2005-09-09 22:10:30 +02:00
|
|
|
/*
|
|
|
|
FUSE: Filesystem in Userspace
|
2006-04-11 21:16:51 +02:00
|
|
|
Copyright (C) 2001-2006 Miklos Szeredi <miklos@szeredi.hu>
|
2005-09-09 22:10:30 +02:00
|
|
|
|
|
|
|
This program can be distributed under the terms of the GNU GPL.
|
|
|
|
See the file COPYING.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "fuse_i.h"
|
|
|
|
|
|
|
|
#include <linux/pagemap.h>
|
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/kernel.h>
|
Detach sched.h from mm.h
First thing mm.h does is including sched.h solely for can_do_mlock() inline
function which has "current" dereference inside. By dealing with can_do_mlock()
mm.h can be detached from sched.h which is good. See below, why.
This patch
a) removes unconditional inclusion of sched.h from mm.h
b) makes can_do_mlock() normal function in mm/mlock.c
c) exports can_do_mlock() to not break compilation
d) adds sched.h inclusions back to files that were getting it indirectly.
e) adds less bloated headers to some files (asm/signal.h, jiffies.h) that were
getting them indirectly
Net result is:
a) mm.h users would get less code to open, read, preprocess, parse, ... if
they don't need sched.h
b) sched.h stops being dependency for significant number of files:
on x86_64 allmodconfig touching sched.h results in recompile of 4083 files,
after patch it's only 3744 (-8.3%).
Cross-compile tested on
all arm defconfigs, all mips defconfigs, all powerpc defconfigs,
alpha alpha-up
arm
i386 i386-up i386-defconfig i386-allnoconfig
ia64 ia64-up
m68k
mips
parisc parisc-up
powerpc powerpc-up
s390 s390-up
sparc sparc-up
sparc64 sparc64-up
um-x86_64
x86_64 x86_64-up x86_64-defconfig x86_64-allnoconfig
as well as my two usual configs.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-20 23:22:52 +02:00
|
|
|
#include <linux/sched.h>
|
2005-09-09 22:10:30 +02:00
|
|
|
|
2006-03-28 11:56:42 +02:00
|
|
|
static const struct file_operations fuse_direct_io_file_operations;
|
2005-09-09 22:10:37 +02:00
|
|
|
|
2005-11-07 09:59:51 +01:00
|
|
|
static int fuse_send_open(struct inode *inode, struct file *file, int isdir,
|
|
|
|
struct fuse_open_out *outargp)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_open_in inarg;
|
2005-11-07 09:59:51 +01:00
|
|
|
struct fuse_req *req;
|
|
|
|
int err;
|
|
|
|
|
2006-04-11 07:54:58 +02:00
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req))
|
|
|
|
return PTR_ERR(req);
|
2005-11-07 09:59:51 +01:00
|
|
|
|
|
|
|
memset(&inarg, 0, sizeof(inarg));
|
2007-10-18 12:07:02 +02:00
|
|
|
inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
|
|
|
|
if (!fc->atomic_o_trunc)
|
|
|
|
inarg.flags &= ~O_TRUNC;
|
2005-11-07 09:59:51 +01:00
|
|
|
req->in.h.opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
|
|
|
|
req->in.h.nodeid = get_node_id(inode);
|
|
|
|
req->in.numargs = 1;
|
|
|
|
req->in.args[0].size = sizeof(inarg);
|
|
|
|
req->in.args[0].value = &inarg;
|
|
|
|
req->out.numargs = 1;
|
|
|
|
req->out.args[0].size = sizeof(*outargp);
|
|
|
|
req->out.args[0].value = outargp;
|
|
|
|
request_send(fc, req);
|
|
|
|
err = req->out.h.error;
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct fuse_file *fuse_file_alloc(void)
|
|
|
|
{
|
|
|
|
struct fuse_file *ff;
|
|
|
|
ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
|
|
|
|
if (ff) {
|
2006-06-25 14:48:52 +02:00
|
|
|
ff->reserved_req = fuse_request_alloc();
|
|
|
|
if (!ff->reserved_req) {
|
2005-11-07 09:59:51 +01:00
|
|
|
kfree(ff);
|
|
|
|
ff = NULL;
|
2007-11-15 02:00:02 +01:00
|
|
|
} else {
|
|
|
|
INIT_LIST_HEAD(&ff->write_entry);
|
|
|
|
atomic_set(&ff->count, 0);
|
2005-11-07 09:59:51 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
return ff;
|
|
|
|
}
|
|
|
|
|
|
|
|
void fuse_file_free(struct fuse_file *ff)
|
|
|
|
{
|
2006-06-25 14:48:52 +02:00
|
|
|
fuse_request_free(ff->reserved_req);
|
2005-11-07 09:59:51 +01:00
|
|
|
kfree(ff);
|
|
|
|
}
|
|
|
|
|
2007-10-17 08:31:00 +02:00
|
|
|
static struct fuse_file *fuse_file_get(struct fuse_file *ff)
|
|
|
|
{
|
|
|
|
atomic_inc(&ff->count);
|
|
|
|
return ff;
|
|
|
|
}
|
|
|
|
|
2007-10-17 08:31:04 +02:00
|
|
|
static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
|
|
|
|
{
|
2008-02-06 10:38:39 +01:00
|
|
|
dput(req->misc.release.dentry);
|
|
|
|
mntput(req->misc.release.vfsmount);
|
2007-10-17 08:31:04 +02:00
|
|
|
fuse_put_request(fc, req);
|
|
|
|
}
|
|
|
|
|
2007-10-17 08:31:00 +02:00
|
|
|
static void fuse_file_put(struct fuse_file *ff)
|
|
|
|
{
|
|
|
|
if (atomic_dec_and_test(&ff->count)) {
|
|
|
|
struct fuse_req *req = ff->reserved_req;
|
2008-02-06 10:38:39 +01:00
|
|
|
struct inode *inode = req->misc.release.dentry->d_inode;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
2007-10-17 08:31:04 +02:00
|
|
|
req->end = fuse_release_end;
|
2007-10-17 08:31:00 +02:00
|
|
|
request_send_background(fc, req);
|
|
|
|
kfree(ff);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2005-11-07 09:59:51 +01:00
|
|
|
void fuse_finish_open(struct inode *inode, struct file *file,
|
|
|
|
struct fuse_file *ff, struct fuse_open_out *outarg)
|
|
|
|
{
|
|
|
|
if (outarg->open_flags & FOPEN_DIRECT_IO)
|
|
|
|
file->f_op = &fuse_direct_io_file_operations;
|
|
|
|
if (!(outarg->open_flags & FOPEN_KEEP_CACHE))
|
2007-10-17 08:31:01 +02:00
|
|
|
invalidate_inode_pages2(inode->i_mapping);
|
2005-11-07 09:59:51 +01:00
|
|
|
ff->fh = outarg->fh;
|
2007-10-17 08:31:00 +02:00
|
|
|
file->private_data = fuse_file_get(ff);
|
2005-11-07 09:59:51 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
int fuse_open_common(struct inode *inode, struct file *file, int isdir)
|
|
|
|
{
|
2005-09-09 22:10:30 +02:00
|
|
|
struct fuse_open_out outarg;
|
|
|
|
struct fuse_file *ff;
|
|
|
|
int err;
|
|
|
|
|
2005-09-30 20:59:02 +02:00
|
|
|
/* VFS checks this, but only _after_ ->open() */
|
|
|
|
if (file->f_flags & O_DIRECT)
|
|
|
|
return -EINVAL;
|
|
|
|
|
2005-09-09 22:10:30 +02:00
|
|
|
err = generic_file_open(inode, file);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
2005-11-07 09:59:51 +01:00
|
|
|
ff = fuse_file_alloc();
|
2005-09-09 22:10:30 +02:00
|
|
|
if (!ff)
|
2005-11-07 09:59:51 +01:00
|
|
|
return -ENOMEM;
|
2005-09-09 22:10:30 +02:00
|
|
|
|
2005-11-07 09:59:51 +01:00
|
|
|
err = fuse_send_open(inode, file, isdir, &outarg);
|
|
|
|
if (err)
|
|
|
|
fuse_file_free(ff);
|
|
|
|
else {
|
|
|
|
if (isdir)
|
|
|
|
outarg.open_flags &= ~FOPEN_DIRECT_IO;
|
|
|
|
fuse_finish_open(inode, file, ff, &outarg);
|
2005-09-09 22:10:30 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2007-10-17 08:31:00 +02:00
|
|
|
void fuse_release_fill(struct fuse_file *ff, u64 nodeid, int flags, int opcode)
|
2006-01-17 07:14:42 +01:00
|
|
|
{
|
2006-06-25 14:48:52 +02:00
|
|
|
struct fuse_req *req = ff->reserved_req;
|
2008-02-06 10:38:39 +01:00
|
|
|
struct fuse_release_in *inarg = &req->misc.release.in;
|
2005-09-09 22:10:30 +02:00
|
|
|
|
|
|
|
inarg->fh = ff->fh;
|
2005-11-07 09:59:51 +01:00
|
|
|
inarg->flags = flags;
|
2006-06-25 14:48:50 +02:00
|
|
|
req->in.h.opcode = opcode;
|
2005-11-07 09:59:51 +01:00
|
|
|
req->in.h.nodeid = nodeid;
|
2005-09-09 22:10:30 +02:00
|
|
|
req->in.numargs = 1;
|
|
|
|
req->in.args[0].size = sizeof(struct fuse_release_in);
|
|
|
|
req->in.args[0].value = inarg;
|
2005-11-07 09:59:51 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
int fuse_release_common(struct inode *inode, struct file *file, int isdir)
|
|
|
|
{
|
|
|
|
struct fuse_file *ff = file->private_data;
|
|
|
|
if (ff) {
|
2007-10-18 12:07:03 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
2008-02-06 10:38:39 +01:00
|
|
|
struct fuse_req *req = ff->reserved_req;
|
2007-10-18 12:07:03 +02:00
|
|
|
|
2007-10-17 08:31:00 +02:00
|
|
|
fuse_release_fill(ff, get_node_id(inode), file->f_flags,
|
|
|
|
isdir ? FUSE_RELEASEDIR : FUSE_RELEASE);
|
2006-06-25 14:48:50 +02:00
|
|
|
|
|
|
|
/* Hold vfsmount and dentry until release is finished */
|
2008-02-06 10:38:39 +01:00
|
|
|
req->misc.release.vfsmount = mntget(file->f_path.mnt);
|
|
|
|
req->misc.release.dentry = dget(file->f_path.dentry);
|
2007-10-18 12:07:03 +02:00
|
|
|
|
|
|
|
spin_lock(&fc->lock);
|
|
|
|
list_del(&ff->write_entry);
|
|
|
|
spin_unlock(&fc->lock);
|
2007-10-17 08:31:00 +02:00
|
|
|
/*
|
|
|
|
* Normally this will send the RELEASE request,
|
|
|
|
* however if some asynchronous READ or WRITE requests
|
|
|
|
* are outstanding, the sending will be delayed
|
|
|
|
*/
|
|
|
|
fuse_file_put(ff);
|
2005-11-07 09:59:51 +01:00
|
|
|
}
|
2005-09-09 22:10:30 +02:00
|
|
|
|
|
|
|
/* Return value is ignored by VFS */
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2005-09-09 22:10:36 +02:00
|
|
|
static int fuse_open(struct inode *inode, struct file *file)
|
|
|
|
{
|
|
|
|
return fuse_open_common(inode, file, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_release(struct inode *inode, struct file *file)
|
|
|
|
{
|
|
|
|
return fuse_release_common(inode, file, 0);
|
|
|
|
}
|
|
|
|
|
2006-06-25 14:48:52 +02:00
|
|
|
/*
|
2006-06-25 14:48:55 +02:00
|
|
|
* Scramble the ID space with XTEA, so that the value of the files_struct
|
|
|
|
* pointer is not exposed to userspace.
|
2006-06-25 14:48:52 +02:00
|
|
|
*/
|
2007-10-18 12:07:04 +02:00
|
|
|
u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
|
2006-06-25 14:48:52 +02:00
|
|
|
{
|
2006-06-25 14:48:55 +02:00
|
|
|
u32 *k = fc->scramble_key;
|
|
|
|
u64 v = (unsigned long) id;
|
|
|
|
u32 v0 = v;
|
|
|
|
u32 v1 = v >> 32;
|
|
|
|
u32 sum = 0;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < 32; i++) {
|
|
|
|
v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
|
|
|
|
sum += 0x9E3779B9;
|
|
|
|
v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
|
|
|
|
}
|
|
|
|
|
|
|
|
return (u64) v0 + ((u64) v1 << 32);
|
2006-06-25 14:48:52 +02:00
|
|
|
}
|
|
|
|
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
/*
|
|
|
|
* Check if page is under writeback
|
|
|
|
*
|
|
|
|
* This is currently done by walking the list of writepage requests
|
|
|
|
* for the inode, which can be pretty inefficient.
|
|
|
|
*/
|
|
|
|
static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
|
|
|
|
{
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
struct fuse_req *req;
|
|
|
|
bool found = false;
|
|
|
|
|
|
|
|
spin_lock(&fc->lock);
|
|
|
|
list_for_each_entry(req, &fi->writepages, writepages_entry) {
|
|
|
|
pgoff_t curr_index;
|
|
|
|
|
|
|
|
BUG_ON(req->inode != inode);
|
|
|
|
curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
|
|
|
|
if (curr_index == index) {
|
|
|
|
found = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
spin_unlock(&fc->lock);
|
|
|
|
|
|
|
|
return found;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Wait for page writeback to be completed.
|
|
|
|
*
|
|
|
|
* Since fuse doesn't rely on the VM writeback tracking, this has to
|
|
|
|
* use some other means.
|
|
|
|
*/
|
|
|
|
static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
|
|
|
|
{
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
|
|
|
|
wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2006-06-23 11:05:12 +02:00
|
|
|
static int fuse_flush(struct file *file, fl_owner_t id)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
2006-12-08 11:37:02 +01:00
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
2005-09-09 22:10:30 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_file *ff = file->private_data;
|
|
|
|
struct fuse_req *req;
|
|
|
|
struct fuse_flush_in inarg;
|
|
|
|
int err;
|
|
|
|
|
2006-01-06 09:19:39 +01:00
|
|
|
if (is_bad_inode(inode))
|
|
|
|
return -EIO;
|
|
|
|
|
2005-09-09 22:10:30 +02:00
|
|
|
if (fc->no_flush)
|
|
|
|
return 0;
|
|
|
|
|
2006-06-25 14:48:52 +02:00
|
|
|
req = fuse_get_req_nofail(fc, file);
|
2005-09-09 22:10:30 +02:00
|
|
|
memset(&inarg, 0, sizeof(inarg));
|
|
|
|
inarg.fh = ff->fh;
|
2006-06-25 14:48:55 +02:00
|
|
|
inarg.lock_owner = fuse_lock_owner_id(fc, id);
|
2005-09-09 22:10:30 +02:00
|
|
|
req->in.h.opcode = FUSE_FLUSH;
|
|
|
|
req->in.h.nodeid = get_node_id(inode);
|
|
|
|
req->in.numargs = 1;
|
|
|
|
req->in.args[0].size = sizeof(inarg);
|
|
|
|
req->in.args[0].value = &inarg;
|
2006-06-25 14:48:52 +02:00
|
|
|
req->force = 1;
|
2005-09-09 22:10:39 +02:00
|
|
|
request_send(fc, req);
|
2005-09-09 22:10:30 +02:00
|
|
|
err = req->out.h.error;
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
if (err == -ENOSYS) {
|
|
|
|
fc->no_flush = 1;
|
|
|
|
err = 0;
|
|
|
|
}
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
/*
|
|
|
|
* Wait for all pending writepages on the inode to finish.
|
|
|
|
*
|
|
|
|
* This is currently done by blocking further writes with FUSE_NOWRITE
|
|
|
|
* and waiting for all sent writes to complete.
|
|
|
|
*
|
|
|
|
* This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
|
|
|
|
* could conflict with truncation.
|
|
|
|
*/
|
|
|
|
static void fuse_sync_writes(struct inode *inode)
|
|
|
|
{
|
|
|
|
fuse_set_nowrite(inode);
|
|
|
|
fuse_release_nowrite(inode);
|
|
|
|
}
|
|
|
|
|
2005-09-09 22:10:38 +02:00
|
|
|
int fuse_fsync_common(struct file *file, struct dentry *de, int datasync,
|
|
|
|
int isdir)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
|
|
|
struct inode *inode = de->d_inode;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_file *ff = file->private_data;
|
|
|
|
struct fuse_req *req;
|
|
|
|
struct fuse_fsync_in inarg;
|
|
|
|
int err;
|
|
|
|
|
2006-01-06 09:19:39 +01:00
|
|
|
if (is_bad_inode(inode))
|
|
|
|
return -EIO;
|
|
|
|
|
2005-09-09 22:10:38 +02:00
|
|
|
if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
|
2005-09-09 22:10:30 +02:00
|
|
|
return 0;
|
|
|
|
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
/*
|
|
|
|
* Start writeback against all dirty pages of the inode, then
|
|
|
|
* wait for all outstanding writes, before sending the FSYNC
|
|
|
|
* request.
|
|
|
|
*/
|
|
|
|
err = write_inode_now(inode, 0);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
|
|
|
fuse_sync_writes(inode);
|
|
|
|
|
2006-04-11 07:54:58 +02:00
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req))
|
|
|
|
return PTR_ERR(req);
|
2005-09-09 22:10:30 +02:00
|
|
|
|
|
|
|
memset(&inarg, 0, sizeof(inarg));
|
|
|
|
inarg.fh = ff->fh;
|
|
|
|
inarg.fsync_flags = datasync ? 1 : 0;
|
2005-09-09 22:10:38 +02:00
|
|
|
req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
|
2005-09-09 22:10:30 +02:00
|
|
|
req->in.h.nodeid = get_node_id(inode);
|
|
|
|
req->in.numargs = 1;
|
|
|
|
req->in.args[0].size = sizeof(inarg);
|
|
|
|
req->in.args[0].value = &inarg;
|
|
|
|
request_send(fc, req);
|
|
|
|
err = req->out.h.error;
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
if (err == -ENOSYS) {
|
2005-09-09 22:10:38 +02:00
|
|
|
if (isdir)
|
|
|
|
fc->no_fsyncdir = 1;
|
|
|
|
else
|
|
|
|
fc->no_fsync = 1;
|
2005-09-09 22:10:30 +02:00
|
|
|
err = 0;
|
|
|
|
}
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2005-09-09 22:10:38 +02:00
|
|
|
static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
|
|
|
|
{
|
|
|
|
return fuse_fsync_common(file, de, datasync, 0);
|
|
|
|
}
|
|
|
|
|
2007-11-29 01:22:00 +01:00
|
|
|
void fuse_read_fill(struct fuse_req *req, struct file *file,
|
2006-01-17 07:14:45 +01:00
|
|
|
struct inode *inode, loff_t pos, size_t count, int opcode)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
2008-04-30 09:54:43 +02:00
|
|
|
struct fuse_read_in *inarg = &req->misc.read.in;
|
2007-11-29 01:22:00 +01:00
|
|
|
struct fuse_file *ff = file->private_data;
|
2005-09-09 22:10:30 +02:00
|
|
|
|
2006-01-17 07:14:45 +01:00
|
|
|
inarg->fh = ff->fh;
|
|
|
|
inarg->offset = pos;
|
|
|
|
inarg->size = count;
|
2007-11-29 01:22:00 +01:00
|
|
|
inarg->flags = file->f_flags;
|
2006-01-17 07:14:45 +01:00
|
|
|
req->in.h.opcode = opcode;
|
2005-09-09 22:10:30 +02:00
|
|
|
req->in.h.nodeid = get_node_id(inode);
|
|
|
|
req->in.numargs = 1;
|
|
|
|
req->in.args[0].size = sizeof(struct fuse_read_in);
|
2006-01-17 07:14:46 +01:00
|
|
|
req->in.args[0].value = inarg;
|
2005-09-09 22:10:30 +02:00
|
|
|
req->out.argpages = 1;
|
|
|
|
req->out.argvar = 1;
|
|
|
|
req->out.numargs = 1;
|
|
|
|
req->out.args[0].size = count;
|
|
|
|
}
|
|
|
|
|
2006-01-17 07:14:28 +01:00
|
|
|
static size_t fuse_send_read(struct fuse_req *req, struct file *file,
|
2007-10-18 12:07:04 +02:00
|
|
|
struct inode *inode, loff_t pos, size_t count,
|
|
|
|
fl_owner_t owner)
|
2005-09-09 22:10:36 +02:00
|
|
|
{
|
2006-01-17 07:14:45 +01:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
2007-10-18 12:07:04 +02:00
|
|
|
|
2007-11-29 01:22:00 +01:00
|
|
|
fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
|
2007-10-18 12:07:04 +02:00
|
|
|
if (owner != NULL) {
|
2008-04-30 09:54:43 +02:00
|
|
|
struct fuse_read_in *inarg = &req->misc.read.in;
|
2007-10-18 12:07:04 +02:00
|
|
|
|
|
|
|
inarg->read_flags |= FUSE_READ_LOCKOWNER;
|
|
|
|
inarg->lock_owner = fuse_lock_owner_id(fc, owner);
|
|
|
|
}
|
2006-01-17 07:14:45 +01:00
|
|
|
request_send(fc, req);
|
|
|
|
return req->out.args[0].size;
|
2005-09-09 22:10:36 +02:00
|
|
|
}
|
|
|
|
|
2008-04-30 09:54:43 +02:00
|
|
|
static void fuse_read_update_size(struct inode *inode, loff_t size,
|
|
|
|
u64 attr_ver)
|
|
|
|
{
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
|
|
|
|
spin_lock(&fc->lock);
|
|
|
|
if (attr_ver == fi->attr_version && size < inode->i_size) {
|
|
|
|
fi->attr_version = ++fc->attr_version;
|
|
|
|
i_size_write(inode, size);
|
|
|
|
}
|
|
|
|
spin_unlock(&fc->lock);
|
|
|
|
}
|
|
|
|
|
2005-09-09 22:10:30 +02:00
|
|
|
static int fuse_readpage(struct file *file, struct page *page)
|
|
|
|
{
|
|
|
|
struct inode *inode = page->mapping->host;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
2006-01-06 09:19:39 +01:00
|
|
|
struct fuse_req *req;
|
2008-04-30 09:54:43 +02:00
|
|
|
size_t num_read;
|
|
|
|
loff_t pos = page_offset(page);
|
|
|
|
size_t count = PAGE_CACHE_SIZE;
|
|
|
|
u64 attr_ver;
|
2006-01-06 09:19:39 +01:00
|
|
|
int err;
|
|
|
|
|
|
|
|
err = -EIO;
|
|
|
|
if (is_bad_inode(inode))
|
|
|
|
goto out;
|
|
|
|
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
/*
|
|
|
|
* Page writeback can extend beyond the liftime of the
|
|
|
|
* page-cache page, so make sure we read a properly synced
|
|
|
|
* page.
|
|
|
|
*/
|
|
|
|
fuse_wait_on_page_writeback(inode, page->index);
|
|
|
|
|
2006-04-11 07:54:58 +02:00
|
|
|
req = fuse_get_req(fc);
|
|
|
|
err = PTR_ERR(req);
|
|
|
|
if (IS_ERR(req))
|
2005-09-09 22:10:30 +02:00
|
|
|
goto out;
|
|
|
|
|
2008-04-30 09:54:43 +02:00
|
|
|
attr_ver = fuse_get_attr_version(fc);
|
|
|
|
|
2005-09-09 22:10:30 +02:00
|
|
|
req->out.page_zeroing = 1;
|
|
|
|
req->num_pages = 1;
|
|
|
|
req->pages[0] = page;
|
2008-04-30 09:54:43 +02:00
|
|
|
num_read = fuse_send_read(req, file, inode, pos, count, NULL);
|
2005-09-09 22:10:30 +02:00
|
|
|
err = req->out.h.error;
|
|
|
|
fuse_put_request(fc, req);
|
2008-04-30 09:54:43 +02:00
|
|
|
|
|
|
|
if (!err) {
|
|
|
|
/*
|
|
|
|
* Short read means EOF. If file size is larger, truncate it
|
|
|
|
*/
|
|
|
|
if (num_read < count)
|
|
|
|
fuse_read_update_size(inode, pos + num_read, attr_ver);
|
|
|
|
|
2005-09-09 22:10:30 +02:00
|
|
|
SetPageUptodate(page);
|
2008-04-30 09:54:43 +02:00
|
|
|
}
|
|
|
|
|
2005-09-09 22:10:38 +02:00
|
|
|
fuse_invalidate_attr(inode); /* atime changed */
|
2005-09-09 22:10:30 +02:00
|
|
|
out:
|
|
|
|
unlock_page(page);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2006-01-17 07:14:46 +01:00
|
|
|
static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
|
2005-09-09 22:10:33 +02:00
|
|
|
{
|
2006-01-17 07:14:46 +01:00
|
|
|
int i;
|
2008-04-30 09:54:43 +02:00
|
|
|
size_t count = req->misc.read.in.size;
|
|
|
|
size_t num_read = req->out.args[0].size;
|
|
|
|
struct inode *inode = req->pages[0]->mapping->host;
|
2006-01-17 07:14:46 +01:00
|
|
|
|
2008-04-30 09:54:43 +02:00
|
|
|
/*
|
|
|
|
* Short read means EOF. If file size is larger, truncate it
|
|
|
|
*/
|
|
|
|
if (!req->out.h.error && num_read < count) {
|
|
|
|
loff_t pos = page_offset(req->pages[0]) + num_read;
|
|
|
|
fuse_read_update_size(inode, pos, req->misc.read.attr_ver);
|
|
|
|
}
|
|
|
|
|
|
|
|
fuse_invalidate_attr(inode); /* atime changed */
|
2006-01-17 07:14:46 +01:00
|
|
|
|
2005-09-09 22:10:33 +02:00
|
|
|
for (i = 0; i < req->num_pages; i++) {
|
|
|
|
struct page *page = req->pages[i];
|
|
|
|
if (!req->out.h.error)
|
|
|
|
SetPageUptodate(page);
|
2006-01-17 07:14:46 +01:00
|
|
|
else
|
|
|
|
SetPageError(page);
|
2005-09-09 22:10:33 +02:00
|
|
|
unlock_page(page);
|
|
|
|
}
|
2007-10-17 08:31:00 +02:00
|
|
|
if (req->ff)
|
|
|
|
fuse_file_put(req->ff);
|
2006-01-17 07:14:46 +01:00
|
|
|
fuse_put_request(fc, req);
|
|
|
|
}
|
|
|
|
|
2007-11-29 01:22:00 +01:00
|
|
|
static void fuse_send_readpages(struct fuse_req *req, struct file *file,
|
2006-01-17 07:14:46 +01:00
|
|
|
struct inode *inode)
|
|
|
|
{
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
loff_t pos = page_offset(req->pages[0]);
|
|
|
|
size_t count = req->num_pages << PAGE_CACHE_SHIFT;
|
|
|
|
req->out.page_zeroing = 1;
|
2007-11-29 01:22:00 +01:00
|
|
|
fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
|
2008-04-30 09:54:43 +02:00
|
|
|
req->misc.read.attr_ver = fuse_get_attr_version(fc);
|
2006-02-01 12:04:40 +01:00
|
|
|
if (fc->async_read) {
|
2007-11-29 01:22:00 +01:00
|
|
|
struct fuse_file *ff = file->private_data;
|
2007-10-17 08:31:00 +02:00
|
|
|
req->ff = fuse_file_get(ff);
|
2006-02-01 12:04:40 +01:00
|
|
|
req->end = fuse_readpages_end;
|
|
|
|
request_send_background(fc, req);
|
|
|
|
} else {
|
|
|
|
request_send(fc, req);
|
|
|
|
fuse_readpages_end(fc, req);
|
|
|
|
}
|
2005-09-09 22:10:33 +02:00
|
|
|
}
|
|
|
|
|
2007-10-17 08:31:00 +02:00
|
|
|
struct fuse_fill_data {
|
2005-09-09 22:10:33 +02:00
|
|
|
struct fuse_req *req;
|
2007-11-29 01:22:00 +01:00
|
|
|
struct file *file;
|
2005-09-09 22:10:33 +02:00
|
|
|
struct inode *inode;
|
|
|
|
};
|
|
|
|
|
|
|
|
static int fuse_readpages_fill(void *_data, struct page *page)
|
|
|
|
{
|
2007-10-17 08:31:00 +02:00
|
|
|
struct fuse_fill_data *data = _data;
|
2005-09-09 22:10:33 +02:00
|
|
|
struct fuse_req *req = data->req;
|
|
|
|
struct inode *inode = data->inode;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
fuse_wait_on_page_writeback(inode, page->index);
|
|
|
|
|
2005-09-09 22:10:33 +02:00
|
|
|
if (req->num_pages &&
|
|
|
|
(req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
|
|
|
|
(req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
|
|
|
|
req->pages[req->num_pages - 1]->index + 1 != page->index)) {
|
2007-11-29 01:22:00 +01:00
|
|
|
fuse_send_readpages(req, data->file, inode);
|
2006-04-11 07:54:58 +02:00
|
|
|
data->req = req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req)) {
|
2005-09-09 22:10:33 +02:00
|
|
|
unlock_page(page);
|
2006-04-11 07:54:58 +02:00
|
|
|
return PTR_ERR(req);
|
2005-09-09 22:10:33 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
req->pages[req->num_pages] = page;
|
|
|
|
req->num_pages ++;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_readpages(struct file *file, struct address_space *mapping,
|
|
|
|
struct list_head *pages, unsigned nr_pages)
|
|
|
|
{
|
|
|
|
struct inode *inode = mapping->host;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
2007-10-17 08:31:00 +02:00
|
|
|
struct fuse_fill_data data;
|
2005-09-09 22:10:33 +02:00
|
|
|
int err;
|
2006-01-06 09:19:39 +01:00
|
|
|
|
2006-08-14 08:24:27 +02:00
|
|
|
err = -EIO;
|
2006-01-06 09:19:39 +01:00
|
|
|
if (is_bad_inode(inode))
|
2006-11-03 07:07:09 +01:00
|
|
|
goto out;
|
2006-01-06 09:19:39 +01:00
|
|
|
|
2007-11-29 01:22:00 +01:00
|
|
|
data.file = file;
|
2005-09-09 22:10:33 +02:00
|
|
|
data.inode = inode;
|
2006-04-11 07:54:58 +02:00
|
|
|
data.req = fuse_get_req(fc);
|
2006-08-14 08:24:27 +02:00
|
|
|
err = PTR_ERR(data.req);
|
2006-04-11 07:54:58 +02:00
|
|
|
if (IS_ERR(data.req))
|
2006-11-03 07:07:09 +01:00
|
|
|
goto out;
|
2005-09-09 22:10:33 +02:00
|
|
|
|
|
|
|
err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
|
2006-04-11 07:54:49 +02:00
|
|
|
if (!err) {
|
|
|
|
if (data.req->num_pages)
|
2007-11-29 01:22:00 +01:00
|
|
|
fuse_send_readpages(data.req, file, inode);
|
2006-04-11 07:54:49 +02:00
|
|
|
else
|
|
|
|
fuse_put_request(fc, data.req);
|
|
|
|
}
|
2006-11-03 07:07:09 +01:00
|
|
|
out:
|
2006-08-14 08:24:27 +02:00
|
|
|
return err;
|
2005-09-09 22:10:33 +02:00
|
|
|
}
|
|
|
|
|
2007-11-29 01:21:59 +01:00
|
|
|
static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
|
|
|
|
unsigned long nr_segs, loff_t pos)
|
|
|
|
{
|
|
|
|
struct inode *inode = iocb->ki_filp->f_mapping->host;
|
|
|
|
|
|
|
|
if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
|
|
|
|
int err;
|
|
|
|
/*
|
|
|
|
* If trying to read past EOF, make sure the i_size
|
|
|
|
* attribute is up-to-date.
|
|
|
|
*/
|
|
|
|
err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
return generic_file_aio_read(iocb, iov, nr_segs, pos);
|
|
|
|
}
|
|
|
|
|
2007-11-29 01:22:00 +01:00
|
|
|
static void fuse_write_fill(struct fuse_req *req, struct file *file,
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
struct fuse_file *ff, struct inode *inode,
|
|
|
|
loff_t pos, size_t count, int writepage)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
2007-10-18 12:07:04 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
2007-10-18 12:07:03 +02:00
|
|
|
struct fuse_write_in *inarg = &req->misc.write.in;
|
|
|
|
struct fuse_write_out *outarg = &req->misc.write.out;
|
2005-09-09 22:10:30 +02:00
|
|
|
|
2007-10-18 12:07:03 +02:00
|
|
|
memset(inarg, 0, sizeof(struct fuse_write_in));
|
|
|
|
inarg->fh = ff->fh;
|
|
|
|
inarg->offset = pos;
|
|
|
|
inarg->size = count;
|
|
|
|
inarg->write_flags = writepage ? FUSE_WRITE_CACHE : 0;
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
inarg->flags = file ? file->f_flags : 0;
|
2005-09-09 22:10:30 +02:00
|
|
|
req->in.h.opcode = FUSE_WRITE;
|
|
|
|
req->in.h.nodeid = get_node_id(inode);
|
|
|
|
req->in.argpages = 1;
|
|
|
|
req->in.numargs = 2;
|
2007-10-18 12:07:04 +02:00
|
|
|
if (fc->minor < 9)
|
|
|
|
req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
|
|
|
|
else
|
|
|
|
req->in.args[0].size = sizeof(struct fuse_write_in);
|
2007-10-18 12:07:03 +02:00
|
|
|
req->in.args[0].value = inarg;
|
2005-09-09 22:10:30 +02:00
|
|
|
req->in.args[1].size = count;
|
|
|
|
req->out.numargs = 1;
|
|
|
|
req->out.args[0].size = sizeof(struct fuse_write_out);
|
2007-10-18 12:07:03 +02:00
|
|
|
req->out.args[0].value = outarg;
|
|
|
|
}
|
|
|
|
|
|
|
|
static size_t fuse_send_write(struct fuse_req *req, struct file *file,
|
2007-10-18 12:07:04 +02:00
|
|
|
struct inode *inode, loff_t pos, size_t count,
|
|
|
|
fl_owner_t owner)
|
2007-10-18 12:07:03 +02:00
|
|
|
{
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
fuse_write_fill(req, file, file->private_data, inode, pos, count, 0);
|
2007-10-18 12:07:04 +02:00
|
|
|
if (owner != NULL) {
|
|
|
|
struct fuse_write_in *inarg = &req->misc.write.in;
|
|
|
|
inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
|
|
|
|
inarg->lock_owner = fuse_lock_owner_id(fc, owner);
|
|
|
|
}
|
2005-09-09 22:10:39 +02:00
|
|
|
request_send(fc, req);
|
2007-10-18 12:07:03 +02:00
|
|
|
return req->misc.write.out.size;
|
2005-09-09 22:10:30 +02:00
|
|
|
}
|
|
|
|
|
2007-10-16 10:25:17 +02:00
|
|
|
static int fuse_write_begin(struct file *file, struct address_space *mapping,
|
|
|
|
loff_t pos, unsigned len, unsigned flags,
|
|
|
|
struct page **pagep, void **fsdata)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
2007-10-16 10:25:17 +02:00
|
|
|
pgoff_t index = pos >> PAGE_CACHE_SHIFT;
|
|
|
|
|
|
|
|
*pagep = __grab_cache_page(mapping, index);
|
|
|
|
if (!*pagep)
|
|
|
|
return -ENOMEM;
|
2005-09-09 22:10:30 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2008-04-30 09:54:41 +02:00
|
|
|
static void fuse_write_update_size(struct inode *inode, loff_t pos)
|
|
|
|
{
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
|
|
|
|
spin_lock(&fc->lock);
|
|
|
|
fi->attr_version = ++fc->attr_version;
|
|
|
|
if (pos > inode->i_size)
|
|
|
|
i_size_write(inode, pos);
|
|
|
|
spin_unlock(&fc->lock);
|
|
|
|
}
|
|
|
|
|
2007-10-16 10:25:17 +02:00
|
|
|
static int fuse_buffered_write(struct file *file, struct inode *inode,
|
|
|
|
loff_t pos, unsigned count, struct page *page)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
|
|
|
int err;
|
2005-09-09 22:10:36 +02:00
|
|
|
size_t nres;
|
2005-09-09 22:10:30 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
2007-10-16 10:25:17 +02:00
|
|
|
unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
|
2006-01-06 09:19:39 +01:00
|
|
|
struct fuse_req *req;
|
|
|
|
|
|
|
|
if (is_bad_inode(inode))
|
|
|
|
return -EIO;
|
|
|
|
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
/*
|
|
|
|
* Make sure writepages on the same page are not mixed up with
|
|
|
|
* plain writes.
|
|
|
|
*/
|
|
|
|
fuse_wait_on_page_writeback(inode, page->index);
|
|
|
|
|
2006-04-11 07:54:58 +02:00
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req))
|
|
|
|
return PTR_ERR(req);
|
2005-09-09 22:10:30 +02:00
|
|
|
|
|
|
|
req->num_pages = 1;
|
|
|
|
req->pages[0] = page;
|
|
|
|
req->page_offset = offset;
|
2007-10-18 12:07:04 +02:00
|
|
|
nres = fuse_send_write(req, file, inode, pos, count, NULL);
|
2005-09-09 22:10:30 +02:00
|
|
|
err = req->out.h.error;
|
|
|
|
fuse_put_request(fc, req);
|
2007-10-16 10:25:17 +02:00
|
|
|
if (!err && !nres)
|
2005-09-09 22:10:30 +02:00
|
|
|
err = -EIO;
|
|
|
|
if (!err) {
|
2007-10-16 10:25:17 +02:00
|
|
|
pos += nres;
|
2008-04-30 09:54:41 +02:00
|
|
|
fuse_write_update_size(inode, pos);
|
2007-10-16 10:25:17 +02:00
|
|
|
if (count == PAGE_CACHE_SIZE)
|
2005-09-09 22:10:30 +02:00
|
|
|
SetPageUptodate(page);
|
2005-09-09 22:10:38 +02:00
|
|
|
}
|
|
|
|
fuse_invalidate_attr(inode);
|
2007-10-16 10:25:17 +02:00
|
|
|
return err ? err : nres;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_write_end(struct file *file, struct address_space *mapping,
|
|
|
|
loff_t pos, unsigned len, unsigned copied,
|
|
|
|
struct page *page, void *fsdata)
|
|
|
|
{
|
|
|
|
struct inode *inode = mapping->host;
|
|
|
|
int res = 0;
|
|
|
|
|
|
|
|
if (copied)
|
|
|
|
res = fuse_buffered_write(file, inode, pos, copied, page);
|
|
|
|
|
|
|
|
unlock_page(page);
|
|
|
|
page_cache_release(page);
|
|
|
|
return res;
|
2005-09-09 22:10:30 +02:00
|
|
|
}
|
|
|
|
|
2008-04-30 09:54:42 +02:00
|
|
|
static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
|
|
|
|
struct inode *inode, loff_t pos,
|
|
|
|
size_t count)
|
|
|
|
{
|
|
|
|
size_t res;
|
|
|
|
unsigned offset;
|
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
for (i = 0; i < req->num_pages; i++)
|
|
|
|
fuse_wait_on_page_writeback(inode, req->pages[i]->index);
|
|
|
|
|
|
|
|
res = fuse_send_write(req, file, inode, pos, count, NULL);
|
|
|
|
|
|
|
|
offset = req->page_offset;
|
|
|
|
count = res;
|
|
|
|
for (i = 0; i < req->num_pages; i++) {
|
|
|
|
struct page *page = req->pages[i];
|
|
|
|
|
|
|
|
if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
|
|
|
|
SetPageUptodate(page);
|
|
|
|
|
|
|
|
if (count > PAGE_CACHE_SIZE - offset)
|
|
|
|
count -= PAGE_CACHE_SIZE - offset;
|
|
|
|
else
|
|
|
|
count = 0;
|
|
|
|
offset = 0;
|
|
|
|
|
|
|
|
unlock_page(page);
|
|
|
|
page_cache_release(page);
|
|
|
|
}
|
|
|
|
|
|
|
|
return res;
|
|
|
|
}
|
|
|
|
|
|
|
|
static ssize_t fuse_fill_write_pages(struct fuse_req *req,
|
|
|
|
struct address_space *mapping,
|
|
|
|
struct iov_iter *ii, loff_t pos)
|
|
|
|
{
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(mapping->host);
|
|
|
|
unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
|
|
|
|
size_t count = 0;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
req->page_offset = offset;
|
|
|
|
|
|
|
|
do {
|
|
|
|
size_t tmp;
|
|
|
|
struct page *page;
|
|
|
|
pgoff_t index = pos >> PAGE_CACHE_SHIFT;
|
|
|
|
size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
|
|
|
|
iov_iter_count(ii));
|
|
|
|
|
|
|
|
bytes = min_t(size_t, bytes, fc->max_write - count);
|
|
|
|
|
|
|
|
again:
|
|
|
|
err = -EFAULT;
|
|
|
|
if (iov_iter_fault_in_readable(ii, bytes))
|
|
|
|
break;
|
|
|
|
|
|
|
|
err = -ENOMEM;
|
|
|
|
page = __grab_cache_page(mapping, index);
|
|
|
|
if (!page)
|
|
|
|
break;
|
|
|
|
|
|
|
|
pagefault_disable();
|
|
|
|
tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
|
|
|
|
pagefault_enable();
|
|
|
|
flush_dcache_page(page);
|
|
|
|
|
|
|
|
if (!tmp) {
|
|
|
|
unlock_page(page);
|
|
|
|
page_cache_release(page);
|
|
|
|
bytes = min(bytes, iov_iter_single_seg_count(ii));
|
|
|
|
goto again;
|
|
|
|
}
|
|
|
|
|
|
|
|
err = 0;
|
|
|
|
req->pages[req->num_pages] = page;
|
|
|
|
req->num_pages++;
|
|
|
|
|
|
|
|
iov_iter_advance(ii, tmp);
|
|
|
|
count += tmp;
|
|
|
|
pos += tmp;
|
|
|
|
offset += tmp;
|
|
|
|
if (offset == PAGE_CACHE_SIZE)
|
|
|
|
offset = 0;
|
|
|
|
|
2008-05-12 23:02:32 +02:00
|
|
|
if (!fc->big_writes)
|
|
|
|
break;
|
2008-04-30 09:54:42 +02:00
|
|
|
} while (iov_iter_count(ii) && count < fc->max_write &&
|
|
|
|
req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
|
|
|
|
|
|
|
|
return count > 0 ? count : err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static ssize_t fuse_perform_write(struct file *file,
|
|
|
|
struct address_space *mapping,
|
|
|
|
struct iov_iter *ii, loff_t pos)
|
|
|
|
{
|
|
|
|
struct inode *inode = mapping->host;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
int err = 0;
|
|
|
|
ssize_t res = 0;
|
|
|
|
|
|
|
|
if (is_bad_inode(inode))
|
|
|
|
return -EIO;
|
|
|
|
|
|
|
|
do {
|
|
|
|
struct fuse_req *req;
|
|
|
|
ssize_t count;
|
|
|
|
|
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req)) {
|
|
|
|
err = PTR_ERR(req);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
count = fuse_fill_write_pages(req, mapping, ii, pos);
|
|
|
|
if (count <= 0) {
|
|
|
|
err = count;
|
|
|
|
} else {
|
|
|
|
size_t num_written;
|
|
|
|
|
|
|
|
num_written = fuse_send_write_pages(req, file, inode,
|
|
|
|
pos, count);
|
|
|
|
err = req->out.h.error;
|
|
|
|
if (!err) {
|
|
|
|
res += num_written;
|
|
|
|
pos += num_written;
|
|
|
|
|
|
|
|
/* break out of the loop on short write */
|
|
|
|
if (num_written != count)
|
|
|
|
err = -EIO;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
} while (!err && iov_iter_count(ii));
|
|
|
|
|
|
|
|
if (res > 0)
|
|
|
|
fuse_write_update_size(inode, pos);
|
|
|
|
|
|
|
|
fuse_invalidate_attr(inode);
|
|
|
|
|
|
|
|
return res > 0 ? res : err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
|
|
|
|
unsigned long nr_segs, loff_t pos)
|
|
|
|
{
|
|
|
|
struct file *file = iocb->ki_filp;
|
|
|
|
struct address_space *mapping = file->f_mapping;
|
|
|
|
size_t count = 0;
|
|
|
|
ssize_t written = 0;
|
|
|
|
struct inode *inode = mapping->host;
|
|
|
|
ssize_t err;
|
|
|
|
struct iov_iter i;
|
|
|
|
|
|
|
|
WARN_ON(iocb->ki_pos != pos);
|
|
|
|
|
|
|
|
err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
|
|
|
mutex_lock(&inode->i_mutex);
|
|
|
|
vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
|
|
|
|
|
|
|
|
/* We can write back this queue in page reclaim */
|
|
|
|
current->backing_dev_info = mapping->backing_dev_info;
|
|
|
|
|
|
|
|
err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
|
|
|
|
if (err)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
if (count == 0)
|
|
|
|
goto out;
|
|
|
|
|
2008-06-24 16:50:14 +02:00
|
|
|
err = file_remove_suid(file);
|
2008-04-30 09:54:42 +02:00
|
|
|
if (err)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
file_update_time(file);
|
|
|
|
|
|
|
|
iov_iter_init(&i, iov, nr_segs, count, 0);
|
|
|
|
written = fuse_perform_write(file, mapping, &i, pos);
|
|
|
|
if (written >= 0)
|
|
|
|
iocb->ki_pos = pos + written;
|
|
|
|
|
|
|
|
out:
|
|
|
|
current->backing_dev_info = NULL;
|
|
|
|
mutex_unlock(&inode->i_mutex);
|
|
|
|
|
|
|
|
return written ? written : err;
|
|
|
|
}
|
|
|
|
|
2005-09-09 22:10:35 +02:00
|
|
|
static void fuse_release_user_pages(struct fuse_req *req, int write)
|
|
|
|
{
|
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
for (i = 0; i < req->num_pages; i++) {
|
|
|
|
struct page *page = req->pages[i];
|
|
|
|
if (write)
|
|
|
|
set_page_dirty_lock(page);
|
|
|
|
put_page(page);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
|
|
|
|
unsigned nbytes, int write)
|
|
|
|
{
|
|
|
|
unsigned long user_addr = (unsigned long) buf;
|
|
|
|
unsigned offset = user_addr & ~PAGE_MASK;
|
|
|
|
int npages;
|
|
|
|
|
|
|
|
/* This doesn't work with nfsd */
|
|
|
|
if (!current->mm)
|
|
|
|
return -EPERM;
|
|
|
|
|
|
|
|
nbytes = min(nbytes, (unsigned) FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
|
|
|
|
npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
|
2008-05-01 13:35:15 +02:00
|
|
|
npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
|
2005-09-09 22:10:35 +02:00
|
|
|
down_read(¤t->mm->mmap_sem);
|
|
|
|
npages = get_user_pages(current, current->mm, user_addr, npages, write,
|
|
|
|
0, req->pages, NULL);
|
|
|
|
up_read(¤t->mm->mmap_sem);
|
|
|
|
if (npages < 0)
|
|
|
|
return npages;
|
|
|
|
|
|
|
|
req->num_pages = npages;
|
|
|
|
req->page_offset = offset;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
|
|
|
|
size_t count, loff_t *ppos, int write)
|
|
|
|
{
|
2006-12-08 11:37:02 +01:00
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
2005-09-09 22:10:35 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
size_t nmax = write ? fc->max_write : fc->max_read;
|
|
|
|
loff_t pos = *ppos;
|
|
|
|
ssize_t res = 0;
|
2006-01-06 09:19:39 +01:00
|
|
|
struct fuse_req *req;
|
|
|
|
|
|
|
|
if (is_bad_inode(inode))
|
|
|
|
return -EIO;
|
|
|
|
|
2006-04-11 07:54:58 +02:00
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req))
|
|
|
|
return PTR_ERR(req);
|
2005-09-09 22:10:35 +02:00
|
|
|
|
|
|
|
while (count) {
|
|
|
|
size_t nres;
|
2008-04-30 09:54:44 +02:00
|
|
|
size_t nbytes_limit = min(count, nmax);
|
|
|
|
size_t nbytes;
|
|
|
|
int err = fuse_get_user_pages(req, buf, nbytes_limit, !write);
|
2005-09-09 22:10:35 +02:00
|
|
|
if (err) {
|
|
|
|
res = err;
|
|
|
|
break;
|
|
|
|
}
|
2006-01-06 09:19:42 +01:00
|
|
|
nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
|
2008-04-30 09:54:44 +02:00
|
|
|
nbytes = min(nbytes_limit, nbytes);
|
2005-09-09 22:10:35 +02:00
|
|
|
if (write)
|
2007-10-18 12:07:04 +02:00
|
|
|
nres = fuse_send_write(req, file, inode, pos, nbytes,
|
|
|
|
current->files);
|
2005-09-09 22:10:35 +02:00
|
|
|
else
|
2007-10-18 12:07:04 +02:00
|
|
|
nres = fuse_send_read(req, file, inode, pos, nbytes,
|
|
|
|
current->files);
|
2005-09-09 22:10:35 +02:00
|
|
|
fuse_release_user_pages(req, !write);
|
|
|
|
if (req->out.h.error) {
|
|
|
|
if (!res)
|
|
|
|
res = req->out.h.error;
|
|
|
|
break;
|
|
|
|
} else if (nres > nbytes) {
|
|
|
|
res = -EIO;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
count -= nres;
|
|
|
|
res += nres;
|
|
|
|
pos += nres;
|
|
|
|
buf += nres;
|
|
|
|
if (nres != nbytes)
|
|
|
|
break;
|
2006-04-11 21:16:51 +02:00
|
|
|
if (count) {
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req))
|
|
|
|
break;
|
|
|
|
}
|
2005-09-09 22:10:35 +02:00
|
|
|
}
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
if (res > 0) {
|
2008-04-30 09:54:41 +02:00
|
|
|
if (write)
|
|
|
|
fuse_write_update_size(inode, pos);
|
2005-09-09 22:10:35 +02:00
|
|
|
*ppos = pos;
|
2005-09-09 22:10:38 +02:00
|
|
|
}
|
|
|
|
fuse_invalidate_attr(inode);
|
2005-09-09 22:10:35 +02:00
|
|
|
|
|
|
|
return res;
|
|
|
|
}
|
|
|
|
|
|
|
|
static ssize_t fuse_direct_read(struct file *file, char __user *buf,
|
|
|
|
size_t count, loff_t *ppos)
|
|
|
|
{
|
|
|
|
return fuse_direct_io(file, buf, count, ppos, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
|
|
|
|
size_t count, loff_t *ppos)
|
|
|
|
{
|
2006-12-08 11:37:02 +01:00
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
2005-09-09 22:10:35 +02:00
|
|
|
ssize_t res;
|
|
|
|
/* Don't allow parallel writes to the same file */
|
2006-01-10 00:59:24 +01:00
|
|
|
mutex_lock(&inode->i_mutex);
|
2007-05-23 22:57:54 +02:00
|
|
|
res = generic_write_checks(file, ppos, &count, 0);
|
|
|
|
if (!res)
|
|
|
|
res = fuse_direct_io(file, buf, count, ppos, 1);
|
2006-01-10 00:59:24 +01:00
|
|
|
mutex_unlock(&inode->i_mutex);
|
2005-09-09 22:10:35 +02:00
|
|
|
return res;
|
|
|
|
}
|
|
|
|
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
__free_page(req->pages[0]);
|
|
|
|
fuse_file_put(req->ff);
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
|
|
|
|
{
|
|
|
|
struct inode *inode = req->inode;
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
|
|
|
|
|
|
|
|
list_del(&req->writepages_entry);
|
|
|
|
dec_bdi_stat(bdi, BDI_WRITEBACK);
|
|
|
|
dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
|
|
|
|
bdi_writeout_inc(bdi);
|
|
|
|
wake_up(&fi->page_waitq);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Called under fc->lock, may release and reacquire it */
|
|
|
|
static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
|
|
|
|
{
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(req->inode);
|
|
|
|
loff_t size = i_size_read(req->inode);
|
|
|
|
struct fuse_write_in *inarg = &req->misc.write.in;
|
|
|
|
|
|
|
|
if (!fc->connected)
|
|
|
|
goto out_free;
|
|
|
|
|
|
|
|
if (inarg->offset + PAGE_CACHE_SIZE <= size) {
|
|
|
|
inarg->size = PAGE_CACHE_SIZE;
|
|
|
|
} else if (inarg->offset < size) {
|
|
|
|
inarg->size = size & (PAGE_CACHE_SIZE - 1);
|
|
|
|
} else {
|
|
|
|
/* Got truncated off completely */
|
|
|
|
goto out_free;
|
2005-09-09 22:10:30 +02:00
|
|
|
}
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
|
|
|
|
req->in.args[1].size = inarg->size;
|
|
|
|
fi->writectr++;
|
|
|
|
request_send_background_locked(fc, req);
|
|
|
|
return;
|
|
|
|
|
|
|
|
out_free:
|
|
|
|
fuse_writepage_finish(fc, req);
|
|
|
|
spin_unlock(&fc->lock);
|
|
|
|
fuse_writepage_free(fc, req);
|
|
|
|
spin_lock(&fc->lock);
|
2005-09-09 22:10:30 +02:00
|
|
|
}
|
|
|
|
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
/*
|
|
|
|
* If fi->writectr is positive (no truncate or fsync going on) send
|
|
|
|
* all queued writepage requests.
|
|
|
|
*
|
|
|
|
* Called with fc->lock
|
|
|
|
*/
|
|
|
|
void fuse_flush_writepages(struct inode *inode)
|
2005-09-09 22:10:30 +02:00
|
|
|
{
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
struct fuse_req *req;
|
|
|
|
|
|
|
|
while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
|
|
|
|
req = list_entry(fi->queued_writes.next, struct fuse_req, list);
|
|
|
|
list_del_init(&req->list);
|
|
|
|
fuse_send_writepage(fc, req);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
|
|
|
|
{
|
|
|
|
struct inode *inode = req->inode;
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
|
|
|
|
mapping_set_error(inode->i_mapping, req->out.h.error);
|
|
|
|
spin_lock(&fc->lock);
|
|
|
|
fi->writectr--;
|
|
|
|
fuse_writepage_finish(fc, req);
|
|
|
|
spin_unlock(&fc->lock);
|
|
|
|
fuse_writepage_free(fc, req);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_writepage_locked(struct page *page)
|
|
|
|
{
|
|
|
|
struct address_space *mapping = page->mapping;
|
|
|
|
struct inode *inode = mapping->host;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
struct fuse_req *req;
|
|
|
|
struct fuse_file *ff;
|
|
|
|
struct page *tmp_page;
|
|
|
|
|
|
|
|
set_page_writeback(page);
|
|
|
|
|
|
|
|
req = fuse_request_alloc_nofs();
|
|
|
|
if (!req)
|
|
|
|
goto err;
|
|
|
|
|
|
|
|
tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
|
|
|
|
if (!tmp_page)
|
|
|
|
goto err_free;
|
|
|
|
|
|
|
|
spin_lock(&fc->lock);
|
|
|
|
BUG_ON(list_empty(&fi->write_files));
|
|
|
|
ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
|
|
|
|
req->ff = fuse_file_get(ff);
|
|
|
|
spin_unlock(&fc->lock);
|
|
|
|
|
|
|
|
fuse_write_fill(req, NULL, ff, inode, page_offset(page), 0, 1);
|
|
|
|
|
|
|
|
copy_highpage(tmp_page, page);
|
|
|
|
req->num_pages = 1;
|
|
|
|
req->pages[0] = tmp_page;
|
|
|
|
req->page_offset = 0;
|
|
|
|
req->end = fuse_writepage_end;
|
|
|
|
req->inode = inode;
|
|
|
|
|
|
|
|
inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
|
|
|
|
inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
|
|
|
|
end_page_writeback(page);
|
|
|
|
|
|
|
|
spin_lock(&fc->lock);
|
|
|
|
list_add(&req->writepages_entry, &fi->writepages);
|
|
|
|
list_add_tail(&req->list, &fi->queued_writes);
|
|
|
|
fuse_flush_writepages(inode);
|
|
|
|
spin_unlock(&fc->lock);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
err_free:
|
|
|
|
fuse_request_free(req);
|
|
|
|
err:
|
|
|
|
end_page_writeback(page);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_writepage(struct page *page, struct writeback_control *wbc)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
|
|
|
|
err = fuse_writepage_locked(page);
|
|
|
|
unlock_page(page);
|
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_launder_page(struct page *page)
|
|
|
|
{
|
|
|
|
int err = 0;
|
|
|
|
if (clear_page_dirty_for_io(page)) {
|
|
|
|
struct inode *inode = page->mapping->host;
|
|
|
|
err = fuse_writepage_locked(page);
|
|
|
|
if (!err)
|
|
|
|
fuse_wait_on_page_writeback(inode, page->index);
|
|
|
|
}
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Write back dirty pages now, because there may not be any suitable
|
|
|
|
* open files later
|
|
|
|
*/
|
|
|
|
static void fuse_vma_close(struct vm_area_struct *vma)
|
|
|
|
{
|
|
|
|
filemap_write_and_wait(vma->vm_file->f_mapping);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Wait for writeback against this page to complete before allowing it
|
|
|
|
* to be marked dirty again, and hence written back again, possibly
|
|
|
|
* before the previous writepage completed.
|
|
|
|
*
|
|
|
|
* Block here, instead of in ->writepage(), so that the userspace fs
|
|
|
|
* can only block processes actually operating on the filesystem.
|
|
|
|
*
|
|
|
|
* Otherwise unprivileged userspace fs would be able to block
|
|
|
|
* unrelated:
|
|
|
|
*
|
|
|
|
* - page migration
|
|
|
|
* - sync(2)
|
|
|
|
* - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
|
|
|
|
*/
|
|
|
|
static int fuse_page_mkwrite(struct vm_area_struct *vma, struct page *page)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
* Don't use page->mapping as it may become NULL from a
|
|
|
|
* concurrent truncate.
|
|
|
|
*/
|
|
|
|
struct inode *inode = vma->vm_file->f_mapping->host;
|
|
|
|
|
|
|
|
fuse_wait_on_page_writeback(inode, page->index);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct vm_operations_struct fuse_file_vm_ops = {
|
|
|
|
.close = fuse_vma_close,
|
|
|
|
.fault = filemap_fault,
|
|
|
|
.page_mkwrite = fuse_page_mkwrite,
|
|
|
|
};
|
|
|
|
|
|
|
|
static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
|
|
|
|
{
|
|
|
|
if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
|
|
|
|
struct inode *inode = file->f_dentry->d_inode;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_inode *fi = get_fuse_inode(inode);
|
|
|
|
struct fuse_file *ff = file->private_data;
|
|
|
|
/*
|
|
|
|
* file may be written through mmap, so chain it onto the
|
|
|
|
* inodes's write_file list
|
|
|
|
*/
|
|
|
|
spin_lock(&fc->lock);
|
|
|
|
if (list_empty(&ff->write_entry))
|
|
|
|
list_add(&ff->write_entry, &fi->write_files);
|
|
|
|
spin_unlock(&fc->lock);
|
|
|
|
}
|
|
|
|
file_accessed(file);
|
|
|
|
vma->vm_ops = &fuse_file_vm_ops;
|
2005-09-09 22:10:30 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2006-06-25 14:48:52 +02:00
|
|
|
static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
|
|
|
|
struct file_lock *fl)
|
|
|
|
{
|
|
|
|
switch (ffl->type) {
|
|
|
|
case F_UNLCK:
|
|
|
|
break;
|
|
|
|
|
|
|
|
case F_RDLCK:
|
|
|
|
case F_WRLCK:
|
|
|
|
if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
|
|
|
|
ffl->end < ffl->start)
|
|
|
|
return -EIO;
|
|
|
|
|
|
|
|
fl->fl_start = ffl->start;
|
|
|
|
fl->fl_end = ffl->end;
|
|
|
|
fl->fl_pid = ffl->pid;
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
return -EIO;
|
|
|
|
}
|
|
|
|
fl->fl_type = ffl->type;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void fuse_lk_fill(struct fuse_req *req, struct file *file,
|
2007-10-18 12:07:02 +02:00
|
|
|
const struct file_lock *fl, int opcode, pid_t pid,
|
|
|
|
int flock)
|
2006-06-25 14:48:52 +02:00
|
|
|
{
|
2006-12-08 11:37:02 +01:00
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
2006-06-25 14:48:55 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
2006-06-25 14:48:52 +02:00
|
|
|
struct fuse_file *ff = file->private_data;
|
|
|
|
struct fuse_lk_in *arg = &req->misc.lk_in;
|
|
|
|
|
|
|
|
arg->fh = ff->fh;
|
2006-06-25 14:48:55 +02:00
|
|
|
arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
|
2006-06-25 14:48:52 +02:00
|
|
|
arg->lk.start = fl->fl_start;
|
|
|
|
arg->lk.end = fl->fl_end;
|
|
|
|
arg->lk.type = fl->fl_type;
|
|
|
|
arg->lk.pid = pid;
|
2007-10-18 12:07:02 +02:00
|
|
|
if (flock)
|
|
|
|
arg->lk_flags |= FUSE_LK_FLOCK;
|
2006-06-25 14:48:52 +02:00
|
|
|
req->in.h.opcode = opcode;
|
|
|
|
req->in.h.nodeid = get_node_id(inode);
|
|
|
|
req->in.numargs = 1;
|
|
|
|
req->in.args[0].size = sizeof(*arg);
|
|
|
|
req->in.args[0].value = arg;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_getlk(struct file *file, struct file_lock *fl)
|
|
|
|
{
|
2006-12-08 11:37:02 +01:00
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
2006-06-25 14:48:52 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_req *req;
|
|
|
|
struct fuse_lk_out outarg;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req))
|
|
|
|
return PTR_ERR(req);
|
|
|
|
|
2007-10-18 12:07:02 +02:00
|
|
|
fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
|
2006-06-25 14:48:52 +02:00
|
|
|
req->out.numargs = 1;
|
|
|
|
req->out.args[0].size = sizeof(outarg);
|
|
|
|
req->out.args[0].value = &outarg;
|
|
|
|
request_send(fc, req);
|
|
|
|
err = req->out.h.error;
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
if (!err)
|
|
|
|
err = convert_fuse_file_lock(&outarg.lk, fl);
|
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2007-10-18 12:07:02 +02:00
|
|
|
static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
|
2006-06-25 14:48:52 +02:00
|
|
|
{
|
2006-12-08 11:37:02 +01:00
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
2006-06-25 14:48:52 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_req *req;
|
|
|
|
int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
|
|
|
|
pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
|
|
|
|
int err;
|
|
|
|
|
2008-07-25 10:49:02 +02:00
|
|
|
if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
|
|
|
|
/* NLM needs asynchronous locks, which we don't support yet */
|
|
|
|
return -ENOLCK;
|
|
|
|
}
|
|
|
|
|
2006-06-25 14:48:52 +02:00
|
|
|
/* Unlock on close is handled by the flush method */
|
|
|
|
if (fl->fl_flags & FL_CLOSE)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req))
|
|
|
|
return PTR_ERR(req);
|
|
|
|
|
2007-10-18 12:07:02 +02:00
|
|
|
fuse_lk_fill(req, file, fl, opcode, pid, flock);
|
2006-06-25 14:48:52 +02:00
|
|
|
request_send(fc, req);
|
|
|
|
err = req->out.h.error;
|
2006-06-25 14:48:54 +02:00
|
|
|
/* locking is restartable */
|
|
|
|
if (err == -EINTR)
|
|
|
|
err = -ERESTARTSYS;
|
2006-06-25 14:48:52 +02:00
|
|
|
fuse_put_request(fc, req);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
|
|
|
|
{
|
2006-12-08 11:37:02 +01:00
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
2006-06-25 14:48:52 +02:00
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
int err;
|
|
|
|
|
2008-07-25 10:49:02 +02:00
|
|
|
if (cmd == F_CANCELLK) {
|
|
|
|
err = 0;
|
|
|
|
} else if (cmd == F_GETLK) {
|
2006-06-25 14:48:52 +02:00
|
|
|
if (fc->no_lock) {
|
2007-02-21 06:55:18 +01:00
|
|
|
posix_test_lock(file, fl);
|
2006-06-25 14:48:52 +02:00
|
|
|
err = 0;
|
|
|
|
} else
|
|
|
|
err = fuse_getlk(file, fl);
|
|
|
|
} else {
|
|
|
|
if (fc->no_lock)
|
2008-07-25 10:49:02 +02:00
|
|
|
err = posix_lock_file(file, fl, NULL);
|
2006-06-25 14:48:52 +02:00
|
|
|
else
|
2007-10-18 12:07:02 +02:00
|
|
|
err = fuse_setlk(file, fl, 0);
|
2006-06-25 14:48:52 +02:00
|
|
|
}
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2007-10-18 12:07:02 +02:00
|
|
|
static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
|
|
|
|
{
|
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
int err;
|
|
|
|
|
|
|
|
if (fc->no_lock) {
|
|
|
|
err = flock_lock_file_wait(file, fl);
|
|
|
|
} else {
|
|
|
|
/* emulate flock with POSIX locks */
|
|
|
|
fl->fl_owner = (fl_owner_t) file;
|
|
|
|
err = fuse_setlk(file, fl, 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2006-12-07 05:35:51 +01:00
|
|
|
static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
|
|
|
|
{
|
|
|
|
struct inode *inode = mapping->host;
|
|
|
|
struct fuse_conn *fc = get_fuse_conn(inode);
|
|
|
|
struct fuse_req *req;
|
|
|
|
struct fuse_bmap_in inarg;
|
|
|
|
struct fuse_bmap_out outarg;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
if (!inode->i_sb->s_bdev || fc->no_bmap)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
req = fuse_get_req(fc);
|
|
|
|
if (IS_ERR(req))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
memset(&inarg, 0, sizeof(inarg));
|
|
|
|
inarg.block = block;
|
|
|
|
inarg.blocksize = inode->i_sb->s_blocksize;
|
|
|
|
req->in.h.opcode = FUSE_BMAP;
|
|
|
|
req->in.h.nodeid = get_node_id(inode);
|
|
|
|
req->in.numargs = 1;
|
|
|
|
req->in.args[0].size = sizeof(inarg);
|
|
|
|
req->in.args[0].value = &inarg;
|
|
|
|
req->out.numargs = 1;
|
|
|
|
req->out.args[0].size = sizeof(outarg);
|
|
|
|
req->out.args[0].value = &outarg;
|
|
|
|
request_send(fc, req);
|
|
|
|
err = req->out.h.error;
|
|
|
|
fuse_put_request(fc, req);
|
|
|
|
if (err == -ENOSYS)
|
|
|
|
fc->no_bmap = 1;
|
|
|
|
|
|
|
|
return err ? 0 : outarg.block;
|
|
|
|
}
|
|
|
|
|
2008-04-30 09:54:45 +02:00
|
|
|
static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
|
|
|
|
{
|
|
|
|
loff_t retval;
|
|
|
|
struct inode *inode = file->f_path.dentry->d_inode;
|
|
|
|
|
|
|
|
mutex_lock(&inode->i_mutex);
|
|
|
|
switch (origin) {
|
|
|
|
case SEEK_END:
|
|
|
|
offset += i_size_read(inode);
|
|
|
|
break;
|
|
|
|
case SEEK_CUR:
|
|
|
|
offset += file->f_pos;
|
|
|
|
}
|
|
|
|
retval = -EINVAL;
|
|
|
|
if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
|
|
|
|
if (offset != file->f_pos) {
|
|
|
|
file->f_pos = offset;
|
|
|
|
file->f_version = 0;
|
|
|
|
}
|
|
|
|
retval = offset;
|
|
|
|
}
|
|
|
|
mutex_unlock(&inode->i_mutex);
|
|
|
|
return retval;
|
|
|
|
}
|
|
|
|
|
2006-03-28 11:56:42 +02:00
|
|
|
static const struct file_operations fuse_file_operations = {
|
2008-04-30 09:54:45 +02:00
|
|
|
.llseek = fuse_file_llseek,
|
2006-10-01 08:28:48 +02:00
|
|
|
.read = do_sync_read,
|
2007-11-29 01:21:59 +01:00
|
|
|
.aio_read = fuse_file_aio_read,
|
2006-10-01 08:28:48 +02:00
|
|
|
.write = do_sync_write,
|
2008-04-30 09:54:42 +02:00
|
|
|
.aio_write = fuse_file_aio_write,
|
2005-09-09 22:10:30 +02:00
|
|
|
.mmap = fuse_file_mmap,
|
|
|
|
.open = fuse_open,
|
|
|
|
.flush = fuse_flush,
|
|
|
|
.release = fuse_release,
|
|
|
|
.fsync = fuse_fsync,
|
2006-06-25 14:48:52 +02:00
|
|
|
.lock = fuse_file_lock,
|
2007-10-18 12:07:02 +02:00
|
|
|
.flock = fuse_file_flock,
|
2007-06-01 11:49:19 +02:00
|
|
|
.splice_read = generic_file_splice_read,
|
2005-09-09 22:10:30 +02:00
|
|
|
};
|
|
|
|
|
2006-03-28 11:56:42 +02:00
|
|
|
static const struct file_operations fuse_direct_io_file_operations = {
|
2008-04-30 09:54:45 +02:00
|
|
|
.llseek = fuse_file_llseek,
|
2005-09-09 22:10:35 +02:00
|
|
|
.read = fuse_direct_read,
|
|
|
|
.write = fuse_direct_write,
|
|
|
|
.open = fuse_open,
|
|
|
|
.flush = fuse_flush,
|
|
|
|
.release = fuse_release,
|
|
|
|
.fsync = fuse_fsync,
|
2006-06-25 14:48:52 +02:00
|
|
|
.lock = fuse_file_lock,
|
2007-10-18 12:07:02 +02:00
|
|
|
.flock = fuse_file_flock,
|
2007-06-01 11:49:19 +02:00
|
|
|
/* no mmap and splice_read */
|
2005-09-09 22:10:35 +02:00
|
|
|
};
|
|
|
|
|
2006-06-28 13:26:44 +02:00
|
|
|
static const struct address_space_operations fuse_file_aops = {
|
2005-09-09 22:10:30 +02:00
|
|
|
.readpage = fuse_readpage,
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
.writepage = fuse_writepage,
|
|
|
|
.launder_page = fuse_launder_page,
|
2007-10-16 10:25:17 +02:00
|
|
|
.write_begin = fuse_write_begin,
|
|
|
|
.write_end = fuse_write_end,
|
2005-09-09 22:10:33 +02:00
|
|
|
.readpages = fuse_readpages,
|
fuse: support writable mmap
Quoting Linus (3 years ago, FUSE inclusion discussions):
"User-space filesystems are hard to get right. I'd claim that they
are almost impossible, unless you limit them somehow (shared
writable mappings are the nastiest part - if you don't have those,
you can reasonably limit your problems by limiting the number of
dirty pages you accept through normal "write()" calls)."
Instead of attempting the impossible, I've just waited for the dirty page
accounting infrastructure to materialize (thanks to Peter Zijlstra and
others). This nicely solved the biggest problem: limiting the number of pages
used for write caching.
Some small details remained, however, which this largish patch attempts to
address. It provides a page writeback implementation for fuse, which is
completely safe against VM related deadlocks. Performance may not be very
good for certain usage patterns, but generally it should be acceptable.
It has been tested extensively with fsx-linux and bash-shared-mapping.
Fuse page writeback design
--------------------------
fuse_writepage() allocates a new temporary page with GFP_NOFS|__GFP_HIGHMEM.
It copies the contents of the original page, and queues a WRITE request to the
userspace filesystem using this temp page.
The writeback is finished instantly from the MM's point of view: the page is
removed from the radix trees, and the PageDirty and PageWriteback flags are
cleared.
For the duration of the actual write, the NR_WRITEBACK_TEMP counter is
incremented. The per-bdi writeback count is not decremented until the actual
write completes.
On dirtying the page, fuse waits for a previous write to finish before
proceeding. This makes sure, there can only be one temporary page used at a
time for one cached page.
This approach is wasteful in both memory and CPU bandwidth, so why is this
complication needed?
The basic problem is that there can be no guarantee about the time in which
the userspace filesystem will complete a write. It may be buggy or even
malicious, and fail to complete WRITE requests. We don't want unrelated parts
of the system to grind to a halt in such cases.
Also a filesystem may need additional resources (particularly memory) to
complete a WRITE request. There's a great danger of a deadlock if that
allocation may wait for the writepage to finish.
Currently there are several cases where the kernel can block on page
writeback:
- allocation order is larger than PAGE_ALLOC_COSTLY_ORDER
- page migration
- throttle_vm_writeout (through NR_WRITEBACK)
- sync(2)
Of course in some cases (fsync, msync) we explicitly want to allow blocking.
So for these cases new code has to be added to fuse, since the VM is not
tracking writeback pages for us any more.
As an extra safetly measure, the maximum dirty ratio allocated to a single
fuse filesystem is set to 1% by default. This way one (or several) buggy or
malicious fuse filesystems cannot slow down the rest of the system by hogging
dirty memory.
With appropriate privileges, this limit can be raised through
'/sys/class/bdi/<bdi>/max_ratio'.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 09:54:41 +02:00
|
|
|
.set_page_dirty = __set_page_dirty_nobuffers,
|
2006-12-07 05:35:51 +01:00
|
|
|
.bmap = fuse_bmap,
|
2005-09-09 22:10:30 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
void fuse_init_file_inode(struct inode *inode)
|
|
|
|
{
|
2005-09-09 22:10:37 +02:00
|
|
|
inode->i_fop = &fuse_file_operations;
|
|
|
|
inode->i_data.a_ops = &fuse_file_aops;
|
2005-09-09 22:10:30 +02:00
|
|
|
}
|