Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma

Pull rdma fixes from Doug Ledford:
"Here's our second -rc pull request. Nothing particularly special in
this one. The client removal deadlock fix is kindy tricky, but we had
multiple eyes on it and no one could find a fault in it. A couple
Spectre V1 fixes too. Otherwise, all just normal -rc fodder:

- A couple Spectre V1 fixes (umad, hfi1)

- Fix a tricky deadlock in the rdma core code with refcounting
instead of locks (client removal patches)

- Build errors (hns)

- Fix a scheduling while atomic issue (mlx5)

- Use after free fix (mad)

- Fix error path return code (hns)

- Null deref fix (siw_crypto_hash)

- A few other misc. minor fixes"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
RDMA/hns: Fix error return code in hns_roce_v1_rsv_lp_qp()
RDMA/mlx5: Release locks during notifier unregister
IB/hfi1: Fix Spectre v1 vulnerability
IB/mad: Fix use-after-free in ib mad completion handling
RDMA/restrack: Track driver QP types in resource tracker
IB/mlx5: Fix MR registration flow to use UMR properly
RDMA/devices: Remove the lock around remove_client_context
RDMA/devices: Do not deadlock during client removal
IB/core: Add mitigation for Spectre V1
Do not dereference 'siw_crypto_shash' before checking
RDMA/qedr: Fix the hca_type and hca_rev returned in device attributes
RDMA/hns: Fix build error

drivers/infiniband/core/core_priv.h

···
 					  struct ib_udata *udata,
 					  struct ib_uobject *uobj)
 {
+	enum ib_qp_type qp_type = attr->qp_type;
 	struct ib_qp *qp;
+	bool is_xrc;
 
 	if (!dev->ops.create_qp)
 		return ERR_PTR(-EOPNOTSUPP);
···
 	 * and more importantly they are created internaly by driver,
 	 * see mlx5 create_dev_resources() as an example.
 	 */
-	if (attr->qp_type < IB_QPT_XRC_INI) {
+	is_xrc = qp_type == IB_QPT_XRC_INI || qp_type == IB_QPT_XRC_TGT;
+	if ((qp_type < IB_QPT_MAX && !is_xrc) || qp_type == IB_QPT_DRIVER) {
 		qp->res.type = RDMA_RESTRACK_QP;
 		if (uobj)
 			rdma_restrack_uadd(&qp->res);

drivers/infiniband/core/device.c

···
 static DECLARE_RWSEM(devices_rwsem);
 #define DEVICE_REGISTERED XA_MARK_1
 
-static LIST_HEAD(client_list);
+static u32 highest_client_id;
 #define CLIENT_REGISTERED XA_MARK_1
 static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
 static DECLARE_RWSEM(clients_rwsem);
+
+static void ib_client_put(struct ib_client *client)
+{
+	if (refcount_dec_and_test(&client->uses))
+		complete(&client->uses_zero);
+}
 
 /*
  * If client_data is registered then the corresponding client must also still
···
 
 	down_write(&device->client_data_rwsem);
 	/*
+	 * So long as the client is registered hold both the client and device
+	 * unregistration locks.
+	 */
+	if (!refcount_inc_not_zero(&client->uses))
+		goto out_unlock;
+	refcount_inc(&device->refcount);
+
+	/*
 	 * Another caller to add_client_context got here first and has already
 	 * completely initialized context.
 	 */
···
 	return 0;
 
 out:
+	ib_device_put(device);
+	ib_client_put(client);
+out_unlock:
 	up_write(&device->client_data_rwsem);
 	return ret;
 }
···
 	client_data = xa_load(&device->client_data, client_id);
 	xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
 	client = xa_load(&clients, client_id);
-	downgrade_write(&device->client_data_rwsem);
+	up_write(&device->client_data_rwsem);
 
 	/*
 	 * Notice we cannot be holding any exclusive locks when calling the
···
 	 *
 	 * For this reason clients and drivers should not call the
 	 * unregistration functions will holdling any locks.
-	 *
-	 * It tempting to drop the client_data_rwsem too, but this is required
-	 * to ensure that unregister_client does not return until all clients
-	 * are completely unregistered, which is required to avoid module
-	 * unloading races.
 	 */
 	if (client->remove)
 		client->remove(device, client_data);
 
 	xa_erase(&device->client_data, client_id);
-	up_read(&device->client_data_rwsem);
+	ib_device_put(device);
+	ib_client_put(client);
 }
 
 static int alloc_port_data(struct ib_device *device)
···
 
 static void disable_device(struct ib_device *device)
 {
-	struct ib_client *client;
+	u32 cid;
 
 	WARN_ON(!refcount_read(&device->refcount));
 
···
 	xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
 	up_write(&devices_rwsem);
 
+	/*
+	 * Remove clients in LIFO order, see assign_client_id. This could be
+	 * more efficient if xarray learns to reverse iterate. Since no new
+	 * clients can be added to this ib_device past this point we only need
+	 * the maximum possible client_id value here.
+	 */
 	down_read(&clients_rwsem);
-	list_for_each_entry_reverse(client, &client_list, list)
-		remove_client_context(device, client->client_id);
+	cid = highest_client_id;
 	up_read(&clients_rwsem);
+	while (cid) {
+		cid--;
+		remove_client_context(device, cid);
+	}
 
 	/* Pairs with refcount_set in enable_device */
 	ib_device_put(device);
···
 	/*
 	 * The add/remove callbacks must be called in FIFO/LIFO order. To
 	 * achieve this we assign client_ids so they are sorted in
-	 * registration order, and retain a linked list we can reverse iterate
-	 * to get the LIFO order. The extra linked list can go away if xarray
-	 * learns to reverse iterate.
+	 * registration order.
 	 */
-	if (list_empty(&client_list)) {
-		client->client_id = 0;
-	} else {
-		struct ib_client *last;
-
-		last = list_last_entry(&client_list, struct ib_client, list);
-		client->client_id = last->client_id + 1;
-	}
+	client->client_id = highest_client_id;
 	ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
 	if (ret)
 		goto out;
 
+	highest_client_id++;
 	xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
-	list_add_tail(&client->list, &client_list);
 
 out:
 	up_write(&clients_rwsem);
 	return ret;
+}
+
+static void remove_client_id(struct ib_client *client)
+{
+	down_write(&clients_rwsem);
+	xa_erase(&clients, client->client_id);
+	for (; highest_client_id; highest_client_id--)
+		if (xa_load(&clients, highest_client_id - 1))
+			break;
+	up_write(&clients_rwsem);
 }
 
 /**
···
 	unsigned long index;
 	int ret;
 
+	refcount_set(&client->uses, 1);
+	init_completion(&client->uses_zero);
 	ret = assign_client_id(client);
 	if (ret)
 		return ret;
···
 	unsigned long index;
 
 	down_write(&clients_rwsem);
+	ib_client_put(client);
 	xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
 	up_write(&clients_rwsem);
-	/*
-	 * Every device still known must be serialized to make sure we are
-	 * done with the client callbacks before we return.
-	 */
-	down_read(&devices_rwsem);
-	xa_for_each (&devices, index, device)
-		remove_client_context(device, client->client_id);
-	up_read(&devices_rwsem);
 
-	down_write(&clients_rwsem);
-	list_del(&client->list);
-	xa_erase(&clients, client->client_id);
-	up_write(&clients_rwsem);
+	/* We do not want to have locks while calling client->remove() */
+	rcu_read_lock();
+	xa_for_each (&devices, index, device) {
+		if (!ib_device_try_get(device))
+			continue;
+		rcu_read_unlock();
+
+		remove_client_context(device, client->client_id);
+
+		ib_device_put(device);
+		rcu_read_lock();
+	}
+	rcu_read_unlock();
+
+	/*
+	 * remove_client_context() is not a fence, it can return even though a
+	 * removal is ongoing. Wait until all removals are completed.
+	 */
+	wait_for_completion(&client->uses_zero);
+	remove_client_id(client);
 }
 EXPORT_SYMBOL(ib_unregister_client);
 

drivers/infiniband/core/mad.c

···
 	if (has_smi)
 		cq_size *= 2;
 
+	port_priv->pd = ib_alloc_pd(device, 0);
+	if (IS_ERR(port_priv->pd)) {
+		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
+		ret = PTR_ERR(port_priv->pd);
+		goto error3;
+	}
+
 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
 			IB_POLL_UNBOUND_WORKQUEUE);
 	if (IS_ERR(port_priv->cq)) {
 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
 		ret = PTR_ERR(port_priv->cq);
-		goto error3;
-	}
-
-	port_priv->pd = ib_alloc_pd(device, 0);
-	if (IS_ERR(port_priv->pd)) {
-		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
-		ret = PTR_ERR(port_priv->pd);
 		goto error4;
 	}
 
···
 error7:
 	destroy_mad_qp(&port_priv->qp_info[0]);
 error6:
-	ib_dealloc_pd(port_priv->pd);
-error4:
 	ib_free_cq(port_priv->cq);
 	cleanup_recv_queue(&port_priv->qp_info[1]);
 	cleanup_recv_queue(&port_priv->qp_info[0]);
+error4:
+	ib_dealloc_pd(port_priv->pd);
 error3:
 	kfree(port_priv);
 
···
 	destroy_workqueue(port_priv->wq);
 	destroy_mad_qp(&port_priv->qp_info[1]);
 	destroy_mad_qp(&port_priv->qp_info[0]);
-	ib_dealloc_pd(port_priv->pd);
 	ib_free_cq(port_priv->cq);
+	ib_dealloc_pd(port_priv->pd);
 	cleanup_recv_queue(&port_priv->qp_info[1]);
 	cleanup_recv_queue(&port_priv->qp_info[0]);
 	/* XXX: Handle deallocation of MAD registration tables */

drivers/infiniband/core/user_mad.c

···
 #include <linux/sched.h>
 #include <linux/semaphore.h>
 #include <linux/slab.h>
+#include <linux/nospec.h>
 
 #include <linux/uaccess.h>
 
···
 
 	if (get_user(id, arg))
 		return -EFAULT;
+	if (id >= IB_UMAD_MAX_AGENTS)
+		return -EINVAL;
 
 	mutex_lock(&file->port->file_mutex);
 	mutex_lock(&file->mutex);
 
-	if (id >= IB_UMAD_MAX_AGENTS || !__get_agent(file, id)) {
+	id = array_index_nospec(id, IB_UMAD_MAX_AGENTS);
+	if (!__get_agent(file, id)) {
 		ret = -EINVAL;
 		goto out;
 	}

drivers/infiniband/hw/hfi1/verbs.c

···
 #include <linux/mm.h>
 #include <linux/vmalloc.h>
 #include <rdma/opa_addr.h>
+#include <linux/nospec.h>
 
 #include "hfi.h"
 #include "common.h"
···
 	sl = rdma_ah_get_sl(ah_attr);
 	if (sl >= ARRAY_SIZE(ibp->sl_to_sc))
 		return -EINVAL;
+	sl = array_index_nospec(sl, ARRAY_SIZE(ibp->sl_to_sc));
 
 	sc5 = ibp->sl_to_sc[sl];
 	if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)

drivers/infiniband/hw/hns/Kconfig

···
 # SPDX-License-Identifier: GPL-2.0-only
 config INFINIBAND_HNS
-	tristate "HNS RoCE Driver"
+	bool "HNS RoCE Driver"
 	depends on NET_VENDOR_HISILICON
 	depends on ARM64 || (COMPILE_TEST && 64BIT)
 	---help---
···
 	  To compile HIP06 or HIP08 driver as module, choose M here.
 
 config INFINIBAND_HNS_HIP06
-	bool "Hisilicon Hip06 Family RoCE support"
+	tristate "Hisilicon Hip06 Family RoCE support"
 	depends on INFINIBAND_HNS && HNS && HNS_DSAF && HNS_ENET
 	---help---
 	  RoCE driver support for Hisilicon RoCE engine in Hisilicon Hip06 and
···
 	  module will be called hns-roce-hw-v1
 
 config INFINIBAND_HNS_HIP08
-	bool "Hisilicon Hip08 Family RoCE support"
+	tristate "Hisilicon Hip08 Family RoCE support"
 	depends on INFINIBAND_HNS && PCI && HNS3
 	---help---
 	  RoCE driver support for Hisilicon RoCE engine in Hisilicon Hip08 SoC.

drivers/infiniband/hw/hns/Makefile

···
 	hns_roce_ah.o hns_roce_hem.o hns_roce_mr.o hns_roce_qp.o \
 	hns_roce_cq.o hns_roce_alloc.o hns_roce_db.o hns_roce_srq.o hns_roce_restrack.o
 
-ifdef CONFIG_INFINIBAND_HNS_HIP06
 hns-roce-hw-v1-objs := hns_roce_hw_v1.o $(hns-roce-objs)
-obj-$(CONFIG_INFINIBAND_HNS) += hns-roce-hw-v1.o
-endif
+obj-$(CONFIG_INFINIBAND_HNS_HIP06) += hns-roce-hw-v1.o
 
-ifdef CONFIG_INFINIBAND_HNS_HIP08
 hns-roce-hw-v2-objs := hns_roce_hw_v2.o hns_roce_hw_v2_dfx.o $(hns-roce-objs)
-obj-$(CONFIG_INFINIBAND_HNS) += hns-roce-hw-v2.o
-endif
+obj-$(CONFIG_INFINIBAND_HNS_HIP08) += hns-roce-hw-v2.o

drivers/infiniband/hw/hns/hns_roce_hw_v1.c

···
 	atomic_set(&free_mr->mr_free_cq->ib_cq.usecnt, 0);
 
 	pd = rdma_zalloc_drv_obj(ibdev, ib_pd);
-	if (!pd)
+	if (!pd) {
+		ret = -ENOMEM;
 		goto alloc_mem_failed;
+	}
 
 	pd->device  = ibdev;
 	ret = hns_roce_alloc_pd(pd, NULL);

drivers/infiniband/hw/mlx5/main.c

···
 		return;
 	}
 
-	if (mpi->mdev_events.notifier_call)
-		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
-	mpi->mdev_events.notifier_call = NULL;
-
 	mpi->ibdev = NULL;
 
 	spin_unlock(&port->mp.mpi_lock);
+	if (mpi->mdev_events.notifier_call)
+		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
+	mpi->mdev_events.notifier_call = NULL;
 	mlx5_remove_netdev_notifier(ibdev, port_num);
 	spin_lock(&port->mp.mpi_lock);
 

drivers/infiniband/hw/mlx5/mr.c

···
 static void dereg_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr);
 static int mr_cache_max_order(struct mlx5_ib_dev *dev);
 static int unreg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr);
-static bool umr_can_modify_entity_size(struct mlx5_ib_dev *dev)
-{
-	return !MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled);
-}
 
 static bool umr_can_use_indirect_mkey(struct mlx5_ib_dev *dev)
 {
 	return !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled);
-}
-
-static bool use_umr(struct mlx5_ib_dev *dev, int order)
-{
-	return order <= mr_cache_max_order(dev) &&
-		umr_can_modify_entity_size(dev);
 }
 
 static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
···
 {
 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
 	struct mlx5_ib_mr *mr = NULL;
-	bool populate_mtts = false;
+	bool use_umr;
 	struct ib_umem *umem;
 	int page_shift;
 	int npages;
···
 	if (err < 0)
 		return ERR_PTR(err);
 
-	if (use_umr(dev, order)) {
+	use_umr = !MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled) &&
+		  (!MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled) ||
+		   !MLX5_CAP_GEN(dev->mdev, atomic));
+
+	if (order <= mr_cache_max_order(dev) && use_umr) {
 		mr = alloc_mr_from_cache(pd, umem, virt_addr, length, ncont,
 					 page_shift, order, access_flags);
 		if (PTR_ERR(mr) == -EAGAIN) {
 			mlx5_ib_dbg(dev, "cache empty for order %d\n", order);
 			mr = NULL;
 		}
-		populate_mtts = false;
 	} else if (!MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset)) {
 		if (access_flags & IB_ACCESS_ON_DEMAND) {
 			err = -EINVAL;
 			pr_err("Got MR registration for ODP MR > 512MB, not supported for Connect-IB\n");
 			goto error;
 		}
-		populate_mtts = true;
+		use_umr = false;
 	}
 
 	if (!mr) {
-		if (!umr_can_modify_entity_size(dev))
-			populate_mtts = true;
 		mutex_lock(&dev->slow_path_mutex);
 		mr = reg_create(NULL, pd, virt_addr, length, umem, ncont,
-				page_shift, access_flags, populate_mtts);
+				page_shift, access_flags, !use_umr);
 		mutex_unlock(&dev->slow_path_mutex);
 	}
 
···
 
 	update_odp_mr(mr);
 
-	if (!populate_mtts) {
+	if (use_umr) {
 		int update_xlt_flags = MLX5_IB_UPD_XLT_ENABLE;
 
 		if (access_flags & IB_ACCESS_ON_DEMAND)

drivers/infiniband/hw/qedr/main.c

···
 	struct qedr_dev *dev =
 		rdma_device_to_drv_device(device, struct qedr_dev, ibdev);
 
-	return scnprintf(buf, PAGE_SIZE, "0x%x\n", dev->pdev->vendor);
+	return scnprintf(buf, PAGE_SIZE, "0x%x\n", dev->attr.hw_ver);
 }
 static DEVICE_ATTR_RO(hw_rev);
 
 static ssize_t hca_type_show(struct device *device,
 			     struct device_attribute *attr, char *buf)
 {
-	return scnprintf(buf, PAGE_SIZE, "%s\n", "HCA_TYPE_TO_SET");
+	struct qedr_dev *dev =
+		rdma_device_to_drv_device(device, struct qedr_dev, ibdev);
+
+	return scnprintf(buf, PAGE_SIZE, "FastLinQ QL%x %s\n",
+			 dev->pdev->device,
+			 rdma_protocol_iwarp(&dev->ibdev, 1) ?
+			 "iWARP" : "RoCE");
 }
 static DEVICE_ATTR_RO(hca_type);
 

drivers/infiniband/sw/siw/siw_qp.c

···
 {
 	struct siw_rx_stream *c_rx = &qp->rx_stream;
 	struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
-	int size = crypto_shash_descsize(siw_crypto_shash) +
-			sizeof(struct shash_desc);
+	int size;
 
 	if (siw_crypto_shash == NULL)
 		return -ENOENT;
+
+	size = crypto_shash_descsize(siw_crypto_shash) +
+		sizeof(struct shash_desc);
 
 	c_tx->mpa_crc_hd = kzalloc(size, GFP_KERNEL);
 	c_rx->mpa_crc_hd = kzalloc(size, GFP_KERNEL);

include/rdma/ib_verbs.h

···
 			const union ib_gid *gid,
 			const struct sockaddr *addr,
 			void *client_data);
-	struct list_head list;
+
+	refcount_t uses;
+	struct completion uses_zero;
 	u32 client_id;
 
 	/* kverbs are not required by the client */