Back to home page

OSCL-LXR
 
 

     

0001 /*
0002  * This file is provided under a dual BSD/GPLv2 license.  When using or
0003  *   redistributing this file, you may do so under either license.
0004  *
0005  *   GPL LICENSE SUMMARY
0006  *
0007  *   Copyright(c) 2015 Intel Corporation. All rights reserved.
0008  *   Copyright(c) 2017 T-Platforms. All Rights Reserved.
0009  *
0010  *   This program is free software; you can redistribute it and/or modify
0011  *   it under the terms of version 2 of the GNU General Public License as
0012  *   published by the Free Software Foundation.
0013  *
0014  *   BSD LICENSE
0015  *
0016  *   Copyright(c) 2015 Intel Corporation. All rights reserved.
0017  *   Copyright(c) 2017 T-Platforms. All Rights Reserved.
0018  *
0019  *   Redistribution and use in source and binary forms, with or without
0020  *   modification, are permitted provided that the following conditions
0021  *   are met:
0022  *
0023  *     * Redistributions of source code must retain the above copyright
0024  *       notice, this list of conditions and the following disclaimer.
0025  *     * Redistributions in binary form must reproduce the above copy
0026  *       notice, this list of conditions and the following disclaimer in
0027  *       the documentation and/or other materials provided with the
0028  *       distribution.
0029  *     * Neither the name of Intel Corporation nor the names of its
0030  *       contributors may be used to endorse or promote products derived
0031  *       from this software without specific prior written permission.
0032  *
0033  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
0034  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
0035  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
0036  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
0037  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
0038  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
0039  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
0040  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
0041  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
0042  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
0043  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
0044  *
0045  * PCIe NTB Perf Linux driver
0046  */
0047 
0048 /*
0049  * How to use this tool, by example.
0050  *
0051  * Assuming $DBG_DIR is something like:
0052  * '/sys/kernel/debug/ntb_perf/0000:00:03.0'
0053  * Suppose aside from local device there is at least one remote device
0054  * connected to NTB with index 0.
0055  *-----------------------------------------------------------------------------
0056  * Eg: install driver with specified chunk/total orders and dma-enabled flag
0057  *
0058  * root@self# insmod ntb_perf.ko chunk_order=19 total_order=28 use_dma
0059  *-----------------------------------------------------------------------------
0060  * Eg: check NTB ports (index) and MW mapping information
0061  *
0062  * root@self# cat $DBG_DIR/info
0063  *-----------------------------------------------------------------------------
0064  * Eg: start performance test with peer (index 0) and get the test metrics
0065  *
0066  * root@self# echo 0 > $DBG_DIR/run
0067  * root@self# cat $DBG_DIR/run
0068  */
0069 
0070 #include <linux/init.h>
0071 #include <linux/kernel.h>
0072 #include <linux/module.h>
0073 #include <linux/sched.h>
0074 #include <linux/wait.h>
0075 #include <linux/dma-mapping.h>
0076 #include <linux/dmaengine.h>
0077 #include <linux/pci.h>
0078 #include <linux/ktime.h>
0079 #include <linux/slab.h>
0080 #include <linux/delay.h>
0081 #include <linux/sizes.h>
0082 #include <linux/workqueue.h>
0083 #include <linux/debugfs.h>
0084 #include <linux/random.h>
0085 #include <linux/ntb.h>
0086 
0087 #define DRIVER_NAME     "ntb_perf"
0088 #define DRIVER_VERSION      "2.0"
0089 
0090 MODULE_LICENSE("Dual BSD/GPL");
0091 MODULE_VERSION(DRIVER_VERSION);
0092 MODULE_AUTHOR("Dave Jiang <dave.jiang@intel.com>");
0093 MODULE_DESCRIPTION("PCIe NTB Performance Measurement Tool");
0094 
0095 #define MAX_THREADS_CNT     32
0096 #define DEF_THREADS_CNT     1
0097 #define MAX_CHUNK_SIZE      SZ_1M
0098 #define MAX_CHUNK_ORDER     20 /* no larger than 1M */
0099 
0100 #define DMA_TRIES       100
0101 #define DMA_MDELAY      10
0102 
0103 #define MSG_TRIES       1000
0104 #define MSG_UDELAY_LOW      1000000
0105 #define MSG_UDELAY_HIGH     2000000
0106 
0107 #define PERF_BUF_LEN 1024
0108 
0109 static unsigned long max_mw_size;
0110 module_param(max_mw_size, ulong, 0644);
0111 MODULE_PARM_DESC(max_mw_size, "Upper limit of memory window size");
0112 
0113 static unsigned char chunk_order = 19; /* 512K */
0114 module_param(chunk_order, byte, 0644);
0115 MODULE_PARM_DESC(chunk_order, "Data chunk order [2^n] to transfer");
0116 
0117 static unsigned char total_order = 30; /* 1G */
0118 module_param(total_order, byte, 0644);
0119 MODULE_PARM_DESC(total_order, "Total data order [2^n] to transfer");
0120 
0121 static bool use_dma; /* default to 0 */
0122 module_param(use_dma, bool, 0644);
0123 MODULE_PARM_DESC(use_dma, "Use DMA engine to measure performance");
0124 
0125 /*==============================================================================
0126  *                         Perf driver data definition
0127  *==============================================================================
0128  */
0129 
0130 enum perf_cmd {
0131     PERF_CMD_INVAL = -1,/* invalid spad command */
0132     PERF_CMD_SSIZE = 0, /* send out buffer size */
0133     PERF_CMD_RSIZE = 1, /* recv in  buffer size */
0134     PERF_CMD_SXLAT = 2, /* send in  buffer xlat */
0135     PERF_CMD_RXLAT = 3, /* recv out buffer xlat */
0136     PERF_CMD_CLEAR = 4, /* clear allocated memory */
0137     PERF_STS_DONE  = 5, /* init is done */
0138     PERF_STS_LNKUP = 6, /* link up state flag */
0139 };
0140 
0141 struct perf_ctx;
0142 
0143 struct perf_peer {
0144     struct perf_ctx *perf;
0145     int pidx;
0146     int gidx;
0147 
0148     /* Outbound MW params */
0149     u64 outbuf_xlat;
0150     resource_size_t outbuf_size;
0151     void __iomem *outbuf;
0152     phys_addr_t out_phys_addr;
0153     dma_addr_t dma_dst_addr;
0154     /* Inbound MW params */
0155     dma_addr_t inbuf_xlat;
0156     resource_size_t inbuf_size;
0157     void        *inbuf;
0158 
0159     /* NTB connection setup service */
0160     struct work_struct  service;
0161     unsigned long       sts;
0162 
0163     struct completion init_comp;
0164 };
0165 #define to_peer_service(__work) \
0166     container_of(__work, struct perf_peer, service)
0167 
0168 struct perf_thread {
0169     struct perf_ctx *perf;
0170     int tidx;
0171 
0172     /* DMA-based test sync parameters */
0173     atomic_t dma_sync;
0174     wait_queue_head_t dma_wait;
0175     struct dma_chan *dma_chan;
0176 
0177     /* Data source and measured statistics */
0178     void *src;
0179     u64 copied;
0180     ktime_t duration;
0181     int status;
0182     struct work_struct work;
0183 };
0184 #define to_thread_work(__work) \
0185     container_of(__work, struct perf_thread, work)
0186 
0187 struct perf_ctx {
0188     struct ntb_dev *ntb;
0189 
0190     /* Global device index and peers descriptors */
0191     int gidx;
0192     int pcnt;
0193     struct perf_peer *peers;
0194 
0195     /* Performance measuring work-threads interface */
0196     unsigned long busy_flag;
0197     wait_queue_head_t twait;
0198     atomic_t tsync;
0199     u8 tcnt;
0200     struct perf_peer *test_peer;
0201     struct perf_thread threads[MAX_THREADS_CNT];
0202 
0203     /* Scratchpad/Message IO operations */
0204     int (*cmd_send)(struct perf_peer *peer, enum perf_cmd cmd, u64 data);
0205     int (*cmd_recv)(struct perf_ctx *perf, int *pidx, enum perf_cmd *cmd,
0206             u64 *data);
0207 
0208     struct dentry *dbgfs_dir;
0209 };
0210 
0211 /*
0212  * Scratchpads-base commands interface
0213  */
0214 #define PERF_SPAD_CNT(_pcnt) \
0215     (3*((_pcnt) + 1))
0216 #define PERF_SPAD_CMD(_gidx) \
0217     (3*(_gidx))
0218 #define PERF_SPAD_LDATA(_gidx) \
0219     (3*(_gidx) + 1)
0220 #define PERF_SPAD_HDATA(_gidx) \
0221     (3*(_gidx) + 2)
0222 #define PERF_SPAD_NOTIFY(_gidx) \
0223     (BIT_ULL(_gidx))
0224 
0225 /*
0226  * Messages-base commands interface
0227  */
0228 #define PERF_MSG_CNT        3
0229 #define PERF_MSG_CMD        0
0230 #define PERF_MSG_LDATA      1
0231 #define PERF_MSG_HDATA      2
0232 
0233 /*==============================================================================
0234  *                           Static data declarations
0235  *==============================================================================
0236  */
0237 
0238 static struct dentry *perf_dbgfs_topdir;
0239 
0240 static struct workqueue_struct *perf_wq __read_mostly;
0241 
0242 /*==============================================================================
0243  *                  NTB cross-link commands execution service
0244  *==============================================================================
0245  */
0246 
0247 static void perf_terminate_test(struct perf_ctx *perf);
0248 
0249 static inline bool perf_link_is_up(struct perf_peer *peer)
0250 {
0251     u64 link;
0252 
0253     link = ntb_link_is_up(peer->perf->ntb, NULL, NULL);
0254     return !!(link & BIT_ULL_MASK(peer->pidx));
0255 }
0256 
0257 static int perf_spad_cmd_send(struct perf_peer *peer, enum perf_cmd cmd,
0258                   u64 data)
0259 {
0260     struct perf_ctx *perf = peer->perf;
0261     int try;
0262     u32 sts;
0263 
0264     dev_dbg(&perf->ntb->dev, "CMD send: %d 0x%llx\n", cmd, data);
0265 
0266     /*
0267      * Perform predefined number of attempts before give up.
0268      * We are sending the data to the port specific scratchpad, so
0269      * to prevent a multi-port access race-condition. Additionally
0270      * there is no need in local locking since only thread-safe
0271      * service work is using this method.
0272      */
0273     for (try = 0; try < MSG_TRIES; try++) {
0274         if (!perf_link_is_up(peer))
0275             return -ENOLINK;
0276 
0277         sts = ntb_peer_spad_read(perf->ntb, peer->pidx,
0278                      PERF_SPAD_CMD(perf->gidx));
0279         if (sts != PERF_CMD_INVAL) {
0280             usleep_range(MSG_UDELAY_LOW, MSG_UDELAY_HIGH);
0281             continue;
0282         }
0283 
0284         ntb_peer_spad_write(perf->ntb, peer->pidx,
0285                     PERF_SPAD_LDATA(perf->gidx),
0286                     lower_32_bits(data));
0287         ntb_peer_spad_write(perf->ntb, peer->pidx,
0288                     PERF_SPAD_HDATA(perf->gidx),
0289                     upper_32_bits(data));
0290         ntb_peer_spad_write(perf->ntb, peer->pidx,
0291                     PERF_SPAD_CMD(perf->gidx),
0292                     cmd);
0293         ntb_peer_db_set(perf->ntb, PERF_SPAD_NOTIFY(peer->gidx));
0294 
0295         dev_dbg(&perf->ntb->dev, "DB ring peer %#llx\n",
0296             PERF_SPAD_NOTIFY(peer->gidx));
0297 
0298         break;
0299     }
0300 
0301     return try < MSG_TRIES ? 0 : -EAGAIN;
0302 }
0303 
0304 static int perf_spad_cmd_recv(struct perf_ctx *perf, int *pidx,
0305                   enum perf_cmd *cmd, u64 *data)
0306 {
0307     struct perf_peer *peer;
0308     u32 val;
0309 
0310     ntb_db_clear(perf->ntb, PERF_SPAD_NOTIFY(perf->gidx));
0311 
0312     /*
0313      * We start scanning all over, since cleared DB may have been set
0314      * by any peer. Yes, it makes peer with smaller index being
0315      * serviced with greater priority, but it's convenient for spad
0316      * and message code unification and simplicity.
0317      */
0318     for (*pidx = 0; *pidx < perf->pcnt; (*pidx)++) {
0319         peer = &perf->peers[*pidx];
0320 
0321         if (!perf_link_is_up(peer))
0322             continue;
0323 
0324         val = ntb_spad_read(perf->ntb, PERF_SPAD_CMD(peer->gidx));
0325         if (val == PERF_CMD_INVAL)
0326             continue;
0327 
0328         *cmd = val;
0329 
0330         val = ntb_spad_read(perf->ntb, PERF_SPAD_LDATA(peer->gidx));
0331         *data = val;
0332 
0333         val = ntb_spad_read(perf->ntb, PERF_SPAD_HDATA(peer->gidx));
0334         *data |= (u64)val << 32;
0335 
0336         /* Next command can be retrieved from now */
0337         ntb_spad_write(perf->ntb, PERF_SPAD_CMD(peer->gidx),
0338                    PERF_CMD_INVAL);
0339 
0340         dev_dbg(&perf->ntb->dev, "CMD recv: %d 0x%llx\n", *cmd, *data);
0341 
0342         return 0;
0343     }
0344 
0345     return -ENODATA;
0346 }
0347 
0348 static int perf_msg_cmd_send(struct perf_peer *peer, enum perf_cmd cmd,
0349                  u64 data)
0350 {
0351     struct perf_ctx *perf = peer->perf;
0352     int try, ret;
0353     u64 outbits;
0354 
0355     dev_dbg(&perf->ntb->dev, "CMD send: %d 0x%llx\n", cmd, data);
0356 
0357     /*
0358      * Perform predefined number of attempts before give up. Message
0359      * registers are free of race-condition problem when accessed
0360      * from different ports, so we don't need splitting registers
0361      * by global device index. We also won't have local locking,
0362      * since the method is used from service work only.
0363      */
0364     outbits = ntb_msg_outbits(perf->ntb);
0365     for (try = 0; try < MSG_TRIES; try++) {
0366         if (!perf_link_is_up(peer))
0367             return -ENOLINK;
0368 
0369         ret = ntb_msg_clear_sts(perf->ntb, outbits);
0370         if (ret)
0371             return ret;
0372 
0373         ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_LDATA,
0374                    lower_32_bits(data));
0375 
0376         if (ntb_msg_read_sts(perf->ntb) & outbits) {
0377             usleep_range(MSG_UDELAY_LOW, MSG_UDELAY_HIGH);
0378             continue;
0379         }
0380 
0381         ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_HDATA,
0382                    upper_32_bits(data));
0383 
0384         /* This call shall trigger peer message event */
0385         ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_CMD, cmd);
0386 
0387         break;
0388     }
0389 
0390     return try < MSG_TRIES ? 0 : -EAGAIN;
0391 }
0392 
0393 static int perf_msg_cmd_recv(struct perf_ctx *perf, int *pidx,
0394                  enum perf_cmd *cmd, u64 *data)
0395 {
0396     u64 inbits;
0397     u32 val;
0398 
0399     inbits = ntb_msg_inbits(perf->ntb);
0400 
0401     if (hweight64(ntb_msg_read_sts(perf->ntb) & inbits) < 3)
0402         return -ENODATA;
0403 
0404     val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_CMD);
0405     *cmd = val;
0406 
0407     val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_LDATA);
0408     *data = val;
0409 
0410     val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_HDATA);
0411     *data |= (u64)val << 32;
0412 
0413     /* Next command can be retrieved from now */
0414     ntb_msg_clear_sts(perf->ntb, inbits);
0415 
0416     dev_dbg(&perf->ntb->dev, "CMD recv: %d 0x%llx\n", *cmd, *data);
0417 
0418     return 0;
0419 }
0420 
0421 static int perf_cmd_send(struct perf_peer *peer, enum perf_cmd cmd, u64 data)
0422 {
0423     struct perf_ctx *perf = peer->perf;
0424 
0425     if (cmd == PERF_CMD_SSIZE || cmd == PERF_CMD_SXLAT)
0426         return perf->cmd_send(peer, cmd, data);
0427 
0428     dev_err(&perf->ntb->dev, "Send invalid command\n");
0429     return -EINVAL;
0430 }
0431 
0432 static int perf_cmd_exec(struct perf_peer *peer, enum perf_cmd cmd)
0433 {
0434     switch (cmd) {
0435     case PERF_CMD_SSIZE:
0436     case PERF_CMD_RSIZE:
0437     case PERF_CMD_SXLAT:
0438     case PERF_CMD_RXLAT:
0439     case PERF_CMD_CLEAR:
0440         break;
0441     default:
0442         dev_err(&peer->perf->ntb->dev, "Exec invalid command\n");
0443         return -EINVAL;
0444     }
0445 
0446     /* No need of memory barrier, since bit ops have invernal lock */
0447     set_bit(cmd, &peer->sts);
0448 
0449     dev_dbg(&peer->perf->ntb->dev, "CMD exec: %d\n", cmd);
0450 
0451     (void)queue_work(system_highpri_wq, &peer->service);
0452 
0453     return 0;
0454 }
0455 
0456 static int perf_cmd_recv(struct perf_ctx *perf)
0457 {
0458     struct perf_peer *peer;
0459     int ret, pidx, cmd;
0460     u64 data;
0461 
0462     while (!(ret = perf->cmd_recv(perf, &pidx, &cmd, &data))) {
0463         peer = &perf->peers[pidx];
0464 
0465         switch (cmd) {
0466         case PERF_CMD_SSIZE:
0467             peer->inbuf_size = data;
0468             return perf_cmd_exec(peer, PERF_CMD_RSIZE);
0469         case PERF_CMD_SXLAT:
0470             peer->outbuf_xlat = data;
0471             return perf_cmd_exec(peer, PERF_CMD_RXLAT);
0472         default:
0473             dev_err(&perf->ntb->dev, "Recv invalid command\n");
0474             return -EINVAL;
0475         }
0476     }
0477 
0478     /* Return 0 if no data left to process, otherwise an error */
0479     return ret == -ENODATA ? 0 : ret;
0480 }
0481 
0482 static void perf_link_event(void *ctx)
0483 {
0484     struct perf_ctx *perf = ctx;
0485     struct perf_peer *peer;
0486     bool lnk_up;
0487     int pidx;
0488 
0489     for (pidx = 0; pidx < perf->pcnt; pidx++) {
0490         peer = &perf->peers[pidx];
0491 
0492         lnk_up = perf_link_is_up(peer);
0493 
0494         if (lnk_up &&
0495             !test_and_set_bit(PERF_STS_LNKUP, &peer->sts)) {
0496             perf_cmd_exec(peer, PERF_CMD_SSIZE);
0497         } else if (!lnk_up &&
0498                test_and_clear_bit(PERF_STS_LNKUP, &peer->sts)) {
0499             perf_cmd_exec(peer, PERF_CMD_CLEAR);
0500         }
0501     }
0502 }
0503 
0504 static void perf_db_event(void *ctx, int vec)
0505 {
0506     struct perf_ctx *perf = ctx;
0507 
0508     dev_dbg(&perf->ntb->dev, "DB vec %d mask %#llx bits %#llx\n", vec,
0509         ntb_db_vector_mask(perf->ntb, vec), ntb_db_read(perf->ntb));
0510 
0511     /* Just receive all available commands */
0512     (void)perf_cmd_recv(perf);
0513 }
0514 
0515 static void perf_msg_event(void *ctx)
0516 {
0517     struct perf_ctx *perf = ctx;
0518 
0519     dev_dbg(&perf->ntb->dev, "Msg status bits %#llx\n",
0520         ntb_msg_read_sts(perf->ntb));
0521 
0522     /* Messages are only sent one-by-one */
0523     (void)perf_cmd_recv(perf);
0524 }
0525 
0526 static const struct ntb_ctx_ops perf_ops = {
0527     .link_event = perf_link_event,
0528     .db_event = perf_db_event,
0529     .msg_event = perf_msg_event
0530 };
0531 
0532 static void perf_free_outbuf(struct perf_peer *peer)
0533 {
0534     (void)ntb_peer_mw_clear_trans(peer->perf->ntb, peer->pidx, peer->gidx);
0535 }
0536 
0537 static int perf_setup_outbuf(struct perf_peer *peer)
0538 {
0539     struct perf_ctx *perf = peer->perf;
0540     int ret;
0541 
0542     /* Outbuf size can be unaligned due to custom max_mw_size */
0543     ret = ntb_peer_mw_set_trans(perf->ntb, peer->pidx, peer->gidx,
0544                     peer->outbuf_xlat, peer->outbuf_size);
0545     if (ret) {
0546         dev_err(&perf->ntb->dev, "Failed to set outbuf translation\n");
0547         return ret;
0548     }
0549 
0550     /* Initialization is finally done */
0551     set_bit(PERF_STS_DONE, &peer->sts);
0552     complete_all(&peer->init_comp);
0553 
0554     return 0;
0555 }
0556 
0557 static void perf_free_inbuf(struct perf_peer *peer)
0558 {
0559     if (!peer->inbuf)
0560         return;
0561 
0562     (void)ntb_mw_clear_trans(peer->perf->ntb, peer->pidx, peer->gidx);
0563     dma_free_coherent(&peer->perf->ntb->pdev->dev, peer->inbuf_size,
0564               peer->inbuf, peer->inbuf_xlat);
0565     peer->inbuf = NULL;
0566 }
0567 
0568 static int perf_setup_inbuf(struct perf_peer *peer)
0569 {
0570     resource_size_t xlat_align, size_align, size_max;
0571     struct perf_ctx *perf = peer->perf;
0572     int ret;
0573 
0574     /* Get inbound MW parameters */
0575     ret = ntb_mw_get_align(perf->ntb, peer->pidx, perf->gidx,
0576                    &xlat_align, &size_align, &size_max);
0577     if (ret) {
0578         dev_err(&perf->ntb->dev, "Couldn't get inbuf restrictions\n");
0579         return ret;
0580     }
0581 
0582     if (peer->inbuf_size > size_max) {
0583         dev_err(&perf->ntb->dev, "Too big inbuf size %pa > %pa\n",
0584             &peer->inbuf_size, &size_max);
0585         return -EINVAL;
0586     }
0587 
0588     peer->inbuf_size = round_up(peer->inbuf_size, size_align);
0589 
0590     perf_free_inbuf(peer);
0591 
0592     peer->inbuf = dma_alloc_coherent(&perf->ntb->pdev->dev,
0593                      peer->inbuf_size, &peer->inbuf_xlat,
0594                      GFP_KERNEL);
0595     if (!peer->inbuf) {
0596         dev_err(&perf->ntb->dev, "Failed to alloc inbuf of %pa\n",
0597             &peer->inbuf_size);
0598         return -ENOMEM;
0599     }
0600     if (!IS_ALIGNED(peer->inbuf_xlat, xlat_align)) {
0601         ret = -EINVAL;
0602         dev_err(&perf->ntb->dev, "Unaligned inbuf allocated\n");
0603         goto err_free_inbuf;
0604     }
0605 
0606     ret = ntb_mw_set_trans(perf->ntb, peer->pidx, peer->gidx,
0607                    peer->inbuf_xlat, peer->inbuf_size);
0608     if (ret) {
0609         dev_err(&perf->ntb->dev, "Failed to set inbuf translation\n");
0610         goto err_free_inbuf;
0611     }
0612 
0613     /*
0614      * We submit inbuf xlat transmission cmd for execution here to follow
0615      * the code architecture, even though this method is called from service
0616      * work itself so the command will be executed right after it returns.
0617      */
0618     (void)perf_cmd_exec(peer, PERF_CMD_SXLAT);
0619 
0620     return 0;
0621 
0622 err_free_inbuf:
0623     perf_free_inbuf(peer);
0624 
0625     return ret;
0626 }
0627 
0628 static void perf_service_work(struct work_struct *work)
0629 {
0630     struct perf_peer *peer = to_peer_service(work);
0631 
0632     if (test_and_clear_bit(PERF_CMD_SSIZE, &peer->sts))
0633         perf_cmd_send(peer, PERF_CMD_SSIZE, peer->outbuf_size);
0634 
0635     if (test_and_clear_bit(PERF_CMD_RSIZE, &peer->sts))
0636         perf_setup_inbuf(peer);
0637 
0638     if (test_and_clear_bit(PERF_CMD_SXLAT, &peer->sts))
0639         perf_cmd_send(peer, PERF_CMD_SXLAT, peer->inbuf_xlat);
0640 
0641     if (test_and_clear_bit(PERF_CMD_RXLAT, &peer->sts))
0642         perf_setup_outbuf(peer);
0643 
0644     if (test_and_clear_bit(PERF_CMD_CLEAR, &peer->sts)) {
0645         init_completion(&peer->init_comp);
0646         clear_bit(PERF_STS_DONE, &peer->sts);
0647         if (test_bit(0, &peer->perf->busy_flag) &&
0648             peer == peer->perf->test_peer) {
0649             dev_warn(&peer->perf->ntb->dev,
0650                 "Freeing while test on-fly\n");
0651             perf_terminate_test(peer->perf);
0652         }
0653         perf_free_outbuf(peer);
0654         perf_free_inbuf(peer);
0655     }
0656 }
0657 
0658 static int perf_init_service(struct perf_ctx *perf)
0659 {
0660     u64 mask;
0661 
0662     if (ntb_peer_mw_count(perf->ntb) < perf->pcnt) {
0663         dev_err(&perf->ntb->dev, "Not enough memory windows\n");
0664         return -EINVAL;
0665     }
0666 
0667     if (ntb_msg_count(perf->ntb) >= PERF_MSG_CNT) {
0668         perf->cmd_send = perf_msg_cmd_send;
0669         perf->cmd_recv = perf_msg_cmd_recv;
0670 
0671         dev_dbg(&perf->ntb->dev, "Message service initialized\n");
0672 
0673         return 0;
0674     }
0675 
0676     dev_dbg(&perf->ntb->dev, "Message service unsupported\n");
0677 
0678     mask = GENMASK_ULL(perf->pcnt, 0);
0679     if (ntb_spad_count(perf->ntb) >= PERF_SPAD_CNT(perf->pcnt) &&
0680         (ntb_db_valid_mask(perf->ntb) & mask) == mask) {
0681         perf->cmd_send = perf_spad_cmd_send;
0682         perf->cmd_recv = perf_spad_cmd_recv;
0683 
0684         dev_dbg(&perf->ntb->dev, "Scratchpad service initialized\n");
0685 
0686         return 0;
0687     }
0688 
0689     dev_dbg(&perf->ntb->dev, "Scratchpad service unsupported\n");
0690 
0691     dev_err(&perf->ntb->dev, "Command services unsupported\n");
0692 
0693     return -EINVAL;
0694 }
0695 
0696 static int perf_enable_service(struct perf_ctx *perf)
0697 {
0698     u64 mask, incmd_bit;
0699     int ret, sidx, scnt;
0700 
0701     mask = ntb_db_valid_mask(perf->ntb);
0702     (void)ntb_db_set_mask(perf->ntb, mask);
0703 
0704     ret = ntb_set_ctx(perf->ntb, perf, &perf_ops);
0705     if (ret)
0706         return ret;
0707 
0708     if (perf->cmd_send == perf_msg_cmd_send) {
0709         u64 inbits, outbits;
0710 
0711         inbits = ntb_msg_inbits(perf->ntb);
0712         outbits = ntb_msg_outbits(perf->ntb);
0713         (void)ntb_msg_set_mask(perf->ntb, inbits | outbits);
0714 
0715         incmd_bit = BIT_ULL(__ffs64(inbits));
0716         ret = ntb_msg_clear_mask(perf->ntb, incmd_bit);
0717 
0718         dev_dbg(&perf->ntb->dev, "MSG sts unmasked %#llx\n", incmd_bit);
0719     } else {
0720         scnt = ntb_spad_count(perf->ntb);
0721         for (sidx = 0; sidx < scnt; sidx++)
0722             ntb_spad_write(perf->ntb, sidx, PERF_CMD_INVAL);
0723         incmd_bit = PERF_SPAD_NOTIFY(perf->gidx);
0724         ret = ntb_db_clear_mask(perf->ntb, incmd_bit);
0725 
0726         dev_dbg(&perf->ntb->dev, "DB bits unmasked %#llx\n", incmd_bit);
0727     }
0728     if (ret) {
0729         ntb_clear_ctx(perf->ntb);
0730         return ret;
0731     }
0732 
0733     ntb_link_enable(perf->ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
0734     /* Might be not necessary */
0735     ntb_link_event(perf->ntb);
0736 
0737     return 0;
0738 }
0739 
0740 static void perf_disable_service(struct perf_ctx *perf)
0741 {
0742     int pidx;
0743 
0744     if (perf->cmd_send == perf_msg_cmd_send) {
0745         u64 inbits;
0746 
0747         inbits = ntb_msg_inbits(perf->ntb);
0748         (void)ntb_msg_set_mask(perf->ntb, inbits);
0749     } else {
0750         (void)ntb_db_set_mask(perf->ntb, PERF_SPAD_NOTIFY(perf->gidx));
0751     }
0752 
0753     ntb_clear_ctx(perf->ntb);
0754 
0755     for (pidx = 0; pidx < perf->pcnt; pidx++)
0756         perf_cmd_exec(&perf->peers[pidx], PERF_CMD_CLEAR);
0757 
0758     for (pidx = 0; pidx < perf->pcnt; pidx++)
0759         flush_work(&perf->peers[pidx].service);
0760 
0761     for (pidx = 0; pidx < perf->pcnt; pidx++) {
0762         struct perf_peer *peer = &perf->peers[pidx];
0763 
0764         ntb_spad_write(perf->ntb, PERF_SPAD_CMD(peer->gidx), 0);
0765     }
0766 
0767     ntb_db_clear(perf->ntb, PERF_SPAD_NOTIFY(perf->gidx));
0768 
0769     ntb_link_disable(perf->ntb);
0770 }
0771 
0772 /*==============================================================================
0773  *                      Performance measuring work-thread
0774  *==============================================================================
0775  */
0776 
0777 static void perf_dma_copy_callback(void *data)
0778 {
0779     struct perf_thread *pthr = data;
0780 
0781     atomic_dec(&pthr->dma_sync);
0782     wake_up(&pthr->dma_wait);
0783 }
0784 
0785 static int perf_copy_chunk(struct perf_thread *pthr,
0786                void __iomem *dst, void *src, size_t len)
0787 {
0788     struct dma_async_tx_descriptor *tx;
0789     struct dmaengine_unmap_data *unmap;
0790     struct device *dma_dev;
0791     int try = 0, ret = 0;
0792     struct perf_peer *peer = pthr->perf->test_peer;
0793     void __iomem *vbase;
0794     void __iomem *dst_vaddr;
0795     dma_addr_t dst_dma_addr;
0796 
0797     if (!use_dma) {
0798         memcpy_toio(dst, src, len);
0799         goto ret_check_tsync;
0800     }
0801 
0802     dma_dev = pthr->dma_chan->device->dev;
0803 
0804     if (!is_dma_copy_aligned(pthr->dma_chan->device, offset_in_page(src),
0805                  offset_in_page(dst), len))
0806         return -EIO;
0807 
0808     vbase = peer->outbuf;
0809     dst_vaddr = dst;
0810     dst_dma_addr = peer->dma_dst_addr + (dst_vaddr - vbase);
0811 
0812     unmap = dmaengine_get_unmap_data(dma_dev, 1, GFP_NOWAIT);
0813     if (!unmap)
0814         return -ENOMEM;
0815 
0816     unmap->len = len;
0817     unmap->addr[0] = dma_map_page(dma_dev, virt_to_page(src),
0818         offset_in_page(src), len, DMA_TO_DEVICE);
0819     if (dma_mapping_error(dma_dev, unmap->addr[0])) {
0820         ret = -EIO;
0821         goto err_free_resource;
0822     }
0823     unmap->to_cnt = 1;
0824 
0825     do {
0826         tx = dmaengine_prep_dma_memcpy(pthr->dma_chan, dst_dma_addr,
0827             unmap->addr[0], len, DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
0828         if (!tx)
0829             msleep(DMA_MDELAY);
0830     } while (!tx && (try++ < DMA_TRIES));
0831 
0832     if (!tx) {
0833         ret = -EIO;
0834         goto err_free_resource;
0835     }
0836 
0837     tx->callback = perf_dma_copy_callback;
0838     tx->callback_param = pthr;
0839     dma_set_unmap(tx, unmap);
0840 
0841     ret = dma_submit_error(dmaengine_submit(tx));
0842     if (ret) {
0843         dmaengine_unmap_put(unmap);
0844         goto err_free_resource;
0845     }
0846 
0847     dmaengine_unmap_put(unmap);
0848 
0849     atomic_inc(&pthr->dma_sync);
0850     dma_async_issue_pending(pthr->dma_chan);
0851 
0852 ret_check_tsync:
0853     return likely(atomic_read(&pthr->perf->tsync) > 0) ? 0 : -EINTR;
0854 
0855 err_free_resource:
0856     dmaengine_unmap_put(unmap);
0857 
0858     return ret;
0859 }
0860 
0861 static bool perf_dma_filter(struct dma_chan *chan, void *data)
0862 {
0863     struct perf_ctx *perf = data;
0864     int node;
0865 
0866     node = dev_to_node(&perf->ntb->dev);
0867 
0868     return node == NUMA_NO_NODE || node == dev_to_node(chan->device->dev);
0869 }
0870 
0871 static int perf_init_test(struct perf_thread *pthr)
0872 {
0873     struct perf_ctx *perf = pthr->perf;
0874     dma_cap_mask_t dma_mask;
0875     struct perf_peer *peer = pthr->perf->test_peer;
0876 
0877     pthr->src = kmalloc_node(perf->test_peer->outbuf_size, GFP_KERNEL,
0878                  dev_to_node(&perf->ntb->dev));
0879     if (!pthr->src)
0880         return -ENOMEM;
0881 
0882     get_random_bytes(pthr->src, perf->test_peer->outbuf_size);
0883 
0884     if (!use_dma)
0885         return 0;
0886 
0887     dma_cap_zero(dma_mask);
0888     dma_cap_set(DMA_MEMCPY, dma_mask);
0889     pthr->dma_chan = dma_request_channel(dma_mask, perf_dma_filter, perf);
0890     if (!pthr->dma_chan) {
0891         dev_err(&perf->ntb->dev, "%d: Failed to get DMA channel\n",
0892             pthr->tidx);
0893         goto err_free;
0894     }
0895     peer->dma_dst_addr =
0896         dma_map_resource(pthr->dma_chan->device->dev,
0897                  peer->out_phys_addr, peer->outbuf_size,
0898                  DMA_FROM_DEVICE, 0);
0899     if (dma_mapping_error(pthr->dma_chan->device->dev,
0900                   peer->dma_dst_addr)) {
0901         dev_err(pthr->dma_chan->device->dev, "%d: Failed to map DMA addr\n",
0902             pthr->tidx);
0903         peer->dma_dst_addr = 0;
0904         dma_release_channel(pthr->dma_chan);
0905         goto err_free;
0906     }
0907     dev_dbg(pthr->dma_chan->device->dev, "%d: Map MMIO %pa to DMA addr %pad\n",
0908             pthr->tidx,
0909             &peer->out_phys_addr,
0910             &peer->dma_dst_addr);
0911 
0912     atomic_set(&pthr->dma_sync, 0);
0913     return 0;
0914 
0915 err_free:
0916     atomic_dec(&perf->tsync);
0917     wake_up(&perf->twait);
0918     kfree(pthr->src);
0919     return -ENODEV;
0920 }
0921 
0922 static int perf_run_test(struct perf_thread *pthr)
0923 {
0924     struct perf_peer *peer = pthr->perf->test_peer;
0925     struct perf_ctx *perf = pthr->perf;
0926     void __iomem *flt_dst, *bnd_dst;
0927     u64 total_size, chunk_size;
0928     void *flt_src;
0929     int ret = 0;
0930 
0931     total_size = 1ULL << total_order;
0932     chunk_size = 1ULL << chunk_order;
0933     chunk_size = min_t(u64, peer->outbuf_size, chunk_size);
0934 
0935     flt_src = pthr->src;
0936     bnd_dst = peer->outbuf + peer->outbuf_size;
0937     flt_dst = peer->outbuf;
0938 
0939     pthr->duration = ktime_get();
0940 
0941     /* Copied field is cleared on test launch stage */
0942     while (pthr->copied < total_size) {
0943         ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size);
0944         if (ret) {
0945             dev_err(&perf->ntb->dev, "%d: Got error %d on test\n",
0946                 pthr->tidx, ret);
0947             return ret;
0948         }
0949 
0950         pthr->copied += chunk_size;
0951 
0952         flt_dst += chunk_size;
0953         flt_src += chunk_size;
0954         if (flt_dst >= bnd_dst || flt_dst < peer->outbuf) {
0955             flt_dst = peer->outbuf;
0956             flt_src = pthr->src;
0957         }
0958 
0959         /* Give up CPU to give a chance for other threads to use it */
0960         schedule();
0961     }
0962 
0963     return 0;
0964 }
0965 
0966 static int perf_sync_test(struct perf_thread *pthr)
0967 {
0968     struct perf_ctx *perf = pthr->perf;
0969 
0970     if (!use_dma)
0971         goto no_dma_ret;
0972 
0973     wait_event(pthr->dma_wait,
0974            (atomic_read(&pthr->dma_sync) == 0 ||
0975             atomic_read(&perf->tsync) < 0));
0976 
0977     if (atomic_read(&perf->tsync) < 0)
0978         return -EINTR;
0979 
0980 no_dma_ret:
0981     pthr->duration = ktime_sub(ktime_get(), pthr->duration);
0982 
0983     dev_dbg(&perf->ntb->dev, "%d: copied %llu bytes\n",
0984         pthr->tidx, pthr->copied);
0985 
0986     dev_dbg(&perf->ntb->dev, "%d: lasted %llu usecs\n",
0987         pthr->tidx, ktime_to_us(pthr->duration));
0988 
0989     dev_dbg(&perf->ntb->dev, "%d: %llu MBytes/s\n", pthr->tidx,
0990         div64_u64(pthr->copied, ktime_to_us(pthr->duration)));
0991 
0992     return 0;
0993 }
0994 
0995 static void perf_clear_test(struct perf_thread *pthr)
0996 {
0997     struct perf_ctx *perf = pthr->perf;
0998 
0999     if (!use_dma)
1000         goto no_dma_notify;
1001 
1002     /*
1003      * If test finished without errors, termination isn't needed.
1004      * We call it anyway just to be sure of the transfers completion.
1005      */
1006     (void)dmaengine_terminate_sync(pthr->dma_chan);
1007     if (pthr->perf->test_peer->dma_dst_addr)
1008         dma_unmap_resource(pthr->dma_chan->device->dev,
1009                    pthr->perf->test_peer->dma_dst_addr,
1010                    pthr->perf->test_peer->outbuf_size,
1011                    DMA_FROM_DEVICE, 0);
1012 
1013     dma_release_channel(pthr->dma_chan);
1014 
1015 no_dma_notify:
1016     atomic_dec(&perf->tsync);
1017     wake_up(&perf->twait);
1018     kfree(pthr->src);
1019 }
1020 
1021 static void perf_thread_work(struct work_struct *work)
1022 {
1023     struct perf_thread *pthr = to_thread_work(work);
1024     int ret;
1025 
1026     /*
1027      * Perform stages in compliance with use_dma flag value.
1028      * Test status is changed only if error happened, otherwise
1029      * status -ENODATA is kept while test is on-fly. Results
1030      * synchronization is performed only if test fininshed
1031      * without an error or interruption.
1032      */
1033     ret = perf_init_test(pthr);
1034     if (ret) {
1035         pthr->status = ret;
1036         return;
1037     }
1038 
1039     ret = perf_run_test(pthr);
1040     if (ret) {
1041         pthr->status = ret;
1042         goto err_clear_test;
1043     }
1044 
1045     pthr->status = perf_sync_test(pthr);
1046 
1047 err_clear_test:
1048     perf_clear_test(pthr);
1049 }
1050 
1051 static int perf_set_tcnt(struct perf_ctx *perf, u8 tcnt)
1052 {
1053     if (tcnt == 0 || tcnt > MAX_THREADS_CNT)
1054         return -EINVAL;
1055 
1056     if (test_and_set_bit_lock(0, &perf->busy_flag))
1057         return -EBUSY;
1058 
1059     perf->tcnt = tcnt;
1060 
1061     clear_bit_unlock(0, &perf->busy_flag);
1062 
1063     return 0;
1064 }
1065 
1066 static void perf_terminate_test(struct perf_ctx *perf)
1067 {
1068     int tidx;
1069 
1070     atomic_set(&perf->tsync, -1);
1071     wake_up(&perf->twait);
1072 
1073     for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
1074         wake_up(&perf->threads[tidx].dma_wait);
1075         cancel_work_sync(&perf->threads[tidx].work);
1076     }
1077 }
1078 
1079 static int perf_submit_test(struct perf_peer *peer)
1080 {
1081     struct perf_ctx *perf = peer->perf;
1082     struct perf_thread *pthr;
1083     int tidx, ret;
1084 
1085     ret = wait_for_completion_interruptible(&peer->init_comp);
1086     if (ret < 0)
1087         return ret;
1088 
1089     if (test_and_set_bit_lock(0, &perf->busy_flag))
1090         return -EBUSY;
1091 
1092     perf->test_peer = peer;
1093     atomic_set(&perf->tsync, perf->tcnt);
1094 
1095     for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
1096         pthr = &perf->threads[tidx];
1097 
1098         pthr->status = -ENODATA;
1099         pthr->copied = 0;
1100         pthr->duration = ktime_set(0, 0);
1101         if (tidx < perf->tcnt)
1102             (void)queue_work(perf_wq, &pthr->work);
1103     }
1104 
1105     ret = wait_event_interruptible(perf->twait,
1106                        atomic_read(&perf->tsync) <= 0);
1107     if (ret == -ERESTARTSYS) {
1108         perf_terminate_test(perf);
1109         ret = -EINTR;
1110     }
1111 
1112     clear_bit_unlock(0, &perf->busy_flag);
1113 
1114     return ret;
1115 }
1116 
1117 static int perf_read_stats(struct perf_ctx *perf, char *buf,
1118                size_t size, ssize_t *pos)
1119 {
1120     struct perf_thread *pthr;
1121     int tidx;
1122 
1123     if (test_and_set_bit_lock(0, &perf->busy_flag))
1124         return -EBUSY;
1125 
1126     (*pos) += scnprintf(buf + *pos, size - *pos,
1127         "    Peer %d test statistics:\n", perf->test_peer->pidx);
1128 
1129     for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
1130         pthr = &perf->threads[tidx];
1131 
1132         if (pthr->status == -ENODATA)
1133             continue;
1134 
1135         if (pthr->status) {
1136             (*pos) += scnprintf(buf + *pos, size - *pos,
1137                 "%d: error status %d\n", tidx, pthr->status);
1138             continue;
1139         }
1140 
1141         (*pos) += scnprintf(buf + *pos, size - *pos,
1142             "%d: copied %llu bytes in %llu usecs, %llu MBytes/s\n",
1143             tidx, pthr->copied, ktime_to_us(pthr->duration),
1144             div64_u64(pthr->copied, ktime_to_us(pthr->duration)));
1145     }
1146 
1147     clear_bit_unlock(0, &perf->busy_flag);
1148 
1149     return 0;
1150 }
1151 
1152 static void perf_init_threads(struct perf_ctx *perf)
1153 {
1154     struct perf_thread *pthr;
1155     int tidx;
1156 
1157     perf->tcnt = DEF_THREADS_CNT;
1158     perf->test_peer = &perf->peers[0];
1159     init_waitqueue_head(&perf->twait);
1160 
1161     for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
1162         pthr = &perf->threads[tidx];
1163 
1164         pthr->perf = perf;
1165         pthr->tidx = tidx;
1166         pthr->status = -ENODATA;
1167         init_waitqueue_head(&pthr->dma_wait);
1168         INIT_WORK(&pthr->work, perf_thread_work);
1169     }
1170 }
1171 
1172 static void perf_clear_threads(struct perf_ctx *perf)
1173 {
1174     perf_terminate_test(perf);
1175 }
1176 
1177 /*==============================================================================
1178  *                               DebugFS nodes
1179  *==============================================================================
1180  */
1181 
1182 static ssize_t perf_dbgfs_read_info(struct file *filep, char __user *ubuf,
1183                     size_t size, loff_t *offp)
1184 {
1185     struct perf_ctx *perf = filep->private_data;
1186     struct perf_peer *peer;
1187     size_t buf_size;
1188     ssize_t pos = 0;
1189     int ret, pidx;
1190     char *buf;
1191 
1192     buf_size = min_t(size_t, size, 0x1000U);
1193 
1194     buf = kmalloc(buf_size, GFP_KERNEL);
1195     if (!buf)
1196         return -ENOMEM;
1197 
1198     pos += scnprintf(buf + pos, buf_size - pos,
1199         "    Performance measuring tool info:\n\n");
1200 
1201     pos += scnprintf(buf + pos, buf_size - pos,
1202         "Local port %d, Global index %d\n", ntb_port_number(perf->ntb),
1203         perf->gidx);
1204     pos += scnprintf(buf + pos, buf_size - pos, "Test status: ");
1205     if (test_bit(0, &perf->busy_flag)) {
1206         pos += scnprintf(buf + pos, buf_size - pos,
1207             "on-fly with port %d (%d)\n",
1208             ntb_peer_port_number(perf->ntb, perf->test_peer->pidx),
1209             perf->test_peer->pidx);
1210     } else {
1211         pos += scnprintf(buf + pos, buf_size - pos, "idle\n");
1212     }
1213 
1214     for (pidx = 0; pidx < perf->pcnt; pidx++) {
1215         peer = &perf->peers[pidx];
1216 
1217         pos += scnprintf(buf + pos, buf_size - pos,
1218             "Port %d (%d), Global index %d:\n",
1219             ntb_peer_port_number(perf->ntb, peer->pidx), peer->pidx,
1220             peer->gidx);
1221 
1222         pos += scnprintf(buf + pos, buf_size - pos,
1223             "\tLink status: %s\n",
1224             test_bit(PERF_STS_LNKUP, &peer->sts) ? "up" : "down");
1225 
1226         pos += scnprintf(buf + pos, buf_size - pos,
1227             "\tOut buffer addr 0x%pK\n", peer->outbuf);
1228 
1229         pos += scnprintf(buf + pos, buf_size - pos,
1230             "\tOut buff phys addr %pa[p]\n", &peer->out_phys_addr);
1231 
1232         pos += scnprintf(buf + pos, buf_size - pos,
1233             "\tOut buffer size %pa\n", &peer->outbuf_size);
1234 
1235         pos += scnprintf(buf + pos, buf_size - pos,
1236             "\tOut buffer xlat 0x%016llx[p]\n", peer->outbuf_xlat);
1237 
1238         if (!peer->inbuf) {
1239             pos += scnprintf(buf + pos, buf_size - pos,
1240                 "\tIn buffer addr: unallocated\n");
1241             continue;
1242         }
1243 
1244         pos += scnprintf(buf + pos, buf_size - pos,
1245             "\tIn buffer addr 0x%pK\n", peer->inbuf);
1246 
1247         pos += scnprintf(buf + pos, buf_size - pos,
1248             "\tIn buffer size %pa\n", &peer->inbuf_size);
1249 
1250         pos += scnprintf(buf + pos, buf_size - pos,
1251             "\tIn buffer xlat %pad[p]\n", &peer->inbuf_xlat);
1252     }
1253 
1254     ret = simple_read_from_buffer(ubuf, size, offp, buf, pos);
1255     kfree(buf);
1256 
1257     return ret;
1258 }
1259 
1260 static const struct file_operations perf_dbgfs_info = {
1261     .open = simple_open,
1262     .read = perf_dbgfs_read_info
1263 };
1264 
1265 static ssize_t perf_dbgfs_read_run(struct file *filep, char __user *ubuf,
1266                    size_t size, loff_t *offp)
1267 {
1268     struct perf_ctx *perf = filep->private_data;
1269     ssize_t ret, pos = 0;
1270     char *buf;
1271 
1272     buf = kmalloc(PERF_BUF_LEN, GFP_KERNEL);
1273     if (!buf)
1274         return -ENOMEM;
1275 
1276     ret = perf_read_stats(perf, buf, PERF_BUF_LEN, &pos);
1277     if (ret)
1278         goto err_free;
1279 
1280     ret = simple_read_from_buffer(ubuf, size, offp, buf, pos);
1281 err_free:
1282     kfree(buf);
1283 
1284     return ret;
1285 }
1286 
1287 static ssize_t perf_dbgfs_write_run(struct file *filep, const char __user *ubuf,
1288                     size_t size, loff_t *offp)
1289 {
1290     struct perf_ctx *perf = filep->private_data;
1291     struct perf_peer *peer;
1292     int pidx, ret;
1293 
1294     ret = kstrtoint_from_user(ubuf, size, 0, &pidx);
1295     if (ret)
1296         return ret;
1297 
1298     if (pidx < 0 || pidx >= perf->pcnt)
1299         return -EINVAL;
1300 
1301     peer = &perf->peers[pidx];
1302 
1303     ret = perf_submit_test(peer);
1304     if (ret)
1305         return ret;
1306 
1307     return size;
1308 }
1309 
1310 static const struct file_operations perf_dbgfs_run = {
1311     .open = simple_open,
1312     .read = perf_dbgfs_read_run,
1313     .write = perf_dbgfs_write_run
1314 };
1315 
1316 static ssize_t perf_dbgfs_read_tcnt(struct file *filep, char __user *ubuf,
1317                     size_t size, loff_t *offp)
1318 {
1319     struct perf_ctx *perf = filep->private_data;
1320     char buf[8];
1321     ssize_t pos;
1322 
1323     pos = scnprintf(buf, sizeof(buf), "%hhu\n", perf->tcnt);
1324 
1325     return simple_read_from_buffer(ubuf, size, offp, buf, pos);
1326 }
1327 
1328 static ssize_t perf_dbgfs_write_tcnt(struct file *filep,
1329                      const char __user *ubuf,
1330                      size_t size, loff_t *offp)
1331 {
1332     struct perf_ctx *perf = filep->private_data;
1333     int ret;
1334     u8 val;
1335 
1336     ret = kstrtou8_from_user(ubuf, size, 0, &val);
1337     if (ret)
1338         return ret;
1339 
1340     ret = perf_set_tcnt(perf, val);
1341     if (ret)
1342         return ret;
1343 
1344     return size;
1345 }
1346 
1347 static const struct file_operations perf_dbgfs_tcnt = {
1348     .open = simple_open,
1349     .read = perf_dbgfs_read_tcnt,
1350     .write = perf_dbgfs_write_tcnt
1351 };
1352 
1353 static void perf_setup_dbgfs(struct perf_ctx *perf)
1354 {
1355     struct pci_dev *pdev = perf->ntb->pdev;
1356 
1357     perf->dbgfs_dir = debugfs_create_dir(pci_name(pdev), perf_dbgfs_topdir);
1358     if (!perf->dbgfs_dir) {
1359         dev_warn(&perf->ntb->dev, "DebugFS unsupported\n");
1360         return;
1361     }
1362 
1363     debugfs_create_file("info", 0600, perf->dbgfs_dir, perf,
1364                 &perf_dbgfs_info);
1365 
1366     debugfs_create_file("run", 0600, perf->dbgfs_dir, perf,
1367                 &perf_dbgfs_run);
1368 
1369     debugfs_create_file("threads_count", 0600, perf->dbgfs_dir, perf,
1370                 &perf_dbgfs_tcnt);
1371 
1372     /* They are made read-only for test exec safety and integrity */
1373     debugfs_create_u8("chunk_order", 0500, perf->dbgfs_dir, &chunk_order);
1374 
1375     debugfs_create_u8("total_order", 0500, perf->dbgfs_dir, &total_order);
1376 
1377     debugfs_create_bool("use_dma", 0500, perf->dbgfs_dir, &use_dma);
1378 }
1379 
1380 static void perf_clear_dbgfs(struct perf_ctx *perf)
1381 {
1382     debugfs_remove_recursive(perf->dbgfs_dir);
1383 }
1384 
1385 /*==============================================================================
1386  *                        Basic driver initialization
1387  *==============================================================================
1388  */
1389 
1390 static struct perf_ctx *perf_create_data(struct ntb_dev *ntb)
1391 {
1392     struct perf_ctx *perf;
1393 
1394     perf = devm_kzalloc(&ntb->dev, sizeof(*perf), GFP_KERNEL);
1395     if (!perf)
1396         return ERR_PTR(-ENOMEM);
1397 
1398     perf->pcnt = ntb_peer_port_count(ntb);
1399     perf->peers = devm_kcalloc(&ntb->dev, perf->pcnt, sizeof(*perf->peers),
1400                   GFP_KERNEL);
1401     if (!perf->peers)
1402         return ERR_PTR(-ENOMEM);
1403 
1404     perf->ntb = ntb;
1405 
1406     return perf;
1407 }
1408 
1409 static int perf_setup_peer_mw(struct perf_peer *peer)
1410 {
1411     struct perf_ctx *perf = peer->perf;
1412     phys_addr_t phys_addr;
1413     int ret;
1414 
1415     /* Get outbound MW parameters and map it */
1416     ret = ntb_peer_mw_get_addr(perf->ntb, perf->gidx, &phys_addr,
1417                    &peer->outbuf_size);
1418     if (ret)
1419         return ret;
1420 
1421     peer->outbuf = devm_ioremap_wc(&perf->ntb->dev, phys_addr,
1422                     peer->outbuf_size);
1423     if (!peer->outbuf)
1424         return -ENOMEM;
1425 
1426     peer->out_phys_addr = phys_addr;
1427 
1428     if (max_mw_size && peer->outbuf_size > max_mw_size) {
1429         peer->outbuf_size = max_mw_size;
1430         dev_warn(&peer->perf->ntb->dev,
1431             "Peer %d outbuf reduced to %pa\n", peer->pidx,
1432             &peer->outbuf_size);
1433     }
1434 
1435     return 0;
1436 }
1437 
1438 static int perf_init_peers(struct perf_ctx *perf)
1439 {
1440     struct perf_peer *peer;
1441     int pidx, lport, ret;
1442 
1443     lport = ntb_port_number(perf->ntb);
1444     perf->gidx = -1;
1445     for (pidx = 0; pidx < perf->pcnt; pidx++) {
1446         peer = &perf->peers[pidx];
1447 
1448         peer->perf = perf;
1449         peer->pidx = pidx;
1450         if (lport < ntb_peer_port_number(perf->ntb, pidx)) {
1451             if (perf->gidx == -1)
1452                 perf->gidx = pidx;
1453             peer->gidx = pidx + 1;
1454         } else {
1455             peer->gidx = pidx;
1456         }
1457         INIT_WORK(&peer->service, perf_service_work);
1458         init_completion(&peer->init_comp);
1459     }
1460     if (perf->gidx == -1)
1461         perf->gidx = pidx;
1462 
1463     /*
1464      * Hardware with only two ports may not have unique port
1465      * numbers. In this case, the gidxs should all be zero.
1466      */
1467     if (perf->pcnt == 1 &&  ntb_port_number(perf->ntb) == 0 &&
1468         ntb_peer_port_number(perf->ntb, 0) == 0) {
1469         perf->gidx = 0;
1470         perf->peers[0].gidx = 0;
1471     }
1472 
1473     for (pidx = 0; pidx < perf->pcnt; pidx++) {
1474         ret = perf_setup_peer_mw(&perf->peers[pidx]);
1475         if (ret)
1476             return ret;
1477     }
1478 
1479     dev_dbg(&perf->ntb->dev, "Global port index %d\n", perf->gidx);
1480 
1481     return 0;
1482 }
1483 
1484 static int perf_probe(struct ntb_client *client, struct ntb_dev *ntb)
1485 {
1486     struct perf_ctx *perf;
1487     int ret;
1488 
1489     perf = perf_create_data(ntb);
1490     if (IS_ERR(perf))
1491         return PTR_ERR(perf);
1492 
1493     ret = perf_init_peers(perf);
1494     if (ret)
1495         return ret;
1496 
1497     perf_init_threads(perf);
1498 
1499     ret = perf_init_service(perf);
1500     if (ret)
1501         return ret;
1502 
1503     ret = perf_enable_service(perf);
1504     if (ret)
1505         return ret;
1506 
1507     perf_setup_dbgfs(perf);
1508 
1509     return 0;
1510 }
1511 
1512 static void perf_remove(struct ntb_client *client, struct ntb_dev *ntb)
1513 {
1514     struct perf_ctx *perf = ntb->ctx;
1515 
1516     perf_clear_dbgfs(perf);
1517 
1518     perf_disable_service(perf);
1519 
1520     perf_clear_threads(perf);
1521 }
1522 
1523 static struct ntb_client perf_client = {
1524     .ops = {
1525         .probe = perf_probe,
1526         .remove = perf_remove
1527     }
1528 };
1529 
1530 static int __init perf_init(void)
1531 {
1532     int ret;
1533 
1534     if (chunk_order > MAX_CHUNK_ORDER) {
1535         chunk_order = MAX_CHUNK_ORDER;
1536         pr_info("Chunk order reduced to %hhu\n", chunk_order);
1537     }
1538 
1539     if (total_order < chunk_order) {
1540         total_order = chunk_order;
1541         pr_info("Total data order reduced to %hhu\n", total_order);
1542     }
1543 
1544     perf_wq = alloc_workqueue("perf_wq", WQ_UNBOUND | WQ_SYSFS, 0);
1545     if (!perf_wq)
1546         return -ENOMEM;
1547 
1548     if (debugfs_initialized())
1549         perf_dbgfs_topdir = debugfs_create_dir(KBUILD_MODNAME, NULL);
1550 
1551     ret = ntb_register_client(&perf_client);
1552     if (ret) {
1553         debugfs_remove_recursive(perf_dbgfs_topdir);
1554         destroy_workqueue(perf_wq);
1555     }
1556 
1557     return ret;
1558 }
1559 module_init(perf_init);
1560 
1561 static void __exit perf_exit(void)
1562 {
1563     ntb_unregister_client(&perf_client);
1564     debugfs_remove_recursive(perf_dbgfs_topdir);
1565     destroy_workqueue(perf_wq);
1566 }
1567 module_exit(perf_exit);
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!