Support for mapping SQ/CQ rings into huge page
Jens Axboe
maintainerv2 cover Support for mapping SQ/CQ rings into huge page
Hi,
io_uring SQ/CQ rings are allocated by the kernel from contigious, normal pages, and then the application mmap()'s the rings into userspace. This works fine, but does require contigious pages to be available for the given SQ and CQ ring sizes. As uptime increases on a given system, so does memory fragmentation. Entropy is invevitable.
This patchset adds support for the application passing in a pre-allocated huge page, and then placing the rings in that. This reduces the need for contigious pages, and also reduces the TLB pressure for larger rings.
The liburing huge.2 branch has support for using this trivially. Applications may use the normal ring init helpers and set IORING_SETUP_NO_MMAP, in which case a huge page will get allocated for them and used. Or they may use io_uring_queue_init_mem() and pass in a pre-allocated huge page, getting the amount of it used returned. This allows placing multiple rings into a single huge page.
Changes since v1:
- Mandate that we're using a single page. May be a normal page if we don't need a lot of memory, or a huge page if the ring itself takes up more space than a single normal page.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH v2 1/4 io_uring: remove sq/cq_off memset
We only have two reserved members we're not clearing, do so manually instead. This is in preparation for using one of these members for a new feature.
--- io_uring/io_uring.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
io_uring/io_uring.c +4 -2
@@ -3887,7 +3887,6 @@ static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,if (ret) goto err; -memset (&p->sq_off,0 ,sizeof (p->sq_off)); p->sq_off.head = offsetof(struct io_rings, sq.head); p->sq_off.tail = offsetof(struct io_rings, sq.tail); p->sq_off.ring_mask = offsetof(struct io_rings , sq_ring_mask);
@@ -3895,8 +3894,9 @@ static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, p->sq_off.flags = offsetof(struct io_rings , sq_flags); p->sq_off.dropped = offsetof(struct io_rings , sq_dropped); p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; + p->sq_off.resv1 =0 ; + p->sq_off.resv2 =0 ; -memset (&p->cq_off,0 ,sizeof (p->cq_off)); p->cq_off.head = offsetof(struct io_rings, cq.head); p->cq_off.tail = offsetof(struct io_rings, cq.tail); p->cq_off.ring_mask = offsetof(struct io_rings , cq_ring_mask);
@@ -3904,6 +3904,8 @@ static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, p->cq_off.overflow = offsetof(struct io_rings , cq_overflow); p->cq_off.cqes = offsetof(struct io_rings , cqes); p->cq_off.flags = offsetof(struct io_rings , cq_flags); + p->cq_off.resv1 =0 ; + p->cq_off.resv2 =0 ; p->features =IORING_FEAT_SINGLE_MMAP |IORING_FEAT_NODROP | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
Signature · 1 line
2.39.2
Jens Axboe
maintainerPATCH v2 2/4 io_uring: return error pointer from io_mem_alloc()
In reply to Jens AxboeIn preparation for having more than one time of ring allocator, make the existing one return valid/error-pointer rather than just NULL.
--- io_uring/io_uring.c | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-)
io_uring/io_uring.c +12 -6
@@ -2712,8 +2712,12 @@ static void io_mem_free(void *ptr)static void *io_mem_alloc (size_t size) {gfp_t gfp =GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; +void *ret; -return (void *)__get_free_pages (gfp,get_order (size)); + ret = (void *)__get_free_pages (gfp,get_order (size)); +if (ret) +return ret; +return ERR_PTR (-ENOMEM ); } static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
@@ -3673,6 +3677,7 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, {struct io_rings *rings;size_t size, sq_array_offset; +void *ptr;/* make sure these are sane, as we already accounted them */ ctx->sq_entries = p->sq_entries;
@@ -3683,8 +3688,8 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,return -EOVERFLOW ; rings =io_mem_alloc (size); -if (!rings) -return -ENOMEM ; +if (IS_ERR (rings)) +return PTR_ERR (rings); ctx->rings = rings; ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
@@ -3703,13 +3708,14 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,return -EOVERFLOW ; } - ctx->sq_sqes =io_mem_alloc (size); -if (!ctx->sq_sqes) { + ptr =io_mem_alloc (size); +if (IS_ERR (ptr)) {io_mem_free (ctx->rings); ctx->rings =NULL ; -return -ENOMEM ; +return PTR_ERR (ptr); } + ctx->sq_sqes =io_mem_alloc (size);return 0 ; }
Signature · 1 line
2.39.2
Re: [PATCH 2/4] io_uring: return error pointer from io_mem_alloc()
Hi Jens,
Quoted 51 lines · Jens Axboe
In preparation for having more than one time of ring allocator, make the existing one return valid/error-pointer rather than just NULL.
Signed-off-by Jens Axboe <[email protected]>--- io_uring/io_uring.c | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-)io_uring/io_uring.c +8
-4@@ -2712,8 +2712,12 @@ static void io_mem_free(void *ptr)static void *io_mem_alloc (size_t size) {gfp_t gfp =GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; +void *ret; -return (void *)__get_free_pages (gfp,get_order (size)); + ret = (void *)__get_free_pages (gfp,get_order (size)); +if (ret) +return ret; +return ERR_PTR (-ENOMEM ); } static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,@@ -3673,6 +3677,7 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, {struct io_rings *rings;size_t size, sq_array_offset; +void *ptr;/* make sure these are sane, as we already accounted them */ ctx->sq_entries = p->sq_entries;@@ -3683,8 +3688,8 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,return -EOVERFLOW ; rings =io_mem_alloc (size); -if (!rings) -return -ENOMEM ; +if (IS_ERR (rings)) +return PTR_ERR (rings); ctx->rings = rings; ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);@@ -3703,13 +3708,14 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,return -EOVERFLOW ; } - ctx->sq_sqes =io_mem_alloc (size);
io_uring/io_uring.c static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, · line 3718
Quoted 10 lines
io_uring/io_uring.c +4
-2-if (!ctx->sq_sqes) { + ptr =io_mem_alloc (size); +if (IS_ERR (ptr)) {io_mem_free (ctx->rings); ctx->rings =NULL ; -return -ENOMEM ; +return PTR_ERR (ptr); } + ctx->sq_sqes =io_mem_alloc (size);
Should be 'ptr' rather than 'io_mem_alloc(size)' here.
Signature · 1 line
Dmitry Kadashev
Jens Axboe
maintainerRe: [PATCH 2/4] io_uring: return error pointer from io_mem_alloc()
Quoted 63 lines · Dmitry Kadashev
Hi Jens,
Quoted 60 lines · On Sat, May 13, 2023 at 9:19?PM Jens Axboe <[email protected]> wrote:
In preparation for having more than one time of ring allocator, make the existing one return valid/error-pointer rather than just NULL.
Signed-off-by Jens Axboe <[email protected]>--- io_uring/io_uring.c | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-)io_uring/io_uring.c +11
-6@@ -2712,8 +2712,12 @@ static void io_mem_free(void *ptr)static void *io_mem_alloc (size_t size) {gfp_t gfp =GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; +void *ret; -return (void *)__get_free_pages (gfp,get_order (size)); + ret = (void *)__get_free_pages (gfp,get_order (size)); +if (ret) +return ret; +return ERR_PTR (-ENOMEM ); } static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,@@ -3673,6 +3677,7 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, {struct io_rings *rings;size_t size, sq_array_offset; +void *ptr;/* make sure these are sane, as we already accounted them */ ctx->sq_entries = p->sq_entries;@@ -3683,8 +3688,8 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,return -EOVERFLOW ; rings =io_mem_alloc (size); -if (!rings) -return -ENOMEM ; +if (IS_ERR (rings)) +return PTR_ERR (rings); ctx->rings = rings; ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);@@ -3703,13 +3708,14 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,return -EOVERFLOW ; } - ctx->sq_sqes =io_mem_alloc (size); -if (!ctx->sq_sqes) { + ptr =io_mem_alloc (size); +if (IS_ERR (ptr)) {io_mem_free (ctx->rings); ctx->rings =NULL ; -return -ENOMEM ; +return PTR_ERR (ptr); }
Quoted 3 lines
io_uring/io_uring.c static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, · line 3718
Quoted 1 line
io_uring/io_uring.c +1
-0+ ctx->sq_sqes =io_mem_alloc (size);Should be 'ptr' rather than '
io_mem_alloc(size)' here.
Indeed, good catch. Patch 4 does correct that so the final result is correct, must've happened during a split rebase a while back. I'll fix up patch 2 and 4 so that it's correct after patch 2 as well.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH v2 3/4 io_uring: add ring freeing helper
In reply to Jens AxboeWe do rings and sqes separately, move them into a helper that does both the freeing and clearing of the memory.
--- io_uring/io_uring.c | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-)
io_uring/io_uring.c +11 -6
@@ -2709,6 +2709,14 @@ static void io_mem_free(void *ptr)free_compound_page (page); } +static void io_rings_free (struct io_ring_ctx *ctx) +{ +io_mem_free (ctx->rings); +io_mem_free (ctx->sq_sqes); + ctx->rings =NULL ; + ctx->sq_sqes =NULL ; +} +static void *io_mem_alloc (size_t size) {gfp_t gfp =GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
@@ -2873,8 +2881,7 @@ static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)mmdrop (ctx->mm_account); ctx->mm_account =NULL ; } -io_mem_free (ctx->rings); -io_mem_free (ctx->sq_sqes); +io_rings_free (ctx);percpu_ref_exit (&ctx->refs);free_uid (ctx->user);
@@ -3703,15 +3710,13 @@ static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,else size =array_size (sizeof (struct io_uring_sqe ), p->sq_entries);if (size ==SIZE_MAX ) { -io_mem_free (ctx->rings); - ctx->rings =NULL ; +io_rings_free (ctx);return -EOVERFLOW ; } ptr =io_mem_alloc (size);if (IS_ERR (ptr)) { -io_mem_free (ctx->rings); - ctx->rings =NULL ; +io_rings_free (ctx);return PTR_ERR (ptr); }
Signature · 1 line
2.39.2
Jens Axboe
maintainerPATCH v2 4/4 io_uring: support for user allocated memory for rings/sqes
In reply to Jens AxboeCurrently io_uring applications must call mmap(2) twice to map the rings themselves, and the sqes array. This works fine, but it does not support using huge pages to back the rings/sqes.
Provide a way for the application to pass in pre-allocated memory for the rings/sqes, which can then suitably be allocated from shmfs or via mmap to get huge page support.
Particularly for larger rings, this reduces the TLBs needed.
If an application wishes to take advantage of that, it must pre-allocate the memory needed for the sq/cq ring, and the sqes. The former must be passed in via the io_uring_params->cq_off.user_data field, while the latter is passed in via the io_uring_params->sq_off.user_data field. Then it must set IORING_SETUP_NO_MMAP in the io_uring_params->flags field, and io_uring will then map the existing memory into the kernel for shared use. The application must not call mmap(2) to map rings as it otherwise would have, that will now fail with -EINVAL if this setup flag was used.
The pages used for the rings and sqes must be contigious. The intent here is clearly that huge pages should be used, otherwise the normal setup procedure works fine as-is. The application may use one huge page for both the rings and sqes.
Outside of those initialization changes, everything works like it did before.
--- include/linux/io_uring_types.h | 10 +++ include/uapi/linux/io_uring.h | 9 ++- io_uring/io_uring.c | 108 ++++++++++++++++++++++++++++++--- 3 files changed, 115 insertions(+), 12 deletions(-)
include/linux/io_uring_types.h +10 -0
@@ -211,6 +211,16 @@ struct io_ring_ctx { unsigned int compat: 1;enum task_work_notify_mode notify_method; + +/* +* If IORING_SETUP_NO_MMAP is used, then the below holds +* the gup'ed pages for the two rings, and the sqes. +*/ +unsigned short n_ring_pages; +unsigned short n_sqe_pages; +struct page **ring_pages; +struct page **sqe_pages; +struct io_rings *rings;struct task_struct *submitter_task;struct percpu_ref refs;
include/uapi/linux/io_uring.h +7 -2
@@ -173,6 +173,11 @@ enum {*/ #define IORING_SETUP_DEFER_TASKRUN (1U << 13) +/* +* Application provides the memory for the rings +*/ +#define IORING_SETUP_NO_MMAP (1U << 14) +enum io_uring_op {IORING_OP_NOP ,IORING_OP_READV ,
@@ -406,7 +411,7 @@ struct io_sqring_offsets {__u32 dropped;__u32 array;__u32 resv1; -__u64 resv2; +__u64 user_addr; };/*
@@ -425,7 +430,7 @@ struct io_cqring_offsets {__u32 cqes;__u32 flags;__u32 resv1; -__u64 resv2; +__u64 user_addr; };/*
io_uring/io_uring.c +98 -10
157 lines not on this page · raw message
Signature · 1 line
2.39.2