From 8cdd312cfed706b067d7ea952603e28cc33c40cc Mon Sep 17 00:00:00 2001 From: Haggai Eran Date: Thu, 11 Dec 2014 17:04:20 +0200 Subject: IB/mlx5: Implement the ODP capability query verb The patch adds infrastructure to query ODP capabilities in the mlx5 driver. The code will read the capabilities from the device, and enable only those capabilities that both the driver and the device supports. At this point ODP is not supported, so no capability is copied from the device, but the patch exposes the global ODP device capability bit. Signed-off-by: Shachar Raindel Signed-off-by: Haggai Eran Signed-off-by: Roland Dreier --- drivers/infiniband/hw/mlx5/main.c | 10 ++++++++++ 1 file changed, 10 insertions(+) (limited to 'drivers/infiniband/hw/mlx5/main.c') diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c index 1ba6c42e4df8..e6d775f2446d 100644 --- a/drivers/infiniband/hw/mlx5/main.c +++ b/drivers/infiniband/hw/mlx5/main.c @@ -244,6 +244,12 @@ static int mlx5_ib_query_device(struct ib_device *ibdev, props->max_mcast_grp; props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */ +#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING + if (dev->mdev->caps.gen.flags & MLX5_DEV_CAP_FLAG_ON_DMND_PG) + props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING; + props->odp_caps = dev->odp_caps; +#endif + out: kfree(in_mad); kfree(out_mad); @@ -1321,6 +1327,8 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev) (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) | (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ) | (1ull << IB_USER_VERBS_CMD_OPEN_QP); + dev->ib_dev.uverbs_ex_cmd_mask = + (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE); dev->ib_dev.query_device = mlx5_ib_query_device; dev->ib_dev.query_port = mlx5_ib_query_port; @@ -1366,6 +1374,8 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev) dev->ib_dev.free_fast_reg_page_list = mlx5_ib_free_fast_reg_page_list; dev->ib_dev.check_mr_status = mlx5_ib_check_mr_status; + mlx5_ib_internal_query_odp_caps(dev); + if (mdev->caps.gen.flags & MLX5_DEV_CAP_FLAG_XRC) { dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd; dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd; -- cgit v1.2.3 From 6aec21f6a8322fa8d43df3ea7f051dfd8967f1b9 Mon Sep 17 00:00:00 2001 From: Haggai Eran Date: Thu, 11 Dec 2014 17:04:23 +0200 Subject: IB/mlx5: Page faults handling infrastructure * Refactor MR registration and cleanup, and fix reg_pages accounting. * Create a work queue to handle page fault events in a kthread context. * Register a fault handler to get events from the core for each QP. The registered fault handler is empty in this patch, and only a later patch implements it. Signed-off-by: Sagi Grimberg Signed-off-by: Shachar Raindel Signed-off-by: Haggai Eran Signed-off-by: Roland Dreier --- drivers/infiniband/hw/mlx5/main.c | 31 +++++++- drivers/infiniband/hw/mlx5/mlx5_ib.h | 67 +++++++++++++++- drivers/infiniband/hw/mlx5/mr.c | 45 +++++++---- drivers/infiniband/hw/mlx5/odp.c | 145 +++++++++++++++++++++++++++++++++++ drivers/infiniband/hw/mlx5/qp.c | 26 ++++++- include/linux/mlx5/driver.h | 2 +- 6 files changed, 294 insertions(+), 22 deletions(-) (limited to 'drivers/infiniband/hw/mlx5/main.c') diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c index e6d775f2446d..a801baa79c8e 100644 --- a/drivers/infiniband/hw/mlx5/main.c +++ b/drivers/infiniband/hw/mlx5/main.c @@ -864,7 +864,7 @@ static ssize_t show_reg_pages(struct device *device, struct mlx5_ib_dev *dev = container_of(device, struct mlx5_ib_dev, ib_dev.dev); - return sprintf(buf, "%d\n", dev->mdev->priv.reg_pages); + return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages)); } static ssize_t show_hca(struct device *device, struct device_attribute *attr, @@ -1389,16 +1389,19 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev) goto err_eqs; mutex_init(&dev->cap_mask_mutex); - spin_lock_init(&dev->mr_lock); err = create_dev_resources(&dev->devr); if (err) goto err_eqs; - err = ib_register_device(&dev->ib_dev, NULL); + err = mlx5_ib_odp_init_one(dev); if (err) goto err_rsrc; + err = ib_register_device(&dev->ib_dev, NULL); + if (err) + goto err_odp; + err = create_umr_res(dev); if (err) goto err_dev; @@ -1420,6 +1423,9 @@ err_umrc: err_dev: ib_unregister_device(&dev->ib_dev); +err_odp: + mlx5_ib_odp_remove_one(dev); + err_rsrc: destroy_dev_resources(&dev->devr); @@ -1435,8 +1441,10 @@ err_dealloc: static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context) { struct mlx5_ib_dev *dev = context; + ib_unregister_device(&dev->ib_dev); destroy_umrc_res(dev); + mlx5_ib_odp_remove_one(dev); destroy_dev_resources(&dev->devr); free_comp_eqs(dev); ib_dealloc_device(&dev->ib_dev); @@ -1450,15 +1458,30 @@ static struct mlx5_interface mlx5_ib_interface = { static int __init mlx5_ib_init(void) { + int err; + if (deprecated_prof_sel != 2) pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n"); - return mlx5_register_interface(&mlx5_ib_interface); + err = mlx5_ib_odp_init(); + if (err) + return err; + + err = mlx5_register_interface(&mlx5_ib_interface); + if (err) + goto clean_odp; + + return err; + +clean_odp: + mlx5_ib_odp_cleanup(); + return err; } static void __exit mlx5_ib_cleanup(void) { mlx5_unregister_interface(&mlx5_ib_interface); + mlx5_ib_odp_cleanup(); } module_init(mlx5_ib_init); diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h index 6856e27bfb6a..c6ceec3e3d6a 100644 --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h @@ -149,6 +149,29 @@ enum { MLX5_QP_EMPTY }; +/* + * Connect-IB can trigger up to four concurrent pagefaults + * per-QP. + */ +enum mlx5_ib_pagefault_context { + MLX5_IB_PAGEFAULT_RESPONDER_READ, + MLX5_IB_PAGEFAULT_REQUESTOR_READ, + MLX5_IB_PAGEFAULT_RESPONDER_WRITE, + MLX5_IB_PAGEFAULT_REQUESTOR_WRITE, + MLX5_IB_PAGEFAULT_CONTEXTS +}; + +static inline enum mlx5_ib_pagefault_context + mlx5_ib_get_pagefault_context(struct mlx5_pagefault *pagefault) +{ + return pagefault->flags & (MLX5_PFAULT_REQUESTOR | MLX5_PFAULT_WRITE); +} + +struct mlx5_ib_pfault { + struct work_struct work; + struct mlx5_pagefault mpfault; +}; + struct mlx5_ib_qp { struct ib_qp ibqp; struct mlx5_core_qp mqp; @@ -194,6 +217,21 @@ struct mlx5_ib_qp { /* Store signature errors */ bool signature_en; + +#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING + /* + * A flag that is true for QP's that are in a state that doesn't + * allow page faults, and shouldn't schedule any more faults. + */ + int disable_page_faults; + /* + * The disable_page_faults_lock protects a QP's disable_page_faults + * field, allowing for a thread to atomically check whether the QP + * allows page faults, and if so schedule a page fault. + */ + spinlock_t disable_page_faults_lock; + struct mlx5_ib_pfault pagefaults[MLX5_IB_PAGEFAULT_CONTEXTS]; +#endif }; struct mlx5_ib_cq_buf { @@ -392,13 +430,17 @@ struct mlx5_ib_dev { struct umr_common umrc; /* sync used page count stats */ - spinlock_t mr_lock; struct mlx5_ib_resources devr; struct mlx5_mr_cache cache; struct timer_list delay_timer; int fill_delay; #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING struct ib_odp_caps odp_caps; + /* + * Sleepable RCU that prevents destruction of MRs while they are still + * being used by a page fault handler. + */ + struct srcu_struct mr_srcu; #endif }; @@ -575,12 +617,33 @@ int mlx5_ib_check_mr_status(struct ib_mr *ibmr, u32 check_mask, struct ib_mr_status *mr_status); #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING +extern struct workqueue_struct *mlx5_ib_page_fault_wq; + int mlx5_ib_internal_query_odp_caps(struct mlx5_ib_dev *dev); -#else +void mlx5_ib_mr_pfault_handler(struct mlx5_ib_qp *qp, + struct mlx5_ib_pfault *pfault); +void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp); +int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev); +void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev); +int __init mlx5_ib_odp_init(void); +void mlx5_ib_odp_cleanup(void); +void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp); +void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp); + +#else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int mlx5_ib_internal_query_odp_caps(struct mlx5_ib_dev *dev) { return 0; } + +static inline void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp) {} +static inline int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev) { return 0; } +static inline void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev) {} +static inline int mlx5_ib_odp_init(void) { return 0; } +static inline void mlx5_ib_odp_cleanup(void) {} +static inline void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp) {} +static inline void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp) {} + #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline void init_query_mad(struct ib_smp *mad) diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c index 38b06267798e..922ac85b7198 100644 --- a/drivers/infiniband/hw/mlx5/mr.c +++ b/drivers/infiniband/hw/mlx5/mr.c @@ -52,6 +52,8 @@ static __be64 mlx5_ib_update_mtt_emergency_buffer[ static DEFINE_MUTEX(mlx5_ib_update_mtt_emergency_buffer_mutex); #endif +static int clean_mr(struct mlx5_ib_mr *mr); + static int order2idx(struct mlx5_ib_dev *dev, int order) { struct mlx5_mr_cache *cache = &dev->cache; @@ -1049,6 +1051,10 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, mlx5_ib_dbg(dev, "cache empty for order %d", order); mr = NULL; } + } else if (access_flags & IB_ACCESS_ON_DEMAND) { + err = -EINVAL; + pr_err("Got MR registration for ODP MR > 512MB, not supported for Connect-IB"); + goto error; } if (!mr) @@ -1064,9 +1070,7 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, mr->umem = umem; mr->npages = npages; - spin_lock(&dev->mr_lock); - dev->mdev->priv.reg_pages += npages; - spin_unlock(&dev->mr_lock); + atomic_add(npages, &dev->mdev->priv.reg_pages); mr->ibmr.lkey = mr->mmr.key; mr->ibmr.rkey = mr->mmr.key; @@ -1110,12 +1114,9 @@ error: return err; } -int mlx5_ib_dereg_mr(struct ib_mr *ibmr) +static int clean_mr(struct mlx5_ib_mr *mr) { - struct mlx5_ib_dev *dev = to_mdev(ibmr->device); - struct mlx5_ib_mr *mr = to_mmr(ibmr); - struct ib_umem *umem = mr->umem; - int npages = mr->npages; + struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device); int umred = mr->umred; int err; @@ -1135,16 +1136,32 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr) free_cached_mr(dev, mr); } + if (!umred) + kfree(mr); + + return 0; +} + +int mlx5_ib_dereg_mr(struct ib_mr *ibmr) +{ + struct mlx5_ib_dev *dev = to_mdev(ibmr->device); + struct mlx5_ib_mr *mr = to_mmr(ibmr); + int npages = mr->npages; + struct ib_umem *umem = mr->umem; + +#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING + if (umem) + /* Wait for all running page-fault handlers to finish. */ + synchronize_srcu(&dev->mr_srcu); +#endif + + clean_mr(mr); + if (umem) { ib_umem_release(umem); - spin_lock(&dev->mr_lock); - dev->mdev->priv.reg_pages -= npages; - spin_unlock(&dev->mr_lock); + atomic_sub(npages, &dev->mdev->priv.reg_pages); } - if (!umred) - kfree(mr); - return 0; } diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c index 66c39ee16aff..63bbdba396f1 100644 --- a/drivers/infiniband/hw/mlx5/odp.c +++ b/drivers/infiniband/hw/mlx5/odp.c @@ -32,6 +32,8 @@ #include "mlx5_ib.h" +struct workqueue_struct *mlx5_ib_page_fault_wq; + #define COPY_ODP_BIT_MLX_TO_IB(reg, ib_caps, field_name, bit_name) do { \ if (be32_to_cpu(reg.field_name) & MLX5_ODP_SUPPORT_##bit_name) \ ib_caps->field_name |= IB_ODP_SUPPORT_##bit_name; \ @@ -58,3 +60,146 @@ int mlx5_ib_internal_query_odp_caps(struct mlx5_ib_dev *dev) out: return err; } + +static struct mlx5_ib_mr *mlx5_ib_odp_find_mr_lkey(struct mlx5_ib_dev *dev, + u32 key) +{ + u32 base_key = mlx5_base_mkey(key); + struct mlx5_core_mr *mmr = __mlx5_mr_lookup(dev->mdev, base_key); + + if (!mmr || mmr->key != key) + return NULL; + + return container_of(mmr, struct mlx5_ib_mr, mmr); +} + +static void mlx5_ib_page_fault_resume(struct mlx5_ib_qp *qp, + struct mlx5_ib_pfault *pfault, + int error) { + struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device); + int ret = mlx5_core_page_fault_resume(dev->mdev, qp->mqp.qpn, + pfault->mpfault.flags, + error); + if (ret) + pr_err("Failed to resolve the page fault on QP 0x%x\n", + qp->mqp.qpn); +} + +void mlx5_ib_mr_pfault_handler(struct mlx5_ib_qp *qp, + struct mlx5_ib_pfault *pfault) +{ + u8 event_subtype = pfault->mpfault.event_subtype; + + switch (event_subtype) { + default: + pr_warn("Invalid page fault event subtype: 0x%x\n", + event_subtype); + mlx5_ib_page_fault_resume(qp, pfault, 1); + break; + } +} + +static void mlx5_ib_qp_pfault_action(struct work_struct *work) +{ + struct mlx5_ib_pfault *pfault = container_of(work, + struct mlx5_ib_pfault, + work); + enum mlx5_ib_pagefault_context context = + mlx5_ib_get_pagefault_context(&pfault->mpfault); + struct mlx5_ib_qp *qp = container_of(pfault, struct mlx5_ib_qp, + pagefaults[context]); + mlx5_ib_mr_pfault_handler(qp, pfault); +} + +void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp) +{ + unsigned long flags; + + spin_lock_irqsave(&qp->disable_page_faults_lock, flags); + qp->disable_page_faults = 1; + spin_unlock_irqrestore(&qp->disable_page_faults_lock, flags); + + /* + * Note that at this point, we are guarenteed that no more + * work queue elements will be posted to the work queue with + * the QP we are closing. + */ + flush_workqueue(mlx5_ib_page_fault_wq); +} + +void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp) +{ + unsigned long flags; + + spin_lock_irqsave(&qp->disable_page_faults_lock, flags); + qp->disable_page_faults = 0; + spin_unlock_irqrestore(&qp->disable_page_faults_lock, flags); +} + +static void mlx5_ib_pfault_handler(struct mlx5_core_qp *qp, + struct mlx5_pagefault *pfault) +{ + /* + * Note that we will only get one fault event per QP per context + * (responder/initiator, read/write), until we resolve the page fault + * with the mlx5_ib_page_fault_resume command. Since this function is + * called from within the work element, there is no risk of missing + * events. + */ + struct mlx5_ib_qp *mibqp = to_mibqp(qp); + enum mlx5_ib_pagefault_context context = + mlx5_ib_get_pagefault_context(pfault); + struct mlx5_ib_pfault *qp_pfault = &mibqp->pagefaults[context]; + + qp_pfault->mpfault = *pfault; + + /* No need to stop interrupts here since we are in an interrupt */ + spin_lock(&mibqp->disable_page_faults_lock); + if (!mibqp->disable_page_faults) + queue_work(mlx5_ib_page_fault_wq, &qp_pfault->work); + spin_unlock(&mibqp->disable_page_faults_lock); +} + +void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp) +{ + int i; + + qp->disable_page_faults = 1; + spin_lock_init(&qp->disable_page_faults_lock); + + qp->mqp.pfault_handler = mlx5_ib_pfault_handler; + + for (i = 0; i < MLX5_IB_PAGEFAULT_CONTEXTS; ++i) + INIT_WORK(&qp->pagefaults[i].work, mlx5_ib_qp_pfault_action); +} + +int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev) +{ + int ret; + + ret = init_srcu_struct(&ibdev->mr_srcu); + if (ret) + return ret; + + return 0; +} + +void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev) +{ + cleanup_srcu_struct(&ibdev->mr_srcu); +} + +int __init mlx5_ib_odp_init(void) +{ + mlx5_ib_page_fault_wq = + create_singlethread_workqueue("mlx5_ib_page_faults"); + if (!mlx5_ib_page_fault_wq) + return -ENOMEM; + + return 0; +} + +void mlx5_ib_odp_cleanup(void) +{ + destroy_workqueue(mlx5_ib_page_fault_wq); +} diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c index 9783c3342dbf..be0cd358b080 100644 --- a/drivers/infiniband/hw/mlx5/qp.c +++ b/drivers/infiniband/hw/mlx5/qp.c @@ -876,6 +876,8 @@ static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd, int inlen = sizeof(*in); int err; + mlx5_ib_odp_create_qp(qp); + gen = &dev->mdev->caps.gen; mutex_init(&qp->mutex); spin_lock_init(&qp->sq.lock); @@ -1160,11 +1162,13 @@ static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp) in = kzalloc(sizeof(*in), GFP_KERNEL); if (!in) return; - if (qp->state != IB_QPS_RESET) + if (qp->state != IB_QPS_RESET) { + mlx5_ib_qp_disable_pagefaults(qp); if (mlx5_core_qp_modify(dev->mdev, to_mlx5_state(qp->state), MLX5_QP_STATE_RST, in, sizeof(*in), &qp->mqp)) mlx5_ib_warn(dev, "mlx5_ib: modify QP %06x to RESET failed\n", qp->mqp.qpn); + } get_cqs(qp, &send_cq, &recv_cq); @@ -1712,6 +1716,15 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, if (mlx5_st < 0) goto out; + /* If moving to a reset or error state, we must disable page faults on + * this QP and flush all current page faults. Otherwise a stale page + * fault may attempt to work on this QP after it is reset and moved + * again to RTS, and may cause the driver and the device to get out of + * sync. */ + if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR && + (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR)) + mlx5_ib_qp_disable_pagefaults(qp); + optpar = ib_mask_to_mlx5_opt(attr_mask); optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st]; in->optparam = cpu_to_be32(optpar); @@ -1721,6 +1734,9 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, if (err) goto out; + if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) + mlx5_ib_qp_enable_pagefaults(qp); + qp->state = new_state; if (attr_mask & IB_QP_ACCESS_FLAGS) @@ -3026,6 +3042,14 @@ int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr int mlx5_state; int err = 0; +#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING + /* + * Wait for any outstanding page faults, in case the user frees memory + * based upon this query's result. + */ + flush_workqueue(mlx5_ib_page_fault_wq); +#endif + mutex_lock(&qp->mutex); outb = kzalloc(sizeof(*outb), GFP_KERNEL); if (!outb) { diff --git a/include/linux/mlx5/driver.h b/include/linux/mlx5/driver.h index 7088dcd19214..166d9315fe4b 100644 --- a/include/linux/mlx5/driver.h +++ b/include/linux/mlx5/driver.h @@ -474,7 +474,7 @@ struct mlx5_priv { struct workqueue_struct *pg_wq; struct rb_root page_root; int fw_pages; - int reg_pages; + atomic_t reg_pages; struct list_head free_list; struct mlx5_core_health health; -- cgit v1.2.3 From b4cfe447d47b5763f630412fd5dc5fbe66e991d1 Mon Sep 17 00:00:00 2001 From: Haggai Eran Date: Thu, 11 Dec 2014 17:04:26 +0200 Subject: IB/mlx5: Implement on demand paging by adding support for MMU notifiers * Implement the relevant invalidation functions (zap MTTs as needed) * Implement interlocking (and rollback in the page fault handlers) for cases of a racing notifier and fault. * With this patch we can now enable the capability bits for supporting RC send/receive/RDMA read/RDMA write, and UD send. Signed-off-by: Sagi Grimberg Signed-off-by: Shachar Raindel Signed-off-by: Haggai Eran Signed-off-by: Roland Dreier --- drivers/infiniband/hw/mlx5/main.c | 4 ++ drivers/infiniband/hw/mlx5/mlx5_ib.h | 3 + drivers/infiniband/hw/mlx5/mr.c | 79 +++++++++++++++++++-- drivers/infiniband/hw/mlx5/odp.c | 128 ++++++++++++++++++++++++++++++++--- 4 files changed, 198 insertions(+), 16 deletions(-) (limited to 'drivers/infiniband/hw/mlx5/main.c') diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c index a801baa79c8e..8a87404e9c76 100644 --- a/drivers/infiniband/hw/mlx5/main.c +++ b/drivers/infiniband/hw/mlx5/main.c @@ -574,6 +574,10 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev, goto out_count; } +#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING + context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range; +#endif + INIT_LIST_HEAD(&context->db_page_list); mutex_init(&context->db_page_mutex); diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h index c6ceec3e3d6a..83f22fe297c8 100644 --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h @@ -325,6 +325,7 @@ struct mlx5_ib_mr { struct mlx5_ib_dev *dev; struct mlx5_create_mkey_mbox_out out; struct mlx5_core_sig_ctx *sig; + int live; }; struct mlx5_ib_fast_reg_page_list { @@ -629,6 +630,8 @@ int __init mlx5_ib_odp_init(void); void mlx5_ib_odp_cleanup(void); void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp); void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp); +void mlx5_ib_invalidate_range(struct ib_umem *umem, unsigned long start, + unsigned long end); #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int mlx5_ib_internal_query_odp_caps(struct mlx5_ib_dev *dev) diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c index 922ac85b7198..32a28bd50b20 100644 --- a/drivers/infiniband/hw/mlx5/mr.c +++ b/drivers/infiniband/hw/mlx5/mr.c @@ -37,6 +37,7 @@ #include #include #include +#include #include #include "mlx5_ib.h" @@ -54,6 +55,18 @@ static DEFINE_MUTEX(mlx5_ib_update_mtt_emergency_buffer_mutex); static int clean_mr(struct mlx5_ib_mr *mr); +static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr) +{ + int err = mlx5_core_destroy_mkey(dev->mdev, &mr->mmr); + +#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING + /* Wait until all page fault handlers using the mr complete. */ + synchronize_srcu(&dev->mr_srcu); +#endif + + return err; +} + static int order2idx(struct mlx5_ib_dev *dev, int order) { struct mlx5_mr_cache *cache = &dev->cache; @@ -191,7 +204,7 @@ static void remove_keys(struct mlx5_ib_dev *dev, int c, int num) ent->cur--; ent->size--; spin_unlock_irq(&ent->lock); - err = mlx5_core_destroy_mkey(dev->mdev, &mr->mmr); + err = destroy_mkey(dev, mr); if (err) mlx5_ib_warn(dev, "failed destroy mkey\n"); else @@ -482,7 +495,7 @@ static void clean_keys(struct mlx5_ib_dev *dev, int c) ent->cur--; ent->size--; spin_unlock_irq(&ent->lock); - err = mlx5_core_destroy_mkey(dev->mdev, &mr->mmr); + err = destroy_mkey(dev, mr); if (err) mlx5_ib_warn(dev, "failed destroy mkey\n"); else @@ -812,6 +825,8 @@ static struct mlx5_ib_mr *reg_umr(struct ib_pd *pd, struct ib_umem *umem, mr->mmr.size = len; mr->mmr.pd = to_mpd(pd)->pdn; + mr->live = 1; + unmap_dma: up(&umrc->sem); dma_unmap_single(ddev, dma, size, DMA_TO_DEVICE); @@ -997,6 +1012,7 @@ static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, u64 virt_addr, goto err_2; } mr->umem = umem; + mr->live = 1; kvfree(in); mlx5_ib_dbg(dev, "mkey = 0x%x\n", mr->mmr.key); @@ -1074,10 +1090,47 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, mr->ibmr.lkey = mr->mmr.key; mr->ibmr.rkey = mr->mmr.key; +#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING + if (umem->odp_data) { + /* + * This barrier prevents the compiler from moving the + * setting of umem->odp_data->private to point to our + * MR, before reg_umr finished, to ensure that the MR + * initialization have finished before starting to + * handle invalidations. + */ + smp_wmb(); + mr->umem->odp_data->private = mr; + /* + * Make sure we will see the new + * umem->odp_data->private value in the invalidation + * routines, before we can get page faults on the + * MR. Page faults can happen once we put the MR in + * the tree, below this line. Without the barrier, + * there can be a fault handling and an invalidation + * before umem->odp_data->private == mr is visible to + * the invalidation handler. + */ + smp_wmb(); + } +#endif + return &mr->ibmr; error: + /* + * Destroy the umem *before* destroying the MR, to ensure we + * will not have any in-flight notifiers when destroying the + * MR. + * + * As the MR is completely invalid to begin with, and this + * error path is only taken if we can't push the mr entry into + * the pagefault tree, this is safe. + */ + ib_umem_release(umem); + /* Kill the MR, and return an error code. */ + clean_mr(mr); return ERR_PTR(err); } @@ -1121,7 +1174,7 @@ static int clean_mr(struct mlx5_ib_mr *mr) int err; if (!umred) { - err = mlx5_core_destroy_mkey(dev->mdev, &mr->mmr); + err = destroy_mkey(dev, mr); if (err) { mlx5_ib_warn(dev, "failed to destroy mkey 0x%x (%d)\n", mr->mmr.key, err); @@ -1150,9 +1203,25 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr) struct ib_umem *umem = mr->umem; #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING - if (umem) + if (umem && umem->odp_data) { + /* Prevent new page faults from succeeding */ + mr->live = 0; /* Wait for all running page-fault handlers to finish. */ synchronize_srcu(&dev->mr_srcu); + /* Destroy all page mappings */ + mlx5_ib_invalidate_range(umem, ib_umem_start(umem), + ib_umem_end(umem)); + /* + * We kill the umem before the MR for ODP, + * so that there will not be any invalidations in + * flight, looking at the *mr struct. + */ + ib_umem_release(umem); + atomic_sub(npages, &dev->mdev->priv.reg_pages); + + /* Avoid double-freeing the umem. */ + umem = NULL; + } #endif clean_mr(mr); @@ -1269,7 +1338,7 @@ int mlx5_ib_destroy_mr(struct ib_mr *ibmr) kfree(mr->sig); } - err = mlx5_core_destroy_mkey(dev->mdev, &mr->mmr); + err = destroy_mkey(dev, mr); if (err) { mlx5_ib_warn(dev, "failed to destroy mkey 0x%x (%d)\n", mr->mmr.key, err); diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c index 936a6cd4ecc7..a2c541c4809a 100644 --- a/drivers/infiniband/hw/mlx5/odp.c +++ b/drivers/infiniband/hw/mlx5/odp.c @@ -37,8 +37,78 @@ #define MAX_PREFETCH_LEN (4*1024*1024U) +/* Timeout in ms to wait for an active mmu notifier to complete when handling + * a pagefault. */ +#define MMU_NOTIFIER_TIMEOUT 1000 + struct workqueue_struct *mlx5_ib_page_fault_wq; +void mlx5_ib_invalidate_range(struct ib_umem *umem, unsigned long start, + unsigned long end) +{ + struct mlx5_ib_mr *mr; + const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT / sizeof(u64)) - 1; + u64 idx = 0, blk_start_idx = 0; + int in_block = 0; + u64 addr; + + if (!umem || !umem->odp_data) { + pr_err("invalidation called on NULL umem or non-ODP umem\n"); + return; + } + + mr = umem->odp_data->private; + + if (!mr || !mr->ibmr.pd) + return; + + start = max_t(u64, ib_umem_start(umem), start); + end = min_t(u64, ib_umem_end(umem), end); + + /* + * Iteration one - zap the HW's MTTs. The notifiers_count ensures that + * while we are doing the invalidation, no page fault will attempt to + * overwrite the same MTTs. Concurent invalidations might race us, + * but they will write 0s as well, so no difference in the end result. + */ + + for (addr = start; addr < end; addr += (u64)umem->page_size) { + idx = (addr - ib_umem_start(umem)) / PAGE_SIZE; + /* + * Strive to write the MTTs in chunks, but avoid overwriting + * non-existing MTTs. The huristic here can be improved to + * estimate the cost of another UMR vs. the cost of bigger + * UMR. + */ + if (umem->odp_data->dma_list[idx] & + (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) { + if (!in_block) { + blk_start_idx = idx; + in_block = 1; + } + } else { + u64 umr_offset = idx & umr_block_mask; + + if (in_block && umr_offset == 0) { + mlx5_ib_update_mtt(mr, blk_start_idx, + idx - blk_start_idx, 1); + in_block = 0; + } + } + } + if (in_block) + mlx5_ib_update_mtt(mr, blk_start_idx, idx - blk_start_idx + 1, + 1); + + /* + * We are now sure that the device will not access the + * memory. We can safely unmap it, and mark it as dirty if + * needed. + */ + + ib_umem_odp_unmap_dma_pages(umem, start, end); +} + #define COPY_ODP_BIT_MLX_TO_IB(reg, ib_caps, field_name, bit_name) do { \ if (be32_to_cpu(reg.field_name) & MLX5_ODP_SUPPORT_##bit_name) \ ib_caps->field_name |= IB_ODP_SUPPORT_##bit_name; \ @@ -59,9 +129,18 @@ int mlx5_ib_internal_query_odp_caps(struct mlx5_ib_dev *dev) if (err) goto out; - /* At this point we would copy the capability bits that the driver - * supports from the hw_caps struct to the caps struct. However, no - * such capabilities are supported so far. */ + caps->general_caps = IB_ODP_SUPPORT; + COPY_ODP_BIT_MLX_TO_IB(hw_caps, caps, per_transport_caps.ud_odp_caps, + SEND); + COPY_ODP_BIT_MLX_TO_IB(hw_caps, caps, per_transport_caps.rc_odp_caps, + SEND); + COPY_ODP_BIT_MLX_TO_IB(hw_caps, caps, per_transport_caps.rc_odp_caps, + RECV); + COPY_ODP_BIT_MLX_TO_IB(hw_caps, caps, per_transport_caps.rc_odp_caps, + WRITE); + COPY_ODP_BIT_MLX_TO_IB(hw_caps, caps, per_transport_caps.rc_odp_caps, + READ); + out: return err; } @@ -71,8 +150,9 @@ static struct mlx5_ib_mr *mlx5_ib_odp_find_mr_lkey(struct mlx5_ib_dev *dev, { u32 base_key = mlx5_base_mkey(key); struct mlx5_core_mr *mmr = __mlx5_mr_lookup(dev->mdev, base_key); + struct mlx5_ib_mr *mr = container_of(mmr, struct mlx5_ib_mr, mmr); - if (!mmr || mmr->key != key) + if (!mmr || mmr->key != key || !mr->live) return NULL; return container_of(mmr, struct mlx5_ib_mr, mmr); @@ -143,6 +223,11 @@ static int pagefault_single_data_segment(struct mlx5_ib_qp *qp, } current_seq = ACCESS_ONCE(mr->umem->odp_data->notifiers_seq); + /* + * Ensure the sequence number is valid for some time before we call + * gup. + */ + smp_rmb(); /* * Avoid branches - this code will perform correctly @@ -165,15 +250,20 @@ static int pagefault_single_data_segment(struct mlx5_ib_qp *qp, if (npages > 0) { mutex_lock(&mr->umem->odp_data->umem_mutex); - /* - * No need to check whether the MTTs really belong to - * this MR, since ib_umem_odp_map_dma_pages already - * checks this. - */ - ret = mlx5_ib_update_mtt(mr, start_idx, npages, 0); + if (!ib_umem_mmu_notifier_retry(mr->umem, current_seq)) { + /* + * No need to check whether the MTTs really belong to + * this MR, since ib_umem_odp_map_dma_pages already + * checks this. + */ + ret = mlx5_ib_update_mtt(mr, start_idx, npages, 0); + } else { + ret = -EAGAIN; + } mutex_unlock(&mr->umem->odp_data->umem_mutex); if (ret < 0) { - pr_err("Failed to update mkey page tables\n"); + if (ret != -EAGAIN) + pr_err("Failed to update mkey page tables\n"); goto srcu_unlock; } @@ -185,6 +275,22 @@ static int pagefault_single_data_segment(struct mlx5_ib_qp *qp, } srcu_unlock: + if (ret == -EAGAIN) { + if (!mr->umem->odp_data->dying) { + struct ib_umem_odp *odp_data = mr->umem->odp_data; + unsigned long timeout = + msecs_to_jiffies(MMU_NOTIFIER_TIMEOUT); + + if (!wait_for_completion_timeout( + &odp_data->notifier_completion, + timeout)) { + pr_warn("timeout waiting for mmu notifier completion\n"); + } + } else { + /* The MR is being killed, kill the QP as well. */ + ret = -EFAULT; + } + } srcu_read_unlock(&mib_dev->mr_srcu, srcu_key); pfault->mpfault.bytes_committed = 0; return ret ? ret : npages; -- cgit v1.2.3