/***********************license start*************** * Author: Cavium Networks * * Contact: support@caviumnetworks.com * This file is part of the OCTEON SDK * * Copyright (c) 2003-2008 Cavium Networks * * 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. * * You should have received a copy of the GNU General Public License * along with this file; if not, write to the Free Software * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA * or visit http://www.gnu.org/licenses/. * * This file may also be available under a different license from Cavium. * Contact Cavium Networks for more information ***********************license end**************************************/ /* * * Support functions for managing command queues used for * various hardware blocks. * * The common command queue infrastructure abstracts out the * software necessary for adding to Octeon's chained queue * structures. These structures are used for commands to the * PKO, ZIP, DFA, RAID, and DMA engine blocks. Although each * hardware unit takes commands and CSRs of different types, * they all use basic linked command buffers to store the * pending request. In general, users of the CVMX API don't * call cvmx-cmd-queue functions directly. Instead the hardware * unit specific wrapper should be used. The wrappers perform * unit specific validation and CSR writes to submit the * commands. * * Even though most software will never directly interact with * cvmx-cmd-queue, knowledge of its internal working can help * in diagnosing performance problems and help with debugging. * * Command queue pointers are stored in a global named block * called "cvmx_cmd_queues". Except for the PKO queues, each * hardware queue is stored in its own cache line to reduce SMP * contention on spin locks. The PKO queues are stored such that * every 16th queue is next to each other in memory. This scheme * allows for queues being in separate cache lines when there * are low number of queues per port. With 16 queues per port, * the first queue for each port is in the same cache area. The * second queues for each port are in another area, etc. This * allows software to implement very efficient lockless PKO with * 16 queues per port using a minimum of cache lines per core. * All queues for a given core will be isolated in the same * cache area. * * In addition to the memory pointer layout, cvmx-cmd-queue * provides an optimized fair ll/sc locking mechanism for the * queues. The lock uses a "ticket / now serving" model to * maintain fair order on contended locks. In addition, it uses * predicted locking time to limit cache contention. When a core * know it must wait in line for a lock, it spins on the * internal cycle counter to completely eliminate any causes of * bus traffic. * */ #ifndef __CVMX_CMD_QUEUE_H__ #define __CVMX_CMD_QUEUE_H__ #include <linux/prefetch.h> #include <asm/compiler.h> #include <asm/octeon/cvmx-fpa.h> /** * By default we disable the max depth support. Most programs * don't use it and it slows down the command queue processing * significantly. */ #ifndef CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH #define CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH 0 #endif /** * Enumeration representing all hardware blocks that use command * queues. Each hardware block has up to 65536 sub identifiers for * multiple command queues. Not all chips support all hardware * units. */ typedef enum { CVMX_CMD_QUEUE_PKO_BASE = 0x00000, #define CVMX_CMD_QUEUE_PKO(queue) \ ((cvmx_cmd_queue_id_t)(CVMX_CMD_QUEUE_PKO_BASE + (0xffff&(queue)))) CVMX_CMD_QUEUE_ZIP = 0x10000, CVMX_CMD_QUEUE_DFA = 0x20000, CVMX_CMD_QUEUE_RAID = 0x30000, CVMX_CMD_QUEUE_DMA_BASE = 0x40000, #define CVMX_CMD_QUEUE_DMA(queue) \ ((cvmx_cmd_queue_id_t)(CVMX_CMD_QUEUE_DMA_BASE + (0xffff&(queue)))) CVMX_CMD_QUEUE_END = 0x50000, } cvmx_cmd_queue_id_t; /** * Command write operations can fail if the command queue needs * a new buffer and the associated FPA pool is empty. It can also * fail if the number of queued command words reaches the maximum * set at initialization. */ typedef enum { CVMX_CMD_QUEUE_SUCCESS = 0, CVMX_CMD_QUEUE_NO_MEMORY = -1, CVMX_CMD_QUEUE_FULL = -2, CVMX_CMD_QUEUE_INVALID_PARAM = -3, CVMX_CMD_QUEUE_ALREADY_SETUP = -4, } cvmx_cmd_queue_result_t; typedef struct { /* You have lock when this is your ticket */ uint8_t now_serving; uint64_t unused1:24; /* Maximum outstanding command words */ uint32_t max_depth; /* FPA pool buffers come from */ uint64_t fpa_pool:3; /* Top of command buffer pointer shifted 7 */ uint64_t base_ptr_div128:29; uint64_t unused2:6; /* FPA buffer size in 64bit words minus 1 */ uint64_t pool_size_m1:13; /* Number of commands already used in buffer */ uint64_t index:13; } __cvmx_cmd_queue_state_t; /** * This structure contains the global state of all command queues. * It is stored in a bootmem named block and shared by all * applications running on Octeon. Tickets are stored in a differnet * cache line that queue information to reduce the contention on the * ll/sc used to get a ticket. If this is not the case, the update * of queue state causes the ll/sc to fail quite often. */ typedef struct { uint64_t ticket[(CVMX_CMD_QUEUE_END >> 16) * 256]; __cvmx_cmd_queue_state_t state[(CVMX_CMD_QUEUE_END >> 16) * 256]; } __cvmx_cmd_queue_all_state_t; /** * Initialize a command queue for use. The initial FPA buffer is * allocated and the hardware unit is configured to point to the * new command queue. * * @queue_id: Hardware command queue to initialize. * @max_depth: Maximum outstanding commands that can be queued. * @fpa_pool: FPA pool the command queues should come from. * @pool_size: Size of each buffer in the FPA pool (bytes) * * Returns CVMX_CMD_QUEUE_SUCCESS or a failure code */ cvmx_cmd_queue_result_t cvmx_cmd_queue_initialize(cvmx_cmd_queue_id_t queue_id, int max_depth, int fpa_pool, int pool_size); /** * Shutdown a queue a free it's command buffers to the FPA. The * hardware connected to the queue must be stopped before this * function is called. * * @queue_id: Queue to shutdown * * Returns CVMX_CMD_QUEUE_SUCCESS or a failure code */ cvmx_cmd_queue_result_t cvmx_cmd_queue_shutdown(cvmx_cmd_queue_id_t queue_id); /** * Return the number of command words pending in the queue. This * function may be relatively slow for some hardware units. * * @queue_id: Hardware command queue to query * * Returns Number of outstanding commands */ int cvmx_cmd_queue_length(cvmx_cmd_queue_id_t queue_id); /** * Return the command buffer to be written to. The purpose of this * function is to allow CVMX routine access t othe low level buffer * for initial hardware setup. User applications should not call this * function directly. * * @queue_id: Command queue to query * * Returns Command buffer or NULL on failure */ void *cvmx_cmd_queue_buffer(cvmx_cmd_queue_id_t queue_id); /** * Get the index into the state arrays for the supplied queue id. * * @queue_id: Queue ID to get an index for * * Returns Index into the state arrays */ static inline int __cvmx_cmd_queue_get_index(cvmx_cmd_queue_id_t queue_id) { /* * Warning: This code currently only works with devices that * have 256 queues or less. Devices with more than 16 queues * are laid out in memory to allow cores quick access to * every 16th queue. This reduces cache thrashing when you are * running 16 queues per port to support lockless operation. */ int unit = queue_id >> 16; int q = (queue_id >> 4) & 0xf; int core = queue_id & 0xf; return unit * 256 + core * 16 + q; } /** * Lock the supplied queue so nobody else is updating it at the same * time as us. * * @queue_id: Queue ID to lock * @qptr: Pointer to the queue's global state */ static inline void __cvmx_cmd_queue_lock(cvmx_cmd_queue_id_t queue_id, __cvmx_cmd_queue_state_t *qptr) { extern __cvmx_cmd_queue_all_state_t *__cvmx_cmd_queue_state_ptr; int tmp; int my_ticket; prefetch(qptr); asm volatile ( ".set push\n" ".set noreorder\n" "1:\n" /* Atomic add one to ticket_ptr */ "ll %[my_ticket], %[ticket_ptr]\n" /* and store the original value */ "li %[ticket], 1\n" /* in my_ticket */ "baddu %[ticket], %[my_ticket]\n" "sc %[ticket], %[ticket_ptr]\n" "beqz %[ticket], 1b\n" " nop\n" /* Load the current now_serving ticket */ "lbu %[ticket], %[now_serving]\n" "2:\n" /* Jump out if now_serving == my_ticket */ "beq %[ticket], %[my_ticket], 4f\n" /* Find out how many tickets are in front of me */ " subu %[ticket], %[my_ticket], %[ticket]\n" /* Use tickets in front of me minus one to delay */ "subu %[ticket], 1\n" /* Delay will be ((tickets in front)-1)*32 loops */ "cins %[ticket], %[ticket], 5, 7\n" "3:\n" /* Loop here until our ticket might be up */ "bnez %[ticket], 3b\n" " subu %[ticket], 1\n" /* Jump back up to check out ticket again */ "b 2b\n" /* Load the current now_serving ticket */ " lbu %[ticket], %[now_serving]\n" "4:\n" ".set pop\n" : [ticket_ptr] "=" GCC_OFF_SMALL_ASM()(__cvmx_cmd_queue_state_ptr->ticket[__cvmx_cmd_queue_get_index(queue_id)]), [now_serving] "=m"(qptr->now_serving), [ticket] "=r"(tmp), [my_ticket] "=r"(my_ticket) ); } /** * Unlock the queue, flushing all writes. * * @qptr: Queue to unlock */ static inline void __cvmx_cmd_queue_unlock(__cvmx_cmd_queue_state_t *qptr) { qptr->now_serving++; CVMX_SYNCWS; } /** * Get the queue state structure for the given queue id * * @queue_id: Queue id to get * * Returns Queue structure or NULL on failure */ static inline __cvmx_cmd_queue_state_t *__cvmx_cmd_queue_get_state(cvmx_cmd_queue_id_t queue_id) { extern __cvmx_cmd_queue_all_state_t *__cvmx_cmd_queue_state_ptr; return &__cvmx_cmd_queue_state_ptr-> state[__cvmx_cmd_queue_get_index(queue_id)]; } /** * Write an arbitrary number of command words to a command queue. * This is a generic function; the fixed number of command word * functions yield higher performance. * * @queue_id: Hardware command queue to write to * @use_locking: * Use internal locking to ensure exclusive access for queue * updates. If you don't use this locking you must ensure * exclusivity some other way. Locking is strongly recommended. * @cmd_count: Number of command words to write * @cmds: Array of commands to write * * Returns CVMX_CMD_QUEUE_SUCCESS or a failure code */ static inline cvmx_cmd_queue_result_t cvmx_cmd_queue_write(cvmx_cmd_queue_id_t queue_id, int use_locking, int cmd_count, uint64_t *cmds) { __cvmx_cmd_queue_state_t *qptr = __cvmx_cmd_queue_get_state(queue_id); /* Make sure nobody else is updating the same queue */ if (likely(use_locking)) __cvmx_cmd_queue_lock(queue_id, qptr); /* * If a max queue length was specified then make sure we don't * exceed it. If any part of the command would be below the * limit we allow it. */ if (CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH && unlikely(qptr->max_depth)) { if (unlikely (cvmx_cmd_queue_length(queue_id) > (int)qptr->max_depth)) { if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_FULL; } } /* * Normally there is plenty of room in the current buffer for * the command. */ if (likely(qptr->index + cmd_count < qptr->pool_size_m1)) { uint64_t *ptr = (uint64_t *) cvmx_phys_to_ptr((uint64_t) qptr-> base_ptr_div128 << 7); ptr += qptr->index; qptr->index += cmd_count; while (cmd_count--) *ptr++ = *cmds++; } else { uint64_t *ptr; int count; /* * We need a new command buffer. Fail if there isn't * one available. */ uint64_t *new_buffer = (uint64_t *) cvmx_fpa_alloc(qptr->fpa_pool); if (unlikely(new_buffer == NULL)) { if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_NO_MEMORY; } ptr = (uint64_t *) cvmx_phys_to_ptr((uint64_t) qptr-> base_ptr_div128 << 7); /* * Figure out how many command words will fit in this * buffer. One location will be needed for the next * buffer pointer. */ count = qptr->pool_size_m1 - qptr->index; ptr += qptr->index; cmd_count -= count; while (count--) *ptr++ = *cmds++; *ptr = cvmx_ptr_to_phys(new_buffer); /* * The current buffer is full and has a link to the * next buffer. Time to write the rest of the commands * into the new buffer. */ qptr->base_ptr_div128 = *ptr >> 7; qptr->index = cmd_count; ptr = new_buffer; while (cmd_count--) *ptr++ = *cmds++; } /* All updates are complete. Release the lock and return */ if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_SUCCESS; } /** * Simple function to write two command words to a command * queue. * * @queue_id: Hardware command queue to write to * @use_locking: * Use internal locking to ensure exclusive access for queue * updates. If you don't use this locking you must ensure * exclusivity some other way. Locking is strongly recommended. * @cmd1: Command * @cmd2: Command * * Returns CVMX_CMD_QUEUE_SUCCESS or a failure code */ static inline cvmx_cmd_queue_result_t cvmx_cmd_queue_write2(cvmx_cmd_queue_id_t queue_id, int use_locking, uint64_t cmd1, uint64_t cmd2) { __cvmx_cmd_queue_state_t *qptr = __cvmx_cmd_queue_get_state(queue_id); /* Make sure nobody else is updating the same queue */ if (likely(use_locking)) __cvmx_cmd_queue_lock(queue_id, qptr); /* * If a max queue length was specified then make sure we don't * exceed it. If any part of the command would be below the * limit we allow it. */ if (CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH && unlikely(qptr->max_depth)) { if (unlikely (cvmx_cmd_queue_length(queue_id) > (int)qptr->max_depth)) { if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_FULL; } } /* * Normally there is plenty of room in the current buffer for * the command. */ if (likely(qptr->index + 2 < qptr->pool_size_m1)) { uint64_t *ptr = (uint64_t *) cvmx_phys_to_ptr((uint64_t) qptr-> base_ptr_div128 << 7); ptr += qptr->index; qptr->index += 2; ptr[0] = cmd1; ptr[1] = cmd2; } else { uint64_t *ptr; /* * Figure out how many command words will fit in this * buffer. One location will be needed for the next * buffer pointer. */ int count = qptr->pool_size_m1 - qptr->index; /* * We need a new command buffer. Fail if there isn't * one available. */ uint64_t *new_buffer = (uint64_t *) cvmx_fpa_alloc(qptr->fpa_pool); if (unlikely(new_buffer == NULL)) { if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_NO_MEMORY; } count--; ptr = (uint64_t *) cvmx_phys_to_ptr((uint64_t) qptr-> base_ptr_div128 << 7); ptr += qptr->index; *ptr++ = cmd1; if (likely(count)) *ptr++ = cmd2; *ptr = cvmx_ptr_to_phys(new_buffer); /* * The current buffer is full and has a link to the * next buffer. Time to write the rest of the commands * into the new buffer. */ qptr->base_ptr_div128 = *ptr >> 7; qptr->index = 0; if (unlikely(count == 0)) { qptr->index = 1; new_buffer[0] = cmd2; } } /* All updates are complete. Release the lock and return */ if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_SUCCESS; } /** * Simple function to write three command words to a command * queue. * * @queue_id: Hardware command queue to write to * @use_locking: * Use internal locking to ensure exclusive access for queue * updates. If you don't use this locking you must ensure * exclusivity some other way. Locking is strongly recommended. * @cmd1: Command * @cmd2: Command * @cmd3: Command * * Returns CVMX_CMD_QUEUE_SUCCESS or a failure code */ static inline cvmx_cmd_queue_result_t cvmx_cmd_queue_write3(cvmx_cmd_queue_id_t queue_id, int use_locking, uint64_t cmd1, uint64_t cmd2, uint64_t cmd3) { __cvmx_cmd_queue_state_t *qptr = __cvmx_cmd_queue_get_state(queue_id); /* Make sure nobody else is updating the same queue */ if (likely(use_locking)) __cvmx_cmd_queue_lock(queue_id, qptr); /* * If a max queue length was specified then make sure we don't * exceed it. If any part of the command would be below the * limit we allow it. */ if (CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH && unlikely(qptr->max_depth)) { if (unlikely (cvmx_cmd_queue_length(queue_id) > (int)qptr->max_depth)) { if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_FULL; } } /* * Normally there is plenty of room in the current buffer for * the command. */ if (likely(qptr->index + 3 < qptr->pool_size_m1)) { uint64_t *ptr = (uint64_t *) cvmx_phys_to_ptr((uint64_t) qptr-> base_ptr_div128 << 7); ptr += qptr->index; qptr->index += 3; ptr[0] = cmd1; ptr[1] = cmd2; ptr[2] = cmd3; } else { uint64_t *ptr; /* * Figure out how many command words will fit in this * buffer. One location will be needed for the next * buffer pointer */ int count = qptr->pool_size_m1 - qptr->index; /* * We need a new command buffer. Fail if there isn't * one available */ uint64_t *new_buffer = (uint64_t *) cvmx_fpa_alloc(qptr->fpa_pool); if (unlikely(new_buffer == NULL)) { if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_NO_MEMORY; } count--; ptr = (uint64_t *) cvmx_phys_to_ptr((uint64_t) qptr-> base_ptr_div128 << 7); ptr += qptr->index; *ptr++ = cmd1; if (count) { *ptr++ = cmd2; if (count > 1) *ptr++ = cmd3; } *ptr = cvmx_ptr_to_phys(new_buffer); /* * The current buffer is full and has a link to the * next buffer. Time to write the rest of the commands * into the new buffer. */ qptr->base_ptr_div128 = *ptr >> 7; qptr->index = 0; ptr = new_buffer; if (count == 0) { *ptr++ = cmd2; qptr->index++; } if (count < 2) { *ptr++ = cmd3; qptr->index++; } } /* All updates are complete. Release the lock and return */ if (likely(use_locking)) __cvmx_cmd_queue_unlock(qptr); return CVMX_CMD_QUEUE_SUCCESS; } #endif /* __CVMX_CMD_QUEUE_H__ */