Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / dma / dmatest.c
diff --git a/kernel/drivers/dma/dmatest.c b/kernel/drivers/dma/dmatest.c
new file mode 100644 (file)
index 0000000..220ee49
--- /dev/null
@@ -0,0 +1,987 @@
+/*
+ * DMA Engine test module
+ *
+ * Copyright (C) 2007 Atmel Corporation
+ * Copyright (C) 2013 Intel Corporation
+ *
+ * This program 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.
+ */
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/delay.h>
+#include <linux/dma-mapping.h>
+#include <linux/dmaengine.h>
+#include <linux/freezer.h>
+#include <linux/init.h>
+#include <linux/kthread.h>
+#include <linux/module.h>
+#include <linux/moduleparam.h>
+#include <linux/random.h>
+#include <linux/slab.h>
+#include <linux/wait.h>
+
+static unsigned int test_buf_size = 16384;
+module_param(test_buf_size, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(test_buf_size, "Size of the memcpy test buffer");
+
+static char test_channel[20];
+module_param_string(channel, test_channel, sizeof(test_channel),
+               S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(channel, "Bus ID of the channel to test (default: any)");
+
+static char test_device[32];
+module_param_string(device, test_device, sizeof(test_device),
+               S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(device, "Bus ID of the DMA Engine to test (default: any)");
+
+static unsigned int threads_per_chan = 1;
+module_param(threads_per_chan, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(threads_per_chan,
+               "Number of threads to start per channel (default: 1)");
+
+static unsigned int max_channels;
+module_param(max_channels, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(max_channels,
+               "Maximum number of channels to use (default: all)");
+
+static unsigned int iterations;
+module_param(iterations, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(iterations,
+               "Iterations before stopping test (default: infinite)");
+
+static unsigned int xor_sources = 3;
+module_param(xor_sources, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(xor_sources,
+               "Number of xor source buffers (default: 3)");
+
+static unsigned int pq_sources = 3;
+module_param(pq_sources, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(pq_sources,
+               "Number of p+q source buffers (default: 3)");
+
+static int timeout = 3000;
+module_param(timeout, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(timeout, "Transfer Timeout in msec (default: 3000), "
+                "Pass -1 for infinite timeout");
+
+static bool noverify;
+module_param(noverify, bool, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(noverify, "Disable random data setup and verification");
+
+static bool verbose;
+module_param(verbose, bool, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(verbose, "Enable \"success\" result messages (default: off)");
+
+/**
+ * struct dmatest_params - test parameters.
+ * @buf_size:          size of the memcpy test buffer
+ * @channel:           bus ID of the channel to test
+ * @device:            bus ID of the DMA Engine to test
+ * @threads_per_chan:  number of threads to start per channel
+ * @max_channels:      maximum number of channels to use
+ * @iterations:                iterations before stopping test
+ * @xor_sources:       number of xor source buffers
+ * @pq_sources:                number of p+q source buffers
+ * @timeout:           transfer timeout in msec, -1 for infinite timeout
+ */
+struct dmatest_params {
+       unsigned int    buf_size;
+       char            channel[20];
+       char            device[32];
+       unsigned int    threads_per_chan;
+       unsigned int    max_channels;
+       unsigned int    iterations;
+       unsigned int    xor_sources;
+       unsigned int    pq_sources;
+       int             timeout;
+       bool            noverify;
+};
+
+/**
+ * struct dmatest_info - test information.
+ * @params:            test parameters
+ * @lock:              access protection to the fields of this structure
+ */
+static struct dmatest_info {
+       /* Test parameters */
+       struct dmatest_params   params;
+
+       /* Internal state */
+       struct list_head        channels;
+       unsigned int            nr_channels;
+       struct mutex            lock;
+       bool                    did_init;
+} test_info = {
+       .channels = LIST_HEAD_INIT(test_info.channels),
+       .lock = __MUTEX_INITIALIZER(test_info.lock),
+};
+
+static int dmatest_run_set(const char *val, const struct kernel_param *kp);
+static int dmatest_run_get(char *val, const struct kernel_param *kp);
+static struct kernel_param_ops run_ops = {
+       .set = dmatest_run_set,
+       .get = dmatest_run_get,
+};
+static bool dmatest_run;
+module_param_cb(run, &run_ops, &dmatest_run, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(run, "Run the test (default: false)");
+
+/* Maximum amount of mismatched bytes in buffer to print */
+#define MAX_ERROR_COUNT                32
+
+/*
+ * Initialization patterns. All bytes in the source buffer has bit 7
+ * set, all bytes in the destination buffer has bit 7 cleared.
+ *
+ * Bit 6 is set for all bytes which are to be copied by the DMA
+ * engine. Bit 5 is set for all bytes which are to be overwritten by
+ * the DMA engine.
+ *
+ * The remaining bits are the inverse of a counter which increments by
+ * one for each byte address.
+ */
+#define PATTERN_SRC            0x80
+#define PATTERN_DST            0x00
+#define PATTERN_COPY           0x40
+#define PATTERN_OVERWRITE      0x20
+#define PATTERN_COUNT_MASK     0x1f
+
+struct dmatest_thread {
+       struct list_head        node;
+       struct dmatest_info     *info;
+       struct task_struct      *task;
+       struct dma_chan         *chan;
+       u8                      **srcs;
+       u8                      **dsts;
+       enum dma_transaction_type type;
+       bool                    done;
+};
+
+struct dmatest_chan {
+       struct list_head        node;
+       struct dma_chan         *chan;
+       struct list_head        threads;
+};
+
+static DECLARE_WAIT_QUEUE_HEAD(thread_wait);
+static bool wait;
+
+static bool is_threaded_test_run(struct dmatest_info *info)
+{
+       struct dmatest_chan *dtc;
+
+       list_for_each_entry(dtc, &info->channels, node) {
+               struct dmatest_thread *thread;
+
+               list_for_each_entry(thread, &dtc->threads, node) {
+                       if (!thread->done)
+                               return true;
+               }
+       }
+
+       return false;
+}
+
+static int dmatest_wait_get(char *val, const struct kernel_param *kp)
+{
+       struct dmatest_info *info = &test_info;
+       struct dmatest_params *params = &info->params;
+
+       if (params->iterations)
+               wait_event(thread_wait, !is_threaded_test_run(info));
+       wait = true;
+       return param_get_bool(val, kp);
+}
+
+static struct kernel_param_ops wait_ops = {
+       .get = dmatest_wait_get,
+       .set = param_set_bool,
+};
+module_param_cb(wait, &wait_ops, &wait, S_IRUGO);
+MODULE_PARM_DESC(wait, "Wait for tests to complete (default: false)");
+
+static bool dmatest_match_channel(struct dmatest_params *params,
+               struct dma_chan *chan)
+{
+       if (params->channel[0] == '\0')
+               return true;
+       return strcmp(dma_chan_name(chan), params->channel) == 0;
+}
+
+static bool dmatest_match_device(struct dmatest_params *params,
+               struct dma_device *device)
+{
+       if (params->device[0] == '\0')
+               return true;
+       return strcmp(dev_name(device->dev), params->device) == 0;
+}
+
+static unsigned long dmatest_random(void)
+{
+       unsigned long buf;
+
+       prandom_bytes(&buf, sizeof(buf));
+       return buf;
+}
+
+static void dmatest_init_srcs(u8 **bufs, unsigned int start, unsigned int len,
+               unsigned int buf_size)
+{
+       unsigned int i;
+       u8 *buf;
+
+       for (; (buf = *bufs); bufs++) {
+               for (i = 0; i < start; i++)
+                       buf[i] = PATTERN_SRC | (~i & PATTERN_COUNT_MASK);
+               for ( ; i < start + len; i++)
+                       buf[i] = PATTERN_SRC | PATTERN_COPY
+                               | (~i & PATTERN_COUNT_MASK);
+               for ( ; i < buf_size; i++)
+                       buf[i] = PATTERN_SRC | (~i & PATTERN_COUNT_MASK);
+               buf++;
+       }
+}
+
+static void dmatest_init_dsts(u8 **bufs, unsigned int start, unsigned int len,
+               unsigned int buf_size)
+{
+       unsigned int i;
+       u8 *buf;
+
+       for (; (buf = *bufs); bufs++) {
+               for (i = 0; i < start; i++)
+                       buf[i] = PATTERN_DST | (~i & PATTERN_COUNT_MASK);
+               for ( ; i < start + len; i++)
+                       buf[i] = PATTERN_DST | PATTERN_OVERWRITE
+                               | (~i & PATTERN_COUNT_MASK);
+               for ( ; i < buf_size; i++)
+                       buf[i] = PATTERN_DST | (~i & PATTERN_COUNT_MASK);
+       }
+}
+
+static void dmatest_mismatch(u8 actual, u8 pattern, unsigned int index,
+               unsigned int counter, bool is_srcbuf)
+{
+       u8              diff = actual ^ pattern;
+       u8              expected = pattern | (~counter & PATTERN_COUNT_MASK);
+       const char      *thread_name = current->comm;
+
+       if (is_srcbuf)
+               pr_warn("%s: srcbuf[0x%x] overwritten! Expected %02x, got %02x\n",
+                       thread_name, index, expected, actual);
+       else if ((pattern & PATTERN_COPY)
+                       && (diff & (PATTERN_COPY | PATTERN_OVERWRITE)))
+               pr_warn("%s: dstbuf[0x%x] not copied! Expected %02x, got %02x\n",
+                       thread_name, index, expected, actual);
+       else if (diff & PATTERN_SRC)
+               pr_warn("%s: dstbuf[0x%x] was copied! Expected %02x, got %02x\n",
+                       thread_name, index, expected, actual);
+       else
+               pr_warn("%s: dstbuf[0x%x] mismatch! Expected %02x, got %02x\n",
+                       thread_name, index, expected, actual);
+}
+
+static unsigned int dmatest_verify(u8 **bufs, unsigned int start,
+               unsigned int end, unsigned int counter, u8 pattern,
+               bool is_srcbuf)
+{
+       unsigned int i;
+       unsigned int error_count = 0;
+       u8 actual;
+       u8 expected;
+       u8 *buf;
+       unsigned int counter_orig = counter;
+
+       for (; (buf = *bufs); bufs++) {
+               counter = counter_orig;
+               for (i = start; i < end; i++) {
+                       actual = buf[i];
+                       expected = pattern | (~counter & PATTERN_COUNT_MASK);
+                       if (actual != expected) {
+                               if (error_count < MAX_ERROR_COUNT)
+                                       dmatest_mismatch(actual, pattern, i,
+                                                        counter, is_srcbuf);
+                               error_count++;
+                       }
+                       counter++;
+               }
+       }
+
+       if (error_count > MAX_ERROR_COUNT)
+               pr_warn("%s: %u errors suppressed\n",
+                       current->comm, error_count - MAX_ERROR_COUNT);
+
+       return error_count;
+}
+
+/* poor man's completion - we want to use wait_event_freezable() on it */
+struct dmatest_done {
+       bool                    done;
+       wait_queue_head_t       *wait;
+};
+
+static void dmatest_callback(void *arg)
+{
+       struct dmatest_done *done = arg;
+
+       done->done = true;
+       wake_up_all(done->wait);
+}
+
+static unsigned int min_odd(unsigned int x, unsigned int y)
+{
+       unsigned int val = min(x, y);
+
+       return val % 2 ? val : val - 1;
+}
+
+static void result(const char *err, unsigned int n, unsigned int src_off,
+                  unsigned int dst_off, unsigned int len, unsigned long data)
+{
+       pr_info("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
+               current->comm, n, err, src_off, dst_off, len, data);
+}
+
+static void dbg_result(const char *err, unsigned int n, unsigned int src_off,
+                      unsigned int dst_off, unsigned int len,
+                      unsigned long data)
+{
+       pr_debug("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
+                current->comm, n, err, src_off, dst_off, len, data);
+}
+
+#define verbose_result(err, n, src_off, dst_off, len, data) ({ \
+       if (verbose)                                            \
+               result(err, n, src_off, dst_off, len, data);    \
+       else                                                    \
+               dbg_result(err, n, src_off, dst_off, len, data);\
+})
+
+static unsigned long long dmatest_persec(s64 runtime, unsigned int val)
+{
+       unsigned long long per_sec = 1000000;
+
+       if (runtime <= 0)
+               return 0;
+
+       /* drop precision until runtime is 32-bits */
+       while (runtime > UINT_MAX) {
+               runtime >>= 1;
+               per_sec <<= 1;
+       }
+
+       per_sec *= val;
+       do_div(per_sec, runtime);
+       return per_sec;
+}
+
+static unsigned long long dmatest_KBs(s64 runtime, unsigned long long len)
+{
+       return dmatest_persec(runtime, len >> 10);
+}
+
+/*
+ * This function repeatedly tests DMA transfers of various lengths and
+ * offsets for a given operation type until it is told to exit by
+ * kthread_stop(). There may be multiple threads running this function
+ * in parallel for a single channel, and there may be multiple channels
+ * being tested in parallel.
+ *
+ * Before each test, the source and destination buffer is initialized
+ * with a known pattern. This pattern is different depending on
+ * whether it's in an area which is supposed to be copied or
+ * overwritten, and different in the source and destination buffers.
+ * So if the DMA engine doesn't copy exactly what we tell it to copy,
+ * we'll notice.
+ */
+static int dmatest_func(void *data)
+{
+       DECLARE_WAIT_QUEUE_HEAD_ONSTACK(done_wait);
+       struct dmatest_thread   *thread = data;
+       struct dmatest_done     done = { .wait = &done_wait };
+       struct dmatest_info     *info;
+       struct dmatest_params   *params;
+       struct dma_chan         *chan;
+       struct dma_device       *dev;
+       unsigned int            error_count;
+       unsigned int            failed_tests = 0;
+       unsigned int            total_tests = 0;
+       dma_cookie_t            cookie;
+       enum dma_status         status;
+       enum dma_ctrl_flags     flags;
+       u8                      *pq_coefs = NULL;
+       int                     ret;
+       int                     src_cnt;
+       int                     dst_cnt;
+       int                     i;
+       ktime_t                 ktime;
+       s64                     runtime = 0;
+       unsigned long long      total_len = 0;
+
+       set_freezable();
+
+       ret = -ENOMEM;
+
+       smp_rmb();
+       info = thread->info;
+       params = &info->params;
+       chan = thread->chan;
+       dev = chan->device;
+       if (thread->type == DMA_MEMCPY)
+               src_cnt = dst_cnt = 1;
+       else if (thread->type == DMA_XOR) {
+               /* force odd to ensure dst = src */
+               src_cnt = min_odd(params->xor_sources | 1, dev->max_xor);
+               dst_cnt = 1;
+       } else if (thread->type == DMA_PQ) {
+               /* force odd to ensure dst = src */
+               src_cnt = min_odd(params->pq_sources | 1, dma_maxpq(dev, 0));
+               dst_cnt = 2;
+
+               pq_coefs = kmalloc(params->pq_sources+1, GFP_KERNEL);
+               if (!pq_coefs)
+                       goto err_thread_type;
+
+               for (i = 0; i < src_cnt; i++)
+                       pq_coefs[i] = 1;
+       } else
+               goto err_thread_type;
+
+       thread->srcs = kcalloc(src_cnt+1, sizeof(u8 *), GFP_KERNEL);
+       if (!thread->srcs)
+               goto err_srcs;
+       for (i = 0; i < src_cnt; i++) {
+               thread->srcs[i] = kmalloc(params->buf_size, GFP_KERNEL);
+               if (!thread->srcs[i])
+                       goto err_srcbuf;
+       }
+       thread->srcs[i] = NULL;
+
+       thread->dsts = kcalloc(dst_cnt+1, sizeof(u8 *), GFP_KERNEL);
+       if (!thread->dsts)
+               goto err_dsts;
+       for (i = 0; i < dst_cnt; i++) {
+               thread->dsts[i] = kmalloc(params->buf_size, GFP_KERNEL);
+               if (!thread->dsts[i])
+                       goto err_dstbuf;
+       }
+       thread->dsts[i] = NULL;
+
+       set_user_nice(current, 10);
+
+       /*
+        * src and dst buffers are freed by ourselves below
+        */
+       flags = DMA_CTRL_ACK | DMA_PREP_INTERRUPT;
+
+       ktime = ktime_get();
+       while (!kthread_should_stop()
+              && !(params->iterations && total_tests >= params->iterations)) {
+               struct dma_async_tx_descriptor *tx = NULL;
+               struct dmaengine_unmap_data *um;
+               dma_addr_t srcs[src_cnt];
+               dma_addr_t *dsts;
+               unsigned int src_off, dst_off, len;
+               u8 align = 0;
+
+               total_tests++;
+
+               /* honor alignment restrictions */
+               if (thread->type == DMA_MEMCPY)
+                       align = dev->copy_align;
+               else if (thread->type == DMA_XOR)
+                       align = dev->xor_align;
+               else if (thread->type == DMA_PQ)
+                       align = dev->pq_align;
+
+               if (1 << align > params->buf_size) {
+                       pr_err("%u-byte buffer too small for %d-byte alignment\n",
+                              params->buf_size, 1 << align);
+                       break;
+               }
+
+               if (params->noverify)
+                       len = params->buf_size;
+               else
+                       len = dmatest_random() % params->buf_size + 1;
+
+               len = (len >> align) << align;
+               if (!len)
+                       len = 1 << align;
+
+               total_len += len;
+
+               if (params->noverify) {
+                       src_off = 0;
+                       dst_off = 0;
+               } else {
+                       src_off = dmatest_random() % (params->buf_size - len + 1);
+                       dst_off = dmatest_random() % (params->buf_size - len + 1);
+
+                       src_off = (src_off >> align) << align;
+                       dst_off = (dst_off >> align) << align;
+
+                       dmatest_init_srcs(thread->srcs, src_off, len,
+                                         params->buf_size);
+                       dmatest_init_dsts(thread->dsts, dst_off, len,
+                                         params->buf_size);
+               }
+
+               um = dmaengine_get_unmap_data(dev->dev, src_cnt+dst_cnt,
+                                             GFP_KERNEL);
+               if (!um) {
+                       failed_tests++;
+                       result("unmap data NULL", total_tests,
+                              src_off, dst_off, len, ret);
+                       continue;
+               }
+
+               um->len = params->buf_size;
+               for (i = 0; i < src_cnt; i++) {
+                       void *buf = thread->srcs[i];
+                       struct page *pg = virt_to_page(buf);
+                       unsigned pg_off = (unsigned long) buf & ~PAGE_MASK;
+
+                       um->addr[i] = dma_map_page(dev->dev, pg, pg_off,
+                                                  um->len, DMA_TO_DEVICE);
+                       srcs[i] = um->addr[i] + src_off;
+                       ret = dma_mapping_error(dev->dev, um->addr[i]);
+                       if (ret) {
+                               dmaengine_unmap_put(um);
+                               result("src mapping error", total_tests,
+                                      src_off, dst_off, len, ret);
+                               failed_tests++;
+                               continue;
+                       }
+                       um->to_cnt++;
+               }
+               /* map with DMA_BIDIRECTIONAL to force writeback/invalidate */
+               dsts = &um->addr[src_cnt];
+               for (i = 0; i < dst_cnt; i++) {
+                       void *buf = thread->dsts[i];
+                       struct page *pg = virt_to_page(buf);
+                       unsigned pg_off = (unsigned long) buf & ~PAGE_MASK;
+
+                       dsts[i] = dma_map_page(dev->dev, pg, pg_off, um->len,
+                                              DMA_BIDIRECTIONAL);
+                       ret = dma_mapping_error(dev->dev, dsts[i]);
+                       if (ret) {
+                               dmaengine_unmap_put(um);
+                               result("dst mapping error", total_tests,
+                                      src_off, dst_off, len, ret);
+                               failed_tests++;
+                               continue;
+                       }
+                       um->bidi_cnt++;
+               }
+
+               if (thread->type == DMA_MEMCPY)
+                       tx = dev->device_prep_dma_memcpy(chan,
+                                                        dsts[0] + dst_off,
+                                                        srcs[0], len, flags);
+               else if (thread->type == DMA_XOR)
+                       tx = dev->device_prep_dma_xor(chan,
+                                                     dsts[0] + dst_off,
+                                                     srcs, src_cnt,
+                                                     len, flags);
+               else if (thread->type == DMA_PQ) {
+                       dma_addr_t dma_pq[dst_cnt];
+
+                       for (i = 0; i < dst_cnt; i++)
+                               dma_pq[i] = dsts[i] + dst_off;
+                       tx = dev->device_prep_dma_pq(chan, dma_pq, srcs,
+                                                    src_cnt, pq_coefs,
+                                                    len, flags);
+               }
+
+               if (!tx) {
+                       dmaengine_unmap_put(um);
+                       result("prep error", total_tests, src_off,
+                              dst_off, len, ret);
+                       msleep(100);
+                       failed_tests++;
+                       continue;
+               }
+
+               done.done = false;
+               tx->callback = dmatest_callback;
+               tx->callback_param = &done;
+               cookie = tx->tx_submit(tx);
+
+               if (dma_submit_error(cookie)) {
+                       dmaengine_unmap_put(um);
+                       result("submit error", total_tests, src_off,
+                              dst_off, len, ret);
+                       msleep(100);
+                       failed_tests++;
+                       continue;
+               }
+               dma_async_issue_pending(chan);
+
+               wait_event_freezable_timeout(done_wait, done.done,
+                                            msecs_to_jiffies(params->timeout));
+
+               status = dma_async_is_tx_complete(chan, cookie, NULL, NULL);
+
+               if (!done.done) {
+                       /*
+                        * We're leaving the timed out dma operation with
+                        * dangling pointer to done_wait.  To make this
+                        * correct, we'll need to allocate wait_done for
+                        * each test iteration and perform "who's gonna
+                        * free it this time?" dancing.  For now, just
+                        * leave it dangling.
+                        */
+                       dmaengine_unmap_put(um);
+                       result("test timed out", total_tests, src_off, dst_off,
+                              len, 0);
+                       failed_tests++;
+                       continue;
+               } else if (status != DMA_COMPLETE) {
+                       dmaengine_unmap_put(um);
+                       result(status == DMA_ERROR ?
+                              "completion error status" :
+                              "completion busy status", total_tests, src_off,
+                              dst_off, len, ret);
+                       failed_tests++;
+                       continue;
+               }
+
+               dmaengine_unmap_put(um);
+
+               if (params->noverify) {
+                       verbose_result("test passed", total_tests, src_off,
+                                      dst_off, len, 0);
+                       continue;
+               }
+
+               pr_debug("%s: verifying source buffer...\n", current->comm);
+               error_count = dmatest_verify(thread->srcs, 0, src_off,
+                               0, PATTERN_SRC, true);
+               error_count += dmatest_verify(thread->srcs, src_off,
+                               src_off + len, src_off,
+                               PATTERN_SRC | PATTERN_COPY, true);
+               error_count += dmatest_verify(thread->srcs, src_off + len,
+                               params->buf_size, src_off + len,
+                               PATTERN_SRC, true);
+
+               pr_debug("%s: verifying dest buffer...\n", current->comm);
+               error_count += dmatest_verify(thread->dsts, 0, dst_off,
+                               0, PATTERN_DST, false);
+               error_count += dmatest_verify(thread->dsts, dst_off,
+                               dst_off + len, src_off,
+                               PATTERN_SRC | PATTERN_COPY, false);
+               error_count += dmatest_verify(thread->dsts, dst_off + len,
+                               params->buf_size, dst_off + len,
+                               PATTERN_DST, false);
+
+               if (error_count) {
+                       result("data error", total_tests, src_off, dst_off,
+                              len, error_count);
+                       failed_tests++;
+               } else {
+                       verbose_result("test passed", total_tests, src_off,
+                                      dst_off, len, 0);
+               }
+       }
+       runtime = ktime_us_delta(ktime_get(), ktime);
+
+       ret = 0;
+err_dstbuf:
+       for (i = 0; thread->dsts[i]; i++)
+               kfree(thread->dsts[i]);
+       kfree(thread->dsts);
+err_dsts:
+err_srcbuf:
+       for (i = 0; thread->srcs[i]; i++)
+               kfree(thread->srcs[i]);
+       kfree(thread->srcs);
+err_srcs:
+       kfree(pq_coefs);
+err_thread_type:
+       pr_info("%s: summary %u tests, %u failures %llu iops %llu KB/s (%d)\n",
+               current->comm, total_tests, failed_tests,
+               dmatest_persec(runtime, total_tests),
+               dmatest_KBs(runtime, total_len), ret);
+
+       /* terminate all transfers on specified channels */
+       if (ret)
+               dmaengine_terminate_all(chan);
+
+       thread->done = true;
+       wake_up(&thread_wait);
+
+       return ret;
+}
+
+static void dmatest_cleanup_channel(struct dmatest_chan *dtc)
+{
+       struct dmatest_thread   *thread;
+       struct dmatest_thread   *_thread;
+       int                     ret;
+
+       list_for_each_entry_safe(thread, _thread, &dtc->threads, node) {
+               ret = kthread_stop(thread->task);
+               pr_debug("thread %s exited with status %d\n",
+                        thread->task->comm, ret);
+               list_del(&thread->node);
+               put_task_struct(thread->task);
+               kfree(thread);
+       }
+
+       /* terminate all transfers on specified channels */
+       dmaengine_terminate_all(dtc->chan);
+
+       kfree(dtc);
+}
+
+static int dmatest_add_threads(struct dmatest_info *info,
+               struct dmatest_chan *dtc, enum dma_transaction_type type)
+{
+       struct dmatest_params *params = &info->params;
+       struct dmatest_thread *thread;
+       struct dma_chan *chan = dtc->chan;
+       char *op;
+       unsigned int i;
+
+       if (type == DMA_MEMCPY)
+               op = "copy";
+       else if (type == DMA_XOR)
+               op = "xor";
+       else if (type == DMA_PQ)
+               op = "pq";
+       else
+               return -EINVAL;
+
+       for (i = 0; i < params->threads_per_chan; i++) {
+               thread = kzalloc(sizeof(struct dmatest_thread), GFP_KERNEL);
+               if (!thread) {
+                       pr_warn("No memory for %s-%s%u\n",
+                               dma_chan_name(chan), op, i);
+                       break;
+               }
+               thread->info = info;
+               thread->chan = dtc->chan;
+               thread->type = type;
+               smp_wmb();
+               thread->task = kthread_create(dmatest_func, thread, "%s-%s%u",
+                               dma_chan_name(chan), op, i);
+               if (IS_ERR(thread->task)) {
+                       pr_warn("Failed to create thread %s-%s%u\n",
+                               dma_chan_name(chan), op, i);
+                       kfree(thread);
+                       break;
+               }
+
+               /* srcbuf and dstbuf are allocated by the thread itself */
+               get_task_struct(thread->task);
+               list_add_tail(&thread->node, &dtc->threads);
+               wake_up_process(thread->task);
+       }
+
+       return i;
+}
+
+static int dmatest_add_channel(struct dmatest_info *info,
+               struct dma_chan *chan)
+{
+       struct dmatest_chan     *dtc;
+       struct dma_device       *dma_dev = chan->device;
+       unsigned int            thread_count = 0;
+       int cnt;
+
+       dtc = kmalloc(sizeof(struct dmatest_chan), GFP_KERNEL);
+       if (!dtc) {
+               pr_warn("No memory for %s\n", dma_chan_name(chan));
+               return -ENOMEM;
+       }
+
+       dtc->chan = chan;
+       INIT_LIST_HEAD(&dtc->threads);
+
+       if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask)) {
+               cnt = dmatest_add_threads(info, dtc, DMA_MEMCPY);
+               thread_count += cnt > 0 ? cnt : 0;
+       }
+       if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
+               cnt = dmatest_add_threads(info, dtc, DMA_XOR);
+               thread_count += cnt > 0 ? cnt : 0;
+       }
+       if (dma_has_cap(DMA_PQ, dma_dev->cap_mask)) {
+               cnt = dmatest_add_threads(info, dtc, DMA_PQ);
+               thread_count += cnt > 0 ? cnt : 0;
+       }
+
+       pr_info("Started %u threads using %s\n",
+               thread_count, dma_chan_name(chan));
+
+       list_add_tail(&dtc->node, &info->channels);
+       info->nr_channels++;
+
+       return 0;
+}
+
+static bool filter(struct dma_chan *chan, void *param)
+{
+       struct dmatest_params *params = param;
+
+       if (!dmatest_match_channel(params, chan) ||
+           !dmatest_match_device(params, chan->device))
+               return false;
+       else
+               return true;
+}
+
+static void request_channels(struct dmatest_info *info,
+                            enum dma_transaction_type type)
+{
+       dma_cap_mask_t mask;
+
+       dma_cap_zero(mask);
+       dma_cap_set(type, mask);
+       for (;;) {
+               struct dmatest_params *params = &info->params;
+               struct dma_chan *chan;
+
+               chan = dma_request_channel(mask, filter, params);
+               if (chan) {
+                       if (dmatest_add_channel(info, chan)) {
+                               dma_release_channel(chan);
+                               break; /* add_channel failed, punt */
+                       }
+               } else
+                       break; /* no more channels available */
+               if (params->max_channels &&
+                   info->nr_channels >= params->max_channels)
+                       break; /* we have all we need */
+       }
+}
+
+static void run_threaded_test(struct dmatest_info *info)
+{
+       struct dmatest_params *params = &info->params;
+
+       /* Copy test parameters */
+       params->buf_size = test_buf_size;
+       strlcpy(params->channel, strim(test_channel), sizeof(params->channel));
+       strlcpy(params->device, strim(test_device), sizeof(params->device));
+       params->threads_per_chan = threads_per_chan;
+       params->max_channels = max_channels;
+       params->iterations = iterations;
+       params->xor_sources = xor_sources;
+       params->pq_sources = pq_sources;
+       params->timeout = timeout;
+       params->noverify = noverify;
+
+       request_channels(info, DMA_MEMCPY);
+       request_channels(info, DMA_XOR);
+       request_channels(info, DMA_PQ);
+}
+
+static void stop_threaded_test(struct dmatest_info *info)
+{
+       struct dmatest_chan *dtc, *_dtc;
+       struct dma_chan *chan;
+
+       list_for_each_entry_safe(dtc, _dtc, &info->channels, node) {
+               list_del(&dtc->node);
+               chan = dtc->chan;
+               dmatest_cleanup_channel(dtc);
+               pr_debug("dropped channel %s\n", dma_chan_name(chan));
+               dma_release_channel(chan);
+       }
+
+       info->nr_channels = 0;
+}
+
+static void restart_threaded_test(struct dmatest_info *info, bool run)
+{
+       /* we might be called early to set run=, defer running until all
+        * parameters have been evaluated
+        */
+       if (!info->did_init)
+               return;
+
+       /* Stop any running test first */
+       stop_threaded_test(info);
+
+       /* Run test with new parameters */
+       run_threaded_test(info);
+}
+
+static int dmatest_run_get(char *val, const struct kernel_param *kp)
+{
+       struct dmatest_info *info = &test_info;
+
+       mutex_lock(&info->lock);
+       if (is_threaded_test_run(info)) {
+               dmatest_run = true;
+       } else {
+               stop_threaded_test(info);
+               dmatest_run = false;
+       }
+       mutex_unlock(&info->lock);
+
+       return param_get_bool(val, kp);
+}
+
+static int dmatest_run_set(const char *val, const struct kernel_param *kp)
+{
+       struct dmatest_info *info = &test_info;
+       int ret;
+
+       mutex_lock(&info->lock);
+       ret = param_set_bool(val, kp);
+       if (ret) {
+               mutex_unlock(&info->lock);
+               return ret;
+       }
+
+       if (is_threaded_test_run(info))
+               ret = -EBUSY;
+       else if (dmatest_run)
+               restart_threaded_test(info, dmatest_run);
+
+       mutex_unlock(&info->lock);
+
+       return ret;
+}
+
+static int __init dmatest_init(void)
+{
+       struct dmatest_info *info = &test_info;
+       struct dmatest_params *params = &info->params;
+
+       if (dmatest_run) {
+               mutex_lock(&info->lock);
+               run_threaded_test(info);
+               mutex_unlock(&info->lock);
+       }
+
+       if (params->iterations && wait)
+               wait_event(thread_wait, !is_threaded_test_run(info));
+
+       /* module parameters are stable, inittime tests are started,
+        * let userspace take over 'run' control
+        */
+       info->did_init = true;
+
+       return 0;
+}
+/* when compiled-in wait for drivers to load first */
+late_initcall(dmatest_init);
+
+static void __exit dmatest_exit(void)
+{
+       struct dmatest_info *info = &test_info;
+
+       mutex_lock(&info->lock);
+       stop_threaded_test(info);
+       mutex_unlock(&info->lock);
+}
+module_exit(dmatest_exit);
+
+MODULE_AUTHOR("Haavard Skinnemoen (Atmel)");
+MODULE_LICENSE("GPL v2");