diff options
author | Mike Marciniszyn <mike.marciniszyn@intel.com> | 2015-07-30 15:17:43 -0400 |
---|---|---|
committer | Doug Ledford <dledford@redhat.com> | 2015-08-28 22:59:36 -0400 |
commit | 7724105686e718ac476a6ad3304fea2fbcfcffde (patch) | |
tree | 743567a9714e95b0f803b74a7278c825e79d529f /drivers/staging/rdma/hfi1/pio.c | |
parent | d4ab347005fb26f414b98b2c8d5ef6de5778c3dc (diff) |
IB/hfi1: add driver files
Signed-off-by: Andrew Friedley <andrew.friedley@intel.com>
Signed-off-by: Arthur Kepner <arthur.kepner@intel.com>
Signed-off-by: Brendan Cunningham <brendan.cunningham@intel.com>
Signed-off-by: Brian Welty <brian.welty@intel.com>
Signed-off-by: Caz Yokoyama <caz.yokoyama@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Jim Snow <jim.m.snow@intel.com>
Signed-off-by: John Gregor <john.a.gregor@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Kevin Pine <kevin.pine@intel.com>
Signed-off-by: Kyle Liddell <kyle.liddell@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Ravi Krishnaswamy <ravi.krishnaswamy@intel.com>
Signed-off-by: Sadanand Warrier <sadanand.warrier@intel.com>
Signed-off-by: Sanath Kumar <sanath.s.kumar@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Vlad Danushevsky <vladimir.danusevsky@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Diffstat (limited to 'drivers/staging/rdma/hfi1/pio.c')
-rw-r--r-- | drivers/staging/rdma/hfi1/pio.c | 1771 |
1 files changed, 1771 insertions, 0 deletions
diff --git a/drivers/staging/rdma/hfi1/pio.c b/drivers/staging/rdma/hfi1/pio.c new file mode 100644 index 000000000000..9991814a8f05 --- /dev/null +++ b/drivers/staging/rdma/hfi1/pio.c @@ -0,0 +1,1771 @@ +/* + * + * 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/delay.h> +#include "hfi.h" +#include "qp.h" +#include "trace.h" + +#define SC_CTXT_PACKET_EGRESS_TIMEOUT 350 /* in chip cycles */ + +#define SC(name) SEND_CTXT_##name +/* + * Send Context functions + */ +static void sc_wait_for_packet_egress(struct send_context *sc, int pause); + +/* + * Set the CM reset bit and wait for it to clear. Use the provided + * sendctrl register. This routine has no locking. + */ +void __cm_reset(struct hfi1_devdata *dd, u64 sendctrl) +{ + write_csr(dd, SEND_CTRL, sendctrl | SEND_CTRL_CM_RESET_SMASK); + while (1) { + udelay(1); + sendctrl = read_csr(dd, SEND_CTRL); + if ((sendctrl & SEND_CTRL_CM_RESET_SMASK) == 0) + break; + } +} + +/* defined in header release 48 and higher */ +#ifndef SEND_CTRL_UNSUPPORTED_VL_SHIFT +#define SEND_CTRL_UNSUPPORTED_VL_SHIFT 3 +#define SEND_CTRL_UNSUPPORTED_VL_MASK 0xffull +#define SEND_CTRL_UNSUPPORTED_VL_SMASK (SEND_CTRL_UNSUPPORTED_VL_MASK \ + << SEND_CTRL_UNSUPPORTED_VL_SHIFT) +#endif + +/* global control of PIO send */ +void pio_send_control(struct hfi1_devdata *dd, int op) +{ + u64 reg, mask; + unsigned long flags; + int write = 1; /* write sendctrl back */ + int flush = 0; /* re-read sendctrl to make sure it is flushed */ + + spin_lock_irqsave(&dd->sendctrl_lock, flags); + + reg = read_csr(dd, SEND_CTRL); + switch (op) { + case PSC_GLOBAL_ENABLE: + reg |= SEND_CTRL_SEND_ENABLE_SMASK; + /* Fall through */ + case PSC_DATA_VL_ENABLE: + /* Disallow sending on VLs not enabled */ + mask = (((~0ull)<<num_vls) & SEND_CTRL_UNSUPPORTED_VL_MASK)<< + SEND_CTRL_UNSUPPORTED_VL_SHIFT; + reg = (reg & ~SEND_CTRL_UNSUPPORTED_VL_SMASK) | mask; + break; + case PSC_GLOBAL_DISABLE: + reg &= ~SEND_CTRL_SEND_ENABLE_SMASK; + break; + case PSC_GLOBAL_VLARB_ENABLE: + reg |= SEND_CTRL_VL_ARBITER_ENABLE_SMASK; + break; + case PSC_GLOBAL_VLARB_DISABLE: + reg &= ~SEND_CTRL_VL_ARBITER_ENABLE_SMASK; + break; + case PSC_CM_RESET: + __cm_reset(dd, reg); + write = 0; /* CSR already written (and flushed) */ + break; + case PSC_DATA_VL_DISABLE: + reg |= SEND_CTRL_UNSUPPORTED_VL_SMASK; + flush = 1; + break; + default: + dd_dev_err(dd, "%s: invalid control %d\n", __func__, op); + break; + } + + if (write) { + write_csr(dd, SEND_CTRL, reg); + if (flush) + (void) read_csr(dd, SEND_CTRL); /* flush write */ + } + + spin_unlock_irqrestore(&dd->sendctrl_lock, flags); +} + +/* number of send context memory pools */ +#define NUM_SC_POOLS 2 + +/* Send Context Size (SCS) wildcards */ +#define SCS_POOL_0 -1 +#define SCS_POOL_1 -2 +/* Send Context Count (SCC) wildcards */ +#define SCC_PER_VL -1 +#define SCC_PER_CPU -2 + +#define SCC_PER_KRCVQ -3 +#define SCC_ACK_CREDITS 32 + +#define PIO_WAIT_BATCH_SIZE 5 + +/* default send context sizes */ +static struct sc_config_sizes sc_config_sizes[SC_MAX] = { + [SC_KERNEL] = { .size = SCS_POOL_0, /* even divide, pool 0 */ + .count = SCC_PER_VL },/* one per NUMA */ + [SC_ACK] = { .size = SCC_ACK_CREDITS, + .count = SCC_PER_KRCVQ }, + [SC_USER] = { .size = SCS_POOL_0, /* even divide, pool 0 */ + .count = SCC_PER_CPU }, /* one per CPU */ + +}; + +/* send context memory pool configuration */ +struct mem_pool_config { + int centipercent; /* % of memory, in 100ths of 1% */ + int absolute_blocks; /* absolute block count */ +}; + +/* default memory pool configuration: 100% in pool 0 */ +static struct mem_pool_config sc_mem_pool_config[NUM_SC_POOLS] = { + /* centi%, abs blocks */ + { 10000, -1 }, /* pool 0 */ + { 0, -1 }, /* pool 1 */ +}; + +/* memory pool information, used when calculating final sizes */ +struct mem_pool_info { + int centipercent; /* 100th of 1% of memory to use, -1 if blocks + already set */ + int count; /* count of contexts in the pool */ + int blocks; /* block size of the pool */ + int size; /* context size, in blocks */ +}; + +/* + * Convert a pool wildcard to a valid pool index. The wildcards + * start at -1 and increase negatively. Map them as: + * -1 => 0 + * -2 => 1 + * etc. + * + * Return -1 on non-wildcard input, otherwise convert to a pool number. + */ +static int wildcard_to_pool(int wc) +{ + if (wc >= 0) + return -1; /* non-wildcard */ + return -wc - 1; +} + +static const char *sc_type_names[SC_MAX] = { + "kernel", + "ack", + "user" +}; + +static const char *sc_type_name(int index) +{ + if (index < 0 || index >= SC_MAX) + return "unknown"; + return sc_type_names[index]; +} + +/* + * Read the send context memory pool configuration and send context + * size configuration. Replace any wildcards and come up with final + * counts and sizes for the send context types. + */ +int init_sc_pools_and_sizes(struct hfi1_devdata *dd) +{ + struct mem_pool_info mem_pool_info[NUM_SC_POOLS] = { { 0 } }; + int total_blocks = (dd->chip_pio_mem_size / PIO_BLOCK_SIZE) - 1; + int total_contexts = 0; + int fixed_blocks; + int pool_blocks; + int used_blocks; + int cp_total; /* centipercent total */ + int ab_total; /* absolute block total */ + int extra; + int i; + + /* + * Step 0: + * - copy the centipercents/absolute sizes from the pool config + * - sanity check these values + * - add up centipercents, then later check for full value + * - add up absolute blocks, then later check for over-commit + */ + cp_total = 0; + ab_total = 0; + for (i = 0; i < NUM_SC_POOLS; i++) { + int cp = sc_mem_pool_config[i].centipercent; + int ab = sc_mem_pool_config[i].absolute_blocks; + + /* + * A negative value is "unused" or "invalid". Both *can* + * be valid, but centipercent wins, so check that first + */ + if (cp >= 0) { /* centipercent valid */ + cp_total += cp; + } else if (ab >= 0) { /* absolute blocks valid */ + ab_total += ab; + } else { /* neither valid */ + dd_dev_err( + dd, + "Send context memory pool %d: both the block count and centipercent are invalid\n", + i); + return -EINVAL; + } + + mem_pool_info[i].centipercent = cp; + mem_pool_info[i].blocks = ab; + } + + /* do not use both % and absolute blocks for different pools */ + if (cp_total != 0 && ab_total != 0) { + dd_dev_err( + dd, + "All send context memory pools must be described as either centipercent or blocks, no mixing between pools\n"); + return -EINVAL; + } + + /* if any percentages are present, they must add up to 100% x 100 */ + if (cp_total != 0 && cp_total != 10000) { + dd_dev_err( + dd, + "Send context memory pool centipercent is %d, expecting 10000\n", + cp_total); + return -EINVAL; + } + + /* the absolute pool total cannot be more than the mem total */ + if (ab_total > total_blocks) { + dd_dev_err( + dd, + "Send context memory pool absolute block count %d is larger than the memory size %d\n", + ab_total, total_blocks); + return -EINVAL; + } + + /* + * Step 2: + * - copy from the context size config + * - replace context type wildcard counts with real values + * - add up non-memory pool block sizes + * - add up memory pool user counts + */ + fixed_blocks = 0; + for (i = 0; i < SC_MAX; i++) { + int count = sc_config_sizes[i].count; + int size = sc_config_sizes[i].size; + int pool; + + /* + * Sanity check count: Either a positive value or + * one of the expected wildcards is valid. The positive + * value is checked later when we compare against total + * memory available. + */ + if (i == SC_ACK) { + count = dd->n_krcv_queues; + } else if (i == SC_KERNEL) { + count = num_vls + 1 /* VL15 */; + } else if (count == SCC_PER_CPU) { + count = dd->num_rcv_contexts - dd->n_krcv_queues; + } else if (count < 0) { + dd_dev_err( + dd, + "%s send context invalid count wildcard %d\n", + sc_type_name(i), count); + return -EINVAL; + } + if (total_contexts + count > dd->chip_send_contexts) + count = dd->chip_send_contexts - total_contexts; + + total_contexts += count; + + /* + * Sanity check pool: The conversion will return a pool + * number or -1 if a fixed (non-negative) value. The fixed + * value is checked later when we compare against + * total memory available. + */ + pool = wildcard_to_pool(size); + if (pool == -1) { /* non-wildcard */ + fixed_blocks += size * count; + } else if (pool < NUM_SC_POOLS) { /* valid wildcard */ + mem_pool_info[pool].count += count; + } else { /* invalid wildcard */ + dd_dev_err( + dd, + "%s send context invalid pool wildcard %d\n", + sc_type_name(i), size); + return -EINVAL; + } + + dd->sc_sizes[i].count = count; + dd->sc_sizes[i].size = size; + } + if (fixed_blocks > total_blocks) { + dd_dev_err( + dd, + "Send context fixed block count, %u, larger than total block count %u\n", + fixed_blocks, total_blocks); + return -EINVAL; + } + + /* step 3: calculate the blocks in the pools, and pool context sizes */ + pool_blocks = total_blocks - fixed_blocks; + if (ab_total > pool_blocks) { + dd_dev_err( + dd, + "Send context fixed pool sizes, %u, larger than pool block count %u\n", + ab_total, pool_blocks); + return -EINVAL; + } + /* subtract off the fixed pool blocks */ + pool_blocks -= ab_total; + + for (i = 0; i < NUM_SC_POOLS; i++) { + struct mem_pool_info *pi = &mem_pool_info[i]; + + /* % beats absolute blocks */ + if (pi->centipercent >= 0) + pi->blocks = (pool_blocks * pi->centipercent) / 10000; + + if (pi->blocks == 0 && pi->count != 0) { + dd_dev_err( + dd, + "Send context memory pool %d has %u contexts, but no blocks\n", + i, pi->count); + return -EINVAL; + } + if (pi->count == 0) { + /* warn about wasted blocks */ + if (pi->blocks != 0) + dd_dev_err( + dd, + "Send context memory pool %d has %u blocks, but zero contexts\n", + i, pi->blocks); + pi->size = 0; + } else { + pi->size = pi->blocks / pi->count; + } + } + + /* step 4: fill in the context type sizes from the pool sizes */ + used_blocks = 0; + for (i = 0; i < SC_MAX; i++) { + if (dd->sc_sizes[i].size < 0) { + unsigned pool = wildcard_to_pool(dd->sc_sizes[i].size); + + WARN_ON_ONCE(pool >= NUM_SC_POOLS); + dd->sc_sizes[i].size = mem_pool_info[pool].size; + } + /* make sure we are not larger than what is allowed by the HW */ +#define PIO_MAX_BLOCKS 1024 + if (dd->sc_sizes[i].size > PIO_MAX_BLOCKS) + dd->sc_sizes[i].size = PIO_MAX_BLOCKS; + + /* calculate our total usage */ + used_blocks += dd->sc_sizes[i].size * dd->sc_sizes[i].count; + } + extra = total_blocks - used_blocks; + if (extra != 0) + dd_dev_info(dd, "unused send context blocks: %d\n", extra); + + return total_contexts; +} + +int init_send_contexts(struct hfi1_devdata *dd) +{ + u16 base; + int ret, i, j, context; + + ret = init_credit_return(dd); + if (ret) + return ret; + + dd->hw_to_sw = kmalloc_array(TXE_NUM_CONTEXTS, sizeof(u8), + GFP_KERNEL); + dd->send_contexts = kcalloc(dd->num_send_contexts, + sizeof(struct send_context_info), + GFP_KERNEL); + if (!dd->send_contexts || !dd->hw_to_sw) { + dd_dev_err(dd, "Unable to allocate send context arrays\n"); + kfree(dd->hw_to_sw); + kfree(dd->send_contexts); + free_credit_return(dd); + return -ENOMEM; + } + + /* hardware context map starts with invalid send context indices */ + for (i = 0; i < TXE_NUM_CONTEXTS; i++) + dd->hw_to_sw[i] = INVALID_SCI; + + /* + * All send contexts have their credit sizes. Allocate credits + * for each context one after another from the global space. + */ + context = 0; + base = 1; + for (i = 0; i < SC_MAX; i++) { + struct sc_config_sizes *scs = &dd->sc_sizes[i]; + + for (j = 0; j < scs->count; j++) { + struct send_context_info *sci = + &dd->send_contexts[context]; + sci->type = i; + sci->base = base; + sci->credits = scs->size; + + context++; + base += scs->size; + } + } + + return 0; +} + +/* + * Allocate a software index and hardware context of the given type. + * + * Must be called with dd->sc_lock held. + */ +static int sc_hw_alloc(struct hfi1_devdata *dd, int type, u32 *sw_index, + u32 *hw_context) +{ + struct send_context_info *sci; + u32 index; + u32 context; + + for (index = 0, sci = &dd->send_contexts[0]; + index < dd->num_send_contexts; index++, sci++) { + if (sci->type == type && sci->allocated == 0) { + sci->allocated = 1; + /* use a 1:1 mapping, but make them non-equal */ + context = dd->chip_send_contexts - index - 1; + dd->hw_to_sw[context] = index; + *sw_index = index; + *hw_context = context; + return 0; /* success */ + } + } + dd_dev_err(dd, "Unable to locate a free type %d send context\n", type); + return -ENOSPC; +} + +/* + * Free the send context given by its software index. + * + * Must be called with dd->sc_lock held. + */ +static void sc_hw_free(struct hfi1_devdata *dd, u32 sw_index, u32 hw_context) +{ + struct send_context_info *sci; + + sci = &dd->send_contexts[sw_index]; + if (!sci->allocated) { + dd_dev_err(dd, "%s: sw_index %u not allocated? hw_context %u\n", + __func__, sw_index, hw_context); + } + sci->allocated = 0; + dd->hw_to_sw[hw_context] = INVALID_SCI; +} + +/* return the base context of a context in a group */ +static inline u32 group_context(u32 context, u32 group) +{ + return (context >> group) << group; +} + +/* return the size of a group */ +static inline u32 group_size(u32 group) +{ + return 1 << group; +} + +/* + * Obtain the credit return addresses, kernel virtual and physical, for the + * given sc. + * + * To understand this routine: + * o va and pa are arrays of struct credit_return. One for each physical + * send context, per NUMA. + * o Each send context always looks in its relative location in a struct + * credit_return for its credit return. + * o Each send context in a group must have its return address CSR programmed + * with the same value. Use the address of the first send context in the + * group. + */ +static void cr_group_addresses(struct send_context *sc, dma_addr_t *pa) +{ + u32 gc = group_context(sc->hw_context, sc->group); + u32 index = sc->hw_context & 0x7; + + sc->hw_free = &sc->dd->cr_base[sc->node].va[gc].cr[index]; + *pa = (unsigned long) + &((struct credit_return *)sc->dd->cr_base[sc->node].pa)[gc]; +} + +/* + * Work queue function triggered in error interrupt routine for + * kernel contexts. + */ +static void sc_halted(struct work_struct *work) +{ + struct send_context *sc; + + sc = container_of(work, struct send_context, halt_work); + sc_restart(sc); +} + +/* + * Calculate PIO block threshold for this send context using the given MTU. + * Trigger a return when one MTU plus optional header of credits remain. + * + * Parameter mtu is in bytes. + * Parameter hdrqentsize is in DWORDs. + * + * Return value is what to write into the CSR: trigger return when + * unreturned credits pass this count. + */ +u32 sc_mtu_to_threshold(struct send_context *sc, u32 mtu, u32 hdrqentsize) +{ + u32 release_credits; + u32 threshold; + + /* add in the header size, then divide by the PIO block size */ + mtu += hdrqentsize << 2; + release_credits = DIV_ROUND_UP(mtu, PIO_BLOCK_SIZE); + + /* check against this context's credits */ + if (sc->credits <= release_credits) + threshold = 1; + else + threshold = sc->credits - release_credits; + + return threshold; +} + +/* + * Calculate credit threshold in terms of percent of the allocated credits. + * Trigger when unreturned credits equal or exceed the percentage of the whole. + * + * Return value is what to write into the CSR: trigger return when + * unreturned credits pass this count. + */ +static u32 sc_percent_to_threshold(struct send_context *sc, u32 percent) +{ + return (sc->credits * percent) / 100; +} + +/* + * Set the credit return threshold. + */ +void sc_set_cr_threshold(struct send_context *sc, u32 new_threshold) +{ + unsigned long flags; + u32 old_threshold; + int force_return = 0; + + spin_lock_irqsave(&sc->credit_ctrl_lock, flags); + + old_threshold = (sc->credit_ctrl >> + SC(CREDIT_CTRL_THRESHOLD_SHIFT)) + & SC(CREDIT_CTRL_THRESHOLD_MASK); + + if (new_threshold != old_threshold) { + sc->credit_ctrl = + (sc->credit_ctrl + & ~SC(CREDIT_CTRL_THRESHOLD_SMASK)) + | ((new_threshold + & SC(CREDIT_CTRL_THRESHOLD_MASK)) + << SC(CREDIT_CTRL_THRESHOLD_SHIFT)); + write_kctxt_csr(sc->dd, sc->hw_context, + SC(CREDIT_CTRL), sc->credit_ctrl); + + /* force a credit return on change to avoid a possible stall */ + force_return = 1; + } + + spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags); + + if (force_return) + sc_return_credits(sc); +} + +/* + * set_pio_integrity + * + * Set the CHECK_ENABLE register for the send context 'sc'. + */ +void set_pio_integrity(struct send_context *sc) +{ + struct hfi1_devdata *dd = sc->dd; + u64 reg = 0; + u32 hw_context = sc->hw_context; + int type = sc->type; + + /* + * No integrity checks if HFI1_CAP_NO_INTEGRITY is set, or if + * we're snooping. + */ + if (likely(!HFI1_CAP_IS_KSET(NO_INTEGRITY)) && + dd->hfi1_snoop.mode_flag != HFI1_PORT_SNOOP_MODE) + reg = hfi1_pkt_default_send_ctxt_mask(dd, type); + + write_kctxt_csr(dd, hw_context, SC(CHECK_ENABLE), reg); +} + +/* + * Allocate a NUMA relative send context structure of the given type along + * with a HW context. + */ +struct send_context *sc_alloc(struct hfi1_devdata *dd, int type, + uint hdrqentsize, int numa) +{ + struct send_context_info *sci; + struct send_context *sc; + dma_addr_t pa; + unsigned long flags; + u64 reg; + u32 thresh; + u32 sw_index; + u32 hw_context; + int ret; + u8 opval, opmask; + + /* do not allocate while frozen */ + if (dd->flags & HFI1_FROZEN) + return NULL; + + sc = kzalloc_node(sizeof(struct send_context), GFP_KERNEL, numa); + if (!sc) { + dd_dev_err(dd, "Cannot allocate send context structure\n"); + return NULL; + } + + spin_lock_irqsave(&dd->sc_lock, flags); + ret = sc_hw_alloc(dd, type, &sw_index, &hw_context); + if (ret) { + spin_unlock_irqrestore(&dd->sc_lock, flags); + kfree(sc); + return NULL; + } + + sci = &dd->send_contexts[sw_index]; + sci->sc = sc; + + sc->dd = dd; + sc->node = numa; + sc->type = type; + spin_lock_init(&sc->alloc_lock); + spin_lock_init(&sc->release_lock); + spin_lock_init(&sc->credit_ctrl_lock); + INIT_LIST_HEAD(&sc->piowait); + INIT_WORK(&sc->halt_work, sc_halted); + atomic_set(&sc->buffers_allocated, 0); + init_waitqueue_head(&sc->halt_wait); + + /* grouping is always single context for now */ + sc->group = 0; + + sc->sw_index = sw_index; + sc->hw_context = hw_context; + cr_group_addresses(sc, &pa); + sc->credits = sci->credits; + +/* PIO Send Memory Address details */ +#define PIO_ADDR_CONTEXT_MASK 0xfful +#define PIO_ADDR_CONTEXT_SHIFT 16 + sc->base_addr = dd->piobase + ((hw_context & PIO_ADDR_CONTEXT_MASK) + << PIO_ADDR_CONTEXT_SHIFT); + + /* set base and credits */ + reg = ((sci->credits & SC(CTRL_CTXT_DEPTH_MASK)) + << SC(CTRL_CTXT_DEPTH_SHIFT)) + | ((sci->base & SC(CTRL_CTXT_BASE_MASK)) + << SC(CTRL_CTXT_BASE_SHIFT)); + write_kctxt_csr(dd, hw_context, SC(CTRL), reg); + + set_pio_integrity(sc); + + /* unmask all errors */ + write_kctxt_csr(dd, hw_context, SC(ERR_MASK), (u64)-1); + + /* set the default partition key */ + write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY), + (DEFAULT_PKEY & + SC(CHECK_PARTITION_KEY_VALUE_MASK)) + << SC(CHECK_PARTITION_KEY_VALUE_SHIFT)); + + /* per context type checks */ + if (type == SC_USER) { + opval = USER_OPCODE_CHECK_VAL; + opmask = USER_OPCODE_CHECK_MASK; + } else { + opval = OPCODE_CHECK_VAL_DISABLED; + opmask = OPCODE_CHECK_MASK_DISABLED; + } + + /* set the send context check opcode mask and value */ + write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE), + ((u64)opmask << SC(CHECK_OPCODE_MASK_SHIFT)) | + ((u64)opval << SC(CHECK_OPCODE_VALUE_SHIFT))); + + /* set up credit return */ + reg = pa & SC(CREDIT_RETURN_ADDR_ADDRESS_SMASK); + write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), reg); + + /* + * Calculate the initial credit return threshold. + * + * For Ack contexts, set a threshold for half the credits. + * For User contexts use the given percentage. This has been + * sanitized on driver start-up. + * For Kernel contexts, use the default MTU plus a header. + */ + if (type == SC_ACK) { + thresh = sc_percent_to_threshold(sc, 50); + } else if (type == SC_USER) { + thresh = sc_percent_to_threshold(sc, + user_credit_return_threshold); + } else { /* kernel */ + thresh = sc_mtu_to_threshold(sc, hfi1_max_mtu, hdrqentsize); + } + reg = thresh << SC(CREDIT_CTRL_THRESHOLD_SHIFT); + /* add in early return */ + if (type == SC_USER && HFI1_CAP_IS_USET(EARLY_CREDIT_RETURN)) + reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK); + else if (HFI1_CAP_IS_KSET(EARLY_CREDIT_RETURN)) /* kernel, ack */ + reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK); + + /* set up write-through credit_ctrl */ + sc->credit_ctrl = reg; + write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), reg); + + /* User send contexts should not allow sending on VL15 */ + if (type == SC_USER) { + reg = 1ULL << 15; + write_kctxt_csr(dd, hw_context, SC(CHECK_VL), reg); + } + + spin_unlock_irqrestore(&dd->sc_lock, flags); + + /* + * Allocate shadow ring to track outstanding PIO buffers _after_ + * unlocking. We don't know the size until the lock is held and + * we can't allocate while the lock is held. No one is using + * the context yet, so allocate it now. + * + * User contexts do not get a shadow ring. + */ + if (type != SC_USER) { + /* + * Size the shadow ring 1 larger than the number of credits + * so head == tail can mean empty. + */ + sc->sr_size = sci->credits + 1; + sc->sr = kzalloc_node(sizeof(union pio_shadow_ring) * + sc->sr_size, GFP_KERNEL, numa); + if (!sc->sr) { + dd_dev_err(dd, + "Cannot allocate send context shadow ring structure\n"); + sc_free(sc); + return NULL; + } + } + + dd_dev_info(dd, + "Send context %u(%u) %s group %u credits %u credit_ctrl 0x%llx threshold %u\n", + sw_index, + hw_context, + sc_type_name(type), + sc->group, + sc->credits, + sc->credit_ctrl, + thresh); + + return sc; +} + +/* free a per-NUMA send context structure */ +void sc_free(struct send_context *sc) +{ + struct hfi1_devdata *dd; + unsigned long flags; + u32 sw_index; + u32 hw_context; + + if (!sc) + return; + + sc->flags |= SCF_IN_FREE; /* ensure no restarts */ + dd = sc->dd; + if (!list_empty(&sc->piowait)) + dd_dev_err(dd, "piowait list not empty!\n"); + sw_index = sc->sw_index; + hw_context = sc->hw_context; + sc_disable(sc); /* make sure the HW is disabled */ + flush_work(&sc->halt_work); + + spin_lock_irqsave(&dd->sc_lock, flags); + dd->send_contexts[sw_index].sc = NULL; + + /* clear/disable all registers set in sc_alloc */ + write_kctxt_csr(dd, hw_context, SC(CTRL), 0); + write_kctxt_csr(dd, hw_context, SC(CHECK_ENABLE), 0); + write_kctxt_csr(dd, hw_context, SC(ERR_MASK), 0); + write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY), 0); + write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE), 0); + write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), 0); + write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), 0); + + /* release the index and context for re-use */ + sc_hw_free(dd, sw_index, hw_context); + spin_unlock_irqrestore(&dd->sc_lock, flags); + + kfree(sc->sr); + kfree(sc); +} + +/* disable the context */ +void sc_disable(struct send_context *sc) +{ + u64 reg; + unsigned long flags; + struct pio_buf *pbuf; + + if (!sc) + return; + + /* do all steps, even if already disabled */ + spin_lock_irqsave(&sc->alloc_lock, flags); + reg = read_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL)); + reg &= ~SC(CTRL_CTXT_ENABLE_SMASK); + sc->flags &= ~SCF_ENABLED; + sc_wait_for_packet_egress(sc, 1); + write_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL), reg); + spin_unlock_irqrestore(&sc->alloc_lock, flags); + + /* + * Flush any waiters. Once the context is disabled, + * credit return interrupts are stopped (although there + * could be one in-process when the context is disabled). + * Wait one microsecond for any lingering interrupts, then + * proceed with the flush. + */ + udelay(1); + spin_lock_irqsave(&sc->release_lock, flags); + if (sc->sr) { /* this context has a shadow ring */ + while (sc->sr_tail != sc->sr_head) { + pbuf = &sc->sr[sc->sr_tail].pbuf; + if (pbuf->cb) + (*pbuf->cb)(pbuf->arg, PRC_SC_DISABLE); + sc->sr_tail++; + if (sc->sr_tail >= sc->sr_size) + sc->sr_tail = 0; + } + } + spin_unlock_irqrestore(&sc->release_lock, flags); +} + +/* return SendEgressCtxtStatus.PacketOccupancy */ +#define packet_occupancy(r) \ + (((r) & SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SMASK)\ + >> SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SHIFT) + +/* is egress halted on the context? */ +#define egress_halted(r) \ + ((r) & SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_HALT_STATUS_SMASK) + +/* wait for packet egress, optionally pause for credit return */ +static void sc_wait_for_packet_egress(struct send_context *sc, int pause) +{ + struct hfi1_devdata *dd = sc->dd; + u64 reg; + u32 loop = 0; + + while (1) { + reg = read_csr(dd, sc->hw_context * 8 + + SEND_EGRESS_CTXT_STATUS); + /* done if egress is stopped */ + if (egress_halted(reg)) + break; + reg = packet_occupancy(reg); + if (reg == 0) + break; + if (loop > 100) { + dd_dev_err(dd, + "%s: context %u(%u) timeout waiting for packets to egress, remaining count %u\n", + __func__, sc->sw_index, + sc->hw_context, (u32)reg); + break; + } + loop++; + udelay(1); + } + + if (pause) + /* Add additional delay to ensure chip returns all credits */ + pause_for_credit_return(dd); +} + +void sc_wait(struct hfi1_devdata *dd) +{ + int i; + + for (i = 0; i < dd->num_send_contexts; i++) { + struct send_context *sc = dd->send_contexts[i].sc; + + if (!sc) + continue; + sc_wait_for_packet_egress(sc, 0); + } +} + +/* + * Restart a context after it has been halted due to error. + * + * If the first step fails - wait for the halt to be asserted, return early. + * Otherwise complain about timeouts but keep going. + * + * It is expected that allocations (enabled flag bit) have been shut off + * already (only applies to kernel contexts). + */ +int sc_restart(struct send_context *sc) +{ + struct hfi1_devdata *dd = sc->dd; + u64 reg; + u32 loop; + int count; + + /* bounce off if not halted, or being free'd */ + if (!(sc->flags & SCF_HALTED) || (sc->flags & SCF_IN_FREE)) + return -EINVAL; + + dd_dev_info(dd, "restarting send context %u(%u)\n", sc->sw_index, + sc->hw_context); + + /* + * Step 1: Wait for the context to actually halt. + * + * The error interrupt is asynchronous to actually setting halt + * on the context. + */ + loop = 0; + while (1) { + reg = read_kctxt_csr(dd, sc->hw_context, SC(STATUS)); + if (reg & SC(STATUS_CTXT_HALTED_SMASK)) + break; + if (loop > 100) { + dd_dev_err(dd, "%s: context %u(%u) not halting, skipping\n", + __func__, sc->sw_index, sc->hw_context); + return -ETIME; + } + loop++; + udelay(1); + } + + /* + * Step 2: Ensure no users are still trying to write to PIO. + * + * For kernel contexts, we have already turned off buffer allocation. + * Now wait for the buffer count to go to zero. + * + * For user contexts, the user handling code has cut off write access + * to the context's PIO pages before calling this routine and will + * restore write access after this routine returns. + */ + if (sc->type != SC_USER) { + /* kernel context */ + loop = 0; + while (1) { + count = atomic_read(&sc->buffers_allocated); + if (count == 0) + break; + if (loop > 100) { + dd_dev_err(dd, + "%s: context %u(%u) timeout waiting for PIO buffers to zero, remaining %d\n", + __func__, sc->sw_index, + sc->hw_context, count); + } + loop++; + udelay(1); + } + } + + /* + * Step 3: Wait for all packets to egress. + * This is done while disabling the send context + * + * Step 4: Disable the context + * + * This is a superset of the halt. After the disable, the + * errors can be cleared. + */ + sc_disable(sc); + + /* + * Step 5: Enable the context + * + * This enable will clear the halted flag and per-send context + * error flags. + */ + return sc_enable(sc); +} + +/* + * PIO freeze processing. To be called after the TXE block is fully frozen. + * Go through all frozen send contexts and disable them. The contexts are + * already stopped by the freeze. + */ +void pio_freeze(struct hfi1_devdata *dd) +{ + struct send_context *sc; + int i; + + for (i = 0; i < dd->num_send_contexts; i++) { + sc = dd->send_contexts[i].sc; + /* + * Don't disable unallocated, unfrozen, or user send contexts. + * User send contexts will be disabled when the process + * calls into the driver to reset its context. + */ + if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER) + continue; + + /* only need to disable, the context is already stopped */ + sc_disable(sc); + } +} + +/* + * Unfreeze PIO for kernel send contexts. The precondition for calling this + * is that all PIO send contexts have been disabled and the SPC freeze has + * been cleared. Now perform the last step and re-enable each kernel context. + * User (PSM) processing will occur when PSM calls into the kernel to + * acknowledge the freeze. + */ +void pio_kernel_unfreeze(struct hfi1_devdata *dd) +{ + struct send_context *sc; + int i; + + for (i = 0; i < dd->num_send_contexts; i++) { + sc = dd->send_contexts[i].sc; + if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER) + continue; + + sc_enable(sc); /* will clear the sc frozen flag */ + } +} + +/* + * Wait for the SendPioInitCtxt.PioInitInProgress bit to clear. + * Returns: + * -ETIMEDOUT - if we wait too long + * -EIO - if there was an |