From 25744f849524e806a13ade17c4fb83f6888fe954 Mon Sep 17 00:00:00 2001 From: Pavel Begunkov Date: Tue, 15 Apr 2025 14:09:45 +0100 Subject: [PATCH 1/8] io_uring/zcrx: return ifq id to the user IORING_OP_RECV_ZC requests take a zcrx object id via sqe::zcrx_ifq_idx, which binds it to the corresponding if / queue. However, we don't return that id back to the user. It's fine as currently there can be only one zcrx and the user assumes that its id should be 0, but as we'll need multiple zcrx objects in the future let's explicitly pass it back on registration. Signed-off-by: Pavel Begunkov Link: https://lore.kernel.org/r/8714667d370651962f7d1a169032e5f02682a73e.1744722517.git.asml.silence@gmail.com Signed-off-by: Jens Axboe --- include/uapi/linux/io_uring.h | 4 +++- io_uring/zcrx.c | 3 ++- 2 files changed, 5 insertions(+), 2 deletions(-) diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h index ed2beb4def3f..8f1fc12bac46 100644 --- a/include/uapi/linux/io_uring.h +++ b/include/uapi/linux/io_uring.h @@ -1010,7 +1010,9 @@ struct io_uring_zcrx_ifq_reg { __u64 region_ptr; /* struct io_uring_region_desc * */ struct io_uring_zcrx_offsets offsets; - __u64 __resv[4]; + __u32 zcrx_id; + __u32 __resv2; + __u64 __resv[3]; }; #ifdef __cplusplus diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c index 0f46e0404c04..d0eccf277a20 100644 --- a/io_uring/zcrx.c +++ b/io_uring/zcrx.c @@ -354,7 +354,8 @@ int io_register_zcrx_ifq(struct io_ring_ctx *ctx, return -EFAULT; if (copy_from_user(&rd, u64_to_user_ptr(reg.region_ptr), sizeof(rd))) return -EFAULT; - if (memchr_inv(®.__resv, 0, sizeof(reg.__resv))) + if (memchr_inv(®.__resv, 0, sizeof(reg.__resv)) || + reg.__resv2 || reg.zcrx_id) return -EINVAL; if (reg.if_rxq == -1 || !reg.rq_entries || reg.flags) return -EINVAL; From 70e4f9bfc13c9abcc97eb9f2feee51cc925524c8 Mon Sep 17 00:00:00 2001 From: Pavel Begunkov Date: Tue, 15 Apr 2025 14:10:16 +0100 Subject: [PATCH 2/8] io_uring/zcrx: add pp to ifq conversion helper It'll likely change how page pools store memory providers, so in preparation for that, keep accesses in one place in io_uring by introducing a helper. Signed-off-by: Pavel Begunkov Link: https://lore.kernel.org/r/3522eb8fa9b4e21bcf32e7e9ae656c616b282210.1744722526.git.asml.silence@gmail.com Signed-off-by: Jens Axboe --- io_uring/zcrx.c | 13 +++++++++---- 1 file changed, 9 insertions(+), 4 deletions(-) diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c index d0eccf277a20..5defbe8f95f9 100644 --- a/io_uring/zcrx.c +++ b/io_uring/zcrx.c @@ -26,6 +26,11 @@ #include "zcrx.h" #include "rsrc.h" +static inline struct io_zcrx_ifq *io_pp_to_ifq(struct page_pool *pp) +{ + return pp->mp_priv; +} + #define IO_DMA_ATTR (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING) static void __io_zcrx_unmap_area(struct io_zcrx_ifq *ifq, @@ -586,7 +591,7 @@ static void io_zcrx_refill_slow(struct page_pool *pp, struct io_zcrx_ifq *ifq) static netmem_ref io_pp_zc_alloc_netmems(struct page_pool *pp, gfp_t gfp) { - struct io_zcrx_ifq *ifq = pp->mp_priv; + struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); /* pp should already be ensuring that */ if (unlikely(pp->alloc.count)) @@ -618,7 +623,7 @@ static bool io_pp_zc_release_netmem(struct page_pool *pp, netmem_ref netmem) static int io_pp_zc_init(struct page_pool *pp) { - struct io_zcrx_ifq *ifq = pp->mp_priv; + struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); if (WARN_ON_ONCE(!ifq)) return -EINVAL; @@ -637,7 +642,7 @@ static int io_pp_zc_init(struct page_pool *pp) static void io_pp_zc_destroy(struct page_pool *pp) { - struct io_zcrx_ifq *ifq = pp->mp_priv; + struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); struct io_zcrx_area *area = ifq->area; if (WARN_ON_ONCE(area->free_count != area->nia.num_niovs)) @@ -792,7 +797,7 @@ static int io_zcrx_recv_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq, niov = netmem_to_net_iov(frag->netmem); if (niov->pp->mp_ops != &io_uring_pp_zc_ops || - niov->pp->mp_priv != ifq) + io_pp_to_ifq(niov->pp) != ifq) return -EFAULT; if (!io_zcrx_queue_cqe(req, niov, ifq, off + skb_frag_off(frag), len)) From 1ac571288822253db32196c49f240739148417e3 Mon Sep 17 00:00:00 2001 From: Pavel Begunkov Date: Thu, 17 Apr 2025 10:32:31 +0100 Subject: [PATCH 3/8] io_uring/rsrc: don't skip offset calculation Don't optimise for requests with offset=0. Large registered buffers are the preference and hence the user is likely to pass an offset, and the adjustments are not expensive and will be made even cheaper in following patches. Signed-off-by: Pavel Begunkov Link: https://lore.kernel.org/r/1c2beb20470ee3c886a363d4d8340d3790db19f3.1744882081.git.asml.silence@gmail.com Signed-off-by: Jens Axboe --- io_uring/rsrc.c | 75 ++++++++++++++++++++++++------------------------- 1 file changed, 37 insertions(+), 38 deletions(-) diff --git a/io_uring/rsrc.c b/io_uring/rsrc.c index b36c8825550e..4d62897d1c89 100644 --- a/io_uring/rsrc.c +++ b/io_uring/rsrc.c @@ -1036,6 +1036,7 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, struct io_mapped_ubuf *imu, u64 buf_addr, size_t len) { + const struct bio_vec *bvec; size_t offset; int ret; @@ -1054,47 +1055,45 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, offset = buf_addr - imu->ubuf; iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); - if (offset) { - /* - * Don't use iov_iter_advance() here, as it's really slow for - * using the latter parts of a big fixed buffer - it iterates - * over each segment manually. We can cheat a bit here for user - * registered nodes, because we know that: - * - * 1) it's a BVEC iter, we set it up - * 2) all bvecs are the same in size, except potentially the - * first and last bvec - * - * So just find our index, and adjust the iterator afterwards. - * If the offset is within the first bvec (or the whole first - * bvec, just use iov_iter_advance(). This makes it easier - * since we can just skip the first segment, which may not - * be folio_size aligned. - */ - const struct bio_vec *bvec = imu->bvec; + /* + * Don't use iov_iter_advance() here, as it's really slow for + * using the latter parts of a big fixed buffer - it iterates + * over each segment manually. We can cheat a bit here for user + * registered nodes, because we know that: + * + * 1) it's a BVEC iter, we set it up + * 2) all bvecs are the same in size, except potentially the + * first and last bvec + * + * So just find our index, and adjust the iterator afterwards. + * If the offset is within the first bvec (or the whole first + * bvec, just use iov_iter_advance(). This makes it easier + * since we can just skip the first segment, which may not + * be folio_size aligned. + */ + bvec = imu->bvec; - /* - * Kernel buffer bvecs, on the other hand, don't necessarily - * have the size property of user registered ones, so we have - * to use the slow iter advance. - */ - if (offset < bvec->bv_len) { - iter->count -= offset; - iter->iov_offset = offset; - } else if (imu->is_kbuf) { - iov_iter_advance(iter, offset); - } else { - unsigned long seg_skip; + /* + * Kernel buffer bvecs, on the other hand, don't necessarily + * have the size property of user registered ones, so we have + * to use the slow iter advance. + */ + if (offset < bvec->bv_len) { + iter->count -= offset; + iter->iov_offset = offset; + } else if (imu->is_kbuf) { + iov_iter_advance(iter, offset); + } else { + unsigned long seg_skip; - /* skip first vec */ - offset -= bvec->bv_len; - seg_skip = 1 + (offset >> imu->folio_shift); + /* skip first vec */ + offset -= bvec->bv_len; + seg_skip = 1 + (offset >> imu->folio_shift); - iter->bvec += seg_skip; - iter->nr_segs -= seg_skip; - iter->count -= bvec->bv_len + offset; - iter->iov_offset = offset & ((1UL << imu->folio_shift) - 1); - } + iter->bvec += seg_skip; + iter->nr_segs -= seg_skip; + iter->count -= bvec->bv_len + offset; + iter->iov_offset = offset & ((1UL << imu->folio_shift) - 1); } return 0; From 50169d07548441e3033b9bbaa06e573e7224f140 Mon Sep 17 00:00:00 2001 From: Pavel Begunkov Date: Thu, 17 Apr 2025 10:32:32 +0100 Subject: [PATCH 4/8] io_uring/rsrc: separate kbuf offset adjustments Kernel registered buffers are special because segments are not uniform in size, and we have a bunch of optimisations based on that uniformity for normal buffers. Handle kbuf separately, it'll be cleaner this way. Signed-off-by: Pavel Begunkov Link: https://lore.kernel.org/r/4e9e5990b0ab5aee723c0be5cd9b5bcf810375f9.1744882081.git.asml.silence@gmail.com Signed-off-by: Jens Axboe --- io_uring/rsrc.c | 19 +++++++------------ 1 file changed, 7 insertions(+), 12 deletions(-) diff --git a/io_uring/rsrc.c b/io_uring/rsrc.c index 4d62897d1c89..fddde8ffe81e 100644 --- a/io_uring/rsrc.c +++ b/io_uring/rsrc.c @@ -1048,11 +1048,14 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, if (!(imu->dir & (1 << ddir))) return -EFAULT; - /* - * Might not be a start of buffer, set size appropriately - * and advance us to the beginning. - */ offset = buf_addr - imu->ubuf; + + if (imu->is_kbuf) { + iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); + iov_iter_advance(iter, offset); + return 0; + } + iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); /* @@ -1072,17 +1075,9 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, * be folio_size aligned. */ bvec = imu->bvec; - - /* - * Kernel buffer bvecs, on the other hand, don't necessarily - * have the size property of user registered ones, so we have - * to use the slow iter advance. - */ if (offset < bvec->bv_len) { iter->count -= offset; iter->iov_offset = offset; - } else if (imu->is_kbuf) { - iov_iter_advance(iter, offset); } else { unsigned long seg_skip; From 59852ebad954c8a3ac8b746930c2ea60febe797a Mon Sep 17 00:00:00 2001 From: Pavel Begunkov Date: Thu, 17 Apr 2025 10:32:33 +0100 Subject: [PATCH 5/8] io_uring/rsrc: refactor io_import_fixed io_import_fixed is a mess. Even though we know the final len of the iterator, we still assign offset + len and do some magic after to correct for that. Do offset calculation first and finalise it with iov_iter_bvec at the end. Signed-off-by: Pavel Begunkov Link: https://lore.kernel.org/r/2d5107fed24f8b23245ef2ede9a5a7f7c426df61.1744882081.git.asml.silence@gmail.com Signed-off-by: Jens Axboe --- io_uring/rsrc.c | 24 +++++++----------------- 1 file changed, 7 insertions(+), 17 deletions(-) diff --git a/io_uring/rsrc.c b/io_uring/rsrc.c index fddde8ffe81e..5cf854318b1d 100644 --- a/io_uring/rsrc.c +++ b/io_uring/rsrc.c @@ -1037,6 +1037,7 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, u64 buf_addr, size_t len) { const struct bio_vec *bvec; + unsigned nr_segs; size_t offset; int ret; @@ -1056,8 +1057,6 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, return 0; } - iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); - /* * Don't use iov_iter_advance() here, as it's really slow for * using the latter parts of a big fixed buffer - it iterates @@ -1067,30 +1066,21 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, * 1) it's a BVEC iter, we set it up * 2) all bvecs are the same in size, except potentially the * first and last bvec - * - * So just find our index, and adjust the iterator afterwards. - * If the offset is within the first bvec (or the whole first - * bvec, just use iov_iter_advance(). This makes it easier - * since we can just skip the first segment, which may not - * be folio_size aligned. */ bvec = imu->bvec; - if (offset < bvec->bv_len) { - iter->count -= offset; - iter->iov_offset = offset; - } else { + if (offset >= bvec->bv_len) { unsigned long seg_skip; /* skip first vec */ offset -= bvec->bv_len; seg_skip = 1 + (offset >> imu->folio_shift); - - iter->bvec += seg_skip; - iter->nr_segs -= seg_skip; - iter->count -= bvec->bv_len + offset; - iter->iov_offset = offset & ((1UL << imu->folio_shift) - 1); + bvec += seg_skip; + offset &= (1UL << imu->folio_shift) - 1; } + nr_segs = imu->nr_bvecs - (bvec - imu->bvec); + iov_iter_bvec(iter, ddir, bvec, nr_segs, len); + iter->iov_offset = offset; return 0; } From 80c7378f94cf193cb3bd2101bbcd5aea78d0e211 Mon Sep 17 00:00:00 2001 From: Nitesh Shetty Date: Thu, 17 Apr 2025 10:32:34 +0100 Subject: [PATCH 6/8] io_uring/rsrc: send exact nr_segs for fixed buffer Sending exact nr_segs, avoids bio split check and processing in block layer, which takes around 5%[1] of overall CPU utilization. In our setup, we see overall improvement of IOPS from 7.15M to 7.65M [2] and 5% less CPU utilization. [1] 3.52% io_uring [kernel.kallsyms] [k] bio_split_rw_at 1.42% io_uring [kernel.kallsyms] [k] bio_split_rw 0.62% io_uring [kernel.kallsyms] [k] bio_submit_split [2] sudo taskset -c 0,1 ./t/io_uring -b512 -d128 -c32 -s32 -p1 -F1 -B1 -n2 -r4 /dev/nvme0n1 /dev/nvme1n1 Signed-off-by: Nitesh Shetty [Pavel: fixed for kbuf, rebased and reworked on top of cleanups] Signed-off-by: Pavel Begunkov Link: https://lore.kernel.org/r/7a1a49a8d053bd617c244291d63dbfbc07afde36.1744882081.git.asml.silence@gmail.com [axboe: fold in fix factoring in buf reg offset] Signed-off-by: Jens Axboe --- io_uring/rsrc.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/io_uring/rsrc.c b/io_uring/rsrc.c index 5cf854318b1d..0c6d7e7415c8 100644 --- a/io_uring/rsrc.c +++ b/io_uring/rsrc.c @@ -1037,6 +1037,7 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, u64 buf_addr, size_t len) { const struct bio_vec *bvec; + size_t folio_mask; unsigned nr_segs; size_t offset; int ret; @@ -1067,6 +1068,7 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, * 2) all bvecs are the same in size, except potentially the * first and last bvec */ + folio_mask = (1UL << imu->folio_shift) - 1; bvec = imu->bvec; if (offset >= bvec->bv_len) { unsigned long seg_skip; @@ -1075,10 +1077,9 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, offset -= bvec->bv_len; seg_skip = 1 + (offset >> imu->folio_shift); bvec += seg_skip; - offset &= (1UL << imu->folio_shift) - 1; + offset &= folio_mask; } - - nr_segs = imu->nr_bvecs - (bvec - imu->bvec); + nr_segs = (offset + len + bvec->bv_offset + folio_mask) >> imu->folio_shift; iov_iter_bvec(iter, ddir, bvec, nr_segs, len); iter->iov_offset = offset; return 0; From b419bed4f0a62c65a57dd495185821dd56bc435c Mon Sep 17 00:00:00 2001 From: Jens Axboe Date: Wed, 16 Apr 2025 16:48:26 -0600 Subject: [PATCH 7/8] io_uring/rsrc: ensure segments counts are correct on kbuf buffers kbuf imports have the front offset adjusted and segments removed, but the tail segments are still included in the segment count that gets passed in the iov_iter. As the segments aren't necessarily all the same size, move importing to a separate helper and iterate the mapped length to get an exact count. Reviewed-by: Nitesh Shetty Signed-off-by: Jens Axboe --- io_uring/rsrc.c | 27 ++++++++++++++++++++++----- 1 file changed, 22 insertions(+), 5 deletions(-) diff --git a/io_uring/rsrc.c b/io_uring/rsrc.c index 0c6d7e7415c8..f80a77c4973f 100644 --- a/io_uring/rsrc.c +++ b/io_uring/rsrc.c @@ -1032,6 +1032,26 @@ static int validate_fixed_range(u64 buf_addr, size_t len, return 0; } +static int io_import_kbuf(int ddir, struct iov_iter *iter, + struct io_mapped_ubuf *imu, size_t len, size_t offset) +{ + size_t count = len + offset; + + iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, count); + iov_iter_advance(iter, offset); + + if (count < imu->len) { + const struct bio_vec *bvec = iter->bvec; + + while (len > bvec->bv_len) { + len -= bvec->bv_len; + bvec++; + } + iter->nr_segs = 1 + bvec - iter->bvec; + } + return 0; +} + static int io_import_fixed(int ddir, struct iov_iter *iter, struct io_mapped_ubuf *imu, u64 buf_addr, size_t len) @@ -1052,11 +1072,8 @@ static int io_import_fixed(int ddir, struct iov_iter *iter, offset = buf_addr - imu->ubuf; - if (imu->is_kbuf) { - iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); - iov_iter_advance(iter, offset); - return 0; - } + if (imu->is_kbuf) + return io_import_kbuf(ddir, iter, imu, len, offset); /* * Don't use iov_iter_advance() here, as it's really slow for From f12ecf5e1c5eca48b8652e893afcdb730384a6aa Mon Sep 17 00:00:00 2001 From: Pavel Begunkov Date: Fri, 18 Apr 2025 13:02:27 +0100 Subject: [PATCH 8/8] io_uring/zcrx: fix late dma unmap for a dead dev There is a problem with page pools not dma-unmapping immediately when the device is going down, and delaying it until the page pool is destroyed, which is not allowed (see links). That just got fixed for normal page pools, and we need to address memory providers as well. Unmap pages in the memory provider uninstall callback, and protect it with a new lock. There is also a gap between when a dma mapping is created and the mp is installed, so if the device is killed in between, io_uring would be holding on to dma mappings to a dead device with no one to call ->uninstall. Move it to page pool init and rely on ->is_mapped to make sure it's only done once. Link: https://lore.kernel.org/lkml/8067f204-1380-4d37-8ffd-007fc6f26738@kernel.org/T/ Link: https://lore.kernel.org/all/20250409-page-pool-track-dma-v9-0-6a9ef2e0cba8@redhat.com/ Fixes: 34a3e60821ab9 ("io_uring/zcrx: implement zerocopy receive pp memory provider") Signed-off-by: Pavel Begunkov Link: https://lore.kernel.org/r/ef9b7db249b14f6e0b570a1bb77ff177389f881c.1744965853.git.asml.silence@gmail.com Signed-off-by: Jens Axboe --- io_uring/zcrx.c | 21 +++++++++++++++++---- io_uring/zcrx.h | 1 + 2 files changed, 18 insertions(+), 4 deletions(-) diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c index 5defbe8f95f9..fe86606b9f30 100644 --- a/io_uring/zcrx.c +++ b/io_uring/zcrx.c @@ -51,14 +51,21 @@ static void __io_zcrx_unmap_area(struct io_zcrx_ifq *ifq, static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area) { + guard(mutex)(&ifq->dma_lock); + if (area->is_mapped) __io_zcrx_unmap_area(ifq, area, area->nia.num_niovs); + area->is_mapped = false; } static int io_zcrx_map_area(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area) { int i; + guard(mutex)(&ifq->dma_lock); + if (area->is_mapped) + return 0; + for (i = 0; i < area->nia.num_niovs; i++) { struct net_iov *niov = &area->nia.niovs[i]; dma_addr_t dma; @@ -280,6 +287,7 @@ static struct io_zcrx_ifq *io_zcrx_ifq_alloc(struct io_ring_ctx *ctx) ifq->ctx = ctx; spin_lock_init(&ifq->lock); spin_lock_init(&ifq->rq_lock); + mutex_init(&ifq->dma_lock); return ifq; } @@ -329,6 +337,7 @@ static void io_zcrx_ifq_free(struct io_zcrx_ifq *ifq) put_device(ifq->dev); io_free_rbuf_ring(ifq); + mutex_destroy(&ifq->dma_lock); kfree(ifq); } @@ -400,10 +409,6 @@ int io_register_zcrx_ifq(struct io_ring_ctx *ctx, goto err; get_device(ifq->dev); - ret = io_zcrx_map_area(ifq, ifq->area); - if (ret) - goto err; - mp_param.mp_ops = &io_uring_pp_zc_ops; mp_param.mp_priv = ifq; ret = net_mp_open_rxq(ifq->netdev, reg.if_rxq, &mp_param); @@ -624,6 +629,7 @@ static bool io_pp_zc_release_netmem(struct page_pool *pp, netmem_ref netmem) static int io_pp_zc_init(struct page_pool *pp) { struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); + int ret; if (WARN_ON_ONCE(!ifq)) return -EINVAL; @@ -636,6 +642,10 @@ static int io_pp_zc_init(struct page_pool *pp) if (pp->p.dma_dir != DMA_FROM_DEVICE) return -EOPNOTSUPP; + ret = io_zcrx_map_area(ifq, ifq->area); + if (ret) + return ret; + percpu_ref_get(&ifq->ctx->refs); return 0; } @@ -671,6 +681,9 @@ static void io_pp_uninstall(void *mp_priv, struct netdev_rx_queue *rxq) struct io_zcrx_ifq *ifq = mp_priv; io_zcrx_drop_netdev(ifq); + if (ifq->area) + io_zcrx_unmap_area(ifq, ifq->area); + p->mp_ops = NULL; p->mp_priv = NULL; } diff --git a/io_uring/zcrx.h b/io_uring/zcrx.h index 47f1c0e8c197..f2bc811f022c 100644 --- a/io_uring/zcrx.h +++ b/io_uring/zcrx.h @@ -38,6 +38,7 @@ struct io_zcrx_ifq { struct net_device *netdev; netdevice_tracker netdev_tracker; spinlock_t lock; + struct mutex dma_lock; }; #if defined(CONFIG_IO_URING_ZCRX)