This time around we have:

- support for rbd data-pool feature, which enables rbd images on
   erasure-coded pools (myself).  CEPH_PG_MAX_SIZE has been bumped to
   allow erasure-coded profiles with k+m up to 32.
 
 - a patch for ceph_d_revalidate() performance regression introduced in
   4.9, along with some cleanups in the area (Jeff Layton)
 
 - a set of fixes for unsafe ->d_parent accesses in CephFS (Jeff Layton)
 
 - buffered reads are now processed in rsize windows instead of rasize
   windows (Andreas Gerstmayr).  The new default for rsize mount option
   is 64M.
 
 - ack vs commit distinction is gone, greatly simplifying ->fsync() and
   MOSDOpReply handling code (myself)
 
 Also a few filesystem bug fixes from Zheng, a CRUSH sync up (CRUSH
 computations are still serialized though) and several minor fixes and
 cleanups all over.
 -----BEGIN PGP SIGNATURE-----
 Version: GnuPG v2
 
 iQEcBAABCAAGBQJYtY0rAAoJEEp/3jgCEfOLQioH/36QKsalquY1FCdJnJve9qj0
 q19OohamIedhv76AYvXhJzBBHlVwerjicE51/bSzuUhxV+ApdATrPPcLC22oLd3i
 h0R9NAUMYjiris1yN/Z9JRiPCSdsxvHuRycsUMRSRbxZhnyP9XdTxFD1A+fLfisU
 Z4osyTzadabVL5Um9maRBbAtXCWh3d9JZzPa5xIvWTEO4CWWk87GtEIIQDcgx+Y6
 8ZSMmrVFDNtskUp9js+LnFYW7/xBsEXyqgsqKaecf5uQqwu1WKRXSKtv9PUmGAIb
 HBrlUdV1PQaCzTYtaoztJshNdYcphM5L7gePzxRG0nXrTNsq8J5eCzI8en5qS8w=
 =CPL/
 -----END PGP SIGNATURE-----

Merge tag 'ceph-for-4.11-rc1' of git://github.com/ceph/ceph-client

Pull ceph updates from Ilya Dryomov:
 "This time around we have:

   - support for rbd data-pool feature, which enables rbd images on
     erasure-coded pools (myself). CEPH_PG_MAX_SIZE has been bumped to
     allow erasure-coded profiles with k+m up to 32.

   - a patch for ceph_d_revalidate() performance regression introduced
     in 4.9, along with some cleanups in the area (Jeff Layton)

   - a set of fixes for unsafe ->d_parent accesses in CephFS (Jeff
     Layton)

   - buffered reads are now processed in rsize windows instead of rasize
     windows (Andreas Gerstmayr). The new default for rsize mount option
     is 64M.

   - ack vs commit distinction is gone, greatly simplifying ->fsync()
     and MOSDOpReply handling code (myself)

  ... also a few filesystem bug fixes from Zheng, a CRUSH sync up (CRUSH
  computations are still serialized though) and several minor fixes and
  cleanups all over"

* tag 'ceph-for-4.11-rc1' of git://github.com/ceph/ceph-client: (52 commits)
  libceph, rbd, ceph: WRITE | ONDISK -> WRITE
  libceph: get rid of ack vs commit
  ceph: remove special ack vs commit behavior
  ceph: tidy some white space in get_nonsnap_parent()
  crush: fix dprintk compilation
  crush: do is_out test only if we do not collide
  ceph: remove req from unsafe list when unregistering it
  rbd: constify device_type structure
  rbd: kill obj_request->object_name and rbd_segment_name_cache
  rbd: store and use obj_request->object_no
  rbd: RBD_V{1,2}_DATA_FORMAT macros
  rbd: factor out __rbd_osd_req_create()
  rbd: set offset and length outside of rbd_obj_request_create()
  rbd: support for data-pool feature
  rbd: introduce rbd_init_layout()
  rbd: use rbd_obj_bytes() more
  rbd: remove now unused rbd_obj_request_wait() and helpers
  rbd: switch rbd_obj_method_sync() to ceph_osdc_call()
  libceph: pass reply buffer length through ceph_osdc_call()
  rbd: do away with obj_request in rbd_obj_read_sync()
  ...

include/linux/ceph/osd_client.h

@@ -22,7 +22,6 @@ struct ceph_osd_client;
  * completion callback for async writepages
  */
 typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *);
-typedef void (*ceph_osdc_unsafe_callback_t)(struct ceph_osd_request *, bool);
 
 #define CEPH_HOMELESS_OSD	-1
 
@@ -170,15 +169,12 @@ struct ceph_osd_request {
 	unsigned int		r_num_ops;
 
 	int               r_result;
-	bool              r_got_reply;
 
 	struct ceph_osd_client *r_osdc;
 	struct kref       r_kref;
 	bool              r_mempool;
-	struct completion r_completion;
-	struct completion r_done_completion;  /* fsync waiter */
+	struct completion r_completion;       /* private to osd_client.c */
 	ceph_osdc_callback_t r_callback;
-	ceph_osdc_unsafe_callback_t r_unsafe_callback;
 	struct list_head  r_unsafe_item;
 
 	struct inode *r_inode;         	      /* for use by callbacks */

include/linux/ceph/osdmap.h

@@ -57,7 +57,7 @@ static inline bool ceph_can_shift_osds(struct ceph_pg_pool_info *pool)
 	case CEPH_POOL_TYPE_EC:
 		return false;
 	default:
-		BUG_ON(1);
+		BUG();
 	}
 }
 
@@ -81,13 +81,6 @@ void ceph_oloc_copy(struct ceph_object_locator *dest,
 		    const struct ceph_object_locator *src);
 void ceph_oloc_destroy(struct ceph_object_locator *oloc);
 
-/*
- * Maximum supported by kernel client object name length
- *
- * (probably outdated: must be >= RBD_MAX_MD_NAME_LEN -- currently 100)
- */
-#define CEPH_MAX_OID_NAME_LEN 100
-
 /*
  * 51-char inline_name is long enough for all cephfs and all but one
  * rbd requests: <imgname> in "<imgname>.rbd"/"rbd_id.<imgname>" can be
@@ -173,8 +166,8 @@ struct ceph_osdmap {
 	 * the list of osds that store+replicate them. */
 	struct crush_map *crush;
 
-	struct mutex crush_scratch_mutex;
-	int crush_scratch_ary[CEPH_PG_MAX_SIZE * 3];
+	struct mutex crush_workspace_mutex;
+	void *crush_workspace;
 };
 
 static inline bool ceph_osd_exists(struct ceph_osdmap *map, int osd)

include/linux/ceph/rados.h

@@ -50,7 +50,7 @@ struct ceph_timespec {
 #define CEPH_PG_LAYOUT_LINEAR 2
 #define CEPH_PG_LAYOUT_HYBRID 3
 
-#define CEPH_PG_MAX_SIZE      16  /* max # osds in a single pg */
+#define CEPH_PG_MAX_SIZE      32  /* max # osds in a single pg */
 
 /*
  * placement group.

include/linux/crush/crush.h

@@ -135,13 +135,6 @@ struct crush_bucket {
 	__u32 size;      /* num items */
 	__s32 *items;
 
-	/*
-	 * cached random permutation: used for uniform bucket and for
-	 * the linear search fallback for the other bucket types.
-	 */
-	__u32 perm_x;  /* @x for which *perm is defined */
-	__u32 perm_n;  /* num elements of *perm that are permuted/defined */
-	__u32 *perm;
 };
 
 struct crush_bucket_uniform {
@@ -211,6 +204,21 @@ struct crush_map {
 	 * device fails. */
 	__u8 chooseleaf_stable;
 
+	/*
+	 * This value is calculated after decode or construction by
+	 * the builder. It is exposed here (rather than having a
+	 * 'build CRUSH working space' function) so that callers can
+	 * reserve a static buffer, allocate space on the stack, or
+	 * otherwise avoid calling into the heap allocator if they
+	 * want to. The size of the working space depends on the map,
+	 * while the size of the scratch vector passed to the mapper
+	 * depends on the size of the desired result set.
+	 *
+	 * Nothing stops the caller from allocating both in one swell
+	 * foop and passing in two points, though.
+	 */
+	size_t working_size;
+
 #ifndef __KERNEL__
 	/*
 	 * version 0 (original) of straw_calc has various flaws.  version 1
@@ -248,4 +256,23 @@ static inline int crush_calc_tree_node(int i)
 	return ((i+1) << 1)-1;
 }
 
+/*
+ * These data structures are private to the CRUSH implementation. They
+ * are exposed in this header file because builder needs their
+ * definitions to calculate the total working size.
+ *
+ * Moving this out of the crush map allow us to treat the CRUSH map as
+ * immutable within the mapper and removes the requirement for a CRUSH
+ * map lock.
+ */
+struct crush_work_bucket {
+	__u32 perm_x; /* @x for which *perm is defined */
+	__u32 perm_n; /* num elements of *perm that are permuted/defined */
+	__u32 *perm;  /* Permutation of the bucket's items */
+};
+
+struct crush_work {
+	struct crush_work_bucket **work; /* Per-bucket working store */
+};
+
 #endif

include/linux/crush/mapper.h

@@ -15,6 +15,20 @@ extern int crush_do_rule(const struct crush_map *map,
 			 int ruleno,
 			 int x, int *result, int result_max,
 			 const __u32 *weights, int weight_max,
-			 int *scratch);
+			 void *cwin);
+
+/*
+ * Returns the exact amount of workspace that will need to be used
+ * for a given combination of crush_map and result_max. The caller can
+ * then allocate this much on its own, either on the stack, in a
+ * per-thread long-lived buffer, or however it likes.
+ */
+static inline size_t crush_work_size(const struct crush_map *map,
+				     int result_max)
+{
+	return map->working_size + result_max * 3 * sizeof(__u32);
+}
+
+void crush_init_workspace(const struct crush_map *map, void *v);
 
 #endif