev->kref); /* assuming it's a bdev, could be a pipe*/
+ /* Either way, we dealloc the in-memory version */
+ inode->i_sb->s_op->dealloc_inode(inode); /* FS-specific clean-up */
kref_put(&inode->i_sb->s_kref);
+ /* TODO: clean this up */
assert(inode->i_mapping == &inode->i_pm);
kmem_cache_free(inode_kcache, inode);
- /* TODO: (BDEV) */
- // kref_put(inode->i_bdev->kref); /* assuming it's a bdev */
}
/* Fills in kstat with the stat information for the inode */
kstat->st_ctime = inode->i_ctime;
}
+void print_kstat(struct kstat *kstat)
+{
+ printk("kstat info for %p:\n", kstat);
+ printk("\tst_dev : %p\n", kstat->st_dev);
+ printk("\tst_ino : %p\n", kstat->st_ino);
+ printk("\tst_mode : %p\n", kstat->st_mode);
+ printk("\tst_nlink : %p\n", kstat->st_nlink);
+ printk("\tst_uid : %p\n", kstat->st_uid);
+ printk("\tst_gid : %p\n", kstat->st_gid);
+ printk("\tst_rdev : %p\n", kstat->st_rdev);
+ printk("\tst_size : %p\n", kstat->st_size);
+ printk("\tst_blksize: %p\n", kstat->st_blksize);
+ printk("\tst_blocks : %p\n", kstat->st_blocks);
+ printk("\tst_atime : %p\n", kstat->st_atime);
+ printk("\tst_mtime : %p\n", kstat->st_mtime);
+ printk("\tst_ctime : %p\n", kstat->st_ctime);
+}
+
+/* Inode Cache management. In general, search on the ino, get a refcnt'd value
+ * back. Remove does not give you a reference back - it should only be called
+ * in inode_release(). */
+struct inode *icache_get(struct super_block *sb, unsigned long ino)
+{
+ /* This is the same style as in pid2proc, it's the "safely create a strong
+ * reference from a weak one, so long as other strong ones exist" pattern */
+ spin_lock(&sb->s_icache_lock);
+ struct inode *inode = hashtable_search(sb->s_icache, (void*)ino);
+ if (inode)
+ if (!kref_get_not_zero(&inode->i_kref, 1))
+ inode = 0;
+ spin_unlock(&sb->s_icache_lock);
+ return inode;
+}
+
+void icache_put(struct super_block *sb, struct inode *inode)
+{
+ spin_lock(&sb->s_icache_lock);
+ /* there's a race in load_ino() that could trigger this */
+ assert(!hashtable_search(sb->s_icache, (void*)inode->i_ino));
+ hashtable_insert(sb->s_icache, (void*)inode->i_ino, inode);
+ spin_unlock(&sb->s_icache_lock);
+}
+
+struct inode *icache_remove(struct super_block *sb, unsigned long ino)
+{
+ struct inode *inode;
+ /* Presumably these hashtable removals could be easier since callers
+ * actually know who they are (same with the pid2proc hash) */
+ spin_lock(&sb->s_icache_lock);
+ inode = hashtable_remove(sb->s_icache, (void*)ino);
+ spin_unlock(&sb->s_icache_lock);
+ assert(inode && !kref_refcnt(&inode->i_kref));
+ return inode;
+}
+
/* File functions */
-/* Read count bytes from the file into buf, starting at *offset, which is increased
- * accordingly, returning the number of bytes transfered. Most filesystems will
- * use this function for their f_op->read. Note, this uses the page cache.
- * Want to try out page remapping later on... */
+/* Read count bytes from the file into buf, starting at *offset, which is
+ * increased accordingly, returning the number of bytes transfered. Most
+ * filesystems will use this function for their f_op->read.
+ * Note, this uses the page cache. */
ssize_t generic_file_read(struct file *file, char *buf, size_t count,
- off_t *offset)
+ off64_t *offset)
{
struct page *page;
int error;
- off_t page_off;
+ off64_t page_off;
unsigned long first_idx, last_idx;
size_t copy_amt;
char *buf_end;
+ /* read in offset, in case of a concurrent reader/writer, so we don't screw
+ * up our math for count, the idxs, etc. */
+ off64_t orig_off = ACCESS_ONCE(*offset);
/* Consider pushing some error checking higher in the VFS */
if (!count)
return 0;
- if (*offset == file->f_dentry->d_inode->i_size)
+ if (orig_off >= file->f_dentry->d_inode->i_size)
return 0; /* EOF */
/* Make sure we don't go past the end of the file */
- if (*offset + count > file->f_dentry->d_inode->i_size) {
- count = file->f_dentry->d_inode->i_size - *offset;
+ if (orig_off + count > file->f_dentry->d_inode->i_size) {
+ count = file->f_dentry->d_inode->i_size - orig_off;
}
- page_off = *offset & (PGSIZE - 1);
- first_idx = *offset >> PGSHIFT;
- last_idx = (*offset + count) >> PGSHIFT;
+ assert((long)count > 0);
+ page_off = orig_off & (PGSIZE - 1);
+ first_idx = orig_off >> PGSHIFT;
+ last_idx = (orig_off + count) >> PGSHIFT;
buf_end = buf + count;
/* For each file page, make sure it's in the page cache, then copy it out.
* TODO: will probably need to consider concurrently truncated files here.*/
for (int i = first_idx; i <= last_idx; i++) {
- error = file_load_page(file, i, &page);
+ error = pm_load_page(file->f_mapping, i, &page);
assert(!error); /* TODO: handle ENOMEM and friends */
copy_amt = MIN(PGSIZE - page_off, buf_end - buf);
/* TODO: (UMEM) think about this. if it's a user buffer, we're relying
}
buf += copy_amt;
page_off = 0;
- page_decref(page); /* it's still in the cache, we just don't need it */
+ pm_put_page(page); /* it's still in the cache, we just don't need it */
}
assert(buf == buf_end);
- *offset += count;
+ /* could have concurrent file ops that screw with offset, so userspace isn't
+ * safe. but at least it'll be a value that one of the concurrent ops could
+ * have produced (compared to *offset_changed_concurrently += count. */
+ *offset = orig_off + count;
return count;
}
-/* Write count bytes from buf to the file, starting at *offset, which is increased
- * accordingly, returning the number of bytes transfered. Most filesystems will
- * use this function for their f_op->write. Note, this uses the page cache.
+/* Write count bytes from buf to the file, starting at *offset, which is
+ * increased accordingly, returning the number of bytes transfered. Most
+ * filesystems will use this function for their f_op->write. Note, this uses
+ * the page cache.
+ *
* Changes don't get flushed to disc til there is an fsync, page cache eviction,
* or other means of trying to writeback the pages. */
ssize_t generic_file_write(struct file *file, const char *buf, size_t count,
- off_t *offset)
+ off64_t *offset)
{
struct page *page;
int error;
- off_t page_off;
+ off64_t page_off;
unsigned long first_idx, last_idx;
size_t copy_amt;
const char *buf_end;
+ off64_t orig_off = ACCESS_ONCE(*offset);
/* Consider pushing some error checking higher in the VFS */
if (!count)
return 0;
- /* Extend the file. Should put more checks in here, and maybe do this per
- * page in the for loop below. */
- if (*offset + count > file->f_dentry->d_inode->i_size)
- file->f_dentry->d_inode->i_size = *offset + count;
- page_off = *offset & (PGSIZE - 1);
- first_idx = *offset >> PGSHIFT;
- last_idx = (*offset + count) >> PGSHIFT;
+ if (file->f_flags & O_APPEND) {
+ spin_lock(&file->f_dentry->d_inode->i_lock);
+ orig_off = file->f_dentry->d_inode->i_size;
+ /* setting the filesize here, instead of during the extend-check, since
+ * we need to atomically reserve space and set our write position. */
+ file->f_dentry->d_inode->i_size += count;
+ spin_unlock(&file->f_dentry->d_inode->i_lock);
+ } else {
+ if (orig_off + count > file->f_dentry->d_inode->i_size) {
+ /* lock for writes to i_size. we allow lockless reads. recheck
+ * i_size in case of concurrent writers since our orig check. */
+ spin_lock(&file->f_dentry->d_inode->i_lock);
+ if (orig_off + count > file->f_dentry->d_inode->i_size)
+ file->f_dentry->d_inode->i_size = orig_off + count;
+ spin_unlock(&file->f_dentry->d_inode->i_lock);
+ }
+ }
+ page_off = orig_off & (PGSIZE - 1);
+ first_idx = orig_off >> PGSHIFT;
+ last_idx = (orig_off + count) >> PGSHIFT;
buf_end = buf + count;
/* For each file page, make sure it's in the page cache, then write it.*/
for (int i = first_idx; i <= last_idx; i++) {
- error = file_load_page(file, i, &page);
+ error = pm_load_page(file->f_mapping, i, &page);
assert(!error); /* TODO: handle ENOMEM and friends */
copy_amt = MIN(PGSIZE - page_off, buf_end - buf);
/* TODO: (UMEM) (KFOP) think about this. if it's a user buffer, we're
}
buf += copy_amt;
page_off = 0;
- page_decref(page); /* it's still in the cache, we just don't need it */
+ atomic_or(&page->pg_flags, PG_DIRTY);
+ pm_put_page(page); /* it's still in the cache, we just don't need it */
}
assert(buf == buf_end);
- *offset += count;
+ *offset = orig_off + count;
return count;
}
* currently expects us to do a readdir (short of doing linux's getdents). Will
* probably need work, based on whatever real programs want. */
ssize_t generic_dir_read(struct file *file, char *u_buf, size_t count,
- off_t *offset)
+ off64_t *offset)
{
struct kdirent dir_r = {0}, *dirent = &dir_r;
- unsigned int num_dirents = count / sizeof(struct kdirent);
int retval = 1;
size_t amt_copied = 0;
char *buf_end = u_buf + count;
- if (!count)
- return 0;
- if (*offset % sizeof(struct kdirent)) {
- printk("[kernel] the f_pos for a directory should be dirent-aligned\n");
- set_errno(EINVAL);
+ if (!S_ISDIR(file->f_dentry->d_inode->i_mode)) {
+ set_errno(ENOTDIR);
return -1;
}
- /* for now, we need to tell readdir which dirent we want */
- dirent->d_off = *offset / sizeof(struct kdirent);
- for (; (u_buf < buf_end) && (retval == 1); u_buf += sizeof(struct kdirent)){
+ if (!count)
+ return 0;
+ /* start readdir from where it left off: */
+ dirent->d_off = *offset;
+ for ( ;
+ u_buf + sizeof(struct kdirent) <= buf_end;
+ u_buf += sizeof(struct kdirent)) {
/* TODO: UMEM/KFOP (pin the u_buf in the syscall, ditch the local copy,
* get rid of this memcpy and reliance on current, etc). Might be
- * tricky with the dirent->d_off */
+ * tricky with the dirent->d_off and trust issues */
retval = file->f_op->readdir(file, dirent);
- if (retval < 0)
+ if (retval < 0) {
+ set_errno(-retval);
break;
+ }
+ /* Slight info exposure: could be extra crap after the name in the
+ * dirent (like the name of a deleted file) */
if (current) {
memcpy_to_user(current, u_buf, dirent, sizeof(struct dirent));
} else {
memcpy(u_buf, dirent, sizeof(struct dirent));
}
amt_copied += sizeof(struct dirent);
- dirent->d_off++;
+ /* 0 signals end of directory */
+ if (retval == 0)
+ break;
}
- *offset += amt_copied;
+ /* Next time read is called, we pick up where we left off */
+ *offset = dirent->d_off; /* UMEM */
+ /* important to tell them how much they got. they often keep going til they
+ * get 0 back (in the case of ls). it's also how much has been read, but it
+ * isn't how much the f_pos has moved (which is opaque to the VFS). */
return amt_copied;
}
struct inode *parent_i;
struct nameidata nd_r = {0}, *nd = &nd_r;
int error;
+ unsigned long nr_pages;
/* The file might exist, lets try to just open it right away */
nd->intent = LOOKUP_OPEN;
error = path_lookup(path, LOOKUP_FOLLOW, nd);
if (!error) {
- /* Still need to make sure we didn't want to O_EXCL create */
+ /* If this is a directory, make sure we are opening with O_RDONLY.
+ * Unfortunately we can't just check for O_RDONLY directly because its
+ * value is 0x0. We instead have to make sure it's not O_WRONLY and
+ * not O_RDWR explicitly. */
+ if (S_ISDIR(nd->dentry->d_inode->i_mode) &&
+ ((flags & O_WRONLY) || (flags & O_RDWR))) {
+ set_errno(EISDIR);
+ goto out_path_only;
+ }
+ /* Also need to make sure we didn't want to O_EXCL create */
if ((flags & O_CREAT) && (flags & O_EXCL)) {
set_errno(EEXIST);
goto out_path_only;
kref_get(&file_d->d_kref, 1);
goto open_the_file;
}
+ if (!(flags & O_CREAT)) {
+ set_errno(-error);
+ goto out_path_only;
+ }
/* So it didn't already exist, release the path from the previous lookup,
* and then we try to create it. */
path_release(nd);
file_d = do_lookup(nd->dentry, nd->last.name);
if (!file_d) {
if (!(flags & O_CREAT)) {
+ warn("Extremely unlikely race, probably a bug");
set_errno(ENOENT);
goto out_path_only;
}
/* Create the inode/file. get a fresh dentry too: */
file_d = get_dentry(nd->dentry->d_sb, nd->dentry, nd->last.name);
+ if (!file_d)
+ goto out_path_only;
parent_i = nd->dentry->d_inode;
/* Note that the mode technically should only apply to future opens,
* but we apply it immediately. */
if (create_file(parent_i, file_d, mode)) /* sets errno */
goto out_file_d;
- dcache_put(file_d);
+ dcache_put(file_d->d_sb, file_d);
} else { /* something already exists */
/* this can happen due to concurrent access, but needs to be thought
* through */
open_the_file:
/* now open the file (freshly created or if it already existed). At this
* point, file_d is a refcnt'd dentry, regardless of which branch we took.*/
- if (flags & O_TRUNC)
- warn("File truncation not supported yet.");
+ if (flags & O_TRUNC) {
+ spin_lock(&file_d->d_inode->i_lock);
+ nr_pages = ROUNDUP(file_d->d_inode->i_size, PGSIZE) >> PGSHIFT;
+ file_d->d_inode->i_size = 0;
+ spin_unlock(&file_d->d_inode->i_lock);
+ pm_remove_contig(file_d->d_inode->i_mapping, 0, nr_pages);
+ }
file = dentry_open(file_d, flags); /* sets errno */
/* Note the fall through to the exit paths. File is 0 by default and if
* dentry_open fails. */
}
/* Doesn't already exist, let's try to make it: */
sym_d = get_dentry(nd->dentry->d_sb, nd->dentry, nd->last.name);
- if (!sym_d) {
- set_errno(ENOMEM);
+ if (!sym_d)
goto out_path_only;
- }
parent_i = nd->dentry->d_inode;
if (create_symlink(parent_i, sym_d, symname, mode))
goto out_sym_d;
- dcache_put(sym_d);
+ dcache_put(sym_d->d_sb, sym_d);
retval = 0; /* Note the fall through to the exit paths */
out_sym_d:
kref_put(&sym_d->d_kref);
/* Doesn't already exist, let's try to make it. Still need to stitch it to
* an inode and set its FS-specific stuff after this.*/
link_d = get_dentry(nd->dentry->d_sb, nd->dentry, nd->last.name);
- if (!link_d) {
- set_errno(ENOMEM);
+ if (!link_d)
goto out_path_only;
- }
/* Now let's get the old_path target */
old_d = lookup_dentry(old_path, LOOKUP_FOLLOW);
if (!old_d) /* errno set by lookup_dentry */
kref_get(&inode->i_kref, 1);
link_d->d_inode = inode;
inode->i_nlink++;
- parent_dir->i_nlink++;
TAILQ_INSERT_TAIL(&inode->i_dentry, link_d, d_alias); /* weak ref */
- dcache_put(link_d);
+ dcache_put(link_d->d_sb, link_d);
retval = 0; /* Note the fall through to the exit paths */
out_both_ds:
kref_put(&old_d->d_kref);
set_errno(-error);
goto out_dentry;
}
- kref_put(&dentry->d_parent->d_kref);
- dentry->d_parent = 0; /* so we don't double-decref it later */
+ /* Now that our parent doesn't track us, we need to make sure we aren't
+ * findable via the dentry cache. DYING, so we will be freed in
+ * dentry_release() */
+ dentry->d_flags |= DENTRY_DYING;
+ dcache_remove(dentry->d_sb, dentry);
dentry->d_inode->i_nlink--; /* TODO: race here, esp with a decref */
/* At this point, the dentry is unlinked from the FS, and the inode has one
* less link. When the in-memory objects (dentry, inode) are going to be
return retval;
}
-int do_chmod(char *path, int mode)
+int do_file_chmod(struct file *file, int mode)
{
- struct nameidata nd_r = {0}, *nd = &nd_r;
- int retval = 0;
- retval = path_lookup(path, 0, nd);
- if (!retval) {
- #if 0
- /* TODO: when we have notions of uid, check for the proc's uid */
- if (nd->dentry->d_inode->i_uid != UID_OF_ME)
- retval = -EPERM;
- else
- #endif
- nd->dentry->d_inode->i_mode |= mode & S_PMASK;
- }
- path_release(nd);
- return retval;
+ int old_mode_ftype = file->f_dentry->d_inode->i_mode & __S_IFMT;
+ #if 0
+ /* TODO: when we have notions of uid, check for the proc's uid */
+ if (file->f_dentry->d_inode->i_uid != UID_OF_ME)
+ retval = -EPERM;
+ else
+ #endif
+ file->f_dentry->d_inode->i_mode = (mode & S_PMASK) | old_mode_ftype;
+ return 0;
}
/* Make a directory at path with mode. Returns -1 and sets errno on errors */
}
/* Doesn't already exist, let's try to make it: */
dentry = get_dentry(nd->dentry->d_sb, nd->dentry, nd->last.name);
- if (!dentry) {
- set_errno(ENOMEM);
+ if (!dentry)
goto out_path_only;
- }
parent_i = nd->dentry->d_inode;
if (create_dir(parent_i, dentry, mode))
goto out_dentry;
- dcache_put(dentry);
+ dcache_put(dentry->d_sb, dentry);
retval = 0; /* Note the fall through to the exit paths */
out_dentry:
kref_put(&dentry->d_kref);
set_errno(ENOTDIR);
goto out_dentry;
}
- /* TODO: make sure we aren't a mount or processes root (EBUSY) */
- /* make sure we are empty. TODO: Race with this, and anything touching
- * i_nlink! */
- if (dentry->d_inode->i_nlink != 1) {
- set_errno(ENOTEMPTY);
+ if (dentry->d_mount_point) {
+ set_errno(EBUSY);
goto out_dentry;
}
- /* now for the removal */
+ /* TODO: make sure we aren't a mount or processes root (EBUSY) */
+ /* Now for the removal. the FSs will check if they are empty */
parent_i = nd->dentry->d_inode;
error = parent_i->i_op->rmdir(parent_i, dentry);
if (error < 0) {
set_errno(-error);
goto out_dentry;
}
+ /* Now that our parent doesn't track us, we need to make sure we aren't
+ * findable via the dentry cache. DYING, so we will be freed in
+ * dentry_release() */
+ dentry->d_flags |= DENTRY_DYING;
+ dcache_remove(dentry->d_sb, dentry);
/* Decref ourselves, so inode_release() knows we are done */
dentry->d_inode->i_nlink--;
TAILQ_REMOVE(&nd->dentry->d_subdirs, dentry, d_subdirs_link);
return retval;
}
-/* Opens and returns the file specified by dentry */
-struct file *dentry_open(struct dentry *dentry, int flags)
+/* Pipes: Doing a simple buffer with reader and writer offsets. Size is power
+ * of two, so we can easily compute its status and whatnot. */
+
+#define PIPE_SZ (1 << PGSHIFT)
+
+static size_t pipe_get_rd_idx(struct pipe_inode_info *pii)
+{
+ return pii->p_rd_off & (PIPE_SZ - 1);
+}
+
+static size_t pipe_get_wr_idx(struct pipe_inode_info *pii)
+{
+
+ return pii->p_wr_off & (PIPE_SZ - 1);
+}
+
+static bool pipe_is_empty(struct pipe_inode_info *pii)
+{
+ return __ring_empty(pii->p_wr_off, pii->p_rd_off);
+}
+
+static bool pipe_is_full(struct pipe_inode_info *pii)
+{
+ return __ring_full(PIPE_SZ, pii->p_wr_off, pii->p_rd_off);
+}
+
+static size_t pipe_nr_full(struct pipe_inode_info *pii)
+{
+ return __ring_nr_full(pii->p_wr_off, pii->p_rd_off);
+}
+
+static size_t pipe_nr_empty(struct pipe_inode_info *pii)
+{
+ return __ring_nr_empty(PIPE_SZ, pii->p_wr_off, pii->p_rd_off);
+}
+
+ssize_t pipe_file_read(struct file *file, char *buf, size_t count,
+ off64_t *offset)
+{
+ struct pipe_inode_info *pii = file->f_dentry->d_inode->i_pipe;
+ size_t copy_amt, amt_copied = 0;
+
+ cv_lock(&pii->p_cv);
+ while (pipe_is_empty(pii)) {
+ /* We wait til the pipe is drained before sending EOF if there are no
+ * writers (instead of aborting immediately) */
+ if (!pii->p_nr_writers) {
+ cv_unlock(&pii->p_cv);
+ return 0;
+ }
+ if (file->f_flags & O_NONBLOCK) {
+ cv_unlock(&pii->p_cv);
+ set_errno(EAGAIN);
+ return -1;
+ }
+ cv_wait(&pii->p_cv);
+ cpu_relax();
+ }
+ /* We might need to wrap-around with our copy, so we'll do the copy in two
+ * passes. This will copy up to the end of the buffer, then on the next
+ * pass will copy the rest to the beginning of the buffer (if necessary) */
+ for (int i = 0; i < 2; i++) {
+ copy_amt = MIN(PIPE_SZ - pipe_get_rd_idx(pii),
+ MIN(pipe_nr_full(pii), count));
+ assert(current); /* shouldn't pipe from the kernel */
+ memcpy_to_user(current, buf, pii->p_buf + pipe_get_rd_idx(pii),
+ copy_amt);
+ buf += copy_amt;
+ count -= copy_amt;
+ pii->p_rd_off += copy_amt;
+ amt_copied += copy_amt;
+ }
+ /* Just using one CV for both readers and writers. We should rarely have
+ * multiple readers or writers. */
+ if (amt_copied)
+ __cv_broadcast(&pii->p_cv);
+ cv_unlock(&pii->p_cv);
+ return amt_copied;
+}
+
+/* Note: we're not dealing with PIPE_BUF and minimum atomic chunks, unless I
+ * have to later. */
+ssize_t pipe_file_write(struct file *file, const char *buf, size_t count,
+ off64_t *offset)
+{
+ struct pipe_inode_info *pii = file->f_dentry->d_inode->i_pipe;
+ size_t copy_amt, amt_copied = 0;
+
+ cv_lock(&pii->p_cv);
+ /* Write aborts right away if there are no readers, regardless of pipe
+ * status. */
+ if (!pii->p_nr_readers) {
+ cv_unlock(&pii->p_cv);
+ set_errno(EPIPE);
+ return -1;
+ }
+ while (pipe_is_full(pii)) {
+ if (file->f_flags & O_NONBLOCK) {
+ cv_unlock(&pii->p_cv);
+ set_errno(EAGAIN);
+ return -1;
+ }
+ cv_wait(&pii->p_cv);
+ cpu_relax();
+ /* Still need to check in the loop, in case the last reader left while
+ * we slept. */
+ if (!pii->p_nr_readers) {
+ cv_unlock(&pii->p_cv);
+ set_errno(EPIPE);
+ return -1;
+ }
+ }
+ /* We might need to wrap-around with our copy, so we'll do the copy in two
+ * passes. This will copy up to the end of the buffer, then on the next
+ * pass will copy the rest to the beginning of the buffer (if necessary) */
+ for (int i = 0; i < 2; i++) {
+ copy_amt = MIN(PIPE_SZ - pipe_get_wr_idx(pii),
+ MIN(pipe_nr_empty(pii), count));
+ assert(current); /* shouldn't pipe from the kernel */
+ memcpy_from_user(current, pii->p_buf + pipe_get_wr_idx(pii), buf,
+ copy_amt);
+ buf += copy_amt;
+ count -= copy_amt;
+ pii->p_wr_off += copy_amt;
+ amt_copied += copy_amt;
+ }
+ /* Just using one CV for both readers and writers. We should rarely have
+ * multiple readers or writers. */
+ if (amt_copied)
+ __cv_broadcast(&pii->p_cv);
+ cv_unlock(&pii->p_cv);
+ return amt_copied;
+}
+
+/* In open and release, we need to track the number of readers and writers,
+ * which we can differentiate by the file flags. */
+int pipe_open(struct inode *inode, struct file *file)
+{
+ struct pipe_inode_info *pii = inode->i_pipe;
+ cv_lock(&pii->p_cv);
+ /* Ugliness due to not using flags for O_RDONLY and friends... */
+ if (file->f_mode == S_IRUSR) {
+ pii->p_nr_readers++;
+ } else if (file->f_mode == S_IWUSR) {
+ pii->p_nr_writers++;
+ } else {
+ warn("Bad pipe file flags 0x%x\n", file->f_flags);
+ }
+ cv_unlock(&pii->p_cv);
+ return 0;
+}
+
+int pipe_release(struct inode *inode, struct file *file)
+{
+ struct pipe_inode_info *pii = inode->i_pipe;
+ cv_lock(&pii->p_cv);
+ /* Ugliness due to not using flags for O_RDONLY and friends... */
+ if (file->f_mode == S_IRUSR) {
+ pii->p_nr_readers--;
+ } else if (file->f_mode == S_IWUSR) {
+ pii->p_nr_writers--;
+ } else {
+ warn("Bad pipe file flags 0x%x\n", file->f_flags);
+ }
+ /* need to wake up any sleeping readers/writers, since we might be done */
+ __cv_broadcast(&pii->p_cv);
+ cv_unlock(&pii->p_cv);
+ return 0;
+}
+
+struct file_operations pipe_f_op = {
+ .read = pipe_file_read,
+ .write = pipe_file_write,
+ .open = pipe_open,
+ .release = pipe_release,
+ 0
+};
+
+void pipe_debug(struct file *f)
+{
+ struct pipe_inode_info *pii = f->f_dentry->d_inode->i_pipe;
+ assert(pii);
+ printk("PIPE %p\n", pii);
+ printk("\trdoff %p\n", pii->p_rd_off);
+ printk("\twroff %p\n", pii->p_wr_off);
+ printk("\tnr_rds %d\n", pii->p_nr_readers);
+ printk("\tnr_wrs %d\n", pii->p_nr_writers);
+ printk("\tcv waiters %d\n", pii->p_cv.nr_waiters);
+
+}
+
+/* General plan: get a dentry/inode to represent the pipe. We'll alloc it from
+ * the default_ns SB, but won't actually link it anywhere. It'll only be held
+ * alive by the krefs, til all the FDs are closed. */
+int do_pipe(struct file **pipe_files, int flags)
+{
+ struct dentry *pipe_d;
+ struct inode *pipe_i;
+ struct file *pipe_f_read, *pipe_f_write;
+ struct super_block *def_sb = default_ns.root->mnt_sb;
+ struct pipe_inode_info *pii;
+
+ pipe_d = get_dentry(def_sb, 0, "pipe");
+ if (!pipe_d)
+ return -1;
+ pipe_d->d_op = &dummy_d_op;
+ pipe_i = get_inode(pipe_d);
+ if (!pipe_i)
+ goto error_post_dentry;
+ /* preemptively mark the dentry for deletion. we have an unlinked dentry
+ * right off the bat, held in only by the kref chain (pipe_d is the ref). */
+ pipe_d->d_flags |= DENTRY_DYING;
+ /* pipe_d->d_inode still has one ref to pipe_i, keeping the inode alive */
+ kref_put(&pipe_i->i_kref);
+ /* init inode fields. note we're using the dummy ops for i_op and d_op */
+ pipe_i->i_mode = S_IRWXU | S_IRWXG | S_IRWXO;
+ SET_FTYPE(pipe_i->i_mode, __S_IFIFO); /* using type == FIFO */
+ pipe_i->i_nlink = 1; /* one for the dentry */
+ pipe_i->i_uid = 0;
+ pipe_i->i_gid = 0;
+ pipe_i->i_size = PGSIZE;
+ pipe_i->i_blocks = 0;
+ pipe_i->i_atime.tv_sec = 0;
+ pipe_i->i_atime.tv_nsec = 0;
+ pipe_i->i_mtime.tv_sec = 0;
+ pipe_i->i_mtime.tv_nsec = 0;
+ pipe_i->i_ctime.tv_sec = 0;
+ pipe_i->i_ctime.tv_nsec = 0;
+ pipe_i->i_fs_info = 0;
+ pipe_i->i_op = &dummy_i_op;
+ pipe_i->i_fop = &pipe_f_op;
+ pipe_i->i_socket = FALSE;
+ /* Actually build the pipe. We're using one page, hanging off the
+ * pipe_inode_info struct. When we release the inode, we free the pipe
+ * memory too */
+ pipe_i->i_pipe = kmalloc(sizeof(struct pipe_inode_info), KMALLOC_WAIT);
+ pii = pipe_i->i_pipe;
+ if (!pii) {
+ set_errno(ENOMEM);
+ goto error_kmalloc;
+ }
+ pii->p_buf = kpage_zalloc_addr();
+ if (!pii->p_buf) {
+ set_errno(ENOMEM);
+ goto error_kpage;
+ }
+ pii->p_rd_off = 0;
+ pii->p_wr_off = 0;
+ pii->p_nr_readers = 0;
+ pii->p_nr_writers = 0;
+ cv_init(&pii->p_cv); /* must do this before dentry_open / pipe_open */
+ /* Now we have an inode for the pipe. We need two files for the read and
+ * write ends of the pipe. */
+ flags &= ~(O_ACCMODE); /* avoid user bugs */
+ pipe_f_read = dentry_open(pipe_d, flags | O_RDONLY);
+ if (!pipe_f_read)
+ goto error_f_read;
+ pipe_f_write = dentry_open(pipe_d, flags | O_WRONLY);
+ if (!pipe_f_write)
+ goto error_f_write;
+ pipe_files[0] = pipe_f_read;
+ pipe_files[1] = pipe_f_write;
+ return 0;
+
+error_f_write:
+ kref_put(&pipe_f_read->f_kref);
+error_f_read:
+ page_decref(kva2page(pii->p_buf));
+error_kpage:
+ kfree(pipe_i->i_pipe);
+error_kmalloc:
+ /* We don't need to free the pipe_i; putting the dentry will free it */
+error_post_dentry:
+ /* Note we only free the dentry on failure. */
+ kref_put(&pipe_d->d_kref);
+ return -1;
+}
+
+int do_rename(char *old_path, char *new_path)
+{
+ struct nameidata nd_old = {0}, *nd_o = &nd_old;
+ struct nameidata nd_new = {0}, *nd_n = &nd_new;
+ struct dentry *old_dir_d, *new_dir_d;
+ struct inode *old_dir_i, *new_dir_i;
+ struct dentry *old_d, *new_d, *unlink_d;
+ int error;
+ int retval = 0;
+ uint64_t now;
+
+ nd_o->intent = LOOKUP_ACCESS; /* maybe, might need another type */
+
+ /* get the parent, but don't follow links */
+ error = path_lookup(old_path, LOOKUP_PARENT | LOOKUP_DIRECTORY, nd_o);
+ if (error) {
+ set_errno(-error);
+ retval = -1;
+ goto out_old_path;
+ }
+ old_dir_d = nd_o->dentry;
+ old_dir_i = old_dir_d->d_inode;
+
+ old_d = do_lookup(old_dir_d, nd_o->last.name);
+ if (!old_d) {
+ set_errno(ENOENT);
+ retval = -1;
+ goto out_old_path;
+ }
+
+ nd_n->intent = LOOKUP_CREATE;
+ error = path_lookup(new_path, LOOKUP_PARENT | LOOKUP_DIRECTORY, nd_n);
+ if (error) {
+ set_errno(-error);
+ retval = -1;
+ goto out_paths_and_src;
+ }
+ new_dir_d = nd_n->dentry;
+ new_dir_i = new_dir_d->d_inode;
+ /* TODO if new_dir == old_dir, we might be able to simplify things */
+
+ if (new_dir_i->i_sb != old_dir_i->i_sb) {
+ set_errno(EXDEV);
+ retval = -1;
+ goto out_paths_and_src;
+ }
+ /* TODO: check_perms is lousy, want to just say "writable" here */
+ if (check_perms(old_dir_i, S_IWUSR) || check_perms(new_dir_i, S_IWUSR)) {
+ set_errno(EPERM);
+ retval = -1;
+ goto out_paths_and_src;
+ }
+ /* TODO: if we're doing a rename that moves a directory, we need to make
+ * sure the new_path doesn't include the old_path. it's not as simple as
+ * just checking, since there could be a concurrent rename that breaks the
+ * check later. e.g. what if new_dir's parent is being moved into a child
+ * of old_dir?
+ *
+ * linux has a per-fs rename mutex for these scenarios, so only one can
+ * proceed at a time. i don't see another way to deal with it either.
+ * maybe something like flagging all dentries on the new_path with "do not
+ * move". */
+
+ /* TODO: this is all very racy. right after we do a new_d lookup, someone
+ * else could create or unlink new_d. need to lock here, or else push this
+ * into the sub-FS.
+ *
+ * For any locking scheme, we probably need to lock both the old and new
+ * dirs. To prevent deadlock, we need a total ordering of all inodes (or
+ * dentries, if we locking them instead). inode number or struct inode*
+ * will work for this. */
+ new_d = do_lookup(new_dir_d, nd_n->last.name);
+ if (new_d) {
+ if (new_d->d_inode == old_d->d_inode)
+ goto out_paths_and_refs; /* rename does nothing */
+ /* TODO: Here's a bunch of other racy checks we need to do, maybe in the
+ * sub-FS:
+ *
+ * if src is a dir, dst must be an empty dir if it exists (RACYx2)
+ * racing on dst being created and it getting new entries
+ * if src is a file, dst must be a file if it exists (RACY)
+ * racing on dst being created and still being a file
+ * racing on dst being unlinked and a new one being added
+ */
+ /* TODO: we should allow empty dirs */
+ if (S_ISDIR(new_d->d_inode->i_mode)) {
+ set_errno(EISDIR);
+ retval = -1;
+ goto out_paths_and_refs;
+ }
+ /* TODO: need this to be atomic with rename */
+ error = new_dir_i->i_op->unlink(new_dir_i, new_d);
+ if (error) {
+ set_errno(-error);
+ retval = -1;
+ goto out_paths_and_refs;
+ }
+ new_d->d_flags |= DENTRY_DYING;
+ /* TODO: racy with other lookups on new_d */
+ dcache_remove(new_d->d_sb, new_d);
+ new_d->d_inode->i_nlink--; /* TODO: race here, esp with a decref */
+ kref_put(&new_d->d_kref);
+ }
+ /* new_d is just a vessel for the name. somewhat lousy. */
+ new_d = get_dentry(new_dir_d->d_sb, new_dir_d, nd_n->last.name);
+
+ /* TODO: more races. need to remove old_d from the dcache, since we're
+ * about to change its parentage. could be readded concurrently. */
+ dcache_remove(old_dir_d->d_sb, old_d);
+ error = new_dir_i->i_op->rename(old_dir_i, old_d, new_dir_i, new_d);
+ if (error) {
+ /* TODO: oh crap, we already unlinked! now we're screwed, and violated
+ * our atomicity requirements. */
+ printk("[kernel] rename failed, you might have lost data\n");
+ set_errno(-error);
+ retval = -1;
+ goto out_paths_and_refs;
+ }
+
+ /* old_dir loses old_d, new_dir gains old_d, renamed to new_d. this is
+ * particularly cumbersome since there are two levels here: the FS has its
+ * info about where things are, and the VFS has its dentry tree. and it's
+ * all racy (TODO). */
+ dentry_set_name(old_d, new_d->d_name.name);
+ old_d->d_parent = new_d->d_parent;
+ if (S_ISDIR(old_d->d_inode->i_mode)) {
+ TAILQ_REMOVE(&old_dir_d->d_subdirs, old_d, d_subdirs_link);
+ old_dir_i->i_nlink--; /* TODO: racy, etc */
+ TAILQ_INSERT_TAIL(&new_dir_d->d_subdirs, old_d, d_subdirs_link);
+ new_dir_i->i_nlink--; /* TODO: racy, etc */
+ }
+
+ /* and then the third level: dcache stuff. we could have old versions of
+ * old_d or negative versions of new_d sitting around. dcache_put should
+ * replace a potentially negative dentry for new_d (now called old_d) */
+ dcache_put(old_dir_d->d_sb, old_d);
+
+ /* TODO could have a helper for this, but it's going away soon */
+ now = epoch_seconds();
+ old_dir_i->i_ctime.tv_sec = now;
+ old_dir_i->i_mtime.tv_sec = now;
+ old_dir_i->i_ctime.tv_nsec = 0;
+ old_dir_i->i_mtime.tv_nsec = 0;
+ new_dir_i->i_ctime.tv_sec = now;
+ new_dir_i->i_mtime.tv_sec = now;
+ new_dir_i->i_ctime.tv_nsec = 0;
+ new_dir_i->i_mtime.tv_nsec = 0;
+
+ /* fall-through */
+out_paths_and_refs:
+ kref_put(&new_d->d_kref);
+out_paths_and_src:
+ kref_put(&old_d->d_kref);
+out_paths:
+ path_release(nd_n);
+out_old_path:
+ path_release(nd_o);
+ return retval;
+}
+
+int do_truncate(struct inode *inode, off64_t len)
+{
+ off64_t old_len;
+ uint64_t now;
+ if (len < 0) {
+ set_errno(EINVAL);
+ return -1;
+ }
+ if (len > PiB) {
+ printk("[kernel] truncate for > petabyte, probably a bug\n");
+ /* continuing, not too concerned. could set EINVAL or EFBIG */
+ }
+ spin_lock(&inode->i_lock);
+ old_len = inode->i_size;
+ if (old_len == len) {
+ spin_unlock(&inode->i_lock);
+ return 0;
+ }
+ inode->i_size = len;
+ /* truncate can't block, since we're holding the spinlock. but it can rely
+ * on that lock being held */
+ inode->i_op->truncate(inode);
+ spin_unlock(&inode->i_lock);
+
+ if (old_len < len) {
+ pm_remove_contig(inode->i_mapping, old_len >> PGSHIFT,
+ (len >> PGSHIFT) - (old_len >> PGSHIFT));
+ }
+ now = epoch_seconds();
+ inode->i_ctime.tv_sec = now;
+ inode->i_mtime.tv_sec = now;
+ inode->i_ctime.tv_nsec = 0;
+ inode->i_mtime.tv_nsec = 0;
+ return 0;
+}
+
+struct file *alloc_file(void)
{
- struct inode *inode;
- int desired_mode;
struct file *file = kmem_cache_alloc(file_kcache, 0);
if (!file) {
set_errno(ENOMEM);
return 0;
}
+ /* one for the ref passed out*/
+ kref_init(&file->f_kref, file_release, 1);
+ return file;
+}
+
+/* Opens and returns the file specified by dentry */
+struct file *dentry_open(struct dentry *dentry, int flags)
+{
+ struct inode *inode;
+ struct file *file;
+ int desired_mode;
inode = dentry->d_inode;
/* Do the mode first, since we can still error out. f_mode stores how the
* OS file is open, which can be more restrictive than the i_mode */
}
if (check_perms(inode, desired_mode))
goto error_access;
+ file = alloc_file();
+ if (!file)
+ return 0;
file->f_mode = desired_mode;
- /* one for the ref passed out, and *none* for the sb TAILQ */
- kref_init(&file->f_kref, file_release, 1);
/* Add to the list of all files of this SB */
TAILQ_INSERT_TAIL(&inode->i_sb->s_files, file, f_list);
kref_get(&dentry->d_kref, 1);
kref_get(&inode->i_sb->s_mount->mnt_kref, 1);
file->f_vfsmnt = inode->i_sb->s_mount; /* saving a ref to the vmnt...*/
file->f_op = inode->i_fop;
- /* Don't store open mode or creation flags */
- file->f_flags = flags & ~(O_ACCMODE | O_CREAT_FLAGS);
+ /* Don't store creation flags */
+ file->f_flags = flags & ~O_CREAT_FLAGS;
file->f_pos = 0;
file->f_uid = inode->i_uid;
file->f_gid = inode->i_gid;
file->f_error = 0;
// struct event_poll_tailq f_ep_links;
spinlock_init(&file->f_ep_lock);
- file->f_fs_info = 0; /* prob overriden by the fs */
+ file->f_privdata = 0; /* prob overriden by the fs */
file->f_mapping = inode->i_mapping;
file->f_op->open(inode, file);
return file;
error_access:
set_errno(EACCES);
- kmem_cache_free(file_kcache, file);
return 0;
}
kmem_cache_free(file_kcache, file);
}
-/* Page cache functions */
+/* Process-related File management functions */
-/* Looks up the index'th page in the page map, returning an incref'd reference,
- * or 0 if it was not in the map. */
-struct page *pm_find_page(struct page_map *pm, unsigned long index)
+/* Given any FD, get the appropriate file, 0 o/w */
+struct file *get_file_from_fd(struct files_struct *open_files, int file_desc)
{
- spin_lock(&pm->pm_tree_lock);
- struct page *page = (struct page*)radix_lookup(&pm->pm_tree, index);
- if (page)
- page_incref(page);
- spin_unlock(&pm->pm_tree_lock);
- return page;
+ struct file *retval = 0;
+ if (file_desc < 0)
+ return 0;
+ spin_lock(&open_files->lock);
+ if (open_files->closed) {
+ spin_unlock(&open_files->lock);
+ return 0;
+ }
+ if (file_desc < open_files->max_fdset) {
+ if (GET_BITMASK_BIT(open_files->open_fds->fds_bits, file_desc)) {
+ /* while max_files and max_fdset might not line up, we should never
+ * have a valid fdset higher than files */
+ assert(file_desc < open_files->max_files);
+ retval = open_files->fd[file_desc].fd_file;
+ /* 9ns might be using this one, in which case file == 0 */
+ if (retval)
+ kref_get(&retval->f_kref, 1);
+ }
+ }
+ spin_unlock(&open_files->lock);
+ return retval;
}
-/* Attempts to insert the page into the page_map, returns 0 for success, or an
- * error code if there was one already (EEXIST) or we ran out of memory
- * (ENOMEM). On success, this will preemptively lock the page, and will also
- * store a reference to the page in the pm. */
-int pm_insert_page(struct page_map *pm, unsigned long index, struct page *page)
+/* Grow the vfs fd set */
+static int grow_fd_set(struct files_struct *open_files)
{
- int error = 0;
- spin_lock(&pm->pm_tree_lock);
- error = radix_insert(&pm->pm_tree, index, page);
- if (!error) {
- page_incref(page);
- page->pg_flags |= PG_LOCKED;
- page->pg_mapping = pm;
- page->pg_index = index;
- pm->pm_num_pages++;
- }
- spin_unlock(&pm->pm_tree_lock);
- return error;
-}
-
-/* Removes the page, including its reference. Not sure yet what interface we
- * want to this (pm and index or page), and this has never been used. There are
- * also issues with when you want to call this, since a page in the cache may be
- * mmap'd by someone else. */
-int pm_remove_page(struct page_map *pm, struct page *page)
-{
- void *retval;
- warn("pm_remove_page() hasn't been thought through or tested.");
- spin_lock(&pm->pm_tree_lock);
- retval = radix_delete(&pm->pm_tree, page->pg_index);
- spin_unlock(&pm->pm_tree_lock);
- assert(retval == (void*)page);
- page_decref(page);
- page->pg_mapping = 0;
- page->pg_index = 0;
- pm->pm_num_pages--;
+ int n;
+ struct file_desc *nfd, *ofd;
+
+ /* Only update open_fds once. If currently pointing to open_fds_init, then
+ * update it to point to a newly allocated fd_set with space for
+ * NR_FILE_DESC_MAX */
+ if (open_files->open_fds == (struct fd_set*)&open_files->open_fds_init) {
+ open_files->open_fds = kzmalloc(sizeof(struct fd_set), 0);
+ memmove(open_files->open_fds, &open_files->open_fds_init,
+ sizeof(struct small_fd_set));
+ }
+
+ /* Grow the open_files->fd array in increments of NR_OPEN_FILES_DEFAULT */
+ n = open_files->max_files + NR_OPEN_FILES_DEFAULT;
+ if (n > NR_FILE_DESC_MAX)
+ return -EMFILE;
+ nfd = kzmalloc(n * sizeof(struct file_desc), 0);
+ if (nfd == NULL)
+ return -ENOMEM;
+
+ /* Move the old array on top of the new one */
+ ofd = open_files->fd;
+ memmove(nfd, ofd, open_files->max_files * sizeof(struct file_desc));
+
+ /* Update the array and the maxes for both max_files and max_fdset */
+ open_files->fd = nfd;
+ open_files->max_files = n;
+ open_files->max_fdset = n;
+
+ /* Only free the old one if it wasn't pointing to open_files->fd_array */
+ if (ofd != open_files->fd_array)
+ kfree(ofd);
return 0;
}
-/* Makes sure the index'th page from file is loaded in the page cache and
- * returns its location via **pp. Note this will give you a refcnt'd reference.
- * This may block! TODO: (BLK) */
-int file_load_page(struct file *file, unsigned long index, struct page **pp)
+/* Free the vfs fd set if necessary */
+static void free_fd_set(struct files_struct *open_files)
{
- struct page_map *pm = file->f_mapping;
- struct page *page;
- int error;
- bool page_was_mapped = TRUE;
-
- page = pm_find_page(pm, index);
- while (!page) {
- /* kpage_alloc, since we want the page to persist after the proc
- * dies (can be used by others, until the inode shuts down). */
- if (kpage_alloc(&page))
- return -ENOMEM;
- /* might want to initialize other things, perhaps in page_alloc() */
- page->pg_flags = 0;
- error = pm_insert_page(pm, index, page);
- switch (error) {
- case 0:
- page_was_mapped = FALSE;
- break;
- case -EEXIST:
- /* the page was mapped already (benign race), just get rid of
- * our page and try again (the only case that uses the while) */
- page_decref(page);
- page = pm_find_page(pm, index);
- break;
- default:
- /* something is wrong, bail out! */
- page_decref(page);
- return error;
- }
+ void *free_me;
+ if (open_files->open_fds != (struct fd_set*)&open_files->open_fds_init) {
+ assert(open_files->fd != open_files->fd_array);
+ /* need to reset the pointers to the internal addrs, in case we take a
+ * look while debugging. 0 them out, since they have old data. our
+ * current versions should all be closed. */
+ memset(&open_files->open_fds_init, 0, sizeof(struct small_fd_set));
+ memset(&open_files->fd_array, 0, sizeof(open_files->fd_array));
+
+ free_me = open_files->open_fds;
+ open_files->open_fds = (struct fd_set*)&open_files->open_fds_init;
+ kfree(free_me);
+
+ free_me = open_files->fd;
+ open_files->fd = open_files->fd_array;
+ kfree(free_me);
}
- *pp = page;
- /* if the page was in the map, we need to do some checks, and might have to
- * read in the page later. If the page was freshly inserted to the pm by
- * us, we skip this since we are the one doing the readpage(). */
- if (page_was_mapped) {
- /* is it already here and up to date? if so, we're done */
- if (page->pg_flags & PG_UPTODATE)
- return 0;
- /* if not, try to lock the page (could BLOCK) */
- lock_page(page);
- /* we got it, is our page still in the cache? check the mapping. if
- * not, start over, perhaps with EAGAIN and outside support */
- if (!page->pg_mapping)
- panic("Page is not in the mapping! Haven't implemented this!");
- /* double check, are we up to date? if so, we're done */
- if (page->pg_flags & PG_UPTODATE) {
- unlock_page(page);
- return 0;
- }
- }
- /* if we're here, the page is locked by us, and it needs to be read in */
- assert(page->pg_mapping == pm);
- error = pm->pm_op->readpage(file, page);
- assert(!error);
- /* Try to sleep on the IO. The page will be unlocked when the IO is done */
- lock_page(page);
- unlock_page(page);
- assert(page->pg_flags & PG_UPTODATE);
- return 0;
}
-/* Process-related File management functions */
-
-/* Given any FD, get the appropriate file, 0 o/w */
-struct file *get_file_from_fd(struct files_struct *open_files, int file_desc)
+/* 9ns: puts back an FD from the VFS-FD-space. */
+int put_fd(struct files_struct *open_files, int file_desc)
{
- struct file *retval = 0;
- if (file_desc < 0)
+ if (file_desc < 0) {
+ warn("Negative FD!\n");
return 0;
+ }
spin_lock(&open_files->lock);
if (file_desc < open_files->max_fdset) {
if (GET_BITMASK_BIT(open_files->open_fds->fds_bits, file_desc)) {
/* while max_files and max_fdset might not line up, we should never
* have a valid fdset higher than files */
assert(file_desc < open_files->max_files);
- retval = open_files->fd[file_desc];
- assert(retval);
- kref_get(&retval->f_kref, 1);
+ CLR_BITMASK_BIT(open_files->open_fds->fds_bits, file_desc);
}
}
spin_unlock(&open_files->lock);
- return retval;
+ return 0;
}
/* Remove FD from the open files, if it was there, and return f. Currently,
/* while max_files and max_fdset might not line up, we should never
* have a valid fdset higher than files */
assert(file_desc < open_files->max_files);
- file = open_files->fd[file_desc];
- open_files->fd[file_desc] = 0;
- assert(file);
+ file = open_files->fd[file_desc].fd_file;
+ open_files->fd[file_desc].fd_file = 0;
+ assert(file); /* 9ns shouldn't call this put */
kref_put(&file->f_kref);
CLR_BITMASK_BIT(open_files->open_fds->fds_bits, file_desc);
}
return file;
}
-/* Inserts the file in the files_struct, returning the corresponding new file
- * descriptor, or an error code. We currently grab the first open FD. */
-int insert_file(struct files_struct *open_files, struct file *file)
+static int __get_fd(struct files_struct *open_files, int low_fd)
{
int slot = -1;
+ int error;
+ if ((low_fd < 0) || (low_fd > NR_FILE_DESC_MAX))
+ return -EINVAL;
+ if (open_files->closed)
+ return -EINVAL; /* won't matter, they are dying */
+
+ /* Loop until we have a valid slot (we grow the fd_array at the bottom of
+ * the loop if we haven't found a slot in the current array */
+ while (slot == -1) {
+ for (low_fd; low_fd < open_files->max_fdset; low_fd++) {
+ if (GET_BITMASK_BIT(open_files->open_fds->fds_bits, low_fd))
+ continue;
+ slot = low_fd;
+ SET_BITMASK_BIT(open_files->open_fds->fds_bits, slot);
+ assert(slot < open_files->max_files &&
+ open_files->fd[slot].fd_file == 0);
+ if (slot >= open_files->next_fd)
+ open_files->next_fd = slot + 1;
+ break;
+ }
+ if (slot == -1) {
+ if ((error = grow_fd_set(open_files)))
+ return error;
+ }
+ }
+ return slot;
+}
+
+/* Gets and claims a free FD, used by 9ns. < 0 == error. cloexec is tracked on
+ * the VFS FD. It's value will be O_CLOEXEC (not 1) or 0. */
+int get_fd(struct files_struct *open_files, int low_fd, int cloexec)
+{
+ int slot;
spin_lock(&open_files->lock);
- for (int i = 0; i < open_files->max_fdset; i++) {
- if (GET_BITMASK_BIT(open_files->open_fds->fds_bits, i))
- continue;
- slot = i;
- SET_BITMASK_BIT(open_files->open_fds->fds_bits, slot);
- assert(slot < open_files->max_files && open_files->fd[slot] == 0);
- kref_get(&file->f_kref, 1);
- open_files->fd[slot] = file;
- if (slot >= open_files->next_fd)
- open_files->next_fd = slot + 1;
- break;
- }
- if (slot == -1) /* should expand the FD array and fd_set */
- warn("Ran out of file descriptors, deal with me!");
+ slot = __get_fd(open_files, low_fd);
+ if (cloexec && (slot >= 0))
+ open_files->fd[slot].fd_flags |= FD_CLOEXEC;
+ spin_unlock(&open_files->lock);
+ return slot;
+}
+
+static int __claim_fd(struct files_struct *open_files, int file_desc)
+{
+ int error;
+ if ((file_desc < 0) || (file_desc > NR_FILE_DESC_MAX))
+ return -EINVAL;
+ if (open_files->closed)
+ return -EINVAL; /* won't matter, they are dying */
+
+ /* Grow the open_files->fd_set until the file_desc can fit inside it */
+ while(file_desc >= open_files->max_files) {
+ if ((error = grow_fd_set(open_files)))
+ return error;
+ cpu_relax();
+ }
+
+ /* If we haven't grown, this could be a problem, so check for it */
+ if (GET_BITMASK_BIT(open_files->open_fds->fds_bits, file_desc))
+ return -ENFILE; /* Should never really happen. Here to catch bugs. */
+
+ SET_BITMASK_BIT(open_files->open_fds->fds_bits, file_desc);
+ assert(file_desc < open_files->max_files &&
+ open_files->fd[file_desc].fd_file == 0);
+ if (file_desc >= open_files->next_fd)
+ open_files->next_fd = file_desc + 1;
+ return 0;
+}
+
+/* Claims a specific FD when duping FDs. used by 9ns. < 0 == error. No need
+ * for cloexec here, since it's not used during dup. */
+int claim_fd(struct files_struct *open_files, int file_desc)
+{
+ int ret;
+ spin_lock(&open_files->lock);
+ ret = __claim_fd(open_files, file_desc);
+ spin_unlock(&open_files->lock);
+ return ret;
+}
+
+/* Inserts the file in the files_struct, returning the corresponding new file
+ * descriptor, or an error code. We start looking for open fds from low_fd.
+ *
+ * Passing cloexec is a bit cheap, since we might want to expand it to support
+ * more FD options in the future. */
+int insert_file(struct files_struct *open_files, struct file *file, int low_fd,
+ bool must, bool cloexec)
+{
+ int slot, ret;
+ spin_lock(&open_files->lock);
+ if (must) {
+ ret = __claim_fd(open_files, low_fd);
+ if (ret < 0) {
+ spin_unlock(&open_files->lock);
+ return ret;
+ }
+ assert(!ret); /* issues with claim_fd returning status, not the fd */
+ slot = low_fd;
+ } else {
+ slot = __get_fd(open_files, low_fd);
+ }
+
+ if (slot < 0) {
+ spin_unlock(&open_files->lock);
+ return slot;
+ }
+ assert(slot < open_files->max_files &&
+ open_files->fd[slot].fd_file == 0);
+ kref_get(&file->f_kref, 1);
+ open_files->fd[slot].fd_file = file;
+ open_files->fd[slot].fd_flags = 0;
+ if (cloexec)
+ open_files->fd[slot].fd_flags |= FD_CLOEXEC;
spin_unlock(&open_files->lock);
return slot;
}
/* Closes all open files. Mostly just a "put" for all files. If cloexec, it
- * will only close files that are opened with O_CLOEXEC. */
+ * will only close the FDs with FD_CLOEXEC (opened with O_CLOEXEC or fcntld). */
void close_all_files(struct files_struct *open_files, bool cloexec)
{
struct file *file;
spin_lock(&open_files->lock);
+ if (open_files->closed) {
+ spin_unlock(&open_files->lock);
+ return;
+ }
for (int i = 0; i < open_files->max_fdset; i++) {
if (GET_BITMASK_BIT(open_files->open_fds->fds_bits, i)) {
/* while max_files and max_fdset might not line up, we should never
* have a valid fdset higher than files */
assert(i < open_files->max_files);
- file = open_files->fd[i];
- if (cloexec && !(file->f_flags | O_CLOEXEC))
+ file = open_files->fd[i].fd_file;
+ /* no file == 9ns uses the FD. they will deal with it */
+ if (!file)
+ continue;
+ if (cloexec && !(open_files->fd[i].fd_flags & FD_CLOEXEC))
continue;
- open_files->fd[i] = 0;
+ /* Actually close the file */
+ open_files->fd[i].fd_file = 0;
assert(file);
kref_put(&file->f_kref);
CLR_BITMASK_BIT(open_files->open_fds->fds_bits, i);
}
}
+ if (!cloexec) {
+ free_fd_set(open_files);
+ open_files->closed = TRUE;
+ }
spin_unlock(&open_files->lock);
}
{
struct file *file;
spin_lock(&src->lock);
+ if (src->closed) {
+ spin_unlock(&src->lock);
+ return;
+ }
spin_lock(&dst->lock);
+ if (dst->closed) {
+ warn("Destination closed before it opened");
+ spin_unlock(&dst->lock);
+ spin_unlock(&src->lock);
+ return;
+ }
for (int i = 0; i < src->max_fdset; i++) {
if (GET_BITMASK_BIT(src->open_fds->fds_bits, i)) {
/* while max_files and max_fdset might not line up, we should never
* have a valid fdset higher than files */
assert(i < src->max_files);
- file = src->fd[i];
+ file = src->fd[i].fd_file;
+ assert(i < dst->max_files && dst->fd[i].fd_file == 0);
SET_BITMASK_BIT(dst->open_fds->fds_bits, i);
- assert(i < dst->max_files && dst->fd[i] == 0);
- dst->fd[i] = file;
- assert(file);
- kref_get(&file->f_kref, 1);
+ dst->fd[i].fd_file = file;
+ /* no file means 9ns is using it, they clone separately */
+ if (file)
+ kref_get(&file->f_kref, 1);
+ if (i >= dst->next_fd)
+ dst->next_fd = i + 1;
}
}
spin_unlock(&dst->lock);
spin_unlock(&src->lock);
}
+static void __chpwd(struct fs_struct *fs_env, struct dentry *new_pwd)
+{
+ struct dentry *old_pwd;
+ kref_get(&new_pwd->d_kref, 1);
+ /* writer lock, make sure we replace pwd with ours. could also CAS.
+ * readers don't lock at all, so they need to either loop, or we need to
+ * delay releasing old_pwd til an RCU grace period. */
+ spin_lock(&fs_env->lock);
+ old_pwd = fs_env->pwd;
+ fs_env->pwd = new_pwd;
+ spin_unlock(&fs_env->lock);
+ kref_put(&old_pwd->d_kref);
+}
+
/* Change the working directory of the given fs env (one per process, at this
- * point). Returns 0 for success, -ERROR for whatever error. */
+ * point). Returns 0 for success, sets errno and returns -1 otherwise. */
int do_chdir(struct fs_struct *fs_env, char *path)
{
struct nameidata nd_r = {0}, *nd = &nd_r;
- int retval;
- retval = path_lookup(path, LOOKUP_DIRECTORY, nd);
- if (!retval) {
- /* nd->dentry is the place we want our PWD to be */
- kref_get(&nd->dentry->d_kref, 1);
- kref_put(&fs_env->pwd->d_kref);
- fs_env->pwd = nd->dentry;
+ int error;
+ error = path_lookup(path, LOOKUP_DIRECTORY, nd);
+ if (error) {
+ set_errno(-error);
+ path_release(nd);
+ return -1;
}
+ /* nd->dentry is the place we want our PWD to be */
+ __chpwd(fs_env, nd->dentry);
path_release(nd);
- return retval;
+ return 0;
+}
+
+int do_fchdir(struct fs_struct *fs_env, struct file *file)
+{
+ if ((file->f_dentry->d_inode->i_mode & __S_IFMT) != __S_IFDIR) {
+ set_errno(ENOTDIR);
+ return -1;
+ }
+ __chpwd(fs_env, file->f_dentry);
+ return 0;
}
/* Returns a null-terminated string of up to length cwd_l containing the
set_errno(ERANGE);
return 0;
}
- path_start -= link_len + 1; /* the 1 is for the \0 */
+ path_start -= link_len;
strncpy(path_start, dentry->d_name.name, link_len);
path_start--;
*path_start = '/';
static void print_dir(struct dentry *dentry, char *buf, int depth)
{
struct dentry *child_d;
- struct dirent next;
+ struct dirent next = {0};
struct file *dir;
int retval;
- int child_num = 0;
if (!S_ISDIR(dentry->d_inode->i_mode)) {
warn("Thought this was only directories!!");
/* Print this dentry */
printk("%s%s/ nlink: %d\n", buf, dentry->d_name.name,
dentry->d_inode->i_nlink);
+ if (dentry->d_mount_point) {
+ dentry = dentry->d_mounted_fs->mnt_root;
+ }
if (depth >= 32)
return;
/* Set buffer for our kids */
panic("Filesystem seems inconsistent - unable to open a dir!");
/* Process every child, recursing on directories */
while (1) {
- next.d_off = child_num++;
retval = dir->f_op->readdir(dir, &next);
if (retval >= 0) {
+ /* Skip .., ., and empty entries */
+ if (!strcmp("..", next.d_name) || !strcmp(".", next.d_name) ||
+ next.d_ino == 0)
+ goto loop_next;
/* there is an entry, now get its dentry */
child_d = do_lookup(dentry, next.d_name);
if (!child_d)
printk("%s%s (char device) nlink: %d\n", buf, next.d_name,
child_d->d_inode->i_nlink);
break;
+ case (__S_IFBLK):
+ printk("%s%s (block device) nlink: %d\n", buf, next.d_name,
+ child_d->d_inode->i_nlink);
+ break;
default:
warn("Look around you! Unknown filetype!");
}
kref_put(&child_d->d_kref);
}
+loop_next:
if (retval <= 0)
break;
}
path_release(nd);
return 0;
}
+
+/* Dummy ops, to catch weird operations we weren't expecting */
+int dummy_create(struct inode *dir, struct dentry *dentry, int mode,
+ struct nameidata *nd)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+struct dentry *dummy_lookup(struct inode *dir, struct dentry *dentry,
+ struct nameidata *nd)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return 0;
+}
+
+int dummy_link(struct dentry *old_dentry, struct inode *dir,
+ struct dentry *new_dentry)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_unlink(struct inode *dir, struct dentry *dentry)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_mkdir(struct inode *dir, struct dentry *dentry, int mode)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_rmdir(struct inode *dir, struct dentry *dentry)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t rdev)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_rename(struct inode *old_dir, struct dentry *old_dentry,
+ struct inode *new_dir, struct dentry *new_dentry)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+char *dummy_readlink(struct dentry *dentry)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return 0;
+}
+
+void dummy_truncate(struct inode *inode)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+}
+
+int dummy_permission(struct inode *inode, int mode, struct nameidata *nd)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_d_revalidate(struct dentry *dir, struct nameidata *nd)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_d_hash(struct dentry *dentry, struct qstr *name)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_d_compare(struct dentry *dir, struct qstr *name1, struct qstr *name2)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_d_delete(struct dentry *dentry)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+int dummy_d_release(struct dentry *dentry)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+ return -1;
+}
+
+void dummy_d_iput(struct dentry *dentry, struct inode *inode)
+{
+ printk("Dummy VFS function %s called!\n", __FUNCTION__);
+}
+
+struct inode_operations dummy_i_op = {
+ dummy_create,
+ dummy_lookup,
+ dummy_link,
+ dummy_unlink,
+ dummy_symlink,
+ dummy_mkdir,
+ dummy_rmdir,
+ dummy_mknod,
+ dummy_rename,
+ dummy_readlink,
+ dummy_truncate,
+ dummy_permission,
+};
+
+struct dentry_operations dummy_d_op = {
+ dummy_d_revalidate,
+ dummy_d_hash,
+ dummy_d_compare,
+ dummy_d_delete,
+ dummy_d_release,
+ dummy_d_iput,
+};
projects / akaros.git / blobdiff