From 2cb13dec8c5e5e104fd2f71c2dee761d6ed9a333 Mon Sep 17 00:00:00 2001 From: Alexander Lobakin Date: Thu, 20 Jun 2024 15:53:34 +0200 Subject: cache: add __cacheline_group_{begin, end}_aligned() (+ couple more) __cacheline_group_begin(), unfortunately, doesn't align the group anyhow. If it is wanted, then you need to do something like __cacheline_group_begin(grp) __aligned(ALIGN) which isn't really convenient nor compact. Add the _aligned() counterparts to align the groups automatically to either the specified alignment (optional) or ``SMP_CACHE_BYTES``. Note that the actual struct layout will then be (on x64 with 64-byte CL): struct x { u32 y; // offset 0, size 4, padding 56 __cacheline_group_begin__grp; // offset 64, size 0 u32 z; // offset 64, size 4, padding 4 __cacheline_group_end__grp; // offset 72, size 0 __cacheline_group_pad__grp; // offset 72, size 0, padding 56 u32 w; // offset 128 }; The end marker is aligned to long, so that you can assert the struct size more strictly, but the offset of the next field in the structure will be aligned to the group alignment, so that the next field won't fall into the group it's not intended to. Add __LARGEST_ALIGN definition and LARGEST_ALIGN() macro. __LARGEST_ALIGN is the value to which the compilers align fields when __aligned_largest is specified. Sometimes, it might be needed to get this value outside of variable definitions. LARGEST_ALIGN() is macro which just aligns a value to __LARGEST_ALIGN. Also add SMP_CACHE_ALIGN(), similar to L1_CACHE_ALIGN(), but using ``SMP_CACHE_BYTES`` instead of ``L1_CACHE_BYTES`` as the former also accounts L2, needed in some cases. Signed-off-by: Alexander Lobakin Reviewed-by: Przemek Kitszel Signed-off-by: Tony Nguyen --- include/linux/cache.h | 59 +++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 59 insertions(+) (limited to 'include') diff --git a/include/linux/cache.h b/include/linux/cache.h index 0ecb17bb6883..ca2a05682a54 100644 --- a/include/linux/cache.h +++ b/include/linux/cache.h @@ -13,6 +13,32 @@ #define SMP_CACHE_BYTES L1_CACHE_BYTES #endif +/** + * SMP_CACHE_ALIGN - align a value to the L2 cacheline size + * @x: value to align + * + * On some architectures, L2 ("SMP") CL size is bigger than L1, and sometimes, + * this needs to be accounted. + * + * Return: aligned value. + */ +#ifndef SMP_CACHE_ALIGN +#define SMP_CACHE_ALIGN(x) ALIGN(x, SMP_CACHE_BYTES) +#endif + +/* + * ``__aligned_largest`` aligns a field to the value most optimal for the + * target architecture to perform memory operations. Get the actual value + * to be able to use it anywhere else. + */ +#ifndef __LARGEST_ALIGN +#define __LARGEST_ALIGN sizeof(struct { long x; } __aligned_largest) +#endif + +#ifndef LARGEST_ALIGN +#define LARGEST_ALIGN(x) ALIGN(x, __LARGEST_ALIGN) +#endif + /* * __read_mostly is used to keep rarely changing variables out of frequently * updated cachelines. Its use should be reserved for data that is used @@ -95,6 +121,39 @@ __u8 __cacheline_group_end__##GROUP[0] #endif +/** + * __cacheline_group_begin_aligned - declare an aligned group start + * @GROUP: name of the group + * @...: optional group alignment + * + * The following block inside a struct: + * + * __cacheline_group_begin_aligned(grp); + * field a; + * field b; + * __cacheline_group_end_aligned(grp); + * + * will always be aligned to either the specified alignment or + * ``SMP_CACHE_BYTES``. + */ +#define __cacheline_group_begin_aligned(GROUP, ...) \ + __cacheline_group_begin(GROUP) \ + __aligned((__VA_ARGS__ + 0) ? : SMP_CACHE_BYTES) + +/** + * __cacheline_group_end_aligned - declare an aligned group end + * @GROUP: name of the group + * @...: optional alignment (same as was in __cacheline_group_begin_aligned()) + * + * Note that the end marker is aligned to sizeof(long) to allow more precise + * size assertion. It also declares a padding at the end to avoid next field + * falling into this cacheline. + */ +#define __cacheline_group_end_aligned(GROUP, ...) \ + __cacheline_group_end(GROUP) __aligned(sizeof(long)); \ + struct { } __cacheline_group_pad__##GROUP \ + __aligned((__VA_ARGS__ + 0) ? : SMP_CACHE_BYTES) + #ifndef CACHELINE_ASSERT_GROUP_MEMBER #define CACHELINE_ASSERT_GROUP_MEMBER(TYPE, GROUP, MEMBER) \ BUILD_BUG_ON(!(offsetof(TYPE, MEMBER) >= \ -- cgit v1.2.3 From 39daa09d34ada1bc7227d68def63e0a2105b5496 Mon Sep 17 00:00:00 2001 From: Alexander Lobakin Date: Thu, 20 Jun 2024 15:53:35 +0200 Subject: page_pool: use __cacheline_group_{begin, end}_aligned() Instead of doing __cacheline_group_begin() __aligned(), use the new __cacheline_group_{begin,end}_aligned(), so that it will take care of the group alignment itself. Also replace open-coded `4 * sizeof(long)` in two places with a definition. Signed-off-by: Alexander Lobakin Signed-off-by: Tony Nguyen --- include/net/page_pool/types.h | 22 ++++++++++++---------- net/core/page_pool.c | 3 ++- 2 files changed, 14 insertions(+), 11 deletions(-) (limited to 'include') diff --git a/include/net/page_pool/types.h b/include/net/page_pool/types.h index b70bcc14ceda..50569fed7868 100644 --- a/include/net/page_pool/types.h +++ b/include/net/page_pool/types.h @@ -129,6 +129,16 @@ struct page_pool_stats { }; #endif +/* The whole frag API block must stay within one cacheline. On 32-bit systems, + * sizeof(long) == sizeof(int), so that the block size is ``3 * sizeof(long)``. + * On 64-bit systems, the actual size is ``2 * sizeof(long) + sizeof(int)``. + * The closest pow-2 to both of them is ``4 * sizeof(long)``, so just use that + * one for simplicity. + * Having it aligned to a cacheline boundary may be excessive and doesn't bring + * any good. + */ +#define PAGE_POOL_FRAG_GROUP_ALIGN (4 * sizeof(long)) + struct page_pool { struct page_pool_params_fast p; @@ -142,19 +152,11 @@ struct page_pool { bool system:1; /* This is a global percpu pool */ #endif - /* The following block must stay within one cacheline. On 32-bit - * systems, sizeof(long) == sizeof(int), so that the block size is - * ``3 * sizeof(long)``. On 64-bit systems, the actual size is - * ``2 * sizeof(long) + sizeof(int)``. The closest pow-2 to both of - * them is ``4 * sizeof(long)``, so just use that one for simplicity. - * Having it aligned to a cacheline boundary may be excessive and - * doesn't bring any good. - */ - __cacheline_group_begin(frag) __aligned(4 * sizeof(long)); + __cacheline_group_begin_aligned(frag, PAGE_POOL_FRAG_GROUP_ALIGN); long frag_users; netmem_ref frag_page; unsigned int frag_offset; - __cacheline_group_end(frag); + __cacheline_group_end_aligned(frag, PAGE_POOL_FRAG_GROUP_ALIGN); struct delayed_work release_dw; void (*disconnect)(void *pool); diff --git a/net/core/page_pool.c b/net/core/page_pool.c index 855271a6cad2..2abe6e919224 100644 --- a/net/core/page_pool.c +++ b/net/core/page_pool.c @@ -178,7 +178,8 @@ static void page_pool_struct_check(void) CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_users); CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_page); CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_offset); - CACHELINE_ASSERT_GROUP_SIZE(struct page_pool, frag, 4 * sizeof(long)); + CACHELINE_ASSERT_GROUP_SIZE(struct page_pool, frag, + PAGE_POOL_FRAG_GROUP_ALIGN); } static int page_pool_init(struct page_pool *pool, -- cgit v1.2.3 From 62c884256ea1f898b00f20729a306a4eeb1840b1 Mon Sep 17 00:00:00 2001 From: Alexander Lobakin Date: Thu, 20 Jun 2024 15:53:36 +0200 Subject: libeth: add cacheline / struct layout assertion helpers Add helpers to assert struct field layout, a bit more crazy and networking-specific than in . They assume you have 3 CL-aligned groups (read-mostly, read-write, cold) in a struct you want to assert, and nothing besides them. For 64-bit with 64-byte cachelines, the assertions are as strict as possible, as the size can then be easily predicted. For the rest, make sure they don't cross the specified bound. Signed-off-by: Alexander Lobakin Signed-off-by: Tony Nguyen --- include/net/libeth/cache.h | 66 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 66 insertions(+) create mode 100644 include/net/libeth/cache.h (limited to 'include') diff --git a/include/net/libeth/cache.h b/include/net/libeth/cache.h new file mode 100644 index 000000000000..bdb0c043ce61 --- /dev/null +++ b/include/net/libeth/cache.h @@ -0,0 +1,66 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ +/* Copyright (C) 2024 Intel Corporation */ + +#ifndef __LIBETH_CACHE_H +#define __LIBETH_CACHE_H + +#include + +/** + * libeth_cacheline_group_assert - make sure cacheline group size is expected + * @type: type of the structure containing the group + * @grp: group name inside the struct + * @sz: expected group size + */ +#if defined(CONFIG_64BIT) && SMP_CACHE_BYTES == 64 +#define libeth_cacheline_group_assert(type, grp, sz) \ + static_assert(offsetof(type, __cacheline_group_end__##grp) - \ + offsetofend(type, __cacheline_group_begin__##grp) == \ + (sz)) +#define __libeth_cacheline_struct_assert(type, sz) \ + static_assert(sizeof(type) == (sz)) +#else /* !CONFIG_64BIT || SMP_CACHE_BYTES != 64 */ +#define libeth_cacheline_group_assert(type, grp, sz) \ + static_assert(offsetof(type, __cacheline_group_end__##grp) - \ + offsetofend(type, __cacheline_group_begin__##grp) <= \ + (sz)) +#define __libeth_cacheline_struct_assert(type, sz) \ + static_assert(sizeof(type) <= (sz)) +#endif /* !CONFIG_64BIT || SMP_CACHE_BYTES != 64 */ + +#define __libeth_cls1(sz1) SMP_CACHE_ALIGN(sz1) +#define __libeth_cls2(sz1, sz2) (SMP_CACHE_ALIGN(sz1) + SMP_CACHE_ALIGN(sz2)) +#define __libeth_cls3(sz1, sz2, sz3) \ + (SMP_CACHE_ALIGN(sz1) + SMP_CACHE_ALIGN(sz2) + SMP_CACHE_ALIGN(sz3)) +#define __libeth_cls(...) \ + CONCATENATE(__libeth_cls, COUNT_ARGS(__VA_ARGS__))(__VA_ARGS__) + +/** + * libeth_cacheline_struct_assert - make sure CL-based struct size is expected + * @type: type of the struct + * @...: from 1 to 3 CL group sizes (read-mostly, read-write, cold) + * + * When a struct contains several CL groups, it's difficult to predict its size + * on different architectures. The macro instead takes sizes of all of the + * groups the structure contains and generates the final struct size. + */ +#define libeth_cacheline_struct_assert(type, ...) \ + __libeth_cacheline_struct_assert(type, __libeth_cls(__VA_ARGS__)); \ + static_assert(__alignof(type) >= SMP_CACHE_BYTES) + +/** + * libeth_cacheline_set_assert - make sure CL-based struct layout is expected + * @type: type of the struct + * @ro: expected size of the read-mostly group + * @rw: expected size of the read-write group + * @c: expected size of the cold group + * + * Check that each group size is expected and then do final struct size check. + */ +#define libeth_cacheline_set_assert(type, ro, rw, c) \ + libeth_cacheline_group_assert(type, read_mostly, ro); \ + libeth_cacheline_group_assert(type, read_write, rw); \ + libeth_cacheline_group_assert(type, cold, c); \ + libeth_cacheline_struct_assert(type, ro, rw, c) + +#endif /* __LIBETH_CACHE_H */ -- cgit v1.2.3 From 5aaac1aece4e4db9d620791a33f7a4173c660e65 Mon Sep 17 00:00:00 2001 From: Alexander Lobakin Date: Thu, 20 Jun 2024 15:53:45 +0200 Subject: libeth: support different types of buffers for Rx Unlike previous generations, idpf requires more buffer types for optimal performance. This includes: header buffers, short buffers, and no-overhead buffers (w/o headroom and tailroom, for TCP zerocopy when the header split is enabled). Introduce libeth Rx buffer type and calculate page_pool params accordingly. All the HW-related details like buffer alignment are still accounted. For the header buffers, pick 256 bytes as in most places in the kernel (have you ever seen frames with bigger headers?). Reviewed-by: Przemek Kitszel Signed-off-by: Alexander Lobakin Signed-off-by: Tony Nguyen --- drivers/net/ethernet/intel/libeth/rx.c | 132 ++++++++++++++++++++++++++++++--- include/net/libeth/rx.h | 19 +++++ 2 files changed, 140 insertions(+), 11 deletions(-) (limited to 'include') diff --git a/drivers/net/ethernet/intel/libeth/rx.c b/drivers/net/ethernet/intel/libeth/rx.c index 6221b88c34ac..d0b158b6e55b 100644 --- a/drivers/net/ethernet/intel/libeth/rx.c +++ b/drivers/net/ethernet/intel/libeth/rx.c @@ -6,7 +6,7 @@ /* Rx buffer management */ /** - * libeth_rx_hw_len - get the actual buffer size to be passed to HW + * libeth_rx_hw_len_mtu - 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 * @@ -14,7 +14,7 @@ * 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) +static u32 libeth_rx_hw_len_mtu(const struct page_pool_params *pp, u32 max_len) { u32 len; @@ -26,6 +26,118 @@ static u32 libeth_rx_hw_len(const struct page_pool_params *pp, u32 max_len) return len; } +/** + * libeth_rx_hw_len_truesize - get the short 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 + * @truesize: desired truesize for the buffers + * + * Return: HW-writeable length per one buffer to pass it to the HW ignoring the + * MTU and closest to the passed truesize. Can be used for "short" buffer + * queues to fragment pages more efficiently. + */ +static u32 libeth_rx_hw_len_truesize(const struct page_pool_params *pp, + u32 max_len, u32 truesize) +{ + u32 min, len; + + min = SKB_HEAD_ALIGN(pp->offset + LIBETH_RX_BUF_STRIDE); + truesize = clamp(roundup_pow_of_two(truesize), roundup_pow_of_two(min), + PAGE_SIZE << LIBETH_RX_PAGE_ORDER); + + len = SKB_WITH_OVERHEAD(truesize - pp->offset); + len = ALIGN_DOWN(len, LIBETH_RX_BUF_STRIDE) ? : LIBETH_RX_BUF_STRIDE; + len = min3(len, ALIGN_DOWN(max_len ? : U32_MAX, LIBETH_RX_BUF_STRIDE), + pp->max_len); + + return len; +} + +/** + * libeth_rx_page_pool_params - calculate params with the stack overhead + * @fq: buffer queue to calculate the size for + * @pp: &page_pool_params of the netdev + * + * Set the PP params to will all needed stack overhead (headroom, tailroom) and + * both the HW buffer length and the truesize for all types of buffers. For + * "short" buffers, truesize never exceeds the "wanted" one; for the rest, + * it can be up to the page size. + * + * Return: true on success, false on invalid input params. + */ +static bool libeth_rx_page_pool_params(struct libeth_fq *fq, + struct page_pool_params *pp) +{ + pp->offset = LIBETH_SKB_HEADROOM; + /* HW-writeable / syncable length per one page */ + pp->max_len = LIBETH_RX_PAGE_LEN(pp->offset); + + /* HW-writeable length per buffer */ + switch (fq->type) { + case LIBETH_FQE_MTU: + fq->buf_len = libeth_rx_hw_len_mtu(pp, fq->buf_len); + break; + case LIBETH_FQE_SHORT: + fq->buf_len = libeth_rx_hw_len_truesize(pp, fq->buf_len, + fq->truesize); + break; + case LIBETH_FQE_HDR: + fq->buf_len = ALIGN(LIBETH_MAX_HEAD, LIBETH_RX_BUF_STRIDE); + break; + default: + return false; + } + + /* Buffer size to allocate */ + fq->truesize = roundup_pow_of_two(SKB_HEAD_ALIGN(pp->offset + + fq->buf_len)); + + return true; +} + +/** + * libeth_rx_page_pool_params_zc - calculate params without the stack overhead + * @fq: buffer queue to calculate the size for + * @pp: &page_pool_params of the netdev + * + * Set the PP params to exclude the stack overhead and both the buffer length + * and the truesize, which are equal for the data buffers. Note that this + * requires separate header buffers to be always active and account the + * overhead. + * With the MTU == ``PAGE_SIZE``, this allows the kernel to enable the zerocopy + * mode. + * + * Return: true on success, false on invalid input params. + */ +static bool libeth_rx_page_pool_params_zc(struct libeth_fq *fq, + struct page_pool_params *pp) +{ + u32 mtu, max; + + pp->offset = 0; + pp->max_len = PAGE_SIZE << LIBETH_RX_PAGE_ORDER; + + switch (fq->type) { + case LIBETH_FQE_MTU: + mtu = READ_ONCE(pp->netdev->mtu); + break; + case LIBETH_FQE_SHORT: + mtu = fq->truesize; + break; + default: + return false; + } + + mtu = roundup_pow_of_two(mtu); + max = min(rounddown_pow_of_two(fq->buf_len ? : U32_MAX), + pp->max_len); + + fq->buf_len = clamp(mtu, LIBETH_RX_BUF_STRIDE, max); + fq->truesize = fq->buf_len; + + return true; +} + /** * libeth_rx_fq_create - create a PP with the default libeth settings * @fq: buffer queue struct to fill @@ -44,19 +156,17 @@ int libeth_rx_fq_create(struct libeth_fq *fq, struct napi_struct *napi) .netdev = napi->dev, .napi = napi, .dma_dir = DMA_FROM_DEVICE, - .offset = LIBETH_SKB_HEADROOM, }; struct libeth_fqe *fqes; struct page_pool *pool; + bool ret; - /* 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)); + if (!fq->hsplit) + ret = libeth_rx_page_pool_params(fq, &pp); + else + ret = libeth_rx_page_pool_params_zc(fq, &pp); + if (!ret) + return -EINVAL; pool = page_pool_create(&pp); if (IS_ERR(pool)) diff --git a/include/net/libeth/rx.h b/include/net/libeth/rx.h index f29ea3e34c6c..43574bd6612f 100644 --- a/include/net/libeth/rx.h +++ b/include/net/libeth/rx.h @@ -17,6 +17,8 @@ #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) +/* Maximum supported L2-L4 header length */ +#define LIBETH_MAX_HEAD roundup_pow_of_two(max(MAX_HEADER, 256)) /* Always use order-0 pages */ #define LIBETH_RX_PAGE_ORDER 0 @@ -43,6 +45,18 @@ struct libeth_fqe { u32 truesize; } __aligned_largest; +/** + * enum libeth_fqe_type - enum representing types of Rx buffers + * @LIBETH_FQE_MTU: buffer size is determined by MTU + * @LIBETH_FQE_SHORT: buffer size is smaller than MTU, for short frames + * @LIBETH_FQE_HDR: buffer size is ```LIBETH_MAX_HEAD```-sized, for headers + */ +enum libeth_fqe_type { + LIBETH_FQE_MTU = 0U, + LIBETH_FQE_SHORT, + LIBETH_FQE_HDR, +}; + /** * struct libeth_fq - structure representing a buffer (fill) queue * @fp: hotpath part of the structure @@ -50,6 +64,8 @@ struct libeth_fqe { * @fqes: array of Rx buffers * @truesize: size to allocate per buffer, w/overhead * @count: number of descriptors/buffers the queue has + * @type: type of the buffers this queue has + * @hsplit: flag whether header split is enabled * @buf_len: HW-writeable length per each buffer * @nid: ID of the closest NUMA node with memory */ @@ -63,6 +79,9 @@ struct libeth_fq { ); /* Cold fields */ + enum libeth_fqe_type type:2; + bool hsplit:1; + u32 buf_len; int nid; }; -- cgit v1.2.3