Re: [PATCH v16 2/6] ring-buffer: Introducing ring-buffer mapping functions

[...]

quoted

+static void rb_update_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
+{
+	struct trace_buffer_meta *meta = cpu_buffer->meta_page;
+
+	meta->reader.read = cpu_buffer->reader_page->read;
+	meta->reader.id = cpu_buffer->reader_page->id;
+	meta->reader.lost_events = cpu_buffer->lost_events;
+
+	meta->entries = local_read(&cpu_buffer->entries);
+	meta->overrun = local_read(&cpu_buffer->overrun);
+	meta->read = cpu_buffer->read;
+
+	/* Some archs do not have data cache coherency between kernel and user-space */
+	flush_dcache_folio(virt_to_folio(cpu_buffer->meta_page));
+}
+
 static void
 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
 {

@@ -5204,6 +5227,9 @@ rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
 	cpu_buffer->lost_events = 0;
 	cpu_buffer->last_overrun = 0;
 
+	if (READ_ONCE(cpu_buffer->mapped))

Isn't the buffer_mutex held when we modify mapped? I believe it's held
here. I don't think we need a READ_ONCE() here. Is there a reason for it?

Hmm, looking down, it looks like you take the buffer->mutex after
setting mapped, is that necessary? If we take the buffer->mutex we can
sync the reset with mapping.

The idea was to not take any of the buffer mutex, reader lock if the refcount is
simply inc/dec. Locks are only used if the meta-page is
installed/uninstalled.

The WRITE_ONCE/READ_ONCE is there only to make sure no compiler optimisation
could lead a reader to wrongly interpret that refcount while it is inc/dec. That
is probably not necessary and I'm happy to either drop it completely or replace
it by taking buffer mutex and reader lock whenever the refcount is inc/dec.

quoted

+		rb_update_meta_page(cpu_buffer);
+
 	rb_head_page_activate(cpu_buffer);
 	cpu_buffer->pages_removed = 0;
 }

@@ -5418,6 +5444,12 @@ int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
 	cpu_buffer_a = buffer_a->buffers[cpu];
 	cpu_buffer_b = buffer_b->buffers[cpu];
 
+	/* It's up to the callers to not try to swap mapped buffers */
+	if (WARN_ON_ONCE(cpu_buffer_a->mapped || cpu_buffer_b->mapped)) {
+		ret = -EBUSY;
+		goto out;
+	}
+
 	/* At least make sure the two buffers are somewhat the same */
 	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
 		goto out;

@@ -5682,7 +5714,8 @@ int ring_buffer_read_page(struct trace_buffer *buffer,
 	 * Otherwise, we can simply swap the page with the one passed in.
 	 */
 	if (read || (len < (commit - read)) ||
-	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
+	    cpu_buffer->reader_page == cpu_buffer->commit_page ||
+	    READ_ONCE(cpu_buffer->mapped)) {

Here the buffer_mutex isn't held, but still, what's the purpose of the READ_ONCE?

I wonder if we just grab the reader_lock when setting the mapped
variable if that would be better than using READ/WRITE_ONCE, which I'm
not sure is helpful.

ditto above.

quoted

 		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
 		unsigned int rpos = read;
 		unsigned int pos = 0;

@@ -5901,6 +5934,11 @@ int ring_buffer_subbuf_order_set(struct trace_buffer *buffer, int order)
 
 		cpu_buffer = buffer->buffers[cpu];
 
+		if (READ_ONCE(cpu_buffer->mapped)) {
+			err = -EBUSY;
+			goto error;
+		}
+
 		/* Update the number of pages to match the new size */
 		nr_pages = old_size * buffer->buffers[cpu]->nr_pages;
 		nr_pages = DIV_ROUND_UP(nr_pages, buffer->subbuf_size);

@@ -6002,6 +6040,304 @@ int ring_buffer_subbuf_order_set(struct trace_buffer *buffer, int order)
 }
 EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_set);
 
+#define subbuf_page(off, start) \
+	virt_to_page((void *)(start + (off << PAGE_SHIFT)))
+
+#define foreach_subbuf_page(sub_order, start, page)		\
+	page = subbuf_page(0, (start));				\
+	for (int __off = 0; __off < (1 << (sub_order));		\
+	     __off++, page = subbuf_page(__off, (start)))
+
+static inline void subbuf_map_prepare(unsigned long subbuf_start, int order)
+{
+	struct page *page;
+
+	/*
+	 * When allocating order > 0 pages, only the first struct page has a
+	 * refcount > 1. Increasing the refcount here ensures none of the struct
+	 * page composing the sub-buffer is freeed when the mapping is closed.
+	 */
+	foreach_subbuf_page(order, subbuf_start, page)
+		page_ref_inc(page);
+}
+
+static inline void subbuf_unmap(unsigned long subbuf_start, int order)
+{
+	struct page *page;
+
+	foreach_subbuf_page(order, subbuf_start, page) {
+		page_ref_dec(page);
+		page->mapping = NULL;
+	}
+}
+
+static void rb_free_subbuf_ids(struct ring_buffer_per_cpu *cpu_buffer)
+{
+	int sub_id;
+
+	for (sub_id = 0; sub_id < cpu_buffer->nr_pages + 1; sub_id++)
+		subbuf_unmap(cpu_buffer->subbuf_ids[sub_id],
+			     cpu_buffer->buffer->subbuf_order);
+
+	kfree(cpu_buffer->subbuf_ids);
+	cpu_buffer->subbuf_ids = NULL;
+}
+
+static int rb_alloc_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
+{
+	if (cpu_buffer->meta_page)
+		return 0;
+
+	cpu_buffer->meta_page = page_to_virt(alloc_page(GFP_USER | __GFP_ZERO));

The above is the main reason I'm looking for a v17. If that
alloc_page() fails, what exactly is going to be returned here? 

#define page_to_virt(x) __va(PFN_PHYS(page_to_pfn(x)))
#define __va(x)                     ((void *)((unsigned long)(x)+PAGE_OFFSET))

I don't think that will be the result you expect. We need to do the
alloc_page(), test the result of that, and then call page_to_virt() on a
page that is not NULL.

Ouch, you're right!

quoted

+	if (!cpu_buffer->meta_page)
+		return -ENOMEM;
+
+	return 0;
+}
+
+static void rb_free_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
+{
+	unsigned long addr = (unsigned long)cpu_buffer->meta_page;
+

I'm thinking for extra safety we could add:

	if (!addr)
		return;

So this function could be called twice without issue.

Ack.

quoted

+	virt_to_page((void *)addr)->mapping = NULL;
+	free_page(addr);
+	cpu_buffer->meta_page = NULL;
+}
+
+static void rb_setup_ids_meta_page(struct ring_buffer_per_cpu *cpu_buffer,
+				   unsigned long *subbuf_ids)
+{
+	struct trace_buffer_meta *meta = cpu_buffer->meta_page;
+	unsigned int nr_subbufs = cpu_buffer->nr_pages + 1;
+	struct buffer_page *first_subbuf, *subbuf;
+	int id = 0;
+
+	subbuf_ids[id] = (unsigned long)cpu_buffer->reader_page->page;
+	subbuf_map_prepare(subbuf_ids[id], cpu_buffer->buffer->subbuf_order);
+	cpu_buffer->reader_page->id = id++;
+
+	first_subbuf = subbuf = rb_set_head_page(cpu_buffer);
+	do {
+		if (id >= nr_subbufs) {
+			WARN_ON(1);
+			break;
+		}

		if (WARN_ON(id >= nr_subbufs))
			break;

quoted

+
+		subbuf_ids[id] = (unsigned long)subbuf->page;
+		subbuf->id = id;
+		subbuf_map_prepare(subbuf_ids[id], cpu_buffer->buffer->subbuf_order);
+
+		rb_inc_page(&subbuf);
+		id++;
+	} while (subbuf != first_subbuf);
+
+	/* install subbuf ID to kern VA translation */
+	cpu_buffer->subbuf_ids = subbuf_ids;
+
+	meta->meta_page_size = PAGE_SIZE;
+	meta->meta_struct_len = sizeof(*meta);
+	meta->nr_subbufs = nr_subbufs;
+	meta->subbuf_size = cpu_buffer->buffer->subbuf_size + BUF_PAGE_HDR_SIZE;
+
+	rb_update_meta_page(cpu_buffer);
+}
+
+static inline struct ring_buffer_per_cpu *
+rb_get_mapped_buffer(struct trace_buffer *buffer, int cpu)
+{
+	struct ring_buffer_per_cpu *cpu_buffer;
+
+	if (!cpumask_test_cpu(cpu, buffer->cpumask))
+		return ERR_PTR(-EINVAL);
+
+	cpu_buffer = buffer->buffers[cpu];
+
+	mutex_lock(&cpu_buffer->mapping_lock);
+
+	if (!cpu_buffer->mapped) {
+		mutex_unlock(&cpu_buffer->mapping_lock);
+		return ERR_PTR(-ENODEV);
+	}
+
+	return cpu_buffer;
+}
+
+static inline void rb_put_mapped_buffer(struct ring_buffer_per_cpu *cpu_buffer)
+{
+	mutex_unlock(&cpu_buffer->mapping_lock);
+}
+
+int ring_buffer_map(struct trace_buffer *buffer, int cpu)
+{
+	struct ring_buffer_per_cpu *cpu_buffer;
+	unsigned long flags, *subbuf_ids;
+	int err = 0;
+
+	if (!cpumask_test_cpu(cpu, buffer->cpumask))
+		return -EINVAL;
+
+	cpu_buffer = buffer->buffers[cpu];
+
+	mutex_lock(&cpu_buffer->mapping_lock);

Is there a reason not to take the buffer->mutex before setting mapped?

quoted

+
+	if (cpu_buffer->mapped) {
+		if (cpu_buffer->mapped == UINT_MAX)
+			err = -EBUSY;
+		else
+			WRITE_ONCE(cpu_buffer->mapped, cpu_buffer->mapped + 1);

As mentioned before, it may be better to take the reader_lock when
setting mapped and that should add the protection we want with mapped
set and other readers.

quoted

+		mutex_unlock(&cpu_buffer->mapping_lock);
+		return err;
+	}
+
+	/* prevent another thread from changing buffer/sub-buffer sizes */
+	mutex_lock(&buffer->mutex);
+
+	err = rb_alloc_meta_page(cpu_buffer);
+	if (err)
+		goto unlock;
+
+	/* subbuf_ids include the reader while nr_pages does not */
+	subbuf_ids = kzalloc(sizeof(*subbuf_ids) * (cpu_buffer->nr_pages + 1),
+			   GFP_KERNEL);

Instead use:

	subbuf_ids = kcalloc(cpu_buffer->nr_pages + 1, sizeof(*subbuf_ids), GFP_KERNEL);

quoted

+	if (!subbuf_ids) {
+		rb_free_meta_page(cpu_buffer);
+		err = -ENOMEM;
+		goto unlock;
+	}
+
+	atomic_inc(&cpu_buffer->resize_disabled);
+
+	/*
+	 * Lock all readers to block any subbuf swap until the subbuf IDs are
+	 * assigned.
+	 */
+	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
+
+	rb_setup_ids_meta_page(cpu_buffer, subbuf_ids);
+	cpu_buffer->mapped = 1;
+
+	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
+unlock:
+	mutex_unlock(&buffer->mutex);
+	mutex_unlock(&cpu_buffer->mapping_lock);
+
+	return err;
+}
+
+int ring_buffer_unmap(struct trace_buffer *buffer, int cpu)
+{
+	struct ring_buffer_per_cpu *cpu_buffer;
+	unsigned long flags;
+	int err = 0;
+
+	if (!cpumask_test_cpu(cpu, buffer->cpumask))
+		return -EINVAL;
+
+	cpu_buffer = buffer->buffers[cpu];
+
+	mutex_lock(&cpu_buffer->mapping_lock);
+
+	if (!cpu_buffer->mapped) {
+		err = -ENODEV;
+		goto out;
+	} else if (cpu_buffer->mapped > 1) {
+		WRITE_ONCE(cpu_buffer->mapped, cpu_buffer->mapped - 1);
+		goto out;
+	}
+
+	mutex_lock(&buffer->mutex);
+	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
+
+	cpu_buffer->mapped = 0;
+
+	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
+
+	rb_free_subbuf_ids(cpu_buffer);
+	rb_free_meta_page(cpu_buffer);
+	atomic_dec(&cpu_buffer->resize_disabled);
+
+	mutex_unlock(&buffer->mutex);
+out:
+	mutex_unlock(&cpu_buffer->mapping_lock);
+
+	return err;
+}
+
+/*
+ *   +--------------+  pgoff == 0
+ *   |   meta page  |
+ *   +--------------+  pgoff == 1
+ *   | subbuffer 0  |
+ *   +--------------+  pgoff == 1 + (1 << subbuf_order)
+ *   | subbuffer 1  |
+ *         ...
+ */
+struct page *ring_buffer_map_fault(struct trace_buffer *buffer, int cpu,
+				   unsigned long pgoff)
+{
+	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
+	unsigned long subbuf_id, subbuf_offset, addr;
+	struct page *page;
+
+	if (!pgoff)
+		return virt_to_page((void *)cpu_buffer->meta_page);
+
+	pgoff--;
+
+	subbuf_id = pgoff >> buffer->subbuf_order;
+	if (subbuf_id > cpu_buffer->nr_pages)
+		return NULL;
+
+	subbuf_offset = pgoff & ((1UL << buffer->subbuf_order) - 1);
+	addr = cpu_buffer->subbuf_ids[subbuf_id] + (subbuf_offset * PAGE_SIZE);
+	page = virt_to_page((void *)addr);
+
+	return page;
+}
+
+int ring_buffer_map_get_reader(struct trace_buffer *buffer, int cpu)
+{
+	struct ring_buffer_per_cpu *cpu_buffer;
+	unsigned long reader_size;
+	unsigned long flags;
+
+	cpu_buffer = rb_get_mapped_buffer(buffer, cpu);
+	if (IS_ERR(cpu_buffer))
+		return (int)PTR_ERR(cpu_buffer);
+
+	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
+consume:
+	if (rb_per_cpu_empty(cpu_buffer))
+		goto out;
+
+	reader_size = rb_page_size(cpu_buffer->reader_page);
+
+	/*
+	 * There are data to be read on the current reader page, we can
+	 * return to the caller. But before that, we assume the latter will read
+	 * everything. Let's update the kernel reader accordingly.
+	 */
+	if (cpu_buffer->reader_page->read < reader_size) {
+		while (cpu_buffer->reader_page->read < reader_size)
+			rb_advance_reader(cpu_buffer);
+		goto out;
+	}
+
+	if (WARN_ON(!rb_get_reader_page(cpu_buffer)))
+		goto out;
+
+	goto consume;
+out:
+	rb_update_meta_page(cpu_buffer);
+	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
+	rb_put_mapped_buffer(cpu_buffer);
+

Hmm, there's no protection here. If this task gets preempted for a long
time, it is possible that the cpu_buffer->reader_page can become NULL,
causing a NULL kernel dereference.

Right, this should be behind the reader lock as well :-\

-- Steve

quoted

+	/* Some archs do not have data cache coherency between kernel and user-space */
+	flush_dcache_folio(virt_to_folio(cpu_buffer->reader_page->page));
+
+	return 0;
+}
+
 /*
  * We only allocate new buffers, never free them if the CPU goes down.
  * If we were to free the buffer, then the user would lose any trace that was in