__u64 pb_slv;
/* VBR: pre-versions */
__u64 pb_pre_versions[PTLRPC_NUM_VERSIONS];
+ __u64 pb_mbits; /**< match bits for bulk request */
/* padding for future needs */
- __u64 pb_padding[4];
+ __u64 pb_padding64_0;
+ __u64 pb_padding64_1;
+ __u64 pb_padding64_2;
char pb_jobid[LUSTRE_JOBID_SIZE];
};
__u64 pb_slv;
/* VBR: pre-versions */
__u64 pb_pre_versions[PTLRPC_NUM_VERSIONS];
+ __u64 pb_mbits; /**< unused in V2 */
/* padding for future needs */
- __u64 pb_padding[4];
+ __u64 pb_padding64_0;
+ __u64 pb_padding64_1;
+ __u64 pb_padding64_2;
};
void lustre_swab_ptlrpc_body(struct ptlrpc_body *pb);
* RPCs in parallel
*/
#define OBD_CONNECT_DIR_STRIPE 0x400000000000000ULL/* striped DNE dir */
+/** bulk matchbits is sent within ptlrpc_body */
+#define OBD_CONNECT_BULK_MBITS 0x2000000000000000ULL
/* XXX README XXX:
* Please DO NOT add flag values here before first ensuring that this same
__u64 rq_transno;
/** xid */
__u64 rq_xid;
+ /** bulk match bits */
+ u64 rq_mbits;
/**
* List item to for replay list. Not yet committed requests get linked
* there.
int bd_nob; /* # bytes covered */
int bd_nob_transferred; /* # bytes GOT/PUT */
- __u64 bd_last_xid;
+ u64 bd_last_mbits;
struct ptlrpc_cb_id bd_cbid; /* network callback info */
lnet_nid_t bd_sender; /* stash event::sender */
void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
+void lustre_msg_set_mbits(struct lustre_msg *msg, u64 mbits);
static inline void
lustre_shrink_reply(struct ptlrpc_request *req, int segment,
OBD_CONNECT_FLOCK_DEAD |
OBD_CONNECT_DISP_STRIPE | OBD_CONNECT_LFSCK |
OBD_CONNECT_OPEN_BY_FID |
- OBD_CONNECT_DIR_STRIPE;
+ OBD_CONNECT_DIR_STRIPE |
+ OBD_CONNECT_BULK_MBITS;
if (sbi->ll_flags & LL_SBI_LRU_RESIZE)
data->ocd_connect_flags |= OBD_CONNECT_LRU_RESIZE;
OBD_CONNECT_64BITHASH | OBD_CONNECT_MAXBYTES |
OBD_CONNECT_EINPROGRESS |
OBD_CONNECT_JOBSTATS | OBD_CONNECT_LVB_TYPE |
- OBD_CONNECT_LAYOUTLOCK | OBD_CONNECT_PINGLESS;
+ OBD_CONNECT_LAYOUTLOCK |
+ OBD_CONNECT_PINGLESS | OBD_CONNECT_LFSCK |
+ OBD_CONNECT_BULK_MBITS;
if (!OBD_FAIL_CHECK(OBD_FAIL_OSC_CONNECT_CKSUM)) {
/* OBD_CONNECT_CKSUM should always be set, even if checksums are
"unlink_close",
"multi_mod_rpcs",
"dir_stripe",
+ "bulk_mbits",
"unknown",
NULL
};
/* We connect to the MGS at setup, and don't disconnect until cleanup */
data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_AT |
OBD_CONNECT_FULL20 | OBD_CONNECT_IMP_RECOV |
- OBD_CONNECT_LVB_TYPE;
+ OBD_CONNECT_LVB_TYPE | OBD_CONNECT_BULK_MBITS;
#if OBD_OCD_VERSION(3, 0, 53, 0) > LUSTRE_VERSION_CODE
data->ocd_connect_flags |= OBD_CONNECT_MNE_SWAB;
spin_unlock(&req->rq_lock);
req->rq_nr_resend++;
- /* allocate new xid to avoid reply reconstruction */
- if (!req->rq_bulk) {
- /* new xid is already allocated for bulk in ptlrpc_check_set() */
- req->rq_xid = ptlrpc_next_xid();
- DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for resend on EINPROGRESS");
- }
-
/* Readjust the timeout for current conditions */
ptlrpc_at_set_req_timeout(req);
/*
spin_lock(&req->rq_lock);
req->rq_resend = 1;
spin_unlock(&req->rq_lock);
- if (req->rq_bulk) {
- __u64 old_xid;
-
- if (!ptlrpc_unregister_bulk(req, 1))
- continue;
-
- /* ensure previous bulk fails */
- old_xid = req->rq_xid;
- req->rq_xid = ptlrpc_next_xid();
- CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
- old_xid, req->rq_xid);
- }
+ if (req->rq_bulk &&
+ !ptlrpc_unregister_bulk(req, 1))
+ continue;
}
/*
* rq_wait_ctx is only touched by ptlrpcd,
req->rq_resend = 1;
req->rq_net_err = 0;
req->rq_timedout = 0;
- if (req->rq_bulk) {
- __u64 old_xid = req->rq_xid;
-
- /* ensure previous bulk fails */
- req->rq_xid = ptlrpc_next_xid();
- CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
- old_xid, req->rq_xid);
- }
ptlrpc_client_wake_req(req);
spin_unlock(&req->rq_lock);
}
return next;
}
+/**
+ * If request has a new allocated XID (new request or EINPROGRESS resend),
+ * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
+ * request to ensure previous bulk fails and avoid problems with lost replies
+ * and therefore several transfers landing into the same buffer from different
+ * sending attempts.
+ */
+void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
+{
+ struct ptlrpc_bulk_desc *bd = req->rq_bulk;
+
+ LASSERT(bd);
+
+ if (!req->rq_resend || req->rq_nr_resend) {
+ /* this request has a new xid, just use it as bulk matchbits */
+ req->rq_mbits = req->rq_xid;
+
+ } else { /* needs to generate a new matchbits for resend */
+ u64 old_mbits = req->rq_mbits;
+
+ if ((bd->bd_import->imp_connect_data.ocd_connect_flags &
+ OBD_CONNECT_BULK_MBITS)) {
+ req->rq_mbits = ptlrpc_next_xid();
+ } else {
+ /* old version transfers rq_xid to peer as matchbits */
+ req->rq_mbits = ptlrpc_next_xid();
+ req->rq_xid = req->rq_mbits;
+ }
+
+ CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
+ old_mbits, req->rq_mbits);
+ }
+
+ /*
+ * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
+ * that server can infer the number of bulks that were prepared,
+ * see LU-1431
+ */
+ req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
+ LNET_MAX_IOV) - 1;
+}
+
/**
* Get a glimpse at what next xid value might have been.
* Returns possible next xid.
int rc2;
int posted_md;
int total_md;
- __u64 xid;
+ u64 mbits;
lnet_handle_me_t me_h;
lnet_md_t md;
LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
LASSERT(desc->bd_cbid.cbid_arg == desc);
- /* An XID is only used for a single request from the client.
- * For retried bulk transfers, a new XID will be allocated in
- * in ptlrpc_check_set() if it needs to be resent, so it is not
- * using the same RDMA match bits after an error.
- *
- * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
- * first bulk XID is power-of-two aligned before rq_xid. LU-1431
- */
- xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
+ total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
+ /* rq_mbits is matchbits of the final bulk */
+ mbits = req->rq_mbits - total_md + 1;
+
+ LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
+ "first mbits = x%llu, last mbits = x%llu\n",
+ mbits, req->rq_mbits);
LASSERTF(!(desc->bd_registered &&
req->rq_send_state != LUSTRE_IMP_REPLAY) ||
- xid != desc->bd_last_xid,
- "registered: %d rq_xid: %llu bd_last_xid: %llu\n",
- desc->bd_registered, xid, desc->bd_last_xid);
+ mbits != desc->bd_last_mbits,
+ "registered: %d rq_mbits: %llu bd_last_mbits: %llu\n",
+ desc->bd_registered, mbits, desc->bd_last_mbits);
- total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
desc->bd_registered = 1;
- desc->bd_last_xid = xid;
+ desc->bd_last_mbits = mbits;
desc->bd_md_count = total_md;
md.user_ptr = &desc->bd_cbid;
md.eq_handle = ptlrpc_eq_h;
md.threshold = 1; /* PUT or GET */
- for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
+ for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
md.options = PTLRPC_MD_OPTIONS |
(ptlrpc_is_bulk_op_get(desc->bd_type) ?
LNET_MD_OP_GET : LNET_MD_OP_PUT);
ptlrpc_fill_bulk_md(&md, desc, posted_md);
- rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
+ rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
LNET_UNLINK, LNET_INS_AFTER, &me_h);
if (rc != 0) {
CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
- desc->bd_import->imp_obd->obd_name, xid,
+ desc->bd_import->imp_obd->obd_name, mbits,
posted_md, rc);
break;
}
&desc->bd_mds[posted_md]);
if (rc != 0) {
CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
- desc->bd_import->imp_obd->obd_name, xid,
+ desc->bd_import->imp_obd->obd_name, mbits,
posted_md, rc);
rc2 = LNetMEUnlink(me_h);
LASSERT(rc2 == 0);
return -ENOMEM;
}
- /* Set rq_xid to matchbits of the final bulk so that server can
- * infer the number of bulks that were prepared
- */
- req->rq_xid = --xid;
- LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
- "bd_last_xid = x%llu, rq_xid = x%llu\n",
- desc->bd_last_xid, req->rq_xid);
-
spin_lock(&desc->bd_lock);
- /* Holler if peer manages to touch buffers before he knows the xid */
+ /* Holler if peer manages to touch buffers before he knows the mbits */
if (desc->bd_md_count != total_md)
CWARN("%s: Peer %s touched %d buffers while I registered\n",
desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
total_md - desc->bd_md_count);
spin_unlock(&desc->bd_lock);
- CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, xid x%#llx-%#llx, portal %u\n",
+ CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
desc->bd_md_count,
ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
desc->bd_iov_count, desc->bd_nob,
- desc->bd_last_xid, req->rq_xid, desc->bd_portal);
+ desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
return 0;
}
lustre_msg_set_conn_cnt(request->rq_reqmsg, imp->imp_conn_cnt);
lustre_msghdr_set_flags(request->rq_reqmsg, imp->imp_msghdr_flags);
+ if (request->rq_nr_resend) {
+ /*
+ * resend for EINPROGRESS, allocate new xid to avoid reply
+ * reconstruction
+ */
+ request->rq_xid = ptlrpc_next_xid();
+ DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for resend on EINPROGRESS");
+ }
+
+ if (request->rq_bulk) {
+ ptlrpc_set_bulk_mbits(request);
+ lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
+ }
+
/**
* For enabled AT all request should have AT_SUPPORT in the
* FULL import state when OBD_CONNECT_AT is set
}
}
+void lustre_msg_set_mbits(struct lustre_msg *msg, __u64 mbits)
+{
+ switch (msg->lm_magic) {
+ case LUSTRE_MSG_MAGIC_V2: {
+ struct ptlrpc_body *pb = lustre_msg_ptlrpc_body(msg);
+
+ LASSERTF(pb, "invalid msg %p: no ptlrpc body!\n", msg);
+ pb->pb_mbits = mbits;
+ return;
+ }
+ default:
+ LASSERTF(0, "incorrect message magic: %08x\n", msg->lm_magic);
+ }
+}
+
void ptlrpc_request_set_replen(struct ptlrpc_request *req)
{
int count = req_capsule_filled_sizes(&req->rq_pill, RCL_SERVER);
__swab64s(&b->pb_pre_versions[1]);
__swab64s(&b->pb_pre_versions[2]);
__swab64s(&b->pb_pre_versions[3]);
+ __swab64s(&b->pb_mbits);
CLASSERT(offsetof(typeof(*b), pb_padding0) != 0);
CLASSERT(offsetof(typeof(*b), pb_padding1) != 0);
- CLASSERT(offsetof(typeof(*b), pb_padding) != 0);
+ CLASSERT(offsetof(typeof(*b), pb_padding64_0) != 0);
+ CLASSERT(offsetof(typeof(*b), pb_padding64_1) != 0);
+ CLASSERT(offsetof(typeof(*b), pb_padding64_2) != 0);
/* While we need to maintain compatibility between
* clients and servers without ptlrpc_body_v2 (< 2.3)
* do not swab any fields beyond pb_jobid, as we are
int ptlrpc_expired_set(void *data);
int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
void ptlrpc_resend_req(struct ptlrpc_request *request);
+void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req);
/* events.c */
int ptlrpc_init_portals(void);
(long long)(int)offsetof(struct ptlrpc_body_v3, pb_pre_versions));
LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_pre_versions) == 32, "found %lld\n",
(long long)(int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_pre_versions));
- LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_padding) == 120, "found %lld\n",
- (long long)(int)offsetof(struct ptlrpc_body_v3, pb_padding));
- LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding) == 32, "found %lld\n",
- (long long)(int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding));
+ LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_mbits) == 120, "found %lld\n",
+ (long long)(int)offsetof(struct ptlrpc_body_v3, pb_mbits));
+ LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_mbits) == 8, "found %lld\n",
+ (long long)(int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_mbits));
+ LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_padding64_0) == 128, "found %lld\n",
+ (long long)(int)offsetof(struct ptlrpc_body_v3, pb_padding64_0));
+ LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_0) == 8, "found %lld\n",
+ (long long)(int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_0));
+ LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_padding64_1) == 136, "found %lld\n",
+ (long long)(int)offsetof(struct ptlrpc_body_v3, pb_padding64_1));
+ LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_1) == 8, "found %lld\n",
+ (long long)(int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_1));
+ LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_padding64_2) == 144, "found %lld\n",
+ (long long)(int)offsetof(struct ptlrpc_body_v3, pb_padding64_2));
+ LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_2) == 8, "found %lld\n",
+ (long long)(int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_2));
CLASSERT(LUSTRE_JOBID_SIZE == 32);
LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_jobid) == 152, "found %lld\n",
(long long)(int)offsetof(struct ptlrpc_body_v3, pb_jobid));
(int)offsetof(struct ptlrpc_body_v3, pb_pre_versions), (int)offsetof(struct ptlrpc_body_v2, pb_pre_versions));
LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_pre_versions) == (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_pre_versions), "%d != %d\n",
(int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_pre_versions), (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_pre_versions));
- LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_padding) == (int)offsetof(struct ptlrpc_body_v2, pb_padding), "%d != %d\n",
- (int)offsetof(struct ptlrpc_body_v3, pb_padding), (int)offsetof(struct ptlrpc_body_v2, pb_padding));
- LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding) == (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_padding), "%d != %d\n",
- (int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding), (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_padding));
+ LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_mbits) == (int)offsetof(struct ptlrpc_body_v2, pb_mbits), "%d != %d\n",
+ (int)offsetof(struct ptlrpc_body_v3, pb_mbits), (int)offsetof(struct ptlrpc_body_v2, pb_mbits));
+ LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_mbits) == (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_mbits), "%d != %d\n",
+ (int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_mbits), (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_mbits));
+ LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_padding64_0) == (int)offsetof(struct ptlrpc_body_v2, pb_padding64_0), "%d != %d\n",
+ (int)offsetof(struct ptlrpc_body_v3, pb_padding64_0), (int)offsetof(struct ptlrpc_body_v2, pb_padding64_0));
+ LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_0) == (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_padding64_0), "%d != %d\n",
+ (int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_0), (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_padding64_0));
+ LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_padding64_1) == (int)offsetof(struct ptlrpc_body_v2, pb_padding64_1), "%d != %d\n",
+ (int)offsetof(struct ptlrpc_body_v3, pb_padding64_1), (int)offsetof(struct ptlrpc_body_v2, pb_padding64_1));
+ LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_1) == (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_padding64_1), "%d != %d\n",
+ (int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_1), (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_padding64_1));
+ LASSERTF((int)offsetof(struct ptlrpc_body_v3, pb_padding64_2) == (int)offsetof(struct ptlrpc_body_v2, pb_padding64_2), "%d != %d\n",
+ (int)offsetof(struct ptlrpc_body_v3, pb_padding64_2), (int)offsetof(struct ptlrpc_body_v2, pb_padding64_2));
+ LASSERTF((int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_2) == (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_padding64_2), "%d != %d\n",
+ (int)sizeof(((struct ptlrpc_body_v3 *)0)->pb_padding64_2), (int)sizeof(((struct ptlrpc_body_v2 *)0)->pb_padding64_2));
LASSERTF(MSG_PTLRPC_BODY_OFF == 0, "found %lld\n",
(long long)MSG_PTLRPC_BODY_OFF);
LASSERTF(REQ_REC_OFF == 1, "found %lld\n",
projects / karo-tx-linux.git / commitdiff