From: Junxian Huang huangjunxian6@hisilicon.com

driver inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/I968IB

----------------------------------------------------------

Support hns roce bonding

Signed-off-by: Junxian Huang huangjunxian6@hisilicon.com Signed-off-by: ChunZhi Hu huchunzhi@huawei.com --- drivers/infiniband/hw/hns/Makefile | 4 +- drivers/infiniband/hw/hns/hns_roce_bond.c | 810 ++++++++++++++++++++ drivers/infiniband/hw/hns/hns_roce_bond.h | 86 +++ drivers/infiniband/hw/hns/hns_roce_device.h | 8 +- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 132 +++- drivers/infiniband/hw/hns/hns_roce_hw_v2.h | 19 + drivers/infiniband/hw/hns/hns_roce_main.c | 73 +- 7 files changed, 1119 insertions(+), 13 deletions(-) create mode 100644 drivers/infiniband/hw/hns/hns_roce_bond.c create mode 100644 drivers/infiniband/hw/hns/hns_roce_bond.h

diff --git a/drivers/infiniband/hw/hns/Makefile b/drivers/infiniband/hw/hns/Makefile index 161615fda..8faa43009 100644 --- a/drivers/infiniband/hw/hns/Makefile +++ b/drivers/infiniband/hw/hns/Makefile @@ -4,11 +4,13 @@ #

ccflags-y := -I $(srctree)/drivers/net/ethernet/hisilicon/hns3 +ccflags-y += -I $(srctree)/drivers/net/ethernet/hisilicon/hns3/hns3pf +ccflags-y += -I $(srctree)/drivers/net/ethernet/hisilicon/hns3/hns3_common

hns-roce-objs := hns_roce_main.o hns_roce_cmd.o hns_roce_pd.o \ 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 \ - hns_roce_debugfs.o hns_roce_sysfs.o + hns_roce_debugfs.o hns_roce_sysfs.o hns_roce_bond.o

ifdef CONFIG_INFINIBAND_HNS_HIP08 hns-roce-hw-v2-objs := hns_roce_hw_v2.o $(hns-roce-objs) diff --git a/drivers/infiniband/hw/hns/hns_roce_bond.c b/drivers/infiniband/hw/hns/hns_roce_bond.c new file mode 100644 index 000000000..d6ebf41c9 --- /dev/null +++ b/drivers/infiniband/hw/hns/hns_roce_bond.c @@ -0,0 +1,810 @@ +// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB +/* + * Copyright (c) 2022 Hisilicon Limited. + */ + +#include <linux/pci.h> +#include "hnae3.h" +#include "hns_roce_device.h" +#include "hns_roce_hw_v2.h" +#include "hns_roce_bond.h" + +static DEFINE_MUTEX(roce_bond_mutex); +static DEFINE_XARRAY(roce_bond_xa); + +static struct hns_roce_dev *hns_roce_get_hrdev_by_netdev(struct net_device *net_dev) +{ + struct ib_device *ibdev = + ib_device_get_by_netdev(net_dev, RDMA_DRIVER_HNS); + struct hns_roce_dev *hr_dev; + + if (!ibdev) + return NULL; + + hr_dev = container_of(ibdev, struct hns_roce_dev, ib_dev); + ib_device_put(ibdev); + + return hr_dev; +} + +static struct net_device *get_upper_dev_from_ndev(struct net_device *net_dev) +{ + struct net_device *upper_dev; + + rcu_read_lock(); + upper_dev = netdev_master_upper_dev_get_rcu(net_dev); + rcu_read_unlock(); + + return upper_dev; +} + +static bool is_netdev_bond_slave(struct net_device *net_dev, + struct hns_roce_bond_group *bond_grp) +{ + int i; + + if (!net_dev || !bond_grp) + return false; + + for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) + if (net_dev == bond_grp->bond_func_info[i].net_dev) + return true; + + return false; +} + +static bool is_hrdev_bond_slave(struct hns_roce_dev *hr_dev, + struct net_device *upper_dev) +{ + struct hns_roce_bond_group *bond_grp; + + if (!hr_dev || !upper_dev) + return false; + + if (!netif_is_lag_master(upper_dev)) + return false; + + if (upper_dev == get_upper_dev_from_ndev(hr_dev->iboe.netdevs[0])) + return true; + + bond_grp = hns_roce_get_bond_grp(hr_dev); + if (bond_grp && upper_dev == bond_grp->upper_dev) + return true; + + return false; +} + +struct hns_roce_bond_group *hns_roce_get_bond_grp(struct hns_roce_dev *hr_dev) +{ + struct hns_roce_die_info *die_info = + xa_load(&roce_bond_xa, hr_dev->pci_dev->bus->number); + struct hns_roce_bond_group *bond_grp; + int i; + + if (!die_info) + return NULL; + + for (i = 0; i < ROCE_BOND_NUM_MAX; i++) { + bond_grp = die_info->bgrps[i]; + if (!bond_grp) + continue; + if (is_netdev_bond_slave(hr_dev->iboe.netdevs[0], bond_grp) || + bond_grp->upper_dev == + get_upper_dev_from_ndev(hr_dev->iboe.netdevs[0])) + return bond_grp; + } + + return NULL; +} + +bool hns_roce_bond_is_active(struct hns_roce_dev *hr_dev) +{ + struct hns_roce_bond_group *bond_grp = hns_roce_get_bond_grp(hr_dev); + + if (bond_grp && + (bond_grp->bond_state == HNS_ROCE_BOND_REGISTERING || + bond_grp->bond_state == HNS_ROCE_BOND_IS_BONDED)) + return true; + + return false; +} + +static inline bool is_active_slave(struct net_device *net_dev, + struct hns_roce_bond_group *bond_grp) +{ + if (!bond_grp || !bond_grp->bond || !bond_grp->bond->curr_active_slave) + return false; + + return net_dev == bond_grp->bond->curr_active_slave->dev; +} + +struct net_device *hns_roce_get_bond_netdev(struct hns_roce_dev *hr_dev) +{ + struct hns_roce_bond_group *bond_grp = hns_roce_get_bond_grp(hr_dev); + struct net_device *net_dev = NULL; + int i; + + if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND)) + return NULL; + + if (!bond_grp) + return NULL; + + mutex_lock(&bond_grp->bond_mutex); + + if (bond_grp->bond_state == HNS_ROCE_BOND_NOT_BONDED) + goto out; + + if (bond_grp->tx_type == NETDEV_LAG_TX_TYPE_ACTIVEBACKUP) { + for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev && is_active_slave(net_dev, bond_grp)) + break; + } + } else { + for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev && get_port_state(net_dev) == IB_PORT_ACTIVE) + break; + } + } + +out: + mutex_unlock(&bond_grp->bond_mutex); + + return net_dev; +} + +static void hns_roce_queue_bond_work(struct hns_roce_bond_group *bond_grp, + unsigned long delay) +{ + schedule_delayed_work(&bond_grp->bond_work, delay); +} + +static void hns_roce_bond_get_active_slave(struct hns_roce_bond_group *bond_grp) +{ + struct net_device *net_dev; + u32 active_slave_map = 0; + u8 active_slave_num = 0; + bool active; + u8 i; + + for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev) { + active = (bond_grp->tx_type == NETDEV_LAG_TX_TYPE_ACTIVEBACKUP) ? + is_active_slave(net_dev, bond_grp) : + (get_port_state(net_dev) == IB_PORT_ACTIVE); + if (active) { + active_slave_num++; + active_slave_map |= (1 << i); + } + } + } + + bond_grp->active_slave_num = active_slave_num; + bond_grp->active_slave_map = active_slave_map; +} + +static void hns_roce_set_bond(struct hns_roce_bond_group *bond_grp) +{ + struct hns_roce_dev *hr_dev = NULL; + struct net_device *net_dev; + int ret; + int i; + + for (i = ROCE_BOND_FUNC_MAX - 1; i >= 0; i--) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev) + hns_roce_bond_uninit_client(bond_grp, i); + } + + bond_grp->bond_state = HNS_ROCE_BOND_REGISTERING; + + for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev) { + hr_dev = hns_roce_bond_init_client(bond_grp, i); + if (hr_dev) { + bond_grp->main_hr_dev = hr_dev; + break; + } + } + } + if (!hr_dev) + return; + + bond_grp->slave_map_diff = 0; + hns_roce_bond_get_active_slave(bond_grp); + ret = hns_roce_cmd_bond(bond_grp, HNS_ROCE_SET_BOND); + if (ret) { + ibdev_err(&hr_dev->ib_dev, "failed to set RoCE bond!\n"); + return; + } + + bond_grp->bond_state = HNS_ROCE_BOND_IS_BONDED; + ibdev_info(&hr_dev->ib_dev, "RoCE set bond finished!\n"); +} + +static void hns_roce_clear_bond(struct hns_roce_bond_group *bond_grp) +{ + u8 main_func_idx = PCI_FUNC(bond_grp->main_hr_dev->pci_dev->devfn); + struct hns_roce_dev *hr_dev = NULL; + struct net_device *net_dev; + int i, ret; + + if (bond_grp->bond_state == HNS_ROCE_BOND_NOT_BONDED) + goto out; + + bond_grp->bond_state = HNS_ROCE_BOND_NOT_BONDED; + bond_grp->main_hr_dev = NULL; + + hns_roce_bond_uninit_client(bond_grp, main_func_idx); + + for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev) { + hr_dev = hns_roce_bond_init_client(bond_grp, i); + if (hr_dev) + bond_grp->main_hr_dev = hr_dev; + } + } + +out: + ret = hns_roce_cleanup_bond(bond_grp); + if (!ret) + ibdev_info(&bond_grp->main_hr_dev->ib_dev, + "RoCE clear bond finished!\n"); +} + +static void hns_roce_slave_changestate(struct hns_roce_bond_group *bond_grp) +{ + int ret; + + hns_roce_bond_get_active_slave(bond_grp); + + ret = hns_roce_cmd_bond(bond_grp, HNS_ROCE_CHANGE_BOND); + if (ret) { + ibdev_err(&bond_grp->main_hr_dev->ib_dev, + "failed to change RoCE bond slave state!\n"); + return; + } + + bond_grp->bond_state = HNS_ROCE_BOND_IS_BONDED; + ibdev_info(&bond_grp->main_hr_dev->ib_dev, + "RoCE slave changestate finished!\n"); +} + +static void hns_roce_slave_inc(struct hns_roce_bond_group *bond_grp) +{ + u32 inc_slave_map = bond_grp->slave_map_diff; + u8 inc_func_idx = 0; + int ret; + + while (inc_slave_map > 0) { + if (inc_slave_map & 1) + hns_roce_bond_uninit_client(bond_grp, inc_func_idx); + inc_slave_map >>= 1; + inc_func_idx++; + } + + bond_grp->slave_map_diff = 0; + hns_roce_bond_get_active_slave(bond_grp); + ret = hns_roce_cmd_bond(bond_grp, HNS_ROCE_CHANGE_BOND); + if (ret) { + ibdev_err(&bond_grp->main_hr_dev->ib_dev, + "failed to increase RoCE bond slave!\n"); + return; + } + + bond_grp->bond_state = HNS_ROCE_BOND_IS_BONDED; + ibdev_info(&bond_grp->main_hr_dev->ib_dev, + "RoCE slave increase finished!\n"); +} + +static void hns_roce_slave_dec(struct hns_roce_bond_group *bond_grp) +{ + u8 main_func_idx = PCI_FUNC(bond_grp->main_hr_dev->pci_dev->devfn); + u32 dec_slave_map = bond_grp->slave_map_diff; + struct hns_roce_dev *hr_dev; + struct net_device *net_dev; + u8 dec_func_idx = 0; + int ret; + int i; + + if (dec_slave_map & (1 << main_func_idx)) { + hns_roce_bond_uninit_client(bond_grp, main_func_idx); + for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (!(dec_slave_map & (1 << i)) && net_dev) { + bond_grp->bond_state = HNS_ROCE_BOND_REGISTERING; + hr_dev = hns_roce_bond_init_client(bond_grp, i); + if (hr_dev) { + bond_grp->main_hr_dev = hr_dev; + break; + } + } + } + } + + while (dec_slave_map > 0) { + if (dec_slave_map & 1) { + bond_grp->bond_func_info[dec_func_idx].net_dev = NULL; + hns_roce_bond_init_client(bond_grp, dec_func_idx); + } + dec_slave_map >>= 1; + dec_func_idx++; + } + + bond_grp->slave_map_diff = 0; + hns_roce_bond_get_active_slave(bond_grp); + ret = hns_roce_cmd_bond(bond_grp, HNS_ROCE_CHANGE_BOND); + if (ret) { + ibdev_err(&bond_grp->main_hr_dev->ib_dev, + "failed to decrease RoCE bond slave!\n"); + return; + } + + bond_grp->bond_state = HNS_ROCE_BOND_IS_BONDED; + ibdev_info(&bond_grp->main_hr_dev->ib_dev, + "RoCE slave decrease finished!\n"); +} + +static void hns_roce_do_bond(struct hns_roce_bond_group *bond_grp) +{ + enum hns_roce_bond_state bond_state = bond_grp->bond_state; + bool bond_ready = bond_grp->bond_ready; + + ibdev_info(&bond_grp->main_hr_dev->ib_dev, + "do_bond: bond_ready - %d, bond_state - %d.\n", + bond_ready, bond_grp->bond_state); + + if (!bond_ready) { + hns_roce_clear_bond(bond_grp); + return; + } + + switch (bond_state) { + case HNS_ROCE_BOND_NOT_BONDED: + hns_roce_set_bond(bond_grp); + return; + case HNS_ROCE_BOND_SLAVE_CHANGESTATE: + hns_roce_slave_changestate(bond_grp); + return; + case HNS_ROCE_BOND_SLAVE_INC: + hns_roce_slave_inc(bond_grp); + return; + case HNS_ROCE_BOND_SLAVE_DEC: + hns_roce_slave_dec(bond_grp); + return; + default: + return; + } +} + +void hns_roce_do_bond_work(struct work_struct *work) +{ + struct delayed_work *delayed_work = to_delayed_work(work); + struct hns_roce_bond_group *bond_grp = + container_of(delayed_work, struct hns_roce_bond_group, + bond_work); + int status; + + status = mutex_trylock(&roce_bond_mutex); + if (!status) { + /* delay 1 sec */ + hns_roce_queue_bond_work(bond_grp, HZ); + return; + } + + hns_roce_do_bond(bond_grp); + mutex_unlock(&roce_bond_mutex); +} + +int hns_roce_bond_init(struct hns_roce_dev *hr_dev) +{ + int ret; + + hr_dev->bond_nb.notifier_call = hns_roce_bond_event; + ret = register_netdevice_notifier(&hr_dev->bond_nb); + if (ret) { + ibdev_err(&hr_dev->ib_dev, + "failed to register notifier for RoCE bond!\n"); + hr_dev->bond_nb.notifier_call = NULL; + } + + return ret; +} + +static struct hns_roce_die_info *alloc_die_info(int bus_num) +{ + struct hns_roce_die_info *die_info; + int ret; + + die_info = kzalloc(sizeof(struct hns_roce_die_info), GFP_KERNEL); + if (!die_info) + return NULL; + + ret = xa_err(xa_store(&roce_bond_xa, bus_num, die_info, GFP_KERNEL)); + if (ret) { + kfree(die_info); + return NULL; + } + + return die_info; +} + +static void dealloc_die_info(struct hns_roce_die_info *die_info, u8 bus_num) +{ + xa_erase(&roce_bond_xa, bus_num); + kvfree(die_info); +} + +static int alloc_bond_id(struct hns_roce_bond_group *bond_grp) +{ + u8 bus_num = bond_grp->bus_num; + struct hns_roce_die_info *die_info = xa_load(&roce_bond_xa, bus_num); + int i; + + if (!die_info) { + die_info = alloc_die_info(bus_num); + if (!die_info) { + ibdev_err(&bond_grp->main_hr_dev->ib_dev, + "failed to alloc die_info.\n"); + return -ENOMEM; + } + } + + for (i = 0; i < ROCE_BOND_NUM_MAX; i++) { + if (die_info->bond_id_mask & BOND_ID(i)) + continue; + + die_info->bond_id_mask |= BOND_ID(i); + die_info->bgrps[i] = bond_grp; + bond_grp->bond_id = i; + + return 0; + } + + return -ENOSPC; +} + +static int remove_bond_id(int bus_num, u8 bond_id) +{ + struct hns_roce_die_info *die_info = xa_load(&roce_bond_xa, bus_num); + + if (bond_id >= ROCE_BOND_NUM_MAX) + return -EINVAL; + + if (!die_info) + return -ENODEV; + + die_info->bond_id_mask &= ~BOND_ID(bond_id); + die_info->bgrps[bond_id] = NULL; + if (!die_info->bond_id_mask) + dealloc_die_info(die_info, bus_num); + + return 0; +} + +int hns_roce_cleanup_bond(struct hns_roce_bond_group *bond_grp) +{ + int ret; + + ret = bond_grp->main_hr_dev ? + hns_roce_cmd_bond(bond_grp, HNS_ROCE_CLEAR_BOND) : -EIO; + if (ret) + ibdev_err(&bond_grp->main_hr_dev->ib_dev, + "failed to clear RoCE bond, ret = %d.\n", ret); + + cancel_delayed_work(&bond_grp->bond_work); + ret = remove_bond_id(bond_grp->bus_num, bond_grp->bond_id); + if (ret) + ibdev_err(&bond_grp->main_hr_dev->ib_dev, + "failed to remove bond ID %d, ret = %d.\n", + bond_grp->bond_id, ret); + kfree(bond_grp); + + return ret; +} + +static bool hns_roce_bond_lowerstate_event(struct hns_roce_dev *hr_dev, + struct hns_roce_bond_group *bond_grp, + struct netdev_notifier_changelowerstate_info *info) +{ + struct net_device *net_dev = + netdev_notifier_info_to_dev((struct netdev_notifier_info *)info); + + if (!netif_is_lag_port(net_dev) || + (!bond_grp || hr_dev != bond_grp->main_hr_dev)) + return false; + + mutex_lock(&bond_grp->bond_mutex); + + if (bond_grp->bond_ready && + bond_grp->bond_state == HNS_ROCE_BOND_IS_BONDED) + bond_grp->bond_state = HNS_ROCE_BOND_SLAVE_CHANGESTATE; + + mutex_unlock(&bond_grp->bond_mutex); + + return true; +} + +static bool is_bond_setting_supported(struct netdev_lag_upper_info *bond_info) +{ + if (!bond_info) + return false; + + if (bond_info->tx_type != NETDEV_LAG_TX_TYPE_ACTIVEBACKUP && + bond_info->tx_type != NETDEV_LAG_TX_TYPE_HASH) + return false; + + if (bond_info->tx_type == NETDEV_LAG_TX_TYPE_HASH && + bond_info->hash_type > NETDEV_LAG_HASH_L23) + return false; + + return true; +} + +static void hns_roce_bond_info_update(struct hns_roce_bond_group *bond_grp, + struct net_device *upper_dev, + bool slave_inc) +{ + struct hns_roce_v2_priv *priv; + struct hns_roce_dev *hr_dev; + struct net_device *net_dev; + u8 func_idx, i; + + if (!slave_inc) { + for (i = 0; i < ROCE_BOND_FUNC_MAX; ++i) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev && upper_dev != + get_upper_dev_from_ndev(net_dev)) { + bond_grp->slave_map_diff |= (1U << i); + bond_grp->slave_map &= ~(1U << i); + } + } + return; + } + + rcu_read_lock(); + for_each_netdev_in_bond_rcu(upper_dev, net_dev) { + hr_dev = hns_roce_get_hrdev_by_netdev(net_dev); + if (hr_dev) { + func_idx = PCI_FUNC(hr_dev->pci_dev->devfn); + if (!bond_grp->bond_func_info[func_idx].net_dev) { + bond_grp->slave_map_diff |= (1U << func_idx); + bond_grp->slave_map |= (1U << func_idx); + priv = hr_dev->priv; + + bond_grp->bond_func_info[func_idx].net_dev = + net_dev; + + bond_grp->bond_func_info[func_idx].handle = + priv->handle; + } + } + } + rcu_read_unlock(); +} + +static bool hns_roce_bond_upper_event(struct hns_roce_bond_group *bond_grp, + struct netdev_notifier_changeupper_info *info) +{ + struct netdev_lag_upper_info *bond_upper_info = NULL; + struct net_device *upper_dev = info->upper_dev; + bool slave_inc = info->linking; + bool changed = false; + + if (!bond_grp || !upper_dev || !netif_is_lag_master(upper_dev)) + return false; + + if (slave_inc) + bond_upper_info = info->upper_info; + + mutex_lock(&bond_grp->bond_mutex); + + if (bond_upper_info) + bond_grp->tx_type = bond_upper_info->tx_type; + + hns_roce_bond_info_update(bond_grp, upper_dev, slave_inc); + + bond_grp->bond = netdev_priv(upper_dev); + + if (bond_grp->bond_state == HNS_ROCE_BOND_NOT_BONDED) { + bond_grp->bond_ready = true; + changed = true; + } else { + bond_grp->bond_state = slave_inc ? + HNS_ROCE_BOND_SLAVE_INC : + HNS_ROCE_BOND_SLAVE_DEC; + bond_grp->bond_ready = true; + changed = true; + } + + mutex_unlock(&bond_grp->bond_mutex); + + return changed; +} + +static struct hns_roce_bond_group *hns_roce_alloc_bond_grp(struct hns_roce_dev *main_hr_dev, + struct net_device *upper_dev) +{ + struct hns_roce_bond_group *bond_grp; + int ret; + + bond_grp = kzalloc(sizeof(*bond_grp), GFP_KERNEL); + if (!bond_grp) + return NULL; + + mutex_init(&bond_grp->bond_mutex); + + INIT_DELAYED_WORK(&bond_grp->bond_work, hns_roce_do_bond_work); + + bond_grp->upper_dev = upper_dev; + bond_grp->main_hr_dev = main_hr_dev; + bond_grp->bond_ready = false; + bond_grp->bond_state = HNS_ROCE_BOND_NOT_BONDED; + bond_grp->bus_num = main_hr_dev->pci_dev->bus->number; + + ret = alloc_bond_id(bond_grp); + if (ret) { + ibdev_err(&main_hr_dev->ib_dev, + "failed to alloc bond ID, ret = %d.\n", ret); + kfree(bond_grp); + return NULL; + } + + hns_roce_bond_info_update(bond_grp, upper_dev, true); + + return bond_grp; +} + +static bool is_dev_bond_supported(struct hns_roce_bond_group *bond_grp, + struct net_device *net_dev, int bus_num) +{ + struct hns_roce_dev *hr_dev = hns_roce_get_hrdev_by_netdev(net_dev); + + if (!hr_dev) { + if (bond_grp && + get_netdev_bond_slave_id(net_dev, bond_grp) >= 0) + return true; + else + return false; + } + + if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND)) + return false; + + if (hr_dev->is_vf || pci_num_vf(hr_dev->pci_dev) > 0) + return false; + + if (bus_num != get_hr_bus_num(hr_dev)) + return false; + + return true; +} + +static bool check_unlinking_bond_support(struct hns_roce_bond_group *bond_grp) +{ + struct net_device *net_dev; + u8 slave_num = 0; + + rcu_read_lock(); + for_each_netdev_in_bond_rcu(bond_grp->upper_dev, net_dev) { + if (get_netdev_bond_slave_id(net_dev, bond_grp) >= 0) + slave_num++; + } + rcu_read_unlock(); + + return (slave_num > 1); +} + +static bool check_linking_bond_support(struct netdev_lag_upper_info *bond_info, + struct hns_roce_bond_group *bond_grp, + struct net_device *upper_dev, + int bus_num) +{ + struct net_device *net_dev; + u8 slave_num = 0; + + if (!is_bond_setting_supported(bond_info)) + return false; + + rcu_read_lock(); + for_each_netdev_in_bond_rcu(upper_dev, net_dev) { + if (is_dev_bond_supported(bond_grp, net_dev, bus_num)) { + slave_num++; + } else { + rcu_read_unlock(); + return false; + } + } + rcu_read_unlock(); + + return (slave_num > 1 && slave_num <= ROCE_BOND_FUNC_MAX); +} + +static enum bond_support_type + check_bond_support(struct hns_roce_dev *hr_dev, + struct net_device **upper_dev, + struct netdev_notifier_changeupper_info *info) +{ + struct hns_roce_bond_group *bond_grp = hns_roce_get_bond_grp(hr_dev); + bool bond_grp_exist = false; + struct net_device *net_dev; + int bus_num = -1; + bool support; + + *upper_dev = info->upper_dev; + if (bond_grp && *upper_dev == bond_grp->upper_dev) + bond_grp_exist = true; + + if (!info->linking && !bond_grp_exist) + return BOND_NOT_SUPPORT; + + if (info->linking) + support = check_linking_bond_support(info->upper_info, bond_grp, + *upper_dev, bus_num); + else + support = check_unlinking_bond_support(bond_grp); + + if (support) + return BOND_SUPPORT; + + return bond_grp_exist ? BOND_EXISTING_NOT_SUPPORT : BOND_NOT_SUPPORT; +} + +int hns_roce_bond_event(struct notifier_block *self, + unsigned long event, void *ptr) +{ + struct net_device *net_dev = netdev_notifier_info_to_dev(ptr); + struct hns_roce_dev *hr_dev = + container_of(self, struct hns_roce_dev, bond_nb); + enum bond_support_type support = BOND_SUPPORT; + struct hns_roce_bond_group *bond_grp; + struct net_device *upper_dev; + bool changed; + + if (event != NETDEV_CHANGEUPPER && event != NETDEV_CHANGELOWERSTATE) + return NOTIFY_DONE; + + if (event == NETDEV_CHANGEUPPER) { + support = check_bond_support(hr_dev, &upper_dev, ptr); + if (support == BOND_NOT_SUPPORT) + return NOTIFY_DONE; + } else { + upper_dev = get_upper_dev_from_ndev(net_dev); + } + + if (upper_dev && !is_hrdev_bond_slave(hr_dev, upper_dev)) + return NOTIFY_DONE; + else if (!upper_dev && hr_dev != hns_roce_get_hrdev_by_netdev(net_dev)) + return NOTIFY_DONE; + + bond_grp = hns_roce_get_bond_grp(hr_dev); + if (event == NETDEV_CHANGEUPPER) { + if (!bond_grp) { + bond_grp = hns_roce_alloc_bond_grp(hr_dev, upper_dev); + if (!bond_grp) { + ibdev_err(&hr_dev->ib_dev, + "failed to alloc RoCE bond_grp!\n"); + return NOTIFY_DONE; + } + } else if (hr_dev != bond_grp->main_hr_dev) { + return NOTIFY_DONE; + } + if (support == BOND_EXISTING_NOT_SUPPORT) { + bond_grp->bond_ready = false; + hns_roce_queue_bond_work(bond_grp, HZ); + return NOTIFY_DONE; + } + changed = hns_roce_bond_upper_event(bond_grp, ptr); + } else { + changed = hns_roce_bond_lowerstate_event(hr_dev, bond_grp, ptr); + } + if (changed) + hns_roce_queue_bond_work(bond_grp, HZ); + + return NOTIFY_DONE; +} diff --git a/drivers/infiniband/hw/hns/hns_roce_bond.h b/drivers/infiniband/hw/hns/hns_roce_bond.h new file mode 100644 index 000000000..94ee5bf36 --- /dev/null +++ b/drivers/infiniband/hw/hns/hns_roce_bond.h @@ -0,0 +1,86 @@ +// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB +/* + * Copyright (c) 2022 Hisilicon Limited. + */ + +#ifndef _HNS_ROCE_BOND_H +#define _HNS_ROCE_BOND_H + +#include <linux/netdevice.h> +#include <net/bonding.h> + +#define ROCE_BOND_FUNC_MAX 4 +#define ROCE_BOND_NUM_MAX 2 + +#define BOND_ID(id) BIT(id) + +enum { + BOND_MODE_1, + BOND_MODE_2_4, +}; + +enum bond_support_type { + BOND_NOT_SUPPORT, + /* + * bond_grp already exists, but in the current + * conditions it's no longer supported + */ + BOND_EXISTING_NOT_SUPPORT, + BOND_SUPPORT, +}; + +enum hns_roce_bond_state { + HNS_ROCE_BOND_NOT_BONDED, + HNS_ROCE_BOND_IS_BONDED, + HNS_ROCE_BOND_REGISTERING, + HNS_ROCE_BOND_SLAVE_INC, + HNS_ROCE_BOND_SLAVE_DEC, + HNS_ROCE_BOND_SLAVE_CHANGESTATE, +}; + +enum hns_roce_bond_cmd_type { + HNS_ROCE_SET_BOND, + HNS_ROCE_CHANGE_BOND, + HNS_ROCE_CLEAR_BOND, +}; + +struct hns_roce_func_info { + struct net_device *net_dev; + struct hnae3_handle *handle; +}; + +struct hns_roce_bond_group { + struct net_device *upper_dev; + struct hns_roce_dev *main_hr_dev; + u8 active_slave_num; + u32 slave_map; + u32 active_slave_map; + u32 slave_map_diff; + u8 bond_id; + u8 bus_num; + struct bonding *bond; + bool bond_ready; + enum hns_roce_bond_state bond_state; + enum netdev_lag_tx_type tx_type; + /* + * A mutex which protect bond_grp info + */ + struct mutex bond_mutex; + struct hns_roce_func_info bond_func_info[ROCE_BOND_FUNC_MAX]; + struct delayed_work bond_work; +}; + +struct hns_roce_die_info { + u8 bond_id_mask; + struct hns_roce_bond_group *bgrps[ROCE_BOND_NUM_MAX]; +}; + +int hns_roce_bond_init(struct hns_roce_dev *hr_dev); +int hns_roce_bond_event(struct notifier_block *self, + unsigned long event, void *ptr); +int hns_roce_cleanup_bond(struct hns_roce_bond_group *bond_grp); +bool hns_roce_bond_is_active(struct hns_roce_dev *hr_dev); +struct net_device *hns_roce_get_bond_netdev(struct hns_roce_dev *hr_dev); +struct hns_roce_bond_group *hns_roce_get_bond_grp(struct hns_roce_dev *hr_dev); + +#endif diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h index 5f69d1f8e..97ac4cac2 100644 --- a/drivers/infiniband/hw/hns/hns_roce_device.h +++ b/drivers/infiniband/hw/hns/hns_roce_device.h @@ -36,6 +36,7 @@ #include <rdma/ib_verbs.h> #include <rdma/hns-abi.h> #include "hns_roce_debugfs.h" +#include "hns_roce_bond.h"

#define PCI_REVISION_ID_HIP08 0x21 #define PCI_REVISION_ID_HIP09 0x30 @@ -147,6 +148,7 @@ enum { HNS_ROCE_CAP_FLAG_SDI_MODE = BIT(14), HNS_ROCE_CAP_FLAG_STASH = BIT(17), HNS_ROCE_CAP_FLAG_CQE_INLINE = BIT(19), + HNS_ROCE_CAP_FLAG_BOND = BIT(21), HNS_ROCE_CAP_FLAG_SRQ_RECORD_DB = BIT(22), };

@@ -969,6 +971,9 @@ struct hns_roce_hw { enum hns_roce_scc_algo algo); int (*query_scc_param)(struct hns_roce_dev *hr_dev, enum hns_roce_scc_algo alog); + int (*bond_init)(struct hns_roce_dev *hr_dev); + bool (*bond_is_active)(struct hns_roce_dev *hr_dev); + struct net_device *(*get_bond_netdev)(struct hns_roce_dev *hr_dev); };

#define HNS_ROCE_SCC_PARAM_SIZE 4 @@ -1045,6 +1050,7 @@ struct hns_roce_dev { struct hns_roce_dev_debugfs dbgfs; atomic64_t *dfx_cnt; struct hns_roce_scc_param *scc_param; + struct notifier_block bond_nb; };

static inline struct hns_roce_dev *to_hr_dev(struct ib_device *ib_dev) @@ -1320,7 +1326,7 @@ void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type); void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type); void hns_roce_handle_device_err(struct hns_roce_dev *hr_dev); int hns_roce_init(struct hns_roce_dev *hr_dev); -void hns_roce_exit(struct hns_roce_dev *hr_dev); +void hns_roce_exit(struct hns_roce_dev *hr_dev, bool bond_cleanup); int hns_roce_fill_res_cq_entry(struct sk_buff *msg, struct ib_cq *ib_cq); int hns_roce_fill_res_cq_entry_raw(struct sk_buff *msg, struct ib_cq *ib_cq); int hns_roce_fill_res_qp_entry(struct sk_buff *msg, struct ib_qp *ib_qp); diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index ec3f78586..d2cdaed84 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -43,6 +43,7 @@ #include <rdma/uverbs_ioctl.h>

#include "hnae3.h" +#include "hclge_main.h" #include "hns_roce_common.h" #include "hns_roce_device.h" #include "hns_roce_cmd.h" @@ -1363,6 +1364,61 @@ static int hns_roce_cmq_send(struct hns_roce_dev *hr_dev, return ret; }

+static inline enum hns_roce_opcode_type + get_bond_opcode(enum hns_roce_bond_cmd_type bond_type) +{ + if (bond_type == HNS_ROCE_SET_BOND) + return HNS_ROCE_OPC_SET_BOND_INFO; + else if (bond_type == HNS_ROCE_CHANGE_BOND) + return HNS_ROCE_OPC_CHANGE_ACTIVE_PORT; + else + return HNS_ROCE_OPC_CLEAR_BOND_INFO; +} + +int hns_roce_cmd_bond(struct hns_roce_bond_group *bond_grp, + enum hns_roce_bond_cmd_type bond_type) +{ + enum hns_roce_opcode_type opcode = get_bond_opcode(bond_type); + struct hns_roce_bond_info *slave_info; + struct hns_roce_cmq_desc desc = {}; + int ret; + + slave_info = (struct hns_roce_bond_info *)desc.data; + hns_roce_cmq_setup_basic_desc(&desc, opcode, false); + + slave_info->bond_id = cpu_to_le32(bond_grp->bond_id); + if (bond_type == HNS_ROCE_CLEAR_BOND) + goto out; + + if (bond_grp->tx_type == NETDEV_LAG_TX_TYPE_ACTIVEBACKUP) { + slave_info->bond_mode = cpu_to_le32(BOND_MODE_1); + if (bond_grp->active_slave_num != 1) + ibdev_err(&bond_grp->main_hr_dev->ib_dev, + "active slave cnt(%d) in Mode 1 is invalid.\n", + bond_grp->active_slave_num); + } else { + slave_info->bond_mode = cpu_to_le32(BOND_MODE_2_4); + slave_info->hash_policy = + cpu_to_le32(bond_grp->bond->params.xmit_policy); + } + + slave_info->active_slave_cnt = + cpu_to_le32(bond_grp->active_slave_num); + slave_info->active_slave_mask = + cpu_to_le32(bond_grp->active_slave_map); + slave_info->slave_mask = + cpu_to_le32(bond_grp->slave_map); + +out: + ret = hns_roce_cmq_send(bond_grp->main_hr_dev, &desc, 1); + if (ret) + ibdev_err(&bond_grp->main_hr_dev->ib_dev, + "cmq bond type(%d) failed, ret = %d.\n", + bond_type, ret); + + return ret; +} + static int config_hem_ba_to_hw(struct hns_roce_dev *hr_dev, dma_addr_t base_addr, u8 cmd, unsigned long tag) { @@ -6842,6 +6898,9 @@ static const struct hns_roce_hw hns_roce_hw_v2 = { .hns_roce_dev_srq_ops = &hns_roce_v2_dev_srq_ops, .config_scc_param = hns_roce_v2_config_scc_param, .query_scc_param = hns_roce_v2_query_scc_param, + .bond_init = hns_roce_bond_init, + .bond_is_active = hns_roce_bond_is_active, + .get_bond_netdev = hns_roce_get_bond_netdev, };

static const struct pci_device_id hns_roce_hw_v2_pci_tbl[] = { @@ -6896,6 +6955,34 @@ static void hns_roce_hw_v2_get_cfg(struct hns_roce_dev *hr_dev, priv->handle = handle; }

+static bool check_vf_support(struct pci_dev *vf) +{ + struct hns_roce_bond_group *bond_grp; + struct pci_dev *pf = pci_physfn(vf); + struct hnae3_ae_dev *ae_dev; + struct hnae3_handle *handle; + struct hns_roce_dev *hr_dev; + struct hclge_dev *hdev; + + /* pf == vf means that the driver is running in VM. */ + if (pf == vf) + return true; + + ae_dev = pci_get_drvdata(pf); + hdev = ae_dev->priv; + handle = &hdev->vport[0].roce; + hr_dev = handle->priv; + + if (!hr_dev) + return false; + + bond_grp = hns_roce_get_bond_grp(hr_dev); + if (bond_grp) + return false; + + return true; +} + static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle) { struct hns_roce_dev *hr_dev; @@ -6913,6 +7000,11 @@ static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)

hns_roce_hw_v2_get_cfg(hr_dev, handle);

+ if (hr_dev->is_vf && !check_vf_support(hr_dev->pci_dev)) { + ret = -EOPNOTSUPP; + goto error_failed_roce_init; + } + ret = hns_roce_init(hr_dev); if (ret) { dev_err(hr_dev->dev, "RoCE Engine init failed!\n"); @@ -6932,7 +7024,7 @@ static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle) return 0;

error_failed_free_mr_init: - hns_roce_exit(hr_dev); + hns_roce_exit(hr_dev, true);

error_failed_roce_init: kfree(hr_dev->priv); @@ -6944,7 +7036,7 @@ error_failed_kzalloc: }

static void __hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle, - bool reset) + bool reset, bool bond_cleanup) { struct hns_roce_dev *hr_dev = handle->priv;

@@ -6959,7 +7051,7 @@ static void __hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle, if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) free_mr_exit(hr_dev);

- hns_roce_exit(hr_dev); + hns_roce_exit(hr_dev, bond_cleanup); kfree(hr_dev->priv); ib_dealloc_device(&hr_dev->ib_dev); } @@ -7015,7 +7107,37 @@ static void hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,

handle->rinfo.instance_state = HNS_ROCE_STATE_UNINIT;

- __hns_roce_hw_v2_uninit_instance(handle, reset); + __hns_roce_hw_v2_uninit_instance(handle, reset, true); + + handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT; +} + +struct hns_roce_dev + *hns_roce_bond_init_client(struct hns_roce_bond_group *bond_grp, + int func_idx) +{ + struct hnae3_handle *handle; + int ret; + + handle = bond_grp->bond_func_info[func_idx].handle; + ret = hns_roce_hw_v2_init_instance(handle); + if (ret) + return NULL; + + return handle->priv; +} + +void hns_roce_bond_uninit_client(struct hns_roce_bond_group *bond_grp, + int func_idx) +{ + struct hnae3_handle *handle = bond_grp->bond_func_info[func_idx].handle; + + if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED) + return; + + handle->rinfo.instance_state = HNS_ROCE_STATE_UNINIT; + + __hns_roce_hw_v2_uninit_instance(handle, false, false);

handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT; } @@ -7080,7 +7202,7 @@ static int hns_roce_hw_v2_reset_notify_uninit(struct hnae3_handle *handle) handle->rinfo.reset_state = HNS_ROCE_STATE_RST_UNINIT; dev_info(&handle->pdev->dev, "In reset process RoCE client uninit.\n"); msleep(HNS_ROCE_V2_HW_RST_UNINT_DELAY); - __hns_roce_hw_v2_uninit_instance(handle, false); + __hns_roce_hw_v2_uninit_instance(handle, false, false);

return 0; } diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h index 3167003c4..074e1f290 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h @@ -226,6 +226,9 @@ enum hns_roce_opcode_type { HNS_ROCE_OPC_CFG_GMV_BT = 0x8510, HNS_ROCE_QUERY_RAM_ECC = 0x8513, HNS_SWITCH_PARAMETER_CFG = 0x1033, + HNS_ROCE_OPC_SET_BOND_INFO = 0x8601, + HNS_ROCE_OPC_CLEAR_BOND_INFO = 0x8602, + HNS_ROCE_OPC_CHANGE_ACTIVE_PORT = 0x8603, };

enum { @@ -1571,7 +1574,23 @@ struct hns_roce_sccc_clr_done { __le32 rsv[5]; };

+struct hns_roce_bond_info { + __le32 bond_id; + __le32 bond_mode; + __le32 active_slave_cnt; + __le32 active_slave_mask; + __le32 slave_mask; + __le32 hash_policy; +}; + +struct hns_roce_dev + *hns_roce_bond_init_client(struct hns_roce_bond_group *bond_grp, + int func_idx); +void hns_roce_bond_uninit_client(struct hns_roce_bond_group *bond_grp, + int func_idx); int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata); +int hns_roce_cmd_bond(struct hns_roce_bond_group *bond_grp, + enum hns_roce_bond_cmd_type bond_type);

static inline void hns_roce_write64(struct hns_roce_dev *hr_dev, __le32 val[2], void __iomem *dest) diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c index 4e9ef6f02..68bd08c87 100644 --- a/drivers/infiniband/hw/hns/hns_roce_main.c +++ b/drivers/infiniband/hw/hns/hns_roce_main.c @@ -37,9 +37,36 @@ #include <rdma/ib_smi.h> #include <rdma/ib_user_verbs.h> #include <rdma/ib_cache.h> + +#include "hnae3.h" #include "hns_roce_common.h" #include "hns_roce_device.h" #include "hns_roce_hem.h" +#include "hns_roce_hw_v2.h" + +static struct net_device *hns_roce_get_netdev(struct ib_device *ib_dev, + u32 port_num) +{ + struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev); + struct net_device *ndev; + + if (port_num < 1 || port_num > hr_dev->caps.num_ports) + return NULL; + + ndev = hr_dev->hw->get_bond_netdev(hr_dev); + + rcu_read_lock(); + + if (!ndev) + ndev = hr_dev->iboe.netdevs[port_num - 1]; + + if (ndev) + dev_hold(ndev); + + rcu_read_unlock(); + + return ndev; +}

static int hns_roce_set_mac(struct hns_roce_dev *hr_dev, u32 port, const u8 *addr) @@ -262,7 +289,9 @@ static int hns_roce_query_port(struct ib_device *ib_dev, u32 port_num,

spin_lock_irqsave(&hr_dev->iboe.lock, flags);

- net_dev = hr_dev->iboe.netdevs[port]; + net_dev = hr_dev->hw->get_bond_netdev(hr_dev); + if (!net_dev) + net_dev = hr_dev->iboe.netdevs[port]; if (!net_dev) { spin_unlock_irqrestore(&hr_dev->iboe.lock, flags); dev_err(dev, "find netdev %u failed!\n", port); @@ -615,9 +644,18 @@ static int hns_roce_get_hw_stats(struct ib_device *device, return num_counters; }

-static void hns_roce_unregister_device(struct hns_roce_dev *hr_dev) +static void hns_roce_unregister_device(struct hns_roce_dev *hr_dev, + bool bond_cleanup) { struct hns_roce_ib_iboe *iboe = &hr_dev->iboe; + struct hns_roce_bond_group *bond_grp; + + if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND) { + unregister_netdevice_notifier(&hr_dev->bond_nb); + bond_grp = hns_roce_get_bond_grp(hr_dev); + if (bond_grp && bond_cleanup) + hns_roce_cleanup_bond(bond_grp); + }

hr_dev->active = false; unregister_netdevice_notifier(&iboe->nb); @@ -647,6 +685,7 @@ static const struct ib_device_ops hns_roce_dev_ops = { .disassociate_ucontext = hns_roce_disassociate_ucontext, .get_dma_mr = hns_roce_get_dma_mr, .get_link_layer = hns_roce_get_link_layer, + .get_netdev = hns_roce_get_netdev, .get_port_immutable = hns_roce_port_immutable, .mmap = hns_roce_mmap, .mmap_free = hns_roce_free_mmap, @@ -766,7 +805,12 @@ static int hns_roce_register_device(struct hns_roce_dev *hr_dev) return ret; } dma_set_max_seg_size(dev, UINT_MAX); - ret = ib_register_device(ib_dev, "hns_%d", dev); + + if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND) && + (hr_dev->hw->bond_is_active(hr_dev))) + ret = ib_register_device(ib_dev, "hns_bond_%d", dev); + else + ret = ib_register_device(ib_dev, "hns_%d", dev); if (ret) { dev_err(dev, "ib_register_device failed!\n"); return ret; @@ -785,9 +829,26 @@ static int hns_roce_register_device(struct hns_roce_dev *hr_dev) goto error_failed_setup_mtu_mac; }

+ if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND) { + ret = hr_dev->hw->bond_init(hr_dev); + if (ret) { + dev_err(dev, "roce bond init failed, ret = %d\n", ret); + /* For non-bond devices, the failure of bond_init does + * not affect other functions. + */ + if (hr_dev->hw->bond_is_active(hr_dev)) + goto error_bond_init; + else + ret = 0; + } + } + hr_dev->active = true; - return 0;

+ return ret; + +error_bond_init: + unregister_netdevice_notifier(&iboe->nb); error_failed_setup_mtu_mac: ib_unregister_device(ib_dev);

@@ -1164,11 +1225,11 @@ error_failed_alloc_dfx_cnt: return ret; }

-void hns_roce_exit(struct hns_roce_dev *hr_dev) +void hns_roce_exit(struct hns_roce_dev *hr_dev, bool bond_cleanup) { hns_roce_unregister_sysfs(hr_dev); + hns_roce_unregister_device(hr_dev, bond_cleanup); hns_roce_unregister_debugfs(hr_dev); - hns_roce_unregister_device(hr_dev);

if (hr_dev->hw->hw_exit) hr_dev->hw->hw_exit(hr_dev);

From: Junxian Huang huangjunxian6@hisilicon.com

driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I968IB

---------------------------------------------------------------

Support dispatching IB event for RoCE bonding. After setting bond, IB_EVENT_PORT_ERR is dispatched in the following situation: 1. bond0 becomes link down; 2. all slaves become link down (as it will lead to a link-down of the upper device).

IB_EVENT_PORT_ACTIVE is dispatched in the following situation: 1. bond0 becomes link up; 2. one slave becomes link up when all slaves were link down (as it will lead to a link-up of the upper device).

Signed-off-by: Junxian Huang huangjunxian6@hisilicon.com Signed-off-by: Juan Zhou zhoujuan51@h-partners.com --- drivers/infiniband/hw/hns/hns_roce_bond.c | 17 +++++++++-------- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 10 ++++++++++ drivers/infiniband/hw/hns/hns_roce_main.c | 18 ++++++++++++------ 3 files changed, 31 insertions(+), 14 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_bond.c b/drivers/infiniband/hw/hns/hns_roce_bond.c index d6ebf41c9..fc179ea52 100644 --- a/drivers/infiniband/hw/hns/hns_roce_bond.c +++ b/drivers/infiniband/hw/hns/hns_roce_bond.c @@ -138,17 +138,18 @@ struct net_device *hns_roce_get_bond_netdev(struct hns_roce_dev *hr_dev) if (bond_grp->tx_type == NETDEV_LAG_TX_TYPE_ACTIVEBACKUP) { for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { net_dev = bond_grp->bond_func_info[i].net_dev; - if (net_dev && is_active_slave(net_dev, bond_grp)) - break; - } - } else { - for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { - net_dev = bond_grp->bond_func_info[i].net_dev; - if (net_dev && get_port_state(net_dev) == IB_PORT_ACTIVE) - break; + if (net_dev && is_active_slave(net_dev, bond_grp) && + get_port_state(net_dev) == IB_PORT_ACTIVE) + goto out; } }

+ for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev && get_port_state(net_dev) == IB_PORT_ACTIVE) + break; + } + out: mutex_unlock(&bond_grp->bond_mutex);

diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index d2cdaed84..de762e4d5 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -7234,6 +7234,7 @@ static void hns_roce_hw_v2_link_status_change(struct hnae3_handle *handle, { struct net_device *netdev = handle->rinfo.netdev; struct hns_roce_dev *hr_dev = handle->priv; + struct hns_roce_bond_group *bond_grp; struct ib_event event; unsigned long flags; u8 phy_port; @@ -7241,6 +7242,15 @@ static void hns_roce_hw_v2_link_status_change(struct hnae3_handle *handle, if (linkup || !hr_dev) return;

+ /* For bond device, the link status depends on the upper netdev, + * and the upper device's link status depends on all the slaves' + * netdev but not only one. So bond device cannot get a correct + * link status from this path. + */ + bond_grp = hns_roce_get_bond_grp(hr_dev); + if (bond_grp) + return; + for (phy_port = 0; phy_port < hr_dev->caps.num_ports; phy_port++) if (netdev == hr_dev->iboe.netdevs[phy_port]) break; diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c index d35ccaf56..6e8196127 100644 --- a/drivers/infiniband/hw/hns/hns_roce_main.c +++ b/drivers/infiniband/hw/hns/hns_roce_main.c @@ -115,17 +115,16 @@ static int hns_roce_del_gid(const struct ib_gid_attr *attr, void **context) return ret; }

-static int handle_en_event(struct hns_roce_dev *hr_dev, u32 port, - unsigned long dev_event) +static int handle_en_event(struct net_device *netdev, + struct hns_roce_dev *hr_dev, + u32 port, unsigned long dev_event) { struct device *dev = hr_dev->dev; enum ib_port_state port_state; - struct net_device *netdev; struct ib_event event; unsigned long flags; int ret = 0;

- netdev = hr_dev->iboe.netdevs[port]; if (!netdev) { dev_err(dev, "can't find netdev on port(%u)!\n", port); return -ENODEV; @@ -173,17 +172,24 @@ static int hns_roce_netdev_event(struct notifier_block *self, unsigned long event, void *ptr) { struct net_device *dev = netdev_notifier_info_to_dev(ptr); + struct hns_roce_bond_group *bond_grp; struct hns_roce_ib_iboe *iboe = NULL; struct hns_roce_dev *hr_dev = NULL; + struct net_device *upper = NULL; int ret; u32 port;

hr_dev = container_of(self, struct hns_roce_dev, iboe.nb); iboe = &hr_dev->iboe; + if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND) { + bond_grp = hns_roce_get_bond_grp(hr_dev); + upper = bond_grp ? bond_grp->upper_dev : NULL; + }

for (port = 0; port < hr_dev->caps.num_ports; port++) { - if (dev == iboe->netdevs[port]) { - ret = handle_en_event(hr_dev, port, event); + if ((!upper && dev == iboe->netdevs[port]) || + (upper && dev == upper)) { + ret = handle_en_event(dev, hr_dev, port, event); if (ret) return NOTIFY_DONE; break;

From: Junxian Huang huangjunxian6@hisilicon.com

driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I968IB

---------------------------------------------------------------

Currently, a RoCE bond device cannot be recovered to a bond device after reset.

Applying this patch, RoCE bonding device 'hns_bond_xx' can be recovered after reset, including the following changes: 1. modify the condition for juding whether bond_grp is active, as the bond_grp may be also holding HNS_ROCE_CHANGE_BOND during reset init. Thus, as long as the bond_grp's state is not HNS_ROCE_BOND_NOT_BONDED, it should be considered active. 2. update the link status of slave in bond_grp from NIC bonding driver right before sending command to firmware, as RoCE driver is uninited for a while in reset process, and during this period bond_grp cannot update the information. 3. After the reset, re-config the bond_grp information to firmware, as the firmware is also reset and the previous configuration is cleared.

Signed-off-by: Junxian Huang huangjunxian6@hisilicon.com Signed-off-by: Juan Zhou zhoujuan51@h-partners.com --- drivers/infiniband/hw/hns/hns_roce_bond.c | 32 ++++++++++++++++++++--- drivers/infiniband/hw/hns/hns_roce_main.c | 10 +++++-- 2 files changed, 36 insertions(+), 6 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_bond.c b/drivers/infiniband/hw/hns/hns_roce_bond.c index a35b982ec..6fdd8e9fd 100644 --- a/drivers/infiniband/hw/hns/hns_roce_bond.c +++ b/drivers/infiniband/hw/hns/hns_roce_bond.c @@ -100,9 +100,7 @@ bool hns_roce_bond_is_active(struct hns_roce_dev *hr_dev) { struct hns_roce_bond_group *bond_grp = hns_roce_get_bond_grp(hr_dev);

- if (bond_grp && - (bond_grp->bond_state == HNS_ROCE_BOND_REGISTERING || - bond_grp->bond_state == HNS_ROCE_BOND_IS_BONDED)) + if (bond_grp && bond_grp->bond_state != HNS_ROCE_BOND_NOT_BONDED) return true;

return false; @@ -186,6 +184,13 @@ static void hns_roce_bond_get_active_slave(struct hns_roce_bond_group *bond_grp) bond_grp->active_slave_map = active_slave_map; }

+static int hns_roce_recover_bond(struct hns_roce_bond_group *bond_grp) +{ + hns_roce_bond_get_active_slave(bond_grp); + + return hns_roce_cmd_bond(bond_grp, HNS_ROCE_SET_BOND); +} + static void hns_roce_set_bond(struct hns_roce_bond_group *bond_grp) { struct hns_roce_dev *hr_dev = NULL; @@ -355,6 +360,9 @@ static void hns_roce_do_bond(struct hns_roce_bond_group *bond_grp) enum hns_roce_bond_state bond_state = bond_grp->bond_state; bool bond_ready = bond_grp->bond_ready;

+ if (!bond_grp->main_hr_dev) + return; + ibdev_info(&bond_grp->main_hr_dev->ib_dev, "do_bond: bond_ready - %d, bond_state - %d.\n", bond_ready, bond_grp->bond_state); @@ -403,13 +411,29 @@ void hns_roce_do_bond_work(struct work_struct *work)

int hns_roce_bond_init(struct hns_roce_dev *hr_dev) { + struct hns_roce_bond_group *bond_grp = hns_roce_get_bond_grp(hr_dev); + struct hns_roce_v2_priv *priv = hr_dev->priv; int ret;

+ if (priv->handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT && + bond_grp) { + bond_grp->main_hr_dev = hr_dev; + ret = hns_roce_recover_bond(bond_grp); + if (ret) { + ibdev_err(&hr_dev->ib_dev, + "failed to recover RoCE bond, ret = %d.\n", + ret); + return ret; + } + bond_grp->bond_state = HNS_ROCE_BOND_IS_BONDED; + } + hr_dev->bond_nb.notifier_call = hns_roce_bond_event; ret = register_netdevice_notifier(&hr_dev->bond_nb); if (ret) { ibdev_err(&hr_dev->ib_dev, - "failed to register notifier for RoCE bond!\n"); + "failed to register notifier for RoCE bond, ret = %d.\n", + ret); hr_dev->bond_nb.notifier_call = NULL; }

diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c index 3f4316d71..22835c331 100644 --- a/drivers/infiniband/hw/hns/hns_roce_main.c +++ b/drivers/infiniband/hw/hns/hns_roce_main.c @@ -671,13 +671,19 @@ static void hns_roce_unregister_device(struct hns_roce_dev *hr_dev, bool bond_cleanup) { struct hns_roce_ib_iboe *iboe = &hr_dev->iboe; + struct hns_roce_v2_priv *priv = hr_dev->priv; struct hns_roce_bond_group *bond_grp;

if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND) { unregister_netdevice_notifier(&hr_dev->bond_nb); bond_grp = hns_roce_get_bond_grp(hr_dev); - if (bond_grp && bond_cleanup) - hns_roce_cleanup_bond(bond_grp); + if (bond_grp) { + if (bond_cleanup) + hns_roce_cleanup_bond(bond_grp); + else if (priv->handle->rinfo.reset_state == + HNS_ROCE_STATE_RST_UNINIT) + bond_grp->main_hr_dev = NULL; + } }

hr_dev->active = false;

From: Junxian Huang huangjunxian6@hisilicon.com

driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I968IB

--------------------------------------------------------------------------

When the resource pf of RoCE bonding is unbinded, the main_hr_dev will be unregistered and the bond resources will be cleaned up. Currently, other slaves will not be re-initialized, and they are not availbale until the whole RoCE ko is removed and inserted again.

To fix this problem, re-initialized all the slaves to hns_* device except the resource slave itself before bond resources cleanup.

Fixes: e62a20278f18 ("RDMA/hns: support RoCE bonding") Signed-off-by: Junxian Huang huangjunxian6@hisilicon.com Signed-off-by: Juan Zhou zhoujuan51@h-partners.com --- drivers/infiniband/hw/hns/hns_roce_bond.c | 22 +++++++++++---- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 3 ++ drivers/infiniband/hw/hns/hns_roce_main.c | 33 ++++++++++++++++------ 3 files changed, 43 insertions(+), 15 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_bond.c b/drivers/infiniband/hw/hns/hns_roce_bond.c index 38f2326f9..911a2ba28 100644 --- a/drivers/infiniband/hw/hns/hns_roce_bond.c +++ b/drivers/infiniband/hw/hns/hns_roce_bond.c @@ -37,19 +37,19 @@ static struct net_device *get_upper_dev_from_ndev(struct net_device *net_dev) return upper_dev; }

-static bool is_netdev_bond_slave(struct net_device *net_dev, - struct hns_roce_bond_group *bond_grp) +static int get_netdev_bond_slave_id(struct net_device *net_dev, + struct hns_roce_bond_group *bond_grp) { int i;

if (!net_dev || !bond_grp) - return false; + return -ENODEV;

for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) if (net_dev == bond_grp->bond_func_info[i].net_dev) - return true; + return i;

- return false; + return -ENOENT; }

static bool is_hrdev_bond_slave(struct hns_roce_dev *hr_dev, @@ -92,7 +92,7 @@ struct hns_roce_bond_group *hns_roce_get_bond_grp(struct net_device *net_dev, bond_grp = die_info->bgrps[i]; if (!bond_grp) continue; - if (is_netdev_bond_slave(net_dev, bond_grp) || + if (get_netdev_bond_slave_id(net_dev, bond_grp) >= 0 || (bond_grp->upper_dev == get_upper_dev_from_ndev(net_dev))) return bond_grp; } @@ -823,6 +823,7 @@ int hns_roce_bond_event(struct notifier_block *self, u8 bus_num = get_hr_bus_num(hr_dev); struct net_device *upper_dev; bool changed; + int slave_id;

if (event != NETDEV_CHANGEUPPER && event != NETDEV_CHANGELOWERSTATE) return NOTIFY_DONE; @@ -852,6 +853,15 @@ int hns_roce_bond_event(struct notifier_block *self, } else if (hr_dev != bond_grp->main_hr_dev) { return NOTIFY_DONE; } + /* In the case of netdev being unregistered, the roce + * instance shouldn't be inited. + */ + if (net_dev->reg_state >= NETREG_UNREGISTERING) { + slave_id = get_netdev_bond_slave_id(net_dev, bond_grp); + if (slave_id >= 0) + bond_grp->bond_func_info[slave_id].handle = NULL; + } + if (support == BOND_EXISTING_NOT_SUPPORT) { bond_grp->bond_ready = false; hns_roce_queue_bond_work(bond_grp, HZ); diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index a05099aed..2153590d0 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -7134,6 +7134,9 @@ struct hns_roce_dev int ret;

handle = bond_grp->bond_func_info[func_idx].handle; + if (!handle || !handle->client) + return NULL; + ret = hns_roce_hw_v2_init_instance(handle); if (ret) return NULL; diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c index 6960d5b2a..c3d057222 100644 --- a/drivers/infiniband/hw/hns/hns_roce_main.c +++ b/drivers/infiniband/hw/hns/hns_roce_main.c @@ -678,19 +678,34 @@ static void hns_roce_unregister_device(struct hns_roce_dev *hr_dev, struct hns_roce_v2_priv *priv = hr_dev->priv; struct hns_roce_bond_group *bond_grp; u8 bus_num = get_hr_bus_num(hr_dev); + int i;

- if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND) { - unregister_netdevice_notifier(&hr_dev->bond_nb); - bond_grp = hns_roce_get_bond_grp(net_dev, bus_num); - if (bond_grp) { - if (bond_cleanup) - hns_roce_cleanup_bond(bond_grp); - else if (priv->handle->rinfo.reset_state == - HNS_ROCE_STATE_RST_UNINIT) - bond_grp->main_hr_dev = NULL; + if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_BOND)) + goto normal_unregister; + + unregister_netdevice_notifier(&hr_dev->bond_nb); + bond_grp = hns_roce_get_bond_grp(net_dev, bus_num); + if (!bond_grp) + goto normal_unregister; + + if (bond_cleanup) { + /* To avoid the loss of other slave devices when main_hr_dev + * is unregistered, re-initialized the remaining slaves before + * the bond resources cleanup. + */ + bond_grp->bond_state = HNS_ROCE_BOND_NOT_BONDED; + for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { + net_dev = bond_grp->bond_func_info[i].net_dev; + if (net_dev && net_dev != iboe->netdevs[0]) + hns_roce_bond_init_client(bond_grp, i); } + hns_roce_cleanup_bond(bond_grp); + } else if (priv->handle->rinfo.reset_state == + HNS_ROCE_STATE_RST_UNINIT) { + bond_grp->main_hr_dev = NULL; }

+normal_unregister: hr_dev->active = false; unregister_netdevice_notifier(&iboe->nb); ib_unregister_device(&hr_dev->ib_dev);