Re: [PATCH 6/6] ceph: perform asynchronous unlink if we have sufficient caps

[Jeff Layton <jlayton@kernel.org>]

On Thu, 2020-01-09 at 17:18 +0800, Yan, Zheng wrote:
> On 1/6/20 11:35 PM, Jeff Layton wrote:
> > From: "Yan, Zheng" <zyan@redhat.com>
> > 
> > The MDS is getting a new lock-caching facility that will allow it
> > to cache the necessary locks to allow asynchronous directory operations.
> > Since the CEPH_CAP_FILE_* caps are currently unused on directories,
> > we can repurpose those bits for this purpose.
> > 
> > When performing an unlink, if we have Fx on the parent directory,
> > and CEPH_CAP_DIR_UNLINK (aka Fr), and we know that the dentry being
> > removed is the primary link, then then we can fire off an unlink
> > request immediately and don't need to wait on reply before returning.
> > 
> > In that situation, just fix up the dcache and link count and return
> > immediately after issuing the call to the MDS. This does mean that we
> > need to hold an extra reference to the inode being unlinked, and extra
> > references to the caps to avoid races. Those references are put and
> > error handling is done in the r_callback routine.
> > 
> > If the operation ends up failing, then set a writeback error on the
> > directory inode that can be fetched later by an fsync on the dir.
> > 
> > Since this feature is very new, also add a new module parameter to
> > enable and disable it (default is disabled).
> > 
> > Signed-off-by: "Yan, Zheng" <zyan@redhat.com>
> > Signed-off-by: Jeff Layton <jlayton@kernel.org>
> > ---
> >   fs/ceph/caps.c               | 35 ++++++++++++------
> >   fs/ceph/dir.c                | 70 +++++++++++++++++++++++++++++++++---
> >   fs/ceph/inode.c              |  8 ++++-
> >   fs/ceph/super.c              |  4 +++
> >   fs/ceph/super.h              |  3 ++
> >   include/linux/ceph/ceph_fs.h |  9 +++++
> >   6 files changed, 113 insertions(+), 16 deletions(-)
> > 
> > diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
> > index d05717397c2a..7fc87b693ba4 100644
> > --- a/fs/ceph/caps.c
> > +++ b/fs/ceph/caps.c
> > @@ -992,7 +992,11 @@ int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
> >   int __ceph_caps_wanted(struct ceph_inode_info *ci)
> >   {
> >   	int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
> > -	if (!S_ISDIR(ci->vfs_inode.i_mode)) {
> > +	if (S_ISDIR(ci->vfs_inode.i_mode)) {
> > +		/* we want EXCL if holding caps of dir ops */
> > +		if (w & CEPH_CAP_ANY_DIR_OPS)
> > +			w |= CEPH_CAP_FILE_EXCL;
> > +	} else {
> >   		/* we want EXCL if dirty data */
> >   		if (w & CEPH_CAP_FILE_BUFFER)
> >   			w |= CEPH_CAP_FILE_EXCL;
> > @@ -1883,10 +1887,13 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
> >   			 * revoking the shared cap on every create/unlink
> >   			 * operation.
> >   			 */
> > -			if (IS_RDONLY(inode))
> > +			if (IS_RDONLY(inode)) {
> >   				want = CEPH_CAP_ANY_SHARED;
> > -			else
> > -				want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
> > +			} else {
> > +				want = CEPH_CAP_ANY_SHARED |
> > +				       CEPH_CAP_FILE_EXCL |
> > +				       CEPH_CAP_ANY_DIR_OPS;
> > +			}
> >   			retain |= want;
> >   		} else {
> >   
> > @@ -2649,7 +2656,10 @@ static int try_get_cap_refs(struct inode *inode, int need, int want,
> >   				}
> >   				snap_rwsem_locked = true;
> >   			}
> > -			*got = need | (have & want);
> > +			if ((have & want) == want)
> > +				*got = need | want;
> > +			else
> > +				*got = need;
> >   			if (S_ISREG(inode->i_mode) &&
> >   			    (need & CEPH_CAP_FILE_RD) &&
> >   			    !(*got & CEPH_CAP_FILE_CACHE))
> > @@ -2739,13 +2749,16 @@ int ceph_try_get_caps(struct inode *inode, int need, int want,
> >   	int ret;
> >   
> >   	BUG_ON(need & ~CEPH_CAP_FILE_RD);
> > -	BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO|CEPH_CAP_FILE_SHARED));
> > -	ret = ceph_pool_perm_check(inode, need);
> > -	if (ret < 0)
> > -		return ret;
> > +	if (need) {
> > +		ret = ceph_pool_perm_check(inode, need);
> > +		if (ret < 0)
> > +			return ret;
> > +	}
> >   
> > -	ret = try_get_cap_refs(inode, need, want, 0,
> > -			       (nonblock ? NON_BLOCKING : 0), got);
> > +	BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
> > +			CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
> > +			CEPH_CAP_ANY_DIR_OPS));
> > +	ret = try_get_cap_refs(inode, need, want, 0, nonblock, got);
> >   	return ret == -EAGAIN ? 0 : ret;
> >   }
> >   
> > diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> > index d0cd0aba5843..10294f07f5f0 100644
> > --- a/fs/ceph/dir.c
> > +++ b/fs/ceph/dir.c
> > @@ -1036,6 +1036,47 @@ static int ceph_link(struct dentry *old_dentry, struct inode *dir,
> >   	return err;
> >   }
> >   
> > +static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
> > +				 struct ceph_mds_request *req)
> > +{
> > +	/* If op failed, set error on parent directory */
> > +	mapping_set_error(req->r_parent->i_mapping, req->r_err);
> > +	if (req->r_err)
> > +		printk("%s: req->r_err = %d\n", __func__, req->r_err);
> > +	ceph_put_cap_refs(ceph_inode(req->r_parent),
> > +			  CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK);
> > +	iput(req->r_old_inode);
> > +}
> > +
> > +static bool get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
> > +{
> > +	struct ceph_inode_info *ci = ceph_inode(dir);
> > +	struct ceph_dentry_info *di;
> > +	int ret, want, got;
> > +
> > +	want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
> > +	ret = ceph_try_get_caps(dir, 0, want, true, &got);
> > +	dout("Fx on %p ret=%d got=%d\n", dir, ret, got);
> > +	if (ret != 1 || got != want)
> > +		return false;
> > +
> > +        spin_lock(&dentry->d_lock);
> > +        di = ceph_dentry(dentry);
> > +	/* - We are holding CEPH_CAP_FILE_EXCL, which implies
> > +	 * CEPH_CAP_FILE_SHARED.
> > +	 * - Only support async unlink for primary linkage */
> > +	if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
> > +	    !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
> > +		ret = 0;
> > +        spin_unlock(&dentry->d_lock);
> > +
> > +	if (!ret) {
> > +		ceph_put_cap_refs(ci, got);
> > +		return false;
> > +	}
> > +	return true;
> > +}
> > +
> >   /*
> >    * rmdir and unlink are differ only by the metadata op code
> >    */
> > @@ -1067,13 +1108,33 @@ static int ceph_unlink(struct inode *dir, struct dentry *dentry)
> >   	req->r_dentry = dget(dentry);
> >   	req->r_num_caps = 2;
> >   	req->r_parent = dir;
> > -	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
> >   	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
> >   	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
> >   	req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
> > -	err = ceph_mdsc_do_request(mdsc, dir, req);
> > -	if (!err && !req->r_reply_info.head->is_dentry)
> > -		d_delete(dentry);
> > +
> > +	if (enable_async_dirops && op == CEPH_MDS_OP_UNLINK &&
> > +	    get_caps_for_async_unlink(dir, dentry)) {
> > +		dout("ceph: Async unlink on %lu/%.*s", dir->i_ino,
> > +		     dentry->d_name.len, dentry->d_name.name);
> > +		req->r_callback = ceph_async_unlink_cb;
> > +		req->r_old_inode = d_inode(dentry);
> > +		ihold(req->r_old_inode);
> > +		err = ceph_mdsc_submit_request(mdsc, dir, req);
> > +		if (!err) {
> > +			/*
> > +			 * We have enough caps, so we assume that the unlink
> > +			 * will succeed. Fix up the target inode and dcache.
> > +			 */
> > +			drop_nlink(inode);
> > +			d_delete(dentry);
> > +		}
> > +	} else {
> > +		set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
> > +		err = ceph_mdsc_do_request(mdsc, dir, req);
> > +		if (!err && !req->r_reply_info.head->is_dentry)
> > +			d_delete(dentry);
> > +	}
> > +
> >   	ceph_mdsc_put_request(req);
> >   out:
> >   	return err;
> > @@ -1411,6 +1472,7 @@ void ceph_invalidate_dentry_lease(struct dentry *dentry)
> >   	spin_lock(&dentry->d_lock);
> >   	di->time = jiffies;
> >   	di->lease_shared_gen = 0;
> > +	di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
> >   	__dentry_lease_unlist(di);
> >   	spin_unlock(&dentry->d_lock);
> >   }
> > diff --git a/fs/ceph/inode.c b/fs/ceph/inode.c
> > index ef9e8281d485..ffef475af72b 100644
> > --- a/fs/ceph/inode.c
> > +++ b/fs/ceph/inode.c
> > @@ -1048,6 +1048,7 @@ static void __update_dentry_lease(struct inode *dir, struct dentry *dentry,
> >   				  struct ceph_mds_session **old_lease_session)
> >   {
> >   	struct ceph_dentry_info *di = ceph_dentry(dentry);
> > +	unsigned mask = le16_to_cpu(lease->mask);
> >   	long unsigned duration = le32_to_cpu(lease->duration_ms);
> >   	long unsigned ttl = from_time + (duration * HZ) / 1000;
> >   	long unsigned half_ttl = from_time + (duration * HZ / 2) / 1000;
> > @@ -1059,8 +1060,13 @@ static void __update_dentry_lease(struct inode *dir, struct dentry *dentry,
> >   	if (ceph_snap(dir) != CEPH_NOSNAP)
> >   		return;
> >   
> > +	if (mask & CEPH_LEASE_PRIMARY_LINK)
> > +		di->flags |= CEPH_DENTRY_PRIMARY_LINK;
> > +	else
> > +		di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
> > +
> >   	di->lease_shared_gen = atomic_read(&ceph_inode(dir)->i_shared_gen);
> > -	if (duration == 0) {
> > +	if (!(mask & CEPH_LEASE_VALID)) {
> >   		__ceph_dentry_dir_lease_touch(di);
> >   		return;
> >   	}
> > diff --git a/fs/ceph/super.c b/fs/ceph/super.c
> > index 112927dbd2f2..c149edb6aa2d 100644
> > --- a/fs/ceph/super.c
> > +++ b/fs/ceph/super.c
> > @@ -1324,6 +1324,10 @@ static void __exit exit_ceph(void)
> >   	destroy_caches();
> >   }
> >   
> > +bool enable_async_dirops;
> > +module_param(enable_async_dirops, bool, 0644);
> > +MODULE_PARM_DESC(enable_async_dirops, "Asynchronous directory operations enabled");
> > +
> 
> why not use mount option
>

I'm open to suggestions here.

I mostly put this in originally to help facilitate performance testing.
A module option makes it easy to change this at runtime (without having
to remount or anything).

That said, we probably _do_ want to have a way for users to enable or
disable this feature. We'll probably want this disabled by default
initially, but I can forsee that changing once get more confidence.

Mount options are a bit of a pain to deal with over time. If the
defaults change, we need to document that in the manpages and online
documentation. If you put a mount option in the fstab, then you have to
deal with breakage if you boot to an earlier kernel that doesn't support
that option.

My thinking is that we should just use a module option initially (for
the really early adopters) and only convert that to a mount option as
the need for it becomes clear.


> >   module_init(init_ceph);
> >   module_exit(exit_ceph);
> >   
> > diff --git a/fs/ceph/super.h b/fs/ceph/super.h
> > index 4b53f32b9324..4083de451710 100644
> > --- a/fs/ceph/super.h
> > +++ b/fs/ceph/super.h
> > @@ -72,6 +72,8 @@
> >   #define CEPH_CAPS_WANTED_DELAY_MIN_DEFAULT      5  /* cap release delay */
> >   #define CEPH_CAPS_WANTED_DELAY_MAX_DEFAULT     60  /* cap release delay */
> >   
> > +extern bool enable_async_dirops;
> > +
> >   struct ceph_mount_options {
> >   	unsigned int flags;
> >   
> > @@ -282,6 +284,7 @@ struct ceph_dentry_info {
> >   #define CEPH_DENTRY_REFERENCED		1
> >   #define CEPH_DENTRY_LEASE_LIST		2
> >   #define CEPH_DENTRY_SHRINK_LIST		4
> > +#define CEPH_DENTRY_PRIMARY_LINK	8
> >   
> >   struct ceph_inode_xattrs_info {
> >   	/*
> > diff --git a/include/linux/ceph/ceph_fs.h b/include/linux/ceph/ceph_fs.h
> > index cb21c5cf12c3..a099f60feb7b 100644
> > --- a/include/linux/ceph/ceph_fs.h
> > +++ b/include/linux/ceph/ceph_fs.h
> > @@ -530,6 +530,9 @@ struct ceph_mds_reply_lease {
> >   	__le32 seq;
> >   } __attribute__ ((packed));
> >   
> > +#define CEPH_LEASE_VALID        (1 | 2) /* old and new bit values */
> > +#define CEPH_LEASE_PRIMARY_LINK 4       /* primary linkage */
> > +
> >   struct ceph_mds_reply_dirfrag {
> >   	__le32 frag;            /* fragment */
> >   	__le32 auth;            /* auth mds, if this is a delegation point */
> > @@ -659,6 +662,12 @@ int ceph_flags_to_mode(int flags);
> >   #define CEPH_CAP_LOCKS (CEPH_LOCK_IFILE | CEPH_LOCK_IAUTH | CEPH_LOCK_ILINK | \
> >   			CEPH_LOCK_IXATTR)
> >   
> > +/* cap masks async dir operations */
> > +#define CEPH_CAP_DIR_CREATE	CEPH_CAP_FILE_CACHE
> > +#define CEPH_CAP_DIR_UNLINK	CEPH_CAP_FILE_RD
> > +#define CEPH_CAP_ANY_DIR_OPS	(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_RD | \
> > +				 CEPH_CAP_FILE_WREXTEND | CEPH_CAP_FILE_LAZYIO)
> > +
> >   int ceph_caps_for_mode(int mode);
> >   
> >   enum {
> > 

-- 
Jeff Layton <jlayton@kernel.org>