diff options
Diffstat (limited to 'drivers/staging/rdma/hfi1/sdma.c')
| -rw-r--r-- | drivers/staging/rdma/hfi1/sdma.c | 2962 |
1 files changed, 2962 insertions, 0 deletions
diff --git a/drivers/staging/rdma/hfi1/sdma.c b/drivers/staging/rdma/hfi1/sdma.c new file mode 100644 index 000000000000..37bd767d6bc0 --- /dev/null +++ b/drivers/staging/rdma/hfi1/sdma.c @@ -0,0 +1,2962 @@ +/* + * + * This file is provided under a dual BSD/GPLv2 license. When using or + * redistributing this file, you may do so under either license. + * + * GPL LICENSE SUMMARY + * + * Copyright(c) 2015 Intel Corporation. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License for more details. + * + * BSD LICENSE + * + * Copyright(c) 2015 Intel Corporation. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in + * the documentation and/or other materials provided with the + * distribution. + * - Neither the name of Intel Corporation nor the names of its + * contributors may be used to endorse or promote products derived + * from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#include <linux/spinlock.h> +#include <linux/seqlock.h> +#include <linux/netdevice.h> +#include <linux/moduleparam.h> +#include <linux/bitops.h> +#include <linux/timer.h> +#include <linux/vmalloc.h> + +#include "hfi.h" +#include "common.h" +#include "qp.h" +#include "sdma.h" +#include "iowait.h" +#include "trace.h" + +/* must be a power of 2 >= 64 <= 32768 */ +#define SDMA_DESCQ_CNT 1024 +#define INVALID_TAIL 0xffff + +static uint sdma_descq_cnt = SDMA_DESCQ_CNT; +module_param(sdma_descq_cnt, uint, S_IRUGO); +MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries"); + +static uint sdma_idle_cnt = 250; +module_param(sdma_idle_cnt, uint, S_IRUGO); +MODULE_PARM_DESC(sdma_idle_cnt, "sdma interrupt idle delay (ns,default 250)"); + +uint mod_num_sdma; +module_param_named(num_sdma, mod_num_sdma, uint, S_IRUGO); +MODULE_PARM_DESC(num_sdma, "Set max number SDMA engines to use"); + +#define SDMA_WAIT_BATCH_SIZE 20 +/* max wait time for a SDMA engine to indicate it has halted */ +#define SDMA_ERR_HALT_TIMEOUT 10 /* ms */ +/* all SDMA engine errors that cause a halt */ + +#define SD(name) SEND_DMA_##name +#define ALL_SDMA_ENG_HALT_ERRS \ + (SD(ENG_ERR_STATUS_SDMA_WRONG_DW_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_GEN_MISMATCH_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_TOO_LONG_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_TAIL_OUT_OF_BOUNDS_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_FIRST_DESC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_MEM_READ_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_LENGTH_MISMATCH_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_PACKET_DESC_OVERFLOW_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_SELECT_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_ADDRESS_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_LENGTH_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_TIMEOUT_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_DESC_TABLE_UNC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_ASSEMBLY_UNC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_PACKET_TRACKING_UNC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_STORAGE_UNC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SMASK)) + +/* sdma_sendctrl operations */ +#define SDMA_SENDCTRL_OP_ENABLE (1U << 0) +#define SDMA_SENDCTRL_OP_INTENABLE (1U << 1) +#define SDMA_SENDCTRL_OP_HALT (1U << 2) +#define SDMA_SENDCTRL_OP_CLEANUP (1U << 3) + +/* handle long defines */ +#define SDMA_EGRESS_PACKET_OCCUPANCY_SMASK \ +SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SMASK +#define SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT \ +SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT + +static const char * const sdma_state_names[] = { + [sdma_state_s00_hw_down] = "s00_HwDown", + [sdma_state_s10_hw_start_up_halt_wait] = "s10_HwStartUpHaltWait", + [sdma_state_s15_hw_start_up_clean_wait] = "s15_HwStartUpCleanWait", + [sdma_state_s20_idle] = "s20_Idle", + [sdma_state_s30_sw_clean_up_wait] = "s30_SwCleanUpWait", + [sdma_state_s40_hw_clean_up_wait] = "s40_HwCleanUpWait", + [sdma_state_s50_hw_halt_wait] = "s50_HwHaltWait", + [sdma_state_s60_idle_halt_wait] = "s60_IdleHaltWait", + [sdma_state_s80_hw_freeze] = "s80_HwFreeze", + [sdma_state_s82_freeze_sw_clean] = "s82_FreezeSwClean", + [sdma_state_s99_running] = "s99_Running", +}; + +static const char * const sdma_event_names[] = { + [sdma_event_e00_go_hw_down] = "e00_GoHwDown", + [sdma_event_e10_go_hw_start] = "e10_GoHwStart", + [sdma_event_e15_hw_halt_done] = "e15_HwHaltDone", + [sdma_event_e25_hw_clean_up_done] = "e25_HwCleanUpDone", + [sdma_event_e30_go_running] = "e30_GoRunning", + [sdma_event_e40_sw_cleaned] = "e40_SwCleaned", + [sdma_event_e50_hw_cleaned] = "e50_HwCleaned", + [sdma_event_e60_hw_halted] = "e60_HwHalted", + [sdma_event_e70_go_idle] = "e70_GoIdle", + [sdma_event_e80_hw_freeze] = "e80_HwFreeze", + [sdma_event_e81_hw_frozen] = "e81_HwFrozen", + [sdma_event_e82_hw_unfreeze] = "e82_HwUnfreeze", + [sdma_event_e85_link_down] = "e85_LinkDown", + [sdma_event_e90_sw_halted] = "e90_SwHalted", +}; + +static const struct sdma_set_state_action sdma_action_table[] = { + [sdma_state_s00_hw_down] = { + .go_s99_running_tofalse = 1, + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s10_hw_start_up_halt_wait] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 1, + .op_cleanup = 0, + }, + [sdma_state_s15_hw_start_up_clean_wait] = { + .op_enable = 0, + .op_intenable = 1, + .op_halt = 0, + .op_cleanup = 1, + }, + [sdma_state_s20_idle] = { + .op_enable = 0, + .op_intenable = 1, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s30_sw_clean_up_wait] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s40_hw_clean_up_wait] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 1, + }, + [sdma_state_s50_hw_halt_wait] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s60_idle_halt_wait] = { + .go_s99_running_tofalse = 1, + .op_enable = 0, + .op_intenable = 0, + .op_halt = 1, + .op_cleanup = 0, + }, + [sdma_state_s80_hw_freeze] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s82_freeze_sw_clean] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s99_running] = { + .op_enable = 1, + .op_intenable = 1, + .op_halt = 0, + .op_cleanup = 0, + .go_s99_running_totrue = 1, + }, +}; + +#define SDMA_TAIL_UPDATE_THRESH 0x1F + +/* declare all statics here rather than keep sorting */ +static void sdma_complete(struct kref *); +static void sdma_finalput(struct sdma_state *); +static void sdma_get(struct sdma_state *); +static void sdma_hw_clean_up_task(unsigned long); +static void sdma_put(struct sdma_state *); +static void sdma_set_state(struct sdma_engine *, enum sdma_states); +static void sdma_start_hw_clean_up(struct sdma_engine *); +static void sdma_start_sw_clean_up(struct sdma_engine *); +static void sdma_sw_clean_up_task(unsigned long); +static void sdma_sendctrl(struct sdma_engine *, unsigned); +static void init_sdma_regs(struct sdma_engine *, u32, uint); +static void sdma_process_event( + struct sdma_engine *sde, + enum sdma_events event); +static void __sdma_process_event( + struct sdma_engine *sde, + enum sdma_events event); +static void dump_sdma_state(struct sdma_engine *sde); +static void sdma_make_progress(struct sdma_engine *sde, u64 status); +static void sdma_desc_avail(struct sdma_engine *sde, unsigned avail); +static void sdma_flush_descq(struct sdma_engine *sde); + +/** + * sdma_state_name() - return state string from enum + * @state: state + */ +static const char *sdma_state_name(enum sdma_states state) +{ + return sdma_state_names[state]; +} + +static void sdma_get(struct sdma_state *ss) +{ + kref_get(&ss->kref); +} + +static void sdma_complete(struct kref *kref) +{ + struct sdma_state *ss = + container_of(kref, struct sdma_state, kref); + + complete(&ss->comp); +} + +static void sdma_put(struct sdma_state *ss) +{ + kref_put(&ss->kref, sdma_complete); +} + +static void sdma_finalput(struct sdma_state *ss) +{ + sdma_put(ss); + wait_for_completion(&ss->comp); +} + +static inline void write_sde_csr( + struct sdma_engine *sde, + u32 offset0, + u64 value) +{ + write_kctxt_csr(sde->dd, sde->this_idx, offset0, value); +} + +static inline u64 read_sde_csr( + struct sdma_engine *sde, + u32 offset0) +{ + return read_kctxt_csr(sde->dd, sde->this_idx, offset0); +} + +/* + * sdma_wait_for_packet_egress() - wait for the VL FIFO occupancy for + * sdma engine 'sde' to drop to 0. + */ +static void sdma_wait_for_packet_egress(struct sdma_engine *sde, + int pause) +{ + u64 off = 8 * sde->this_idx; + struct hfi1_devdata *dd = sde->dd; + int lcnt = 0; + + while (1) { + u64 reg = read_csr(dd, off + SEND_EGRESS_SEND_DMA_STATUS); + + reg &= SDMA_EGRESS_PACKET_OCCUPANCY_SMASK; + reg >>= SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT; + if (reg == 0) + break; + if (lcnt++ > 100) { + dd_dev_err(dd, "%s: engine %u timeout waiting for packets to egress, remaining count %u\n", + __func__, sde->this_idx, (u32)reg); + break; + } + udelay(1); + } +} + +/* + * sdma_wait() - wait for packet egress to complete for all SDMA engines, + * and pause for credit return. + */ +void sdma_wait(struct hfi1_devdata *dd) +{ + int i; + + for (i = 0; i < dd->num_sdma; i++) { + struct sdma_engine *sde = &dd->per_sdma[i]; + + sdma_wait_for_packet_egress(sde, 0); + } +} + +static inline void sdma_set_desc_cnt(struct sdma_engine *sde, unsigned cnt) +{ + u64 reg; + + if (!(sde->dd->flags & HFI1_HAS_SDMA_TIMEOUT)) + return; + reg = cnt; + reg &= SD(DESC_CNT_CNT_MASK); + reg <<= SD(DESC_CNT_CNT_SHIFT); + write_sde_csr(sde, SD(DESC_CNT), reg); +} + +/* + * Complete all the sdma requests with a SDMA_TXREQ_S_ABORTED status + * + * Depending on timing there can be txreqs in two places: + * - in the descq ring + * - in the flush list + * + * To avoid ordering issues the descq ring needs to be flushed + * first followed by the flush list. + * + * This routine is called from two places + * - From a work queue item + * - Directly from the state machine just before setting the + * state to running + * + * Must be called with head_lock held + * + */ +static void sdma_flush(struct sdma_engine *sde) +{ + struct sdma_txreq *txp, *txp_next; + LIST_HEAD(flushlist); + + /* flush from head to tail */ + sdma_flush_descq(sde); + spin_lock(&sde->flushlist_lock); + /* copy flush list */ + list_for_each_entry_safe(txp, txp_next, &sde->flushlist, list) { + list_del_init(&txp->list); + list_add_tail(&txp->list, &flushlist); + } + spin_unlock(&sde->flushlist_lock); + /* flush from flush list */ + list_for_each_entry_safe(txp, txp_next, &flushlist, list) { + int drained = 0; + /* protect against complete modifying */ + struct iowait *wait = txp->wait; + + list_del_init(&txp->list); +#ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER + trace_hfi1_sdma_out_sn(sde, txp->sn); + if (WARN_ON_ONCE(sde->head_sn != txp->sn)) + dd_dev_err(sde->dd, "expected %llu got %llu\n", + sde->head_sn, txp->sn); + sde->head_sn++; +#endif + sdma_txclean(sde->dd, txp); + if (wait) + drained = atomic_dec_and_test(&wait->sdma_busy); + if (txp->complete) + (*txp->complete)(txp, SDMA_TXREQ_S_ABORTED, drained); + if (wait && drained) + iowait_drain_wakeup(wait); + } +} + +/* + * Fields a work request for flushing the descq ring + * and the flush list + * + * If the engine has been brought to running during + * the scheduling delay, the flush is ignored, assuming + * that the process of bringing the engine to running + * would have done this flush prior to going to running. + * + */ +static void sdma_field_flush(struct work_struct *work) +{ + unsigned long flags; + struct sdma_engine *sde = + container_of(work, struct sdma_engine, flush_worker); + + write_seqlock_irqsave(&sde->head_lock, flags); + if (!__sdma_running(sde)) + sdma_flush(sde); + write_sequnlock_irqrestore(&sde->head_lock, flags); +} + +static void sdma_err_halt_wait(struct work_struct *work) +{ + struct sdma_engine *sde = container_of(work, struct sdma_engine, + err_halt_worker); + u64 statuscsr; + unsigned long timeout; + + timeout = jiffies + msecs_to_jiffies(SDMA_ERR_HALT_TIMEOUT); + while (1) { + statuscsr = read_sde_csr(sde, SD(STATUS)); + statuscsr &= SD(STATUS_ENG_HALTED_SMASK); + if (statuscsr) + break; + if (time_after(jiffies, timeout)) { + dd_dev_err(sde->dd, + "SDMA engine %d - timeout waiting for engine to halt\n", + sde->this_idx); + /* + * Continue anyway. This could happen if there was + * an uncorrectable error in the wrong spot. + */ + break; + } + usleep_range(80, 120); + } + + sdma_process_event(sde, sdma_event_e15_hw_halt_done); +} + +static void sdma_start_err_halt_wait(struct sdma_engine *sde) +{ + schedule_work(&sde->err_halt_worker); +} + + +static void sdma_err_progress_check_schedule(struct sdma_engine *sde) +{ + if (!is_bx(sde->dd) && HFI1_CAP_IS_KSET(SDMA_AHG)) { + + unsigned index; + struct hfi1_devdata *dd = sde->dd; + + for (index = 0; index < dd->num_sdma; index++) { + struct sdma_engine *curr_sdma = &dd->per_sdma[index]; + + if (curr_sdma != sde) + curr_sdma->progress_check_head = + curr_sdma->descq_head; + } + dd_dev_err(sde->dd, + "SDMA engine %d - check scheduled\n", + sde->this_idx); + mod_timer(&sde->err_progress_check_timer, jiffies + 10); + } +} + +static void sdma_err_progress_check(unsigned long data) +{ + unsigned index; + struct sdma_engine *sde = (struct sdma_engine *)data; + + dd_dev_err(sde->dd, "SDE progress check event\n"); + for (index = 0; index < sde->dd->num_sdma; index++) { + struct sdma_engine *curr_sde = &sde->dd->per_sdma[index]; + unsigned long flags; + + /* check progress on each engine except the current one */ + if (curr_sde == sde) + continue; + /* + * We must lock interrupts when acquiring sde->lock, + * to avoid a deadlock if interrupt triggers and spins on + * the same lock on same CPU + */ + spin_lock_irqsave(&curr_sde->tail_lock, flags); + write_seqlock(&curr_sde->head_lock); + + /* skip non-running queues */ + if (curr_sde->state.current_state != sdma_state_s99_running) { + write_sequnlock(&curr_sde->head_lock); + spin_unlock_irqrestore(&curr_sde->tail_lock, flags); + continue; + } + + if ((curr_sde->descq_head != curr_sde->descq_tail) && + (curr_sde->descq_head == + curr_sde->progress_check_head)) + __sdma_process_event(curr_sde, + sdma_event_e90_sw_halted); + write_sequnlock(&curr_sde->head_lock); + spin_unlock_irqrestore(&curr_sde->tail_lock, flags); + } + schedule_work(&sde->err_halt_worker); +} + +static void sdma_hw_clean_up_task(unsigned long opaque) +{ + struct sdma_engine *sde = (struct sdma_engine *) opaque; + u64 statuscsr; + + while (1) { +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", + sde->this_idx, slashstrip(__FILE__), __LINE__, + __func__); +#endif + statuscsr = read_sde_csr(sde, SD(STATUS)); + statuscsr &= SD(STATUS_ENG_CLEANED_UP_SMASK); + if (statuscsr) + break; + udelay(10); + } + + sdma_process_event(sde, sdma_event_e25_hw_clean_up_done); +} + +static inline struct sdma_txreq *get_txhead(struct sdma_engine *sde) +{ + smp_read_barrier_depends(); /* see sdma_update_tail() */ + return sde->tx_ring[sde->tx_head & sde->sdma_mask]; +} + +/* + * flush ring for recovery + */ +static void sdma_flush_descq(struct sdma_engine *sde) +{ + u16 head, tail; + int progress = 0; + struct sdma_txreq *txp = get_txhead(sde); + + /* The reason for some of the complexity of this code is that + * not all descriptors have corresponding txps. So, we have to + * be able to skip over descs until we wander into the range of + * the next txp on the list. + */ + head = sde->descq_head & sde->sdma_mask; + tail = sde->descq_tail & sde->sdma_mask; + while (head != tail) { + /* advance head, wrap if needed */ + head = ++sde->descq_head & sde->sdma_mask; + /* if now past this txp's descs, do the callback */ + if (txp && txp->next_descq_idx == head) { + int drained = 0; + /* protect against complete modifying */ + struct iowait *wait = txp->wait; + + /* remove from list */ + sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL; + if (wait) + drained = atomic_dec_and_test(&wait->sdma_busy); +#ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER + trace_hfi1_sdma_out_sn(sde, txp->sn); + if (WARN_ON_ONCE(sde->head_sn != txp->sn)) + dd_dev_err(sde->dd, "expected %llu got %llu\n", + sde->head_sn, txp->sn); + sde->head_sn++; +#endif + sdma_txclean(sde->dd, txp); + trace_hfi1_sdma_progress(sde, head, tail, txp); + if (txp->complete) + (*txp->complete)( + txp, + SDMA_TXREQ_S_ABORTED, + drained); + if (wait && drained) + iowait_drain_wakeup(wait); + /* see if there is another txp */ + txp = get_txhead(sde); + } + progress++; + } + if (progress) + sdma_desc_avail(sde, sdma_descq_freecnt(sde)); +} + +static void sdma_sw_clean_up_task(unsigned long opaque) +{ + struct sdma_engine *sde = (struct sdma_engine *) opaque; + unsigned long flags; + + spin_lock_irqsave(&sde->tail_lock, flags); + write_seqlock(&sde->head_lock); + + /* + * At this point, the following should always be true: + * - We are halted, so no more descriptors are getting retired. + * - We are not running, so no one is submitting new work. + * - Only we can send the e40_sw_cleaned, so we can't start + * running again until we say so. So, the active list and + * descq are ours to play with. + */ + + + /* + * In the error clean up sequence, software clean must be called + * before the hardware clean so we can use the hardware head in + * the progress routine. A hardware clean or SPC unfreeze will + * reset the hardware head. + * + * Process all retired requests. The progress routine will use the + * latest physical hardware head - we are not running so speed does + * not matter. + */ + sdma_make_progress(sde, 0); + + sdma_flush(sde); + + /* + * Reset our notion of head and tail. + * Note that the HW registers have been reset via an earlier + * clean up. + */ + sde->descq_tail = 0; + sde->descq_head = 0; + sde->desc_avail = sdma_descq_freecnt(sde); + *sde->head_dma = 0; + + __sdma_process_event(sde, sdma_event_e40_sw_cleaned); + + write_sequnlock(&sde->head_lock); + spin_unlock_irqrestore(&sde->tail_lock, flags); +} + +static void sdma_sw_tear_down(struct sdma_engine *sde) +{ + struct sdma_state *ss = &sde->state; + + /* Releasing this reference means the state machine has stopped. */ + sdma_put(ss); + + /* stop waiting for all unfreeze events to complete */ + atomic_set(&sde->dd->sdma_unfreeze_count, -1); + wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); +} + +static void sdma_start_hw_clean_up(struct sdma_engine *sde) +{ + tasklet_hi_schedule(&sde->sdma_hw_clean_up_task); +} + +static void sdma_start_sw_clean_up(struct sdma_engine *sde) +{ + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); +} + +static void sdma_set_state(struct sdma_engine *sde, + enum sdma_states next_state) +{ + struct sdma_state *ss = &sde->state; + const struct sdma_set_state_action *action = sdma_action_table; + unsigned op = 0; + + trace_hfi1_sdma_state( + sde, + sdma_state_names[ss->current_state], + sdma_state_names[next_state]); + + /* debugging bookkeeping */ + ss->previous_state = ss->current_state; + ss->previous_op = ss->current_op; + ss->current_state = next_state; + + if (ss->previous_state != sdma_state_s99_running + && next_state == sdma_state_s99_running) + sdma_flush(sde); + + if (action[next_state].op_enable) + op |= SDMA_SENDCTRL_OP_ENABLE; + + if (action[next_state].op_intenable) + op |= SDMA_SENDCTRL_OP_INTENABLE; + + if (action[next_state].op_halt) + op |= SDMA_SENDCTRL_OP_HALT; + + if (action[next_state].op_cleanup) + op |= SDMA_SENDCTRL_OP_CLEANUP; + + if (action[next_state].go_s99_running_tofalse) + ss->go_s99_running = 0; + + if (action[next_state].go_s99_running_totrue) + ss->go_s99_running = 1; + + ss->current_op = op; + sdma_sendctrl(sde, ss->current_op); +} + +/** + * sdma_get_descq_cnt() - called when device probed + * + * Return a validated descq count. + * + * This is currently only used in the verbs initialization to build the tx + * list. + * + * This will probably be deleted in favor of a more scalable approach to + * alloc tx's. + * + */ +u16 sdma_get_descq_cnt(void) +{ + u16 count = sdma_descq_cnt; + + if (!count) + return SDMA_DESCQ_CNT; + /* count must be a power of 2 greater than 64 and less than + * 32768. Otherwise return default. + */ + if (!is_power_of_2(count)) + return SDMA_DESCQ_CNT; + if (count < 64 && count > 32768) + return SDMA_DESCQ_CNT; + return count; +} +/** + * sdma_select_engine_vl() - select sdma engine + * @dd: devdata + * @selector: a spreading factor + * @vl: this vl + * + * + * This function returns an engine based on the selector and a vl. The + * mapping fields are protected by RCU. + */ +struct sdma_engine *sdma_select_engine_vl( + struct hfi1_devdata *dd, + u32 selector, + u8 vl) +{ + struct sdma_vl_map *m; + struct sdma_map_elem *e; + struct sdma_engine *rval; + + if (WARN_ON(vl > 8)) + return NULL; + + rcu_read_lock(); + m = rcu_dereference(dd->sdma_map); + if (unlikely(!m)) { + rcu_read_unlock(); + return NULL; + } + e = m->map[vl & m->mask]; + rval = e->sde[selector & e->mask]; + rcu_read_unlock(); + + trace_hfi1_sdma_engine_select(dd, selector, vl, rval->this_idx); + return rval; +} + +/** + * sdma_select_engine_sc() - select sdma engine + * @dd: devdata + * @selector: a spreading factor + * @sc5: the 5 bit sc + * + * + * This function returns an engine based on the selector and an sc. + */ +struct sdma_engine *sdma_select_engine_sc( + struct hfi1_devdata *dd, + u32 selector, + u8 sc5) +{ + u8 vl = sc_to_vlt(dd, sc5); + + return sdma_select_engine_vl(dd, selector, vl); +} + +/* + * Free the indicated map struct + */ +static void sdma_map_free(struct sdma_vl_map *m) +{ + int i; + + for (i = 0; m && i < m->actual_vls; i++) + kfree(m->map[i]); + kfree(m); +} + +/* + * Handle RCU callback + */ +static void sdma_map_rcu_callback(struct rcu_head *list) +{ + struct sdma_vl_map *m = container_of(list, struct sdma_vl_map, list); + + sdma_map_free(m); +} + +/** + * sdma_map_init - called when # vls change + * @dd: hfi1_devdata + * @port: port number + * @num_vls: number of vls + * @vl_engines: per vl engine mapping (optional) + * + * This routine changes the mapping based on the number of vls. + * + * vl_engines is used to specify a non-uniform vl/engine loading. NULL + * implies auto computing the loading and giving each VLs a uniform + * distribution of engines per VL. + * + * The auto algorithm computes the sde_per_vl and the number of extra + * engines. Any extra engines are added from the last VL on down. + * + * rcu locking is used here to control access to the mapping fields. + * + * If either the num_vls or num_sdma are non-power of 2, the array sizes + * in the struct sdma_vl_map and the struct sdma_map_elem are rounded + * up to the next highest power of 2 and the first entry is reused + * in a round robin fashion. + * + * If an error occurs the map change is not done and the mapping is + * not changed. + * + */ +int sdma_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_engines) +{ + int i, j; + int extra, sde_per_vl; + int engine = 0; + u8 lvl_engines[OPA_MAX_VLS]; + struct sdma_vl_map *oldmap, *newmap; + + if (!(dd->flags & HFI1_HAS_SEND_DMA)) + return 0; + + if (!vl_engines) { + /* truncate divide */ + sde_per_vl = dd->num_sdma / num_vls; + /* extras */ + extra = dd->num_sdma % num_vls; + vl_engines = lvl_engines; + /* add extras from last vl down */ + for (i = num_vls - 1; i >= 0; i--, extra--) + vl_engines[i] = sde_per_vl + (extra > 0 ? 1 : 0); + } + /* build new map */ + newmap = kzalloc( + sizeof(struct sdma_vl_map) + + roundup_pow_of_two(num_vls) * + sizeof(struct sdma_map_elem *), + GFP_KERNEL); + if (!newmap) + goto bail; + newmap->actual_vls = num_vls; + newmap->vls = roundup_pow_of_two(num_vls); + newmap->mask = (1 << ilog2(newmap->vls)) - 1; + for (i = 0; i < newmap->vls; i++) { + /* save for wrap around */ + int first_engine = engine; + + if (i < newmap->actual_vls) { + int sz = roundup_pow_of_two(vl_engines[i]); + + /* only allocate once */ + newmap->map[i] = kzalloc( + sizeof(struct sdma_map_elem) + + sz * sizeof(struct sdma_engine *), + GFP_KERNEL); + if (!newmap->map[i]) + goto bail; + newmap->map[i]->mask = (1 << ilog2(sz)) - 1; + /* assign engines */ + for (j = 0; j < sz; j++) { + newmap->map[i]->sde[j] = + &dd->per_sdma[engine]; + if (++engine >= first_engine + vl_engines[i]) + /* wrap back to first engine */ + engine = first_engine; + } + } else { + /* just re-use entry without allocating */ + newmap->map[i] = newmap->map[i % num_vls]; + } + engine = first_engine + vl_engines[i]; + } + /* newmap in hand, save old map */ + spin_lock_irq(&dd->sde_map_lock); + oldmap = rcu_dereference_protected(dd->sdma_map, + lockdep_is_held(&dd->sde_map_lock)); + + /* publish newmap */ + rcu_assign_pointer(dd->sdma_map, newmap); + + spin_unlock_irq(&dd->sde_map_lock); + /* success, free any old map after grace period */ + if (oldmap) + call_rcu(&oldmap->list, sdma_map_rcu_callback); + return 0; +bail: + /* free any partial allocation */ + sdma_map_free(newmap); + return -ENOMEM; +} + +/* + * Clean up allocated memory. + * + * This routine is can be called regardless of the success of sdma_init() + * + */ +static void sdma_clean(struct hfi1_devdata *dd, size_t num_engines) +{ + size_t i; + struct sdma_engine *sde; + + if (dd->sdma_pad_dma) { + dma_free_coherent(&dd->pcidev->dev, 4, + (void *)dd->sdma_pad_dma, + dd->sdma_pad_phys); + dd->sdma_pad_dma = NULL; + dd->sdma_pad_phys = 0; + } + if (dd->sdma_heads_dma) { + dma_free_coherent(&dd->pcidev->dev, dd->sdma_heads_size, + (void *)dd->sdma_heads_dma, + dd->sdma_heads_phys); + dd->sdma_heads_dma = NULL; + dd->sdma_heads_phys = 0; + } + for (i = 0; dd->per_sdma && i < num_engines; ++i) { + sde = &dd->per_sdma[i]; + + sde->head_dma = NULL; + sde->head_phys = 0; + + if (sde->descq) { + dma_free_coherent( + &dd->pcidev->dev, + sde->descq_cnt * sizeof(u64[2]), + sde->descq, + sde->descq_phys + ); + sde->descq = NULL; + sde->descq_phys = 0; + } + if (is_vmalloc_addr(sde->tx_ring)) + vfree(sde->tx_ring); + else + kfree(sde->tx_ring); + sde->tx_ring = NULL; + } + spin_lock_irq(&dd->sde_map_lock); + kfree(rcu_access_pointer(dd->sdma_map)); + RCU_INIT_POINTER(dd->sdma_map, NULL); + spin_unlock_irq(&dd->sde_map_lock); + synchronize_rcu(); + kfree(dd->per_sdma); + dd->per_sdma = NULL; +} + +/** + * sdma_init() - called when device probed + * @dd: hfi1_devdata + * @port: port number (currently only zero) + * + * sdma_init initializes the specified number of engines. + * + * The code initializes each sde, its csrs. Interrupts + * are not required to be enabled. + * + * Returns: + * 0 - success, -errno on failure + */ +int sdma_init(struct hfi1_devdata *dd, u8 port) +{ + unsigned this_idx; + struct sdma_engine *sde; + u16 descq_cnt; + void *curr_head; + struct hfi1_pportdata *ppd = dd->pport + port; + u32 per_sdma_credits; + uint idle_cnt = sdma_idle_cnt; + size_t num_engines = dd->chip_sdma_engines; + + if (!HFI1_CAP_IS_KSET(SDMA)) { + HFI1_CAP_CLEAR(SDMA_AHG); + return 0; + } + if (mod_num_sdma && + /* can't exceed chip support */ + mod_num_sdma <= dd->chip_sdma_engines && + /* count must be >= vls */ + mod_num_sdma >= num_vls) + num_engines = mod_num_sdma; + + dd_dev_info(dd, "SDMA mod_num_sdma: %u\n", mod_num_sdma); + dd_dev_info(dd, "SDMA chip_sdma_engines: %u\n", dd->chip_sdma_engines); + dd_dev_info(dd, "SDMA chip_sdma_mem_size: %u\n", + dd->chip_sdma_mem_size); + + per_sdma_credits = + dd->chip_sdma_mem_size/(num_engines * SDMA_BLOCK_SIZE); + + /* set up freeze waitqueue */ + init_waitqueue_head(&dd->sdma_unfreeze_wq); + atomic_set(&dd->sdma_unfreeze_count, 0); + + descq_cnt = sdma_get_descq_cnt(); + dd_dev_info(dd, "SDMA engines %zu descq_cnt %u\n", + num_engines, descq_cnt); + + /* alloc memory for array of send engines */ + dd->per_sdma = kcalloc(num_engines, sizeof(*dd->per_sdma), GFP_KERNEL); + if (!dd->per_sdma) + return -ENOMEM; + + idle_cnt = ns_to_cclock(dd, idle_cnt); + /* Allocate memory for SendDMA descriptor FIFOs */ + for (this_idx = 0; this_idx < num_engines; ++this_idx) { + sde = &dd->per_sdma[this_idx]; + sde->dd = dd; + sde->ppd = ppd; + sde->this_idx = this_idx; + sde->descq_cnt = descq_cnt; + sde->desc_avail = sdma_descq_freecnt(sde); + sde->sdma_shift = ilog2(descq_cnt); + sde->sdma_mask = (1 << sde->sdma_shift) - 1; + sde->descq_full_count = 0; + + /* Create a mask for all 3 chip interrupt sources */ + sde->imask = (u64)1 << (0*TXE_NUM_SDMA_ENGINES + this_idx) + | (u64)1 << (1*TXE_NUM_SDMA_ENGINES + this_idx) + | (u64)1 << (2*TXE_NUM_SDMA_ENGINES + this_idx); + /* Create a mask specifically for sdma_idle */ + sde->idle_mask = + (u64)1 << (2*TXE_NUM_SDMA_ENGINES + this_idx); + /* Create a mask specifically for sdma_progress */ + sde->progress_mask = + (u64)1 << (TXE_NUM_SDMA_ENGINES + this_idx); + spin_lock_init(&sde->tail_lock); + seqlock_init(&sde->head_lock); + spin_lock_init(&sde->senddmactrl_lock); + spin_lock_init(&sde->flushlist_lock); + /* insure there is always a zero bit */ + sde->ahg_bits = 0xfffffffe00000000ULL; + + sdma_set_state(sde, sdma_state_s00_hw_down); + + /* set up reference counting */ + kref_init(&sde->state.kref); + init_completion(&sde->state.comp); + + INIT_LIST_HEAD(&sde->flushlist); + INIT_LIST_HEAD(&sde->dmawait); + + sde->tail_csr = + get_kctxt_csr_addr(dd, this_idx, SD(TAIL)); + + if (idle_cnt) + dd->default_desc1 = + SDMA_DESC1_HEAD_TO_HOST_FLAG; + else + dd->default_desc1 = + SDMA_DESC1_INT_REQ_FLAG; + + tasklet_init(&sde->sdma_hw_clean_up_task, sdma_hw_clean_up_task, + (unsigned long)sde); + + tasklet_init(&sde->sdma_sw_clean_up_task, sdma_sw_clean_up_task, + (unsigned long)sde); + INIT_WORK(&sde->err_halt_worker, sdma_err_halt_wait); + INIT_WORK(&sde->flush_worker, sdma_field_flush); + + sde->progress_check_head = 0; + + init_timer(&sde->err_progress_check_timer); + sde->err_progress_check_timer.function = + sdma_err_progress_check; + sde->err_progress_check_timer.data = (unsigned long)sde; + + sde->descq = dma_zalloc_coherent( + &dd->pcidev->dev, + descq_cnt * sizeof(u64[2]), + &sde->descq_phys, + GFP_KERNEL + ); + if (!sde->descq) + goto bail; + sde->tx_ring = + kcalloc(descq_cnt, sizeof(struct sdma_txreq *), + GFP_KERNEL); + if (!sde->tx_ring) + sde->tx_ring = + vzalloc( + sizeof(struct sdma_txreq *) * + descq_cnt); + if (!sde->tx_ring) + goto bail; + } + + dd->sdma_heads_size = L1_CACHE_BYTES * num_engines; + /* Allocate memory for DMA of head registers to memory */ + dd->sdma_heads_dma = dma_zalloc_coherent( + &dd->pcidev->dev, + dd->sdma_heads_size, + &dd->sdma_heads_phys, + GFP_KERNEL + ); + if (!dd->sdma_heads_dma) { + dd_dev_err(dd, "failed to allocate SendDMA head memory\n"); + goto bail; + } + + /* Allocate memory for pad */ + dd->sdma_pad_dma = dma_zalloc_coherent( + &dd->pcidev->dev, + sizeof(u32), + &dd->sdma_pad_phys, + GFP_KERNEL + ); + if (!dd->sdma_pad_dma) { + dd_dev_err(dd, "failed to allocate SendDMA pad memory\n"); + goto bail; + } + + /* assign each engine to different cacheline and init registers */ + curr_head = (void *)dd->sdma_heads_dma; + for (this_idx = 0; this_idx < num_engines; ++this_idx) { + unsigned long phys_offset; + + sde = &dd->per_sdma[this_idx]; + + sde->head_dma = curr_head; |