From: Leon Romanovsky leonro@nvidia.com
Changelog: v5: * Use sg_init_table to allocate table and avoid changes is __sg_alloc_table * Fix offset issue v4: https://lore.kernel.org/lkml/20200927064647.3106737-1-leon@kernel.org * Fixed formatting in first patch. * Added fix (clear tmp_netnts) in first patch to fix i915 failure. * Added test patches v3: https://lore.kernel.org/linux-rdma/20200922083958.2150803-1-leon@kernel.org/ * Squashed Christopher's suggestion to avoid introduced new API, but extend existing one. v2: https://lore.kernel.org/linux-rdma/20200916140726.839377-1-leon@kernel.org * Fixed indentations and comments * Deleted sg_alloc_next() * Squashed lib/scatterlist patches into one v1: https://lore.kernel.org/lkml/20200910134259.1304543-1-leon@kernel.org * Changed _sg_chain to be __sg_chain * Added dependency on ARCH_NO_SG_CHAIN * Removed struct sg_append v0: * https://lore.kernel.org/lkml/20200903121853.1145976-1-leon@kernel.org
--------------------------------------------------------------------------
From Maor:
This series extends __sg_alloc_table_from_pages to allow chaining of new pages to already initialized SG table.
This allows for the drivers to utilize the optimization of merging contiguous pages without a need to pre allocate all the pages and hold them in a very large temporary buffer prior to the call to SG table initialization.
The second patch changes the Infiniband driver to use the new API. It removes duplicate functionality from the code and benefits the optimization of allocating dynamic SG table from pages.
In huge pages system of 2MB page size, without this change, the SG table would contain x512 SG entries. E.g. for 100GB memory registration:
Number of entries Size Before 26214400 600.0MB After 51200 1.2MB
Thanks
Maor Gottlieb (2): lib/scatterlist: Add support in dynamic allocation of SG table from pages RDMA/umem: Move to allocate SG table from pages
Tvrtko Ursulin (2): tools/testing/scatterlist: Rejuvenate bit-rotten test tools/testing/scatterlist: Show errors in human readable form
drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 12 +- drivers/gpu/drm/vmwgfx/vmwgfx_ttm_buffer.c | 15 ++- drivers/infiniband/core/umem.c | 94 ++------------- include/linux/scatterlist.h | 38 +++--- lib/scatterlist.c | 125 ++++++++++++++++---- tools/testing/scatterlist/Makefile | 3 +- tools/testing/scatterlist/linux/mm.h | 35 ++++++ tools/testing/scatterlist/main.c | 53 ++++++--- 8 files changed, 225 insertions(+), 150 deletions(-)
-- 2.26.2
From: Maor Gottlieb maorg@nvidia.com
Extend __sg_alloc_table_from_pages to support dynamic allocation of SG table from pages. It should be used by drivers that can't supply all the pages at one time.
This function returns the last populated SGE in the table. Users should pass it as an argument to the function from the second call and forward. As before, nents will be equal to the number of populated SGEs (chunks).
With this new extension, drivers can benefit the optimization of merging contiguous pages without a need to allocate all pages in advance and hold them in a large buffer.
E.g. with the Infiniband driver that allocates a single page for hold the pages. For 1TB memory registration, the temporary buffer would consume only 4KB, instead of 2GB.
Signed-off-by: Maor Gottlieb maorg@nvidia.com Reviewed-by: Christoph Hellwig hch@lst.de Signed-off-by: Leon Romanovsky leonro@nvidia.com --- drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 12 +- drivers/gpu/drm/vmwgfx/vmwgfx_ttm_buffer.c | 15 ++- include/linux/scatterlist.h | 38 +++--- lib/scatterlist.c | 125 ++++++++++++++++---- tools/testing/scatterlist/main.c | 9 +- 5 files changed, 142 insertions(+), 57 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c index 12b30075134a..f2eaed6aca3d 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c @@ -403,6 +403,7 @@ __i915_gem_userptr_alloc_pages(struct drm_i915_gem_object *obj, unsigned int max_segment = i915_sg_segment_size(); struct sg_table *st; unsigned int sg_page_sizes; + struct scatterlist *sg; int ret;
st = kmalloc(sizeof(*st), GFP_KERNEL); @@ -410,13 +411,12 @@ __i915_gem_userptr_alloc_pages(struct drm_i915_gem_object *obj, return ERR_PTR(-ENOMEM);
alloc_table: - ret = __sg_alloc_table_from_pages(st, pvec, num_pages, - 0, num_pages << PAGE_SHIFT, - max_segment, - GFP_KERNEL); - if (ret) { + sg = __sg_alloc_table_from_pages(st, pvec, num_pages, 0, + num_pages << PAGE_SHIFT, max_segment, + NULL, 0, GFP_KERNEL); + if (IS_ERR(sg)) { kfree(st); - return ERR_PTR(ret); + return ERR_CAST(sg); }
ret = i915_gem_gtt_prepare_pages(obj, st); diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_buffer.c b/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_buffer.c index ab524ab3b0b4..f22acd398b1f 100644 --- a/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_buffer.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_buffer.c @@ -419,6 +419,7 @@ static int vmw_ttm_map_dma(struct vmw_ttm_tt *vmw_tt) int ret = 0; static size_t sgl_size; static size_t sgt_size; + struct scatterlist *sg;
if (vmw_tt->mapped) return 0; @@ -441,13 +442,15 @@ static int vmw_ttm_map_dma(struct vmw_ttm_tt *vmw_tt) if (unlikely(ret != 0)) return ret;
- ret = __sg_alloc_table_from_pages - (&vmw_tt->sgt, vsgt->pages, vsgt->num_pages, 0, - (unsigned long) vsgt->num_pages << PAGE_SHIFT, - dma_get_max_seg_size(dev_priv->dev->dev), - GFP_KERNEL); - if (unlikely(ret != 0)) + sg = __sg_alloc_table_from_pages(&vmw_tt->sgt, vsgt->pages, + vsgt->num_pages, 0, + (unsigned long) vsgt->num_pages << PAGE_SHIFT, + dma_get_max_seg_size(dev_priv->dev->dev), + NULL, 0, GFP_KERNEL); + if (IS_ERR(sg)) { + ret = PTR_ERR(sg); goto out_sg_alloc_fail; + }
if (vsgt->num_pages > vmw_tt->sgt.nents) { uint64_t over_alloc = diff --git a/include/linux/scatterlist.h b/include/linux/scatterlist.h index 45cf7b69d852..36c47e7e66a2 100644 --- a/include/linux/scatterlist.h +++ b/include/linux/scatterlist.h @@ -165,6 +165,22 @@ static inline void sg_set_buf(struct scatterlist *sg, const void *buf, #define for_each_sgtable_dma_sg(sgt, sg, i) \ for_each_sg((sgt)->sgl, sg, (sgt)->nents, i)
+static inline void __sg_chain(struct scatterlist *chain_sg, + struct scatterlist *sgl) +{ + /* + * offset and length are unused for chain entry. Clear them. + */ + chain_sg->offset = 0; + chain_sg->length = 0; + + /* + * Set lowest bit to indicate a link pointer, and make sure to clear + * the termination bit if it happens to be set. + */ + chain_sg->page_link = ((unsigned long) sgl | SG_CHAIN) & ~SG_END; +} + /** * sg_chain - Chain two sglists together * @prv: First scatterlist @@ -178,18 +194,7 @@ static inline void sg_set_buf(struct scatterlist *sg, const void *buf, static inline void sg_chain(struct scatterlist *prv, unsigned int prv_nents, struct scatterlist *sgl) { - /* - * offset and length are unused for chain entry. Clear them. - */ - prv[prv_nents - 1].offset = 0; - prv[prv_nents - 1].length = 0; - - /* - * Set lowest bit to indicate a link pointer, and make sure to clear - * the termination bit if it happens to be set. - */ - prv[prv_nents - 1].page_link = ((unsigned long) sgl | SG_CHAIN) - & ~SG_END; + __sg_chain(&prv[prv_nents - 1], sgl); }
/** @@ -286,10 +291,11 @@ void sg_free_table(struct sg_table *); int __sg_alloc_table(struct sg_table *, unsigned int, unsigned int, struct scatterlist *, unsigned int, gfp_t, sg_alloc_fn *); int sg_alloc_table(struct sg_table *, unsigned int, gfp_t); -int __sg_alloc_table_from_pages(struct sg_table *sgt, struct page **pages, - unsigned int n_pages, unsigned int offset, - unsigned long size, unsigned int max_segment, - gfp_t gfp_mask); +struct scatterlist *__sg_alloc_table_from_pages(struct sg_table *sgt, + struct page **pages, unsigned int n_pages, unsigned int offset, + unsigned long size, unsigned int max_segment, + struct scatterlist *prv, unsigned int left_pages, + gfp_t gfp_mask); int sg_alloc_table_from_pages(struct sg_table *sgt, struct page **pages, unsigned int n_pages, unsigned int offset, unsigned long size, gfp_t gfp_mask); diff --git a/lib/scatterlist.c b/lib/scatterlist.c index 5d63a8857f36..e102fdfaa75b 100644 --- a/lib/scatterlist.c +++ b/lib/scatterlist.c @@ -365,6 +365,37 @@ int sg_alloc_table(struct sg_table *table, unsigned int nents, gfp_t gfp_mask) } EXPORT_SYMBOL(sg_alloc_table);
+static struct scatterlist *get_next_sg(struct sg_table *table, + struct scatterlist *cur, + unsigned long needed_sges, + gfp_t gfp_mask) +{ + struct scatterlist *new_sg, *next_sg; + unsigned int alloc_size; + + if (cur) { + next_sg = sg_next(cur); + /* Check if last entry should be keeped for chainning */ + if (!sg_is_last(next_sg) || needed_sges == 1) + return next_sg; + } + + alloc_size = min_t(unsigned long, needed_sges, SG_MAX_SINGLE_ALLOC); + new_sg = sg_kmalloc(alloc_size, gfp_mask); + if (!new_sg) + return ERR_PTR(-ENOMEM); + sg_init_table(new_sg, alloc_size); + if (cur) { + __sg_chain(next_sg, new_sg); + table->orig_nents += alloc_size - 1; + } else { + table->sgl = new_sg; + table->orig_nents = alloc_size; + table->nents = 0; + } + return new_sg; +} + /** * __sg_alloc_table_from_pages - Allocate and initialize an sg table from * an array of pages @@ -374,29 +405,63 @@ EXPORT_SYMBOL(sg_alloc_table); * @offset: Offset from start of the first page to the start of a buffer * @size: Number of valid bytes in the buffer (after offset) * @max_segment: Maximum size of a scatterlist node in bytes (page aligned) + * @prv: Last populated sge in sgt + * @left_pages: Left pages caller have to set after this call * @gfp_mask: GFP allocation mask * - * Description: - * Allocate and initialize an sg table from a list of pages. Contiguous - * ranges of the pages are squashed into a single scatterlist node up to the - * maximum size specified in @max_segment. An user may provide an offset at a - * start and a size of valid data in a buffer specified by the page array. - * The returned sg table is released by sg_free_table. + * Description: + * If @prv is NULL, allocate and initialize an sg table from a list of pages, + * else reuse the scatterlist passed in at @prv. + * Contiguous ranges of the pages are squashed into a single scatterlist + * entry up to the maximum size specified in @max_segment. A user may + * provide an offset at a start and a size of valid data in a buffer + * specified by the page array. * * Returns: - * 0 on success, negative error on failure + * Last SGE in sgt on success, PTR_ERR on otherwise. + * The allocation in @sgt must be released by sg_free_table. + * + * Notes: + * If this function returns non-0 (eg failure), the caller must call + * sg_free_table() to cleanup any leftover allocations. */ -int __sg_alloc_table_from_pages(struct sg_table *sgt, struct page **pages, - unsigned int n_pages, unsigned int offset, - unsigned long size, unsigned int max_segment, - gfp_t gfp_mask) +struct scatterlist *__sg_alloc_table_from_pages(struct sg_table *sgt, + struct page **pages, unsigned int n_pages, unsigned int offset, + unsigned long size, unsigned int max_segment, + struct scatterlist *prv, unsigned int left_pages, + gfp_t gfp_mask) { - unsigned int chunks, cur_page, seg_len, i; - int ret; - struct scatterlist *s; + unsigned int chunks, cur_page, seg_len, i, prv_len = 0; + unsigned int added_nents = 0; + struct scatterlist *s = prv;
if (WARN_ON(!max_segment || offset_in_page(max_segment))) - return -EINVAL; + return ERR_PTR(-EINVAL); + + if (IS_ENABLED(CONFIG_ARCH_NO_SG_CHAIN) && prv) + return ERR_PTR(-EOPNOTSUPP); + + if (prv) { + unsigned long paddr = (page_to_pfn(sg_page(prv)) * PAGE_SIZE + + prv->offset + prv->length) / + PAGE_SIZE; + + if (WARN_ON(offset)) + return ERR_PTR(-EINVAL); + + /* Merge contiguous pages into the last SG */ + prv_len = prv->length; + while (n_pages && page_to_pfn(pages[0]) == paddr) { + if (prv->length + PAGE_SIZE > max_segment) + break; + prv->length += PAGE_SIZE; + paddr++; + pages++; + n_pages--; + } + if (!n_pages) + goto out; + }
/* compute number of contiguous chunks */ chunks = 1; @@ -410,13 +475,9 @@ int __sg_alloc_table_from_pages(struct sg_table *sgt, struct page **pages, } }
- ret = sg_alloc_table(sgt, chunks, gfp_mask); - if (unlikely(ret)) - return ret; - /* merging chunks and putting them into the scatterlist */ cur_page = 0; - for_each_sg(sgt->sgl, s, sgt->orig_nents, i) { + for (i = 0; i < chunks; i++) { unsigned int j, chunk_size;
/* look for the end of the current chunk */ @@ -429,15 +490,30 @@ int __sg_alloc_table_from_pages(struct sg_table *sgt, struct page **pages, break; }
+ /* Pass how many chunks might be left */ + s = get_next_sg(sgt, s, chunks - i + left_pages, gfp_mask); + if (IS_ERR(s)) { + /* + * Adjust entry length to be as before function was + * called. + */ + if (prv) + prv->length = prv_len; + return s; + } chunk_size = ((j - cur_page) << PAGE_SHIFT) - offset; sg_set_page(s, pages[cur_page], min_t(unsigned long, size, chunk_size), offset); + added_nents++; size -= chunk_size; offset = 0; cur_page = j; } - - return 0; + sgt->nents += added_nents; +out: + if (!left_pages) + sg_mark_end(s); + return s; } EXPORT_SYMBOL(__sg_alloc_table_from_pages);
@@ -465,8 +541,9 @@ int sg_alloc_table_from_pages(struct sg_table *sgt, struct page **pages, unsigned int n_pages, unsigned int offset, unsigned long size, gfp_t gfp_mask) { - return __sg_alloc_table_from_pages(sgt, pages, n_pages, offset, size, - SCATTERLIST_MAX_SEGMENT, gfp_mask); + return PTR_ERR_OR_ZERO(__sg_alloc_table_from_pages(sgt, pages, n_pages, + offset, size, SCATTERLIST_MAX_SEGMENT, + NULL, 0, gfp_mask)); } EXPORT_SYMBOL(sg_alloc_table_from_pages);
diff --git a/tools/testing/scatterlist/main.c b/tools/testing/scatterlist/main.c index 0a1464181226..4899359a31ac 100644 --- a/tools/testing/scatterlist/main.c +++ b/tools/testing/scatterlist/main.c @@ -55,14 +55,13 @@ int main(void) for (i = 0, test = tests; test->expected_segments; test++, i++) { struct page *pages[MAX_PAGES]; struct sg_table st; - int ret; + struct scatterlist *sg;
set_pages(pages, test->pfn, test->num_pages);
- ret = __sg_alloc_table_from_pages(&st, pages, test->num_pages, - 0, test->size, test->max_seg, - GFP_KERNEL); - assert(ret == test->alloc_ret); + sg = __sg_alloc_table_from_pages(&st, pages, test->num_pages, 0, + test->size, test->max_seg, NULL, 0, GFP_KERNEL); + assert(PTR_ERR_OR_ZERO(sg) == test->alloc_ret);
if (test->alloc_ret) continue; -- 2.26.2
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
A couple small tweaks are needed to make the test build and run on current kernels.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Maor Gottlieb maorg@nvidia.com Signed-off-by: Leon Romanovsky leonro@nvidia.com --- tools/testing/scatterlist/Makefile | 3 ++- tools/testing/scatterlist/linux/mm.h | 35 ++++++++++++++++++++++++++++ 2 files changed, 37 insertions(+), 1 deletion(-)
diff --git a/tools/testing/scatterlist/Makefile b/tools/testing/scatterlist/Makefile index cbb003d9305e..c65233876622 100644 --- a/tools/testing/scatterlist/Makefile +++ b/tools/testing/scatterlist/Makefile @@ -14,7 +14,7 @@ targets: include $(TARGETS) main: $(OFILES)
clean: - $(RM) $(TARGETS) $(OFILES) scatterlist.c linux/scatterlist.h linux/highmem.h linux/kmemleak.h asm/io.h + $(RM) $(TARGETS) $(OFILES) scatterlist.c linux/scatterlist.h linux/highmem.h linux/kmemleak.h linux/slab.h asm/io.h @rmdir asm
scatterlist.c: ../../../lib/scatterlist.c @@ -28,4 +28,5 @@ include: ../../../include/linux/scatterlist.h @touch asm/io.h @touch linux/highmem.h @touch linux/kmemleak.h + @touch linux/slab.h @cp $< linux/scatterlist.h diff --git a/tools/testing/scatterlist/linux/mm.h b/tools/testing/scatterlist/linux/mm.h index 6f9ac14aa800..6ae907f375d2 100644 --- a/tools/testing/scatterlist/linux/mm.h +++ b/tools/testing/scatterlist/linux/mm.h @@ -114,6 +114,12 @@ static inline void *kmalloc(unsigned int size, unsigned int flags) return malloc(size); }
+static inline void * +kmalloc_array(unsigned int n, unsigned int size, unsigned int flags) +{ + return malloc(n * size); +} + #define kfree(x) free(x)
#define kmemleak_alloc(a, b, c, d) @@ -122,4 +128,33 @@ static inline void *kmalloc(unsigned int size, unsigned int flags) #define PageSlab(p) (0) #define flush_kernel_dcache_page(p)
+#define MAX_ERRNO 4095 + +#define IS_ERR_VALUE(x) unlikely((unsigned long)(void *)(x) >= (unsigned long)-MAX_ERRNO) + +static inline void * __must_check ERR_PTR(long error) +{ + return (void *) error; +} + +static inline long __must_check PTR_ERR(__force const void *ptr) +{ + return (long) ptr; +} + +static inline bool __must_check IS_ERR(__force const void *ptr) +{ + return IS_ERR_VALUE((unsigned long)ptr); +} + +static inline int __must_check PTR_ERR_OR_ZERO(__force const void *ptr) +{ + if (IS_ERR(ptr)) + return PTR_ERR(ptr); + else + return 0; +} + +#define IS_ENABLED(x) (0) + #endif -- 2.26.2
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Instead of just asserting dump some more useful info about what the test saw versus what it expected to see.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Maor Gottlieb maorg@nvidia.com Signed-off-by: Leon Romanovsky leonro@nvidia.com --- tools/testing/scatterlist/main.c | 44 ++++++++++++++++++++++++-------- 1 file changed, 34 insertions(+), 10 deletions(-)
diff --git a/tools/testing/scatterlist/main.c b/tools/testing/scatterlist/main.c index 4899359a31ac..b2c7e9f7b8d3 100644 --- a/tools/testing/scatterlist/main.c +++ b/tools/testing/scatterlist/main.c @@ -5,6 +5,15 @@
#define MAX_PAGES (64)
+struct test { + int alloc_ret; + unsigned num_pages; + unsigned *pfn; + unsigned size; + unsigned int max_seg; + unsigned int expected_segments; +}; + static void set_pages(struct page **pages, const unsigned *array, unsigned num) { unsigned int i; @@ -17,17 +26,32 @@ static void set_pages(struct page **pages, const unsigned *array, unsigned num)
#define pfn(...) (unsigned []){ __VA_ARGS__ }
+static void fail(struct test *test, struct sg_table *st, const char *cond) +{ + unsigned int i; + + fprintf(stderr, "Failed on '%s'!\n\n", cond); + + printf("size = %u, max segment = %u, expected nents = %u\nst->nents = %u, st->orig_nents= %u\n", + test->size, test->max_seg, test->expected_segments, st->nents, + st->orig_nents); + + printf("%u input PFNs:", test->num_pages); + for (i = 0; i < test->num_pages; i++) + printf(" %x", test->pfn[i]); + printf("\n"); + + exit(1); +} + +#define VALIDATE(cond, st, test) \ + if (!(cond)) \ + fail((test), (st), #cond); + int main(void) { const unsigned int sgmax = SCATTERLIST_MAX_SEGMENT; - struct test { - int alloc_ret; - unsigned num_pages; - unsigned *pfn; - unsigned size; - unsigned int max_seg; - unsigned int expected_segments; - } *test, tests[] = { + struct test *test, tests[] = { { -EINVAL, 1, pfn(0), PAGE_SIZE, PAGE_SIZE + 1, 1 }, { -EINVAL, 1, pfn(0), PAGE_SIZE, 0, 1 }, { -EINVAL, 1, pfn(0), PAGE_SIZE, sgmax + 1, 1 }, @@ -66,8 +90,8 @@ int main(void) if (test->alloc_ret) continue;
- assert(st.nents == test->expected_segments); - assert(st.orig_nents == test->expected_segments); + VALIDATE(st.nents == test->expected_segments, &st, test); + VALIDATE(st.orig_nents == test->expected_segments, &st, test);
sg_free_table(&st); } -- 2.26.2
From: Maor Gottlieb maorg@nvidia.com
Remove the implementation of ib_umem_add_sg_table and instead call to __sg_alloc_table_from_pages which already has the logic to merge contiguous pages.
Besides that it removes duplicated functionality, it reduces the memory consumption of the SG table significantly. Prior to this patch, the SG table was allocated in advance regardless consideration of contiguous pages.
In huge pages system of 2MB page size, without this change, the SG table would contain x512 SG entries. E.g. for 100GB memory registration:
Number of entries Size Before 26214400 600.0MB After 51200 1.2MB
Signed-off-by: Maor Gottlieb maorg@nvidia.com Signed-off-by: Leon Romanovsky leonro@nvidia.com --- drivers/infiniband/core/umem.c | 94 +++++----------------------------- 1 file changed, 12 insertions(+), 82 deletions(-)
diff --git a/drivers/infiniband/core/umem.c b/drivers/infiniband/core/umem.c index c1ab6a4f2bc3..e9fecbdf391b 100644 --- a/drivers/infiniband/core/umem.c +++ b/drivers/infiniband/core/umem.c @@ -61,73 +61,6 @@ static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d sg_free_table(&umem->sg_head); }
-/* ib_umem_add_sg_table - Add N contiguous pages to scatter table - * - * sg: current scatterlist entry - * page_list: array of npage struct page pointers - * npages: number of pages in page_list - * max_seg_sz: maximum segment size in bytes - * nents: [out] number of entries in the scatterlist - * - * Return new end of scatterlist - */ -static struct scatterlist *ib_umem_add_sg_table(struct scatterlist *sg, - struct page **page_list, - unsigned long npages, - unsigned int max_seg_sz, - int *nents) -{ - unsigned long first_pfn; - unsigned long i = 0; - bool update_cur_sg = false; - bool first = !sg_page(sg); - - /* Check if new page_list is contiguous with end of previous page_list. - * sg->length here is a multiple of PAGE_SIZE and sg->offset is 0. - */ - if (!first && (page_to_pfn(sg_page(sg)) + (sg->length >> PAGE_SHIFT) == - page_to_pfn(page_list[0]))) - update_cur_sg = true; - - while (i != npages) { - unsigned long len; - struct page *first_page = page_list[i]; - - first_pfn = page_to_pfn(first_page); - - /* Compute the number of contiguous pages we have starting - * at i - */ - for (len = 0; i != npages && - first_pfn + len == page_to_pfn(page_list[i]) && - len < (max_seg_sz >> PAGE_SHIFT); - len++) - i++; - - /* Squash N contiguous pages from page_list into current sge */ - if (update_cur_sg) { - if ((max_seg_sz - sg->length) >= (len << PAGE_SHIFT)) { - sg_set_page(sg, sg_page(sg), - sg->length + (len << PAGE_SHIFT), - 0); - update_cur_sg = false; - continue; - } - update_cur_sg = false; - } - - /* Squash N contiguous pages into next sge or first sge */ - if (!first) - sg = sg_next(sg); - - (*nents)++; - sg_set_page(sg, first_page, len << PAGE_SHIFT, 0); - first = false; - } - - return sg; -} - /** * ib_umem_find_best_pgsz - Find best HW page size to use for this MR * @@ -217,7 +150,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr, struct mm_struct *mm; unsigned long npages; int ret; - struct scatterlist *sg; + struct scatterlist *sg = NULL; unsigned int gup_flags = FOLL_WRITE;
/* @@ -272,15 +205,9 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
cur_base = addr & PAGE_MASK;
- ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL); - if (ret) - goto vma; - if (!umem->writable) gup_flags |= FOLL_FORCE;
- sg = umem->sg_head.sgl; - while (npages) { cond_resched(); ret = pin_user_pages_fast(cur_base, @@ -292,15 +219,19 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr, goto umem_release;
cur_base += ret * PAGE_SIZE; - npages -= ret; - - sg = ib_umem_add_sg_table(sg, page_list, ret, - dma_get_max_seg_size(device->dma_device), - &umem->sg_nents); + npages -= ret; + sg = __sg_alloc_table_from_pages( + &umem->sg_head, page_list, ret, 0, ret << PAGE_SHIFT, + dma_get_max_seg_size(device->dma_device), sg, npages, + GFP_KERNEL); + umem->sg_nents = umem->sg_head.nents; + if (IS_ERR(sg)) { + unpin_user_pages_dirty_lock(page_list, ret, 0); + ret = PTR_ERR(sg); + goto umem_release; + } }
- sg_mark_end(sg); - if (access & IB_ACCESS_RELAXED_ORDERING) dma_attr |= DMA_ATTR_WEAK_ORDERING;
@@ -318,7 +249,6 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
umem_release: __ib_umem_release(device, umem, 0); -vma: atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm); out: free_page((unsigned long) page_list); -- 2.26.2
On Sun, Oct 04, 2020 at 06:43:36PM +0300, Leon Romanovsky wrote:
This series extends __sg_alloc_table_from_pages to allow chaining of new pages to already initialized SG table.
This allows for the drivers to utilize the optimization of merging contiguous pages without a need to pre allocate all the pages and hold them in a very large temporary buffer prior to the call to SG table initialization.
The second patch changes the Infiniband driver to use the new API. It removes duplicate functionality from the code and benefits the optimization of allocating dynamic SG table from pages.
In huge pages system of 2MB page size, without this change, the SG table would contain x512 SG entries. E.g. for 100GB memory registration:
Number of entries Size Before 26214400 600.0MB After 51200 1.2MBThanks
Maor Gottlieb (2): lib/scatterlist: Add support in dynamic allocation of SG table from pages RDMA/umem: Move to allocate SG table from pages
Tvrtko Ursulin (2): tools/testing/scatterlist: Rejuvenate bit-rotten test tools/testing/scatterlist: Show errors in human readable form
This looks OK, I'm going to send it into linux-next on the hmm tree for awhile to see if anything gets broken. If there is more remarks/tags/etc please continue
Thanks, Jason
On Mon, Oct 05, 2020 at 08:56:50PM -0300, Jason Gunthorpe wrote:
On Sun, Oct 04, 2020 at 06:43:36PM +0300, Leon Romanovsky wrote:
This series extends __sg_alloc_table_from_pages to allow chaining of new pages to already initialized SG table.
This allows for the drivers to utilize the optimization of merging contiguous pages without a need to pre allocate all the pages and hold them in a very large temporary buffer prior to the call to SG table initialization.
The second patch changes the Infiniband driver to use the new API. It removes duplicate functionality from the code and benefits the optimization of allocating dynamic SG table from pages.
In huge pages system of 2MB page size, without this change, the SG table would contain x512 SG entries. E.g. for 100GB memory registration:
Number of entries Size Before 26214400 600.0MB After 51200 1.2MBThanks
Maor Gottlieb (2): lib/scatterlist: Add support in dynamic allocation of SG table from pages RDMA/umem: Move to allocate SG table from pages
Tvrtko Ursulin (2): tools/testing/scatterlist: Rejuvenate bit-rotten test tools/testing/scatterlist: Show errors in human readable form
This looks OK, I'm going to send it into linux-next on the hmm tree for awhile to see if anything gets broken. If there is more remarks/tags/etc please continue
An idea that just crossed my mind: A pin_user_pages_sgt might be useful for both rdma and drm, since this would avoid the possible huge interim struct pages array for thp pages. Or anything else that could be coalesced down into a single sg entry.
Not sure it's worth it, but would at least give a slightly neater interface I think. -Daniel
On Tue, Oct 06, 2020 at 12:41:22PM +0200, Daniel Vetter wrote:
On Mon, Oct 05, 2020 at 08:56:50PM -0300, Jason Gunthorpe wrote:
On Sun, Oct 04, 2020 at 06:43:36PM +0300, Leon Romanovsky wrote:
This series extends __sg_alloc_table_from_pages to allow chaining of new pages to already initialized SG table.
This allows for the drivers to utilize the optimization of merging contiguous pages without a need to pre allocate all the pages and hold them in a very large temporary buffer prior to the call to SG table initialization.
The second patch changes the Infiniband driver to use the new API. It removes duplicate functionality from the code and benefits the optimization of allocating dynamic SG table from pages.
In huge pages system of 2MB page size, without this change, the SG table would contain x512 SG entries. E.g. for 100GB memory registration:
Number of entries Size Before 26214400 600.0MB After 51200 1.2MBThanks
Maor Gottlieb (2): lib/scatterlist: Add support in dynamic allocation of SG table from pages RDMA/umem: Move to allocate SG table from pages
Tvrtko Ursulin (2): tools/testing/scatterlist: Rejuvenate bit-rotten test tools/testing/scatterlist: Show errors in human readable form
This looks OK, I'm going to send it into linux-next on the hmm tree for awhile to see if anything gets broken. If there is more remarks/tags/etc please continue
An idea that just crossed my mind: A pin_user_pages_sgt might be useful for both rdma and drm, since this would avoid the possible huge interim struct pages array for thp pages. Or anything else that could be coalesced down into a single sg entry.
Not sure it's worth it, but would at least give a slightly neater interface I think.
We've talked about it. Christoph wants to see this area move to a biovec interface instead of sgl, but it might still be worthwhile to have an interm step at least as an API consolidation.
Avoiding the page list would be complicated as we'd somehow have to code share the page table iterator scheme.
Jason
On Wed, Oct 7, 2020 at 9:22 AM Jason Gunthorpe jgg@ziepe.ca wrote:
On Tue, Oct 06, 2020 at 12:41:22PM +0200, Daniel Vetter wrote:
On Mon, Oct 05, 2020 at 08:56:50PM -0300, Jason Gunthorpe wrote:
On Sun, Oct 04, 2020 at 06:43:36PM +0300, Leon Romanovsky wrote:
This series extends __sg_alloc_table_from_pages to allow chaining of new pages to already initialized SG table.
This allows for the drivers to utilize the optimization of merging contiguous pages without a need to pre allocate all the pages and hold them in a very large temporary buffer prior to the call to SG table initialization.
The second patch changes the Infiniband driver to use the new API. It removes duplicate functionality from the code and benefits the optimization of allocating dynamic SG table from pages.
In huge pages system of 2MB page size, without this change, the SG table would contain x512 SG entries. E.g. for 100GB memory registration:
Number of entries Size Before 26214400 600.0MB After 51200 1.2MBThanks
Maor Gottlieb (2): lib/scatterlist: Add support in dynamic allocation of SG table from pages RDMA/umem: Move to allocate SG table from pages
Tvrtko Ursulin (2): tools/testing/scatterlist: Rejuvenate bit-rotten test tools/testing/scatterlist: Show errors in human readable form
This looks OK, I'm going to send it into linux-next on the hmm tree for awhile to see if anything gets broken. If there is more remarks/tags/etc please continue
An idea that just crossed my mind: A pin_user_pages_sgt might be useful for both rdma and drm, since this would avoid the possible huge interim struct pages array for thp pages. Or anything else that could be coalesced down into a single sg entry.
Not sure it's worth it, but would at least give a slightly neater interface I think.
We've talked about it. Christoph wants to see this area move to a biovec interface instead of sgl, but it might still be worthwhile to have an interm step at least as an API consolidation.
Hm but then we'd need a new struct for the mapped side of things (which would still be what you get from dma-buf). That would be quite a bit of work to roll out everywhere, and sgt isn't such a huge misfit for passing buffer object mappings and system memory backing storage around, and hence what we (very slowly) converging drivers/gpu towards over the past 10 years or so.
And moving the dma_map step out of dma-buf doesn't work, because some of the use-cases we have is for very special iommus which are managed by the gpu driver directly. Stuff that e.g. rotates/retiles/compresses on the fly, and is accessible by other (gfx related like video code, camera, ..) devices. Not something I expect to ever be relevant for rdma since this exist mostly on some small soc, but it's a thing. Without that dma-buf could hand out biovec for struct_page backed stuff, or some pfn_vec for the p2p stuff.
Anyway was just an idea, I guess we'll have to live with some impedance mismatch since rolling out the one an only iovec structure which suits everyone is I think impossible :-)
Avoiding the page list would be complicated as we'd somehow have to code share the page table iterator scheme.
We're (slowly) getting towards thp for vram mappings and everything so I guess for drivers/gpu we might make that happen. But yeah it'd be not so pretty I think. -Daniel
dri-devel@lists.freedesktop.org
-
Daniel Vetter -
Jason Gunthorpe -
Jason Gunthorpe -
Leon Romanovsky