[PATCH v16 2/2] nvmet: support reservation feature

Sagi Grimberg sagi at grimberg.me
Mon Oct 21 00:01:23 PDT 2024




On 21/10/2024 8:58, Guixin Liu wrote:
>
> 在 2024/10/21 07:03, Sagi Grimberg 写道:
>>
>>
>>
>> On 17/10/2024 7:48, Guixin Liu wrote:
>>> This patch implements the reservation feature, including:
>>>    1. reservation register(register, unregister and replace).
>>>    2. reservation acquire(acquire, preempt, preempt and abort).
>>>    3. reservation release(release and clear).
>>>    4. reservation report.
>>>    5. set feature and get feature of reservation notify mask.
>>>    6. get log page of reservation event.
>>>
>>> Not supported:
>>>    1. persistent reservation through power loss.
>>>
>>> Test cases:
>>>    Use nvme-cli and fio to test all implemented sub features:
>>>    1. use nvme resv-register to register host a registrant or
>>>       unregister or replace a new key.
>>>    2. use nvme resv-acquire to set host to the holder, and use fio
>>>       to send read and write io in all reservation type. And also
>>>       test preempt and "preempt and abort".
>>>    3. use nvme resv-report to show all registrants and reservation
>>>       status.
>>>    4. use nvme resv-release to release all registrants.
>>>    5. use nvme get-log to get events generated by the preceding
>>>       operations.
>>>
>>> In addition, make reservation configurable, one can set ns to
>>> support reservation before enable ns. The default of resv_enable
>>> is false.
>>>
>>> Signed-off-by: Guixin Liu <kanie at linux.alibaba.com>
>>> Reviewed-by: Dmitry Bogdanov <d.bogdanov at yadro.com>
>>> Reviewed-by: Christoph Hellwig <hch at lst.de>
>>> Tested-by: Chaitanya Kulkarni <kch at nvidia.com>
>>> Reviewed-by: Chaitanya Kulkarni <kch at nvidia.com>
>>> ---
>>>   drivers/nvme/target/Makefile      |    2 +-
>>>   drivers/nvme/target/admin-cmd.c   |   24 +-
>>>   drivers/nvme/target/configfs.c    |   27 +
>>>   drivers/nvme/target/core.c        |   58 +-
>>>   drivers/nvme/target/fabrics-cmd.c |    4 +-
>>>   drivers/nvme/target/nvmet.h       |   55 +-
>>>   drivers/nvme/target/pr.c          | 1172 
>>> +++++++++++++++++++++++++++++
>>>   7 files changed, 1330 insertions(+), 12 deletions(-)
>>>   create mode 100644 drivers/nvme/target/pr.c
>>>
>>> diff --git a/drivers/nvme/target/Makefile 
>>> b/drivers/nvme/target/Makefile
>>> index c402c44350b2..f2b025bbe10c 100644
>>> --- a/drivers/nvme/target/Makefile
>>> +++ b/drivers/nvme/target/Makefile
>>> @@ -10,7 +10,7 @@ obj-$(CONFIG_NVME_TARGET_FCLOOP)    += nvme-fcloop.o
>>>   obj-$(CONFIG_NVME_TARGET_TCP)        += nvmet-tcp.o
>>>     nvmet-y        += core.o configfs.o admin-cmd.o fabrics-cmd.o \
>>> -            discovery.o io-cmd-file.o io-cmd-bdev.o
>>> +            discovery.o io-cmd-file.o io-cmd-bdev.o pr.o
>>>   nvmet-$(CONFIG_NVME_TARGET_DEBUGFS)    += debugfs.o
>>>   nvmet-$(CONFIG_NVME_TARGET_PASSTHRU)    += passthru.o
>>>   nvmet-$(CONFIG_BLK_DEV_ZONED)        += zns.o
>>> diff --git a/drivers/nvme/target/admin-cmd.c 
>>> b/drivers/nvme/target/admin-cmd.c
>>> index 081f0473cd9e..19428745c795 100644
>>> --- a/drivers/nvme/target/admin-cmd.c
>>> +++ b/drivers/nvme/target/admin-cmd.c
>>> @@ -176,6 +176,10 @@ static void nvmet_get_cmd_effects_nvm(struct 
>>> nvme_effects_log *log)
>>>       log->iocs[nvme_cmd_read] =
>>>       log->iocs[nvme_cmd_flush] =
>>>       log->iocs[nvme_cmd_dsm]    =
>>> +    log->iocs[nvme_cmd_resv_acquire] =
>>> +    log->iocs[nvme_cmd_resv_register] =
>>> +    log->iocs[nvme_cmd_resv_release] =
>>> +    log->iocs[nvme_cmd_resv_report] =
>>>           cpu_to_le32(NVME_CMD_EFFECTS_CSUPP);
>>>       log->iocs[nvme_cmd_write] =
>>>       log->iocs[nvme_cmd_write_zeroes] =
>>> @@ -340,6 +344,8 @@ static void nvmet_execute_get_log_page(struct 
>>> nvmet_req *req)
>>>           return nvmet_execute_get_log_cmd_effects_ns(req);
>>>       case NVME_LOG_ANA:
>>>           return nvmet_execute_get_log_page_ana(req);
>>> +    case NVME_LOG_RESERVATION:
>>> +        return nvmet_execute_get_log_page_resv(req);
>>>       }
>>>       pr_debug("unhandled lid %d on qid %d\n",
>>>              req->cmd->get_log_page.lid, req->sq->qid);
>>> @@ -433,7 +439,8 @@ static void nvmet_execute_identify_ctrl(struct 
>>> nvmet_req *req)
>>>       id->nn = cpu_to_le32(NVMET_MAX_NAMESPACES);
>>>       id->mnan = cpu_to_le32(NVMET_MAX_NAMESPACES);
>>>       id->oncs = cpu_to_le16(NVME_CTRL_ONCS_DSM |
>>> -            NVME_CTRL_ONCS_WRITE_ZEROES);
>>> +            NVME_CTRL_ONCS_WRITE_ZEROES |
>>> +            NVME_CTRL_ONCS_RESERVATIONS);
>>>         /* XXX: don't report vwc if the underlying device is write 
>>> through */
>>>       id->vwc = NVME_CTRL_VWC_PRESENT;
>>> @@ -551,6 +558,15 @@ static void nvmet_execute_identify_ns(struct 
>>> nvmet_req *req)
>>>       id->nmic = NVME_NS_NMIC_SHARED;
>>>       id->anagrpid = cpu_to_le32(req->ns->anagrpid);
>>>   +    if (req->ns->pr.enable)
>>> +        id->rescap = NVME_PR_SUPPORT_WRITE_EXCLUSIVE |
>>> +            NVME_PR_SUPPORT_EXCLUSIVE_ACCESS |
>>> +            NVME_PR_SUPPORT_WRITE_EXCLUSIVE_REG_ONLY |
>>> +            NVME_PR_SUPPORT_EXCLUSIVE_ACCESS_REG_ONLY |
>>> +            NVME_PR_SUPPORT_WRITE_EXCLUSIVE_ALL_REGS |
>>> +            NVME_PR_SUPPORT_EXCLUSIVE_ACCESS_ALL_REGS |
>>> +            NVME_PR_SUPPORT_IEKEY_VER_1_3_DEF;
>>> +
>>>       memcpy(&id->nguid, &req->ns->nguid, sizeof(id->nguid));
>>>         id->lbaf[0].ds = req->ns->blksize_shift;
>>> @@ -861,6 +877,9 @@ void nvmet_execute_set_features(struct nvmet_req 
>>> *req)
>>>       case NVME_FEAT_WRITE_PROTECT:
>>>           status = nvmet_set_feat_write_protect(req);
>>>           break;
>>> +    case NVME_FEAT_RESV_MASK:
>>> +        status = nvmet_set_feat_resv_notif_mask(req, cdw11);
>>> +        break;
>>>       default:
>>>           req->error_loc = offsetof(struct nvme_common_command, cdw10);
>>>           status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
>>> @@ -959,6 +978,9 @@ void nvmet_execute_get_features(struct nvmet_req 
>>> *req)
>>>       case NVME_FEAT_WRITE_PROTECT:
>>>           status = nvmet_get_feat_write_protect(req);
>>>           break;
>>> +    case NVME_FEAT_RESV_MASK:
>>> +        status = nvmet_get_feat_resv_notif_mask(req);
>>> +        break;
>>>       default:
>>>           req->error_loc =
>>>               offsetof(struct nvme_common_command, cdw10);
>>> diff --git a/drivers/nvme/target/configfs.c 
>>> b/drivers/nvme/target/configfs.c
>>> index 685e89b35d33..eeee9e9b854c 100644
>>> --- a/drivers/nvme/target/configfs.c
>>> +++ b/drivers/nvme/target/configfs.c
>>> @@ -769,6 +769,32 @@ static ssize_t 
>>> nvmet_ns_revalidate_size_store(struct config_item *item,
>>>     CONFIGFS_ATTR_WO(nvmet_ns_, revalidate_size);
>>>   +static ssize_t nvmet_ns_resv_enable_show(struct config_item 
>>> *item, char *page)
>>> +{
>>> +    return sysfs_emit(page, "%d\n", to_nvmet_ns(item)->pr.enable);
>>> +}
>>> +
>>> +static ssize_t nvmet_ns_resv_enable_store(struct config_item *item,
>>> +        const char *page, size_t count)
>>> +{
>>> +    struct nvmet_ns *ns = to_nvmet_ns(item);
>>> +    bool val;
>>> +
>>> +    if (kstrtobool(page, &val))
>>> +        return -EINVAL;
>>> +
>>> +    mutex_lock(&ns->subsys->lock);
>>> +    if (ns->enabled) {
>>> +        pr_err("the ns:%d is already enabled.\n", ns->nsid);
>>> +        mutex_unlock(&ns->subsys->lock);
>>> +        return -EINVAL;
>>> +    }
>>> +    ns->pr.enable = val;
>>> +    mutex_unlock(&ns->subsys->lock);
>>> +    return count;
>>> +}
>>> +CONFIGFS_ATTR(nvmet_ns_, resv_enable);
>>> +
>>>   static struct configfs_attribute *nvmet_ns_attrs[] = {
>>>       &nvmet_ns_attr_device_path,
>>>       &nvmet_ns_attr_device_nguid,
>>> @@ -777,6 +803,7 @@ static struct configfs_attribute 
>>> *nvmet_ns_attrs[] = {
>>>       &nvmet_ns_attr_enable,
>>>       &nvmet_ns_attr_buffered_io,
>>>       &nvmet_ns_attr_revalidate_size,
>>> +    &nvmet_ns_attr_resv_enable,
>>>   #ifdef CONFIG_PCI_P2PDMA
>>>       &nvmet_ns_attr_p2pmem,
>>>   #endif
>>> diff --git a/drivers/nvme/target/core.c b/drivers/nvme/target/core.c
>>> index ed2424f8a396..a2af7a9bc3b9 100644
>>> --- a/drivers/nvme/target/core.c
>>> +++ b/drivers/nvme/target/core.c
>>> @@ -611,6 +611,12 @@ int nvmet_ns_enable(struct nvmet_ns *ns)
>>>       if (ret)
>>>           goto out_restore_subsys_maxnsid;
>>>   +    if (ns->pr.enable) {
>>> +        ret = nvmet_pr_init_ns(ns);
>>> +        if (ret)
>>> +            goto out_remove_from_subsys;
>>> +    }
>>> +
>>>       subsys->nr_namespaces++;
>>>         nvmet_ns_changed(subsys, ns->nsid);
>>> @@ -620,6 +626,8 @@ int nvmet_ns_enable(struct nvmet_ns *ns)
>>>       mutex_unlock(&subsys->lock);
>>>       return ret;
>>>   +out_remove_from_subsys:
>>> +    xa_erase(&subsys->namespaces, ns->nsid);
>>>   out_restore_subsys_maxnsid:
>>>       subsys->max_nsid = nvmet_max_nsid(subsys);
>>>       percpu_ref_exit(&ns->ref);
>>> @@ -663,6 +671,9 @@ void nvmet_ns_disable(struct nvmet_ns *ns)
>>>       wait_for_completion(&ns->disable_done);
>>>       percpu_ref_exit(&ns->ref);
>>>   +    if (ns->pr.enable)
>>> +        nvmet_pr_exit_ns(ns);
>>> +
>>>       mutex_lock(&subsys->lock);
>>>         subsys->nr_namespaces--;
>>> @@ -766,6 +777,7 @@ static void __nvmet_req_complete(struct 
>>> nvmet_req *req, u16 status)
>>>       trace_nvmet_req_complete(req);
>>>         req->ops->queue_response(req);
>>> +    nvmet_pr_put_ns_pc_ref(req);
>>>       if (ns)
>>>           nvmet_put_namespace(ns);
>>>   }
>>> @@ -929,18 +941,39 @@ static u16 nvmet_parse_io_cmd(struct nvmet_req 
>>> *req)
>>>           return ret;
>>>       }
>>>   +    if (req->ns->pr.enable) {
>>> +        ret = nvmet_parse_pr_cmd(req);
>>> +        if (!ret)
>>> +            return ret;
>>> +    }
>>> +
>>>       switch (req->ns->csi) {
>>>       case NVME_CSI_NVM:
>>>           if (req->ns->file)
>>> -            return nvmet_file_parse_io_cmd(req);
>>> -        return nvmet_bdev_parse_io_cmd(req);
>>> +            ret = nvmet_file_parse_io_cmd(req);
>>> +        else
>>> +            ret = nvmet_bdev_parse_io_cmd(req);
>>> +        break;
>>>       case NVME_CSI_ZNS:
>>>           if (IS_ENABLED(CONFIG_BLK_DEV_ZONED))
>>> -            return nvmet_bdev_zns_parse_io_cmd(req);
>>> -        return NVME_SC_INVALID_IO_CMD_SET;
>>> +            ret = nvmet_bdev_zns_parse_io_cmd(req);
>>> +        else
>>> +            ret = NVME_SC_INVALID_IO_CMD_SET;
>>> +        break;
>>>       default:
>>> -        return NVME_SC_INVALID_IO_CMD_SET;
>>> +        ret = NVME_SC_INVALID_IO_CMD_SET;
>>> +    }
>>> +    if (ret)
>>> +        return ret;
>>> +
>>> +    if (req->ns->pr.enable) {
>>> +        ret = nvmet_pr_check_cmd_access(req);
>>> +        if (ret)
>>> +            return ret;
>>> +
>>> +        ret = nvmet_pr_get_ns_pc_ref(req);
>>>       }
>>> +    return ret;
>>>   }
>>>     bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
>>> @@ -964,6 +997,7 @@ bool nvmet_req_init(struct nvmet_req *req, 
>>> struct nvmet_cq *cq,
>>>       req->ns = NULL;
>>>       req->error_loc = NVMET_NO_ERROR_LOC;
>>>       req->error_slba = 0;
>>> +    req->pc_ref = NULL;
>>>         /* no support for fused commands yet */
>>>       if (unlikely(flags & (NVME_CMD_FUSE_FIRST | 
>>> NVME_CMD_FUSE_SECOND))) {
>>> @@ -1015,6 +1049,7 @@ EXPORT_SYMBOL_GPL(nvmet_req_init);
>>>   void nvmet_req_uninit(struct nvmet_req *req)
>>>   {
>>>       percpu_ref_put(&req->sq->ref);
>>> +    nvmet_pr_put_ns_pc_ref(req);
>>>       if (req->ns)
>>>           nvmet_put_namespace(req->ns);
>>>   }
>>> @@ -1383,7 +1418,8 @@ static void nvmet_fatal_error_handler(struct 
>>> work_struct *work)
>>>   }
>>>     u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
>>> -        struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
>>> +        struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp,
>>> +        uuid_t *hostid)
>>>   {
>>>       struct nvmet_subsys *subsys;
>>>       struct nvmet_ctrl *ctrl;
>>> @@ -1462,6 +1498,8 @@ u16 nvmet_alloc_ctrl(const char *subsysnqn, 
>>> const char *hostnqn,
>>>       }
>>>       ctrl->cntlid = ret;
>>>   +    uuid_copy(&ctrl->hostid, hostid);
>>> +
>>>       /*
>>>        * Discovery controllers may use some arbitrary high value
>>>        * in order to cleanup stale discovery sessions
>>> @@ -1478,6 +1516,9 @@ u16 nvmet_alloc_ctrl(const char *subsysnqn, 
>>> const char *hostnqn,
>>>       nvmet_start_keep_alive_timer(ctrl);
>>>         mutex_lock(&subsys->lock);
>>> +    ret = nvmet_ctrl_init_pr(ctrl);
>>> +    if (ret)
>>> +        goto init_pr_fail;
>>>       list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
>>>       nvmet_setup_p2p_ns_map(ctrl, req);
>>>       nvmet_debugfs_ctrl_setup(ctrl);
>>> @@ -1486,6 +1527,10 @@ u16 nvmet_alloc_ctrl(const char *subsysnqn, 
>>> const char *hostnqn,
>>>       *ctrlp = ctrl;
>>>       return 0;
>>>   +init_pr_fail:
>>> +    mutex_unlock(&subsys->lock);
>>> +    nvmet_stop_keep_alive_timer(ctrl);
>>> +    ida_free(&cntlid_ida, ctrl->cntlid);
>>>   out_free_sqs:
>>>       kfree(ctrl->sqs);
>>>   out_free_changed_ns_list:
>>> @@ -1504,6 +1549,7 @@ static void nvmet_ctrl_free(struct kref *ref)
>>>       struct nvmet_subsys *subsys = ctrl->subsys;
>>>         mutex_lock(&subsys->lock);
>>> +    nvmet_ctrl_destroy_pr(ctrl);
>>>       nvmet_release_p2p_ns_map(ctrl);
>>>       list_del(&ctrl->subsys_entry);
>>>       mutex_unlock(&subsys->lock);
>>> diff --git a/drivers/nvme/target/fabrics-cmd.c 
>>> b/drivers/nvme/target/fabrics-cmd.c
>>> index c4b2eddd5666..28a84af1b4c0 100644
>>> --- a/drivers/nvme/target/fabrics-cmd.c
>>> +++ b/drivers/nvme/target/fabrics-cmd.c
>>> @@ -245,12 +245,10 @@ static void nvmet_execute_admin_connect(struct 
>>> nvmet_req *req)
>>>       d->subsysnqn[NVMF_NQN_FIELD_LEN - 1] = '\0';
>>>       d->hostnqn[NVMF_NQN_FIELD_LEN - 1] = '\0';
>>>       status = nvmet_alloc_ctrl(d->subsysnqn, d->hostnqn, req,
>>> -                  le32_to_cpu(c->kato), &ctrl);
>>> +                  le32_to_cpu(c->kato), &ctrl, &d->hostid);
>>>       if (status)
>>>           goto out;
>>>   -    uuid_copy(&ctrl->hostid, &d->hostid);
>>> -
>>>       dhchap_status = nvmet_setup_auth(ctrl);
>>>       if (dhchap_status) {
>>>           pr_err("Failed to setup authentication, dhchap status %u\n",
>>> diff --git a/drivers/nvme/target/nvmet.h b/drivers/nvme/target/nvmet.h
>>> index 190f55e6d753..51552f0430b4 100644
>>> --- a/drivers/nvme/target/nvmet.h
>>> +++ b/drivers/nvme/target/nvmet.h
>>> @@ -20,6 +20,7 @@
>>>   #include <linux/blkdev.h>
>>>   #include <linux/radix-tree.h>
>>>   #include <linux/t10-pi.h>
>>> +#include <linux/kfifo.h>
>>>     #define NVMET_DEFAULT_VS        NVME_VS(1, 3, 0)
>>>   @@ -30,6 +31,7 @@
>>>   #define NVMET_MN_MAX_SIZE        40
>>>   #define NVMET_SN_MAX_SIZE        20
>>>   #define NVMET_FR_MAX_SIZE        8
>>> +#define NVMET_PR_LOG_QUEUE_SIZE        64
>>>     /*
>>>    * Supported optional AENs:
>>> @@ -56,6 +58,30 @@
>>>   #define IPO_IATTR_CONNECT_SQE(x)    \
>>>       (cpu_to_le32(offsetof(struct nvmf_connect_command, x)))
>>>   +struct nvmet_pr_registrant {
>>> +    u64            rkey;
>>> +    uuid_t            hostid;
>>> +    enum nvme_pr_type    rtype;
>>> +    struct list_head    entry;
>>> +    struct rcu_head        rcu;
>>> +};
>>> +
>>> +struct nvmet_pr {
>>> +    bool            enable;
>>> +    unsigned long        notify_mask;
>>> +    atomic_t        generation;
>>> +    struct nvmet_pr_registrant __rcu *holder;
>>> +    struct mutex        pr_lock;
>>> +    struct list_head    registrant_list;
>>> +};
>>> +
>>> +struct nvmet_pr_per_ctrl_ref {
>>> +    struct percpu_ref    ref;
>>> +    struct completion    free_done;
>>> +    struct completion    confirm_done;
>>> +    uuid_t            hostid;
>>> +};
>>> +
>>>   struct nvmet_ns {
>>>       struct percpu_ref    ref;
>>>       struct file        *bdev_file;
>>> @@ -85,6 +111,8 @@ struct nvmet_ns {
>>>       int            pi_type;
>>>       int            metadata_size;
>>>       u8            csi;
>>> +    struct nvmet_pr        pr;
>>> +    struct xarray        pr_per_ctrl_refs;
>>>   };
>>>     static inline struct nvmet_ns *to_nvmet_ns(struct config_item 
>>> *item)
>>> @@ -191,6 +219,13 @@ static inline bool 
>>> nvmet_port_secure_channel_required(struct nvmet_port *port)
>>>       return nvmet_port_disc_addr_treq_secure_channel(port) == 
>>> NVMF_TREQ_REQUIRED;
>>>   }
>>>   +struct nvmet_pr_log_mgr {
>>> +    struct mutex        lock;
>>> +    u64            lost_count;
>>> +    u64            counter;
>>> +    DECLARE_KFIFO(log_queue, struct nvme_pr_log, 
>>> NVMET_PR_LOG_QUEUE_SIZE);
>>> +};
>>> +
>>>   struct nvmet_ctrl {
>>>       struct nvmet_subsys    *subsys;
>>>       struct nvmet_sq        **sqs;
>>> @@ -246,6 +281,7 @@ struct nvmet_ctrl {
>>>       u8            *dh_key;
>>>       size_t            dh_keysize;
>>>   #endif
>>> +    struct nvmet_pr_log_mgr pr_log_mgr;
>>>   };
>>>     struct nvmet_subsys {
>>> @@ -412,6 +448,7 @@ struct nvmet_req {
>>>       struct device        *p2p_client;
>>>       u16            error_loc;
>>>       u64            error_slba;
>>> +    struct nvmet_pr_per_ctrl_ref *pc_ref;
>>>   };
>>>     #define NVMET_MAX_MPOOL_BVEC        16
>>> @@ -498,7 +535,8 @@ void nvmet_ctrl_fatal_error(struct nvmet_ctrl 
>>> *ctrl);
>>>     void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new);
>>>   u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
>>> -        struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp);
>>> +        struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp,
>>> +        uuid_t *hostid);
>>>   struct nvmet_ctrl *nvmet_ctrl_find_get(const char *subsysnqn,
>>>                          const char *hostnqn, u16 cntlid,
>>>                          struct nvmet_req *req);
>>> @@ -761,4 +799,19 @@ static inline bool nvmet_has_auth(struct 
>>> nvmet_ctrl *ctrl)
>>>   static inline const char *nvmet_dhchap_dhgroup_name(u8 dhgid) { 
>>> return NULL; }
>>>   #endif
>>>   +int nvmet_pr_init_ns(struct nvmet_ns *ns);
>>> +u16 nvmet_parse_pr_cmd(struct nvmet_req *req);
>>> +u16 nvmet_pr_check_cmd_access(struct nvmet_req *req);
>>> +int nvmet_ctrl_init_pr(struct nvmet_ctrl *ctrl);
>>> +void nvmet_ctrl_destroy_pr(struct nvmet_ctrl *ctrl);
>>> +void nvmet_pr_exit_ns(struct nvmet_ns *ns);
>>> +void nvmet_execute_get_log_page_resv(struct nvmet_req *req);
>>> +u16 nvmet_set_feat_resv_notif_mask(struct nvmet_req *req, u32 mask);
>>> +u16 nvmet_get_feat_resv_notif_mask(struct nvmet_req *req);
>>> +u16 nvmet_pr_get_ns_pc_ref(struct nvmet_req *req);
>>> +static inline void nvmet_pr_put_ns_pc_ref(struct nvmet_req *req)
>>> +{
>>> +    if (req->pc_ref)
>>> +        percpu_ref_put(&req->pc_ref->ref);
>>> +}
>>>   #endif /* _NVMET_H */
>>> diff --git a/drivers/nvme/target/pr.c b/drivers/nvme/target/pr.c
>>> new file mode 100644
>>> index 000000000000..7795a103dd3b
>>> --- /dev/null
>>> +++ b/drivers/nvme/target/pr.c
>>> @@ -0,0 +1,1172 @@
>>> +// SPDX-License-Identifier: GPL-2.0
>>> +/*
>>> + * NVMe over Fabrics Persist Reservation.
>>> + * Copyright (c) 2024 Guixin Liu, Alibaba Group.
>>> + * All rights reserved.
>>> + */
>>> +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
>>> +#include <linux/unaligned.h>
>>> +#include <linux/lockdep.h>
>>> +#include "nvmet.h"
>>> +
>>> +#define NVMET_PR_NOTIFI_MASK_ALL \
>>> +    (1 << NVME_PR_NOTIFY_BIT_REG_PREEMPTED | \
>>> +     1 << NVME_PR_NOTIFY_BIT_RESV_RELEASED | \
>>> +     1 << NVME_PR_NOTIFY_BIT_RESV_PREEMPTED)
>>> +
>>> +static inline bool nvmet_pr_parse_ignore_key(u32 cdw10)
>>> +{
>>> +    /* Ignore existing key, bit 03. */
>>> +    return (cdw10 >> 3) & 1;
>>> +}
>>> +
>>> +static inline struct nvmet_ns *nvmet_pr_to_ns(struct nvmet_pr *pr)
>>> +{
>>> +    return container_of(pr, struct nvmet_ns, pr);
>>> +}
>>> +
>>> +static struct nvmet_pr_registrant *
>>> +nvmet_pr_find_registrant(struct nvmet_pr *pr, uuid_t *hostid)
>>> +{
>>> +    struct nvmet_pr_registrant *reg;
>>> +
>>> +    list_for_each_entry_rcu(reg, &pr->registrant_list, entry,
>>> +                rcu_read_lock_held() ||
>>> +                lockdep_is_held(&pr->pr_lock)) {
>>> +        if (uuid_equal(&reg->hostid, hostid))
>>> +            return reg;
>>> +    }
>>> +    return NULL;
>>> +}
>>> +
>>> +u16 nvmet_set_feat_resv_notif_mask(struct nvmet_req *req, u32 mask)
>>> +{
>>> +    u32 nsid = le32_to_cpu(req->cmd->common.nsid);
>>> +    struct nvmet_ctrl *ctrl = req->sq->ctrl;
>>> +    struct nvmet_ns *ns;
>>> +    unsigned long idx;
>>> +    u16 status;
>>> +
>>> +    if (mask & ~(NVMET_PR_NOTIFI_MASK_ALL)) {
>>> +        req->error_loc = offsetof(struct nvme_common_command, cdw11);
>>> +        return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
>>> +    }
>>> +
>>> +    if (nsid != U32_MAX) {
>>> +        status = nvmet_req_find_ns(req);
>>> +        if (status)
>>> +            return status;
>>> +        if (!req->ns->pr.enable)
>>> +            return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
>>> +
>>> +        WRITE_ONCE(req->ns->pr.notify_mask, mask);
>>> +        goto success;
>>> +    }
>>> +
>>> +    xa_for_each(&ctrl->subsys->namespaces, idx, ns) {
>>> +        if (ns->pr.enable)
>>> +            WRITE_ONCE(ns->pr.notify_mask, mask);
>>> +    }
>>> +
>>> +success:
>>> +    nvmet_set_result(req, mask);
>>> +    return NVME_SC_SUCCESS;
>>> +}
>>> +
>>> +u16 nvmet_get_feat_resv_notif_mask(struct nvmet_req *req)
>>> +{
>>> +    u16 status;
>>> +
>>> +    status = nvmet_req_find_ns(req);
>>> +    if (status)
>>> +        return status;
>>> +
>>> +    if (!req->ns->pr.enable)
>>> +        return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
>>> +
>>> +    nvmet_set_result(req, READ_ONCE(req->ns->pr.notify_mask));
>>> +    return status;
>>> +}
>>> +
>>> +void nvmet_execute_get_log_page_resv(struct nvmet_req *req)
>>> +{
>>> +    struct nvmet_pr_log_mgr *log_mgr = &req->sq->ctrl->pr_log_mgr;
>>> +    struct nvme_pr_log next_log = {0};
>>> +    struct nvme_pr_log log = {0};
>>> +    u16 status = NVME_SC_SUCCESS;
>>> +    u64 lost_count;
>>> +    u64 cur_count;
>>> +    u64 next_count;
>>> +
>>> +    mutex_lock(&log_mgr->lock);
>>> +    if (!kfifo_get(&log_mgr->log_queue, &log))
>>> +        goto out;
>>> +
>>> +    /*
>>> +     * We can't get the last in kfifo.
>>> +     * Utilize the current count and the count from the next log to
>>> +     * calculate the number of lost logs, while also addressing cases
>>> +     * of overflow. If there is no subsequent log, the number of lost
>>> +     * logs is equal to the lost_count within the nvmet_pr_log_mgr.
>>> +     */
>>> +    cur_count = le64_to_cpu(log.count);
>>> +    if (kfifo_peek(&log_mgr->log_queue, &next_log)) {
>>> +        next_count = le64_to_cpu(next_log.count);
>>> +        if (next_count > cur_count)
>>> +            lost_count = next_count - cur_count - 1;
>>> +        else
>>> +            lost_count = U64_MAX - cur_count + next_count - 1;
>>> +    } else {
>>> +        lost_count = log_mgr->lost_count;
>>> +    }
>>> +
>>> +    log.count = cpu_to_le64((cur_count + lost_count) == 0 ?
>>> +                1 : (cur_count + lost_count));
>>> +    log_mgr->lost_count -= lost_count;
>>> +
>>> +    log.nr_pages = kfifo_len(&log_mgr->log_queue);
>>> +
>>> +out:
>>> +    status = nvmet_copy_to_sgl(req, 0, &log, sizeof(log));
>>> +    mutex_unlock(&log_mgr->lock);
>>> +    nvmet_req_complete(req, status);
>>> +}
>>> +
>>> +static void nvmet_pr_add_resv_log(struct nvmet_ctrl *ctrl, u8 
>>> log_type,
>>> +                  u32 nsid)
>>> +{
>>> +    struct nvmet_pr_log_mgr *log_mgr = &ctrl->pr_log_mgr;
>>> +    struct nvme_pr_log log = {0};
>>> +
>>> +    mutex_lock(&log_mgr->lock);
>>> +    log_mgr->counter++;
>>> +    if (log_mgr->counter == 0)
>>> +        log_mgr->counter = 1;
>>> +
>>> +    log.count = cpu_to_le64(log_mgr->counter);
>>> +    log.type = log_type;
>>> +    log.nsid = cpu_to_le32(nsid);
>>> +
>>> +    if (!kfifo_put(&log_mgr->log_queue, log)) {
>>> +        pr_info("a reservation log lost, cntlid:%d, log_type:%d, 
>>> nsid:%d\n",
>>> +            ctrl->cntlid, log_type, nsid);
>>> +        log_mgr->lost_count++;
>>> +    }
>>> +
>>> +    mutex_unlock(&log_mgr->lock);
>>> +}
>>> +
>>> +static void nvmet_pr_resv_released(struct nvmet_pr *pr, uuid_t 
>>> *hostid)
>>> +{
>>> +    struct nvmet_ns *ns = nvmet_pr_to_ns(pr);
>>> +    struct nvmet_subsys *subsys = ns->subsys;
>>> +    struct nvmet_ctrl *ctrl;
>>> +
>>> +    if (test_bit(NVME_PR_NOTIFY_BIT_RESV_RELEASED, &pr->notify_mask))
>>> +        return;
>>> +
>>> +    mutex_lock(&subsys->lock);
>>> +    list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
>>> +        if (!uuid_equal(&ctrl->hostid, hostid) &&
>>> +            nvmet_pr_find_registrant(pr, &ctrl->hostid)) {
>>> +            nvmet_pr_add_resv_log(ctrl,
>>> +                NVME_PR_LOG_RESERVATION_RELEASED, ns->nsid);
>>> +            nvmet_add_async_event(ctrl, NVME_AER_CSS,
>>> +                NVME_AEN_RESV_LOG_PAGE_AVALIABLE,
>>> +                NVME_LOG_RESERVATION);
>>> +        }
>>> +    }
>>> +    mutex_unlock(&subsys->lock);
>>> +}
>>> +
>>> +static void nvmet_pr_send_event_to_host(struct nvmet_pr *pr, uuid_t 
>>> *hostid,
>>> +                      u8 log_type)
>>> +{
>>> +    struct nvmet_ns *ns = nvmet_pr_to_ns(pr);
>>> +    struct nvmet_subsys *subsys = ns->subsys;
>>> +    struct nvmet_ctrl *ctrl;
>>> +
>>> +    mutex_lock(&subsys->lock);
>>> +    list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
>>> +        if (uuid_equal(hostid, &ctrl->hostid)) {
>>> +            nvmet_pr_add_resv_log(ctrl, log_type, ns->nsid);
>>> +            nvmet_add_async_event(ctrl, NVME_AER_CSS,
>>> +                NVME_AEN_RESV_LOG_PAGE_AVALIABLE,
>>> +                NVME_LOG_RESERVATION);
>>> +        }
>>> +    }
>>> +    mutex_unlock(&subsys->lock);
>>> +}
>>> +
>>> +static void nvmet_pr_resv_preempted(struct nvmet_pr *pr, uuid_t 
>>> *hostid)
>>> +{
>>> +    if (test_bit(NVME_PR_NOTIFY_BIT_RESV_PREEMPTED, &pr->notify_mask))
>>> +        return;
>>> +
>>> +    nvmet_pr_send_event_to_host(pr, hostid,
>>> +        NVME_PR_LOG_RESERVATOPM_PREEMPTED);
>>> +}
>>> +
>>> +static void nvmet_pr_registration_preempted(struct nvmet_pr *pr,
>>> +                        uuid_t *hostid)
>>> +{
>>> +    if (test_bit(NVME_PR_NOTIFY_BIT_REG_PREEMPTED, &pr->notify_mask))
>>> +        return;
>>> +
>>> +    nvmet_pr_send_event_to_host(pr, hostid,
>>> +        NVME_PR_LOG_REGISTRATION_PREEMPTED);
>>> +}
>>> +
>>> +static inline void nvmet_pr_set_new_holder(struct nvmet_pr *pr, u8 
>>> new_rtype,
>>> +                       struct nvmet_pr_registrant *reg)
>>> +{
>>> +    reg->rtype = new_rtype;
>>> +    rcu_assign_pointer(pr->holder, reg);
>>> +}
>>> +
>>> +static u16 nvmet_pr_register(struct nvmet_req *req,
>>> +                 struct nvmet_pr_register_data *d)
>>> +{
>>> +    struct nvmet_ctrl *ctrl = req->sq->ctrl;
>>> +    struct nvmet_pr_registrant *new, *reg;
>>> +    struct nvmet_pr *pr = &req->ns->pr;
>>> +    u16 status = NVME_SC_SUCCESS;
>>> +    u64 nrkey = le64_to_cpu(d->nrkey);
>>> +
>>> +    new = kmalloc(sizeof(*new), GFP_KERNEL);
>>> +    if (!new)
>>> +        return NVME_SC_INTERNAL;
>>> +
>>> +    mutex_lock(&pr->pr_lock);
>>> +    reg = nvmet_pr_find_registrant(pr, &ctrl->hostid);
>>> +    if (reg) {
>>> +        if (reg->rkey != nrkey)
>>> +            status = NVME_SC_RESERVATION_CONFLICT | NVME_STATUS_DNR;
>>> +        kfree(new);
>>> +        goto out;
>>> +    }
>>> +
>>> +    memset(new, 0, sizeof(*new));
>>> +    INIT_LIST_HEAD(&new->entry);
>>> +    new->rkey = nrkey;
>>> +    uuid_copy(&new->hostid, &ctrl->hostid);
>>> +    list_add_tail_rcu(&new->entry, &pr->registrant_list);
>>> +
>>> +out:
>>> +    mutex_unlock(&pr->pr_lock);
>>> +    return status;
>>> +}
>>> +
>>> +static void nvmet_pr_unregister_one(struct nvmet_pr *pr,
>>> +                    struct nvmet_pr_registrant *reg)
>>> +{
>>> +    struct nvmet_pr_registrant *first_reg;
>>> +    struct nvmet_pr_registrant *holder;
>>> +    u8 original_rtype;
>>> +
>>> +    lockdep_assert_held(&pr->pr_lock);
>>> +    list_del_rcu(&reg->entry);
>>> +
>>> +    holder = rcu_dereference_protected(pr->holder,
>>> +            lockdep_is_held(&pr->pr_lock));
>>> +    if (reg != holder)
>>> +        goto out;
>>> +
>>> +    original_rtype = holder->rtype;
>>> +    if (original_rtype == NVME_PR_WRITE_EXCLUSIVE_ALL_REGS ||
>>> +        original_rtype == NVME_PR_EXCLUSIVE_ACCESS_ALL_REGS) {
>>> +        first_reg = list_first_or_null_rcu(&pr->registrant_list,
>>> +                struct nvmet_pr_registrant, entry);
>>> +        if (first_reg)
>>> +            first_reg->rtype = original_rtype;
>>> +        rcu_assign_pointer(pr->holder, first_reg);
>>> +    } else {
>>> +        rcu_assign_pointer(pr->holder, NULL);
>>> +
>>> +        if (original_rtype == NVME_PR_WRITE_EXCLUSIVE_REG_ONLY ||
>>> +            original_rtype == NVME_PR_EXCLUSIVE_ACCESS_REG_ONLY)
>>> +            nvmet_pr_resv_released(pr, &reg->hostid);
>>> +    }
>>> +out:
>>> +    kfree_rcu(reg, rcu);
>>> +}
>>> +
>>> +static u16 nvmet_pr_unregister(struct nvmet_req *req,
>>> +                   struct nvmet_pr_register_data *d,
>>> +                   bool ignore_key)
>>> +{
>>> +    u16 status = NVME_SC_RESERVATION_CONFLICT | NVME_STATUS_DNR;
>>> +    struct nvmet_ctrl *ctrl = req->sq->ctrl;
>>> +    struct nvmet_pr *pr = &req->ns->pr;
>>> +    struct nvmet_pr_registrant *reg;
>>> +
>>> +    mutex_lock(&pr->pr_lock);
>>> +    list_for_each_entry_rcu(reg, &pr->registrant_list, entry,
>>> +                lockdep_is_held(&pr->pr_lock)) {
>>> +        if (uuid_equal(&reg->hostid, &ctrl->hostid)) {
>>> +            if (ignore_key || reg->rkey == le64_to_cpu(d->crkey)) {
>>> +                status = NVME_SC_SUCCESS;
>>> +                nvmet_pr_unregister_one(pr, reg);
>>> +            }
>>> +            break;
>>> +        }
>>> +    }
>>> +    mutex_unlock(&pr->pr_lock);
>>> +
>>> +    return status;
>>> +}
>>> +
>>> +static void nvmet_pr_update_reg_rkey(struct nvmet_pr_registrant *reg,
>>> +                     void *attr)
>>> +{
>>> +    reg->rkey = *(u64 *)attr;
>>> +}
>>> +
>>> +static u16 nvmet_pr_update_reg_attr(struct nvmet_pr *pr,
>>> +            struct nvmet_pr_registrant *reg,
>>> +            void (*change_attr)(struct nvmet_pr_registrant *reg,
>>> +            void *attr),
>>> +            void *attr)
>>> +{
>>> +    struct nvmet_pr_registrant *holder;
>>> +    struct nvmet_pr_registrant *new;
>>> +
>>> +    lockdep_assert_held(&pr->pr_lock);
>>> +    holder = rcu_dereference_protected(pr->holder,
>>> +            lockdep_is_held(&pr->pr_lock));
>>> +    if (reg != holder) {
>>> +        change_attr(reg, attr);
>>> +        return NVME_SC_SUCCESS;
>>> +    }
>>> +
>>> +    new = kmalloc(sizeof(*new), GFP_ATOMIC);
>>> +    if (!new)
>>> +        return NVME_SC_INTERNAL;
>>> +
>>> +    new->rkey = holder->rkey;
>>> +    new->rtype = holder->rtype;
>>> +    uuid_copy(&new->hostid, &holder->hostid);
>>> +    INIT_LIST_HEAD(&new->entry);
>>> +
>>> +    change_attr(new, attr);
>>> +    list_replace_rcu(&holder->entry, &new->entry);
>>> +    rcu_assign_pointer(pr->holder, new);
>>> +    kfree_rcu(holder, rcu);
>>> +
>>> +    return NVME_SC_SUCCESS;
>>> +}
>>> +
>>> +static u16 nvmet_pr_replace(struct nvmet_req *req,
>>> +                struct nvmet_pr_register_data *d,
>>> +                bool ignore_key)
>>> +{
>>> +    u16 status = NVME_SC_RESERVATION_CONFLICT | NVME_STATUS_DNR;
>>> +    struct nvmet_ctrl *ctrl = req->sq->ctrl;
>>> +    struct nvmet_pr *pr = &req->ns->pr;
>>> +    struct nvmet_pr_registrant *reg;
>>> +    u64 nrkey = le64_to_cpu(d->nrkey);
>>> +
>>> +    mutex_lock(&pr->pr_lock);
>>> +    list_for_each_entry_rcu(reg, &pr->registrant_list, entry,
>>> +                lockdep_is_held(&pr->pr_lock)) {
>>> +        if (uuid_equal(&reg->hostid, &ctrl->hostid)) {
>>> +            if (ignore_key || reg->rkey == le64_to_cpu(d->crkey))
>>> +                status = nvmet_pr_update_reg_attr(pr, reg,
>>> +                        nvmet_pr_update_reg_rkey,
>>> +                        &nrkey);
>>> +            break;
>>> +        }
>>> +    }
>>> +    mutex_unlock(&pr->pr_lock);
>>> +    return status;
>>> +}
>>> +
>>> +static void nvmet_execute_pr_register(struct nvmet_req *req)
>>> +{
>>> +    u32 cdw10 = le32_to_cpu(req->cmd->common.cdw10);
>>> +    bool ignore_key = nvmet_pr_parse_ignore_key(cdw10);
>>> +    struct nvmet_pr_register_data *d;
>>> +    u8 reg_act = cdw10 & 0x07; /* Reservation Register Action, bit 
>>> 02:00 */
>>> +    u16 status;
>>> +
>>> +    d = kmalloc(sizeof(*d), GFP_KERNEL);
>>> +    if (!d) {
>>> +        status = NVME_SC_INTERNAL;
>>> +        goto out;
>>> +    }
>>> +
>>> +    status = nvmet_copy_from_sgl(req, 0, d, sizeof(*d));
>>> +    if (status)
>>> +        goto free_data;
>>> +
>>> +    switch (reg_act) {
>>> +    case NVME_PR_REGISTER_ACT_REG:
>>> +        status = nvmet_pr_register(req, d);
>>> +        break;
>>> +    case NVME_PR_REGISTER_ACT_UNREG:
>>> +        status = nvmet_pr_unregister(req, d, ignore_key);
>>> +        break;
>>> +    case NVME_PR_REGISTER_ACT_REPLACE:
>>> +        status = nvmet_pr_replace(req, d, ignore_key);
>>> +        break;
>>> +    default:
>>> +        req->error_loc = offsetof(struct nvme_common_command, cdw10);
>>> +        status = NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
>>> +        break;
>>> +    }
>>> +free_data:
>>> +    kfree(d);
>>> +out:
>>> +    if (!status)
>>> +        atomic_inc(&req->ns->pr.generation);
>>> +    nvmet_req_complete(req, status);
>>> +}
>>> +
>>> +static u16 nvmet_pr_acquire(struct nvmet_req *req,
>>> +                struct nvmet_pr_registrant *reg,
>>> +                u8 rtype)
>>> +{
>>> +    struct nvmet_pr *pr = &req->ns->pr;
>>> +    struct nvmet_pr_registrant *holder;
>>> +
>>> +    lockdep_assert_held(&pr->pr_lock);
>>> +    holder = rcu_dereference_protected(pr->holder,
>>> +            lockdep_is_held(&pr->pr_lock));
>>> +    if (holder && reg != holder)
>>> +        return  NVME_SC_RESERVATION_CONFLICT | NVME_STATUS_DNR;
>>> +    if (holder && reg == holder) {
>>> +        if (holder->rtype == rtype)
>>> +            return NVME_SC_SUCCESS;
>>> +        return NVME_SC_RESERVATION_CONFLICT | NVME_STATUS_DNR;
>>> +    }
>>> +
>>> +    nvmet_pr_set_new_holder(pr, rtype, reg);
>>> +    return NVME_SC_SUCCESS;
>>> +}
>>> +
>>> +static void nvmet_pr_confirm_ns_pc_ref(struct percpu_ref *ref)
>>> +{
>>> +    struct nvmet_pr_per_ctrl_ref *pc_ref =
>>> +        container_of(ref, struct nvmet_pr_per_ctrl_ref, ref);
>>> +
>>> +    complete(&pc_ref->confirm_done);
>>> +}
>>> +
>>> +static void nvmet_pr_set_ctrl_to_abort(struct nvmet_req *req, 
>>> uuid_t *hostid)
>>> +{
>>> +    struct nvmet_pr_per_ctrl_ref *pc_ref;
>>> +    struct nvmet_ns *ns = req->ns;
>>> +    unsigned long idx;
>>> +
>>> +    xa_for_each(&ns->pr_per_ctrl_refs, idx, pc_ref) {
>>> +        if (uuid_equal(&pc_ref->hostid, hostid)) {
>>> +            percpu_ref_kill_and_confirm(&pc_ref->ref,
>>> +                        nvmet_pr_confirm_ns_pc_ref);
>>> + wait_for_completion(&pc_ref->confirm_done);
>>> +        }
>>> +    }
>>> +}
>>> +
>>> +static u16 nvmet_pr_unreg_all_host_by_prkey(struct nvmet_req *req, 
>>> u64 prkey,
>>> +                        uuid_t *send_hostid,
>>> +                        bool abort)
>>> +{
>>> +    u16 status = NVME_SC_RESERVATION_CONFLICT | NVME_STATUS_DNR;
>>> +    struct nvmet_pr_registrant *reg, *tmp;
>>> +    struct nvmet_pr *pr = &req->ns->pr;
>>> +    uuid_t hostid;
>>> +
>>> +    lockdep_assert_held(&pr->pr_lock);
>>> +
>>> +    list_for_each_entry_safe(reg, tmp, &pr->registrant_list, entry) {
>>> +        if (reg->rkey == prkey) {
>>> +            status = NVME_SC_SUCCESS;
>>> +            uuid_copy(&hostid, &reg->hostid);
>>> +            if (abort)
>>> +                nvmet_pr_set_ctrl_to_abort(req, &hostid);
>>> +            nvmet_pr_unregister_one(pr, reg);
>>> +            if (!uuid_equal(&hostid, send_hostid))
>>> +                nvmet_pr_registration_preempted(pr, &hostid);
>>> +        }
>>> +    }
>>> +    return status;
>>> +}
>>> +
>>> +static void nvmet_pr_unreg_all_others_by_prkey(struct nvmet_req *req,
>>> +                           u64 prkey,
>>> +                           uuid_t *send_hostid,
>>> +                           bool abort)
>>> +{
>>> +    struct nvmet_pr_registrant *reg, *tmp;
>>> +    struct nvmet_pr *pr = &req->ns->pr;
>>> +    uuid_t hostid;
>>> +
>>> +    lockdep_assert_held(&pr->pr_lock);
>>> +
>>> +    list_for_each_entry_safe(reg, tmp, &pr->registrant_list, entry) {
>>> +        if (reg->rkey == prkey &&
>>> +            !uuid_equal(&reg->hostid, send_hostid)) {
>>> +            uuid_copy(&hostid, &reg->hostid);
>>> +            if (abort)
>>> +                nvmet_pr_set_ctrl_to_abort(req, &hostid);
>>> +            nvmet_pr_unregister_one(pr, reg);
>>> +            nvmet_pr_registration_preempted(pr, &hostid);
>>> +        }
>>> +    }
>>> +}
>>> +
>>> +static void nvmet_pr_unreg_all_others(struct nvmet_req *req,
>>> +                      uuid_t *send_hostid,
>>> +                      bool abort)
>>> +{
>>> +    struct nvmet_pr_registrant *reg, *tmp;
>>> +    struct nvmet_pr *pr = &req->ns->pr;
>>> +    uuid_t hostid;
>>> +
>>> +    lockdep_assert_held(&pr->pr_lock);
>>> +
>>> +    list_for_each_entry_safe(reg, tmp, &pr->registrant_list, entry) {
>>> +        if (!uuid_equal(&reg->hostid, send_hostid)) {
>>> +            uuid_copy(&hostid, &reg->hostid);
>>> +            if (abort)
>>> +                nvmet_pr_set_ctrl_to_abort(req, &hostid);
>>> +            nvmet_pr_unregister_one(pr, reg);
>>> +            nvmet_pr_registration_preempted(pr, &hostid);
>>> +        }
>>> +    }
>>> +}
>>> +
>>> +static void nvmet_pr_update_holder_rtype(struct nvmet_pr_registrant 
>>> *reg,
>>> +                     void *attr)
>>> +{
>>> +    u8 new_rtype = *(u8 *)attr;
>>> +
>>> +    reg->rtype = new_rtype;
>>> +}
>>> +
>>> +static u16 nvmet_pr_preempt(struct nvmet_req *req,
>>> +                struct nvmet_pr_registrant *reg,
>>> +                u8 rtype,
>>> +                struct nvmet_pr_acquire_data *d,
>>> +                bool abort)
>>> +{
>>> +    struct nvmet_ctrl *ctrl = req->sq->ctrl;
>>> +    struct nvmet_pr *pr = &req->ns->pr;
>>> +    struct nvmet_pr_registrant *holder;
>>> +    enum nvme_pr_type original_rtype;
>>> +    u64 prkey = le64_to_cpu(d->prkey);
>>> +    u16 status;
>>> +
>>> +    lockdep_assert_held(&pr->pr_lock);
>>> +    holder = rcu_dereference_protected(pr->holder,
>>> +            lockdep_is_held(&pr->pr_lock));
>>> +    if (!holder)
>>> +        return nvmet_pr_unreg_all_host_by_prkey(req, prkey,
>>> +                    &ctrl->hostid, abort);
>>> +
>>> +    original_rtype = holder->rtype;
>>> +    if (original_rtype == NVME_PR_WRITE_EXCLUSIVE_ALL_REGS ||
>>> +        original_rtype == NVME_PR_EXCLUSIVE_ACCESS_ALL_REGS) {
>>> +        if (!prkey) {
>>> +            /*
>>> +             * To prevent possible access from other hosts, and
>>> +             * avoid terminate the holder, set the new holder
>>> +             * first before unregistering.
>>> +             */
>>> +            nvmet_pr_set_new_holder(pr, rtype, reg);
>>> +            nvmet_pr_unreg_all_others(req, &ctrl->hostid, abort);
>>> +            return NVME_SC_SUCCESS;
>>> +        }
>>> +        return nvmet_pr_unreg_all_host_by_prkey(req, prkey,
>>> +                &ctrl->hostid, abort);
>>> +    }
>>> +
>>> +    if (holder == reg) {
>>> +        status = nvmet_pr_update_reg_attr(pr, holder,
>>> +                nvmet_pr_update_holder_rtype, &rtype);
>>> +        if (!status && original_rtype != rtype)
>>> +            nvmet_pr_resv_released(pr, &reg->hostid);
>>> +        return status;
>>> +    }
>>> +
>>> +    if (prkey == holder->rkey) {
>>> +        /*
>>> +         * Same as before, set the new holder first.
>>> +         */
>>> +        nvmet_pr_set_new_holder(pr, rtype, reg);
>>> +        nvmet_pr_unreg_all_others_by_prkey(req, prkey, &ctrl->hostid,
>>> +                        abort);
>>> +        if (original_rtype != rtype)
>>> +            nvmet_pr_resv_released(pr, &reg->hostid);
>>> +        return NVME_SC_SUCCESS;
>>> +    }
>>> +
>>> +    if (prkey)
>>> +        return nvmet_pr_unreg_all_host_by_prkey(req, prkey,
>>> +                    &ctrl->hostid, abort);
>>> +    return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
>>> +}
>>> +
>>> +static void nvmet_pr_do_abort(struct nvmet_req *req)
>>> +{
>>> +    struct nvmet_pr_per_ctrl_ref *pc_ref;
>>> +    struct nvmet_ns *ns = req->ns;
>>> +    unsigned long idx;
>>> +
>>> +    /*
>>> +     * The target does not support abort, just wait per-controller 
>>> ref to 0.
>>> +     */
>>> +    xa_for_each(&ns->pr_per_ctrl_refs, idx, pc_ref) {
>>> +        if (percpu_ref_is_dying(&pc_ref->ref)) {
>>
>> Is this possible that we get here without the ref dying? Maybe a warn
>> is appropriate here? just feels incorrect to blindly do nothing and just
>> skip...
>>
> The pc_ref will only be in a dying state when the command is preempt 
> and abort,
>
> and when the controller's reservation is preempted or released.
>
> I killed it in the nvmet_pr_set_ctrl_to_abort() function. As for
>
> the pc_ref of other controllers that are not affected, they are
>
> not in a dying state. In this case, we don't need to wait;
>
> skipping is the correct approach.

Umm, not fully clear. in which case do you have preempt_and_abort which not
all ns->pr_per_ctrl_refs need to quiesce (i.e. do the abort)?

>
>>> + wait_for_completion(&pc_ref->free_done);
>>> +            reinit_completion(&pc_ref->confirm_done);
>>> +            reinit_completion(&pc_ref->free_done);
>>> +            percpu_ref_resurrect(&pc_ref->ref);
>>> +        }
>>> +    }
>>
>> I'm wandering if the do_abort should be deferred to a wq? this will
>> block the transport thread for a long time, preventing progress
>> for other namespaces that can accept IO perhaps?
>>
> Here we are already in wq, ib_cq_poll_work and nvmet_tcp_io_work,

I know, but the transport wq context will not process any other command 
until
this one is done. See for example how nvmet file buffered io is done.

>
> still we still need another wq? If need, I will send the v17 patch to 
> change.

You have nvmet_wq, you can queue it there.



More information about the Linux-nvme mailing list