On 23/09/2021 07:13, Anshuman Khandual wrote:
On 9/21/21 7:11 PM, Suzuki K Poulose wrote:
ARM Neoverse-N2 (#2139208) and Cortex-A710(##2119858) suffers from an erratum, which when triggered, might cause the TRBE to overwrite the trace data already collected in FILL mode, in the event of a WRAP. i.e, the TRBE doesn't stop writing the data, instead wraps to the base and could write upto 3 cache line size worth trace. Thus, this could corrupt the trace at the "BASE" pointer.
The workaround is to program the write pointer 256bytes from the
3 cache lines = 256 bytes on all implementation which might have TRBE ? OR this skid bytes should be derived from the platform cache line size instead.
256bytes is the aligned (to the power of 2) value for the safe guard. Not 3 cache lines. Ideally, if there is another CPU that has larger cache line size, affected by the erratum, yes, we must do that. But for now this is sufficient.
base, such that if the erratum is triggered, it doesn't overwrite the trace data that was captured. This skipped region could be padded with ignore packets at the end of the session, so that the decoder sees a continuous buffer with some padding at the beginning. The trace data written at the base is considered lost as the limit could have been in the middle of the perf ring buffer, and jumping to the "base" is not acceptable. We set the flags already to indicate that some amount of trace was lost during the FILL event IRQ. So this is fine.
Via PERF_AUX_FLAG_TRUNCATED ? Should be specified here to be clear.
Please note that setting the flag is not a side effect of the work around. And as such I don't think this needs to be mentioned here. e.g, we changed this to COLLISION recently for WRAP events. It makes sense to keep the details to the driver.
One important change with the work around is, we program the TRBBASER_EL1 to current page where we are allowed to write. Otherwise, it could overwrite a region that may be consumed by the perf. Towards this, we always make sure that the "handle->head" and thus the trbe_write is PAGE_SIZE aligned, so that we can set the BASE to the PAGE base and move the TRBPTR to the 256bytes offset.
Cc: Mike Leach mike.leach@linaro.org Cc: Mathieu Poirier mathieu.poirier@linaro.org Cc: Anshuman Khandual anshuman.khandual@arm.com Cc: Leo Yan leo.yan@linaro.org Signed-off-by: Suzuki K Poulose suzuki.poulose@arm.com
Change since v1:
- Updated comment with ASCII art
- Add _BYTES suffix for the space to skip for the work around.
drivers/hwtracing/coresight/coresight-trbe.c | 144 +++++++++++++++++-- 1 file changed, 132 insertions(+), 12 deletions(-)
diff --git a/drivers/hwtracing/coresight/coresight-trbe.c b/drivers/hwtracing/coresight/coresight-trbe.c index f569010c672b..983dd5039e52 100644 --- a/drivers/hwtracing/coresight/coresight-trbe.c +++ b/drivers/hwtracing/coresight/coresight-trbe.c @@ -16,6 +16,7 @@ #define pr_fmt(fmt) DRVNAME ": " fmt #include <asm/barrier.h> +#include <asm/cpufeature.h> #include <asm/cputype.h> #include "coresight-self-hosted-trace.h" @@ -84,9 +85,17 @@ struct trbe_buf {
- per TRBE instance, we keep track of the list of errata that
- affects the given instance of the TRBE.
*/ -#define TRBE_ERRATA_MAX 0 +#define TRBE_WORKAROUND_OVERWRITE_FILL_MODE 0 +#define TRBE_ERRATA_MAX 1
+/*
- Safe limit for the number of bytes that may be overwritten
- when the erratum is triggered.
- */
+#define TRBE_WORKAROUND_OVERWRITE_FILL_MODE_SKIP_BYTES 256
As mentioned earlier, does it depend on the platform cache line size ? Otherwise if the skip bytes is something platform independent, should be mentioned here in a comment.
I could add in a comment.
static unsigned long trbe_errata_cpucaps[TRBE_ERRATA_MAX] = {
- [TRBE_WORKAROUND_OVERWRITE_FILL_MODE] = ARM64_WORKAROUND_TRBE_OVERWRITE_FILL_MODE, };
/* @@ -519,10 +528,13 @@ static void trbe_enable_hw(struct trbe_buf *buf) set_trbe_limit_pointer_enabled(buf->trbe_limit); } -static enum trbe_fault_action trbe_get_fault_act(u64 trbsr) +static enum trbe_fault_action trbe_get_fault_act(struct perf_output_handle *handle,
u64 trbsr)- struct trbe_buf *buf = etm_perf_sink_config(handle);
- struct trbe_cpudata *cpudata = buf->cpudata;
Passing down the perf handle to derive trbe_cpudata seems to be right.
WARN_ON(is_trbe_running(trbsr)); if (is_trbe_trg(trbsr) || is_trbe_abort(trbsr)) @@ -531,10 +543,16 @@ static enum trbe_fault_action trbe_get_fault_act(u64 trbsr) if ((ec == TRBE_EC_STAGE1_ABORT) || (ec == TRBE_EC_STAGE2_ABORT)) return TRBE_FAULT_ACT_FATAL;
- if (is_trbe_wrap(trbsr) && (ec == TRBE_EC_OTHERS) && (bsc == TRBE_BSC_FILLED)) {
if (get_trbe_write_pointer() == get_trbe_base_pointer())return TRBE_FAULT_ACT_WRAP;- }
- /*
* If the trbe is affected by TRBE_WORKAROUND_OVERWRITE_FILL_MODE,* it might write data after a WRAP event in the fill mode.* Thus the check TRBPTR == TRBBASER will not be honored.*/Needs bit formatting/alignment cleanup.
- if ((is_trbe_wrap(trbsr) && (ec == TRBE_EC_OTHERS) && (bsc == TRBE_BSC_FILLED)) &&
(trbe_has_erratum(cpudata, TRBE_WORKAROUND_OVERWRITE_FILL_MODE) ||get_trbe_write_pointer() == get_trbe_base_pointer()))return TRBE_FAULT_ACT_WRAP;Right, TRBE without the errata should continue to have the write pointer = base pointer check. Could all TRBE errata checks like the following be shortened (without the workaround index) for better readability ? But not something very important.
trbe_has_erratum(cpudata, TRBE_WORKAROUND_OVERWRITE_FILL_MODE)
Do you mean something like :
trbe_has_erratum(cpudata, TRBE_WORKAROUND_OVERWRITE_FILL_MODE) -> trbe_may_overwrite_in_fill_mode(cpudata) ?
return TRBE_FAULT_ACT_SPURIOUS; } @@ -544,6 +562,8 @@ static unsigned long trbe_get_trace_size(struct perf_output_handle *handle, { u64 write; u64 start_off, end_off;
- u64 size;
- u64 overwrite_skip = TRBE_WORKAROUND_OVERWRITE_FILL_MODE_SKIP_BYTES;
/* * If the TRBE has wrapped around the write pointer has @@ -559,7 +579,18 @@ static unsigned long trbe_get_trace_size(struct perf_output_handle *handle, if (WARN_ON_ONCE(end_off < start_off)) return 0;
- return (end_off - start_off);
- size = end_off - start_off;
- /*
* If the TRBE is affected by the following erratum, we must fill* the space we skipped with IGNORE packets. And we are always* guaranteed to have at least a PAGE_SIZE space in the buffer.*/- if (trbe_has_erratum(buf->cpudata, TRBE_WORKAROUND_OVERWRITE_FILL_MODE) &&
!WARN_ON(size < overwrite_skip))__trbe_pad_buf(buf, start_off, overwrite_skip);- return size; }
static void *arm_trbe_alloc_buffer(struct coresight_device *csdev, @@ -678,7 +709,7 @@ static unsigned long arm_trbe_update_buffer(struct coresight_device *csdev, clr_trbe_irq(); isb();
act = trbe_get_fault_act(status);
act = trbe_get_fault_act(handle, status);
- If this was not due to a WRAP event, we have some
- errors and as such buffer is empty.
@@ -702,21 +733,95 @@ static unsigned long arm_trbe_update_buffer(struct coresight_device *csdev, return size; }
+static int trbe_apply_work_around_before_enable(struct trbe_buf *buf) +{
- /*
* TRBE_WORKAROUND_OVERWRITE_FILL_MODE causes the TRBE to overwrite a few cachefew cache lines = 3 cache lines ?
Yes, upto 3.
* line size from the "TRBBASER_EL1" in the event of a "FILL".* Thus, we could loose some amount of the trace at the base.** Before Fix:** normal-BASE head normal-PTR tail normal-LIMIT* | \/ /* -------------------------------------------------------------* | | |xyzdefghij..|... tuvw| |* -------------------------------------------------------------* / | \* After Fix-> TRBBASER TRBPTR TRBLIMITR.LIMIT** In the normal course of action, we would set the TRBBASER to the* beginning of the ring-buffer (normal-BASE). But with the erratum,* the TRBE could overwrite the contents at the "normal-BASE", after* hitting the "normal-LIMIT", since it doesn't stop as expected. And* this is wrong. So we must always make sure that the TRBBASER is* within the region [head, head+size].** Also, we would set the TRBPTR to head (after adjusting for* alignment) at normal-PTR. This would mean that the last few bytes* of the trace (say, "xyz") might overwrite the first few bytes of* trace written ("abc"). More importantly they will appear in what\* userspace sees as the beginning of the trace, which is wrong. We may* not always have space to move the latest trace "xyz" to the correct* order as it must appear beyond the LIMIT. (i.e, [head..head+size].* Thus it is easier to ignore those bytes than to complicate the* driver to move it, assuming that the erratum was triggered and doing* additional checks to see if there is indeed allowed space at* TRBLIMITR.LIMIT.** To summarize, with the work around:** - We always align the offset for the next session to PAGE_SIZE* (This is to ensure we can program the TRBBASER to this offset* within the region [head...head+size]).** - At TRBE enable:* - Set the TRBBASER to the page aligned offset of the current* proposed write offset. (which is guaranteed to be aligned* as above)* - Move the TRBPTR to skip first 256bytes (that might be* overwritten with the erratum). This ensures that the trace* generated in the session is not re-written.** - At trace collection:* - Pad the 256bytes skipped above again with IGNORE packets.*/- if (trbe_has_erratum(buf->cpudata, TRBE_WORKAROUND_OVERWRITE_FILL_MODE)) {
if (WARN_ON(!IS_ALIGNED(buf->trbe_write, PAGE_SIZE)))return -EINVAL;buf->trbe_hw_base = buf->trbe_write;buf->trbe_write += TRBE_WORKAROUND_OVERWRITE_FILL_MODE_SKIP_BYTES;- }
- return 0;
+}
- static int __arm_trbe_enable(struct trbe_buf *buf, struct perf_output_handle *handle) {
- int ret = 0;
- perf_aux_output_flag(handle, PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW); buf->trbe_limit = compute_trbe_buffer_limit(handle); buf->trbe_write = buf->trbe_base + PERF_IDX2OFF(handle->head, buf); if (buf->trbe_limit == buf->trbe_base) {
trbe_stop_and_truncate_event(handle);return -ENOSPC;
ret = -ENOSPC;goto err;- ret = trbe_apply_work_around_before_enable(buf);
- if (ret)
goto err;- *this_cpu_ptr(buf->cpudata->drvdata->handle) = handle; trbe_enable_hw(buf); return 0;
+err:
- trbe_stop_and_truncate_event(handle);
- return ret; }
static int arm_trbe_enable(struct coresight_device *csdev, u32 mode, void *data) @@ -860,7 +965,7 @@ static irqreturn_t arm_trbe_irq_handler(int irq, void *dev) if (!is_perf_trbe(handle)) return IRQ_NONE;
- act = trbe_get_fault_act(status);
- act = trbe_get_fault_act(handle, status); switch (act) { case TRBE_FAULT_ACT_WRAP: truncated = !!trbe_handle_overflow(handle);
@@ -1000,7 +1105,22 @@ static void arm_trbe_probe_cpu(void *info) } trbe_check_errata(cpudata);
- cpudata->trbe_align = cpudata->trbe_hw_align;
- /*
* If the TRBE is affected by erratum TRBE_WORKAROUND_OVERWRITE_FILL_MODE,* we must always program the TBRPTR_EL1, 256bytes from a page* boundary, with TRBBASER_EL1 set to the page, to prevent* TRBE over-writing 256bytes at TRBBASER_EL1 on FILL event.** Thus make sure we always align our write pointer to a PAGE_SIZE,* which also guarantees that we have at least a PAGE_SIZE space in* the buffer (TRBLIMITR is PAGE aligned) and thus we can skip* the required bytes at the base.*/- if (trbe_has_erratum(cpudata, TRBE_WORKAROUND_OVERWRITE_FILL_MODE))
cpudata->trbe_align = PAGE_SIZE;- else
cpudata->trbe_align = cpudata->trbe_hw_align;But like trbe_apply_work_around_before_enable(), trbe_align assignment should also be wrapped inside a new helper which should contain these comments and conditional block. Because it makes sense to have errata work arounds in the leaf level helper functions, rather than TRBE core operations.
That would imply we re-initialize the trbe_align in the new helper after setting the value here for all other unaffected TRBEs. I would rather leave it as it is, until we have more work arounds that touch this area. This is one of code called per TRBE instance.
Suzuki
-
Suzuki K Poulose