Default to mmap'ed provided buffers for hppa

Architectures like parisc cannot guarantee that the kernel and userspace
virtual address map to the same cache color, when we have the application
allocate the ring and then the kernel mapping it.

Add a helper to use the buffer ring registration that has the kernel
allocate the memory, and then have the application mmap(2) it. With
that, we can have mmap pick an appropriate virtual address that so we
avoid any coloring mismatches.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

src/include/liburing/io_uring.h

@@ -389,6 +389,9 @@ enum {
 #define IORING_OFF_SQ_RING		0ULL
 #define IORING_OFF_CQ_RING		0x8000000ULL
 #define IORING_OFF_SQES			0x10000000ULL
+#define IORING_OFF_PBUF_RING		0x80000000ULL
+#define IORING_OFF_PBUF_SHIFT		16
+#define IORING_OFF_MMAP_MASK		0xf8000000ULL
 
 /*
  * Filled with the offset for mmap(2)
@@ -635,12 +638,26 @@ struct io_uring_buf_ring {
 	};
 };
 
+/*
+ * Flags for IORING_REGISTER_PBUF_RING.
+ *
+ * IOU_PBUF_RING_MMAP:	If set, kernel will allocate the memory for the ring.
+ *			The application must not set a ring_addr in struct
+ *			io_uring_buf_reg, instead it must subsequently call
+ *			mmap(2) with the offset set as:
+ *			IORING_OFF_PBUF_RING | (bgid << IORING_OFF_PBUF_SHIFT)
+ *			to get a virtual mapping for the ring.
+ */
+enum {
+	IOU_PBUF_RING_MMAP	= 1,
+};
+
 /* argument for IORING_(UN)REGISTER_PBUF_RING */
 struct io_uring_buf_reg {
 	__u64	ring_addr;
 	__u32	ring_entries;
 	__u16	bgid;
-	__u16	pad;
+	__u16	flags;
 	__u64	resv[3];
 };
 

src/setup.c

@@ -369,25 +369,57 @@ __cold ssize_t io_uring_mlock_size(unsigned entries, unsigned flags)
 	return io_uring_mlock_size_params(entries, &p);
 }
 
-struct io_uring_buf_ring *io_uring_setup_buf_ring(struct io_uring *ring,
-						  unsigned int nentries,
-						  int bgid, unsigned int flags,
-						  int *ret)
+#if defined(__hppa__)
+static struct io_uring_buf_ring *br_setup(struct io_uring *ring,
+					  unsigned int nentries, int bgid,
+					  unsigned int flags, int *ret)
 {
 	struct io_uring_buf_reg reg = { };
 	struct io_uring_buf_ring *br;
 	size_t ring_size;
+	off_t off;
 	int lret;
 
+	reg.ring_entries = nentries;
+	reg.bgid = bgid;
+	reg.flags = IOU_PBUF_RING_MMAP;
+
+	*ret = 0;
+	lret = io_uring_register_buf_ring(ring, &reg, flags);
+	if (lret) {
+		*ret = lret;
+		return NULL;
+	}
+
+	off = IORING_OFF_PBUF_RING | (unsigned long long) bgid << IORING_OFF_PBUF_SHIFT;
 	ring_size = nentries * sizeof(struct io_uring_buf);
 	br = __sys_mmap(NULL, ring_size, PROT_READ | PROT_WRITE,
-			MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+			MAP_SHARED | MAP_POPULATE, ring->ring_fd, off);
 	if (IS_ERR(br)) {
 		*ret = PTR_ERR(br);
 		return NULL;
 	}
 
-	io_uring_buf_ring_init(br);
+	return br;
+
+}
+#else
+static struct io_uring_buf_ring *br_setup(struct io_uring *ring,
+					  unsigned int nentries, int bgid,
+					  unsigned int flags, int *ret)
+{
+	struct io_uring_buf_reg reg = { };
+	struct io_uring_buf_ring *br;
+	size_t ring_size;
+	int lret;
+
+	ring_size = nentries * sizeof(struct io_uring_buf);
+	br = __sys_mmap(NULL, ring_size, PROT_READ | PROT_WRITE,
+			MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+	if (IS_ERR(br)) {
+		*ret = PTR_ERR(br);
+		return NULL;
+	}
 
 	reg.ring_addr = (unsigned long) (uintptr_t) br;
 	reg.ring_entries = nentries;
@@ -402,6 +434,22 @@ struct io_uring_buf_ring *io_uring_setup_buf_ring(struct io_uring *ring,
 	}
 
 	return br;
+
+}
+#endif
+
+struct io_uring_buf_ring *io_uring_setup_buf_ring(struct io_uring *ring,
+						  unsigned int nentries,
+						  int bgid, unsigned int flags,
+						  int *ret)
+{
+	struct io_uring_buf_ring *br;
+
+	br = br_setup(ring, nentries, bgid, flags, ret);
+	if (br)
+		io_uring_buf_ring_init(br);
+
+	return br;
 }
 
 int io_uring_free_buf_ring(struct io_uring *ring, struct io_uring_buf_ring *br,