linux/include/rdma/rdma_vt.h

533 lines
14 KiB
C
Raw Normal View History

/* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
/*
* Copyright(c) 2016 - 2019 Intel Corporation.
*/
#ifndef DEF_RDMA_VT_H
#define DEF_RDMA_VT_H
/*
* Structure that low level drivers will populate in order to register with the
* rdmavt layer.
*/
#include <linux/spinlock.h>
#include <linux/list.h>
#include <linux/hash.h>
#include <rdma/ib_verbs.h>
#include <rdma/ib_mad.h>
#include <rdma/rdmavt_mr.h>
#define RVT_MAX_PKEY_VALUES 16
#define RVT_MAX_TRAP_LEN 100 /* Limit pending trap list */
#define RVT_MAX_TRAP_LISTS 5 /*((IB_NOTICE_TYPE_INFO & 0x0F) + 1)*/
#define RVT_TRAP_TIMEOUT 4096 /* 4.096 usec */
struct trap_list {
u32 list_len;
struct list_head list;
};
struct rvt_qp;
struct rvt_qpn_table;
struct rvt_ibport {
struct rvt_qp __rcu *qp[2];
struct ib_mad_agent *send_agent; /* agent for SMI (traps) */
struct rb_root mcast_tree;
spinlock_t lock; /* protect changes in this struct */
/* non-zero when timer is set */
unsigned long mkey_lease_timeout;
unsigned long trap_timeout;
__be64 gid_prefix; /* in network order */
__be64 mkey;
u64 tid;
u32 port_cap_flags;
u16 port_cap3_flags;
u32 pma_sample_start;
u32 pma_sample_interval;
__be16 pma_counter_select[5];
u16 pma_tag;
u16 mkey_lease_period;
u32 sm_lid;
u8 sm_sl;
u8 mkeyprot;
u8 subnet_timeout;
u8 vl_high_limit;
/*
* Driver is expected to keep these up to date. These
* counters are informational only and not required to be
* completely accurate.
*/
u64 n_rc_resends;
u64 n_seq_naks;
u64 n_rdma_seq;
u64 n_rnr_naks;
u64 n_other_naks;
u64 n_loop_pkts;
u64 n_pkt_drops;
u64 n_vl15_dropped;
u64 n_rc_timeouts;
u64 n_dmawait;
u64 n_unaligned;
u64 n_rc_dupreq;
u64 n_rc_seqnak;
u64 n_rc_crwaits;
u16 pkey_violations;
u16 qkey_violations;
u16 mkey_violations;
/* Hot-path per CPU counters to avoid cacheline trading to update */
u64 z_rc_acks;
u64 z_rc_qacks;
u64 z_rc_delayed_comp;
u64 __percpu *rc_acks;
u64 __percpu *rc_qacks;
u64 __percpu *rc_delayed_comp;
void *priv; /* driver private data */
/*
* The pkey table is allocated and maintained by the driver. Drivers
* need to have access to this before registering with rdmav. However
* rdmavt will need access to it so drivers need to provide this during
* the attach port API call.
*/
u16 *pkey_table;
struct rvt_ah *sm_ah;
/*
* Keep a list of traps that have not been repressed. They will be
* resent based on trap_timer.
*/
struct trap_list trap_lists[RVT_MAX_TRAP_LISTS];
struct timer_list trap_timer;
};
#define RVT_CQN_MAX 16 /* maximum length of cq name */
#define RVT_SGE_COPY_MEMCPY 0
#define RVT_SGE_COPY_CACHELESS 1
#define RVT_SGE_COPY_ADAPTIVE 2
/*
* Things that are driver specific, module parameters in hfi1 and qib
*/
struct rvt_driver_params {
struct ib_device_attr props;
/*
* Anything driver specific that is not covered by props
* For instance special module parameters. Goes here.
*/
unsigned int lkey_table_size;
unsigned int qp_table_size;
unsigned int sge_copy_mode;
unsigned int wss_threshold;
unsigned int wss_clean_period;
int qpn_start;
int qpn_inc;
int qpn_res_start;
int qpn_res_end;
int nports;
int npkeys;
int node;
int psn_mask;
int psn_shift;
int psn_modify_mask;
u32 core_cap_flags;
u32 max_mad_size;
u8 qos_shift;
u8 max_rdma_atomic;
u8 extra_rdma_atomic;
u8 reserved_operations;
};
/* User context */
struct rvt_ucontext {
struct ib_ucontext ibucontext;
};
/* Protection domain */
struct rvt_pd {
struct ib_pd ibpd;
bool user;
};
/* Address handle */
struct rvt_ah {
struct ib_ah ibah;
struct rdma_ah_attr attr;
u8 vl;
u8 log_pmtu;
};
/*
* This structure is used by rvt_mmap() to validate an offset
* when an mmap() request is made. The vm_area_struct then uses
* this as its vm_private_data.
*/
struct rvt_mmap_info {
struct list_head pending_mmaps;
struct ib_ucontext *context;
void *obj;
__u64 offset;
struct kref ref;
u32 size;
};
/* memory working set size */
struct rvt_wss {
unsigned long *entries;
atomic_t total_count;
atomic_t clean_counter;
atomic_t clean_entry;
int threshold;
int num_entries;
long pages_mask;
unsigned int clean_period;
};
struct rvt_dev_info;
struct rvt_swqe;
struct rvt_driver_provided {
/*
* Which functions are required depends on which verbs rdmavt is
* providing and which verbs the driver is overriding. See
* check_support() for details.
*/
/* hot path calldowns in a single cacheline */
/*
* Give the driver a notice that there is send work to do. It is up to
* the driver to generally push the packets out, this just queues the
* work with the driver. There are two variants here. The no_lock
* version requires the s_lock not to be held. The other assumes the
* s_lock is held.
*/
bool (*schedule_send)(struct rvt_qp *qp);
bool (*schedule_send_no_lock)(struct rvt_qp *qp);
/*
* Driver specific work request setup and checking.
* This function is allowed to perform any setup, checks, or
* adjustments required to the SWQE in order to be usable by
* underlying protocols. This includes private data structure
* allocations.
*/
int (*setup_wqe)(struct rvt_qp *qp, struct rvt_swqe *wqe,
bool *call_send);
/*
* Sometimes rdmavt needs to kick the driver's send progress. That is
* done by this call back.
*/
void (*do_send)(struct rvt_qp *qp);
/*
* Returns a pointer to the underlying hardware's PCI device. This is
* used to display information as to what hardware is being referenced
* in an output message
*/
struct pci_dev * (*get_pci_dev)(struct rvt_dev_info *rdi);
/*
* Allocate a private queue pair data structure for driver specific
* information which is opaque to rdmavt. Errors are returned via
* ERR_PTR(err). The driver is free to return NULL or a valid
* pointer.
*/
void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
/*
* Init a structure allocated with qp_priv_alloc(). This should be
* called after all qp fields have been initialized in rdmavt.
*/
int (*qp_priv_init)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
struct ib_qp_init_attr *init_attr);
/*
* Free the driver's private qp structure.
*/
void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
/*
* Inform the driver the particular qp in question has been reset so
* that it can clean up anything it needs to.
*/
void (*notify_qp_reset)(struct rvt_qp *qp);
/*
* Get a path mtu from the driver based on qp attributes.
*/
int (*get_pmtu_from_attr)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
struct ib_qp_attr *attr);
/*
* Notify driver that it needs to flush any outstanding IO requests that
* are waiting on a qp.
*/
void (*flush_qp_waiters)(struct rvt_qp *qp);
/*
* Notify driver to stop its queue of sending packets. Nothing else
* should be posted to the queue pair after this has been called.
*/
void (*stop_send_queue)(struct rvt_qp *qp);
/*
* Have the driver drain any in progress operations
*/
void (*quiesce_qp)(struct rvt_qp *qp);
/*
* Inform the driver a qp has went to error state.
*/
void (*notify_error_qp)(struct rvt_qp *qp);
/*
* Get an MTU for a qp.
*/
u32 (*mtu_from_qp)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
u32 pmtu);
/*
* Convert an mtu to a path mtu
*/
int (*mtu_to_path_mtu)(u32 mtu);
/*
* Get the guid of a port in big endian byte order
*/
int (*get_guid_be)(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
int guid_index, __be64 *guid);
/*
* Query driver for the state of the port.
*/
RDMA: Support more than 255 rdma ports Current code uses many different types when dealing with a port of a RDMA device: u8, unsigned int and u32. Switch to u32 to clean up the logic. This allows us to make (at least) the core view consistent and use the same type. Unfortunately not all places can be converted. Many uverbs functions expect port to be u8 so keep those places in order not to break UAPIs. HW/Spec defined values must also not be changed. With the switch to u32 we now can support devices with more than 255 ports. U32_MAX is reserved to make control logic a bit easier to deal with. As a device with U32_MAX ports probably isn't going to happen any time soon this seems like a non issue. When a device with more than 255 ports is created uverbs will report the RDMA device as having 255 ports as this is the max currently supported. The verbs interface is not changed yet because the IBTA spec limits the port size in too many places to be u8 and all applications that relies in verbs won't be able to cope with this change. At this stage, we are extending the interfaces that are using vendor channel solely Once the limitation is lifted mlx5 in switchdev mode will be able to have thousands of SFs created by the device. As the only instance of an RDMA device that reports more than 255 ports will be a representor device and it exposes itself as a RAW Ethernet only device CM/MAD/IPoIB and other ULPs aren't effected by this change and their sysfs/interfaces that are exposes to userspace can remain unchanged. While here cleanup some alignment issues and remove unneeded sanity checks (mainly in rdmavt), Link: https://lore.kernel.org/r/20210301070420.439400-1-leon@kernel.org Signed-off-by: Mark Bloch <mbloch@nvidia.com> Signed-off-by: Leon Romanovsky <leonro@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
2021-03-01 07:04:20 +00:00
int (*query_port_state)(struct rvt_dev_info *rdi, u32 port_num,
struct ib_port_attr *props);
/*
* Tell driver to shutdown a port
*/
RDMA: Support more than 255 rdma ports Current code uses many different types when dealing with a port of a RDMA device: u8, unsigned int and u32. Switch to u32 to clean up the logic. This allows us to make (at least) the core view consistent and use the same type. Unfortunately not all places can be converted. Many uverbs functions expect port to be u8 so keep those places in order not to break UAPIs. HW/Spec defined values must also not be changed. With the switch to u32 we now can support devices with more than 255 ports. U32_MAX is reserved to make control logic a bit easier to deal with. As a device with U32_MAX ports probably isn't going to happen any time soon this seems like a non issue. When a device with more than 255 ports is created uverbs will report the RDMA device as having 255 ports as this is the max currently supported. The verbs interface is not changed yet because the IBTA spec limits the port size in too many places to be u8 and all applications that relies in verbs won't be able to cope with this change. At this stage, we are extending the interfaces that are using vendor channel solely Once the limitation is lifted mlx5 in switchdev mode will be able to have thousands of SFs created by the device. As the only instance of an RDMA device that reports more than 255 ports will be a representor device and it exposes itself as a RAW Ethernet only device CM/MAD/IPoIB and other ULPs aren't effected by this change and their sysfs/interfaces that are exposes to userspace can remain unchanged. While here cleanup some alignment issues and remove unneeded sanity checks (mainly in rdmavt), Link: https://lore.kernel.org/r/20210301070420.439400-1-leon@kernel.org Signed-off-by: Mark Bloch <mbloch@nvidia.com> Signed-off-by: Leon Romanovsky <leonro@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
2021-03-01 07:04:20 +00:00
int (*shut_down_port)(struct rvt_dev_info *rdi, u32 port_num);
/* Tell driver to send a trap for changed port capabilities */
RDMA: Support more than 255 rdma ports Current code uses many different types when dealing with a port of a RDMA device: u8, unsigned int and u32. Switch to u32 to clean up the logic. This allows us to make (at least) the core view consistent and use the same type. Unfortunately not all places can be converted. Many uverbs functions expect port to be u8 so keep those places in order not to break UAPIs. HW/Spec defined values must also not be changed. With the switch to u32 we now can support devices with more than 255 ports. U32_MAX is reserved to make control logic a bit easier to deal with. As a device with U32_MAX ports probably isn't going to happen any time soon this seems like a non issue. When a device with more than 255 ports is created uverbs will report the RDMA device as having 255 ports as this is the max currently supported. The verbs interface is not changed yet because the IBTA spec limits the port size in too many places to be u8 and all applications that relies in verbs won't be able to cope with this change. At this stage, we are extending the interfaces that are using vendor channel solely Once the limitation is lifted mlx5 in switchdev mode will be able to have thousands of SFs created by the device. As the only instance of an RDMA device that reports more than 255 ports will be a representor device and it exposes itself as a RAW Ethernet only device CM/MAD/IPoIB and other ULPs aren't effected by this change and their sysfs/interfaces that are exposes to userspace can remain unchanged. While here cleanup some alignment issues and remove unneeded sanity checks (mainly in rdmavt), Link: https://lore.kernel.org/r/20210301070420.439400-1-leon@kernel.org Signed-off-by: Mark Bloch <mbloch@nvidia.com> Signed-off-by: Leon Romanovsky <leonro@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
2021-03-01 07:04:20 +00:00
void (*cap_mask_chg)(struct rvt_dev_info *rdi, u32 port_num);
/*
* The following functions can be safely ignored completely. Any use of
* these is checked for NULL before blindly calling. Rdmavt should also
* be functional if drivers omit these.
*/
/* Called to inform the driver that all qps should now be freed. */
unsigned (*free_all_qps)(struct rvt_dev_info *rdi);
/* Driver specific AH validation */
int (*check_ah)(struct ib_device *, struct rdma_ah_attr *);
/* Inform the driver a new AH has been created */
void (*notify_new_ah)(struct ib_device *, struct rdma_ah_attr *,
struct rvt_ah *);
/* Let the driver pick the next queue pair number*/
int (*alloc_qpn)(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
RDMA: Support more than 255 rdma ports Current code uses many different types when dealing with a port of a RDMA device: u8, unsigned int and u32. Switch to u32 to clean up the logic. This allows us to make (at least) the core view consistent and use the same type. Unfortunately not all places can be converted. Many uverbs functions expect port to be u8 so keep those places in order not to break UAPIs. HW/Spec defined values must also not be changed. With the switch to u32 we now can support devices with more than 255 ports. U32_MAX is reserved to make control logic a bit easier to deal with. As a device with U32_MAX ports probably isn't going to happen any time soon this seems like a non issue. When a device with more than 255 ports is created uverbs will report the RDMA device as having 255 ports as this is the max currently supported. The verbs interface is not changed yet because the IBTA spec limits the port size in too many places to be u8 and all applications that relies in verbs won't be able to cope with this change. At this stage, we are extending the interfaces that are using vendor channel solely Once the limitation is lifted mlx5 in switchdev mode will be able to have thousands of SFs created by the device. As the only instance of an RDMA device that reports more than 255 ports will be a representor device and it exposes itself as a RAW Ethernet only device CM/MAD/IPoIB and other ULPs aren't effected by this change and their sysfs/interfaces that are exposes to userspace can remain unchanged. While here cleanup some alignment issues and remove unneeded sanity checks (mainly in rdmavt), Link: https://lore.kernel.org/r/20210301070420.439400-1-leon@kernel.org Signed-off-by: Mark Bloch <mbloch@nvidia.com> Signed-off-by: Leon Romanovsky <leonro@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
2021-03-01 07:04:20 +00:00
enum ib_qp_type type, u32 port_num);
/* Determine if its safe or allowed to modify the qp */
int (*check_modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr,
int attr_mask, struct ib_udata *udata);
/* Driver specific QP modification/notification-of */
void (*modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr,
int attr_mask, struct ib_udata *udata);
/* Notify driver a mad agent has been created */
void (*notify_create_mad_agent)(struct rvt_dev_info *rdi, int port_idx);
/* Notify driver a mad agent has been removed */
void (*notify_free_mad_agent)(struct rvt_dev_info *rdi, int port_idx);
/* Notify driver to restart rc */
void (*notify_restart_rc)(struct rvt_qp *qp, u32 psn, int wait);
IB/{hfi1, rdmavt, qib}: Implement CQ completion vector support Currently the driver doesn't support completion vectors. These are used to indicate which sets of CQs should be grouped together into the same vector. A vector is a CQ processing thread that runs on a specific CPU. If an application has several CQs bound to different completion vectors, and each completion vector runs on different CPUs, then the completion queue workload is balanced. This helps scale as more nodes are used. Implement CQ completion vector support using a global workqueue where a CQ entry is queued to the CPU corresponding to the CQ's completion vector. Since the workqueue is global, it's guaranteed to always be there when queueing CQ entries; Therefore, the RCU locking for cq->rdi->worker in the hot path is superfluous. Each completion vector is assigned to a different CPU. The number of completion vectors available is computed by taking the number of online, physical CPUs from the local NUMA node and subtracting the CPUs used for kernel receive queues and the general interrupt. Special use cases: * If there are no CPUs left for completion vectors, the same CPU for the general interrupt is used; Therefore, there would only be one completion vector available. * For multi-HFI systems, the number of completion vectors available for each device is the total number of completion vectors in the local NUMA node divided by the number of devices in the same NUMA node. If there's a division remainder, the first device to get initialized gets an extra completion vector. Upon a CQ creation, an invalid completion vector could be specified. Handle it as follows: * If the completion vector is less than 0, set it to 0. * Set the completion vector to the result of the passed completion vector moded with the number of device completion vectors available. Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com> Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by: Doug Ledford <dledford@redhat.com>
2018-05-02 13:43:55 +00:00
/* Get and return CPU to pin CQ processing thread */
int (*comp_vect_cpu_lookup)(struct rvt_dev_info *rdi, int comp_vect);
};
struct rvt_dev_info {
struct ib_device ibdev; /* Keep this first. Nothing above here */
/*
* Prior to calling for registration the driver will be responsible for
* allocating space for this structure.
*
* The driver will also be responsible for filling in certain members of
* dparms.props. The driver needs to fill in dparms exactly as it would
* want values reported to a ULP. This will be returned to the caller
* in rdmavt's device. The driver should also therefore refrain from
* modifying this directly after registration with rdmavt.
*/
/* Driver specific properties */
struct rvt_driver_params dparms;
/* post send table */
const struct rvt_operation_params *post_parms;
/* opcode translation table */
const enum ib_wc_opcode *wc_opcode;
/* Driver specific helper functions */
struct rvt_driver_provided driver_f;
struct rvt_mregion __rcu *dma_mr;
struct rvt_lkey_table lkey_table;
/* Internal use */
int n_pds_allocated;
spinlock_t n_pds_lock; /* Protect pd allocated count */
int n_ahs_allocated;
spinlock_t n_ahs_lock; /* Protect ah allocated count */
u32 n_srqs_allocated;
spinlock_t n_srqs_lock; /* Protect srqs allocated count */
int flags;
struct rvt_ibport **ports;
/* QP */
struct rvt_qp_ibdev *qp_dev;
u32 n_qps_allocated; /* number of QPs allocated for device */
u32 n_rc_qps; /* number of RC QPs allocated for device */
u32 busy_jiffies; /* timeout scaling based on RC QP count */
spinlock_t n_qps_lock; /* protect qps, rc qps and busy jiffy counts */
/* memory maps */
struct list_head pending_mmaps;
spinlock_t mmap_offset_lock; /* protect mmap_offset */
u32 mmap_offset;
spinlock_t pending_lock; /* protect pending mmap list */
/* CQ */
u32 n_cqs_allocated; /* number of CQs allocated for device */
spinlock_t n_cqs_lock; /* protect count of in use cqs */
/* Multicast */
u32 n_mcast_grps_allocated; /* number of mcast groups allocated */
spinlock_t n_mcast_grps_lock;
/* Memory Working Set Size */
struct rvt_wss *wss;
};
/**
* rvt_set_ibdev_name - Craft an IB device name from client info
* @rdi: pointer to the client rvt_dev_info structure
* @name: client specific name
* @unit: client specific unit number.
*/
static inline void rvt_set_ibdev_name(struct rvt_dev_info *rdi,
const char *fmt, const char *name,
const int unit)
{
/*
* FIXME: rvt and its users want to touch the ibdev before
* registration and have things like the name work. We don't have the
* infrastructure in the core to support this directly today, hack it
* to work by setting the name manually here.
*/
dev_set_name(&rdi->ibdev.dev, fmt, name, unit);
strlcpy(rdi->ibdev.name, dev_name(&rdi->ibdev.dev), IB_DEVICE_NAME_MAX);
}
/**
* rvt_get_ibdev_name - return the IB name
* @rdi: rdmavt device
*
* Return the registered name of the device.
*/
static inline const char *rvt_get_ibdev_name(const struct rvt_dev_info *rdi)
{
return dev_name(&rdi->ibdev.dev);
}
static inline struct rvt_pd *ibpd_to_rvtpd(struct ib_pd *ibpd)
{
return container_of(ibpd, struct rvt_pd, ibpd);
}
static inline struct rvt_ah *ibah_to_rvtah(struct ib_ah *ibah)
{
return container_of(ibah, struct rvt_ah, ibah);
}
static inline struct rvt_dev_info *ib_to_rvt(struct ib_device *ibdev)
{
return container_of(ibdev, struct rvt_dev_info, ibdev);
}
static inline unsigned rvt_get_npkeys(struct rvt_dev_info *rdi)
{
/*
* All ports have same number of pkeys.
*/
return rdi->dparms.npkeys;
}
/*
* Return the max atomic suitable for determining
* the size of the ack ring buffer in a QP.
*/
static inline unsigned int rvt_max_atomic(struct rvt_dev_info *rdi)
{
return rdi->dparms.max_rdma_atomic +
rdi->dparms.extra_rdma_atomic + 1;
}
static inline unsigned int rvt_size_atomic(struct rvt_dev_info *rdi)
{
return rdi->dparms.max_rdma_atomic +
rdi->dparms.extra_rdma_atomic;
}
/*
* Return the indexed PKEY from the port PKEY table.
*/
static inline u16 rvt_get_pkey(struct rvt_dev_info *rdi,
int port_index,
unsigned index)
{
if (index >= rvt_get_npkeys(rdi))
return 0;
else
return rdi->ports[port_index]->pkey_table[index];
}
struct rvt_dev_info *rvt_alloc_device(size_t size, int nports);
void rvt_dealloc_device(struct rvt_dev_info *rdi);
int rvt_register_device(struct rvt_dev_info *rvd);
void rvt_unregister_device(struct rvt_dev_info *rvd);
int rvt_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr);
int rvt_init_port(struct rvt_dev_info *rdi, struct rvt_ibport *port,
int port_index, u16 *pkey_table);
int rvt_fast_reg_mr(struct rvt_qp *qp, struct ib_mr *ibmr, u32 key,
int access);
int rvt_invalidate_rkey(struct rvt_qp *qp, u32 rkey);
int rvt_rkey_ok(struct rvt_qp *qp, struct rvt_sge *sge,
u32 len, u64 vaddr, u32 rkey, int acc);
int rvt_lkey_ok(struct rvt_lkey_table *rkt, struct rvt_pd *pd,
struct rvt_sge *isge, struct rvt_sge *last_sge,
struct ib_sge *sge, int acc);
struct rvt_mcast *rvt_mcast_find(struct rvt_ibport *ibp, union ib_gid *mgid,
u16 lid);
#endif /* DEF_RDMA_VT_H */