scsi: elx: efct: Hardware I/O and SGL initialization
authorJames Smart <jsmart2021@gmail.com>
Tue, 1 Jun 2021 23:55:02 +0000 (16:55 -0700)
committerMartin K. Petersen <martin.petersen@oracle.com>
Wed, 16 Jun 2021 03:39:31 +0000 (23:39 -0400)
Add driver definitions for:

 - Routines to create I/O interfaces (wqs, etc), SGL initialization, and
   configure hardware features.

Link: https://lore.kernel.org/r/20210601235512.20104-22-jsmart2021@gmail.com
Reviewed-by: Daniel Wagner <dwagner@suse.de>
Co-developed-by: Ram Vegesna <ram.vegesna@broadcom.com>
Signed-off-by: Ram Vegesna <ram.vegesna@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
drivers/scsi/elx/efct/efct_hw.c
drivers/scsi/elx/efct/efct_hw.h

index d4b1fcefcfe09a5bf9982bd2c4ed57aadb39374d..41eb0f9e3334f8263fbb53ea3ce22479c4f19fb6 100644 (file)
@@ -1557,3 +1557,577 @@ efct_issue_mbox_rqst(void *base, void *cmd, void *cb, void *arg)
 
        return 0;
 }
+
+static inline struct efct_hw_io *
+_efct_hw_io_alloc(struct efct_hw *hw)
+{
+       struct efct_hw_io *io = NULL;
+
+       if (!list_empty(&hw->io_free)) {
+               io = list_first_entry(&hw->io_free, struct efct_hw_io,
+                                     list_entry);
+               list_del(&io->list_entry);
+       }
+       if (io) {
+               INIT_LIST_HEAD(&io->list_entry);
+               list_add_tail(&io->list_entry, &hw->io_inuse);
+               io->state = EFCT_HW_IO_STATE_INUSE;
+               io->abort_reqtag = U32_MAX;
+               io->wq = hw->wq_cpu_array[raw_smp_processor_id()];
+               if (!io->wq) {
+                       efc_log_err(hw->os, "WQ not assigned for cpu:%d\n",
+                                   raw_smp_processor_id());
+                       io->wq = hw->hw_wq[0];
+               }
+               kref_init(&io->ref);
+               io->release = efct_hw_io_free_internal;
+       } else {
+               atomic_add(1, &hw->io_alloc_failed_count);
+       }
+
+       return io;
+}
+
+struct efct_hw_io *
+efct_hw_io_alloc(struct efct_hw *hw)
+{
+       struct efct_hw_io *io = NULL;
+       unsigned long flags = 0;
+
+       spin_lock_irqsave(&hw->io_lock, flags);
+       io = _efct_hw_io_alloc(hw);
+       spin_unlock_irqrestore(&hw->io_lock, flags);
+
+       return io;
+}
+
+static void
+efct_hw_io_free_move_correct_list(struct efct_hw *hw,
+                                 struct efct_hw_io *io)
+{
+       /*
+        * When an IO is freed, depending on the exchange busy flag,
+        * move it to the correct list.
+        */
+       if (io->xbusy) {
+               /*
+                * add to wait_free list and wait for XRI_ABORTED CQEs to clean
+                * up
+                */
+               INIT_LIST_HEAD(&io->list_entry);
+               list_add_tail(&io->list_entry, &hw->io_wait_free);
+               io->state = EFCT_HW_IO_STATE_WAIT_FREE;
+       } else {
+               /* IO not busy, add to free list */
+               INIT_LIST_HEAD(&io->list_entry);
+               list_add_tail(&io->list_entry, &hw->io_free);
+               io->state = EFCT_HW_IO_STATE_FREE;
+       }
+}
+
+static inline void
+efct_hw_io_free_common(struct efct_hw *hw, struct efct_hw_io *io)
+{
+       /* initialize IO fields */
+       efct_hw_init_free_io(io);
+
+       /* Restore default SGL */
+       efct_hw_io_restore_sgl(hw, io);
+}
+
+void
+efct_hw_io_free_internal(struct kref *arg)
+{
+       unsigned long flags = 0;
+       struct efct_hw_io *io = container_of(arg, struct efct_hw_io, ref);
+       struct efct_hw *hw = io->hw;
+
+       /* perform common cleanup */
+       efct_hw_io_free_common(hw, io);
+
+       spin_lock_irqsave(&hw->io_lock, flags);
+       /* remove from in-use list */
+       if (!list_empty(&io->list_entry) && !list_empty(&hw->io_inuse)) {
+               list_del_init(&io->list_entry);
+               efct_hw_io_free_move_correct_list(hw, io);
+       }
+       spin_unlock_irqrestore(&hw->io_lock, flags);
+}
+
+int
+efct_hw_io_free(struct efct_hw *hw, struct efct_hw_io *io)
+{
+       return kref_put(&io->ref, io->release);
+}
+
+struct efct_hw_io *
+efct_hw_io_lookup(struct efct_hw *hw, u32 xri)
+{
+       u32 ioindex;
+
+       ioindex = xri - hw->sli.ext[SLI4_RSRC_XRI].base[0];
+       return hw->io[ioindex];
+}
+
+int
+efct_hw_io_init_sges(struct efct_hw *hw, struct efct_hw_io *io,
+                    enum efct_hw_io_type type)
+{
+       struct sli4_sge *data = NULL;
+       u32 i = 0;
+       u32 skips = 0;
+       u32 sge_flags = 0;
+
+       if (!io) {
+               efc_log_err(hw->os, "bad parameter hw=%p io=%p\n", hw, io);
+               return -EIO;
+       }
+
+       /* Clear / reset the scatter-gather list */
+       io->sgl = &io->def_sgl;
+       io->sgl_count = io->def_sgl_count;
+       io->first_data_sge = 0;
+
+       memset(io->sgl->virt, 0, 2 * sizeof(struct sli4_sge));
+       io->n_sge = 0;
+       io->sge_offset = 0;
+
+       io->type = type;
+
+       data = io->sgl->virt;
+
+       /*
+        * Some IO types have underlying hardware requirements on the order
+        * of SGEs. Process all special entries here.
+        */
+       switch (type) {
+       case EFCT_HW_IO_TARGET_WRITE:
+
+               /* populate host resident XFER_RDY buffer */
+               sge_flags = le32_to_cpu(data->dw2_flags);
+               sge_flags &= (~SLI4_SGE_TYPE_MASK);
+               sge_flags |= (SLI4_SGE_TYPE_DATA << SLI4_SGE_TYPE_SHIFT);
+               data->buffer_address_high =
+                       cpu_to_le32(upper_32_bits(io->xfer_rdy.phys));
+               data->buffer_address_low  =
+                       cpu_to_le32(lower_32_bits(io->xfer_rdy.phys));
+               data->buffer_length = cpu_to_le32(io->xfer_rdy.size);
+               data->dw2_flags = cpu_to_le32(sge_flags);
+               data++;
+
+               skips = EFCT_TARGET_WRITE_SKIPS;
+
+               io->n_sge = 1;
+               break;
+       case EFCT_HW_IO_TARGET_READ:
+               /*
+                * For FCP_TSEND64, the first 2 entries are SKIP SGE's
+                */
+               skips = EFCT_TARGET_READ_SKIPS;
+               break;
+       case EFCT_HW_IO_TARGET_RSP:
+               /*
+                * No skips, etc. for FCP_TRSP64
+                */
+               break;
+       default:
+               efc_log_err(hw->os, "unsupported IO type %#x\n", type);
+               return -EIO;
+       }
+
+       /*
+        * Write skip entries
+        */
+       for (i = 0; i < skips; i++) {
+               sge_flags = le32_to_cpu(data->dw2_flags);
+               sge_flags &= (~SLI4_SGE_TYPE_MASK);
+               sge_flags |= (SLI4_SGE_TYPE_SKIP << SLI4_SGE_TYPE_SHIFT);
+               data->dw2_flags = cpu_to_le32(sge_flags);
+               data++;
+       }
+
+       io->n_sge += skips;
+
+       /*
+        * Set last
+        */
+       sge_flags = le32_to_cpu(data->dw2_flags);
+       sge_flags |= SLI4_SGE_LAST;
+       data->dw2_flags = cpu_to_le32(sge_flags);
+
+       return 0;
+}
+
+int
+efct_hw_io_add_sge(struct efct_hw *hw, struct efct_hw_io *io,
+                  uintptr_t addr, u32 length)
+{
+       struct sli4_sge *data = NULL;
+       u32 sge_flags = 0;
+
+       if (!io || !addr || !length) {
+               efc_log_err(hw->os,
+                           "bad parameter hw=%p io=%p addr=%lx length=%u\n",
+                           hw, io, addr, length);
+               return -EIO;
+       }
+
+       if (length > hw->sli.sge_supported_length) {
+               efc_log_err(hw->os,
+                           "length of SGE %d bigger than allowed %d\n",
+                           length, hw->sli.sge_supported_length);
+               return -EIO;
+       }
+
+       data = io->sgl->virt;
+       data += io->n_sge;
+
+       sge_flags = le32_to_cpu(data->dw2_flags);
+       sge_flags &= ~SLI4_SGE_TYPE_MASK;
+       sge_flags |= SLI4_SGE_TYPE_DATA << SLI4_SGE_TYPE_SHIFT;
+       sge_flags &= ~SLI4_SGE_DATA_OFFSET_MASK;
+       sge_flags |= SLI4_SGE_DATA_OFFSET_MASK & io->sge_offset;
+
+       data->buffer_address_high = cpu_to_le32(upper_32_bits(addr));
+       data->buffer_address_low  = cpu_to_le32(lower_32_bits(addr));
+       data->buffer_length = cpu_to_le32(length);
+
+       /*
+        * Always assume this is the last entry and mark as such.
+        * If this is not the first entry unset the "last SGE"
+        * indication for the previous entry
+        */
+       sge_flags |= SLI4_SGE_LAST;
+       data->dw2_flags = cpu_to_le32(sge_flags);
+
+       if (io->n_sge) {
+               sge_flags = le32_to_cpu(data[-1].dw2_flags);
+               sge_flags &= ~SLI4_SGE_LAST;
+               data[-1].dw2_flags = cpu_to_le32(sge_flags);
+       }
+
+       /* Set first_data_bde if not previously set */
+       if (io->first_data_sge == 0)
+               io->first_data_sge = io->n_sge;
+
+       io->sge_offset += length;
+       io->n_sge++;
+
+       return 0;
+}
+
+void
+efct_hw_io_abort_all(struct efct_hw *hw)
+{
+       struct efct_hw_io *io_to_abort  = NULL;
+       struct efct_hw_io *next_io = NULL;
+
+       list_for_each_entry_safe(io_to_abort, next_io,
+                                &hw->io_inuse, list_entry) {
+               efct_hw_io_abort(hw, io_to_abort, true, NULL, NULL);
+       }
+}
+
+static void
+efct_hw_wq_process_abort(void *arg, u8 *cqe, int status)
+{
+       struct efct_hw_io *io = arg;
+       struct efct_hw *hw = io->hw;
+       u32 ext = 0;
+       u32 len = 0;
+       struct hw_wq_callback *wqcb;
+
+       /*
+        * For IOs that were aborted internally, we may need to issue the
+        * callback here depending on whether a XRI_ABORTED CQE is expected ot
+        * not. If the status is Local Reject/No XRI, then
+        * issue the callback now.
+        */
+       ext = sli_fc_ext_status(&hw->sli, cqe);
+       if (status == SLI4_FC_WCQE_STATUS_LOCAL_REJECT &&
+           ext == SLI4_FC_LOCAL_REJECT_NO_XRI && io->done) {
+               efct_hw_done_t done = io->done;
+
+               io->done = NULL;
+
+               /*
+                * Use latched status as this is always saved for an internal
+                * abort Note: We won't have both a done and abort_done
+                * function, so don't worry about
+                *       clobbering the len, status and ext fields.
+                */
+               status = io->saved_status;
+               len = io->saved_len;
+               ext = io->saved_ext;
+               io->status_saved = false;
+               done(io, len, status, ext, io->arg);
+       }
+
+       if (io->abort_done) {
+               efct_hw_done_t done = io->abort_done;
+
+               io->abort_done = NULL;
+               done(io, len, status, ext, io->abort_arg);
+       }
+
+       /* clear abort bit to indicate abort is complete */
+       io->abort_in_progress = false;
+
+       /* Free the WQ callback */
+       if (io->abort_reqtag == U32_MAX) {
+               efc_log_err(hw->os, "HW IO already freed\n");
+               return;
+       }
+
+       wqcb = efct_hw_reqtag_get_instance(hw, io->abort_reqtag);
+       efct_hw_reqtag_free(hw, wqcb);
+
+       /*
+        * Call efct_hw_io_free() because this releases the WQ reservation as
+        * well as doing the refcount put. Don't duplicate the code here.
+        */
+       (void)efct_hw_io_free(hw, io);
+}
+
+static void
+efct_hw_fill_abort_wqe(struct efct_hw *hw, struct efct_hw_wqe *wqe)
+{
+       struct sli4_abort_wqe *abort = (void *)wqe->wqebuf;
+
+       memset(abort, 0, hw->sli.wqe_size);
+
+       abort->criteria = SLI4_ABORT_CRITERIA_XRI_TAG;
+       abort->ia_ir_byte |= wqe->send_abts ? 0 : 1;
+
+       /* Suppress ABTS retries */
+       abort->ia_ir_byte |= SLI4_ABRT_WQE_IR;
+
+       abort->t_tag  = cpu_to_le32(wqe->id);
+       abort->command = SLI4_WQE_ABORT;
+       abort->request_tag = cpu_to_le16(wqe->abort_reqtag);
+
+       abort->dw10w0_flags = cpu_to_le16(SLI4_ABRT_WQE_QOSD);
+
+       abort->cq_id = cpu_to_le16(SLI4_CQ_DEFAULT);
+}
+
+int
+efct_hw_io_abort(struct efct_hw *hw, struct efct_hw_io *io_to_abort,
+                bool send_abts, void *cb, void *arg)
+{
+       struct hw_wq_callback *wqcb;
+       unsigned long flags = 0;
+
+       if (!io_to_abort) {
+               efc_log_err(hw->os, "bad parameter hw=%p io=%p\n",
+                           hw, io_to_abort);
+               return -EIO;
+       }
+
+       if (hw->state != EFCT_HW_STATE_ACTIVE) {
+               efc_log_err(hw->os, "cannot send IO abort, HW state=%d\n",
+                           hw->state);
+               return -EIO;
+       }
+
+       /* take a reference on IO being aborted */
+       if (kref_get_unless_zero(&io_to_abort->ref) == 0) {
+               /* command no longer active */
+               efc_log_debug(hw->os,
+                             "io not active xri=0x%x tag=0x%x\n",
+                             io_to_abort->indicator, io_to_abort->reqtag);
+               return -ENOENT;
+       }
+
+       /* Must have a valid WQ reference */
+       if (!io_to_abort->wq) {
+               efc_log_debug(hw->os, "io_to_abort xri=0x%x not active on WQ\n",
+                             io_to_abort->indicator);
+               /* efct_ref_get(): same function */
+               kref_put(&io_to_abort->ref, io_to_abort->release);
+               return -ENOENT;
+       }
+
+       /*
+        * Validation checks complete; now check to see if already being
+        * aborted, if not set the flag.
+        */
+       if (cmpxchg(&io_to_abort->abort_in_progress, false, true)) {
+               /* efct_ref_get(): same function */
+               kref_put(&io_to_abort->ref, io_to_abort->release);
+               efc_log_debug(hw->os,
+                             "io already being aborted xri=0x%x tag=0x%x\n",
+                             io_to_abort->indicator, io_to_abort->reqtag);
+               return -EINPROGRESS;
+       }
+
+       /*
+        * If we got here, the possibilities are:
+        * - host owned xri
+        *      - io_to_abort->wq_index != U32_MAX
+        *              - submit ABORT_WQE to same WQ
+        * - port owned xri:
+        *      - rxri: io_to_abort->wq_index == U32_MAX
+        *              - submit ABORT_WQE to any WQ
+        *      - non-rxri
+        *              - io_to_abort->index != U32_MAX
+        *                      - submit ABORT_WQE to same WQ
+        *              - io_to_abort->index == U32_MAX
+        *                      - submit ABORT_WQE to any WQ
+        */
+       io_to_abort->abort_done = cb;
+       io_to_abort->abort_arg  = arg;
+
+       /* Allocate a request tag for the abort portion of this IO */
+       wqcb = efct_hw_reqtag_alloc(hw, efct_hw_wq_process_abort, io_to_abort);
+       if (!wqcb) {
+               efc_log_err(hw->os, "can't allocate request tag\n");
+               return -ENOSPC;
+       }
+
+       io_to_abort->abort_reqtag = wqcb->instance_index;
+       io_to_abort->wqe.send_abts = send_abts;
+       io_to_abort->wqe.id = io_to_abort->indicator;
+       io_to_abort->wqe.abort_reqtag = io_to_abort->abort_reqtag;
+
+       /*
+        * If the wqe is on the pending list, then set this wqe to be
+        * aborted when the IO's wqe is removed from the list.
+        */
+       if (io_to_abort->wq) {
+               spin_lock_irqsave(&io_to_abort->wq->queue->lock, flags);
+               if (io_to_abort->wqe.list_entry.next) {
+                       io_to_abort->wqe.abort_wqe_submit_needed = true;
+                       spin_unlock_irqrestore(&io_to_abort->wq->queue->lock,
+                                              flags);
+                       return 0;
+               }
+               spin_unlock_irqrestore(&io_to_abort->wq->queue->lock, flags);
+       }
+
+       efct_hw_fill_abort_wqe(hw, &io_to_abort->wqe);
+
+       /* ABORT_WQE does not actually utilize an XRI on the Port,
+        * therefore, keep xbusy as-is to track the exchange's state,
+        * not the ABORT_WQE's state
+        */
+       if (efct_hw_wq_write(io_to_abort->wq, &io_to_abort->wqe)) {
+               io_to_abort->abort_in_progress = false;
+               /* efct_ref_get(): same function */
+               kref_put(&io_to_abort->ref, io_to_abort->release);
+               return -EIO;
+       }
+
+       return 0;
+}
+
+void
+efct_hw_reqtag_pool_free(struct efct_hw *hw)
+{
+       u32 i;
+       struct reqtag_pool *reqtag_pool = hw->wq_reqtag_pool;
+       struct hw_wq_callback *wqcb = NULL;
+
+       if (reqtag_pool) {
+               for (i = 0; i < U16_MAX; i++) {
+                       wqcb = reqtag_pool->tags[i];
+                       if (!wqcb)
+                               continue;
+
+                       kfree(wqcb);
+               }
+               kfree(reqtag_pool);
+               hw->wq_reqtag_pool = NULL;
+       }
+}
+
+struct reqtag_pool *
+efct_hw_reqtag_pool_alloc(struct efct_hw *hw)
+{
+       u32 i = 0;
+       struct reqtag_pool *reqtag_pool;
+       struct hw_wq_callback *wqcb;
+
+       reqtag_pool = kzalloc(sizeof(*reqtag_pool), GFP_KERNEL);
+       if (!reqtag_pool)
+               return NULL;
+
+       INIT_LIST_HEAD(&reqtag_pool->freelist);
+       /* initialize reqtag pool lock */
+       spin_lock_init(&reqtag_pool->lock);
+       for (i = 0; i < U16_MAX; i++) {
+               wqcb = kmalloc(sizeof(*wqcb), GFP_KERNEL);
+               if (!wqcb)
+                       break;
+
+               reqtag_pool->tags[i] = wqcb;
+               wqcb->instance_index = i;
+               wqcb->callback = NULL;
+               wqcb->arg = NULL;
+               INIT_LIST_HEAD(&wqcb->list_entry);
+               list_add_tail(&wqcb->list_entry, &reqtag_pool->freelist);
+       }
+
+       return reqtag_pool;
+}
+
+struct hw_wq_callback *
+efct_hw_reqtag_alloc(struct efct_hw *hw,
+                    void (*callback)(void *arg, u8 *cqe, int status),
+                    void *arg)
+{
+       struct hw_wq_callback *wqcb = NULL;
+       struct reqtag_pool *reqtag_pool = hw->wq_reqtag_pool;
+       unsigned long flags = 0;
+
+       if (!callback)
+               return wqcb;
+
+       spin_lock_irqsave(&reqtag_pool->lock, flags);
+
+       if (!list_empty(&reqtag_pool->freelist)) {
+               wqcb = list_first_entry(&reqtag_pool->freelist,
+                                       struct hw_wq_callback, list_entry);
+       }
+
+       if (wqcb) {
+               list_del_init(&wqcb->list_entry);
+               spin_unlock_irqrestore(&reqtag_pool->lock, flags);
+               wqcb->callback = callback;
+               wqcb->arg = arg;
+       } else {
+               spin_unlock_irqrestore(&reqtag_pool->lock, flags);
+       }
+
+       return wqcb;
+}
+
+void
+efct_hw_reqtag_free(struct efct_hw *hw, struct hw_wq_callback *wqcb)
+{
+       unsigned long flags = 0;
+       struct reqtag_pool *reqtag_pool = hw->wq_reqtag_pool;
+
+       if (!wqcb->callback)
+               efc_log_err(hw->os, "WQCB is already freed\n");
+
+       spin_lock_irqsave(&reqtag_pool->lock, flags);
+       wqcb->callback = NULL;
+       wqcb->arg = NULL;
+       INIT_LIST_HEAD(&wqcb->list_entry);
+       list_add(&wqcb->list_entry, &hw->wq_reqtag_pool->freelist);
+       spin_unlock_irqrestore(&reqtag_pool->lock, flags);
+}
+
+struct hw_wq_callback *
+efct_hw_reqtag_get_instance(struct efct_hw *hw, u32 instance_index)
+{
+       struct hw_wq_callback *wqcb;
+
+       wqcb = hw->wq_reqtag_pool->tags[instance_index];
+       if (!wqcb)
+               efc_log_err(hw->os, "wqcb for instance %d is null\n",
+                           instance_index);
+
+       return wqcb;
+}
index e396dfcd8cdb963913089b267576f4778c867f28..90c4f4aea77446bb777cd42e1cc54bee3f97f614 100644 (file)
@@ -611,4 +611,45 @@ efct_hw_command(struct efct_hw *hw, u8 *cmd, u32 opts, void *cb,
 int
 efct_issue_mbox_rqst(void *base, void *cmd, void *cb, void *arg);
 
+struct efct_hw_io *efct_hw_io_alloc(struct efct_hw *hw);
+int efct_hw_io_free(struct efct_hw *hw, struct efct_hw_io *io);
+u8 efct_hw_io_inuse(struct efct_hw *hw, struct efct_hw_io *io);
+int
+efct_hw_io_send(struct efct_hw *hw, enum efct_hw_io_type type,
+               struct efct_hw_io *io, union efct_hw_io_param_u *iparam,
+               void *cb, void *arg);
+int
+efct_hw_io_register_sgl(struct efct_hw *hw, struct efct_hw_io *io,
+                       struct efc_dma *sgl,
+                       u32 sgl_count);
+int
+efct_hw_io_init_sges(struct efct_hw *hw,
+                    struct efct_hw_io *io, enum efct_hw_io_type type);
+
+int
+efct_hw_io_add_sge(struct efct_hw *hw, struct efct_hw_io *io,
+                  uintptr_t addr, u32 length);
+int
+efct_hw_io_abort(struct efct_hw *hw, struct efct_hw_io *io_to_abort,
+                bool send_abts, void *cb, void *arg);
+u32
+efct_hw_io_get_count(struct efct_hw *hw,
+                    enum efct_hw_io_count_type io_count_type);
+struct efct_hw_io
+*efct_hw_io_lookup(struct efct_hw *hw, u32 indicator);
+void efct_hw_io_abort_all(struct efct_hw *hw);
+void efct_hw_io_free_internal(struct kref *arg);
+
+/* HW WQ request tag API */
+struct reqtag_pool *efct_hw_reqtag_pool_alloc(struct efct_hw *hw);
+void efct_hw_reqtag_pool_free(struct efct_hw *hw);
+struct hw_wq_callback
+*efct_hw_reqtag_alloc(struct efct_hw *hw,
+                       void (*callback)(void *arg, u8 *cqe,
+                                        int status), void *arg);
+void
+efct_hw_reqtag_free(struct efct_hw *hw, struct hw_wq_callback *wqcb);
+struct hw_wq_callback
+*efct_hw_reqtag_get_instance(struct efct_hw *hw, u32 instance_index);
+
 #endif /* __EFCT_H__ */