]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blobdiff - drivers/scsi/cxlflash/main.c
scsi: cxlflash: Separate AFU internal command handling from AFU sync specifics
[mirror_ubuntu-zesty-kernel.git] / drivers / scsi / cxlflash / main.c
index 3c4a83307e6e534c230aaabf0ccc1d658d636b9d..7732dfc099c42d29745d0ab79d6f9aa1a24f21d4 100644 (file)
@@ -34,6 +34,10 @@ MODULE_AUTHOR("Manoj N. Kumar <manoj@linux.vnet.ibm.com>");
 MODULE_AUTHOR("Matthew R. Ochs <mrochs@linux.vnet.ibm.com>");
 MODULE_LICENSE("GPL");
 
+static struct class *cxlflash_class;
+static u32 cxlflash_major;
+static DECLARE_BITMAP(cxlflash_minor, CXLFLASH_MAX_ADAPTERS);
+
 /**
  * process_cmd_err() - command error handler
  * @cmd:       AFU command that experienced the error.
@@ -162,8 +166,13 @@ static void cmd_complete(struct afu_cmd *cmd)
        struct afu *afu = cmd->parent;
        struct cxlflash_cfg *cfg = afu->parent;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq = get_hwq(afu, cmd->hwq_index);
        bool cmd_is_tmf;
 
+       spin_lock_irqsave(&hwq->hsq_slock, lock_flags);
+       list_del(&cmd->list);
+       spin_unlock_irqrestore(&hwq->hsq_slock, lock_flags);
+
        if (cmd->scp) {
                scp = cmd->scp;
                if (unlikely(cmd->sa.ioasc))
@@ -189,53 +198,101 @@ static void cmd_complete(struct afu_cmd *cmd)
 }
 
 /**
- * context_reset() - reset command owner context via specified register
- * @cmd:       AFU command that timed out.
+ * flush_pending_cmds() - flush all pending commands on this hardware queue
+ * @hwq:       Hardware queue to flush.
+ *
+ * The hardware send queue lock associated with this hardware queue must be
+ * held when calling this routine.
+ */
+static void flush_pending_cmds(struct hwq *hwq)
+{
+       struct afu_cmd *cmd, *tmp;
+       struct scsi_cmnd *scp;
+
+       list_for_each_entry_safe(cmd, tmp, &hwq->pending_cmds, list) {
+               /* Bypass command when on a doneq, cmd_complete() will handle */
+               if (!list_empty(&cmd->queue))
+                       continue;
+
+               list_del(&cmd->list);
+
+               if (cmd->scp) {
+                       scp = cmd->scp;
+                       scp->result = (DID_IMM_RETRY << 16);
+                       scp->scsi_done(scp);
+               } else {
+                       cmd->cmd_aborted = true;
+                       complete(&cmd->cevent);
+               }
+       }
+}
+
+/**
+ * context_reset() - reset context via specified register
+ * @hwq:       Hardware queue owning the context to be reset.
  * @reset_reg: MMIO register to perform reset.
+ *
+ * When the reset is successful, the SISLite specification guarantees that
+ * the AFU has aborted all currently pending I/O. Accordingly, these commands
+ * must be flushed.
+ *
+ * Return: 0 on success, -errno on failure
  */
-static void context_reset(struct afu_cmd *cmd, __be64 __iomem *reset_reg)
+static int context_reset(struct hwq *hwq, __be64 __iomem *reset_reg)
 {
-       int nretry = 0;
-       u64 rrin = 0x1;
-       struct afu *afu = cmd->parent;
-       struct cxlflash_cfg *cfg = afu->parent;
+       struct cxlflash_cfg *cfg = hwq->afu->parent;
        struct device *dev = &cfg->dev->dev;
+       int rc = -ETIMEDOUT;
+       int nretry = 0;
+       u64 val = 0x1;
+       ulong lock_flags;
+
+       dev_dbg(dev, "%s: hwq=%p\n", __func__, hwq);
 
-       dev_dbg(dev, "%s: cmd=%p\n", __func__, cmd);
+       spin_lock_irqsave(&hwq->hsq_slock, lock_flags);
 
-       writeq_be(rrin, reset_reg);
+       writeq_be(val, reset_reg);
        do {
-               rrin = readq_be(reset_reg);
-               if (rrin != 0x1)
+               val = readq_be(reset_reg);
+               if ((val & 0x1) == 0x0) {
+                       rc = 0;
                        break;
+               }
+
                /* Double delay each time */
                udelay(1 << nretry);
        } while (nretry++ < MC_ROOM_RETRY_CNT);
 
-       dev_dbg(dev, "%s: returning rrin=%016llx nretry=%d\n",
-               __func__, rrin, nretry);
+       if (!rc)
+               flush_pending_cmds(hwq);
+
+       spin_unlock_irqrestore(&hwq->hsq_slock, lock_flags);
+
+       dev_dbg(dev, "%s: returning rc=%d, val=%016llx nretry=%d\n",
+               __func__, rc, val, nretry);
+       return rc;
 }
 
 /**
- * context_reset_ioarrin() - reset command owner context via IOARRIN register
- * @cmd:       AFU command that timed out.
+ * context_reset_ioarrin() - reset context via IOARRIN register
+ * @hwq:       Hardware queue owning the context to be reset.
+ *
+ * Return: 0 on success, -errno on failure
  */
-static void context_reset_ioarrin(struct afu_cmd *cmd)
+static int context_reset_ioarrin(struct hwq *hwq)
 {
-       struct afu *afu = cmd->parent;
-
-       context_reset(cmd, &afu->host_map->ioarrin);
+       return context_reset(hwq, &hwq->host_map->ioarrin);
 }
 
 /**
- * context_reset_sq() - reset command owner context w/ SQ Context Reset register
- * @cmd:       AFU command that timed out.
+ * context_reset_sq() - reset context via SQ_CONTEXT_RESET register
+ * @hwq:       Hardware queue owning the context to be reset.
+ *
+ * Return: 0 on success, -errno on failure
  */
-static void context_reset_sq(struct afu_cmd *cmd)
+static int context_reset_sq(struct hwq *hwq)
 {
-       struct afu *afu = cmd->parent;
-
-       context_reset(cmd, &afu->host_map->sq_ctx_reset);
+       return context_reset(hwq, &hwq->host_map->sq_ctx_reset);
 }
 
 /**
@@ -250,6 +307,7 @@ static int send_cmd_ioarrin(struct afu *afu, struct afu_cmd *cmd)
 {
        struct cxlflash_cfg *cfg = afu->parent;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq = get_hwq(afu, cmd->hwq_index);
        int rc = 0;
        s64 room;
        ulong lock_flags;
@@ -258,23 +316,24 @@ static int send_cmd_ioarrin(struct afu *afu, struct afu_cmd *cmd)
         * To avoid the performance penalty of MMIO, spread the update of
         * 'room' over multiple commands.
         */
-       spin_lock_irqsave(&afu->rrin_slock, lock_flags);
-       if (--afu->room < 0) {
-               room = readq_be(&afu->host_map->cmd_room);
+       spin_lock_irqsave(&hwq->hsq_slock, lock_flags);
+       if (--hwq->room < 0) {
+               room = readq_be(&hwq->host_map->cmd_room);
                if (room <= 0) {
                        dev_dbg_ratelimited(dev, "%s: no cmd_room to send "
                                            "0x%02X, room=0x%016llX\n",
                                            __func__, cmd->rcb.cdb[0], room);
-                       afu->room = 0;
+                       hwq->room = 0;
                        rc = SCSI_MLQUEUE_HOST_BUSY;
                        goto out;
                }
-               afu->room = room - 1;
+               hwq->room = room - 1;
        }
 
-       writeq_be((u64)&cmd->rcb, &afu->host_map->ioarrin);
+       list_add(&cmd->list, &hwq->pending_cmds);
+       writeq_be((u64)&cmd->rcb, &hwq->host_map->ioarrin);
 out:
-       spin_unlock_irqrestore(&afu->rrin_slock, lock_flags);
+       spin_unlock_irqrestore(&hwq->hsq_slock, lock_flags);
        dev_dbg(dev, "%s: cmd=%p len=%u ea=%016llx rc=%d\n", __func__,
                cmd, cmd->rcb.data_len, cmd->rcb.data_ea, rc);
        return rc;
@@ -292,11 +351,12 @@ static int send_cmd_sq(struct afu *afu, struct afu_cmd *cmd)
 {
        struct cxlflash_cfg *cfg = afu->parent;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq = get_hwq(afu, cmd->hwq_index);
        int rc = 0;
        int newval;
        ulong lock_flags;
 
-       newval = atomic_dec_if_positive(&afu->hsq_credits);
+       newval = atomic_dec_if_positive(&hwq->hsq_credits);
        if (newval <= 0) {
                rc = SCSI_MLQUEUE_HOST_BUSY;
                goto out;
@@ -304,22 +364,24 @@ static int send_cmd_sq(struct afu *afu, struct afu_cmd *cmd)
 
        cmd->rcb.ioasa = &cmd->sa;
 
-       spin_lock_irqsave(&afu->hsq_slock, lock_flags);
+       spin_lock_irqsave(&hwq->hsq_slock, lock_flags);
 
-       *afu->hsq_curr = cmd->rcb;
-       if (afu->hsq_curr < afu->hsq_end)
-               afu->hsq_curr++;
+       *hwq->hsq_curr = cmd->rcb;
+       if (hwq->hsq_curr < hwq->hsq_end)
+               hwq->hsq_curr++;
        else
-               afu->hsq_curr = afu->hsq_start;
-       writeq_be((u64)afu->hsq_curr, &afu->host_map->sq_tail);
+               hwq->hsq_curr = hwq->hsq_start;
+
+       list_add(&cmd->list, &hwq->pending_cmds);
+       writeq_be((u64)hwq->hsq_curr, &hwq->host_map->sq_tail);
 
-       spin_unlock_irqrestore(&afu->hsq_slock, lock_flags);
+       spin_unlock_irqrestore(&hwq->hsq_slock, lock_flags);
 out:
        dev_dbg(dev, "%s: cmd=%p len=%u ea=%016llx ioasa=%p rc=%d curr=%p "
               "head=%016llx tail=%016llx\n", __func__, cmd, cmd->rcb.data_len,
-              cmd->rcb.data_ea, cmd->rcb.ioasa, rc, afu->hsq_curr,
-              readq_be(&afu->host_map->sq_head),
-              readq_be(&afu->host_map->sq_tail));
+              cmd->rcb.data_ea, cmd->rcb.ioasa, rc, hwq->hsq_curr,
+              readq_be(&hwq->host_map->sq_head),
+              readq_be(&hwq->host_map->sq_tail));
        return rc;
 }
 
@@ -328,8 +390,7 @@ out:
  * @afu:       AFU associated with the host.
  * @cmd:       AFU command that was sent.
  *
- * Return:
- *     0 on success, -1 on timeout/error
+ * Return: 0 on success, -errno on failure
  */
 static int wait_resp(struct afu *afu, struct afu_cmd *cmd)
 {
@@ -339,20 +400,58 @@ static int wait_resp(struct afu *afu, struct afu_cmd *cmd)
        ulong timeout = msecs_to_jiffies(cmd->rcb.timeout * 2 * 1000);
 
        timeout = wait_for_completion_timeout(&cmd->cevent, timeout);
-       if (!timeout) {
-               afu->context_reset(cmd);
-               rc = -1;
-       }
+       if (!timeout)
+               rc = -ETIMEDOUT;
+
+       if (cmd->cmd_aborted)
+               rc = -EAGAIN;
 
        if (unlikely(cmd->sa.ioasc != 0)) {
                dev_err(dev, "%s: cmd %02x failed, ioasc=%08x\n",
                        __func__, cmd->rcb.cdb[0], cmd->sa.ioasc);
-               rc = -1;
+               rc = -EIO;
        }
 
        return rc;
 }
 
+/**
+ * cmd_to_target_hwq() - selects a target hardware queue for a SCSI command
+ * @host:      SCSI host associated with device.
+ * @scp:       SCSI command to send.
+ * @afu:       SCSI command to send.
+ *
+ * Hashes a command based upon the hardware queue mode.
+ *
+ * Return: Trusted index of target hardware queue
+ */
+static u32 cmd_to_target_hwq(struct Scsi_Host *host, struct scsi_cmnd *scp,
+                            struct afu *afu)
+{
+       u32 tag;
+       u32 hwq = 0;
+
+       if (afu->num_hwqs == 1)
+               return 0;
+
+       switch (afu->hwq_mode) {
+       case HWQ_MODE_RR:
+               hwq = afu->hwq_rr_count++ % afu->num_hwqs;
+               break;
+       case HWQ_MODE_TAG:
+               tag = blk_mq_unique_tag(scp->request);
+               hwq = blk_mq_unique_tag_to_hwq(tag);
+               break;
+       case HWQ_MODE_CPU:
+               hwq = smp_processor_id() % afu->num_hwqs;
+               break;
+       default:
+               WARN_ON_ONCE(1);
+       }
+
+       return hwq;
+}
+
 /**
  * send_tmf() - sends a Task Management Function (TMF)
  * @afu:       AFU to checkout from.
@@ -364,9 +463,12 @@ static int wait_resp(struct afu *afu, struct afu_cmd *cmd)
  */
 static int send_tmf(struct afu *afu, struct scsi_cmnd *scp, u64 tmfcmd)
 {
-       struct cxlflash_cfg *cfg = shost_priv(scp->device->host);
+       struct Scsi_Host *host = scp->device->host;
+       struct cxlflash_cfg *cfg = shost_priv(host);
        struct afu_cmd *cmd = sc_to_afucz(scp);
        struct device *dev = &cfg->dev->dev;
+       int hwq_index = cmd_to_target_hwq(host, scp, afu);
+       struct hwq *hwq = get_hwq(afu, hwq_index);
        ulong lock_flags;
        int rc = 0;
        ulong to;
@@ -383,8 +485,9 @@ static int send_tmf(struct afu *afu, struct scsi_cmnd *scp, u64 tmfcmd)
        cmd->scp = scp;
        cmd->parent = afu;
        cmd->cmd_tmf = true;
+       cmd->hwq_index = hwq_index;
 
-       cmd->rcb.ctx_id = afu->ctx_hndl;
+       cmd->rcb.ctx_id = hwq->ctx_hndl;
        cmd->rcb.msi = SISL_MSI_RRQ_UPDATED;
        cmd->rcb.port_sel = CHAN2PORTMASK(scp->device->channel);
        cmd->rcb.lun_id = lun_to_lunid(scp->device->lun);
@@ -442,6 +545,8 @@ static int cxlflash_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scp)
        struct device *dev = &cfg->dev->dev;
        struct afu_cmd *cmd = sc_to_afucz(scp);
        struct scatterlist *sg = scsi_sglist(scp);
+       int hwq_index = cmd_to_target_hwq(host, scp, afu);
+       struct hwq *hwq = get_hwq(afu, hwq_index);
        u16 req_flags = SISL_REQ_FLAGS_SUP_UNDERRUN;
        ulong lock_flags;
        int rc = 0;
@@ -491,8 +596,9 @@ static int cxlflash_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scp)
 
        cmd->scp = scp;
        cmd->parent = afu;
+       cmd->hwq_index = hwq_index;
 
-       cmd->rcb.ctx_id = afu->ctx_hndl;
+       cmd->rcb.ctx_id = hwq->ctx_hndl;
        cmd->rcb.msi = SISL_MSI_RRQ_UPDATED;
        cmd->rcb.port_sel = CHAN2PORTMASK(scp->device->channel);
        cmd->rcb.lun_id = lun_to_lunid(scp->device->lun);
@@ -536,6 +642,20 @@ static void free_mem(struct cxlflash_cfg *cfg)
        }
 }
 
+/**
+ * cxlflash_reset_sync() - synchronizing point for asynchronous resets
+ * @cfg:       Internal structure associated with the host.
+ */
+static void cxlflash_reset_sync(struct cxlflash_cfg *cfg)
+{
+       if (cfg->async_reset_cookie == 0)
+               return;
+
+       /* Wait until all async calls prior to this cookie have completed */
+       async_synchronize_cookie(cfg->async_reset_cookie + 1);
+       cfg->async_reset_cookie = 0;
+}
+
 /**
  * stop_afu() - stops the AFU command timers and unmaps the MMIO space
  * @cfg:       Internal structure associated with the host.
@@ -548,14 +668,25 @@ static void free_mem(struct cxlflash_cfg *cfg)
 static void stop_afu(struct cxlflash_cfg *cfg)
 {
        struct afu *afu = cfg->afu;
+       struct hwq *hwq;
+       int i;
 
        cancel_work_sync(&cfg->work_q);
+       if (!current_is_async())
+               cxlflash_reset_sync(cfg);
 
        if (likely(afu)) {
                while (atomic_read(&afu->cmds_active))
                        ssleep(1);
-               if (afu_is_irqpoll_enabled(afu))
-                       irq_poll_disable(&afu->irqpoll);
+
+               if (afu_is_irqpoll_enabled(afu)) {
+                       for (i = 0; i < afu->num_hwqs; i++) {
+                               hwq = get_hwq(afu, i);
+
+                               irq_poll_disable(&hwq->irqpoll);
+                       }
+               }
+
                if (likely(afu->afu_map)) {
                        cxl_psa_unmap((void __iomem *)afu->afu_map);
                        afu->afu_map = NULL;
@@ -567,28 +698,40 @@ static void stop_afu(struct cxlflash_cfg *cfg)
  * term_intr() - disables all AFU interrupts
  * @cfg:       Internal structure associated with the host.
  * @level:     Depth of allocation, where to begin waterfall tear down.
+ * @index:     Index of the hardware queue.
  *
  * Safe to call with AFU/MC in partially allocated/initialized state.
  */
-static void term_intr(struct cxlflash_cfg *cfg, enum undo_level level)
+static void term_intr(struct cxlflash_cfg *cfg, enum undo_level level,
+                     u32 index)
 {
        struct afu *afu = cfg->afu;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq;
+
+       if (!afu) {
+               dev_err(dev, "%s: returning with NULL afu\n", __func__);
+               return;
+       }
 
-       if (!afu || !cfg->mcctx) {
-               dev_err(dev, "%s: returning with NULL afu or MC\n", __func__);
+       hwq = get_hwq(afu, index);
+
+       if (!hwq->ctx) {
+               dev_err(dev, "%s: returning with NULL MC\n", __func__);
                return;
        }
 
        switch (level) {
        case UNMAP_THREE:
-               cxl_unmap_afu_irq(cfg->mcctx, 3, afu);
+               /* SISL_MSI_ASYNC_ERROR is setup only for the primary HWQ */
+               if (index == PRIMARY_HWQ)
+                       cxl_unmap_afu_irq(hwq->ctx, 3, hwq);
        case UNMAP_TWO:
-               cxl_unmap_afu_irq(cfg->mcctx, 2, afu);
+               cxl_unmap_afu_irq(hwq->ctx, 2, hwq);
        case UNMAP_ONE:
-               cxl_unmap_afu_irq(cfg->mcctx, 1, afu);
+               cxl_unmap_afu_irq(hwq->ctx, 1, hwq);
        case FREE_IRQ:
-               cxl_free_afu_irqs(cfg->mcctx);
+               cxl_free_afu_irqs(hwq->ctx);
                /* fall through */
        case UNDO_NOOP:
                /* No action required */
@@ -599,24 +742,37 @@ static void term_intr(struct cxlflash_cfg *cfg, enum undo_level level)
 /**
  * term_mc() - terminates the master context
  * @cfg:       Internal structure associated with the host.
- * @level:     Depth of allocation, where to begin waterfall tear down.
+ * @index:     Index of the hardware queue.
  *
  * Safe to call with AFU/MC in partially allocated/initialized state.
  */
-static void term_mc(struct cxlflash_cfg *cfg)
+static void term_mc(struct cxlflash_cfg *cfg, u32 index)
 {
-       int rc = 0;
        struct afu *afu = cfg->afu;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq;
+       ulong lock_flags;
 
-       if (!afu || !cfg->mcctx) {
-               dev_err(dev, "%s: returning with NULL afu or MC\n", __func__);
+       if (!afu) {
+               dev_err(dev, "%s: returning with NULL afu\n", __func__);
                return;
        }
 
-       rc = cxl_stop_context(cfg->mcctx);
-       WARN_ON(rc);
-       cfg->mcctx = NULL;
+       hwq = get_hwq(afu, index);
+
+       if (!hwq->ctx) {
+               dev_err(dev, "%s: returning with NULL MC\n", __func__);
+               return;
+       }
+
+       WARN_ON(cxl_stop_context(hwq->ctx));
+       if (index != PRIMARY_HWQ)
+               WARN_ON(cxl_release_context(hwq->ctx));
+       hwq->ctx = NULL;
+
+       spin_lock_irqsave(&hwq->hsq_slock, lock_flags);
+       flush_pending_cmds(hwq);
+       spin_unlock_irqrestore(&hwq->hsq_slock, lock_flags);
 }
 
 /**
@@ -628,21 +784,25 @@ static void term_mc(struct cxlflash_cfg *cfg)
 static void term_afu(struct cxlflash_cfg *cfg)
 {
        struct device *dev = &cfg->dev->dev;
+       int k;
 
        /*
         * Tear down is carefully orchestrated to ensure
         * no interrupts can come in when the problem state
         * area is unmapped.
         *
-        * 1) Disable all AFU interrupts
+        * 1) Disable all AFU interrupts for each master
         * 2) Unmap the problem state area
-        * 3) Stop the master context
+        * 3) Stop each master context
         */
-       term_intr(cfg, UNMAP_THREE);
+       for (k = cfg->afu->num_hwqs - 1; k >= 0; k--)
+               term_intr(cfg, UNMAP_THREE, k);
+
        if (cfg->afu)
                stop_afu(cfg);
 
-       term_mc(cfg);
+       for (k = cfg->afu->num_hwqs - 1; k >= 0; k--)
+               term_mc(cfg, k);
 
        dev_dbg(dev, "%s: returning\n", __func__);
 }
@@ -706,6 +866,47 @@ static void notify_shutdown(struct cxlflash_cfg *cfg, bool wait)
        }
 }
 
+/**
+ * cxlflash_get_minor() - gets the first available minor number
+ *
+ * Return: Unique minor number that can be used to create the character device.
+ */
+static int cxlflash_get_minor(void)
+{
+       int minor;
+       long bit;
+
+       bit = find_first_zero_bit(cxlflash_minor, CXLFLASH_MAX_ADAPTERS);
+       if (bit >= CXLFLASH_MAX_ADAPTERS)
+               return -1;
+
+       minor = bit & MINORMASK;
+       set_bit(minor, cxlflash_minor);
+       return minor;
+}
+
+/**
+ * cxlflash_put_minor() - releases the minor number
+ * @minor:     Minor number that is no longer needed.
+ */
+static void cxlflash_put_minor(int minor)
+{
+       clear_bit(minor, cxlflash_minor);
+}
+
+/**
+ * cxlflash_release_chrdev() - release the character device for the host
+ * @cfg:       Internal structure associated with the host.
+ */
+static void cxlflash_release_chrdev(struct cxlflash_cfg *cfg)
+{
+       put_device(cfg->chardev);
+       device_unregister(cfg->chardev);
+       cfg->chardev = NULL;
+       cdev_del(&cfg->cdev);
+       cxlflash_put_minor(MINOR(cfg->cdev.dev));
+}
+
 /**
  * cxlflash_remove() - PCI entry point to tear down host
  * @pdev:      PCI device associated with the host.
@@ -741,6 +942,8 @@ static void cxlflash_remove(struct pci_dev *pdev)
        cxlflash_stop_term_user_contexts(cfg);
 
        switch (cfg->init_state) {
+       case INIT_STATE_CDEV:
+               cxlflash_release_chrdev(cfg);
        case INIT_STATE_SCSI:
                cxlflash_term_local_luns(cfg);
                scsi_remove_host(cfg->host);
@@ -782,6 +985,7 @@ static int alloc_mem(struct cxlflash_cfg *cfg)
                goto out;
        }
        cfg->afu->parent = cfg;
+       cfg->afu->desired_hwqs = CXLFLASH_DEF_HWQS;
        cfg->afu->afu_map = NULL;
 out:
        return rc;
@@ -1017,52 +1221,6 @@ static void afu_link_reset(struct afu *afu, int port, __be64 __iomem *fc_regs)
        dev_dbg(dev, "%s: returning port_sel=%016llx\n", __func__, port_sel);
 }
 
-/*
- * Asynchronous interrupt information table
- *
- * NOTE: The checkpatch script considers the BUILD_SISL_ASTATUS_FC_PORT macro
- * as complex and complains because it is not wrapped with parentheses/braces.
- */
-#define ASTATUS_FC(_a, _b, _c, _d)                                      \
-       { SISL_ASTATUS_FC##_a##_##_b, _c, _a, (_d) }
-
-#define BUILD_SISL_ASTATUS_FC_PORT(_a)                                  \
-       ASTATUS_FC(_a, OTHER, "other error", CLR_FC_ERROR | LINK_RESET), \
-       ASTATUS_FC(_a, LOGO, "target initiated LOGO", 0),                \
-       ASTATUS_FC(_a, CRC_T, "CRC threshold exceeded", LINK_RESET),     \
-       ASTATUS_FC(_a, LOGI_R, "login timed out, retrying", LINK_RESET), \
-       ASTATUS_FC(_a, LOGI_F, "login failed", CLR_FC_ERROR),            \
-       ASTATUS_FC(_a, LOGI_S, "login succeeded", SCAN_HOST),            \
-       ASTATUS_FC(_a, LINK_DN, "link down", 0),                         \
-       ASTATUS_FC(_a, LINK_UP, "link up", 0)
-
-static const struct asyc_intr_info ainfo[] = {
-       BUILD_SISL_ASTATUS_FC_PORT(2),
-       BUILD_SISL_ASTATUS_FC_PORT(3),
-       BUILD_SISL_ASTATUS_FC_PORT(0),
-       BUILD_SISL_ASTATUS_FC_PORT(1),
-       { 0x0, "", 0, 0 }
-};
-
-/**
- * find_ainfo() - locates and returns asynchronous interrupt information
- * @status:    Status code set by AFU on error.
- *
- * Return: The located information or NULL when the status code is invalid.
- */
-static const struct asyc_intr_info *find_ainfo(u64 status)
-{
-       const struct asyc_intr_info *info;
-
-       BUILD_BUG_ON(ainfo[ARRAY_SIZE(ainfo) - 1].status != 0);
-
-       for (info = &ainfo[0]; info->status; info++)
-               if (info->status == status)
-                       return info;
-
-       return NULL;
-}
-
 /**
  * afu_err_intr_init() - clears and initializes the AFU for error interrupts
  * @afu:       AFU associated with the host.
@@ -1072,6 +1230,7 @@ static void afu_err_intr_init(struct afu *afu)
        struct cxlflash_cfg *cfg = afu->parent;
        __be64 __iomem *fc_port_regs;
        int i;
+       struct hwq *hwq = get_hwq(afu, PRIMARY_HWQ);
        u64 reg;
 
        /* global async interrupts: AFU clears afu_ctrl on context exit
@@ -1083,8 +1242,8 @@ static void afu_err_intr_init(struct afu *afu)
 
        /* mask all */
        writeq_be(-1ULL, &afu->afu_map->global.regs.aintr_mask);
-       /* set LISN# to send and point to master context */
-       reg = ((u64) (((afu->ctx_hndl << 8) | SISL_MSI_ASYNC_ERROR)) << 40);
+       /* set LISN# to send and point to primary master context */
+       reg = ((u64) (((hwq->ctx_hndl << 8) | SISL_MSI_ASYNC_ERROR)) << 40);
 
        if (afu->internal_lun)
                reg |= 1;       /* Bit 63 indicates local lun */
@@ -1120,8 +1279,12 @@ static void afu_err_intr_init(struct afu *afu)
        /* IOARRIN yet), so there is nothing to clear. */
 
        /* set LISN#, it is always sent to the context that wrote IOARRIN */
-       writeq_be(SISL_MSI_SYNC_ERROR, &afu->host_map->ctx_ctrl);
-       writeq_be(SISL_ISTATUS_MASK, &afu->host_map->intr_mask);
+       for (i = 0; i < afu->num_hwqs; i++) {
+               hwq = get_hwq(afu, i);
+
+               writeq_be(SISL_MSI_SYNC_ERROR, &hwq->host_map->ctx_ctrl);
+               writeq_be(SISL_ISTATUS_MASK, &hwq->host_map->intr_mask);
+       }
 }
 
 /**
@@ -1133,13 +1296,13 @@ static void afu_err_intr_init(struct afu *afu)
  */
 static irqreturn_t cxlflash_sync_err_irq(int irq, void *data)
 {
-       struct afu *afu = (struct afu *)data;
-       struct cxlflash_cfg *cfg = afu->parent;
+       struct hwq *hwq = (struct hwq *)data;
+       struct cxlflash_cfg *cfg = hwq->afu->parent;
        struct device *dev = &cfg->dev->dev;
        u64 reg;
        u64 reg_unmasked;
 
-       reg = readq_be(&afu->host_map->intr_status);
+       reg = readq_be(&hwq->host_map->intr_status);
        reg_unmasked = (reg & SISL_ISTATUS_UNMASK);
 
        if (reg_unmasked == 0UL) {
@@ -1151,7 +1314,7 @@ static irqreturn_t cxlflash_sync_err_irq(int irq, void *data)
        dev_err(dev, "%s: unexpected interrupt, intr_status=%016llx\n",
                __func__, reg);
 
-       writeq_be(reg_unmasked, &afu->host_map->intr_clear);
+       writeq_be(reg_unmasked, &hwq->host_map->intr_clear);
 
 cxlflash_sync_err_irq_exit:
        return IRQ_HANDLED;
@@ -1167,17 +1330,18 @@ cxlflash_sync_err_irq_exit:
  *
  * Return: The number of entries processed.
  */
-static int process_hrrq(struct afu *afu, struct list_head *doneq, int budget)
+static int process_hrrq(struct hwq *hwq, struct list_head *doneq, int budget)
 {
+       struct afu *afu = hwq->afu;
        struct afu_cmd *cmd;
        struct sisl_ioasa *ioasa;
        struct sisl_ioarcb *ioarcb;
-       bool toggle = afu->toggle;
+       bool toggle = hwq->toggle;
        int num_hrrq = 0;
        u64 entry,
-           *hrrq_start = afu->hrrq_start,
-           *hrrq_end = afu->hrrq_end,
-           *hrrq_curr = afu->hrrq_curr;
+           *hrrq_start = hwq->hrrq_start,
+           *hrrq_end = hwq->hrrq_end,
+           *hrrq_curr = hwq->hrrq_curr;
 
        /* Process ready RRQ entries up to the specified budget (if any) */
        while (true) {
@@ -1206,15 +1370,15 @@ static int process_hrrq(struct afu *afu, struct list_head *doneq, int budget)
                        toggle ^= SISL_RESP_HANDLE_T_BIT;
                }
 
-               atomic_inc(&afu->hsq_credits);
+               atomic_inc(&hwq->hsq_credits);
                num_hrrq++;
 
                if (budget > 0 && num_hrrq >= budget)
                        break;
        }
 
-       afu->hrrq_curr = hrrq_curr;
-       afu->toggle = toggle;
+       hwq->hrrq_curr = hrrq_curr;
+       hwq->toggle = toggle;
 
        return num_hrrq;
 }
@@ -1244,18 +1408,18 @@ static void process_cmd_doneq(struct list_head *doneq)
  */
 static int cxlflash_irqpoll(struct irq_poll *irqpoll, int budget)
 {
-       struct afu *afu = container_of(irqpoll, struct afu, irqpoll);
+       struct hwq *hwq = container_of(irqpoll, struct hwq, irqpoll);
        unsigned long hrrq_flags;
        LIST_HEAD(doneq);
        int num_entries = 0;
 
-       spin_lock_irqsave(&afu->hrrq_slock, hrrq_flags);
+       spin_lock_irqsave(&hwq->hrrq_slock, hrrq_flags);
 
-       num_entries = process_hrrq(afu, &doneq, budget);
+       num_entries = process_hrrq(hwq, &doneq, budget);
        if (num_entries < budget)
                irq_poll_complete(irqpoll);
 
-       spin_unlock_irqrestore(&afu->hrrq_slock, hrrq_flags);
+       spin_unlock_irqrestore(&hwq->hrrq_slock, hrrq_flags);
 
        process_cmd_doneq(&doneq);
        return num_entries;
@@ -1270,21 +1434,22 @@ static int cxlflash_irqpoll(struct irq_poll *irqpoll, int budget)
  */
 static irqreturn_t cxlflash_rrq_irq(int irq, void *data)
 {
-       struct afu *afu = (struct afu *)data;
+       struct hwq *hwq = (struct hwq *)data;
+       struct afu *afu = hwq->afu;
        unsigned long hrrq_flags;
        LIST_HEAD(doneq);
        int num_entries = 0;
 
-       spin_lock_irqsave(&afu->hrrq_slock, hrrq_flags);
+       spin_lock_irqsave(&hwq->hrrq_slock, hrrq_flags);
 
        if (afu_is_irqpoll_enabled(afu)) {
-               irq_poll_sched(&afu->irqpoll);
-               spin_unlock_irqrestore(&afu->hrrq_slock, hrrq_flags);
+               irq_poll_sched(&hwq->irqpoll);
+               spin_unlock_irqrestore(&hwq->hrrq_slock, hrrq_flags);
                return IRQ_HANDLED;
        }
 
-       num_entries = process_hrrq(afu, &doneq, -1);
-       spin_unlock_irqrestore(&afu->hrrq_slock, hrrq_flags);
+       num_entries = process_hrrq(hwq, &doneq, -1);
+       spin_unlock_irqrestore(&hwq->hrrq_slock, hrrq_flags);
 
        if (num_entries == 0)
                return IRQ_NONE;
@@ -1293,6 +1458,35 @@ static irqreturn_t cxlflash_rrq_irq(int irq, void *data)
        return IRQ_HANDLED;
 }
 
+/*
+ * Asynchronous interrupt information table
+ *
+ * NOTE:
+ *     - Order matters here as this array is indexed by bit position.
+ *
+ *     - The checkpatch script considers the BUILD_SISL_ASTATUS_FC_PORT macro
+ *       as complex and complains due to a lack of parentheses/braces.
+ */
+#define ASTATUS_FC(_a, _b, _c, _d)                                      \
+       { SISL_ASTATUS_FC##_a##_##_b, _c, _a, (_d) }
+
+#define BUILD_SISL_ASTATUS_FC_PORT(_a)                                  \
+       ASTATUS_FC(_a, LINK_UP, "link up", 0),                           \
+       ASTATUS_FC(_a, LINK_DN, "link down", 0),                         \
+       ASTATUS_FC(_a, LOGI_S, "login succeeded", SCAN_HOST),            \
+       ASTATUS_FC(_a, LOGI_F, "login failed", CLR_FC_ERROR),            \
+       ASTATUS_FC(_a, LOGI_R, "login timed out, retrying", LINK_RESET), \
+       ASTATUS_FC(_a, CRC_T, "CRC threshold exceeded", LINK_RESET),     \
+       ASTATUS_FC(_a, LOGO, "target initiated LOGO", 0),                \
+       ASTATUS_FC(_a, OTHER, "other error", CLR_FC_ERROR | LINK_RESET)
+
+static const struct asyc_intr_info ainfo[] = {
+       BUILD_SISL_ASTATUS_FC_PORT(1),
+       BUILD_SISL_ASTATUS_FC_PORT(0),
+       BUILD_SISL_ASTATUS_FC_PORT(3),
+       BUILD_SISL_ASTATUS_FC_PORT(2)
+};
+
 /**
  * cxlflash_async_err_irq() - interrupt handler for asynchronous errors
  * @irq:       Interrupt number.
@@ -1302,21 +1496,22 @@ static irqreturn_t cxlflash_rrq_irq(int irq, void *data)
  */
 static irqreturn_t cxlflash_async_err_irq(int irq, void *data)
 {
-       struct afu *afu = (struct afu *)data;
+       struct hwq *hwq = (struct hwq *)data;
+       struct afu *afu = hwq->afu;
        struct cxlflash_cfg *cfg = afu->parent;
        struct device *dev = &cfg->dev->dev;
-       u64 reg_unmasked;
        const struct asyc_intr_info *info;
        struct sisl_global_map __iomem *global = &afu->afu_map->global;
        __be64 __iomem *fc_port_regs;
+       u64 reg_unmasked;
        u64 reg;
+       u64 bit;
        u8 port;
-       int i;
 
        reg = readq_be(&global->regs.aintr_status);
        reg_unmasked = (reg & SISL_ASTATUS_UNMASK);
 
-       if (reg_unmasked == 0) {
+       if (unlikely(reg_unmasked == 0)) {
                dev_err(dev, "%s: spurious interrupt, aintr_status=%016llx\n",
                        __func__, reg);
                goto out;
@@ -1326,10 +1521,17 @@ static irqreturn_t cxlflash_async_err_irq(int irq, void *data)
        writeq_be(reg_unmasked, &global->regs.aintr_clear);
 
        /* Check each bit that is on */
-       for (i = 0; reg_unmasked; i++, reg_unmasked = (reg_unmasked >> 1)) {
-               info = find_ainfo(1ULL << i);
-               if (((reg_unmasked & 0x1) == 0) || !info)
+       for_each_set_bit(bit, (ulong *)&reg_unmasked, BITS_PER_LONG) {
+               if (unlikely(bit >= ARRAY_SIZE(ainfo))) {
+                       WARN_ON_ONCE(1);
+                       continue;
+               }
+
+               info = &ainfo[bit];
+               if (unlikely(info->status != 1ULL << bit)) {
+                       WARN_ON_ONCE(1);
                        continue;
+               }
 
                port = info->port;
                fc_port_regs = get_fc_port_regs(cfg, port);
@@ -1378,16 +1580,18 @@ out:
 /**
  * start_context() - starts the master context
  * @cfg:       Internal structure associated with the host.
+ * @index:     Index of the hardware queue.
  *
  * Return: A success or failure value from CXL services.
  */
-static int start_context(struct cxlflash_cfg *cfg)
+static int start_context(struct cxlflash_cfg *cfg, u32 index)
 {
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq = get_hwq(cfg->afu, index);
        int rc = 0;
 
-       rc = cxl_start_context(cfg->mcctx,
-                              cfg->afu->work.work_element_descriptor,
+       rc = cxl_start_context(hwq->ctx,
+                              hwq->work.work_element_descriptor,
                               NULL);
 
        dev_dbg(dev, "%s: returning rc=%d\n", __func__, rc);
@@ -1497,6 +1701,7 @@ static void init_pcr(struct cxlflash_cfg *cfg)
 {
        struct afu *afu = cfg->afu;
        struct sisl_ctrl_map __iomem *ctrl_map;
+       struct hwq *hwq;
        int i;
 
        for (i = 0; i < MAX_CONTEXT; i++) {
@@ -1508,13 +1713,17 @@ static void init_pcr(struct cxlflash_cfg *cfg)
                writeq_be(0, &ctrl_map->ctx_cap);
        }
 
-       /* Copy frequently used fields into afu */
-       afu->ctx_hndl = (u16) cxl_process_element(cfg->mcctx);
-       afu->host_map = &afu->afu_map->hosts[afu->ctx_hndl].host;
-       afu->ctrl_map = &afu->afu_map->ctrls[afu->ctx_hndl].ctrl;
+       /* Copy frequently used fields into hwq */
+       for (i = 0; i < afu->num_hwqs; i++) {
+               hwq = get_hwq(afu, i);
+
+               hwq->ctx_hndl = (u16) cxl_process_element(hwq->ctx);
+               hwq->host_map = &afu->afu_map->hosts[hwq->ctx_hndl].host;
+               hwq->ctrl_map = &afu->afu_map->ctrls[hwq->ctx_hndl].ctrl;
 
-       /* Program the Endian Control for the master context */
-       writeq_be(SISL_ENDIAN_CTRL, &afu->host_map->endian_ctrl);
+               /* Program the Endian Control for the master context */
+               writeq_be(SISL_ENDIAN_CTRL, &hwq->host_map->endian_ctrl);
+       }
 }
 
 /**
@@ -1525,6 +1734,8 @@ static int init_global(struct cxlflash_cfg *cfg)
 {
        struct afu *afu = cfg->afu;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq;
+       struct sisl_host_map __iomem *hmap;
        __be64 __iomem *fc_port_regs;
        u64 wwpn[MAX_FC_PORTS]; /* wwpn of AFU ports */
        int i = 0, num_ports = 0;
@@ -1537,13 +1748,18 @@ static int init_global(struct cxlflash_cfg *cfg)
                goto out;
        }
 
-       /* Set up RRQ and SQ in AFU for master issued cmds */
-       writeq_be((u64) afu->hrrq_start, &afu->host_map->rrq_start);
-       writeq_be((u64) afu->hrrq_end, &afu->host_map->rrq_end);
+       /* Set up RRQ and SQ in HWQ for master issued cmds */
+       for (i = 0; i < afu->num_hwqs; i++) {
+               hwq = get_hwq(afu, i);
+               hmap = hwq->host_map;
 
-       if (afu_is_sq_cmd_mode(afu)) {
-               writeq_be((u64)afu->hsq_start, &afu->host_map->sq_start);
-               writeq_be((u64)afu->hsq_end, &afu->host_map->sq_end);
+               writeq_be((u64) hwq->hrrq_start, &hmap->rrq_start);
+               writeq_be((u64) hwq->hrrq_end, &hmap->rrq_end);
+
+               if (afu_is_sq_cmd_mode(afu)) {
+                       writeq_be((u64)hwq->hsq_start, &hmap->sq_start);
+                       writeq_be((u64)hwq->hsq_end, &hmap->sq_end);
+               }
        }
 
        /* AFU configuration */
@@ -1587,11 +1803,15 @@ static int init_global(struct cxlflash_cfg *cfg)
        /* Set up master's own CTX_CAP to allow real mode, host translation */
        /* tables, afu cmds and read/write GSCSI cmds. */
        /* First, unlock ctx_cap write by reading mbox */
-       (void)readq_be(&afu->ctrl_map->mbox_r); /* unlock ctx_cap */
-       writeq_be((SISL_CTX_CAP_REAL_MODE | SISL_CTX_CAP_HOST_XLATE |
-                  SISL_CTX_CAP_READ_CMD | SISL_CTX_CAP_WRITE_CMD |
-                  SISL_CTX_CAP_AFU_CMD | SISL_CTX_CAP_GSCSI_CMD),
-                 &afu->ctrl_map->ctx_cap);
+       for (i = 0; i < afu->num_hwqs; i++) {
+               hwq = get_hwq(afu, i);
+
+               (void)readq_be(&hwq->ctrl_map->mbox_r); /* unlock ctx_cap */
+               writeq_be((SISL_CTX_CAP_REAL_MODE | SISL_CTX_CAP_HOST_XLATE |
+                       SISL_CTX_CAP_READ_CMD | SISL_CTX_CAP_WRITE_CMD |
+                       SISL_CTX_CAP_AFU_CMD | SISL_CTX_CAP_GSCSI_CMD),
+                       &hwq->ctrl_map->ctx_cap);
+       }
        /* Initialize heartbeat */
        afu->hb = readq_be(&afu->afu_map->global.regs.afu_hb);
 out:
@@ -1606,33 +1826,45 @@ static int start_afu(struct cxlflash_cfg *cfg)
 {
        struct afu *afu = cfg->afu;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq;
        int rc = 0;
+       int i;
 
        init_pcr(cfg);
 
-       /* Initialize RRQ */
-       memset(&afu->rrq_entry, 0, sizeof(afu->rrq_entry));
-       afu->hrrq_start = &afu->rrq_entry[0];
-       afu->hrrq_end = &afu->rrq_entry[NUM_RRQ_ENTRY - 1];
-       afu->hrrq_curr = afu->hrrq_start;
-       afu->toggle = 1;
-       spin_lock_init(&afu->hrrq_slock);
+       /* Initialize each HWQ */
+       for (i = 0; i < afu->num_hwqs; i++) {
+               hwq = get_hwq(afu, i);
 
-       /* Initialize SQ */
-       if (afu_is_sq_cmd_mode(afu)) {
-               memset(&afu->sq, 0, sizeof(afu->sq));
-               afu->hsq_start = &afu->sq[0];
-               afu->hsq_end = &afu->sq[NUM_SQ_ENTRY - 1];
-               afu->hsq_curr = afu->hsq_start;
+               /* After an AFU reset, RRQ entries are stale, clear them */
+               memset(&hwq->rrq_entry, 0, sizeof(hwq->rrq_entry));
 
-               spin_lock_init(&afu->hsq_slock);
-               atomic_set(&afu->hsq_credits, NUM_SQ_ENTRY - 1);
-       }
+               /* Initialize RRQ pointers */
+               hwq->hrrq_start = &hwq->rrq_entry[0];
+               hwq->hrrq_end = &hwq->rrq_entry[NUM_RRQ_ENTRY - 1];
+               hwq->hrrq_curr = hwq->hrrq_start;
+               hwq->toggle = 1;
+
+               /* Initialize spin locks */
+               spin_lock_init(&hwq->hrrq_slock);
+               spin_lock_init(&hwq->hsq_slock);
+
+               /* Initialize SQ */
+               if (afu_is_sq_cmd_mode(afu)) {
+                       memset(&hwq->sq, 0, sizeof(hwq->sq));
+                       hwq->hsq_start = &hwq->sq[0];
+                       hwq->hsq_end = &hwq->sq[NUM_SQ_ENTRY - 1];
+                       hwq->hsq_curr = hwq->hsq_start;
+
+                       atomic_set(&hwq->hsq_credits, NUM_SQ_ENTRY - 1);
+               }
+
+               /* Initialize IRQ poll */
+               if (afu_is_irqpoll_enabled(afu))
+                       irq_poll_init(&hwq->irqpoll, afu->irqpoll_weight,
+                                     cxlflash_irqpoll);
 
-       /* Initialize IRQ poll */
-       if (afu_is_irqpoll_enabled(afu))
-               irq_poll_init(&afu->irqpoll, afu->irqpoll_weight,
-                             cxlflash_irqpoll);
+       }
 
        rc = init_global(cfg);
 
@@ -1643,18 +1875,21 @@ static int start_afu(struct cxlflash_cfg *cfg)
 /**
  * init_intr() - setup interrupt handlers for the master context
  * @cfg:       Internal structure associated with the host.
+ * @hwq:       Hardware queue to initialize.
  *
  * Return: 0 on success, -errno on failure
  */
 static enum undo_level init_intr(struct cxlflash_cfg *cfg,
-                                struct cxl_context *ctx)
+                                struct hwq *hwq)
 {
-       struct afu *afu = cfg->afu;
        struct device *dev = &cfg->dev->dev;
+       struct cxl_context *ctx = hwq->ctx;
        int rc = 0;
        enum undo_level level = UNDO_NOOP;
+       bool is_primary_hwq = (hwq->index == PRIMARY_HWQ);
+       int num_irqs = is_primary_hwq ? 3 : 2;
 
-       rc = cxl_allocate_afu_irqs(ctx, 3);
+       rc = cxl_allocate_afu_irqs(ctx, num_irqs);
        if (unlikely(rc)) {
                dev_err(dev, "%s: allocate_afu_irqs failed rc=%d\n",
                        __func__, rc);
@@ -1662,7 +1897,7 @@ static enum undo_level init_intr(struct cxlflash_cfg *cfg,
                goto out;
        }
 
-       rc = cxl_map_afu_irq(ctx, 1, cxlflash_sync_err_irq, afu,
+       rc = cxl_map_afu_irq(ctx, 1, cxlflash_sync_err_irq, hwq,
                             "SISL_MSI_SYNC_ERROR");
        if (unlikely(rc <= 0)) {
                dev_err(dev, "%s: SISL_MSI_SYNC_ERROR map failed\n", __func__);
@@ -1670,7 +1905,7 @@ static enum undo_level init_intr(struct cxlflash_cfg *cfg,
                goto out;
        }
 
-       rc = cxl_map_afu_irq(ctx, 2, cxlflash_rrq_irq, afu,
+       rc = cxl_map_afu_irq(ctx, 2, cxlflash_rrq_irq, hwq,
                             "SISL_MSI_RRQ_UPDATED");
        if (unlikely(rc <= 0)) {
                dev_err(dev, "%s: SISL_MSI_RRQ_UPDATED map failed\n", __func__);
@@ -1678,7 +1913,11 @@ static enum undo_level init_intr(struct cxlflash_cfg *cfg,
                goto out;
        }
 
-       rc = cxl_map_afu_irq(ctx, 3, cxlflash_async_err_irq, afu,
+       /* SISL_MSI_ASYNC_ERROR is setup only for the primary HWQ */
+       if (!is_primary_hwq)
+               goto out;
+
+       rc = cxl_map_afu_irq(ctx, 3, cxlflash_async_err_irq, hwq,
                             "SISL_MSI_ASYNC_ERROR");
        if (unlikely(rc <= 0)) {
                dev_err(dev, "%s: SISL_MSI_ASYNC_ERROR map failed\n", __func__);
@@ -1692,55 +1931,74 @@ out:
 /**
  * init_mc() - create and register as the master context
  * @cfg:       Internal structure associated with the host.
+ * index:      HWQ Index of the master context.
  *
  * Return: 0 on success, -errno on failure
  */
-static int init_mc(struct cxlflash_cfg *cfg)
+static int init_mc(struct cxlflash_cfg *cfg, u32 index)
 {
        struct cxl_context *ctx;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq = get_hwq(cfg->afu, index);
        int rc = 0;
        enum undo_level level;
 
-       ctx = cxl_get_context(cfg->dev);
+       hwq->afu = cfg->afu;
+       hwq->index = index;
+       INIT_LIST_HEAD(&hwq->pending_cmds);
+
+       if (index == PRIMARY_HWQ)
+               ctx = cxl_get_context(cfg->dev);
+       else
+               ctx = cxl_dev_context_init(cfg->dev);
        if (unlikely(!ctx)) {
                rc = -ENOMEM;
-               goto ret;
+               goto err1;
        }
-       cfg->mcctx = ctx;
+
+       WARN_ON(hwq->ctx);
+       hwq->ctx = ctx;
 
        /* Set it up as a master with the CXL */
        cxl_set_master(ctx);
 
-       /* During initialization reset the AFU to start from a clean slate */
-       rc = cxl_afu_reset(cfg->mcctx);
-       if (unlikely(rc)) {
-               dev_err(dev, "%s: AFU reset failed rc=%d\n", __func__, rc);
-               goto ret;
+       /* Reset AFU when initializing primary context */
+       if (index == PRIMARY_HWQ) {
+               rc = cxl_afu_reset(ctx);
+               if (unlikely(rc)) {
+                       dev_err(dev, "%s: AFU reset failed rc=%d\n",
+                                     __func__, rc);
+                       goto err1;
+               }
        }
 
-       level = init_intr(cfg, ctx);
+       level = init_intr(cfg, hwq);
        if (unlikely(level)) {
                dev_err(dev, "%s: interrupt init failed rc=%d\n", __func__, rc);
-               goto out;
+               goto err2;
        }
 
        /* This performs the equivalent of the CXL_IOCTL_START_WORK.
         * The CXL_IOCTL_GET_PROCESS_ELEMENT is implicit in the process
         * element (pe) that is embedded in the context (ctx)
         */
-       rc = start_context(cfg);
+       rc = start_context(cfg, index);
        if (unlikely(rc)) {
                dev_err(dev, "%s: start context failed rc=%d\n", __func__, rc);
                level = UNMAP_THREE;
-               goto out;
+               goto err2;
        }
-ret:
+
+out:
        dev_dbg(dev, "%s: returning rc=%d\n", __func__, rc);
        return rc;
-out:
-       term_intr(cfg, level);
-       goto ret;
+err2:
+       term_intr(cfg, level, index);
+       if (index != PRIMARY_HWQ)
+               cxl_release_context(ctx);
+err1:
+       hwq->ctx = NULL;
+       goto out;
 }
 
 /**
@@ -1791,18 +2049,24 @@ static int init_afu(struct cxlflash_cfg *cfg)
        int rc = 0;
        struct afu *afu = cfg->afu;
        struct device *dev = &cfg->dev->dev;
+       struct hwq *hwq;
+       int i;
 
        cxl_perst_reloads_same_image(cfg->cxl_afu, true);
 
-       rc = init_mc(cfg);
-       if (rc) {
-               dev_err(dev, "%s: init_mc failed rc=%d\n",
-                       __func__, rc);
-               goto out;
+       afu->num_hwqs = afu->desired_hwqs;
+       for (i = 0; i < afu->num_hwqs; i++) {
+               rc = init_mc(cfg, i);
+               if (rc) {
+                       dev_err(dev, "%s: init_mc failed rc=%d index=%d\n",
+                               __func__, rc, i);
+                       goto err1;
+               }
        }
 
-       /* Map the entire MMIO space of the AFU */
-       afu->afu_map = cxl_psa_map(cfg->mcctx);
+       /* Map the entire MMIO space of the AFU using the first context */
+       hwq = get_hwq(afu, PRIMARY_HWQ);
+       afu->afu_map = cxl_psa_map(hwq->ctx);
        if (!afu->afu_map) {
                dev_err(dev, "%s: cxl_psa_map failed\n", __func__);
                rc = -ENOMEM;
@@ -1842,8 +2106,11 @@ static int init_afu(struct cxlflash_cfg *cfg)
        }
 
        afu_err_intr_init(cfg->afu);
-       spin_lock_init(&afu->rrin_slock);
-       afu->room = readq_be(&afu->host_map->cmd_room);
+       for (i = 0; i < afu->num_hwqs; i++) {
+               hwq = get_hwq(afu, i);
+
+               hwq->room = readq_be(&hwq->host_map->cmd_room);
+       }
 
        /* Restore the LUN mappings */
        cxlflash_restore_luntable(cfg);
@@ -1852,83 +2119,179 @@ out:
        return rc;
 
 err1:
-       term_intr(cfg, UNMAP_THREE);
-       term_mc(cfg);
+       for (i = afu->num_hwqs - 1; i >= 0; i--) {
+               term_intr(cfg, UNMAP_THREE, i);
+               term_mc(cfg, i);
+       }
        goto out;
 }
 
 /**
- * cxlflash_afu_sync() - builds and sends an AFU sync command
- * @afu:       AFU associated with the host.
- * @ctx_hndl_u:        Identifies context requesting sync.
- * @res_hndl_u:        Identifies resource requesting sync.
- * @mode:      Type of sync to issue (lightweight, heavyweight, global).
- *
- * The AFU can only take 1 sync command at a time. This routine enforces this
- * limitation by using a mutex to provide exclusive access to the AFU during
- * the sync. This design point requires calling threads to not be on interrupt
- * context due to the possibility of sleeping during concurrent sync operations.
- *
- * AFU sync operations are only necessary and allowed when the device is
- * operating normally. When not operating normally, sync requests can occur as
- * part of cleaning up resources associated with an adapter prior to removal.
- * In this scenario, these requests are simply ignored (safe due to the AFU
- * going away).
+ * afu_reset() - resets the AFU
+ * @cfg:       Internal structure associated with the host.
  *
- * Return:
- *     0 on success
- *     -1 on failure
+ * Return: 0 on success, -errno on failure
  */
-int cxlflash_afu_sync(struct afu *afu, ctx_hndl_t ctx_hndl_u,
-                     res_hndl_t res_hndl_u, u8 mode)
+static int afu_reset(struct cxlflash_cfg *cfg)
 {
-       struct cxlflash_cfg *cfg = afu->parent;
        struct device *dev = &cfg->dev->dev;
-       struct afu_cmd *cmd = NULL;
-       char *buf = NULL;
        int rc = 0;
-       static DEFINE_MUTEX(sync_active);
 
-       if (cfg->state != STATE_NORMAL) {
-               dev_dbg(dev, "%s: Sync not required state=%u\n",
-                       __func__, cfg->state);
-               return 0;
-       }
+       /* Stop the context before the reset. Since the context is
+        * no longer available restart it after the reset is complete
+        */
+       term_afu(cfg);
 
-       mutex_lock(&sync_active);
-       atomic_inc(&afu->cmds_active);
-       buf = kzalloc(sizeof(*cmd) + __alignof__(*cmd) - 1, GFP_KERNEL);
-       if (unlikely(!buf)) {
-               dev_err(dev, "%s: no memory for command\n", __func__);
-               rc = -1;
+       rc = init_afu(cfg);
+
+       dev_dbg(dev, "%s: returning rc=%d\n", __func__, rc);
+       return rc;
+}
+
+/**
+ * drain_ioctls() - wait until all currently executing ioctls have completed
+ * @cfg:       Internal structure associated with the host.
+ *
+ * Obtain write access to read/write semaphore that wraps ioctl
+ * handling to 'drain' ioctls currently executing.
+ */
+static void drain_ioctls(struct cxlflash_cfg *cfg)
+{
+       down_write(&cfg->ioctl_rwsem);
+       up_write(&cfg->ioctl_rwsem);
+}
+
+/**
+ * cxlflash_async_reset_host() - asynchronous host reset handler
+ * @data:      Private data provided while scheduling reset.
+ * @cookie:    Cookie that can be used for checkpointing.
+ */
+static void cxlflash_async_reset_host(void *data, async_cookie_t cookie)
+{
+       struct cxlflash_cfg *cfg = data;
+       struct device *dev = &cfg->dev->dev;
+       int rc = 0;
+
+       if (cfg->state != STATE_RESET) {
+               dev_dbg(dev, "%s: Not performing a reset, state=%d\n",
+                       __func__, cfg->state);
                goto out;
        }
 
-       cmd = (struct afu_cmd *)PTR_ALIGN(buf, __alignof__(*cmd));
-       init_completion(&cmd->cevent);
-       cmd->parent = afu;
+       drain_ioctls(cfg);
+       cxlflash_mark_contexts_error(cfg);
+       rc = afu_reset(cfg);
+       if (rc)
+               cfg->state = STATE_FAILTERM;
+       else
+               cfg->state = STATE_NORMAL;
+       wake_up_all(&cfg->reset_waitq);
 
-       dev_dbg(dev, "%s: afu=%p cmd=%p %d\n", __func__, afu, cmd, ctx_hndl_u);
+out:
+       scsi_unblock_requests(cfg->host);
+}
 
-       cmd->rcb.req_flags = SISL_REQ_FLAGS_AFU_CMD;
-       cmd->rcb.ctx_id = afu->ctx_hndl;
-       cmd->rcb.msi = SISL_MSI_RRQ_UPDATED;
-       cmd->rcb.timeout = MC_AFU_SYNC_TIMEOUT;
+/**
+ * cxlflash_schedule_async_reset() - schedule an asynchronous host reset
+ * @cfg:       Internal structure associated with the host.
+ */
+static void cxlflash_schedule_async_reset(struct cxlflash_cfg *cfg)
+{
+       struct device *dev = &cfg->dev->dev;
 
-       cmd->rcb.cdb[0] = 0xC0; /* AFU Sync */
-       cmd->rcb.cdb[1] = mode;
+       if (cfg->state != STATE_NORMAL) {
+               dev_dbg(dev, "%s: Not performing reset state=%d\n",
+                       __func__, cfg->state);
+               return;
+       }
 
-       /* The cdb is aligned, no unaligned accessors required */
-       *((__be16 *)&cmd->rcb.cdb[2]) = cpu_to_be16(ctx_hndl_u);
-       *((__be32 *)&cmd->rcb.cdb[4]) = cpu_to_be32(res_hndl_u);
+       cfg->state = STATE_RESET;
+       scsi_block_requests(cfg->host);
+       cfg->async_reset_cookie = async_schedule(cxlflash_async_reset_host,
+                                                cfg);
+}
+
+/**
+ * send_afu_cmd() - builds and sends an internal AFU command
+ * @afu:       AFU associated with the host.
+ * @rcb:       Pre-populated IOARCB describing command to send.
+ *
+ * The AFU can only take one internal AFU command at a time. This limitation is
+ * enforced by using a mutex to provide exclusive access to the AFU during the
+ * operation. This design point requires calling threads to not be on interrupt
+ * context due to the possibility of sleeping during concurrent AFU operations.
+ *
+ * The command status is optionally passed back to the caller when the caller
+ * populates the IOASA field of the IOARCB with a pointer to an IOASA structure.
+ *
+ * Return:
+ *     0 on success, -errno on failure
+ */
+static int send_afu_cmd(struct afu *afu, struct sisl_ioarcb *rcb)
+{
+       struct cxlflash_cfg *cfg = afu->parent;
+       struct device *dev = &cfg->dev->dev;
+       struct afu_cmd *cmd = NULL;
+       struct hwq *hwq = get_hwq(afu, PRIMARY_HWQ);
+       char *buf = NULL;
+       int rc = 0;
+       int nretry = 0;
+       static DEFINE_MUTEX(sync_active);
+
+       if (cfg->state != STATE_NORMAL) {
+               dev_dbg(dev, "%s: Sync not required state=%u\n",
+                       __func__, cfg->state);
+               return 0;
+       }
+
+       mutex_lock(&sync_active);
+       atomic_inc(&afu->cmds_active);
+       buf = kmalloc(sizeof(*cmd) + __alignof__(*cmd) - 1, GFP_KERNEL);
+       if (unlikely(!buf)) {
+               dev_err(dev, "%s: no memory for command\n", __func__);
+               rc = -ENOMEM;
+               goto out;
+       }
+
+       cmd = (struct afu_cmd *)PTR_ALIGN(buf, __alignof__(*cmd));
+
+retry:
+       memset(cmd, 0, sizeof(*cmd));
+       memcpy(&cmd->rcb, rcb, sizeof(*rcb));
+       INIT_LIST_HEAD(&cmd->queue);
+       init_completion(&cmd->cevent);
+       cmd->parent = afu;
+       cmd->hwq_index = hwq->index;
+       cmd->rcb.ctx_id = hwq->ctx_hndl;
+
+       dev_dbg(dev, "%s: afu=%p cmd=%p type=%02x nretry=%d\n",
+               __func__, afu, cmd, cmd->rcb.cdb[0], nretry);
 
        rc = afu->send_cmd(afu, cmd);
-       if (unlikely(rc))
+       if (unlikely(rc)) {
+               rc = -ENOBUFS;
                goto out;
+       }
 
        rc = wait_resp(afu, cmd);
-       if (unlikely(rc))
-               rc = -1;
+       switch (rc) {
+       case -ETIMEDOUT:
+               rc = afu->context_reset(hwq);
+               if (rc) {
+                       cxlflash_schedule_async_reset(cfg);
+                       break;
+               }
+               /* fall through to retry */
+       case -EAGAIN:
+               if (++nretry < 2)
+                       goto retry;
+               /* fall through to exit */
+       default:
+               break;
+       }
+
+       if (rcb->ioasa)
+               *rcb->ioasa = cmd->sa;
 out:
        atomic_dec(&afu->cmds_active);
        mutex_unlock(&sync_active);
@@ -1938,38 +2301,88 @@ out:
 }
 
 /**
- * afu_reset() - resets the AFU
- * @cfg:       Internal structure associated with the host.
+ * cxlflash_afu_sync() - builds and sends an AFU sync command
+ * @afu:       AFU associated with the host.
+ * @ctx:       Identifies context requesting sync.
+ * @res:       Identifies resource requesting sync.
+ * @mode:      Type of sync to issue (lightweight, heavyweight, global).
  *
- * Return: 0 on success, -errno on failure
+ * AFU sync operations are only necessary and allowed when the device is
+ * operating normally. When not operating normally, sync requests can occur as
+ * part of cleaning up resources associated with an adapter prior to removal.
+ * In this scenario, these requests are simply ignored (safe due to the AFU
+ * going away).
+ *
+ * Return:
+ *     0 on success, -errno on failure
  */
-static int afu_reset(struct cxlflash_cfg *cfg)
+int cxlflash_afu_sync(struct afu *afu, ctx_hndl_t ctx, res_hndl_t res, u8 mode)
 {
+       struct cxlflash_cfg *cfg = afu->parent;
        struct device *dev = &cfg->dev->dev;
-       int rc = 0;
+       struct sisl_ioarcb rcb = { 0 };
 
-       /* Stop the context before the reset. Since the context is
-        * no longer available restart it after the reset is complete
-        */
-       term_afu(cfg);
+       dev_dbg(dev, "%s: afu=%p ctx=%u res=%u mode=%u\n",
+               __func__, afu, ctx, res, mode);
 
-       rc = init_afu(cfg);
+       rcb.req_flags = SISL_REQ_FLAGS_AFU_CMD;
+       rcb.msi = SISL_MSI_RRQ_UPDATED;
+       rcb.timeout = MC_AFU_SYNC_TIMEOUT;
 
-       dev_dbg(dev, "%s: returning rc=%d\n", __func__, rc);
-       return rc;
+       rcb.cdb[0] = SISL_AFU_CMD_SYNC;
+       rcb.cdb[1] = mode;
+       put_unaligned_be16(ctx, &rcb.cdb[2]);
+       put_unaligned_be32(res, &rcb.cdb[4]);
+
+       return send_afu_cmd(afu, &rcb);
 }
 
 /**
- * drain_ioctls() - wait until all currently executing ioctls have completed
- * @cfg:       Internal structure associated with the host.
+ * cxlflash_eh_abort_handler() - abort a SCSI command
+ * @scp:       SCSI command to abort.
  *
- * Obtain write access to read/write semaphore that wraps ioctl
- * handling to 'drain' ioctls currently executing.
+ * CXL Flash devices do not support a single command abort. Reset the context
+ * as per SISLite specification. Flush any pending commands in the hardware
+ * queue before the reset.
+ *
+ * Return: SUCCESS/FAILED as defined in scsi/scsi.h
  */
-static void drain_ioctls(struct cxlflash_cfg *cfg)
+static int cxlflash_eh_abort_handler(struct scsi_cmnd *scp)
 {
-       down_write(&cfg->ioctl_rwsem);
-       up_write(&cfg->ioctl_rwsem);
+       int rc = FAILED;
+       struct Scsi_Host *host = scp->device->host;
+       struct cxlflash_cfg *cfg = shost_priv(host);
+       struct afu_cmd *cmd = sc_to_afuc(scp);
+       struct device *dev = &cfg->dev->dev;
+       struct afu *afu = cfg->afu;
+       struct hwq *hwq = get_hwq(afu, cmd->hwq_index);
+
+       dev_dbg(dev, "%s: (scp=%p) %d/%d/%d/%llu "
+               "cdb=(%08x-%08x-%08x-%08x)\n", __func__, scp, host->host_no,
+               scp->device->channel, scp->device->id, scp->device->lun,
+               get_unaligned_be32(&((u32 *)scp->cmnd)[0]),
+               get_unaligned_be32(&((u32 *)scp->cmnd)[1]),
+               get_unaligned_be32(&((u32 *)scp->cmnd)[2]),
+               get_unaligned_be32(&((u32 *)scp->cmnd)[3]));
+
+       /* When the state is not normal, another reset/reload is in progress.
+        * Return failed and the mid-layer will invoke host reset handler.
+        */
+       if (cfg->state != STATE_NORMAL) {
+               dev_dbg(dev, "%s: Invalid state for abort, state=%d\n",
+                       __func__, cfg->state);
+               goto out;
+       }
+
+       rc = afu->context_reset(hwq);
+       if (unlikely(rc))
+               goto out;
+
+       rc = SUCCESS;
+
+out:
+       dev_dbg(dev, "%s: returning rc=%d\n", __func__, rc);
+       return rc;
 }
 
 /**
@@ -2424,8 +2837,9 @@ static ssize_t irqpoll_weight_store(struct device *dev,
        struct cxlflash_cfg *cfg = shost_priv(class_to_shost(dev));
        struct device *cfgdev = &cfg->dev->dev;
        struct afu *afu = cfg->afu;
+       struct hwq *hwq;
        u32 weight;
-       int rc;
+       int rc, i;
 
        rc = kstrtouint(buf, 10, &weight);
        if (rc)
@@ -2443,13 +2857,173 @@ static ssize_t irqpoll_weight_store(struct device *dev,
                return -EINVAL;
        }
 
-       if (afu_is_irqpoll_enabled(afu))
-               irq_poll_disable(&afu->irqpoll);
+       if (afu_is_irqpoll_enabled(afu)) {
+               for (i = 0; i < afu->num_hwqs; i++) {
+                       hwq = get_hwq(afu, i);
+
+                       irq_poll_disable(&hwq->irqpoll);
+               }
+       }
 
        afu->irqpoll_weight = weight;
 
-       if (weight > 0)
-               irq_poll_init(&afu->irqpoll, weight, cxlflash_irqpoll);
+       if (weight > 0) {
+               for (i = 0; i < afu->num_hwqs; i++) {
+                       hwq = get_hwq(afu, i);
+
+                       irq_poll_init(&hwq->irqpoll, weight, cxlflash_irqpoll);
+               }
+       }
+
+       return count;
+}
+
+/**
+ * num_hwqs_show() - presents the number of hardware queues for the host
+ * @dev:       Generic device associated with the host.
+ * @attr:      Device attribute representing the number of hardware queues.
+ * @buf:       Buffer of length PAGE_SIZE to report back the number of hardware
+ *             queues in ASCII.
+ *
+ * Return: The size of the ASCII string returned in @buf.
+ */
+static ssize_t num_hwqs_show(struct device *dev,
+                            struct device_attribute *attr, char *buf)
+{
+       struct cxlflash_cfg *cfg = shost_priv(class_to_shost(dev));
+       struct afu *afu = cfg->afu;
+
+       return scnprintf(buf, PAGE_SIZE, "%u\n", afu->num_hwqs);
+}
+
+/**
+ * num_hwqs_store() - sets the number of hardware queues for the host
+ * @dev:       Generic device associated with the host.
+ * @attr:      Device attribute representing the number of hardware queues.
+ * @buf:       Buffer of length PAGE_SIZE containing the number of hardware
+ *             queues in ASCII.
+ * @count:     Length of data resizing in @buf.
+ *
+ * n > 0: num_hwqs = n
+ * n = 0: num_hwqs = num_online_cpus()
+ * n < 0: num_online_cpus() / abs(n)
+ *
+ * Return: The size of the ASCII string returned in @buf.
+ */
+static ssize_t num_hwqs_store(struct device *dev,
+                             struct device_attribute *attr,
+                             const char *buf, size_t count)
+{
+       struct cxlflash_cfg *cfg = shost_priv(class_to_shost(dev));
+       struct afu *afu = cfg->afu;
+       int rc;
+       int nhwqs, num_hwqs;
+
+       rc = kstrtoint(buf, 10, &nhwqs);
+       if (rc)
+               return -EINVAL;
+
+       if (nhwqs >= 1)
+               num_hwqs = nhwqs;
+       else if (nhwqs == 0)
+               num_hwqs = num_online_cpus();
+       else
+               num_hwqs = num_online_cpus() / abs(nhwqs);
+
+       afu->desired_hwqs = min(num_hwqs, CXLFLASH_MAX_HWQS);
+       WARN_ON_ONCE(afu->desired_hwqs == 0);
+
+retry:
+       switch (cfg->state) {
+       case STATE_NORMAL:
+               cfg->state = STATE_RESET;
+               drain_ioctls(cfg);
+               cxlflash_mark_contexts_error(cfg);
+               rc = afu_reset(cfg);
+               if (rc)
+                       cfg->state = STATE_FAILTERM;
+               else
+                       cfg->state = STATE_NORMAL;
+               wake_up_all(&cfg->reset_waitq);
+               break;
+       case STATE_RESET:
+               wait_event(cfg->reset_waitq, cfg->state != STATE_RESET);
+               if (cfg->state == STATE_NORMAL)
+                       goto retry;
+       default:
+               /* Ideally should not happen */
+               dev_err(dev, "%s: Device is not ready, state=%d\n",
+                       __func__, cfg->state);
+               break;
+       }
+
+       return count;
+}
+
+static const char *hwq_mode_name[MAX_HWQ_MODE] = { "rr", "tag", "cpu" };
+
+/**
+ * hwq_mode_show() - presents the HWQ steering mode for the host
+ * @dev:       Generic device associated with the host.
+ * @attr:      Device attribute representing the HWQ steering mode.
+ * @buf:       Buffer of length PAGE_SIZE to report back the HWQ steering mode
+ *             as a character string.
+ *
+ * Return: The size of the ASCII string returned in @buf.
+ */
+static ssize_t hwq_mode_show(struct device *dev,
+                            struct device_attribute *attr, char *buf)
+{
+       struct cxlflash_cfg *cfg = shost_priv(class_to_shost(dev));
+       struct afu *afu = cfg->afu;
+
+       return scnprintf(buf, PAGE_SIZE, "%s\n", hwq_mode_name[afu->hwq_mode]);
+}
+
+/**
+ * hwq_mode_store() - sets the HWQ steering mode for the host
+ * @dev:       Generic device associated with the host.
+ * @attr:      Device attribute representing the HWQ steering mode.
+ * @buf:       Buffer of length PAGE_SIZE containing the HWQ steering mode
+ *             as a character string.
+ * @count:     Length of data resizing in @buf.
+ *
+ * rr = Round-Robin
+ * tag = Block MQ Tagging
+ * cpu = CPU Affinity
+ *
+ * Return: The size of the ASCII string returned in @buf.
+ */
+static ssize_t hwq_mode_store(struct device *dev,
+                             struct device_attribute *attr,
+                             const char *buf, size_t count)
+{
+       struct Scsi_Host *shost = class_to_shost(dev);
+       struct cxlflash_cfg *cfg = shost_priv(shost);
+       struct device *cfgdev = &cfg->dev->dev;
+       struct afu *afu = cfg->afu;
+       int i;
+       u32 mode = MAX_HWQ_MODE;
+
+       for (i = 0; i < MAX_HWQ_MODE; i++) {
+               if (!strncmp(hwq_mode_name[i], buf, strlen(hwq_mode_name[i]))) {
+                       mode = i;
+                       break;
+               }
+       }
+
+       if (mode >= MAX_HWQ_MODE) {
+               dev_info(cfgdev, "Invalid HWQ steering mode.\n");
+               return -EINVAL;
+       }
+
+       if ((mode == HWQ_MODE_TAG) && !shost_use_blk_mq(shost)) {
+               dev_info(cfgdev, "SCSI-MQ is not enabled, use a different "
+                        "HWQ steering mode.\n");
+               return -EINVAL;
+       }
+
+       afu->hwq_mode = mode;
 
        return count;
 }
@@ -2485,6 +3059,8 @@ static DEVICE_ATTR_RO(port1_lun_table);
 static DEVICE_ATTR_RO(port2_lun_table);
 static DEVICE_ATTR_RO(port3_lun_table);
 static DEVICE_ATTR_RW(irqpoll_weight);
+static DEVICE_ATTR_RW(num_hwqs);
+static DEVICE_ATTR_RW(hwq_mode);
 
 static struct device_attribute *cxlflash_host_attrs[] = {
        &dev_attr_port0,
@@ -2498,6 +3074,8 @@ static struct device_attribute *cxlflash_host_attrs[] = {
        &dev_attr_port2_lun_table,
        &dev_attr_port3_lun_table,
        &dev_attr_irqpoll_weight,
+       &dev_attr_num_hwqs,
+       &dev_attr_hwq_mode,
        NULL
 };
 
@@ -2521,6 +3099,7 @@ static struct scsi_host_template driver_template = {
        .ioctl = cxlflash_ioctl,
        .proc_name = CXLFLASH_NAME,
        .queuecommand = cxlflash_queuecommand,
+       .eh_abort_handler = cxlflash_eh_abort_handler,
        .eh_device_reset_handler = cxlflash_eh_device_reset_handler,
        .eh_host_reset_handler = cxlflash_eh_host_reset_handler,
        .change_queue_depth = cxlflash_change_queue_depth,
@@ -2610,6 +3189,86 @@ static void cxlflash_worker_thread(struct work_struct *work)
                scsi_scan_host(cfg->host);
 }
 
+/**
+ * cxlflash_chr_open() - character device open handler
+ * @inode:     Device inode associated with this character device.
+ * @file:      File pointer for this device.
+ *
+ * Only users with admin privileges are allowed to open the character device.
+ *
+ * Return: 0 on success, -errno on failure
+ */
+static int cxlflash_chr_open(struct inode *inode, struct file *file)
+{
+       struct cxlflash_cfg *cfg;
+
+       if (!capable(CAP_SYS_ADMIN))
+               return -EACCES;
+
+       cfg = container_of(inode->i_cdev, struct cxlflash_cfg, cdev);
+       file->private_data = cfg;
+
+       return 0;
+}
+
+/*
+ * Character device file operations
+ */
+static const struct file_operations cxlflash_chr_fops = {
+       .owner          = THIS_MODULE,
+       .open           = cxlflash_chr_open,
+};
+
+/**
+ * init_chrdev() - initialize the character device for the host
+ * @cfg:       Internal structure associated with the host.
+ *
+ * Return: 0 on success, -errno on failure
+ */
+static int init_chrdev(struct cxlflash_cfg *cfg)
+{
+       struct device *dev = &cfg->dev->dev;
+       struct device *char_dev;
+       dev_t devno;
+       int minor;
+       int rc = 0;
+
+       minor = cxlflash_get_minor();
+       if (unlikely(minor < 0)) {
+               dev_err(dev, "%s: Exhausted allowed adapters\n", __func__);
+               rc = -ENOSPC;
+               goto out;
+       }
+
+       devno = MKDEV(cxlflash_major, minor);
+       cdev_init(&cfg->cdev, &cxlflash_chr_fops);
+
+       rc = cdev_add(&cfg->cdev, devno, 1);
+       if (rc) {
+               dev_err(dev, "%s: cdev_add failed rc=%d\n", __func__, rc);
+               goto err1;
+       }
+
+       char_dev = device_create(cxlflash_class, NULL, devno,
+                                NULL, "cxlflash%d", minor);
+       if (IS_ERR(char_dev)) {
+               rc = PTR_ERR(char_dev);
+               dev_err(dev, "%s: device_create failed rc=%d\n",
+                       __func__, rc);
+               goto err2;
+       }
+
+       cfg->chardev = char_dev;
+out:
+       dev_dbg(dev, "%s: returning rc=%d\n", __func__, rc);
+       return rc;
+err2:
+       cdev_del(&cfg->cdev);
+err1:
+       cxlflash_put_minor(minor);
+       goto out;
+}
+
 /**
  * cxlflash_probe() - PCI entry point to add host
  * @pdev:      PCI device associated with the host.
@@ -2720,6 +3379,13 @@ static int cxlflash_probe(struct pci_dev *pdev,
        }
        cfg->init_state = INIT_STATE_SCSI;
 
+       rc = init_chrdev(cfg);
+       if (rc) {
+               dev_err(dev, "%s: init_chrdev failed rc=%d\n", __func__, rc);
+               goto out_remove;
+       }
+       cfg->init_state = INIT_STATE_CDEV;
+
        if (wq_has_sleeper(&cfg->reset_waitq)) {
                cfg->state = STATE_PROBED;
                wake_up_all(&cfg->reset_waitq);
@@ -2822,6 +3488,63 @@ static void cxlflash_pci_resume(struct pci_dev *pdev)
        scsi_unblock_requests(cfg->host);
 }
 
+/**
+ * cxlflash_devnode() - provides devtmpfs for devices in the cxlflash class
+ * @dev:       Character device.
+ * @mode:      Mode that can be used to verify access.
+ *
+ * Return: Allocated string describing the devtmpfs structure.
+ */
+static char *cxlflash_devnode(struct device *dev, umode_t *mode)
+{
+       return kasprintf(GFP_KERNEL, "cxlflash/%s", dev_name(dev));
+}
+
+/**
+ * cxlflash_class_init() - create character device class
+ *
+ * Return: 0 on success, -errno on failure
+ */
+static int cxlflash_class_init(void)
+{
+       dev_t devno;
+       int rc = 0;
+
+       rc = alloc_chrdev_region(&devno, 0, CXLFLASH_MAX_ADAPTERS, "cxlflash");
+       if (unlikely(rc)) {
+               pr_err("%s: alloc_chrdev_region failed rc=%d\n", __func__, rc);
+               goto out;
+       }
+
+       cxlflash_major = MAJOR(devno);
+
+       cxlflash_class = class_create(THIS_MODULE, "cxlflash");
+       if (IS_ERR(cxlflash_class)) {
+               rc = PTR_ERR(cxlflash_class);
+               pr_err("%s: class_create failed rc=%d\n", __func__, rc);
+               goto err;
+       }
+
+       cxlflash_class->devnode = cxlflash_devnode;
+out:
+       pr_debug("%s: returning rc=%d\n", __func__, rc);
+       return rc;
+err:
+       unregister_chrdev_region(devno, CXLFLASH_MAX_ADAPTERS);
+       goto out;
+}
+
+/**
+ * cxlflash_class_exit() - destroy character device class
+ */
+static void cxlflash_class_exit(void)
+{
+       dev_t devno = MKDEV(cxlflash_major, 0);
+
+       class_destroy(cxlflash_class);
+       unregister_chrdev_region(devno, CXLFLASH_MAX_ADAPTERS);
+}
+
 static const struct pci_error_handlers cxlflash_err_handler = {
        .error_detected = cxlflash_pci_error_detected,
        .slot_reset = cxlflash_pci_slot_reset,
@@ -2847,9 +3570,23 @@ static struct pci_driver cxlflash_driver = {
  */
 static int __init init_cxlflash(void)
 {
+       int rc;
+
+       check_sizes();
        cxlflash_list_init();
+       rc = cxlflash_class_init();
+       if (unlikely(rc))
+               goto out;
 
-       return pci_register_driver(&cxlflash_driver);
+       rc = pci_register_driver(&cxlflash_driver);
+       if (unlikely(rc))
+               goto err;
+out:
+       pr_debug("%s: returning rc=%d\n", __func__, rc);
+       return rc;
+err:
+       cxlflash_class_exit();
+       goto out;
 }
 
 /**
@@ -2861,6 +3598,7 @@ static void __exit exit_cxlflash(void)
        cxlflash_free_errpage();
 
        pci_unregister_driver(&cxlflash_driver);
+       cxlflash_class_exit();
 }
 
 module_init(init_cxlflash);