Re: [bug] kpti, perf_event, bts: sporadic truncated trace

[Hugh Dickins <hughd@google.com>]

On Thu, 12 Jul 2018, Metzger, Markus T wrote:

> Hello,
> 
> Starting with 4.15 I noticed that BTS is sporadically missing the tail
> of the trace in the perf_event data buffer.  It shows as
> 
>     [decode error (1): instruction overflow]
> 
> in GDB.  Chances to see this are higher the longer the debuggee is
> running.  With this [1] tiny patch to one of GDB's tests, I am able to
> reproduce it reliably on my box.  To run the test, use:
> 
>     $ make -s check RUNTESTFLAGS="gdb.btrace/exception.exp"
> 
> from the gdb/ sub-directory in the GDB build directory.
> 
> The issue remains when I use 'nopti' on the kernel command-line.
> 
> 
> Bisecting yielded commit
> 
>     c1961a4 x86/events/intel/ds: Map debug buffers in cpu_entry_area
> 
> I reverted the commit on top of v4.17 [2] and the issue disappears
> when I use 'nopti' on the kernel command-line.
> 
> regards,
> markus.
> 
> 
> [1]
> diff --git a/gdb/testsuite/gdb.btrace/exception.exp b/gdb/testsuite/gdb.btrace/exception.exp
> index 9408d61..a24ddd3 100755
> --- a/gdb/testsuite/gdb.btrace/exception.exp
> +++ b/gdb/testsuite/gdb.btrace/exception.exp
> @@ -36,16 +36,12 @@ if ![runto_main] {
>  gdb_test_no_output "set record function-call-history-size 0"
>  
>  # set bp
> -set bp_1 [gdb_get_line_number "bp.1" $srcfile]
>  set bp_2 [gdb_get_line_number "bp.2" $srcfile]
> -gdb_breakpoint $bp_1
>  gdb_breakpoint $bp_2
>  
> -# trace the code between the two breakpoints
> -gdb_continue_to_breakpoint "cont to bp.1" ".*$srcfile:$bp_1\r\n.*"
>  # increase the BTS buffer size - the trace can be quite big
> -gdb_test_no_output "set record btrace bts buffer-size 128000"
> -gdb_test_no_output "record btrace"
> +gdb_test_no_output "set record btrace bts buffer-size 1024000"
> +gdb_test_no_output "record btrace bts"
>  gdb_continue_to_breakpoint "cont to bp.2" ".*$srcfile:$bp_2\r\n.*"
>  
>  # show the flat branch trace
> 
> 
> [2]
> diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
[ snipped the revert ]

Although my name was kept on that commit as a generous courtesy, it
did change a lot after leaving my fingers - and I was never the best
person to be making perf changes in the first place!

I'm sorry to hear that it's breaking you, I've spent a little while
looking through its final state, most of it looks fine to me, but I
notice one discrepancy: whose effect I cannot predict at all, but
there's a chance that it has something to do with what you're seeing.

A little "optimization" crept into alloc_bts_buffer() along the way,
which now places bts_interrupt_threshold not on a record boundary.
And Stephane has shown me the sentence in Vol 3B, 17.4.9, which says
"This address must point to an offset from the BTS buffer base that
is a multiple of the BTS record size."

Please give the patch below a try, and let us know if it helps (if it
does not, then I think we'll need perfier expertise than I can give).

Hugh

--- 4.18-rc4/arch/x86/events/intel/ds.c	2018-06-03 14:15:21.000000000 -0700
+++ linux/arch/x86/events/intel/ds.c	2018-07-12 17:38:28.471378616 -0700
@@ -408,9 +408,11 @@ static int alloc_bts_buffer(int cpu)
 	ds->bts_buffer_base = (unsigned long) cea;
 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
 	ds->bts_index = ds->bts_buffer_base;
-	max = BTS_RECORD_SIZE * (BTS_BUFFER_SIZE / BTS_RECORD_SIZE);
-	ds->bts_absolute_maximum = ds->bts_buffer_base + max;
-	ds->bts_interrupt_threshold = ds->bts_absolute_maximum - (max / 16);
+	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
+	ds->bts_absolute_maximum = ds->bts_buffer_base +
+					max * BTS_RECORD_SIZE;
+	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
+					(max / 16) * BTS_RECORD_SIZE;
 	return 0;
 }