[PATCH v2 0/7] tracing: Kill the buggy trace_cpu

[Oleg Nesterov <oleg@redhat.com>]

Hello.

This series kills tracing_open_generic_tc/trace_cpu/etc.

trace_array_get(inode->i_private) is mostly fine, we do not
dereference this pointer untill trace_array_get() succeeds.
But trace_array_get(tc->tr) is wrong by the same reason why
tracing_open_generic_file/etc are wrong, see 1/7.

(Steven, I think you are right and we can also remove the
 list_for_each_entry() loop from trace_array_get() later,
 and avoid the (harmless but still) race with rmdir + mkdir.
 Lets discuss this after we fix the problems with file/call).

Changes:

	1/7: Fix the whitespace problems.

	     Add the comment above tracing_get_cpu() and update
	     the changelog. tracing_get_cpu() is always safe but
	     the code like

		cpu = tracing_get_cpu(inode);
		if (trace_array_get(tr) == 0)
			do_something(tr, cpu);

	     is racy, it can wrongly use RING_BUFFER_ALL_CPUS.
	     IOW, tracing_get_cpu() should be used only after
	     trace_array_get() or something else which takes
	     trace_types_lock at least once.

	5/7: No changes. checkpatch.pl complains, but that
	     "line over 80 characters" in tracing_entries_read()
	     was not added by this patch and I do not know how
	     can I make checkpatch.pl happy without adding a
	     helper for per_cpu_ptr(tr->trace_buffer.data).

	6/7: Incorporated the fix from Steven (thanks!).

Oleg.

 kernel/trace/trace.c |  194 +++++++++++++++++++++-----------------------------
 kernel/trace/trace.h |    8 --
 2 files changed, 80 insertions(+), 122 deletions(-)