Skip to content

Commit 2e499bb

Browse files
mathieupoiriergregkh
authored andcommitted
coresight: tmc: implementing TMC-ETF AUX space API
This patch implement the AUX area interfaces required to use the TMC (configured as an ETF) from the Perf sub-system. The heuristic is heavily borrowed from the ETB10 implementation. Signed-off-by: Mathieu Poirier <[email protected]> Reviewed-by: Suzuki K Poulose <[email protected]> Signed-off-by: Greg Kroah-Hartman <[email protected]>
1 parent a02e81f commit 2e499bb

File tree

2 files changed

+200
-0
lines changed

2 files changed

+200
-0
lines changed

drivers/hwtracing/coresight/coresight-tmc-etf.c

+199
Original file line numberDiff line numberDiff line change
@@ -15,7 +15,9 @@
1515
* this program. If not, see <http://www.gnu.org/licenses/>.
1616
*/
1717

18+
#include <linux/circ_buf.h>
1819
#include <linux/coresight.h>
20+
#include <linux/perf_event.h>
1921
#include <linux/slab.h>
2022
#include "coresight-priv.h"
2123
#include "coresight-tmc.h"
@@ -282,9 +284,206 @@ static void tmc_disable_etf_link(struct coresight_device *csdev,
282284
dev_info(drvdata->dev, "TMC disabled\n");
283285
}
284286

287+
static void *tmc_alloc_etf_buffer(struct coresight_device *csdev, int cpu,
288+
void **pages, int nr_pages, bool overwrite)
289+
{
290+
int node;
291+
struct cs_buffers *buf;
292+
293+
if (cpu == -1)
294+
cpu = smp_processor_id();
295+
node = cpu_to_node(cpu);
296+
297+
/* Allocate memory structure for interaction with Perf */
298+
buf = kzalloc_node(sizeof(struct cs_buffers), GFP_KERNEL, node);
299+
if (!buf)
300+
return NULL;
301+
302+
buf->snapshot = overwrite;
303+
buf->nr_pages = nr_pages;
304+
buf->data_pages = pages;
305+
306+
return buf;
307+
}
308+
309+
static void tmc_free_etf_buffer(void *config)
310+
{
311+
struct cs_buffers *buf = config;
312+
313+
kfree(buf);
314+
}
315+
316+
static int tmc_set_etf_buffer(struct coresight_device *csdev,
317+
struct perf_output_handle *handle,
318+
void *sink_config)
319+
{
320+
int ret = 0;
321+
unsigned long head;
322+
struct cs_buffers *buf = sink_config;
323+
324+
/* wrap head around to the amount of space we have */
325+
head = handle->head & ((buf->nr_pages << PAGE_SHIFT) - 1);
326+
327+
/* find the page to write to */
328+
buf->cur = head / PAGE_SIZE;
329+
330+
/* and offset within that page */
331+
buf->offset = head % PAGE_SIZE;
332+
333+
local_set(&buf->data_size, 0);
334+
335+
return ret;
336+
}
337+
338+
static unsigned long tmc_reset_etf_buffer(struct coresight_device *csdev,
339+
struct perf_output_handle *handle,
340+
void *sink_config, bool *lost)
341+
{
342+
long size = 0;
343+
struct cs_buffers *buf = sink_config;
344+
345+
if (buf) {
346+
/*
347+
* In snapshot mode ->data_size holds the new address of the
348+
* ring buffer's head. The size itself is the whole address
349+
* range since we want the latest information.
350+
*/
351+
if (buf->snapshot)
352+
handle->head = local_xchg(&buf->data_size,
353+
buf->nr_pages << PAGE_SHIFT);
354+
/*
355+
* Tell the tracer PMU how much we got in this run and if
356+
* something went wrong along the way. Nobody else can use
357+
* this cs_buffers instance until we are done. As such
358+
* resetting parameters here and squaring off with the ring
359+
* buffer API in the tracer PMU is fine.
360+
*/
361+
*lost = !!local_xchg(&buf->lost, 0);
362+
size = local_xchg(&buf->data_size, 0);
363+
}
364+
365+
return size;
366+
}
367+
368+
static void tmc_update_etf_buffer(struct coresight_device *csdev,
369+
struct perf_output_handle *handle,
370+
void *sink_config)
371+
{
372+
int i, cur;
373+
u32 *buf_ptr;
374+
u32 read_ptr, write_ptr;
375+
u32 status, to_read;
376+
unsigned long offset;
377+
struct cs_buffers *buf = sink_config;
378+
struct tmc_drvdata *drvdata = dev_get_drvdata(csdev->dev.parent);
379+
380+
if (!buf)
381+
return;
382+
383+
/* This shouldn't happen */
384+
if (WARN_ON_ONCE(local_read(&drvdata->mode) != CS_MODE_PERF))
385+
return;
386+
387+
CS_UNLOCK(drvdata->base);
388+
389+
tmc_flush_and_stop(drvdata);
390+
391+
read_ptr = readl_relaxed(drvdata->base + TMC_RRP);
392+
write_ptr = readl_relaxed(drvdata->base + TMC_RWP);
393+
394+
/*
395+
* Get a hold of the status register and see if a wrap around
396+
* has occurred. If so adjust things accordingly.
397+
*/
398+
status = readl_relaxed(drvdata->base + TMC_STS);
399+
if (status & TMC_STS_FULL) {
400+
local_inc(&buf->lost);
401+
to_read = drvdata->size;
402+
} else {
403+
to_read = CIRC_CNT(write_ptr, read_ptr, drvdata->size);
404+
}
405+
406+
/*
407+
* The TMC RAM buffer may be bigger than the space available in the
408+
* perf ring buffer (handle->size). If so advance the RRP so that we
409+
* get the latest trace data.
410+
*/
411+
if (to_read > handle->size) {
412+
u32 mask = 0;
413+
414+
/*
415+
* The value written to RRP must be byte-address aligned to
416+
* the width of the trace memory databus _and_ to a frame
417+
* boundary (16 byte), whichever is the biggest. For example,
418+
* for 32-bit, 64-bit and 128-bit wide trace memory, the four
419+
* LSBs must be 0s. For 256-bit wide trace memory, the five
420+
* LSBs must be 0s.
421+
*/
422+
switch (drvdata->memwidth) {
423+
case TMC_MEM_INTF_WIDTH_32BITS:
424+
case TMC_MEM_INTF_WIDTH_64BITS:
425+
case TMC_MEM_INTF_WIDTH_128BITS:
426+
mask = GENMASK(31, 5);
427+
break;
428+
case TMC_MEM_INTF_WIDTH_256BITS:
429+
mask = GENMASK(31, 6);
430+
break;
431+
}
432+
433+
/*
434+
* Make sure the new size is aligned in accordance with the
435+
* requirement explained above.
436+
*/
437+
to_read = handle->size & mask;
438+
/* Move the RAM read pointer up */
439+
read_ptr = (write_ptr + drvdata->size) - to_read;
440+
/* Make sure we are still within our limits */
441+
if (read_ptr > (drvdata->size - 1))
442+
read_ptr -= drvdata->size;
443+
/* Tell the HW */
444+
writel_relaxed(read_ptr, drvdata->base + TMC_RRP);
445+
local_inc(&buf->lost);
446+
}
447+
448+
cur = buf->cur;
449+
offset = buf->offset;
450+
451+
/* for every byte to read */
452+
for (i = 0; i < to_read; i += 4) {
453+
buf_ptr = buf->data_pages[cur] + offset;
454+
*buf_ptr = readl_relaxed(drvdata->base + TMC_RRD);
455+
456+
offset += 4;
457+
if (offset >= PAGE_SIZE) {
458+
offset = 0;
459+
cur++;
460+
/* wrap around at the end of the buffer */
461+
cur &= buf->nr_pages - 1;
462+
}
463+
}
464+
465+
/*
466+
* In snapshot mode all we have to do is communicate to
467+
* perf_aux_output_end() the address of the current head. In full
468+
* trace mode the same function expects a size to move rb->aux_head
469+
* forward.
470+
*/
471+
if (buf->snapshot)
472+
local_set(&buf->data_size, (cur * PAGE_SIZE) + offset);
473+
else
474+
local_add(to_read, &buf->data_size);
475+
476+
CS_LOCK(drvdata->base);
477+
}
478+
285479
static const struct coresight_ops_sink tmc_etf_sink_ops = {
286480
.enable = tmc_enable_etf_sink,
287481
.disable = tmc_disable_etf_sink,
482+
.alloc_buffer = tmc_alloc_etf_buffer,
483+
.free_buffer = tmc_free_etf_buffer,
484+
.set_buffer = tmc_set_etf_buffer,
485+
.reset_buffer = tmc_reset_etf_buffer,
486+
.update_buffer = tmc_update_etf_buffer,
288487
};
289488

290489
static const struct coresight_ops_link tmc_etf_link_ops = {

drivers/hwtracing/coresight/coresight-tmc.h

+1
Original file line numberDiff line numberDiff line change
@@ -52,6 +52,7 @@
5252
#define TMC_CTL_CAPT_EN BIT(0)
5353
/* TMC_STS - 0x00C */
5454
#define TMC_STS_TMCREADY_BIT 2
55+
#define TMC_STS_FULL BIT(0)
5556
#define TMC_STS_TRIGGERED BIT(1)
5657
/* TMC_AXICTL - 0x110 */
5758
#define TMC_AXICTL_PROT_CTL_B0 BIT(0)

0 commit comments

Comments
 (0)