@@ -3,6 +3,7 @@
config LIBETH
tristate
+ select PAGE_POOL
help
libeth is a common library containing routines shared between several
drivers, but not yet promoted to the generic kernel API.
@@ -3,6 +3,104 @@
#include <net/libeth/rx.h>
+/* Rx buffer management */
+
+/**
+ * libeth_rx_hw_len - get the actual buffer size to be passed to HW
+ * @pp: &page_pool_params of the netdev to calculate the size for
+ * @max_len: maximum buffer size for a single descriptor
+ *
+ * Return: HW-writeable length per one buffer to pass it to the HW accounting:
+ * MTU the @dev has, HW required alignment, minimum and maximum allowed values,
+ * and system's page size.
+ */
+static u32 libeth_rx_hw_len(const struct page_pool_params *pp, u32 max_len)
+{
+ u32 len;
+
+ len = READ_ONCE(pp->netdev->mtu) + LIBETH_RX_LL_LEN;
+ len = ALIGN(len, LIBETH_RX_BUF_STRIDE);
+ len = min3(len, ALIGN_DOWN(max_len ? : U32_MAX, LIBETH_RX_BUF_STRIDE),
+ pp->max_len);
+
+ return len;
+}
+
+/**
+ * libeth_rx_fq_create - create a PP with the default libeth settings
+ * @fq: buffer queue struct to fill
+ * @napi: &napi_struct covering this PP (no usage outside its poll loops)
+ *
+ * Return: %0 on success, -%errno on failure.
+ */
+int libeth_rx_fq_create(struct libeth_fq *fq, struct napi_struct *napi)
+{
+ struct page_pool_params pp = {
+ .flags = PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV,
+ .order = LIBETH_RX_PAGE_ORDER,
+ .pool_size = fq->count,
+ .nid = fq->nid,
+ .dev = napi->dev->dev.parent,
+ .netdev = napi->dev,
+ .napi = napi,
+ .dma_dir = DMA_FROM_DEVICE,
+ .offset = LIBETH_SKB_HEADROOM,
+ };
+ struct libeth_fqe *fqes;
+ struct page_pool *pool;
+
+ /* HW-writeable / syncable length per one page */
+ pp.max_len = LIBETH_RX_PAGE_LEN(pp.offset);
+
+ /* HW-writeable length per buffer */
+ fq->buf_len = libeth_rx_hw_len(&pp, fq->buf_len);
+ /* Buffer size to allocate */
+ fq->truesize = roundup_pow_of_two(SKB_HEAD_ALIGN(pp.offset +
+ fq->buf_len));
+
+ pool = page_pool_create(&pp);
+ if (IS_ERR(pool))
+ return PTR_ERR(pool);
+
+ fqes = kvcalloc_node(fq->count, sizeof(*fqes), GFP_KERNEL, fq->nid);
+ if (!fqes)
+ goto err_buf;
+
+ fq->fqes = fqes;
+ fq->pp = pool;
+
+ return 0;
+
+err_buf:
+ page_pool_destroy(pool);
+
+ return -ENOMEM;
+}
+EXPORT_SYMBOL_NS_GPL(libeth_rx_fq_create, LIBETH);
+
+/**
+ * libeth_rx_fq_destroy - destroy a &page_pool created by libeth
+ * @fq: buffer queue to process
+ */
+void libeth_rx_fq_destroy(struct libeth_fq *fq)
+{
+ kvfree(fq->fqes);
+ page_pool_destroy(fq->pp);
+}
+EXPORT_SYMBOL_NS_GPL(libeth_rx_fq_destroy, LIBETH);
+
+/**
+ * libeth_rx_recycle_slow - recycle a libeth page from the NAPI context
+ * @page: page to recycle
+ *
+ * To be used on exceptions or rare cases not requiring fast inline recycling.
+ */
+void libeth_rx_recycle_slow(struct page *page)
+{
+ page_pool_recycle_direct(page->pp, page);
+}
+EXPORT_SYMBOL_NS_GPL(libeth_rx_recycle_slow, LIBETH);
+
/* Converting abstract packet type numbers into a software structure with
* the packet parameters to do O(1) lookup on Rx.
*/
@@ -4,8 +4,125 @@
#ifndef __LIBETH_RX_H
#define __LIBETH_RX_H
+#include <linux/if_vlan.h>
+
+#include <net/page_pool/helpers.h>
#include <net/xdp.h>
+/* Rx buffer management */
+
+/* Space reserved in front of each frame */
+#define LIBETH_SKB_HEADROOM (NET_SKB_PAD + NET_IP_ALIGN)
+/* Maximum headroom for worst-case calculations */
+#define LIBETH_MAX_HEADROOM LIBETH_SKB_HEADROOM
+/* Link layer / L2 overhead: Ethernet, 2 VLAN tags (C + S), FCS */
+#define LIBETH_RX_LL_LEN (ETH_HLEN + 2 * VLAN_HLEN + ETH_FCS_LEN)
+
+/* Always use order-0 pages */
+#define LIBETH_RX_PAGE_ORDER 0
+/* Pick a sane buffer stride and align to a cacheline boundary */
+#define LIBETH_RX_BUF_STRIDE SKB_DATA_ALIGN(128)
+/* HW-writeable space in one buffer: truesize - headroom/tailroom, aligned */
+#define LIBETH_RX_PAGE_LEN(hr) \
+ ALIGN_DOWN(SKB_MAX_ORDER(hr, LIBETH_RX_PAGE_ORDER), \
+ LIBETH_RX_BUF_STRIDE)
+
+/**
+ * struct libeth_fqe - structure representing an Rx buffer (fill queue element)
+ * @page: page holding the buffer
+ * @offset: offset from the page start (to the headroom)
+ * @truesize: total space occupied by the buffer (w/ headroom and tailroom)
+ *
+ * Depending on the MTU, API switches between one-page-per-frame and shared
+ * page model (to conserve memory on bigger-page platforms). In case of the
+ * former, @offset is always 0 and @truesize is always ```PAGE_SIZE```.
+ */
+struct libeth_fqe {
+ struct page *page;
+ u32 offset;
+ u32 truesize;
+} __aligned_largest;
+
+/**
+ * struct libeth_fq - structure representing a buffer (fill) queue
+ * @fp: hotpath part of the structure
+ * @pp: &page_pool for buffer management
+ * @fqes: array of Rx buffers
+ * @truesize: size to allocate per buffer, w/overhead
+ * @count: number of descriptors/buffers the queue has
+ * @buf_len: HW-writeable length per each buffer
+ * @nid: ID of the closest NUMA node with memory
+ */
+struct libeth_fq {
+ struct_group_tagged(libeth_fq_fp, fp,
+ struct page_pool *pp;
+ struct libeth_fqe *fqes;
+
+ u32 truesize;
+ u32 count;
+ );
+
+ /* Cold fields */
+ u32 buf_len;
+ int nid;
+};
+
+int libeth_rx_fq_create(struct libeth_fq *fq, struct napi_struct *napi);
+void libeth_rx_fq_destroy(struct libeth_fq *fq);
+
+/**
+ * libeth_rx_alloc - allocate a new Rx buffer
+ * @fq: fill queue to allocate for
+ * @i: index of the buffer within the queue
+ *
+ * Return: DMA address to be passed to HW for Rx on successful allocation,
+ * ```DMA_MAPPING_ERROR``` otherwise.
+ */
+static inline dma_addr_t libeth_rx_alloc(const struct libeth_fq_fp *fq, u32 i)
+{
+ struct libeth_fqe *buf = &fq->fqes[i];
+
+ buf->truesize = fq->truesize;
+ buf->page = page_pool_dev_alloc(fq->pp, &buf->offset, &buf->truesize);
+ if (unlikely(!buf->page))
+ return DMA_MAPPING_ERROR;
+
+ return page_pool_get_dma_addr(buf->page) + buf->offset +
+ fq->pp->p.offset;
+}
+
+void libeth_rx_recycle_slow(struct page *page);
+
+/**
+ * libeth_rx_sync_for_cpu - synchronize or recycle buffer post DMA
+ * @fqe: buffer to process
+ * @len: frame length from the descriptor
+ *
+ * Process the buffer after it's written by HW. The regular path is to
+ * synchronize DMA for CPU, but in case of no data it will be immediately
+ * recycled back to its PP.
+ *
+ * Return: true when there's data to process, false otherwise.
+ */
+static inline bool libeth_rx_sync_for_cpu(const struct libeth_fqe *fqe,
+ u32 len)
+{
+ struct page *page = fqe->page;
+
+ /* Very rare, but possible case. The most common reason:
+ * the last fragment contained FCS only, which was then
+ * stripped by the HW.
+ */
+ if (unlikely(!len)) {
+ libeth_rx_recycle_slow(page);
+ return false;
+ }
+
+ page_pool_dma_sync_for_cpu(page->pp, page, fqe->offset, len);
+
+ return true;
+}
+
/* Converting abstract packet type numbers into a software structure with
* the packet parameters to do O(1) lookup on Rx.
*/