kern/src/ns/fs_file.c

/* Copyright (c) 2018 Google Inc
 * Barret Rhoden <brho@cs.berkeley.edu>
 * See LICENSE for details.
 *
 * fs_file: structs and helpers for files for 9ns devices
 */

#include <fs_file.h>
#include <kmalloc.h>
#include <string.h>
#include <stdio.h>
#include <assert.h>
#include <error.h>
#include <umem.h>
#include <pmap.h>

/* Initializes a zalloced fs_file.  The caller is responsible for filling in
 * dir, except for name.  Most fields are fine with being zeroed.  Note the kref
 * == 0 too. */
void fs_file_init(struct fs_file *f, const char *name, struct fs_file_ops *ops)
{
	qlock_init(&f->qlock);
	fs_file_set_basename(f, name);
	f->ops = ops;
	/* TODO: consider holding off on initializing the PM, since only walked
	 * and opened entries could use it.  pm == NULL means no PM yet.
	 * Negative entries will never be used in this manner.  Doing it now
	 * avoids races, though it's mostly zeroing cache-hot fields. */
	f->pm = &f->static_pm;
	pm_init(f->pm, (struct page_map_operations*)ops, f);
}

void fs_file_set_basename(struct fs_file *f, const char *name)
{
	size_t name_len = strlen(name) + 1;

	if (name_len > KNAMELEN)
		f->dir.name = kzmalloc(name_len, MEM_WAIT);
	else
		f->dir.name = f->static_name;
	memcpy(f->dir.name, name, name_len);
}

/* Technically, a reader could see the old string pointer and read it.  That
 * memory could be realloced and used for something else.  But thanks to the
 * seqctr, the reader will retry.  Otherwise, we might not need the seqctr,
 * since we never change_basename when a file is in a tree.  So far.
 *
 * The only reader that races with setting the name is stat.  Regular lookups
 * won't see the file, since it was removed from the HT, and readdirs won't see
 * it due to the parent's qlock. */
void fs_file_change_basename(struct fs_file *f, const char *name)
{
	char *old_name = NULL;
	char *new_name = NULL;
	size_t name_len = strlen(name) + 1;

	if (name_len > KNAMELEN)
		new_name = kzmalloc(name_len, MEM_WAIT);
	qlock(&f->qlock);
	if (f->dir.name != f->static_name)
		old_name = f->dir.name;
	if (new_name)
		f->dir.name = new_name;
	else
		f->dir.name = f->static_name;
	memcpy(f->dir.name, name, name_len);
	/* TODO: if we store the hash of the name in the file, do so here. */
	qunlock(&f->qlock);
	kfree(old_name);
}

/* Helper for building a dir.  Caller sets qid path and vers.  YMMV. */
void fs_file_init_dir(struct fs_file *f, uint16_t dir_type, uint32_t dir_dev,
                      struct username *user, int perm)
{
	struct dir *dir = &f->dir;

	if (perm & DMDIR)
		dir->qid.type |= QTDIR;
	if (perm & DMAPPEND)
		dir->qid.type |= QTAPPEND;
	if (perm & DMEXCL)
		dir->qid.type |= QTEXCL;
	if (perm & DMSYMLINK)
		dir->qid.type |= QTSYMLINK;
	dir->type = dir_type;
	dir->dev = dir_dev;
	/* dir->mode stores all the DM bits, but note that userspace can only
	 * affect the permissions (S_PMASK) bits. */
	dir->mode = perm;
	__set_acmtime(f, FSF_ATIME | FSF_BTIME | FSF_MTIME | FSF_CTIME);
	dir->length = 0;
	/* TODO: this is a mess if you use anything other than eve.  If you use
	 * a process, that memory is sitting in the proc struct, but we have
	 * weak refs on it.  What happens when that proc exits?  Disaster. */
	assert(user == &eve);
	dir->uid = user->name;
	dir->gid = user->name;
	dir->muid = user->name;
}

static char *copy_str(const char *s)
{
	char *ret;
	size_t sz;

	if (!s)
		return NULL;
	sz = strlen(s) + 1;
	ret = kmalloc(sz, MEM_WAIT);
	memcpy(ret, s, sz);
	return ret;
}

/* Deep copies the contents of dir into the fs_file's dir. */
void fs_file_copy_from_dir(struct fs_file *f, struct dir *dir)
{
	memcpy(&f->dir, dir, sizeof(struct dir));
	fs_file_set_basename(f, dir->name);
	/* TODO: sort out usernames.  Not only are these just eve, but they are
	 * not struct user or something and they ignore whatever the name was
	 * from the remote end. */
	f->dir.uid = eve.name;
	f->dir.gid = eve.name;
	f->dir.muid = eve.name;
	f->dir.ext = copy_str(dir->ext);
}

void cleanup_fs_file(struct fs_file *f)
{
	if (f->dir.name != f->static_name)
		kfree(f->dir.name);
	/* TODO: Not sure if these will be refcounted objects in the future.
	 * Keep this in sync with other code that manages/sets uid/gid/muid. */
	f->dir.uid = NULL;
	f->dir.gid = NULL;
	f->dir.muid = NULL;
	if (f->dir.ext)
		kfree(f->dir.ext);
	f->dir.ext = NULL;
	pm_destroy(f->pm);
	/* Might share mappings in the future.  Catch it here. */
	assert(f->pm == &f->static_pm);
}

void __set_acmtime_to(struct fs_file *f, int which, struct timespec *t)
{
	/* WRITE_ONCE, due to lockless peakers */
	if (which & FSF_ATIME) {
		WRITE_ONCE(f->dir.atime.tv_sec, t->tv_sec);
		WRITE_ONCE(f->dir.atime.tv_nsec, t->tv_nsec);
	}
	if (which & FSF_BTIME) {
		WRITE_ONCE(f->dir.btime.tv_sec, t->tv_sec);
		WRITE_ONCE(f->dir.btime.tv_nsec, t->tv_nsec);
	}
	if (which & FSF_CTIME) {
		WRITE_ONCE(f->dir.ctime.tv_sec, t->tv_sec);
		WRITE_ONCE(f->dir.ctime.tv_nsec, t->tv_nsec);
	}
	if (which & FSF_MTIME) {
		WRITE_ONCE(f->dir.mtime.tv_sec, t->tv_sec);
		WRITE_ONCE(f->dir.mtime.tv_nsec, t->tv_nsec);
	}
}

/* Caller should hold f's qlock */
void __set_acmtime(struct fs_file *f, int which)
{
	struct timespec now = nsec2timespec(epoch_nsec());

	__set_acmtime_to(f, which, &now);
}

/* Recall that the frontend always has the most up-to-date info.  This gets
 * synced to the backend when we flush or fsync. */
void set_acmtime_to(struct fs_file *f, int which, struct timespec *t)
{
	ERRSTACK(1);

	qlock(&f->qlock);
	if (waserror()) {
		qunlock(&f->qlock);
		nexterror();
	}
	if ((which & FSF_ATIME) && !caller_has_file_perms(f, O_READ))
		error(EPERM, "insufficient perms to set atime");
	if ((which & FSF_BTIME) && !caller_is_username(f->dir.uid))
		error(EPERM, "insufficient perms to set btime");
	if ((which & FSF_CTIME) && !caller_has_file_perms(f, O_WRITE))
		error(EPERM, "insufficient perms to set ctime");
	if ((which & FSF_MTIME) && !caller_has_file_perms(f, O_WRITE))
		error(EPERM, "insufficient perms to set mtime");
	__set_acmtime_to(f, which, t);
	qunlock(&f->qlock);
	poperror();
}

void set_acmtime_noperm(struct fs_file *f, int which)
{
	struct timespec now = nsec2timespec(epoch_nsec());

	/* <3 atime.  We'll go with an hour resolution, like NTFS. */
	if (which == FSF_ATIME) {
		if (now.tv_sec < ACCESS_ONCE(f->dir.atime.tv_sec) + 3600)
			return;
	}
	qlock(&f->qlock);
	__set_acmtime_to(f, which, &now);
	qunlock(&f->qlock);
}

size_t fs_file_stat(struct fs_file *f, uint8_t *m_buf, size_t m_buf_sz)
{
	size_t ret;

	qlock(&f->qlock);
	ret = convD2M(&f->dir, m_buf, m_buf_sz);
	qunlock(&f->qlock);
	if (ret <= BIT16SZ)
		error(EINVAL, "buffer too small for stat");
	return ret;
}

/* Helper: update file metadata after a write */
static void write_metadata(struct fs_file *f, off64_t offset,
                           bool always_update_len)
{
	qlock(&f->qlock);
	f->flags |= FSF_DIRTY;
	if (always_update_len || (offset > f->dir.length))
		WRITE_ONCE(f->dir.length, offset);
	__set_acmtime(f, FSF_MTIME | FSF_CTIME);
	qunlock(&f->qlock);
}

/* Punches a hole from begin to end.  Pages completely in the hole will be
 * removed.  Otherwise, the edges will be zeroed.
 *
 * Concurrent truncates with reads and writes can lead to weird data.
 * truncate()/punch_hole will attempt to remove data in page-sized chunks.
 * Concurrent users (with a PM refcnt, under the current code) will prevent
 * removal.  punch_hole will memset those areas to 0, similar to a concurrent
 * write.
 *
 * read() will return data that is up to the file's length.  write() and
 * punch_hole() will add or remove data and set the length.  When adding data
 * (write), we add it first, then set the len.  When removing data, we set the
 * len, then remove it.  If you mix those ops, the len could be set above an
 * area where the data is still being mucked with.  read/write/mmap all grab
 * references to the PM's page slot, locking the page in the page cache for a
 * little.  Truncate often won't remove those pages, but it will try to zero
 * them.  reads and mmaps will check the length on their own, while it is being
 * changed by other ops.
 *
 * A few examples:
 * - Trunc to 0 during write to N.  A reader might get zeros instead of the data
 *   written (trunc was dropping/zeroing the pages after write wrote them).
 * - Trunc to 0, trunc back to N, with concurrent reads/mmaps of the area in
 *   between: a reader might see the old data or tears in the data.
 * - mmaps of pages in a region that gets hole-punched and faults at the same
 *   time might not get a SIGBUS / ESPIPE.  That length check is best effort.
 * - After we remove hole pages from the page cache, but before we tell the
 *   backend, a read/write/mmap-fault to a page in the hole could fetch the old
 *   data from the backend before the FS op removes the data from the backend,
 *   and we'd end up with some old data.  The root issue here is that the
 *   frontend is a cache, and it has the most recent version of the data.  In
 *   the case of hole punching, we want there to be an absence of data.
 *   Technically, we could have zeroed pages, but we don't want the overhead of
 *   that.  So we drop the pages - that situation looks the same as not having
 *   the data in the cache/frontend.
 *
 * To prevent these things, we could qlock the entire file during all ops, or
 * even just for trunc, write, and loading pages into the PM for read.  That was
 * my first version of this.  But you can imagine backends that don't require
 * this sort of serialization (e.g. ramfs, future #mnts, etc), and it
 * complicates some mmap / pagemap code.  If you want the qlock to protect the
 * invariant (one of which is that the file's contents below len are always
 * valid; another is that hole punches don't keep old data), then we can add
 * some sort of file locking.
 */
static void fs_file_punch_hole(struct fs_file *f, off64_t begin, off64_t end)
{
	size_t first_pg_idx, last_pg_idx, nr_pages, zero_amt;
	struct page *page;
	int error;

	/* Caller should check for this */
	assert((long)begin >= 0);
	assert((long)end >= 0);
	if (end <= begin)
		return;
	/* We're punching for the range [begin, end), but inclusive for the
	 * pages: [first_pg_idx, last_pg_idx]. */
	first_pg_idx = LA2PPN(begin);
	last_pg_idx = LA2PPN(ROUNDUP(end, PGSIZE)) - 1;
	nr_pages = last_pg_idx - first_pg_idx + 1;
	if (PGOFF(begin)) {
		error = pm_load_page(f->pm, first_pg_idx, &page);
		if (error)
			error(-error, "punch_hole pm_load_page failed");
		zero_amt = MIN(PGSIZE - PGOFF(begin), end - begin);
		memset(page2kva(page) + PGOFF(begin), 0, zero_amt);
		atomic_or(&page->pg_flags, PG_DIRTY);
		pm_put_page(page);
		first_pg_idx++;
		nr_pages--;
		if (!nr_pages)
			return;
	}
	if (PGOFF(end)) {
		/* if this unaligned end is beyond the EOF, we might pull in a
		 * page of zeros, then zero the first part of it. */
		error = pm_load_page(f->pm, last_pg_idx, &page);
		if (error)
			error(-error, "punch_hole pm_load_page failed");
		memset(page2kva(page), 0, PGOFF(end));
		atomic_or(&page->pg_flags, PG_DIRTY);
		pm_put_page(page);
		last_pg_idx--;
		nr_pages--;
		if (!nr_pages)
			return;
	}
	pm_remove_or_zero_pages(f->pm, first_pg_idx, nr_pages);
	/* After we removed the pages from the PM, but before we tell the
	 * backend, someone could load a backend page.  Note that we only tell
	 * the backend about the intermediate pages - we already dealt with the
	 * edge pages above, and the PM has the latest, dirty version of them.
	 * */
	f->ops->punch_hole(f, first_pg_idx << PGSHIFT,
	                   (first_pg_idx + nr_pages) << PGSHIFT);
}

void fs_file_truncate(struct fs_file *f, off64_t to)
{
	off64_t old_len = fs_file_get_length(f);

	fs_file_perm_check(f, O_WRITE);
	if ((to > old_len) && !f->ops->can_grow_to(f, to))
		error(EINVAL, "can't grow file to %lu bytes", to);
	write_metadata(f, to, true);
	if (to < old_len) {
		/* Round up the old_len to avoid making an unnecessary partial
		 * page of zeros at the end of the file. */
		fs_file_punch_hole(f, to, ROUNDUP(old_len, PGSIZE));
	}
}

/* Standard read.  We sync with write, in that once the length is set, we'll
 * attempt to read those bytes. */
size_t fs_file_read(struct fs_file *f, uint8_t *buf, size_t count,
                    off64_t offset)
{
	ERRSTACK(1);
	struct page *page;
	size_t copy_amt, pg_off, pg_idx, total_remaining;
	volatile size_t so_far = 0;		/* volatile for waserror */
	const uint8_t *buf_end = buf + count;
	int error;

	/* These errors should have been caught by higher level code */
	if ((uintptr_t)buf + count < (uintptr_t)buf)
		panic("Bad buf %p + count %p", buf, count);
	if (offset + count < offset)
		panic("Bad offset %p + count %p", offset, count);
	if (waserror()) {
		if (so_far) {
			poperror();
			return so_far;
		}
		nexterror();
	}
	while (buf < buf_end) {
		/* Check early, so we don't load pages beyond length needlessly.
		 * The PM/FSF op might just create zeroed pages when asked. */
		if (offset + so_far >= fs_file_get_length(f))
			break;
		pg_off = PGOFF(offset + so_far);
		pg_idx = LA2PPN(offset + so_far);
		error = pm_load_page(f->pm, pg_idx, &page);
		if (error)
			error(-error, "read pm_load_page failed");
		copy_amt = MIN(PGSIZE - pg_off, buf_end - buf);
		/* Lockless peak.  Check the len so we don't read beyond EOF.
		 * We have a page, but we don't necessarily have access to all
		 * of it. */
		total_remaining = fs_file_get_length(f) - (offset + so_far);
		if (copy_amt > total_remaining) {
			copy_amt = total_remaining;
			buf_end = buf + copy_amt;
		}
		memcpy_to_safe(buf, page2kva(page) + pg_off, copy_amt);
		buf += copy_amt;
		so_far += copy_amt;
		pm_put_page(page);
	}
	if (so_far)
		set_acmtime_noperm(f, FSF_ATIME);
	poperror();
	return so_far;
}

size_t fs_file_write(struct fs_file *f, const uint8_t *buf, size_t count,
                     off64_t offset)
{
	ERRSTACK(1);
	struct page *page;
	size_t copy_amt, pg_off, pg_idx;
	volatile size_t so_far = 0;		/* volatile for waserror */
	const uint8_t *buf_end = buf + count;
	int error;

	/* These errors should have been caught by higher level code */
	if ((uintptr_t)buf + count < (uintptr_t)buf)
		panic("Bad buf %p + count %p", buf, count);
	if (offset + count < offset)
		panic("Bad offset %p + count %p", offset, count);
	if (waserror()) {
		if (so_far) {
			write_metadata(f, offset + so_far, false);
			poperror();
			return so_far;
		}
		nexterror();
	};
	if (offset + count > fs_file_get_length(f)) {
		if (!f->ops->can_grow_to(f, offset + count))
			error(EINVAL, "can't write file to %lu bytes", offset +
			      count);
	}
	while (buf < buf_end) {
		pg_off = PGOFF(offset + so_far);
		pg_idx = LA2PPN(offset + so_far);
		error = pm_load_page(f->pm, pg_idx, &page);
		if (error)
			error(-error, "write pm_load_page failed");
		copy_amt = MIN(PGSIZE - pg_off, buf_end - buf);
		/* TODO: If you ask the kernel to write from a user address that
		 * will page fault, the memcpy will fail and we'll move on to
		 * the next region.  To avoid leaving a chunk of uninitialized
		 * memory, we'll zero it out in the page cache.  Otherwise the
		 * user could come back and read old kernel data.
		 *
		 * The real fix will be to have the kernel throw an error if it
		 * was a segfault or block if it wasn't.  Note that when the
		 * kernel attempts to access the user's page, it does so with a
		 * handle_page_fault_nofile, which won't attempt to handle
		 * file-backed VMRs *even if* the file is in the page cache.
		 * Yikes! */
		if (memcpy_from_safe(page2kva(page) + pg_off, buf, copy_amt))
			memset(page2kva(page) + pg_off, 0, copy_amt);
		buf += copy_amt;
		so_far += copy_amt;
		atomic_or(&page->pg_flags, PG_DIRTY);
		pm_put_page(page);
	}
	assert(buf == buf_end);
	assert(count == so_far);
	/* We set the len *after* writing for our lockless reads.  If we set len
	 * before, then read() could start as soon as we loaded the page (all
	 * zeros), but before we wrote the actual data.  They'd get zeros
	 * instead of what we added. */
	write_metadata(f, offset + so_far, false);
	poperror();
	return so_far;
}

static void wstat_mode(struct fs_file *f, int new_mode)
{
	ERRSTACK(1);
	int mode;

	qlock(&f->qlock);
	if (waserror()) {
		qunlock(&f->qlock);
		nexterror();
	}
	if (!caller_is_username(f->dir.uid))
		error(EPERM, "wrong user for wstat, need %s", f->dir.uid);
	/* Only allowing changes in permissions, not random stuff like whether
	 * it is a directory or symlink. */
	static_assert(!(DMMODE_BITS & S_PMASK));
	mode = (f->dir.mode & ~S_PMASK) | (new_mode & S_PMASK);
	WRITE_ONCE(f->dir.mode, mode);
	__set_acmtime(f, FSF_CTIME);
	qunlock(&f->qlock);
	poperror();
}

size_t fs_file_wstat(struct fs_file *f, uint8_t *m_buf, size_t m_buf_sz)
{
	struct dir *m_dir;
	size_t m_sz;

	/* common trick in wstats.  we want the dir and any strings in the M.
	 * the strings are smaller than the entire M (which is strings plus the
	 * real dir M).  the strings will be placed right after the dir
	 * (dir[1]). */
	m_dir = kzmalloc(sizeof(struct dir) + m_buf_sz, MEM_WAIT);
	m_sz = convM2D(m_buf, m_buf_sz, &m_dir[0], (char*)&m_dir[1]);
	if (!m_sz) {
		kfree(m_dir);
		error(ENODATA, "couldn't convM2D");
	}
	/* We'll probably have similar issues for all of the strings.  At that
	 * point, we might not even bother reading the strings in. */
	if (!emptystr(m_dir->name))
		error(EINVAL, "do not rename with wstat");
	if (m_dir->mode != -1)
		wstat_mode(f, m_dir->mode);
	if (m_dir->length != -1)
		fs_file_truncate(f, m_dir->length);
	if ((int64_t)m_dir->atime.tv_sec != -1)
		set_acmtime_to(f, FSF_ATIME, &m_dir->atime);
	if ((int64_t)m_dir->btime.tv_sec != -1)
		set_acmtime_to(f, FSF_BTIME, &m_dir->btime);
	if ((int64_t)m_dir->ctime.tv_sec != -1)
		set_acmtime_to(f, FSF_CTIME, &m_dir->ctime);
	if ((int64_t)m_dir->mtime.tv_sec != -1)
		set_acmtime_to(f, FSF_MTIME, &m_dir->mtime);
	/* TODO: handle uid/gid/muid changes */
	kfree(m_dir);
	return m_sz;
}