ring-buffer: Align meta-page to sub-buffers for improved TLB usage

Previously, the mapped ring-buffer layout caused misalignment between
the meta-page and sub-buffers when the sub-buffer size was not a
multiple of PAGE_SIZE. This prevented hardware with larger TLB entries
from utilizing them effectively.

Add a padding with the zero-page between the meta-page and sub-buffers.
Also update the ring-buffer map_test to verify that padding.

Link: https://lore.kernel.org/[email protected]
Signed-off-by: Vincent Donnefort <[email protected]>
Signed-off-by: Steven Rostedt (Google) <[email protected]>

Author: vdonnefort

kernel/trace/ring_buffer.c

@@ -6852,10 +6852,10 @@ static void rb_setup_ids_meta_page(struct ring_buffer_per_cpu *cpu_buffer,
 	/* 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;
+	meta->meta_page_size = meta->subbuf_size;
 
 	rb_update_meta_page(cpu_buffer);
 }
@@ -6949,6 +6949,12 @@ static int __rb_map_vma(struct ring_buffer_per_cpu *cpu_buffer,
 	    !(vma->vm_flags & VM_MAYSHARE))
 		return -EPERM;
 
+	subbuf_order = cpu_buffer->buffer->subbuf_order;
+	subbuf_pages = 1 << subbuf_order;
+
+	if (subbuf_order && pgoff % subbuf_pages)
+		return -EINVAL;
+
 	/*
 	 * Make sure the mapping cannot become writable later. Also tell the VM
 	 * to not touch these pages (VM_DONTCOPY | VM_DONTEXPAND).
@@ -6958,11 +6964,8 @@ static int __rb_map_vma(struct ring_buffer_per_cpu *cpu_buffer,
 
 	lockdep_assert_held(&cpu_buffer->mapping_lock);
 
-	subbuf_order = cpu_buffer->buffer->subbuf_order;
-	subbuf_pages = 1 << subbuf_order;
-
 	nr_subbufs = cpu_buffer->nr_pages + 1; /* + reader-subbuf */
-	nr_pages = ((nr_subbufs) << subbuf_order) - pgoff + 1; /* + meta-page */
+	nr_pages = ((nr_subbufs + 1) << subbuf_order) - pgoff; /* + meta-page */
 
 	nr_vma_pages = vma_pages(vma);
 	if (!nr_vma_pages || nr_vma_pages > nr_pages)
@@ -6975,20 +6978,24 @@ static int __rb_map_vma(struct ring_buffer_per_cpu *cpu_buffer,
 		return -ENOMEM;
 
 	if (!pgoff) {
+		unsigned long meta_page_padding;
+
 		pages[p++] = virt_to_page(cpu_buffer->meta_page);
 
 		/*
-		 * TODO: Align sub-buffers on their size, once
-		 * vm_insert_pages() supports the zero-page.
+		 * Pad with the zero-page to align the meta-page with the
+		 * sub-buffers.
 		 */
-	} else {
-		/* Skip the meta-page */
-		pgoff--;
+		meta_page_padding = subbuf_pages - 1;
+		while (meta_page_padding-- && p < nr_pages) {
+			unsigned long __maybe_unused zero_addr =
+				vma->vm_start + (PAGE_SIZE * p);
 
-		if (pgoff % subbuf_pages) {
-			err = -EINVAL;
-			goto out;
+			pages[p++] = ZERO_PAGE(zero_addr);
 		}
+	} else {
+		/* Skip the meta-page */
+		pgoff -= subbuf_pages;
 
 		s += pgoff / subbuf_pages;
 	}

tools/testing/selftests/ring-buffer/map_test.c

@@ -228,6 +228,20 @@ TEST_F(map, data_mmap)
 	data = mmap(NULL, data_len, PROT_READ, MAP_SHARED,
 		    desc->cpu_fd, meta_len);
 	ASSERT_EQ(data, MAP_FAILED);
+
+	/* Verify meta-page padding */
+	if (desc->meta->meta_page_size > getpagesize()) {
+		void *addr;
+
+		data_len = desc->meta->meta_page_size;
+		data = mmap(NULL, data_len,
+			    PROT_READ, MAP_SHARED, desc->cpu_fd, 0);
+		ASSERT_NE(data, MAP_FAILED);
+
+		addr = (void *)((unsigned long)data + getpagesize());
+		ASSERT_EQ(*((int *)addr), 0);
+		munmap(data, data_len);
+	}
 }
 
 FIXTURE(snapshot) {