These changes are the raw update to linux-4.4.6-rt14. Kernel sources
[kvmfornfv.git] / kernel / drivers / net / ethernet / cavium / liquidio / request_manager.c
diff --git a/kernel/drivers/net/ethernet/cavium/liquidio/request_manager.c b/kernel/drivers/net/ethernet/cavium/liquidio/request_manager.c
new file mode 100644 (file)
index 0000000..a2a2465
--- /dev/null
@@ -0,0 +1,765 @@
+/**********************************************************************
+ * Author: Cavium, Inc.
+ *
+ * Contact: support@cavium.com
+ *          Please include "LiquidIO" in the subject.
+ *
+ * Copyright (c) 2003-2015 Cavium, Inc.
+ *
+ * This file is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, Version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This file is distributed in the hope that it will be useful, but
+ * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty
+ * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or
+ * NONINFRINGEMENT.  See the GNU General Public License for more
+ * details.
+ *
+ * This file may also be available under a different license from Cavium.
+ * Contact Cavium, Inc. for more information
+ **********************************************************************/
+#include <linux/version.h>
+#include <linux/types.h>
+#include <linux/list.h>
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/kthread.h>
+#include <linux/netdevice.h>
+#include <linux/vmalloc.h>
+#include "octeon_config.h"
+#include "liquidio_common.h"
+#include "octeon_droq.h"
+#include "octeon_iq.h"
+#include "response_manager.h"
+#include "octeon_device.h"
+#include "octeon_nic.h"
+#include "octeon_main.h"
+#include "octeon_network.h"
+#include "cn66xx_regs.h"
+#include "cn66xx_device.h"
+#include "cn68xx_regs.h"
+#include "cn68xx_device.h"
+#include "liquidio_image.h"
+
+#define INCR_INSTRQUEUE_PKT_COUNT(octeon_dev_ptr, iq_no, field, count)  \
+       (octeon_dev_ptr->instr_queue[iq_no]->stats.field += count)
+
+struct iq_post_status {
+       int status;
+       int index;
+};
+
+static void check_db_timeout(struct work_struct *work);
+static void  __check_db_timeout(struct octeon_device *oct, unsigned long iq_no);
+
+static void (*reqtype_free_fn[MAX_OCTEON_DEVICES][REQTYPE_LAST + 1]) (void *);
+
+static inline int IQ_INSTR_MODE_64B(struct octeon_device *oct, int iq_no)
+{
+       struct octeon_instr_queue *iq =
+           (struct octeon_instr_queue *)oct->instr_queue[iq_no];
+       return iq->iqcmd_64B;
+}
+
+#define IQ_INSTR_MODE_32B(oct, iq_no)  (!IQ_INSTR_MODE_64B(oct, iq_no))
+
+/* Define this to return the request status comaptible to old code */
+/*#define OCTEON_USE_OLD_REQ_STATUS*/
+
+/* Return 0 on success, 1 on failure */
+int octeon_init_instr_queue(struct octeon_device *oct,
+                           u32 iq_no, u32 num_descs)
+{
+       struct octeon_instr_queue *iq;
+       struct octeon_iq_config *conf = NULL;
+       u32 q_size;
+       struct cavium_wq *db_wq;
+
+       if (OCTEON_CN6XXX(oct))
+               conf = &(CFG_GET_IQ_CFG(CHIP_FIELD(oct, cn6xxx, conf)));
+
+       if (!conf) {
+               dev_err(&oct->pci_dev->dev, "Unsupported Chip %x\n",
+                       oct->chip_id);
+               return 1;
+       }
+
+       if (num_descs & (num_descs - 1)) {
+               dev_err(&oct->pci_dev->dev,
+                       "Number of descriptors for instr queue %d not in power of 2.\n",
+                       iq_no);
+               return 1;
+       }
+
+       q_size = (u32)conf->instr_type * num_descs;
+
+       iq = oct->instr_queue[iq_no];
+
+       iq->base_addr = lio_dma_alloc(oct, q_size,
+                                     (dma_addr_t *)&iq->base_addr_dma);
+       if (!iq->base_addr) {
+               dev_err(&oct->pci_dev->dev, "Cannot allocate memory for instr queue %d\n",
+                       iq_no);
+               return 1;
+       }
+
+       iq->max_count = num_descs;
+
+       /* Initialize a list to holds requests that have been posted to Octeon
+        * but has yet to be fetched by octeon
+        */
+       iq->request_list = vmalloc(sizeof(*iq->request_list) * num_descs);
+       if (!iq->request_list) {
+               lio_dma_free(oct, q_size, iq->base_addr, iq->base_addr_dma);
+               dev_err(&oct->pci_dev->dev, "Alloc failed for IQ[%d] nr free list\n",
+                       iq_no);
+               return 1;
+       }
+
+       memset(iq->request_list, 0, sizeof(*iq->request_list) * num_descs);
+
+       dev_dbg(&oct->pci_dev->dev, "IQ[%d]: base: %p basedma: %llx count: %d\n",
+               iq_no, iq->base_addr, iq->base_addr_dma, iq->max_count);
+
+       iq->iq_no = iq_no;
+       iq->fill_threshold = (u32)conf->db_min;
+       iq->fill_cnt = 0;
+       iq->host_write_index = 0;
+       iq->octeon_read_index = 0;
+       iq->flush_index = 0;
+       iq->last_db_time = 0;
+       iq->do_auto_flush = 1;
+       iq->db_timeout = (u32)conf->db_timeout;
+       atomic_set(&iq->instr_pending, 0);
+
+       /* Initialize the spinlock for this instruction queue */
+       spin_lock_init(&iq->lock);
+
+       oct->io_qmask.iq |= (1 << iq_no);
+
+       /* Set the 32B/64B mode for each input queue */
+       oct->io_qmask.iq64B |= ((conf->instr_type == 64) << iq_no);
+       iq->iqcmd_64B = (conf->instr_type == 64);
+
+       oct->fn_list.setup_iq_regs(oct, iq_no);
+
+       oct->check_db_wq[iq_no].wq = create_workqueue("check_iq_db");
+       if (!oct->check_db_wq[iq_no].wq) {
+               lio_dma_free(oct, q_size, iq->base_addr, iq->base_addr_dma);
+               dev_err(&oct->pci_dev->dev, "check db wq create failed for iq %d\n",
+                       iq_no);
+               return 1;
+       }
+
+       db_wq = &oct->check_db_wq[iq_no];
+
+       INIT_DELAYED_WORK(&db_wq->wk.work, check_db_timeout);
+       db_wq->wk.ctxptr = oct;
+       db_wq->wk.ctxul = iq_no;
+       queue_delayed_work(db_wq->wq, &db_wq->wk.work, msecs_to_jiffies(1));
+
+       return 0;
+}
+
+int octeon_delete_instr_queue(struct octeon_device *oct, u32 iq_no)
+{
+       u64 desc_size = 0, q_size;
+       struct octeon_instr_queue *iq = oct->instr_queue[iq_no];
+
+       cancel_delayed_work_sync(&oct->check_db_wq[iq_no].wk.work);
+       flush_workqueue(oct->check_db_wq[iq_no].wq);
+       destroy_workqueue(oct->check_db_wq[iq_no].wq);
+
+       if (OCTEON_CN6XXX(oct))
+               desc_size =
+                   CFG_GET_IQ_INSTR_TYPE(CHIP_FIELD(oct, cn6xxx, conf));
+
+       vfree(iq->request_list);
+
+       if (iq->base_addr) {
+               q_size = iq->max_count * desc_size;
+               lio_dma_free(oct, (u32)q_size, iq->base_addr,
+                            iq->base_addr_dma);
+               return 0;
+       }
+       return 1;
+}
+
+/* Return 0 on success, 1 on failure */
+int octeon_setup_iq(struct octeon_device *oct,
+                   u32 iq_no,
+                   u32 num_descs,
+                   void *app_ctx)
+{
+       if (oct->instr_queue[iq_no]) {
+               dev_dbg(&oct->pci_dev->dev, "IQ is in use. Cannot create the IQ: %d again\n",
+                       iq_no);
+               oct->instr_queue[iq_no]->app_ctx = app_ctx;
+               return 0;
+       }
+       oct->instr_queue[iq_no] =
+           vmalloc(sizeof(struct octeon_instr_queue));
+       if (!oct->instr_queue[iq_no])
+               return 1;
+
+       memset(oct->instr_queue[iq_no], 0,
+              sizeof(struct octeon_instr_queue));
+
+       oct->instr_queue[iq_no]->app_ctx = app_ctx;
+       if (octeon_init_instr_queue(oct, iq_no, num_descs)) {
+               vfree(oct->instr_queue[iq_no]);
+               oct->instr_queue[iq_no] = NULL;
+               return 1;
+       }
+
+       oct->num_iqs++;
+       oct->fn_list.enable_io_queues(oct);
+       return 0;
+}
+
+int lio_wait_for_instr_fetch(struct octeon_device *oct)
+{
+       int i, retry = 1000, pending, instr_cnt = 0;
+
+       do {
+               instr_cnt = 0;
+
+               /*for (i = 0; i < oct->num_iqs; i++) {*/
+               for (i = 0; i < MAX_OCTEON_INSTR_QUEUES; i++) {
+                       if (!(oct->io_qmask.iq & (1UL << i)))
+                               continue;
+                       pending =
+                           atomic_read(&oct->
+                                              instr_queue[i]->instr_pending);
+                       if (pending)
+                               __check_db_timeout(oct, i);
+                       instr_cnt += pending;
+               }
+
+               if (instr_cnt == 0)
+                       break;
+
+               schedule_timeout_uninterruptible(1);
+
+       } while (retry-- && instr_cnt);
+
+       return instr_cnt;
+}
+
+static inline void
+ring_doorbell(struct octeon_device *oct, struct octeon_instr_queue *iq)
+{
+       if (atomic_read(&oct->status) == OCT_DEV_RUNNING) {
+               writel(iq->fill_cnt, iq->doorbell_reg);
+               /* make sure doorbell write goes through */
+               mmiowb();
+               iq->fill_cnt = 0;
+               iq->last_db_time = jiffies;
+               return;
+       }
+}
+
+static inline void __copy_cmd_into_iq(struct octeon_instr_queue *iq,
+                                     u8 *cmd)
+{
+       u8 *iqptr, cmdsize;
+
+       cmdsize = ((iq->iqcmd_64B) ? 64 : 32);
+       iqptr = iq->base_addr + (cmdsize * iq->host_write_index);
+
+       memcpy(iqptr, cmd, cmdsize);
+}
+
+static inline int
+__post_command(struct octeon_device *octeon_dev __attribute__((unused)),
+              struct octeon_instr_queue *iq,
+              u32 force_db __attribute__((unused)), u8 *cmd)
+{
+       u32 index = -1;
+
+       /* This ensures that the read index does not wrap around to the same
+        * position if queue gets full before Octeon could fetch any instr.
+        */
+       if (atomic_read(&iq->instr_pending) >= (s32)(iq->max_count - 1))
+               return -1;
+
+       __copy_cmd_into_iq(iq, cmd);
+
+       /* "index" is returned, host_write_index is modified. */
+       index = iq->host_write_index;
+       INCR_INDEX_BY1(iq->host_write_index, iq->max_count);
+       iq->fill_cnt++;
+
+       /* Flush the command into memory. We need to be sure the data is in
+        * memory before indicating that the instruction is pending.
+        */
+       wmb();
+
+       atomic_inc(&iq->instr_pending);
+
+       return index;
+}
+
+static inline struct iq_post_status
+__post_command2(struct octeon_device *octeon_dev __attribute__((unused)),
+               struct octeon_instr_queue *iq,
+               u32 force_db __attribute__((unused)), u8 *cmd)
+{
+       struct iq_post_status st;
+
+       st.status = IQ_SEND_OK;
+
+       /* This ensures that the read index does not wrap around to the same
+        * position if queue gets full before Octeon could fetch any instr.
+        */
+       if (atomic_read(&iq->instr_pending) >= (s32)(iq->max_count - 1)) {
+               st.status = IQ_SEND_FAILED;
+               st.index = -1;
+               return st;
+       }
+
+       if (atomic_read(&iq->instr_pending) >= (s32)(iq->max_count - 2))
+               st.status = IQ_SEND_STOP;
+
+       __copy_cmd_into_iq(iq, cmd);
+
+       /* "index" is returned, host_write_index is modified. */
+       st.index = iq->host_write_index;
+       INCR_INDEX_BY1(iq->host_write_index, iq->max_count);
+       iq->fill_cnt++;
+
+       /* Flush the command into memory. We need to be sure the data is in
+        * memory before indicating that the instruction is pending.
+        */
+       wmb();
+
+       atomic_inc(&iq->instr_pending);
+
+       return st;
+}
+
+int
+octeon_register_reqtype_free_fn(struct octeon_device *oct, int reqtype,
+                               void (*fn)(void *))
+{
+       if (reqtype > REQTYPE_LAST) {
+               dev_err(&oct->pci_dev->dev, "%s: Invalid reqtype: %d\n",
+                       __func__, reqtype);
+               return -EINVAL;
+       }
+
+       reqtype_free_fn[oct->octeon_id][reqtype] = fn;
+
+       return 0;
+}
+
+static inline void
+__add_to_request_list(struct octeon_instr_queue *iq,
+                     int idx, void *buf, int reqtype)
+{
+       iq->request_list[idx].buf = buf;
+       iq->request_list[idx].reqtype = reqtype;
+}
+
+int
+lio_process_iq_request_list(struct octeon_device *oct,
+                           struct octeon_instr_queue *iq)
+{
+       int reqtype;
+       void *buf;
+       u32 old = iq->flush_index;
+       u32 inst_count = 0;
+       unsigned pkts_compl = 0, bytes_compl = 0;
+       struct octeon_soft_command *sc;
+       struct octeon_instr_irh *irh;
+
+       while (old != iq->octeon_read_index) {
+               reqtype = iq->request_list[old].reqtype;
+               buf     = iq->request_list[old].buf;
+
+               if (reqtype == REQTYPE_NONE)
+                       goto skip_this;
+
+               octeon_update_tx_completion_counters(buf, reqtype, &pkts_compl,
+                                                    &bytes_compl);
+
+               switch (reqtype) {
+               case REQTYPE_NORESP_NET:
+               case REQTYPE_NORESP_NET_SG:
+               case REQTYPE_RESP_NET_SG:
+                       reqtype_free_fn[oct->octeon_id][reqtype](buf);
+                       break;
+               case REQTYPE_RESP_NET:
+               case REQTYPE_SOFT_COMMAND:
+                       sc = buf;
+
+                       irh = (struct octeon_instr_irh *)&sc->cmd.irh;
+                       if (irh->rflag) {
+                               /* We're expecting a response from Octeon.
+                                * It's up to lio_process_ordered_list() to
+                                * process  sc. Add sc to the ordered soft
+                                * command response list because we expect
+                                * a response from Octeon.
+                                */
+                               spin_lock_bh(&oct->response_list
+                                       [OCTEON_ORDERED_SC_LIST].lock);
+                               atomic_inc(&oct->response_list
+                                       [OCTEON_ORDERED_SC_LIST].
+                                       pending_req_count);
+                               list_add_tail(&sc->node, &oct->response_list
+                                       [OCTEON_ORDERED_SC_LIST].head);
+                               spin_unlock_bh(&oct->response_list
+                                       [OCTEON_ORDERED_SC_LIST].lock);
+                       } else {
+                               if (sc->callback) {
+                                       sc->callback(oct, OCTEON_REQUEST_DONE,
+                                                    sc->callback_arg);
+                               }
+                       }
+                       break;
+               default:
+                       dev_err(&oct->pci_dev->dev,
+                               "%s Unknown reqtype: %d buf: %p at idx %d\n",
+                               __func__, reqtype, buf, old);
+               }
+
+               iq->request_list[old].buf = NULL;
+               iq->request_list[old].reqtype = 0;
+
+ skip_this:
+               inst_count++;
+               INCR_INDEX_BY1(old, iq->max_count);
+       }
+       if (bytes_compl)
+               octeon_report_tx_completion_to_bql(iq->app_ctx, pkts_compl,
+                                                  bytes_compl);
+       iq->flush_index = old;
+
+       return inst_count;
+}
+
+static inline void
+update_iq_indices(struct octeon_device *oct, struct octeon_instr_queue *iq)
+{
+       u32 inst_processed = 0;
+
+       /* Calculate how many commands Octeon has read and move the read index
+        * accordingly.
+        */
+       iq->octeon_read_index = oct->fn_list.update_iq_read_idx(oct, iq);
+
+       /* Move the NORESPONSE requests to the per-device completion list. */
+       if (iq->flush_index != iq->octeon_read_index)
+               inst_processed = lio_process_iq_request_list(oct, iq);
+
+       if (inst_processed) {
+               atomic_sub(inst_processed, &iq->instr_pending);
+               iq->stats.instr_processed += inst_processed;
+       }
+}
+
+static void
+octeon_flush_iq(struct octeon_device *oct, struct octeon_instr_queue *iq,
+               u32 pending_thresh)
+{
+       if (atomic_read(&iq->instr_pending) >= (s32)pending_thresh) {
+               spin_lock_bh(&iq->lock);
+               update_iq_indices(oct, iq);
+               spin_unlock_bh(&iq->lock);
+       }
+}
+
+static void __check_db_timeout(struct octeon_device *oct, unsigned long iq_no)
+{
+       struct octeon_instr_queue *iq;
+       u64 next_time;
+
+       if (!oct)
+               return;
+       iq = oct->instr_queue[iq_no];
+       if (!iq)
+               return;
+
+       /* If jiffies - last_db_time < db_timeout do nothing  */
+       next_time = iq->last_db_time + iq->db_timeout;
+       if (!time_after(jiffies, (unsigned long)next_time))
+               return;
+       iq->last_db_time = jiffies;
+
+       /* Get the lock and prevent tasklets. This routine gets called from
+        * the poll thread. Instructions can now be posted in tasklet context
+        */
+       spin_lock_bh(&iq->lock);
+       if (iq->fill_cnt != 0)
+               ring_doorbell(oct, iq);
+
+       spin_unlock_bh(&iq->lock);
+
+       /* Flush the instruction queue */
+       if (iq->do_auto_flush)
+               octeon_flush_iq(oct, iq, 1);
+}
+
+/* Called by the Poll thread at regular intervals to check the instruction
+ * queue for commands to be posted and for commands that were fetched by Octeon.
+ */
+static void check_db_timeout(struct work_struct *work)
+{
+       struct cavium_wk *wk = (struct cavium_wk *)work;
+       struct octeon_device *oct = (struct octeon_device *)wk->ctxptr;
+       unsigned long iq_no = wk->ctxul;
+       struct cavium_wq *db_wq = &oct->check_db_wq[iq_no];
+
+       __check_db_timeout(oct, iq_no);
+       queue_delayed_work(db_wq->wq, &db_wq->wk.work, msecs_to_jiffies(1));
+}
+
+int
+octeon_send_command(struct octeon_device *oct, u32 iq_no,
+                   u32 force_db, void *cmd, void *buf,
+                   u32 datasize, u32 reqtype)
+{
+       struct iq_post_status st;
+       struct octeon_instr_queue *iq = oct->instr_queue[iq_no];
+
+       spin_lock_bh(&iq->lock);
+
+       st = __post_command2(oct, iq, force_db, cmd);
+
+       if (st.status != IQ_SEND_FAILED) {
+               octeon_report_sent_bytes_to_bql(buf, reqtype);
+               __add_to_request_list(iq, st.index, buf, reqtype);
+               INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, bytes_sent, datasize);
+               INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, instr_posted, 1);
+
+               if (iq->fill_cnt >= iq->fill_threshold || force_db)
+                       ring_doorbell(oct, iq);
+       } else {
+               INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, instr_dropped, 1);
+       }
+
+       spin_unlock_bh(&iq->lock);
+
+       if (iq->do_auto_flush)
+               octeon_flush_iq(oct, iq, 2);
+
+       return st.status;
+}
+
+void
+octeon_prepare_soft_command(struct octeon_device *oct,
+                           struct octeon_soft_command *sc,
+                           u8 opcode,
+                           u8 subcode,
+                           u32 irh_ossp,
+                           u64 ossp0,
+                           u64 ossp1)
+{
+       struct octeon_config *oct_cfg;
+       struct octeon_instr_ih *ih;
+       struct octeon_instr_irh *irh;
+       struct octeon_instr_rdp *rdp;
+
+       BUG_ON(opcode > 15);
+       BUG_ON(subcode > 127);
+
+       oct_cfg = octeon_get_conf(oct);
+
+       ih          = (struct octeon_instr_ih *)&sc->cmd.ih;
+       ih->tagtype = ATOMIC_TAG;
+       ih->tag     = LIO_CONTROL;
+       ih->raw     = 1;
+       ih->grp     = CFG_GET_CTRL_Q_GRP(oct_cfg);
+
+       if (sc->datasize) {
+               ih->dlengsz = sc->datasize;
+               ih->rs = 1;
+       }
+
+       irh            = (struct octeon_instr_irh *)&sc->cmd.irh;
+       irh->opcode    = opcode;
+       irh->subcode   = subcode;
+
+       /* opcode/subcode specific parameters (ossp) */
+       irh->ossp       = irh_ossp;
+       sc->cmd.ossp[0] = ossp0;
+       sc->cmd.ossp[1] = ossp1;
+
+       if (sc->rdatasize) {
+               rdp            = (struct octeon_instr_rdp *)&sc->cmd.rdp;
+               rdp->pcie_port = oct->pcie_port;
+               rdp->rlen      = sc->rdatasize;
+
+               irh->rflag =  1;
+               irh->len   =  4;
+               ih->fsz    = 40; /* irh+ossp[0]+ossp[1]+rdp+rptr = 40 bytes */
+       } else {
+               irh->rflag =  0;
+               irh->len   =  2;
+               ih->fsz    = 24; /* irh + ossp[0] + ossp[1] = 24 bytes */
+       }
+
+       while (!(oct->io_qmask.iq & (1 << sc->iq_no)))
+               sc->iq_no++;
+}
+
+int octeon_send_soft_command(struct octeon_device *oct,
+                            struct octeon_soft_command *sc)
+{
+       struct octeon_instr_ih *ih;
+       struct octeon_instr_irh *irh;
+       struct octeon_instr_rdp *rdp;
+
+       ih = (struct octeon_instr_ih *)&sc->cmd.ih;
+       if (ih->dlengsz) {
+               BUG_ON(!sc->dmadptr);
+               sc->cmd.dptr = sc->dmadptr;
+       }
+
+       irh = (struct octeon_instr_irh *)&sc->cmd.irh;
+       if (irh->rflag) {
+               BUG_ON(!sc->dmarptr);
+               BUG_ON(!sc->status_word);
+               *sc->status_word = COMPLETION_WORD_INIT;
+
+               rdp = (struct octeon_instr_rdp *)&sc->cmd.rdp;
+
+               sc->cmd.rptr = sc->dmarptr;
+       }
+
+       if (sc->wait_time)
+               sc->timeout = jiffies + sc->wait_time;
+
+       return octeon_send_command(oct, sc->iq_no, 1, &sc->cmd, sc,
+                                  (u32)ih->dlengsz, REQTYPE_SOFT_COMMAND);
+}
+
+int octeon_setup_sc_buffer_pool(struct octeon_device *oct)
+{
+       int i;
+       u64 dma_addr;
+       struct octeon_soft_command *sc;
+
+       INIT_LIST_HEAD(&oct->sc_buf_pool.head);
+       spin_lock_init(&oct->sc_buf_pool.lock);
+       atomic_set(&oct->sc_buf_pool.alloc_buf_count, 0);
+
+       for (i = 0; i < MAX_SOFT_COMMAND_BUFFERS; i++) {
+               sc = (struct octeon_soft_command *)
+                       lio_dma_alloc(oct,
+                                     SOFT_COMMAND_BUFFER_SIZE,
+                                         (dma_addr_t *)&dma_addr);
+               if (!sc)
+                       return 1;
+
+               sc->dma_addr = dma_addr;
+               sc->size = SOFT_COMMAND_BUFFER_SIZE;
+
+               list_add_tail(&sc->node, &oct->sc_buf_pool.head);
+       }
+
+       return 0;
+}
+
+int octeon_free_sc_buffer_pool(struct octeon_device *oct)
+{
+       struct list_head *tmp, *tmp2;
+       struct octeon_soft_command *sc;
+
+       spin_lock(&oct->sc_buf_pool.lock);
+
+       list_for_each_safe(tmp, tmp2, &oct->sc_buf_pool.head) {
+               list_del(tmp);
+
+               sc = (struct octeon_soft_command *)tmp;
+
+               lio_dma_free(oct, sc->size, sc, sc->dma_addr);
+       }
+
+       INIT_LIST_HEAD(&oct->sc_buf_pool.head);
+
+       spin_unlock(&oct->sc_buf_pool.lock);
+
+       return 0;
+}
+
+struct octeon_soft_command *octeon_alloc_soft_command(struct octeon_device *oct,
+                                                     u32 datasize,
+                                                     u32 rdatasize,
+                                                     u32 ctxsize)
+{
+       u64 dma_addr;
+       u32 size;
+       u32 offset = sizeof(struct octeon_soft_command);
+       struct octeon_soft_command *sc = NULL;
+       struct list_head *tmp;
+
+       BUG_ON((offset + datasize + rdatasize + ctxsize) >
+              SOFT_COMMAND_BUFFER_SIZE);
+
+       spin_lock(&oct->sc_buf_pool.lock);
+
+       if (list_empty(&oct->sc_buf_pool.head)) {
+               spin_unlock(&oct->sc_buf_pool.lock);
+               return NULL;
+       }
+
+       list_for_each(tmp, &oct->sc_buf_pool.head)
+               break;
+
+       list_del(tmp);
+
+       atomic_inc(&oct->sc_buf_pool.alloc_buf_count);
+
+       spin_unlock(&oct->sc_buf_pool.lock);
+
+       sc = (struct octeon_soft_command *)tmp;
+
+       dma_addr = sc->dma_addr;
+       size = sc->size;
+
+       memset(sc, 0, sc->size);
+
+       sc->dma_addr = dma_addr;
+       sc->size = size;
+
+       if (ctxsize) {
+               sc->ctxptr = (u8 *)sc + offset;
+               sc->ctxsize = ctxsize;
+       }
+
+       /* Start data at 128 byte boundary */
+       offset = (offset + ctxsize + 127) & 0xffffff80;
+
+       if (datasize) {
+               sc->virtdptr = (u8 *)sc + offset;
+               sc->dmadptr = dma_addr + offset;
+               sc->datasize = datasize;
+       }
+
+       /* Start rdata at 128 byte boundary */
+       offset = (offset + datasize + 127) & 0xffffff80;
+
+       if (rdatasize) {
+               BUG_ON(rdatasize < 16);
+               sc->virtrptr = (u8 *)sc + offset;
+               sc->dmarptr = dma_addr + offset;
+               sc->rdatasize = rdatasize;
+               sc->status_word = (u64 *)((u8 *)(sc->virtrptr) + rdatasize - 8);
+       }
+
+       return sc;
+}
+
+void octeon_free_soft_command(struct octeon_device *oct,
+                             struct octeon_soft_command *sc)
+{
+       spin_lock(&oct->sc_buf_pool.lock);
+
+       list_add_tail(&sc->node, &oct->sc_buf_pool.head);
+
+       atomic_dec(&oct->sc_buf_pool.alloc_buf_count);
+
+       spin_unlock(&oct->sc_buf_pool.lock);
+}