ceph: OSD client
authorSage Weil <sage@newdream.net>
Tue, 6 Oct 2009 18:31:10 +0000 (11:31 -0700)
committerSage Weil <sage@newdream.net>
Tue, 6 Oct 2009 18:31:10 +0000 (11:31 -0700)
The OSD client is responsible for reading and writing data from/to the
object storage pool.  This includes determining where objects are
stored in the cluster, and ensuring that requests are retried or
redirected in the event of a node failure or data migration.

If an OSD does not respond before a timeout expires, keepalive
messages are sent across the lossless, ordered communications channel
to ensure that any break in the TCP is discovered.  If the session
does reset, a reconnection is attempted and affected requests are
resent (by the message transport layer).

Signed-off-by: Sage Weil <sage@newdream.net>
fs/ceph/osd_client.c [new file with mode: 0644]
fs/ceph/osd_client.h [new file with mode: 0644]
fs/ceph/osdmap.c [new file with mode: 0644]
fs/ceph/osdmap.h [new file with mode: 0644]

diff --git a/fs/ceph/osd_client.c b/fs/ceph/osd_client.c
new file mode 100644 (file)
index 0000000..978593a
--- /dev/null
@@ -0,0 +1,1294 @@
+#include "ceph_debug.h"
+
+#include <linux/err.h>
+#include <linux/highmem.h>
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+#include <linux/slab.h>
+#include <linux/uaccess.h>
+
+#include "super.h"
+#include "osd_client.h"
+#include "messenger.h"
+#include "decode.h"
+
+const static struct ceph_connection_operations osd_con_ops;
+
+static void kick_requests(struct ceph_osd_client *osdc, struct ceph_osd *osd);
+
+/*
+ * Implement client access to distributed object storage cluster.
+ *
+ * All data objects are stored within a cluster/cloud of OSDs, or
+ * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
+ * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
+ * remote daemons serving up and coordinating consistent and safe
+ * access to storage.
+ *
+ * Cluster membership and the mapping of data objects onto storage devices
+ * are described by the osd map.
+ *
+ * We keep track of pending OSD requests (read, write), resubmit
+ * requests to different OSDs when the cluster topology/data layout
+ * change, or retry the affected requests when the communications
+ * channel with an OSD is reset.
+ */
+
+/*
+ * calculate the mapping of a file extent onto an object, and fill out the
+ * request accordingly.  shorten extent as necessary if it crosses an
+ * object boundary.
+ *
+ * fill osd op in request message.
+ */
+static void calc_layout(struct ceph_osd_client *osdc,
+                       struct ceph_vino vino, struct ceph_file_layout *layout,
+                       u64 off, u64 *plen,
+                       struct ceph_osd_request *req)
+{
+       struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
+       struct ceph_osd_op *op = (void *)(reqhead + 1);
+       u64 orig_len = *plen;
+       u64 objoff, objlen;    /* extent in object */
+       u64 bno;
+
+       reqhead->snapid = cpu_to_le64(vino.snap);
+
+       /* object extent? */
+       ceph_calc_file_object_mapping(layout, off, plen, &bno,
+                                     &objoff, &objlen);
+       if (*plen < orig_len)
+               dout(" skipping last %llu, final file extent %llu~%llu\n",
+                    orig_len - *plen, off, *plen);
+
+       sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
+       req->r_oid_len = strlen(req->r_oid);
+
+       op->extent.offset = cpu_to_le64(objoff);
+       op->extent.length = cpu_to_le64(objlen);
+       req->r_num_pages = calc_pages_for(off, *plen);
+
+       dout("calc_layout %s (%d) %llu~%llu (%d pages)\n",
+            req->r_oid, req->r_oid_len, objoff, objlen, req->r_num_pages);
+}
+
+
+/*
+ * requests
+ */
+void ceph_osdc_put_request(struct ceph_osd_request *req)
+{
+       dout("osdc put_request %p %d -> %d\n", req, atomic_read(&req->r_ref),
+            atomic_read(&req->r_ref)-1);
+       BUG_ON(atomic_read(&req->r_ref) <= 0);
+       if (atomic_dec_and_test(&req->r_ref)) {
+               if (req->r_request)
+                       ceph_msg_put(req->r_request);
+               if (req->r_reply)
+                       ceph_msg_put(req->r_reply);
+               if (req->r_own_pages)
+                       ceph_release_page_vector(req->r_pages,
+                                                req->r_num_pages);
+               ceph_put_snap_context(req->r_snapc);
+               if (req->r_mempool)
+                       mempool_free(req, req->r_osdc->req_mempool);
+               else
+                       kfree(req);
+       }
+}
+
+/*
+ * build new request AND message, calculate layout, and adjust file
+ * extent as needed.
+ *
+ * if the file was recently truncated, we include information about its
+ * old and new size so that the object can be updated appropriately.  (we
+ * avoid synchronously deleting truncated objects because it's slow.)
+ *
+ * if @do_sync, include a 'startsync' command so that the osd will flush
+ * data quickly.
+ */
+struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
+                                              struct ceph_file_layout *layout,
+                                              struct ceph_vino vino,
+                                              u64 off, u64 *plen,
+                                              int opcode, int flags,
+                                              struct ceph_snap_context *snapc,
+                                              int do_sync,
+                                              u32 truncate_seq,
+                                              u64 truncate_size,
+                                              struct timespec *mtime,
+                                              bool use_mempool, int num_reply)
+{
+       struct ceph_osd_request *req;
+       struct ceph_msg *msg;
+       struct ceph_osd_request_head *head;
+       struct ceph_osd_op *op;
+       void *p;
+       int do_trunc = truncate_seq && (off + *plen > truncate_size);
+       int num_op = 1 + do_sync + do_trunc;
+       size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
+       int err, i;
+       u64 prevofs;
+
+       if (use_mempool) {
+               req = mempool_alloc(osdc->req_mempool, GFP_NOFS);
+               memset(req, 0, sizeof(*req));
+       } else {
+               req = kzalloc(sizeof(*req), GFP_NOFS);
+       }
+       if (req == NULL)
+               return ERR_PTR(-ENOMEM);
+
+       err = ceph_msgpool_resv(&osdc->msgpool_op_reply, num_reply);
+       if (err) {
+               ceph_osdc_put_request(req);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       req->r_osdc = osdc;
+       req->r_mempool = use_mempool;
+       atomic_set(&req->r_ref, 1);
+       init_completion(&req->r_completion);
+       init_completion(&req->r_safe_completion);
+       INIT_LIST_HEAD(&req->r_unsafe_item);
+       req->r_flags = flags;
+
+       WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
+
+       /* create message; allow space for oid */
+       msg_size += 40;
+       if (snapc)
+               msg_size += sizeof(u64) * snapc->num_snaps;
+       if (use_mempool)
+               msg = ceph_msgpool_get(&osdc->msgpool_op);
+       else
+               msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, 0, 0, NULL);
+       if (IS_ERR(msg)) {
+               ceph_msgpool_resv(&osdc->msgpool_op_reply, num_reply);
+               ceph_osdc_put_request(req);
+               return ERR_PTR(PTR_ERR(msg));
+       }
+       msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
+       memset(msg->front.iov_base, 0, msg->front.iov_len);
+       head = msg->front.iov_base;
+       op = (void *)(head + 1);
+       p = (void *)(op + num_op);
+
+       req->r_request = msg;
+       req->r_snapc = ceph_get_snap_context(snapc);
+
+       head->client_inc = cpu_to_le32(1); /* always, for now. */
+       head->flags = cpu_to_le32(flags);
+       if (flags & CEPH_OSD_FLAG_WRITE)
+               ceph_encode_timespec(&head->mtime, mtime);
+       head->num_ops = cpu_to_le16(num_op);
+       op->op = cpu_to_le16(opcode);
+
+       /* calculate max write size */
+       calc_layout(osdc, vino, layout, off, plen, req);
+       req->r_file_layout = *layout;  /* keep a copy */
+
+       if (flags & CEPH_OSD_FLAG_WRITE) {
+               req->r_request->hdr.data_off = cpu_to_le16(off);
+               req->r_request->hdr.data_len = cpu_to_le32(*plen);
+               op->payload_len = cpu_to_le32(*plen);
+       }
+
+       /* fill in oid */
+       head->object_len = cpu_to_le32(req->r_oid_len);
+       memcpy(p, req->r_oid, req->r_oid_len);
+       p += req->r_oid_len;
+
+       /* additional ops */
+       if (do_trunc) {
+               op++;
+               op->op = cpu_to_le16(opcode == CEPH_OSD_OP_READ ?
+                            CEPH_OSD_OP_MASKTRUNC : CEPH_OSD_OP_SETTRUNC);
+               op->trunc.truncate_seq = cpu_to_le32(truncate_seq);
+               prevofs = le64_to_cpu((op-1)->extent.offset);
+               op->trunc.truncate_size = cpu_to_le64(truncate_size -
+                                                     (off-prevofs));
+       }
+       if (do_sync) {
+               op++;
+               op->op = cpu_to_le16(CEPH_OSD_OP_STARTSYNC);
+       }
+       if (snapc) {
+               head->snap_seq = cpu_to_le64(snapc->seq);
+               head->num_snaps = cpu_to_le32(snapc->num_snaps);
+               for (i = 0; i < snapc->num_snaps; i++) {
+                       put_unaligned_le64(snapc->snaps[i], p);
+                       p += sizeof(u64);
+               }
+       }
+
+       BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
+       return req;
+}
+
+/*
+ * We keep osd requests in an rbtree, sorted by ->r_tid.
+ */
+static void __insert_request(struct ceph_osd_client *osdc,
+                            struct ceph_osd_request *new)
+{
+       struct rb_node **p = &osdc->requests.rb_node;
+       struct rb_node *parent = NULL;
+       struct ceph_osd_request *req = NULL;
+
+       while (*p) {
+               parent = *p;
+               req = rb_entry(parent, struct ceph_osd_request, r_node);
+               if (new->r_tid < req->r_tid)
+                       p = &(*p)->rb_left;
+               else if (new->r_tid > req->r_tid)
+                       p = &(*p)->rb_right;
+               else
+                       BUG();
+       }
+
+       rb_link_node(&new->r_node, parent, p);
+       rb_insert_color(&new->r_node, &osdc->requests);
+}
+
+static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
+                                                u64 tid)
+{
+       struct ceph_osd_request *req;
+       struct rb_node *n = osdc->requests.rb_node;
+
+       while (n) {
+               req = rb_entry(n, struct ceph_osd_request, r_node);
+               if (tid < req->r_tid)
+                       n = n->rb_left;
+               else if (tid > req->r_tid)
+                       n = n->rb_right;
+               else
+                       return req;
+       }
+       return NULL;
+}
+
+static struct ceph_osd_request *
+__lookup_request_ge(struct ceph_osd_client *osdc,
+                   u64 tid)
+{
+       struct ceph_osd_request *req;
+       struct rb_node *n = osdc->requests.rb_node;
+
+       while (n) {
+               req = rb_entry(n, struct ceph_osd_request, r_node);
+               if (tid < req->r_tid) {
+                       if (!n->rb_left)
+                               return req;
+                       n = n->rb_left;
+               } else if (tid > req->r_tid) {
+                       n = n->rb_right;
+               } else {
+                       return req;
+               }
+       }
+       return NULL;
+}
+
+
+/*
+ * The messaging layer will reconnect to the osd as needed.  If the
+ * session has dropped, the OSD will have dropped the session state,
+ * and we'll get notified by the messaging layer.  If that happens, we
+ * need to resubmit all requests for that osd.
+ */
+static void osd_reset(struct ceph_connection *con)
+{
+       struct ceph_osd *osd = con->private;
+       struct ceph_osd_client *osdc;
+
+       if (!osd)
+               return;
+       dout("osd_reset osd%d\n", osd->o_osd);
+       osdc = osd->o_osdc;
+       osd->o_incarnation++;
+       down_read(&osdc->map_sem);
+       kick_requests(osdc, osd);
+       up_read(&osdc->map_sem);
+}
+
+/*
+ * Track open sessions with osds.
+ */
+static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
+{
+       struct ceph_osd *osd;
+
+       osd = kzalloc(sizeof(*osd), GFP_NOFS);
+       if (!osd)
+               return NULL;
+
+       atomic_set(&osd->o_ref, 1);
+       osd->o_osdc = osdc;
+       INIT_LIST_HEAD(&osd->o_requests);
+       osd->o_incarnation = 1;
+
+       ceph_con_init(osdc->client->msgr, &osd->o_con);
+       osd->o_con.private = osd;
+       osd->o_con.ops = &osd_con_ops;
+       osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
+       return osd;
+}
+
+static struct ceph_osd *get_osd(struct ceph_osd *osd)
+{
+       if (atomic_inc_not_zero(&osd->o_ref)) {
+               dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
+                    atomic_read(&osd->o_ref));
+               return osd;
+       } else {
+               dout("get_osd %p FAIL\n", osd);
+               return NULL;
+       }
+}
+
+static void put_osd(struct ceph_osd *osd)
+{
+       dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
+            atomic_read(&osd->o_ref) - 1);
+       if (atomic_dec_and_test(&osd->o_ref)) {
+               ceph_con_shutdown(&osd->o_con);
+               kfree(osd);
+       }
+}
+
+/*
+ * remove an osd from our map
+ */
+static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
+{
+       dout("remove_osd %p\n", osd);
+       BUG_ON(!list_empty(&osd->o_requests));
+       rb_erase(&osd->o_node, &osdc->osds);
+       ceph_con_close(&osd->o_con);
+       put_osd(osd);
+}
+
+/*
+ * reset osd connect
+ */
+static int reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
+{
+       int ret = 0;
+
+       dout("reset_osd %p osd%d\n", osd, osd->o_osd);
+       if (list_empty(&osd->o_requests)) {
+               remove_osd(osdc, osd);
+       } else {
+               ceph_con_close(&osd->o_con);
+               ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
+               osd->o_incarnation++;
+       }
+       return ret;
+}
+
+static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
+{
+       struct rb_node **p = &osdc->osds.rb_node;
+       struct rb_node *parent = NULL;
+       struct ceph_osd *osd = NULL;
+
+       while (*p) {
+               parent = *p;
+               osd = rb_entry(parent, struct ceph_osd, o_node);
+               if (new->o_osd < osd->o_osd)
+                       p = &(*p)->rb_left;
+               else if (new->o_osd > osd->o_osd)
+                       p = &(*p)->rb_right;
+               else
+                       BUG();
+       }
+
+       rb_link_node(&new->o_node, parent, p);
+       rb_insert_color(&new->o_node, &osdc->osds);
+}
+
+static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
+{
+       struct ceph_osd *osd;
+       struct rb_node *n = osdc->osds.rb_node;
+
+       while (n) {
+               osd = rb_entry(n, struct ceph_osd, o_node);
+               if (o < osd->o_osd)
+                       n = n->rb_left;
+               else if (o > osd->o_osd)
+                       n = n->rb_right;
+               else
+                       return osd;
+       }
+       return NULL;
+}
+
+
+/*
+ * Register request, assign tid.  If this is the first request, set up
+ * the timeout event.
+ */
+static void register_request(struct ceph_osd_client *osdc,
+                            struct ceph_osd_request *req)
+{
+       struct ceph_osd_request_head *head = req->r_request->front.iov_base;
+
+       mutex_lock(&osdc->request_mutex);
+       req->r_tid = ++osdc->last_tid;
+       head->tid = cpu_to_le64(req->r_tid);
+
+       dout("register_request %p tid %lld\n", req, req->r_tid);
+       __insert_request(osdc, req);
+       ceph_osdc_get_request(req);
+       osdc->num_requests++;
+
+       req->r_timeout_stamp =
+               jiffies + osdc->client->mount_args.osd_timeout*HZ;
+
+       if (osdc->num_requests == 1) {
+               osdc->timeout_tid = req->r_tid;
+               dout("  timeout on tid %llu at %lu\n", req->r_tid,
+                    req->r_timeout_stamp);
+               schedule_delayed_work(&osdc->timeout_work,
+                     round_jiffies_relative(req->r_timeout_stamp - jiffies));
+       }
+       mutex_unlock(&osdc->request_mutex);
+}
+
+/*
+ * called under osdc->request_mutex
+ */
+static void __unregister_request(struct ceph_osd_client *osdc,
+                                struct ceph_osd_request *req)
+{
+       dout("__unregister_request %p tid %lld\n", req, req->r_tid);
+       rb_erase(&req->r_node, &osdc->requests);
+       osdc->num_requests--;
+
+       list_del_init(&req->r_osd_item);
+       if (list_empty(&req->r_osd->o_requests))
+               remove_osd(osdc, req->r_osd);
+       req->r_osd = NULL;
+
+       ceph_osdc_put_request(req);
+
+       if (req->r_tid == osdc->timeout_tid) {
+               if (osdc->num_requests == 0) {
+                       dout("no requests, canceling timeout\n");
+                       osdc->timeout_tid = 0;
+                       cancel_delayed_work(&osdc->timeout_work);
+               } else {
+                       req = rb_entry(rb_first(&osdc->requests),
+                                      struct ceph_osd_request, r_node);
+                       osdc->timeout_tid = req->r_tid;
+                       dout("rescheduled timeout on tid %llu at %lu\n",
+                            req->r_tid, req->r_timeout_stamp);
+                       schedule_delayed_work(&osdc->timeout_work,
+                             round_jiffies_relative(req->r_timeout_stamp -
+                                                    jiffies));
+               }
+       }
+}
+
+/*
+ * Cancel a previously queued request message
+ */
+static void __cancel_request(struct ceph_osd_request *req)
+{
+       if (req->r_sent) {
+               ceph_con_revoke(&req->r_osd->o_con, req->r_request);
+               req->r_sent = 0;
+       }
+}
+
+/*
+ * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
+ * (as needed), and set the request r_osd appropriately.  If there is
+ * no up osd, set r_osd to NULL.
+ *
+ * Return 0 if unchanged, 1 if changed, or negative on error.
+ *
+ * Caller should hold map_sem for read and request_mutex.
+ */
+static int __map_osds(struct ceph_osd_client *osdc,
+                     struct ceph_osd_request *req)
+{
+       struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
+       union ceph_pg pgid;
+       int o = -1;
+       int err;
+       struct ceph_osd *newosd = NULL;
+
+       dout("map_osds %p tid %lld\n", req, req->r_tid);
+       err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
+                                     &req->r_file_layout, osdc->osdmap);
+       if (err)
+               return err;
+       pgid.pg64 = le64_to_cpu(reqhead->layout.ol_pgid);
+       o = ceph_calc_pg_primary(osdc->osdmap, pgid);
+
+       if ((req->r_osd && req->r_osd->o_osd == o &&
+            req->r_sent >= req->r_osd->o_incarnation) ||
+           (req->r_osd == NULL && o == -1))
+               return 0;  /* no change */
+
+       dout("map_osds tid %llu pgid %llx pool %d osd%d (was osd%d)\n",
+            req->r_tid, pgid.pg64, pgid.pg.pool, o,
+            req->r_osd ? req->r_osd->o_osd : -1);
+
+       if (req->r_osd) {
+               __cancel_request(req);
+               list_del_init(&req->r_osd_item);
+               if (list_empty(&req->r_osd->o_requests)) {
+                       /* try to re-use r_osd if possible */
+                       newosd = get_osd(req->r_osd);
+                       remove_osd(osdc, newosd);
+               }
+               req->r_osd = NULL;
+       }
+
+       req->r_osd = __lookup_osd(osdc, o);
+       if (!req->r_osd && o >= 0) {
+               if (newosd) {
+                       req->r_osd = newosd;
+                       newosd = NULL;
+               } else {
+                       err = -ENOMEM;
+                       req->r_osd = create_osd(osdc);
+                       if (!req->r_osd)
+                               goto out;
+               }
+
+               dout("map_osds osd %p is osd%d\n", req->r_osd, o);
+               req->r_osd->o_osd = o;
+               req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
+               __insert_osd(osdc, req->r_osd);
+
+               ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
+       }
+
+       if (req->r_osd)
+               list_add(&req->r_osd_item, &req->r_osd->o_requests);
+       err = 1;   /* osd changed */
+
+out:
+       if (newosd)
+               put_osd(newosd);
+       return err;
+}
+
+/*
+ * caller should hold map_sem (for read) and request_mutex
+ */
+static int __send_request(struct ceph_osd_client *osdc,
+                         struct ceph_osd_request *req)
+{
+       struct ceph_osd_request_head *reqhead;
+       int err;
+
+       err = __map_osds(osdc, req);
+       if (err < 0)
+               return err;
+       if (req->r_osd == NULL) {
+               dout("send_request %p no up osds in pg\n", req);
+               ceph_monc_request_next_osdmap(&osdc->client->monc);
+               return 0;
+       }
+
+       dout("send_request %p tid %llu to osd%d flags %d\n",
+            req, req->r_tid, req->r_osd->o_osd, req->r_flags);
+
+       reqhead = req->r_request->front.iov_base;
+       reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
+       reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
+       reqhead->reassert_version = req->r_reassert_version;
+
+       req->r_timeout_stamp = jiffies+osdc->client->mount_args.osd_timeout*HZ;
+
+       ceph_msg_get(req->r_request); /* send consumes a ref */
+       ceph_con_send(&req->r_osd->o_con, req->r_request);
+       req->r_sent = req->r_osd->o_incarnation;
+       return 0;
+}
+
+/*
+ * Timeout callback, called every N seconds when 1 or more osd
+ * requests has been active for more than N seconds.  When this
+ * happens, we ping all OSDs with requests who have timed out to
+ * ensure any communications channel reset is detected.  Reset the
+ * request timeouts another N seconds in the future as we go.
+ * Reschedule the timeout event another N seconds in future (unless
+ * there are no open requests).
+ */
+static void handle_timeout(struct work_struct *work)
+{
+       struct ceph_osd_client *osdc =
+               container_of(work, struct ceph_osd_client, timeout_work.work);
+       struct ceph_osd_request *req;
+       struct ceph_osd *osd;
+       unsigned long timeout = osdc->client->mount_args.osd_timeout * HZ;
+       unsigned long next_timeout = timeout + jiffies;
+       struct rb_node *p;
+
+       dout("timeout\n");
+       down_read(&osdc->map_sem);
+
+       ceph_monc_request_next_osdmap(&osdc->client->monc);
+
+       mutex_lock(&osdc->request_mutex);
+       for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
+               req = rb_entry(p, struct ceph_osd_request, r_node);
+
+               if (req->r_resend) {
+                       int err;
+
+                       dout("osdc resending prev failed %lld\n", req->r_tid);
+                       err = __send_request(osdc, req);
+                       if (err)
+                               dout("osdc failed again on %lld\n", req->r_tid);
+                       else
+                               req->r_resend = false;
+                       continue;
+               }
+       }
+       for (p = rb_first(&osdc->osds); p; p = rb_next(p)) {
+               osd = rb_entry(p, struct ceph_osd, o_node);
+               if (list_empty(&osd->o_requests))
+                       continue;
+               req = list_first_entry(&osd->o_requests,
+                                      struct ceph_osd_request, r_osd_item);
+               if (time_before(jiffies, req->r_timeout_stamp))
+                       continue;
+
+               dout(" tid %llu (at least) timed out on osd%d\n",
+                    req->r_tid, osd->o_osd);
+               req->r_timeout_stamp = next_timeout;
+               ceph_con_keepalive(&osd->o_con);
+       }
+
+       if (osdc->timeout_tid)
+               schedule_delayed_work(&osdc->timeout_work,
+                                     round_jiffies_relative(timeout));
+
+       mutex_unlock(&osdc->request_mutex);
+
+       up_read(&osdc->map_sem);
+}
+
+/*
+ * handle osd op reply.  either call the callback if it is specified,
+ * or do the completion to wake up the waiting thread.
+ */
+static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg)
+{
+       struct ceph_osd_reply_head *rhead = msg->front.iov_base;
+       struct ceph_osd_request *req;
+       u64 tid;
+       int numops, object_len, flags;
+
+       if (msg->front.iov_len < sizeof(*rhead))
+               goto bad;
+       tid = le64_to_cpu(rhead->tid);
+       numops = le32_to_cpu(rhead->num_ops);
+       object_len = le32_to_cpu(rhead->object_len);
+       if (msg->front.iov_len != sizeof(*rhead) + object_len +
+           numops * sizeof(struct ceph_osd_op))
+               goto bad;
+       dout("handle_reply %p tid %llu\n", msg, tid);
+
+       /* lookup */
+       mutex_lock(&osdc->request_mutex);
+       req = __lookup_request(osdc, tid);
+       if (req == NULL) {
+               dout("handle_reply tid %llu dne\n", tid);
+               mutex_unlock(&osdc->request_mutex);
+               return;
+       }
+       ceph_osdc_get_request(req);
+       flags = le32_to_cpu(rhead->flags);
+
+       if (req->r_reply) {
+               /*
+                * once we see the message has been received, we don't
+                * need a ref (which is only needed for revoking
+                * pages)
+                */
+               ceph_msg_put(req->r_reply);
+               req->r_reply = NULL;
+       }
+
+       if (!req->r_got_reply) {
+               unsigned bytes;
+
+               req->r_result = le32_to_cpu(rhead->result);
+               bytes = le32_to_cpu(msg->hdr.data_len);
+               dout("handle_reply result %d bytes %d\n", req->r_result,
+                    bytes);
+               if (req->r_result == 0)
+                       req->r_result = bytes;
+
+               /* in case this is a write and we need to replay, */
+               req->r_reassert_version = rhead->reassert_version;
+
+               req->r_got_reply = 1;
+       } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
+               dout("handle_reply tid %llu dup ack\n", tid);
+               goto done;
+       }
+
+       dout("handle_reply tid %llu flags %d\n", tid, flags);
+
+       /* either this is a read, or we got the safe response */
+       if ((flags & CEPH_OSD_FLAG_ONDISK) ||
+           ((flags & CEPH_OSD_FLAG_WRITE) == 0))
+               __unregister_request(osdc, req);
+
+       mutex_unlock(&osdc->request_mutex);
+
+       if (req->r_callback)
+               req->r_callback(req, msg);
+       else
+               complete(&req->r_completion);
+
+       if (flags & CEPH_OSD_FLAG_ONDISK) {
+               if (req->r_safe_callback)
+                       req->r_safe_callback(req, msg);
+               complete(&req->r_safe_completion);  /* fsync waiter */
+       }
+
+done:
+       ceph_osdc_put_request(req);
+       return;
+
+bad:
+       pr_err("corrupt osd_op_reply got %d %d expected %d\n",
+              (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
+              (int)sizeof(*rhead));
+}
+
+
+/*
+ * Resubmit osd requests whose osd or osd address has changed.  Request
+ * a new osd map if osds are down, or we are otherwise unable to determine
+ * how to direct a request.
+ *
+ * Close connections to down osds.
+ *
+ * If @who is specified, resubmit requests for that specific osd.
+ *
+ * Caller should hold map_sem for read and request_mutex.
+ */
+static void kick_requests(struct ceph_osd_client *osdc,
+                         struct ceph_osd *kickosd)
+{
+       struct ceph_osd_request *req;
+       struct rb_node *p, *n;
+       int needmap = 0;
+       int err;
+
+       dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
+       mutex_lock(&osdc->request_mutex);
+       if (!kickosd) {
+               for (p = rb_first(&osdc->osds); p; p = n) {
+                       struct ceph_osd *osd =
+                               rb_entry(p, struct ceph_osd, o_node);
+
+                       n = rb_next(p);
+                       if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
+                           !ceph_entity_addr_equal(&osd->o_con.peer_addr,
+                                           ceph_osd_addr(osdc->osdmap,
+                                                         osd->o_osd)))
+                               reset_osd(osdc, osd);
+               }
+       }
+
+       for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
+               req = rb_entry(p, struct ceph_osd_request, r_node);
+
+               if (req->r_resend) {
+                       dout(" r_resend set on tid %llu\n", req->r_tid);
+                       goto kick;
+               }
+               if (req->r_osd && kickosd == req->r_osd)
+                       goto kick;
+
+               err = __map_osds(osdc, req);
+               if (err == 0)
+                       continue;  /* no change */
+               if (err < 0) {
+                       /*
+                        * FIXME: really, we should set the request
+                        * error and fail if this isn't a 'nofail'
+                        * request, but that's a fair bit more
+                        * complicated to do.  So retry!
+                        */
+                       dout(" setting r_resend on %llu\n", req->r_tid);
+                       req->r_resend = true;
+                       continue;
+               }
+               if (req->r_osd == NULL) {
+                       dout("tid %llu maps to no valid osd\n", req->r_tid);
+                       needmap++;  /* request a newer map */
+                       continue;
+               }
+
+kick:
+               dout("kicking tid %llu osd%d\n", req->r_tid, req->r_osd->o_osd);
+               req->r_flags |= CEPH_OSD_FLAG_RETRY;
+               err = __send_request(osdc, req);
+               if (err) {
+                       dout(" setting r_resend on %llu\n", req->r_tid);
+                       req->r_resend = true;
+               }
+       }
+       mutex_unlock(&osdc->request_mutex);
+
+       if (needmap) {
+               dout("%d requests for down osds, need new map\n", needmap);
+               ceph_monc_request_next_osdmap(&osdc->client->monc);
+       }
+}
+
+/*
+ * Process updated osd map.
+ *
+ * The message contains any number of incremental and full maps, normally
+ * indicating some sort of topology change in the cluster.  Kick requests
+ * off to different OSDs as needed.
+ */
+void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
+{
+       void *p, *end, *next;
+       u32 nr_maps, maplen;
+       u32 epoch;
+       struct ceph_osdmap *newmap = NULL, *oldmap;
+       int err;
+       struct ceph_fsid fsid;
+
+       dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
+       p = msg->front.iov_base;
+       end = p + msg->front.iov_len;
+
+       /* verify fsid */
+       ceph_decode_need(&p, end, sizeof(fsid), bad);
+       ceph_decode_copy(&p, &fsid, sizeof(fsid));
+       if (ceph_fsid_compare(&fsid, &osdc->client->monc.monmap->fsid)) {
+               pr_err("got osdmap with wrong fsid, ignoring\n");
+               return;
+       }
+
+       down_write(&osdc->map_sem);
+
+       /* incremental maps */
+       ceph_decode_32_safe(&p, end, nr_maps, bad);
+       dout(" %d inc maps\n", nr_maps);
+       while (nr_maps > 0) {
+               ceph_decode_need(&p, end, 2*sizeof(u32), bad);
+               ceph_decode_32(&p, epoch);
+               ceph_decode_32(&p, maplen);
+               ceph_decode_need(&p, end, maplen, bad);
+               next = p + maplen;
+               if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
+                       dout("applying incremental map %u len %d\n",
+                            epoch, maplen);
+                       newmap = osdmap_apply_incremental(&p, next,
+                                                         osdc->osdmap,
+                                                         osdc->client->msgr);
+                       if (IS_ERR(newmap)) {
+                               err = PTR_ERR(newmap);
+                               goto bad;
+                       }
+                       if (newmap != osdc->osdmap) {
+                               ceph_osdmap_destroy(osdc->osdmap);
+                               osdc->osdmap = newmap;
+                       }
+               } else {
+                       dout("ignoring incremental map %u len %d\n",
+                            epoch, maplen);
+               }
+               p = next;
+               nr_maps--;
+       }
+       if (newmap)
+               goto done;
+
+       /* full maps */
+       ceph_decode_32_safe(&p, end, nr_maps, bad);
+       dout(" %d full maps\n", nr_maps);
+       while (nr_maps) {
+               ceph_decode_need(&p, end, 2*sizeof(u32), bad);
+               ceph_decode_32(&p, epoch);
+               ceph_decode_32(&p, maplen);
+               ceph_decode_need(&p, end, maplen, bad);
+               if (nr_maps > 1) {
+                       dout("skipping non-latest full map %u len %d\n",
+                            epoch, maplen);
+               } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
+                       dout("skipping full map %u len %d, "
+                            "older than our %u\n", epoch, maplen,
+                            osdc->osdmap->epoch);
+               } else {
+                       dout("taking full map %u len %d\n", epoch, maplen);
+                       newmap = osdmap_decode(&p, p+maplen);
+                       if (IS_ERR(newmap)) {
+                               err = PTR_ERR(newmap);
+                               goto bad;
+                       }
+                       oldmap = osdc->osdmap;
+                       osdc->osdmap = newmap;
+                       if (oldmap)
+                               ceph_osdmap_destroy(oldmap);
+               }
+               p += maplen;
+               nr_maps--;
+       }
+
+done:
+       downgrade_write(&osdc->map_sem);
+       ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
+       if (newmap)
+               kick_requests(osdc, NULL);
+       up_read(&osdc->map_sem);
+       return;
+
+bad:
+       pr_err("osdc handle_map corrupt msg\n");
+       up_write(&osdc->map_sem);
+       return;
+}
+
+
+/*
+ * A read request prepares specific pages that data is to be read into.
+ * When a message is being read off the wire, we call prepare_pages to
+ * find those pages.
+ *  0 = success, -1 failure.
+ */
+static int prepare_pages(struct ceph_connection *con, struct ceph_msg *m,
+                        int want)
+{
+       struct ceph_osd *osd = con->private;
+       struct ceph_osd_client *osdc;
+       struct ceph_osd_reply_head *rhead = m->front.iov_base;
+       struct ceph_osd_request *req;
+       u64 tid;
+       int ret = -1;
+       int type = le16_to_cpu(m->hdr.type);
+
+       if (!osd)
+               return -1;
+       osdc = osd->o_osdc;
+
+       dout("prepare_pages on msg %p want %d\n", m, want);
+       if (unlikely(type != CEPH_MSG_OSD_OPREPLY))
+               return -1;  /* hmm! */
+
+       tid = le64_to_cpu(rhead->tid);
+       mutex_lock(&osdc->request_mutex);
+       req = __lookup_request(osdc, tid);
+       if (!req) {
+               dout("prepare_pages unknown tid %llu\n", tid);
+               goto out;
+       }
+       dout("prepare_pages tid %llu has %d pages, want %d\n",
+            tid, req->r_num_pages, want);
+       if (likely(req->r_num_pages >= want && !req->r_prepared_pages)) {
+               m->pages = req->r_pages;
+               m->nr_pages = req->r_num_pages;
+               req->r_reply = m;  /* only for duration of read over socket */
+               ceph_msg_get(m);
+               req->r_prepared_pages = 1;
+               ret = 0; /* success */
+       }
+out:
+       mutex_unlock(&osdc->request_mutex);
+       return ret;
+}
+
+/*
+ * Register request, send initial attempt.
+ */
+int ceph_osdc_start_request(struct ceph_osd_client *osdc,
+                           struct ceph_osd_request *req,
+                           bool nofail)
+{
+       int rc;
+
+       req->r_request->pages = req->r_pages;
+       req->r_request->nr_pages = req->r_num_pages;
+
+       register_request(osdc, req);
+
+       down_read(&osdc->map_sem);
+       mutex_lock(&osdc->request_mutex);
+       rc = __send_request(osdc, req);
+       if (rc) {
+               if (nofail) {
+                       dout("osdc_start_request failed send, marking %lld\n",
+                            req->r_tid);
+                       req->r_resend = true;
+                       rc = 0;
+               } else {
+                       __unregister_request(osdc, req);
+               }
+       }
+       mutex_unlock(&osdc->request_mutex);
+       up_read(&osdc->map_sem);
+       return rc;
+}
+
+/*
+ * wait for a request to complete
+ */
+int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
+                          struct ceph_osd_request *req)
+{
+       int rc;
+
+       rc = wait_for_completion_interruptible(&req->r_completion);
+       if (rc < 0) {
+               mutex_lock(&osdc->request_mutex);
+               __cancel_request(req);
+               mutex_unlock(&osdc->request_mutex);
+               dout("wait_request tid %llu timed out\n", req->r_tid);
+               return rc;
+       }
+
+       dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
+       return req->r_result;
+}
+
+/*
+ * sync - wait for all in-flight requests to flush.  avoid starvation.
+ */
+void ceph_osdc_sync(struct ceph_osd_client *osdc)
+{
+       struct ceph_osd_request *req;
+       u64 last_tid, next_tid = 0;
+
+       mutex_lock(&osdc->request_mutex);
+       last_tid = osdc->last_tid;
+       while (1) {
+               req = __lookup_request_ge(osdc, next_tid);
+               if (!req)
+                       break;
+               if (req->r_tid > last_tid)
+                       break;
+
+               next_tid = req->r_tid + 1;
+               if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
+                       continue;
+
+               ceph_osdc_get_request(req);
+               mutex_unlock(&osdc->request_mutex);
+               dout("sync waiting on tid %llu (last is %llu)\n",
+                    req->r_tid, last_tid);
+               wait_for_completion(&req->r_safe_completion);
+               mutex_lock(&osdc->request_mutex);
+               ceph_osdc_put_request(req);
+       }
+       mutex_unlock(&osdc->request_mutex);
+       dout("sync done (thru tid %llu)\n", last_tid);
+}
+
+/*
+ * init, shutdown
+ */
+int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
+{
+       int err;
+
+       dout("init\n");
+       osdc->client = client;
+       osdc->osdmap = NULL;
+       init_rwsem(&osdc->map_sem);
+       init_completion(&osdc->map_waiters);
+       osdc->last_requested_map = 0;
+       mutex_init(&osdc->request_mutex);
+       osdc->timeout_tid = 0;
+       osdc->last_tid = 0;
+       osdc->osds = RB_ROOT;
+       osdc->requests = RB_ROOT;
+       osdc->num_requests = 0;
+       INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
+
+       osdc->req_mempool = mempool_create_kmalloc_pool(10,
+                                       sizeof(struct ceph_osd_request));
+       if (!osdc->req_mempool)
+               return -ENOMEM;
+
+       err = ceph_msgpool_init(&osdc->msgpool_op, 4096, 10, true);
+       if (err < 0)
+               return -ENOMEM;
+       err = ceph_msgpool_init(&osdc->msgpool_op_reply, 512, 0, false);
+       if (err < 0)
+               return -ENOMEM;
+
+       return 0;
+}
+
+void ceph_osdc_stop(struct ceph_osd_client *osdc)
+{
+       cancel_delayed_work_sync(&osdc->timeout_work);
+       if (osdc->osdmap) {
+               ceph_osdmap_destroy(osdc->osdmap);
+               osdc->osdmap = NULL;
+       }
+       mempool_destroy(osdc->req_mempool);
+       ceph_msgpool_destroy(&osdc->msgpool_op);
+       ceph_msgpool_destroy(&osdc->msgpool_op_reply);
+}
+
+/*
+ * Read some contiguous pages.  If we cross a stripe boundary, shorten
+ * *plen.  Return number of bytes read, or error.
+ */
+int ceph_osdc_readpages(struct ceph_osd_client *osdc,
+                       struct ceph_vino vino, struct ceph_file_layout *layout,
+                       u64 off, u64 *plen,
+                       u32 truncate_seq, u64 truncate_size,
+                       struct page **pages, int num_pages)
+{
+       struct ceph_osd_request *req;
+       int rc = 0;
+
+       dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
+            vino.snap, off, *plen);
+       req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
+                                   CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
+                                   NULL, 0, truncate_seq, truncate_size, NULL,
+                                   false, 1);
+       if (IS_ERR(req))
+               return PTR_ERR(req);
+
+       /* it may be a short read due to an object boundary */
+       req->r_pages = pages;
+       num_pages = calc_pages_for(off, *plen);
+       req->r_num_pages = num_pages;
+
+       dout("readpages  final extent is %llu~%llu (%d pages)\n",
+            off, *plen, req->r_num_pages);
+
+       rc = ceph_osdc_start_request(osdc, req, false);
+       if (!rc)
+               rc = ceph_osdc_wait_request(osdc, req);
+
+       ceph_osdc_put_request(req);
+       dout("readpages result %d\n", rc);
+       return rc;
+}
+
+/*
+ * do a synchronous write on N pages
+ */
+int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
+                        struct ceph_file_layout *layout,
+                        struct ceph_snap_context *snapc,
+                        u64 off, u64 len,
+                        u32 truncate_seq, u64 truncate_size,
+                        struct timespec *mtime,
+                        struct page **pages, int num_pages,
+                        int flags, int do_sync, bool nofail)
+{
+       struct ceph_osd_request *req;
+       int rc = 0;
+
+       BUG_ON(vino.snap != CEPH_NOSNAP);
+       req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
+                                   CEPH_OSD_OP_WRITE,
+                                   flags | CEPH_OSD_FLAG_ONDISK |
+                                           CEPH_OSD_FLAG_WRITE,
+                                   snapc, do_sync,
+                                   truncate_seq, truncate_size, mtime,
+                                   nofail, 1);
+       if (IS_ERR(req))
+               return PTR_ERR(req);
+
+       /* it may be a short write due to an object boundary */
+       req->r_pages = pages;
+       req->r_num_pages = calc_pages_for(off, len);
+       dout("writepages %llu~%llu (%d pages)\n", off, len,
+            req->r_num_pages);
+
+       rc = ceph_osdc_start_request(osdc, req, nofail);
+       if (!rc)
+               rc = ceph_osdc_wait_request(osdc, req);
+
+       ceph_osdc_put_request(req);
+       if (rc == 0)
+               rc = len;
+       dout("writepages result %d\n", rc);
+       return rc;
+}
+
+/*
+ * handle incoming message
+ */
+static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
+{
+       struct ceph_osd *osd = con->private;
+       struct ceph_osd_client *osdc = osd->o_osdc;
+       int type = le16_to_cpu(msg->hdr.type);
+
+       if (!osd)
+               return;
+
+       switch (type) {
+       case CEPH_MSG_OSD_MAP:
+               ceph_osdc_handle_map(osdc, msg);
+               break;
+       case CEPH_MSG_OSD_OPREPLY:
+               handle_reply(osdc, msg);
+               break;
+
+       default:
+               pr_err("received unknown message type %d %s\n", type,
+                      ceph_msg_type_name(type));
+       }
+       ceph_msg_put(msg);
+}
+
+static struct ceph_msg *alloc_msg(struct ceph_connection *con,
+                                 struct ceph_msg_header *hdr)
+{
+       struct ceph_osd *osd = con->private;
+       struct ceph_osd_client *osdc = osd->o_osdc;
+       int type = le16_to_cpu(hdr->type);
+
+       switch (type) {
+       case CEPH_MSG_OSD_OPREPLY:
+               return ceph_msgpool_get(&osdc->msgpool_op_reply);
+       }
+       return ceph_alloc_msg(con, hdr);
+}
+
+/*
+ * Wrappers to refcount containing ceph_osd struct
+ */
+static struct ceph_connection *get_osd_con(struct ceph_connection *con)
+{
+       struct ceph_osd *osd = con->private;
+       if (get_osd(osd))
+               return con;
+       return NULL;
+}
+
+static void put_osd_con(struct ceph_connection *con)
+{
+       struct ceph_osd *osd = con->private;
+       put_osd(osd);
+}
+
+const static struct ceph_connection_operations osd_con_ops = {
+       .get = get_osd_con,
+       .put = put_osd_con,
+       .dispatch = dispatch,
+       .alloc_msg = alloc_msg,
+       .peer_reset = osd_reset,
+       .alloc_middle = ceph_alloc_middle,
+       .prepare_pages = prepare_pages,
+};
diff --git a/fs/ceph/osd_client.h b/fs/ceph/osd_client.h
new file mode 100644 (file)
index 0000000..9a4addf
--- /dev/null
@@ -0,0 +1,144 @@
+#ifndef _FS_CEPH_OSD_CLIENT_H
+#define _FS_CEPH_OSD_CLIENT_H
+
+#include <linux/completion.h>
+#include <linux/mempool.h>
+#include <linux/rbtree.h>
+
+#include "types.h"
+#include "osdmap.h"
+#include "messenger.h"
+
+struct ceph_msg;
+struct ceph_snap_context;
+struct ceph_osd_request;
+struct ceph_osd_client;
+
+/*
+ * completion callback for async writepages
+ */
+typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *,
+                                    struct ceph_msg *);
+
+/* a given osd we're communicating with */
+struct ceph_osd {
+       atomic_t o_ref;
+       struct ceph_osd_client *o_osdc;
+       int o_osd;
+       int o_incarnation;
+       struct rb_node o_node;
+       struct ceph_connection o_con;
+       struct list_head o_requests;
+};
+
+/* an in-flight request */
+struct ceph_osd_request {
+       u64             r_tid;              /* unique for this client */
+       struct rb_node  r_node;
+       struct list_head r_osd_item;
+       struct ceph_osd *r_osd;
+
+       struct ceph_msg  *r_request, *r_reply;
+       int               r_result;
+       int               r_flags;     /* any additional flags for the osd */
+       u32               r_sent;      /* >0 if r_request is sending/sent */
+       int r_prepared_pages, r_got_reply;
+
+       struct ceph_osd_client *r_osdc;
+       atomic_t          r_ref;
+       bool              r_mempool;
+       struct completion r_completion, r_safe_completion;
+       ceph_osdc_callback_t r_callback, r_safe_callback;
+       struct ceph_eversion r_reassert_version;
+       struct list_head  r_unsafe_item;
+
+       struct inode *r_inode;                /* for use by callbacks */
+       struct writeback_control *r_wbc;      /* ditto */
+
+       char              r_oid[40];          /* object name */
+       int               r_oid_len;
+       unsigned long     r_timeout_stamp;
+       bool              r_resend;           /* msg send failed, needs retry */
+
+       struct ceph_file_layout r_file_layout;
+       struct ceph_snap_context *r_snapc;    /* snap context for writes */
+       unsigned          r_num_pages;        /* size of page array (follows) */
+       struct page     **r_pages;            /* pages for data payload */
+       int               r_pages_from_pool;
+       int               r_own_pages;        /* if true, i own page list */
+};
+
+struct ceph_osd_client {
+       struct ceph_client     *client;
+
+       struct ceph_osdmap     *osdmap;       /* current map */
+       struct rw_semaphore    map_sem;
+       struct completion      map_waiters;
+       u64                    last_requested_map;
+
+       struct mutex           request_mutex;
+       struct rb_root         osds;          /* osds */
+       u64                    timeout_tid;   /* tid of timeout triggering rq */
+       u64                    last_tid;      /* tid of last request */
+       struct rb_root         requests;      /* pending requests */
+       int                    num_requests;
+       struct delayed_work    timeout_work;
+       struct dentry          *debugfs_file;
+
+       mempool_t              *req_mempool;
+
+       struct ceph_msgpool   msgpool_op;
+       struct ceph_msgpool   msgpool_op_reply;
+};
+
+extern int ceph_osdc_init(struct ceph_osd_client *osdc,
+                         struct ceph_client *client);
+extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
+
+extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
+                                  struct ceph_msg *msg);
+extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
+                                struct ceph_msg *msg);
+
+extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
+                                     struct ceph_file_layout *layout,
+                                     struct ceph_vino vino,
+                                     u64 offset, u64 *len, int op, int flags,
+                                     struct ceph_snap_context *snapc,
+                                     int do_sync, u32 truncate_seq,
+                                     u64 truncate_size,
+                                     struct timespec *mtime,
+                                     bool use_mempool, int num_reply);
+
+static inline void ceph_osdc_get_request(struct ceph_osd_request *req)
+{
+       atomic_inc(&req->r_ref);
+}
+extern void ceph_osdc_put_request(struct ceph_osd_request *req);
+
+extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
+                                  struct ceph_osd_request *req,
+                                  bool nofail);
+extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
+                                 struct ceph_osd_request *req);
+extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
+
+extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
+                              struct ceph_vino vino,
+                              struct ceph_file_layout *layout,
+                              u64 off, u64 *plen,
+                              u32 truncate_seq, u64 truncate_size,
+                              struct page **pages, int nr_pages);
+
+extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
+                               struct ceph_vino vino,
+                               struct ceph_file_layout *layout,
+                               struct ceph_snap_context *sc,
+                               u64 off, u64 len,
+                               u32 truncate_seq, u64 truncate_size,
+                               struct timespec *mtime,
+                               struct page **pages, int nr_pages,
+                               int flags, int do_sync, bool nofail);
+
+#endif
+
diff --git a/fs/ceph/osdmap.c b/fs/ceph/osdmap.c
new file mode 100644 (file)
index 0000000..e38fe63
--- /dev/null
@@ -0,0 +1,875 @@
+
+#include <asm/div64.h>
+
+#include "super.h"
+#include "osdmap.h"
+#include "crush/hash.h"
+#include "crush/mapper.h"
+#include "decode.h"
+#include "ceph_debug.h"
+
+char *ceph_osdmap_state_str(char *str, int len, int state)
+{
+       int flag = 0;
+
+       if (!len)
+               goto done;
+
+       *str = '\0';
+       if (state) {
+               if (state & CEPH_OSD_EXISTS) {
+                       snprintf(str, len, "exists");
+                       flag = 1;
+               }
+               if (state & CEPH_OSD_UP) {
+                       snprintf(str, len, "%s%s%s", str, (flag ? ", " : ""),
+                                "up");
+                       flag = 1;
+               }
+       } else {
+               snprintf(str, len, "doesn't exist");
+       }
+done:
+       return str;
+}
+
+/* maps */
+
+static int calc_bits_of(unsigned t)
+{
+       int b = 0;
+       while (t) {
+               t = t >> 1;
+               b++;
+       }
+       return b;
+}
+
+/*
+ * the foo_mask is the smallest value 2^n-1 that is >= foo.
+ */
+static void calc_pg_masks(struct ceph_pg_pool_info *pi)
+{
+       pi->pg_num_mask = (1 << calc_bits_of(le32_to_cpu(pi->v.pg_num)-1)) - 1;
+       pi->pgp_num_mask =
+               (1 << calc_bits_of(le32_to_cpu(pi->v.pgp_num)-1)) - 1;
+       pi->lpg_num_mask =
+               (1 << calc_bits_of(le32_to_cpu(pi->v.lpg_num)-1)) - 1;
+       pi->lpgp_num_mask =
+               (1 << calc_bits_of(le32_to_cpu(pi->v.lpgp_num)-1)) - 1;
+}
+
+/*
+ * decode crush map
+ */
+static int crush_decode_uniform_bucket(void **p, void *end,
+                                      struct crush_bucket_uniform *b)
+{
+       dout("crush_decode_uniform_bucket %p to %p\n", *p, end);
+       ceph_decode_need(p, end, (1+b->h.size) * sizeof(u32), bad);
+       ceph_decode_32(p, b->item_weight);
+       return 0;
+bad:
+       return -EINVAL;
+}
+
+static int crush_decode_list_bucket(void **p, void *end,
+                                   struct crush_bucket_list *b)
+{
+       int j;
+       dout("crush_decode_list_bucket %p to %p\n", *p, end);
+       b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
+       if (b->item_weights == NULL)
+               return -ENOMEM;
+       b->sum_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
+       if (b->sum_weights == NULL)
+               return -ENOMEM;
+       ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad);
+       for (j = 0; j < b->h.size; j++) {
+               ceph_decode_32(p, b->item_weights[j]);
+               ceph_decode_32(p, b->sum_weights[j]);
+       }
+       return 0;
+bad:
+       return -EINVAL;
+}
+
+static int crush_decode_tree_bucket(void **p, void *end,
+                                   struct crush_bucket_tree *b)
+{
+       int j;
+       dout("crush_decode_tree_bucket %p to %p\n", *p, end);
+       ceph_decode_32_safe(p, end, b->num_nodes, bad);
+       b->node_weights = kcalloc(b->num_nodes, sizeof(u32), GFP_NOFS);
+       if (b->node_weights == NULL)
+               return -ENOMEM;
+       ceph_decode_need(p, end, b->num_nodes * sizeof(u32), bad);
+       for (j = 0; j < b->num_nodes; j++)
+               ceph_decode_32(p, b->node_weights[j]);
+       return 0;
+bad:
+       return -EINVAL;
+}
+
+static int crush_decode_straw_bucket(void **p, void *end,
+                                    struct crush_bucket_straw *b)
+{
+       int j;
+       dout("crush_decode_straw_bucket %p to %p\n", *p, end);
+       b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
+       if (b->item_weights == NULL)
+               return -ENOMEM;
+       b->straws = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
+       if (b->straws == NULL)
+               return -ENOMEM;
+       ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad);
+       for (j = 0; j < b->h.size; j++) {
+               ceph_decode_32(p, b->item_weights[j]);
+               ceph_decode_32(p, b->straws[j]);
+       }
+       return 0;
+bad:
+       return -EINVAL;
+}
+
+static struct crush_map *crush_decode(void *pbyval, void *end)
+{
+       struct crush_map *c;
+       int err = -EINVAL;
+       int i, j;
+       void **p = &pbyval;
+       void *start = pbyval;
+       u32 magic;
+
+       dout("crush_decode %p to %p len %d\n", *p, end, (int)(end - *p));
+
+       c = kzalloc(sizeof(*c), GFP_NOFS);
+       if (c == NULL)
+               return ERR_PTR(-ENOMEM);
+
+       ceph_decode_need(p, end, 4*sizeof(u32), bad);
+       ceph_decode_32(p, magic);
+       if (magic != CRUSH_MAGIC) {
+               pr_err("crush_decode magic %x != current %x\n",
+                      (unsigned)magic, (unsigned)CRUSH_MAGIC);
+               goto bad;
+       }
+       ceph_decode_32(p, c->max_buckets);
+       ceph_decode_32(p, c->max_rules);
+       ceph_decode_32(p, c->max_devices);
+
+       c->device_parents = kcalloc(c->max_devices, sizeof(u32), GFP_NOFS);
+       if (c->device_parents == NULL)
+               goto badmem;
+       c->bucket_parents = kcalloc(c->max_buckets, sizeof(u32), GFP_NOFS);
+       if (c->bucket_parents == NULL)
+               goto badmem;
+
+       c->buckets = kcalloc(c->max_buckets, sizeof(*c->buckets), GFP_NOFS);
+       if (c->buckets == NULL)
+               goto badmem;
+       c->rules = kcalloc(c->max_rules, sizeof(*c->rules), GFP_NOFS);
+       if (c->rules == NULL)
+               goto badmem;
+
+       /* buckets */
+       for (i = 0; i < c->max_buckets; i++) {
+               int size = 0;
+               u32 alg;
+               struct crush_bucket *b;
+
+               ceph_decode_32_safe(p, end, alg, bad);
+               if (alg == 0) {
+                       c->buckets[i] = NULL;
+                       continue;
+               }
+               dout("crush_decode bucket %d off %x %p to %p\n",
+                    i, (int)(*p-start), *p, end);
+
+               switch (alg) {
+               case CRUSH_BUCKET_UNIFORM:
+                       size = sizeof(struct crush_bucket_uniform);
+                       break;
+               case CRUSH_BUCKET_LIST:
+                       size = sizeof(struct crush_bucket_list);
+                       break;
+               case CRUSH_BUCKET_TREE:
+                       size = sizeof(struct crush_bucket_tree);
+                       break;
+               case CRUSH_BUCKET_STRAW:
+                       size = sizeof(struct crush_bucket_straw);
+                       break;
+               default:
+                       goto bad;
+               }
+               BUG_ON(size == 0);
+               b = c->buckets[i] = kzalloc(size, GFP_NOFS);
+               if (b == NULL)
+                       goto badmem;
+
+               ceph_decode_need(p, end, 4*sizeof(u32), bad);
+               ceph_decode_32(p, b->id);
+               ceph_decode_16(p, b->type);
+               ceph_decode_16(p, b->alg);
+               ceph_decode_32(p, b->weight);
+               ceph_decode_32(p, b->size);
+
+               dout("crush_decode bucket size %d off %x %p to %p\n",
+                    b->size, (int)(*p-start), *p, end);
+
+               b->items = kcalloc(b->size, sizeof(__s32), GFP_NOFS);
+               if (b->items == NULL)
+                       goto badmem;
+               b->perm = kcalloc(b->size, sizeof(u32), GFP_NOFS);
+               if (b->perm == NULL)
+                       goto badmem;
+               b->perm_n = 0;
+
+               ceph_decode_need(p, end, b->size*sizeof(u32), bad);
+               for (j = 0; j < b->size; j++)
+                       ceph_decode_32(p, b->items[j]);
+
+               switch (b->alg) {
+               case CRUSH_BUCKET_UNIFORM:
+                       err = crush_decode_uniform_bucket(p, end,
+                                 (struct crush_bucket_uniform *)b);
+                       if (err < 0)
+                               goto bad;
+                       break;
+               case CRUSH_BUCKET_LIST:
+                       err = crush_decode_list_bucket(p, end,
+                              (struct crush_bucket_list *)b);
+                       if (err < 0)
+                               goto bad;
+                       break;
+               case CRUSH_BUCKET_TREE:
+                       err = crush_decode_tree_bucket(p, end,
+                               (struct crush_bucket_tree *)b);
+                       if (err < 0)
+                               goto bad;
+                       break;
+               case CRUSH_BUCKET_STRAW:
+                       err = crush_decode_straw_bucket(p, end,
+                               (struct crush_bucket_straw *)b);
+                       if (err < 0)
+                               goto bad;
+                       break;
+               }
+       }
+
+       /* rules */
+       dout("rule vec is %p\n", c->rules);
+       for (i = 0; i < c->max_rules; i++) {
+               u32 yes;
+               struct crush_rule *r;
+
+               ceph_decode_32_safe(p, end, yes, bad);
+               if (!yes) {
+                       dout("crush_decode NO rule %d off %x %p to %p\n",
+                            i, (int)(*p-start), *p, end);
+                       c->rules[i] = NULL;
+                       continue;
+               }
+
+               dout("crush_decode rule %d off %x %p to %p\n",
+                    i, (int)(*p-start), *p, end);
+
+               /* len */
+               ceph_decode_32_safe(p, end, yes, bad);
+#if BITS_PER_LONG == 32
+               if (yes > ULONG_MAX / sizeof(struct crush_rule_step))
+                       goto bad;
+#endif
+               r = c->rules[i] = kmalloc(sizeof(*r) +
+                                         yes*sizeof(struct crush_rule_step),
+                                         GFP_NOFS);
+               if (r == NULL)
+                       goto badmem;
+               dout(" rule %d is at %p\n", i, r);
+               r->len = yes;
+               ceph_decode_copy_safe(p, end, &r->mask, 4, bad); /* 4 u8's */
+               ceph_decode_need(p, end, r->len*3*sizeof(u32), bad);
+               for (j = 0; j < r->len; j++) {
+                       ceph_decode_32(p, r->steps[j].op);
+                       ceph_decode_32(p, r->steps[j].arg1);
+                       ceph_decode_32(p, r->steps[j].arg2);
+               }
+       }
+
+       /* ignore trailing name maps. */
+
+       dout("crush_decode success\n");
+       return c;
+
+badmem:
+       err = -ENOMEM;
+bad:
+       dout("crush_decode fail %d\n", err);
+       crush_destroy(c);
+       return ERR_PTR(err);
+}
+
+
+/*
+ * osd map
+ */
+void ceph_osdmap_destroy(struct ceph_osdmap *map)
+{
+       dout("osdmap_destroy %p\n", map);
+       if (map->crush)
+               crush_destroy(map->crush);
+       while (!RB_EMPTY_ROOT(&map->pg_temp))
+               rb_erase(rb_first(&map->pg_temp), &map->pg_temp);
+       kfree(map->osd_state);
+       kfree(map->osd_weight);
+       kfree(map->pg_pool);
+       kfree(map->osd_addr);
+       kfree(map);
+}
+
+/*
+ * adjust max osd value.  reallocate arrays.
+ */
+static int osdmap_set_max_osd(struct ceph_osdmap *map, int max)
+{
+       u8 *state;
+       struct ceph_entity_addr *addr;
+       u32 *weight;
+
+       state = kcalloc(max, sizeof(*state), GFP_NOFS);
+       addr = kcalloc(max, sizeof(*addr), GFP_NOFS);
+       weight = kcalloc(max, sizeof(*weight), GFP_NOFS);
+       if (state == NULL || addr == NULL || weight == NULL) {
+               kfree(state);
+               kfree(addr);
+               kfree(weight);
+               return -ENOMEM;
+       }
+
+       /* copy old? */
+       if (map->osd_state) {
+               memcpy(state, map->osd_state, map->max_osd*sizeof(*state));
+               memcpy(addr, map->osd_addr, map->max_osd*sizeof(*addr));
+               memcpy(weight, map->osd_weight, map->max_osd*sizeof(*weight));
+               kfree(map->osd_state);
+               kfree(map->osd_addr);
+               kfree(map->osd_weight);
+       }
+
+       map->osd_state = state;
+       map->osd_weight = weight;
+       map->osd_addr = addr;
+       map->max_osd = max;
+       return 0;
+}
+
+/*
+ * Insert a new pg_temp mapping
+ */
+static void __insert_pg_mapping(struct ceph_pg_mapping *new,
+                               struct rb_root *root)
+{
+       struct rb_node **p = &root->rb_node;
+       struct rb_node *parent = NULL;
+       struct ceph_pg_mapping *pg = NULL;
+
+       while (*p) {
+               parent = *p;
+               pg = rb_entry(parent, struct ceph_pg_mapping, node);
+               if (new->pgid < pg->pgid)
+                       p = &(*p)->rb_left;
+               else if (new->pgid > pg->pgid)
+                       p = &(*p)->rb_right;
+               else
+                       BUG();
+       }
+
+       rb_link_node(&new->node, parent, p);
+       rb_insert_color(&new->node, root);
+}
+
+/*
+ * decode a full map.
+ */
+struct ceph_osdmap *osdmap_decode(void **p, void *end)
+{
+       struct ceph_osdmap *map;
+       u16 version;
+       u32 len, max, i;
+       int err = -EINVAL;
+       void *start = *p;
+
+       dout("osdmap_decode %p to %p len %d\n", *p, end, (int)(end - *p));
+
+       map = kzalloc(sizeof(*map), GFP_NOFS);
+       if (map == NULL)
+               return ERR_PTR(-ENOMEM);
+       map->pg_temp = RB_ROOT;
+
+       ceph_decode_16_safe(p, end, version, bad);
+
+       ceph_decode_need(p, end, 2*sizeof(u64)+6*sizeof(u32), bad);
+       ceph_decode_copy(p, &map->fsid, sizeof(map->fsid));
+       ceph_decode_32(p, map->epoch);
+       ceph_decode_copy(p, &map->created, sizeof(map->created));
+       ceph_decode_copy(p, &map->modified, sizeof(map->modified));
+
+       ceph_decode_32(p, map->num_pools);
+       map->pg_pool = kcalloc(map->num_pools, sizeof(*map->pg_pool),
+                              GFP_NOFS);
+       if (!map->pg_pool) {
+               err = -ENOMEM;
+               goto bad;
+       }
+       ceph_decode_32_safe(p, end, max, bad);
+       while (max--) {
+               ceph_decode_need(p, end, 4+sizeof(map->pg_pool->v), bad);
+               ceph_decode_32(p, i);
+               if (i >= map->num_pools)
+                       goto bad;
+               ceph_decode_copy(p, &map->pg_pool[i].v,
+                                sizeof(map->pg_pool->v));
+               calc_pg_masks(&map->pg_pool[i]);
+               p += le32_to_cpu(map->pg_pool[i].v.num_snaps) * sizeof(u64);
+               p += le32_to_cpu(map->pg_pool[i].v.num_removed_snap_intervals)
+                       * sizeof(u64) * 2;
+       }
+
+       ceph_decode_32_safe(p, end, map->flags, bad);
+
+       ceph_decode_32(p, max);
+
+       /* (re)alloc osd arrays */
+       err = osdmap_set_max_osd(map, max);
+       if (err < 0)
+               goto bad;
+       dout("osdmap_decode max_osd = %d\n", map->max_osd);
+
+       /* osds */
+       err = -EINVAL;
+       ceph_decode_need(p, end, 3*sizeof(u32) +
+                        map->max_osd*(1 + sizeof(*map->osd_weight) +
+                                      sizeof(*map->osd_addr)), bad);
+       *p += 4; /* skip length field (should match max) */
+       ceph_decode_copy(p, map->osd_state, map->max_osd);
+
+       *p += 4; /* skip length field (should match max) */
+       for (i = 0; i < map->max_osd; i++)
+               ceph_decode_32(p, map->osd_weight[i]);
+
+       *p += 4; /* skip length field (should match max) */
+       ceph_decode_copy(p, map->osd_addr, map->max_osd*sizeof(*map->osd_addr));
+
+       /* pg_temp */
+       ceph_decode_32_safe(p, end, len, bad);
+       for (i = 0; i < len; i++) {
+               int n, j;
+               u64 pgid;
+               struct ceph_pg_mapping *pg;
+
+               ceph_decode_need(p, end, sizeof(u32) + sizeof(u64), bad);
+               ceph_decode_64(p, pgid);
+               ceph_decode_32(p, n);
+               ceph_decode_need(p, end, n * sizeof(u32), bad);
+               pg = kmalloc(sizeof(*pg) + n*sizeof(u32), GFP_NOFS);
+               if (!pg) {
+                       err = -ENOMEM;
+                       goto bad;
+               }
+               pg->pgid = pgid;
+               pg->len = n;
+               for (j = 0; j < n; j++)
+                       ceph_decode_32(p, pg->osds[j]);
+
+               __insert_pg_mapping(pg, &map->pg_temp);
+               dout(" added pg_temp %llx len %d\n", pgid, len);
+       }
+
+       /* crush */
+       ceph_decode_32_safe(p, end, len, bad);
+       dout("osdmap_decode crush len %d from off 0x%x\n", len,
+            (int)(*p - start));
+       ceph_decode_need(p, end, len, bad);
+       map->crush = crush_decode(*p, end);
+       *p += len;
+       if (IS_ERR(map->crush)) {
+               err = PTR_ERR(map->crush);
+               map->crush = NULL;
+               goto bad;
+       }
+
+       /* ignore the rest of the map */
+       *p = end;
+
+       dout("osdmap_decode done %p %p\n", *p, end);
+       return map;
+
+bad:
+       dout("osdmap_decode fail\n");
+       ceph_osdmap_destroy(map);
+       return ERR_PTR(err);
+}
+
+/*
+ * decode and apply an incremental map update.
+ */
+struct ceph_osdmap *osdmap_apply_incremental(void **p, void *end,
+                                            struct ceph_osdmap *map,
+                                            struct ceph_messenger *msgr)
+{
+       struct ceph_osdmap *newmap = map;
+       struct crush_map *newcrush = NULL;
+       struct ceph_fsid fsid;
+       u32 epoch = 0;
+       struct ceph_timespec modified;
+       u32 len, pool;
+       __s32 new_flags, max;
+       void *start = *p;
+       int err = -EINVAL;
+       u16 version;
+       struct rb_node *rbp;
+
+       ceph_decode_16_safe(p, end, version, bad);
+
+       ceph_decode_need(p, end, sizeof(fsid)+sizeof(modified)+2*sizeof(u32),
+                        bad);
+       ceph_decode_copy(p, &fsid, sizeof(fsid));
+       ceph_decode_32(p, epoch);
+       BUG_ON(epoch != map->epoch+1);
+       ceph_decode_copy(p, &modified, sizeof(modified));
+       ceph_decode_32(p, new_flags);
+
+       /* full map? */
+       ceph_decode_32_safe(p, end, len, bad);
+       if (len > 0) {
+               dout("apply_incremental full map len %d, %p to %p\n",
+                    len, *p, end);
+               newmap = osdmap_decode(p, min(*p+len, end));
+               return newmap;  /* error or not */
+       }
+
+       /* new crush? */
+       ceph_decode_32_safe(p, end, len, bad);
+       if (len > 0) {
+               dout("apply_incremental new crush map len %d, %p to %p\n",
+                    len, *p, end);
+               newcrush = crush_decode(*p, min(*p+len, end));
+               if (IS_ERR(newcrush))
+                       return ERR_PTR(PTR_ERR(newcrush));
+       }
+
+       /* new flags? */
+       if (new_flags >= 0)
+               map->flags = new_flags;
+
+       ceph_decode_need(p, end, 5*sizeof(u32), bad);
+
+       /* new max? */
+       ceph_decode_32(p, max);
+       if (max >= 0) {
+               err = osdmap_set_max_osd(map, max);
+               if (err < 0)
+                       goto bad;
+       }
+
+       map->epoch++;
+       map->modified = map->modified;
+       if (newcrush) {
+               if (map->crush)
+                       crush_destroy(map->crush);
+               map->crush = newcrush;
+               newcrush = NULL;
+       }
+
+       /* new_pool */
+       ceph_decode_32_safe(p, end, len, bad);
+       while (len--) {
+               ceph_decode_32_safe(p, end, pool, bad);
+               if (pool >= map->num_pools) {
+                       void *pg_pool = kcalloc(pool + 1,
+                                               sizeof(*map->pg_pool),
+                                               GFP_NOFS);
+                       if (!pg_pool) {
+                               err = -ENOMEM;
+                               goto bad;
+                       }
+                       memcpy(pg_pool, map->pg_pool,
+                              map->num_pools * sizeof(*map->pg_pool));
+                       kfree(map->pg_pool);
+                       map->pg_pool = pg_pool;
+                       map->num_pools = pool+1;
+               }
+               ceph_decode_copy(p, &map->pg_pool[pool].v,
+                                sizeof(map->pg_pool->v));
+               calc_pg_masks(&map->pg_pool[pool]);
+       }
+
+       /* old_pool (ignore) */
+       ceph_decode_32_safe(p, end, len, bad);
+       *p += len * sizeof(u32);
+
+       /* new_up */
+       err = -EINVAL;
+       ceph_decode_32_safe(p, end, len, bad);
+       while (len--) {
+               u32 osd;
+               struct ceph_entity_addr addr;
+               ceph_decode_32_safe(p, end, osd, bad);
+               ceph_decode_copy_safe(p, end, &addr, sizeof(addr), bad);
+               pr_info("osd%d up\n", osd);
+               BUG_ON(osd >= map->max_osd);
+               map->osd_state[osd] |= CEPH_OSD_UP;
+               map->osd_addr[osd] = addr;
+       }
+
+       /* new_down */
+       ceph_decode_32_safe(p, end, len, bad);
+       while (len--) {
+               u32 osd;
+               ceph_decode_32_safe(p, end, osd, bad);
+               (*p)++;  /* clean flag */
+               pr_info("ceph osd%d down\n", osd);
+               if (osd < map->max_osd)
+                       map->osd_state[osd] &= ~CEPH_OSD_UP;
+       }
+
+       /* new_weight */
+       ceph_decode_32_safe(p, end, len, bad);
+       while (len--) {
+               u32 osd, off;
+               ceph_decode_need(p, end, sizeof(u32)*2, bad);
+               ceph_decode_32(p, osd);
+               ceph_decode_32(p, off);
+               pr_info("osd%d weight 0x%x %s\n", osd, off,
+                    off == CEPH_OSD_IN ? "(in)" :
+                    (off == CEPH_OSD_OUT ? "(out)" : ""));
+               if (osd < map->max_osd)
+                       map->osd_weight[osd] = off;
+       }
+
+       /* new_pg_temp */
+       rbp = rb_first(&map->pg_temp);
+       ceph_decode_32_safe(p, end, len, bad);
+       while (len--) {
+               struct ceph_pg_mapping *pg;
+               int j;
+               u64 pgid;
+               u32 pglen;
+               ceph_decode_need(p, end, sizeof(u64) + sizeof(u32), bad);
+               ceph_decode_64(p, pgid);
+               ceph_decode_32(p, pglen);
+
+               /* remove any? */
+               while (rbp && rb_entry(rbp, struct ceph_pg_mapping,
+                                      node)->pgid <= pgid) {
+                       struct rb_node *cur = rbp;
+                       rbp = rb_next(rbp);
+                       dout(" removed pg_temp %llx\n",
+                            rb_entry(cur, struct ceph_pg_mapping, node)->pgid);
+                       rb_erase(cur, &map->pg_temp);
+               }
+
+               if (pglen) {
+                       /* insert */
+                       ceph_decode_need(p, end, pglen*sizeof(u32), bad);
+                       pg = kmalloc(sizeof(*pg) + sizeof(u32)*pglen, GFP_NOFS);
+                       if (!pg) {
+                               err = -ENOMEM;
+                               goto bad;
+                       }
+                       pg->pgid = pgid;
+                       pg->len = pglen;
+                       for (j = 0; j < len; j++)
+                               ceph_decode_32(p, pg->osds[j]);
+                       __insert_pg_mapping(pg, &map->pg_temp);
+                       dout(" added pg_temp %llx len %d\n", pgid, pglen);
+               }
+       }
+       while (rbp) {
+               struct rb_node *cur = rbp;
+               rbp = rb_next(rbp);
+               dout(" removed pg_temp %llx\n",
+                    rb_entry(cur, struct ceph_pg_mapping, node)->pgid);
+               rb_erase(cur, &map->pg_temp);
+       }
+
+       /* ignore the rest */
+       *p = end;
+       return map;
+
+bad:
+       pr_err("corrupt inc osdmap epoch %d off %d (%p of %p-%p)\n",
+              epoch, (int)(*p - start), *p, start, end);
+       if (newcrush)
+               crush_destroy(newcrush);
+       return ERR_PTR(err);
+}
+
+
+
+
+/*
+ * calculate file layout from given offset, length.
+ * fill in correct oid, logical length, and object extent
+ * offset, length.
+ *
+ * for now, we write only a single su, until we can
+ * pass a stride back to the caller.
+ */
+void ceph_calc_file_object_mapping(struct ceph_file_layout *layout,
+                                  u64 off, u64 *plen,
+                                  u64 *bno,
+                                  u64 *oxoff, u64 *oxlen)
+{
+       u32 osize = le32_to_cpu(layout->fl_object_size);
+       u32 su = le32_to_cpu(layout->fl_stripe_unit);
+       u32 sc = le32_to_cpu(layout->fl_stripe_count);
+       u32 bl, stripeno, stripepos, objsetno;
+       u32 su_per_object;
+       u64 t;
+
+       dout("mapping %llu~%llu  osize %u fl_su %u\n", off, *plen,
+            osize, su);
+       su_per_object = osize / le32_to_cpu(layout->fl_stripe_unit);
+       dout("osize %u / su %u = su_per_object %u\n", osize, su,
+            su_per_object);
+
+       BUG_ON((su & ~PAGE_MASK) != 0);
+       /* bl = *off / su; */
+       t = off;
+       do_div(t, su);
+       bl = t;
+       dout("off %llu / su %u = bl %u\n", off, su, bl);
+
+       stripeno = bl / sc;
+       stripepos = bl % sc;
+       objsetno = stripeno / su_per_object;
+
+       *bno = objsetno * sc + stripepos;
+       dout("objset %u * sc %u = bno %u\n", objsetno, sc, (unsigned)*bno);
+       /* *oxoff = *off / layout->fl_stripe_unit; */
+       t = off;
+       *oxoff = do_div(t, su);
+       *oxlen = min_t(u64, *plen, su - *oxoff);
+       *plen = *oxlen;
+
+       dout(" obj extent %llu~%llu\n", *oxoff, *oxlen);
+}
+
+/*
+ * calculate an object layout (i.e. pgid) from an oid,
+ * file_layout, and osdmap
+ */
+int ceph_calc_object_layout(struct ceph_object_layout *ol,
+                           const char *oid,
+                           struct ceph_file_layout *fl,
+                           struct ceph_osdmap *osdmap)
+{
+       unsigned num, num_mask;
+       union ceph_pg pgid;
+       s32 preferred = (s32)le32_to_cpu(fl->fl_pg_preferred);
+       int poolid = le32_to_cpu(fl->fl_pg_pool);
+       struct ceph_pg_pool_info *pool;
+
+       if (poolid >= osdmap->num_pools)
+               return -EIO;
+       pool = &osdmap->pg_pool[poolid];
+
+       if (preferred >= 0) {
+               num = le32_to_cpu(pool->v.lpg_num);
+               num_mask = pool->lpg_num_mask;
+       } else {
+               num = le32_to_cpu(pool->v.pg_num);
+               num_mask = pool->pg_num_mask;
+       }
+
+       pgid.pg64 = 0;   /* start with it zeroed out */
+       pgid.pg.ps = ceph_full_name_hash(oid, strlen(oid));
+       pgid.pg.preferred = preferred;
+       pgid.pg.pool = le32_to_cpu(fl->fl_pg_pool);
+       if (preferred >= 0)
+               dout("calc_object_layout '%s' pgid %d.%xp%d (%llx)\n", oid,
+                    pgid.pg.pool, pgid.pg.ps, (int)preferred, pgid.pg64);
+       else
+               dout("calc_object_layout '%s' pgid %d.%x (%llx)\n", oid,
+                    pgid.pg.pool, pgid.pg.ps, pgid.pg64);
+
+       ol->ol_pgid = cpu_to_le64(pgid.pg64);
+       ol->ol_stripe_unit = fl->fl_object_stripe_unit;
+
+       return 0;
+}
+
+/*
+ * Calculate raw osd vector for the given pgid.  Return pointer to osd
+ * array, or NULL on failure.
+ */
+static int *calc_pg_raw(struct ceph_osdmap *osdmap, union ceph_pg pgid,
+                       int *osds, int *num)
+{
+       struct rb_node *n = osdmap->pg_temp.rb_node;
+       struct ceph_pg_mapping *pg;
+       struct ceph_pg_pool_info *pool;
+       int ruleno;
+       unsigned pps; /* placement ps */
+
+       /* pg_temp? */
+       while (n) {
+               pg = rb_entry(n, struct ceph_pg_mapping, node);
+               if (pgid.pg64 < pg->pgid)
+                       n = n->rb_left;
+               else if (pgid.pg64 > pg->pgid)
+                       n = n->rb_right;
+               else {
+                       *num = pg->len;
+                       return pg->osds;
+               }
+       }
+
+       /* crush */
+       if (pgid.pg.pool >= osdmap->num_pools)
+               return NULL;
+       pool = &osdmap->pg_pool[pgid.pg.pool];
+       ruleno = crush_find_rule(osdmap->crush, pool->v.crush_ruleset,
+                                pool->v.type, pool->v.size);
+       if (ruleno < 0) {
+               pr_err("no crush rule pool %d type %d size %d\n",
+                      pgid.pg.pool, pool->v.type, pool->v.size);
+               return NULL;
+       }
+
+       if (pgid.pg.preferred >= 0)
+               pps = ceph_stable_mod(pgid.pg.ps,
+                                     le32_to_cpu(pool->v.lpgp_num),
+                                     pool->lpgp_num_mask);
+       else
+               pps = ceph_stable_mod(pgid.pg.ps,
+                                     le32_to_cpu(pool->v.pgp_num),
+                                     pool->pgp_num_mask);
+       pps += pgid.pg.pool;
+       *num = crush_do_rule(osdmap->crush, ruleno, pps, osds,
+                            min_t(int, pool->v.size, *num),
+                            pgid.pg.preferred, osdmap->osd_weight);
+       return osds;
+}
+
+/*
+ * Return primary osd for given pgid, or -1 if none.
+ */
+int ceph_calc_pg_primary(struct ceph_osdmap *osdmap, union ceph_pg pgid)
+{
+       int rawosds[10], *osds;
+       int i, num = ARRAY_SIZE(rawosds);
+
+       osds = calc_pg_raw(osdmap, pgid, rawosds, &num);
+       if (!osds)
+               return -1;
+
+       /* primary is first up osd */
+       for (i = 0; i < num; i++)
+               if (ceph_osd_is_up(osdmap, osds[i])) {
+                       return osds[i];
+                       break;
+               }
+       return -1;
+}
diff --git a/fs/ceph/osdmap.h b/fs/ceph/osdmap.h
new file mode 100644 (file)
index 0000000..07127c6
--- /dev/null
@@ -0,0 +1,123 @@
+#ifndef _FS_CEPH_OSDMAP_H
+#define _FS_CEPH_OSDMAP_H
+
+#include <linux/rbtree.h>
+#include "types.h"
+#include "ceph_fs.h"
+#include "crush/crush.h"
+
+/*
+ * The osd map describes the current membership of the osd cluster and
+ * specifies the mapping of objects to placement groups and placement
+ * groups to (sets of) osds.  That is, it completely specifies the
+ * (desired) distribution of all data objects in the system at some
+ * point in time.
+ *
+ * Each map version is identified by an epoch, which increases monotonically.
+ *
+ * The map can be updated either via an incremental map (diff) describing
+ * the change between two successive epochs, or as a fully encoded map.
+ */
+struct ceph_pg_pool_info {
+       struct ceph_pg_pool v;
+       int pg_num_mask, pgp_num_mask, lpg_num_mask, lpgp_num_mask;
+};
+
+struct ceph_pg_mapping {
+       struct rb_node node;
+       u64 pgid;
+       int len;
+       int osds[];
+};
+
+struct ceph_osdmap {
+       struct ceph_fsid fsid;
+       u32 epoch;
+       u32 mkfs_epoch;
+       struct ceph_timespec created, modified;
+
+       u32 flags;         /* CEPH_OSDMAP_* */
+
+       u32 max_osd;       /* size of osd_state, _offload, _addr arrays */
+       u8 *osd_state;     /* CEPH_OSD_* */
+       u32 *osd_weight;   /* 0 = failed, 0x10000 = 100% normal */
+       struct ceph_entity_addr *osd_addr;
+
+       struct rb_root pg_temp;
+
+       u32 num_pools;
+       struct ceph_pg_pool_info *pg_pool;
+
+       /* the CRUSH map specifies the mapping of placement groups to
+        * the list of osds that store+replicate them. */
+       struct crush_map *crush;
+};
+
+/*
+ * file layout helpers
+ */
+#define ceph_file_layout_su(l) ((__s32)le32_to_cpu((l).fl_stripe_unit))
+#define ceph_file_layout_stripe_count(l) \
+       ((__s32)le32_to_cpu((l).fl_stripe_count))
+#define ceph_file_layout_object_size(l) ((__s32)le32_to_cpu((l).fl_object_size))
+#define ceph_file_layout_cas_hash(l) ((__s32)le32_to_cpu((l).fl_cas_hash))
+#define ceph_file_layout_object_su(l) \
+       ((__s32)le32_to_cpu((l).fl_object_stripe_unit))
+#define ceph_file_layout_pg_preferred(l) \
+       ((__s32)le32_to_cpu((l).fl_pg_preferred))
+#define ceph_file_layout_pg_pool(l) \
+       ((__s32)le32_to_cpu((l).fl_pg_pool))
+
+static inline unsigned ceph_file_layout_stripe_width(struct ceph_file_layout *l)
+{
+       return le32_to_cpu(l->fl_stripe_unit) *
+               le32_to_cpu(l->fl_stripe_count);
+}
+
+/* "period" == bytes before i start on a new set of objects */
+static inline unsigned ceph_file_layout_period(struct ceph_file_layout *l)
+{
+       return le32_to_cpu(l->fl_object_size) *
+               le32_to_cpu(l->fl_stripe_count);
+}
+
+
+static inline int ceph_osd_is_up(struct ceph_osdmap *map, int osd)
+{
+       return (osd < map->max_osd) && (map->osd_state[osd] & CEPH_OSD_UP);
+}
+
+static inline bool ceph_osdmap_flag(struct ceph_osdmap *map, int flag)
+{
+       return map && (map->flags & flag);
+}
+
+extern char *ceph_osdmap_state_str(char *str, int len, int state);
+
+static inline struct ceph_entity_addr *ceph_osd_addr(struct ceph_osdmap *map,
+                                                    int osd)
+{
+       if (osd >= map->max_osd)
+               return NULL;
+       return &map->osd_addr[osd];
+}
+
+extern struct ceph_osdmap *osdmap_decode(void **p, void *end);
+extern struct ceph_osdmap *osdmap_apply_incremental(void **p, void *end,
+                                           struct ceph_osdmap *map,
+                                           struct ceph_messenger *msgr);
+extern void ceph_osdmap_destroy(struct ceph_osdmap *map);
+
+/* calculate mapping of a file extent to an object */
+extern void ceph_calc_file_object_mapping(struct ceph_file_layout *layout,
+                                         u64 off, u64 *plen,
+                                         u64 *bno, u64 *oxoff, u64 *oxlen);
+
+/* calculate mapping of object to a placement group */
+extern int ceph_calc_object_layout(struct ceph_object_layout *ol,
+                                  const char *oid,
+                                  struct ceph_file_layout *fl,
+                                  struct ceph_osdmap *osdmap);
+extern int ceph_calc_pg_primary(struct ceph_osdmap *osdmap, union ceph_pg pgid);
+
+#endif