From a9c3c4c597704b3a1a2b9bef990e7d8a881f6533 Mon Sep 17 00:00:00 2001 From: James Erwin Date: Fri, 1 Nov 2019 15:20:59 -0400 Subject: IB/hfi1: Ensure full Gen3 speed in a Gen4 system If an hfi1 card is inserted in a Gen4 systems, the driver will avoid the gen3 speed bump and the card will operate at half speed. This is because the driver avoids the gen3 speed bump when the parent bus speed isn't identical to gen3, 8.0GT/s. This is not compatible with gen4 and newer speeds. Fix by relaxing the test to explicitly look for the lower capability speeds which inherently allows for gen4 and all future speeds. Fixes: 7724105686e7 ("IB/hfi1: add driver files") Link: https://lore.kernel.org/r/20191101192059.106248.1699.stgit@awfm-01.aw.intel.com Cc: Reviewed-by: Dennis Dalessandro Reviewed-by: Kaike Wan Signed-off-by: James Erwin Signed-off-by: Mike Marciniszyn Signed-off-by: Dennis Dalessandro Signed-off-by: Jason Gunthorpe --- drivers/infiniband/hw/hfi1/pcie.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) (limited to 'drivers/infiniband') diff --git a/drivers/infiniband/hw/hfi1/pcie.c b/drivers/infiniband/hw/hfi1/pcie.c index 61aa5504d7c3..61362bd6d3ce 100644 --- a/drivers/infiniband/hw/hfi1/pcie.c +++ b/drivers/infiniband/hw/hfi1/pcie.c @@ -319,7 +319,9 @@ int pcie_speeds(struct hfi1_devdata *dd) /* * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed */ - if (parent && dd->pcidev->bus->max_bus_speed != PCIE_SPEED_8_0GT) { + if (parent && + (dd->pcidev->bus->max_bus_speed == PCIE_SPEED_2_5GT || + dd->pcidev->bus->max_bus_speed == PCIE_SPEED_5_0GT)) { dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n"); dd->link_gen3_capable = 0; } -- cgit From c1abd865bd125015783286b353abb8da51644f59 Mon Sep 17 00:00:00 2001 From: Kaike Wan Date: Fri, 25 Oct 2019 15:58:30 -0400 Subject: IB/hfi1: Ensure r_tid_ack is valid before building TID RDMA ACK packet The index r_tid_ack is used to indicate the next TID RDMA WRITE request to acknowledge in the ring s_ack_queue[] on the responder side and should be set to a valid index other than its initial value before r_tid_tail is advanced to the next TID RDMA WRITE request and particularly before a TID RDMA ACK is built. Otherwise, a NULL pointer dereference may result: BUG: unable to handle kernel paging request at ffff9a32d27abff8 IP: [] hfi1_make_tid_rdma_pkt+0x476/0xcb0 [hfi1] PGD 2749032067 PUD 0 Oops: 0000 1 SMP Modules linked in: osp(OE) ofd(OE) lfsck(OE) ost(OE) mgc(OE) osd_zfs(OE) lquota(OE) lustre(OE) lmv(OE) mdc(OE) lov(OE) fid(OE) fld(OE) ko2iblnd(OE) ptlrpc(OE) obdclass(OE) lnet(OE) libcfs(OE) ib_ipoib(OE) hfi1(OE) rdmavt(OE) nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache ib_isert iscsi_target_mod target_core_mod ib_ucm dm_mirror dm_region_hash dm_log mlx5_ib dm_mod zfs(POE) rpcrdma sunrpc rdma_ucm ib_uverbs opa_vnic ib_iser zunicode(POE) ib_umad zavl(POE) icp(POE) sb_edac intel_powerclamp coretemp rdma_cm intel_rapl iosf_mbi iw_cm libiscsi scsi_transport_iscsi kvm ib_cm iTCO_wdt mxm_wmi iTCO_vendor_support irqbypass crc32_pclmul ghash_clmulni_intel aesni_intel lrw gf128mul glue_helper ablk_helper cryptd zcommon(POE) znvpair(POE) pcspkr spl(OE) mei_me sg mei ioatdma lpc_ich joydev i2c_i801 shpchp ipmi_si ipmi_devintf ipmi_msghandler wmi acpi_power_meter ip_tables xfs libcrc32c sd_mod crc_t10dif crct10dif_generic mgag200 mlx5_core drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ixgbe ahci ttm mlxfw ib_core libahci devlink mdio crct10dif_pclmul crct10dif_common drm ptp libata megaraid_sas crc32c_intel i2c_algo_bit pps_core i2c_core dca [last unloaded: rdmavt] CPU: 15 PID: 68691 Comm: kworker/15:2H Kdump: loaded Tainted: P W OE ------------ 3.10.0-862.2.3.el7_lustre.x86_64 #1 Hardware name: Intel Corporation S2600WTT/S2600WTT, BIOS SE5C610.86B.01.01.0016.033120161139 03/31/2016 Workqueue: hfi0_0 _hfi1_do_tid_send [hfi1] task: ffff9a01f47faf70 ti: ffff9a11776a8000 task.ti: ffff9a11776a8000 RIP: 0010:[] [] hfi1_make_tid_rdma_pkt+0x476/0xcb0 [hfi1] RSP: 0018:ffff9a11776abd08 EFLAGS: 00010002 RAX: ffff9a32d27abfc0 RBX: ffff99f2d27aa000 RCX: 00000000ffffffff RDX: 0000000000000000 RSI: 0000000000000220 RDI: ffff99f2ffc05300 RBP: ffff9a11776abd88 R08: 000000000001c310 R09: ffffffffc0d87ad4 R10: 0000000000000000 R11: 0000000000000000 R12: ffff9a117a423c00 R13: ffff9a117a423c00 R14: ffff9a03500c0000 R15: ffff9a117a423cb8 FS: 0000000000000000(0000) GS:ffff9a117e9c0000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: ffff9a32d27abff8 CR3: 0000002748a0e000 CR4: 00000000001607e0 Call Trace: [] _hfi1_do_tid_send+0x194/0x320 [hfi1] [] process_one_work+0x17f/0x440 [] worker_thread+0x126/0x3c0 [] ? manage_workers.isra.24+0x2a0/0x2a0 [] kthread+0xd1/0xe0 [] ? insert_kthread_work+0x40/0x40 [] ret_from_fork_nospec_begin+0x21/0x21 [] ? insert_kthread_work+0x40/0x40 hfi1 0000:05:00.0: hfi1_0: reserved_op: opcode 0xf2, slot 2, rsv_used 1, rsv_ops 1 Code: 00 00 41 8b 8d d8 02 00 00 89 c8 48 89 45 b0 48 c1 65 b0 06 48 8b 83 a0 01 00 00 48 01 45 b0 48 8b 45 b0 41 80 bd 10 03 00 00 00 <48> 8b 50 38 4c 8d 7a 50 74 45 8b b2 d0 00 00 00 85 f6 0f 85 72 RIP [] hfi1_make_tid_rdma_pkt+0x476/0xcb0 [hfi1] RSP CR2: ffff9a32d27abff8 This problem can happen if a RESYNC request is received before r_tid_ack is modified. This patch fixes the issue by making sure that r_tid_ack is set to a valid value before a TID RDMA ACK is built. Functions are defined to simplify the code. Fixes: 07b923701e38 ("IB/hfi1: Add functions to receive TID RDMA WRITE request") Fixes: 7cf0ad679de4 ("IB/hfi1: Add a function to receive TID RDMA RESYNC packet") Link: https://lore.kernel.org/r/20191025195830.106825.44022.stgit@awfm-01.aw.intel.com Cc: Reviewed-by: Mike Marciniszyn Reviewed-by: Dennis Dalessandro Signed-off-by: Kaike Wan Signed-off-by: Dennis Dalessandro Signed-off-by: Jason Gunthorpe --- drivers/infiniband/hw/hfi1/tid_rdma.c | 44 +++++++++++++++++++++-------------- 1 file changed, 27 insertions(+), 17 deletions(-) (limited to 'drivers/infiniband') diff --git a/drivers/infiniband/hw/hfi1/tid_rdma.c b/drivers/infiniband/hw/hfi1/tid_rdma.c index f21fca3617d5..73bc78ba288e 100644 --- a/drivers/infiniband/hw/hfi1/tid_rdma.c +++ b/drivers/infiniband/hw/hfi1/tid_rdma.c @@ -136,6 +136,26 @@ static void update_r_next_psn_fecn(struct hfi1_packet *packet, struct tid_rdma_flow *flow, bool fecn); +static void validate_r_tid_ack(struct hfi1_qp_priv *priv) +{ + if (priv->r_tid_ack == HFI1_QP_WQE_INVALID) + priv->r_tid_ack = priv->r_tid_tail; +} + +static void tid_rdma_schedule_ack(struct rvt_qp *qp) +{ + struct hfi1_qp_priv *priv = qp->priv; + + priv->s_flags |= RVT_S_ACK_PENDING; + hfi1_schedule_tid_send(qp); +} + +static void tid_rdma_trigger_ack(struct rvt_qp *qp) +{ + validate_r_tid_ack(qp->priv); + tid_rdma_schedule_ack(qp); +} + static u64 tid_rdma_opfn_encode(struct tid_rdma_params *p) { return @@ -3005,10 +3025,7 @@ nak_psn: qpriv->s_nak_state = IB_NAK_PSN_ERROR; /* We are NAK'ing the next expected PSN */ qpriv->s_nak_psn = mask_psn(flow->flow_state.r_next_psn); - qpriv->s_flags |= RVT_S_ACK_PENDING; - if (qpriv->r_tid_ack == HFI1_QP_WQE_INVALID) - qpriv->r_tid_ack = qpriv->r_tid_tail; - hfi1_schedule_tid_send(qp); + tid_rdma_trigger_ack(qp); } goto unlock; } @@ -3526,7 +3543,7 @@ static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx) /* * If overtaking req->acked_tail, send an RNR NAK. Because the * QP is not queued in this case, and the issue can only be - * caused due a delay in scheduling the second leg which we + * caused by a delay in scheduling the second leg which we * cannot estimate, we use a rather arbitrary RNR timeout of * (MAX_FLOWS / 2) segments */ @@ -3534,8 +3551,7 @@ static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx) MAX_FLOWS)) { ret = -EAGAIN; to_seg = MAX_FLOWS >> 1; - qpriv->s_flags |= RVT_S_ACK_PENDING; - hfi1_schedule_tid_send(qp); + tid_rdma_trigger_ack(qp); break; } @@ -4335,8 +4351,7 @@ void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet) trace_hfi1_tid_req_rcv_write_data(qp, 0, e->opcode, e->psn, e->lpsn, req); trace_hfi1_tid_write_rsp_rcv_data(qp); - if (priv->r_tid_ack == HFI1_QP_WQE_INVALID) - priv->r_tid_ack = priv->r_tid_tail; + validate_r_tid_ack(priv); if (opcode == TID_OP(WRITE_DATA_LAST)) { release_rdma_sge_mr(e); @@ -4375,8 +4390,7 @@ void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet) } done: - priv->s_flags |= RVT_S_ACK_PENDING; - hfi1_schedule_tid_send(qp); + tid_rdma_schedule_ack(qp); exit: priv->r_next_psn_kdeth = flow->flow_state.r_next_psn; if (fecn) @@ -4388,10 +4402,7 @@ send_nak: if (!priv->s_nak_state) { priv->s_nak_state = IB_NAK_PSN_ERROR; priv->s_nak_psn = flow->flow_state.r_next_psn; - priv->s_flags |= RVT_S_ACK_PENDING; - if (priv->r_tid_ack == HFI1_QP_WQE_INVALID) - priv->r_tid_ack = priv->r_tid_tail; - hfi1_schedule_tid_send(qp); + tid_rdma_trigger_ack(qp); } goto done; } @@ -4939,8 +4950,7 @@ void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet) qpriv->resync = true; /* RESYNC request always gets a TID RDMA ACK. */ qpriv->s_nak_state = 0; - qpriv->s_flags |= RVT_S_ACK_PENDING; - hfi1_schedule_tid_send(qp); + tid_rdma_trigger_ack(qp); bail: if (fecn) qp->s_flags |= RVT_S_ECN; -- cgit From c2be3865a1763c4be39574937e1aae27e917af4d Mon Sep 17 00:00:00 2001 From: Kaike Wan Date: Fri, 25 Oct 2019 15:58:36 -0400 Subject: IB/hfi1: Calculate flow weight based on QP MTU for TID RDMA For a TID RDMA WRITE request, a QP on the responder side could be put into a queue when a hardware flow is not available. A RNR NAK will be returned to the requester with a RNR timeout value based on the position of the QP in the queue. The tid_rdma_flow_wt variable is used to calculate the timeout value and is determined by using a MTU of 4096 at the module loading time. This could reduce the timeout value by half from the desired value, leading to excessive RNR retries. This patch fixes the issue by calculating the flow weight with the real MTU assigned to the QP. Fixes: 07b923701e38 ("IB/hfi1: Add functions to receive TID RDMA WRITE request") Link: https://lore.kernel.org/r/20191025195836.106825.77769.stgit@awfm-01.aw.intel.com Cc: Reviewed-by: Mike Marciniszyn Reviewed-by: Dennis Dalessandro Signed-off-by: Kaike Wan Signed-off-by: Dennis Dalessandro Signed-off-by: Jason Gunthorpe --- drivers/infiniband/hw/hfi1/init.c | 1 - drivers/infiniband/hw/hfi1/tid_rdma.c | 13 +++++-------- drivers/infiniband/hw/hfi1/tid_rdma.h | 3 +-- 3 files changed, 6 insertions(+), 11 deletions(-) (limited to 'drivers/infiniband') diff --git a/drivers/infiniband/hw/hfi1/init.c b/drivers/infiniband/hw/hfi1/init.c index 71cb9525c074..26b792bb1027 100644 --- a/drivers/infiniband/hw/hfi1/init.c +++ b/drivers/infiniband/hw/hfi1/init.c @@ -1489,7 +1489,6 @@ static int __init hfi1_mod_init(void) goto bail_dev; } - hfi1_compute_tid_rdma_flow_wt(); /* * These must be called before the driver is registered with * the PCI subsystem. diff --git a/drivers/infiniband/hw/hfi1/tid_rdma.c b/drivers/infiniband/hw/hfi1/tid_rdma.c index 73bc78ba288e..e53f542b60af 100644 --- a/drivers/infiniband/hw/hfi1/tid_rdma.c +++ b/drivers/infiniband/hw/hfi1/tid_rdma.c @@ -107,8 +107,6 @@ static u32 mask_generation(u32 a) * C - Capcode */ -static u32 tid_rdma_flow_wt; - static void tid_rdma_trigger_resume(struct work_struct *work); static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req); static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req, @@ -3388,18 +3386,17 @@ u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe, return sizeof(ohdr->u.tid_rdma.w_req) / sizeof(u32); } -void hfi1_compute_tid_rdma_flow_wt(void) +static u32 hfi1_compute_tid_rdma_flow_wt(struct rvt_qp *qp) { /* * Heuristic for computing the RNR timeout when waiting on the flow * queue. Rather than a computationaly expensive exact estimate of when * a flow will be available, we assume that if a QP is at position N in * the flow queue it has to wait approximately (N + 1) * (number of - * segments between two sync points), assuming PMTU of 4K. The rationale - * for this is that flows are released and recycled at each sync point. + * segments between two sync points). The rationale for this is that + * flows are released and recycled at each sync point. */ - tid_rdma_flow_wt = MAX_TID_FLOW_PSN * enum_to_mtu(OPA_MTU_4096) / - TID_RDMA_MAX_SEGMENT_SIZE; + return (MAX_TID_FLOW_PSN * qp->pmtu) >> TID_RDMA_SEGMENT_SHIFT; } static u32 position_in_queue(struct hfi1_qp_priv *qpriv, @@ -3522,7 +3519,7 @@ static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx) if (qpriv->flow_state.index >= RXE_NUM_TID_FLOWS) { ret = hfi1_kern_setup_hw_flow(qpriv->rcd, qp); if (ret) { - to_seg = tid_rdma_flow_wt * + to_seg = hfi1_compute_tid_rdma_flow_wt(qp) * position_in_queue(qpriv, &rcd->flow_queue); break; diff --git a/drivers/infiniband/hw/hfi1/tid_rdma.h b/drivers/infiniband/hw/hfi1/tid_rdma.h index 1c536185261e..6e82df2190b7 100644 --- a/drivers/infiniband/hw/hfi1/tid_rdma.h +++ b/drivers/infiniband/hw/hfi1/tid_rdma.h @@ -17,6 +17,7 @@ #define TID_RDMA_MIN_SEGMENT_SIZE BIT(18) /* 256 KiB (for now) */ #define TID_RDMA_MAX_SEGMENT_SIZE BIT(18) /* 256 KiB (for now) */ #define TID_RDMA_MAX_PAGES (BIT(18) >> PAGE_SHIFT) +#define TID_RDMA_SEGMENT_SHIFT 18 /* * Bit definitions for priv->s_flags. @@ -274,8 +275,6 @@ u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe, struct ib_other_headers *ohdr, u32 *bth1, u32 *bth2, u32 *len); -void hfi1_compute_tid_rdma_flow_wt(void); - void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet); u32 hfi1_build_tid_rdma_write_resp(struct rvt_qp *qp, struct rvt_ack_entry *e, -- cgit From ce8e8087cf3b5b4f19d29248bfc7deef95525490 Mon Sep 17 00:00:00 2001 From: Kaike Wan Date: Fri, 25 Oct 2019 15:58:42 -0400 Subject: IB/hfi1: TID RDMA WRITE should not return IB_WC_RNR_RETRY_EXC_ERR Normal RDMA WRITE request never returns IB_WC_RNR_RETRY_EXC_ERR to ULPs because it does not need post receive buffer on the responder side. Consequently, as an enhancement to normal RDMA WRITE request inside the hfi1 driver, TID RDMA WRITE request should not return such an error status to ULPs, although it does receive RNR NAKs from the responder when TID resources are not available. This behavior is violated when qp->s_rnr_retry_cnt is set in current hfi1 implementation. This patch enforces these semantics by avoiding any reaction to the updates of the RNR QP attributes. Fixes: 3c6cb20a0d17 ("IB/hfi1: Add TID RDMA WRITE functionality into RDMA verbs") Link: https://lore.kernel.org/r/20191025195842.106825.71532.stgit@awfm-01.aw.intel.com Cc: Reviewed-by: Mike Marciniszyn Reviewed-by: Dennis Dalessandro Signed-off-by: Kaike Wan Signed-off-by: Dennis Dalessandro Signed-off-by: Jason Gunthorpe --- drivers/infiniband/hw/hfi1/rc.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) (limited to 'drivers/infiniband') diff --git a/drivers/infiniband/hw/hfi1/rc.c b/drivers/infiniband/hw/hfi1/rc.c index 513a8aac9ccd..1a3c647675a7 100644 --- a/drivers/infiniband/hw/hfi1/rc.c +++ b/drivers/infiniband/hw/hfi1/rc.c @@ -2209,15 +2209,15 @@ int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode, if (qp->s_flags & RVT_S_WAIT_RNR) goto bail_stop; rdi = ib_to_rvt(qp->ibqp.device); - if (qp->s_rnr_retry == 0 && - !((rdi->post_parms[wqe->wr.opcode].flags & - RVT_OPERATION_IGN_RNR_CNT) && - qp->s_rnr_retry_cnt == 0)) { - status = IB_WC_RNR_RETRY_EXC_ERR; - goto class_b; + if (!(rdi->post_parms[wqe->wr.opcode].flags & + RVT_OPERATION_IGN_RNR_CNT)) { + if (qp->s_rnr_retry == 0) { + status = IB_WC_RNR_RETRY_EXC_ERR; + goto class_b; + } + if (qp->s_rnr_retry_cnt < 7 && qp->s_rnr_retry_cnt > 0) + qp->s_rnr_retry--; } - if (qp->s_rnr_retry_cnt < 7 && qp->s_rnr_retry_cnt > 0) - qp->s_rnr_retry--; /* * The last valid PSN is the previous PSN. For TID RDMA WRITE -- cgit From 531eb45b3da4267fc2a64233ba256c8ffb02edd2 Mon Sep 17 00:00:00 2001 From: Sirong Wang Date: Fri, 1 Nov 2019 10:33:29 +0800 Subject: RDMA/hns: Correct the value of HNS_ROCE_HEM_CHUNK_LEN Size of pointer to buf field of struct hns_roce_hem_chunk should be considered when calculating HNS_ROCE_HEM_CHUNK_LEN, or sg table size will be larger than expected when allocating hem. Fixes: 9a4435375cd1 ("IB/hns: Add driver files for hns RoCE driver") Link: https://lore.kernel.org/r/1572575610-52530-2-git-send-email-liweihang@hisilicon.com Signed-off-by: Sirong Wang Signed-off-by: Weihang Li Signed-off-by: Jason Gunthorpe --- drivers/infiniband/hw/hns/hns_roce_hem.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/infiniband') diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.h b/drivers/infiniband/hw/hns/hns_roce_hem.h index 86783276fb1f..3bb8f78fb7b0 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hem.h +++ b/drivers/infiniband/hw/hns/hns_roce_hem.h @@ -59,7 +59,7 @@ enum { #define HNS_ROCE_HEM_CHUNK_LEN \ ((256 - sizeof(struct list_head) - 2 * sizeof(int)) / \ - (sizeof(struct scatterlist))) + (sizeof(struct scatterlist) + sizeof(void *))) #define check_whether_bt_num_3(type, hop_num) \ (type < HEM_TYPE_MTT && hop_num == 2) -- cgit From 411c1e6774e2e1f96b1ccce4f119376b94ade3e4 Mon Sep 17 00:00:00 2001 From: Wenpeng Liang Date: Fri, 1 Nov 2019 10:33:30 +0800 Subject: RDMA/hns: Correct the value of srq_desc_size srq_desc_size should be rounded up to pow of two before used, or related calculation may cause allocating wrong size of memory for srq buffer. Fixes: c7bcb13442e1 ("RDMA/hns: Add SRQ support for hip08 kernel mode") Link: https://lore.kernel.org/r/1572575610-52530-3-git-send-email-liweihang@hisilicon.com Signed-off-by: Wenpeng Liang Signed-off-by: Weihang Li Signed-off-by: Jason Gunthorpe --- drivers/infiniband/hw/hns/hns_roce_srq.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/infiniband') diff --git a/drivers/infiniband/hw/hns/hns_roce_srq.c b/drivers/infiniband/hw/hns/hns_roce_srq.c index 9591457eb768..43ea2c13b212 100644 --- a/drivers/infiniband/hw/hns/hns_roce_srq.c +++ b/drivers/infiniband/hw/hns/hns_roce_srq.c @@ -376,7 +376,7 @@ int hns_roce_create_srq(struct ib_srq *ib_srq, srq->max = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1); srq->max_gs = srq_init_attr->attr.max_sge; - srq_desc_size = max(16, 16 * srq->max_gs); + srq_desc_size = roundup_pow_of_two(max(16, 16 * srq->max_gs)); srq->wqe_shift = ilog2(srq_desc_size); -- cgit