Re: [PATCH 1/1] tools/dtrace: initial implementation of DTrace

[Kris Van Hees]

On Thu, Jul 04, 2019 at 03:03:36PM +0200, Peter Zijlstra wrote:
> On Wed, Jul 03, 2019 at 08:14:30PM -0700, Kris Van Hees wrote:
> > +/*
> > + * Read the data_head offset from the header page of the ring buffer.  The
> > + * argument is declared 'volatile' because it references a memory mapped page
> > + * that the kernel may be writing to while we access it here.
> > + */
> > +static u64 read_rb_head(volatile struct perf_event_mmap_page *rb_page)
> > +{
> > +	u64	head = rb_page->data_head;
> > +
> > +	asm volatile("" ::: "memory");
> > +
> > +	return head;
> > +}
> > +
> > +/*
> > + * Write the data_tail offset in the header page of the ring buffer.  The
> > + * argument is declared 'volatile' because it references a memory mapped page
> > + * that the kernel may be writing to while we access it here.
> 
> s/writing/reading/

Thanks!

> > + */
> > +static void write_rb_tail(volatile struct perf_event_mmap_page *rb_page,
> > +			  u64 tail)
> > +{
> > +	asm volatile("" ::: "memory");
> > +
> > +	rb_page->data_tail = tail;
> > +}
> 
> That volatile usage is atrocious (kernel style would have you use
> {READ,WRITE}_ONCE()). Also your comments fail to mark these as
> load_acquire and store_release. And by only using a compiler barrier
> you're hard assuming TSO, which is somewhat fragile at best.
> 
> Alternatively, you can use the C11 bits and write:
> 
> 	return __atomic_load_n(&rb_page->data_head, __ATOMIC_ACQUIRE);
> 
> 	__atomic_store_n(&rb_page->data_tail, tail, __ATOMIC_RELEASE);

Perhaps I should just use ring_buffer_read_head() and ring_buffer_write_tail()
since they are provided in tools/include/linux/ring_buffer.h?  I expect that
would be even more preferable over __atomic_load_n() and __atomic_store_n()?

> > +/*
> > + * Process and output the probe data at the supplied address.
> > + */
> > +static int output_event(int cpu, u64 *buf)
> > +{
> > +	u8				*data = (u8 *)buf;
> > +	struct perf_event_header	*hdr;
> > +
> > +	hdr = (struct perf_event_header *)data;
> > +	data += sizeof(struct perf_event_header);
> > +
> > +	if (hdr->type == PERF_RECORD_SAMPLE) {
> > +		u8		*ptr = data;
> > +		u32		i, size, probe_id;
> > +
> > +		/*
> > +		 * struct {
> > +		 *	struct perf_event_header	header;
> > +		 *	u32				size;
> > +		 *	u32				probe_id;
> > +		 *	u32				gap;
> > +		 *	u64				data[n];
> > +		 * }
> > +		 * and data points to the 'size' member at this point.
> > +		 */
> > +		if (ptr > (u8 *)buf + hdr->size) {
> > +			fprintf(stderr, "BAD: corrupted sample header\n");
> > +			goto out;
> > +		}
> > +
> > +		size = *(u32 *)data;
> > +		data += sizeof(size);
> > +		ptr += sizeof(size) + size;
> > +		if (ptr != (u8 *)buf + hdr->size) {
> > +			fprintf(stderr, "BAD: invalid sample size\n");
> > +			goto out;
> > +		}
> > +
> > +		probe_id = *(u32 *)data;
> > +		data += sizeof(probe_id);
> > +		size -= sizeof(probe_id);
> > +		data += sizeof(u32);		/* skip 32-bit gap */
> > +		size -= sizeof(u32);
> > +		buf = (u64 *)data;
> > +
> > +		printf("%3d %6d ", cpu, probe_id);
> > +		for (i = 0, size /= sizeof(u64); i < size; i++)
> > +			printf("%#016lx ", buf[i]);
> > +		printf("\n");
> > +	} else if (hdr->type == PERF_RECORD_LOST) {
> > +		u64	lost;
> > +
> > +		/*
> > +		 * struct {
> > +		 *	struct perf_event_header	header;
> > +		 *	u64				id;
> > +		 *	u64				lost;
> > +		 * }
> > +		 * and data points to the 'id' member at this point.
> > +		 */
> > +		lost = *(u64 *)(data + sizeof(u64));
> > +
> > +		printf("[%ld probes dropped]\n", lost);
> > +	} else
> > +		fprintf(stderr, "UNKNOWN: record type %d\n", hdr->type);
> > +
> > +out:
> > +	return hdr->size;
> > +}
> 
> I see a distinct lack of wrapping support. AFAICT when buf+hdr->size
> wraps you're doing out-of-bounds accesses.

Yes, that is correct.  I'm actually trying to figure out why it didn't actually
cause a SEGV when I tested this because I'm clearly reading past the end of
the mmap'd memory.  Thank you for noticing this - I was trying to be too
minimal in the code I was putting out and really didn't pay attention to this.

Fixed in the V2 I am preparing.

> > +/*
> > + * Process the available probe data in the given buffer.
> > + */
> > +static void process_data(struct dtrace_buffer *buf)
> > +{
> > +	/* This is volatile because the kernel may be updating the content. */
> > +	volatile struct perf_event_mmap_page	*rb_page = buf->base;
> > +	u8					*base = (u8 *)buf->base +
> > +							buf->page_size;
> > +	u64					head = read_rb_head(rb_page);
> > +
> > +	while (rb_page->data_tail != head) {
> > +		u64	tail = rb_page->data_tail;
> > +		u64	*ptr = (u64 *)(base + tail % buf->data_size);
> > +		int	len;
> > +
> > +		len = output_event(buf->cpu, ptr);
> > +
> > +		write_rb_tail(rb_page, tail + len);
> > +		head = read_rb_head(rb_page);
> > +	}
> > +}
> 
> more volatile yuck.
> 
> Also:
> 
> 	for (;;) {
> 		head = __atomic_load_n(&rb_page->data_head, __ATOMIC_ACQUIRE);
> 		tail = __atomic_load_n(&rb_page->data_tail, __ATOMIC_RELAXED);
> 
> 		if (head == tail)
> 			break;
> 
> 		do {
> 			hdr = buf->base + (tail & ((1UL << buf->data_shift) - 1));
> 			if ((tail >> buf->data_shift) !=
> 			    ((tail + hdr->size) >> buf->data_shift))
> 				/* handle wrap case */
> 			else
> 				/* normal case */
> 
> 			tail += hdr->size;
> 		} while (tail != head);
> 
> 		__atomic_store_n(&rb_page->data_tail, tail, __ATOMIC_RELEASE);
> 	}
> 
> Or something.

Thank you for this suggestion.  As mentioned above, I lean towards using the
provided ring_buffer_(read_head,write_tail) implementations since that is the
'other end' of the ring buffer head/tail mechanism that is going to be kept
in sync with any changes that might happen on the kernel side, right?

> > +/*
> > + * Wait for data to become available in any of the buffers.
> > + */
> > +int dt_buffer_poll(int epoll_fd, int timeout)
> > +{
> > +	struct epoll_event	events[dt_numcpus];
> > +	int			i, cnt;
> > +
> > +	cnt = epoll_wait(epoll_fd, events, dt_numcpus, timeout);
> > +	if (cnt < 0)
> > +		return -errno;
> > +
> > +	for (i = 0; i < cnt; i++)
> > +		process_data((struct dtrace_buffer *)events[i].data.ptr);
> > +
> > +	return cnt;
> > +}
> 
> Or make sure to read on the CPU by having a poll thread per CPU, then
> you can do away with the memory barriers.

That is definitely something for the todo list for future optimizations.

Thanks for your review and code suggestions.

	Kris