vmalloc() consists of the following: (1) find empty space in the vmalloc space -> (2) get physical pages from the buddy system -> (3) map the pages into the pagetable. It turns out that the cost of (1) and (3) is pretty insignificant. Hence, the cost of vmalloc becomes highly sensitive to physical memory allocation time. Currently, if we decide to use huge mappings, apart from aligning the start of the target vm_struct region to the huge-alignment, we also align the size. This does not seem to produce any benefit (apart from simplification of the code), and there is a clear disadvantage - as mentioned above, the main cost of vmalloc comes from its interaction with the buddy system, and thus requesting more memory than was requested by the caller is suboptimal and unnecessary. This change is also motivated due to the next patch ("arm64/mm: Enable vmalloc-huge by default"). Suppose that some user of vmalloc maps 17 pages, uses that mapping for an extremely short time, and vfree's it. That patch, without this patch, on arm64 will ultimately map 16 * 2 = 32 pages in a contiguous way. Since the mapping is used for a very short time, it is likely that the extra cost of mapping 15 pages defeats any benefit from reduced TLB pressure, and regresses that code path. Signed-off-by: Dev Jain --- mm/vmalloc.c | 38 ++++++++++++++++++++++++++++++-------- 1 file changed, 30 insertions(+), 8 deletions(-) diff --git a/mm/vmalloc.c b/mm/vmalloc.c index 798b2ed21e46..ddd9294a4634 100644 --- a/mm/vmalloc.c +++ b/mm/vmalloc.c @@ -647,7 +647,7 @@ static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end, int __vmap_pages_range_noflush(unsigned long addr, unsigned long end, pgprot_t prot, struct page **pages, unsigned int page_shift) { - unsigned int i, nr = (end - addr) >> PAGE_SHIFT; + unsigned int i, step, nr = (end - addr) >> PAGE_SHIFT; WARN_ON(page_shift < PAGE_SHIFT); @@ -655,7 +655,8 @@ int __vmap_pages_range_noflush(unsigned long addr, unsigned long end, page_shift == PAGE_SHIFT) return vmap_small_pages_range_noflush(addr, end, prot, pages); - for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) { + step = 1U << (page_shift - PAGE_SHIFT); + for (i = 0; i < ALIGN_DOWN(nr, step); i += step) { int err; err = vmap_range_noflush(addr, addr + (1UL << page_shift), @@ -666,8 +667,9 @@ int __vmap_pages_range_noflush(unsigned long addr, unsigned long end, addr += 1UL << page_shift; } - - return 0; + if (IS_ALIGNED(nr, step)) + return 0; + return vmap_small_pages_range_noflush(addr, end, prot, pages + i); } int vmap_pages_range_noflush(unsigned long addr, unsigned long end, @@ -3171,7 +3173,7 @@ struct vm_struct *__get_vm_area_node(unsigned long size, unsigned long requested_size = size; BUG_ON(in_interrupt()); - size = ALIGN(size, 1ul << shift); + size = PAGE_ALIGN(size); if (unlikely(!size)) return NULL; @@ -3327,7 +3329,7 @@ static void vm_reset_perms(struct vm_struct *area) * Find the start and end range of the direct mappings to make sure that * the vm_unmap_aliases() flush includes the direct map. */ - for (i = 0; i < area->nr_pages; i += 1U << page_order) { + for (i = 0; i < ALIGN_DOWN(area->nr_pages, 1U << page_order); i += (1U << page_order)) { unsigned long addr = (unsigned long)page_address(area->pages[i]); if (addr) { @@ -3339,6 +3341,18 @@ static void vm_reset_perms(struct vm_struct *area) flush_dmap = 1; } } + for (; i < area->nr_pages; ++i) { + unsigned long addr = (unsigned long)page_address(area->pages[i]); + + if (addr) { + unsigned long page_size; + + page_size = PAGE_SIZE; + start = min(addr, start); + end = max(addr + page_size, end); + flush_dmap = 1; + } + } /* * Set direct map to something invalid so that it won't be cached if @@ -3602,6 +3616,7 @@ vm_area_alloc_pages(gfp_t gfp, int nid, * more permissive. */ if (!order) { +page_map: while (nr_allocated < nr_pages) { unsigned int nr, nr_pages_request; @@ -3633,13 +3648,18 @@ vm_area_alloc_pages(gfp_t gfp, int nid, * If zero or pages were obtained partly, * fallback to a single page allocator. */ - if (nr != nr_pages_request) + if (nr != nr_pages_request) { + order = 0; break; + } } } /* High-order pages or fallback path if "bulk" fails. */ while (nr_allocated < nr_pages) { + if (nr_pages - nr_allocated < (1UL << order)) + goto page_map; + if (!(gfp & __GFP_NOFAIL) && fatal_signal_pending(current)) break; @@ -5024,7 +5044,9 @@ static void show_numa_info(struct seq_file *m, struct vm_struct *v, memset(counters, 0, nr_node_ids * sizeof(unsigned int)); - for (nr = 0; nr < v->nr_pages; nr += step) + for (nr = 0; nr < ALIGN_DOWN(v->nr_pages, step); nr += step) + counters[page_to_nid(v->pages[nr])] += step; + for (; nr < v->nr_pages; ++nr) counters[page_to_nid(v->pages[nr])] += step; for_each_node_state(nr, N_HIGH_MEMORY) if (counters[nr]) -- 2.30.2 For BBML2-noabort arm64 systems, enable vmalloc cont mappings and PMD mappings by default. There is benefit to be gained in any code path which maps >= 16 pages using vmalloc, since any usage of that mapping will now come with reduced TLB pressure. Currently, I am not being able to produce a reliable statistically significant improvement for the benchmarks which we have. I am optimistic that xfs benchmarks should give some benefit. Upon running test_vmalloc.sh, this series produces an optimization and some regressions. I conclude that we should ignore the results of this testsuite. I explain the regression in the long_busy_list_alloc_test below: upon running ./test_vmalloc.sh run_test_mask=4 nr_threads=1, a regression of approx 17% is observed (which increases to 31% if we do *not* apply the previous patch ("mm/vmalloc: Do not align size to huge size")). The long_busy_list_alloc_test first maps a lot of single pages to fragment the vmalloc space. Then, it does the following in a loop: map 100 pages, map a single page, then vfree both of them. My investigation reveals that the majority of time is *not* spent in finding a free space in the vmalloc region (which is exactly the time which the setup of this particular test wants to increase), but in the interaction with the physical memory allocator. It turns out that mapping 100 pages in a contiguous way is *faster* than bulk mapping 100 single pages. The regression is actually carried by vfree(). When we contpte map 100 pages, we get 6 * 16 = 96 pages from the free lists of the buddy allocator, and not the pcp lists. Then, vmalloc subsystem splits this page into individual pages because drivers can operate on individual pages, messing up the refcounts. As a result, vfree frees these pages as single 4k pages, freeing them into the pcp lists. Thus, now we have got a behaviour of taking from the freelists of the buddy, and freeing into the pcp lists, which forces pcp draining into the freelists. By playing with the following code in mm/page_alloc.c: high = nr_pcp_high(pcp, zone, batch, free_high); if (pcp->count < high) return; The time taken by the test is highly sensitive to the value returned by nr_pcp_high (although, increasing the value of high does not reduce the regression). Summarizing, the regression is due to messing up the state of the buddy system by rapidly stealing from the freelists and not giving back to them. If we insert an msleep(1) just before we vfree() both the regions, the regression reduces. This proves that the regression is due to the unnatural behaviour of the test - it allocates memory, does absolutely nothing with that memory, and releases it. No workload is expected to map memory without actually utilizing it for some time. The time between vmalloc() and vfree() will give time for the buddy to stabilize, and the regression is eliminated. The optimization is observed in fix_size_alloc_test with nr_pages = 512, because both vmalloc() and vfree() will now operate to and from the pcp. Signed-off-by: Dev Jain --- arch/arm64/include/asm/vmalloc.h | 6 ++++++ arch/arm64/mm/pageattr.c | 4 +--- include/linux/vmalloc.h | 7 +++++++ mm/vmalloc.c | 6 +++++- 4 files changed, 19 insertions(+), 4 deletions(-) diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h index 4ec1acd3c1b3..c72ae9bd7360 100644 --- a/arch/arm64/include/asm/vmalloc.h +++ b/arch/arm64/include/asm/vmalloc.h @@ -6,6 +6,12 @@ #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP +#define arch_wants_vmalloc_huge_always arch_wants_vmalloc_huge_always +static inline bool arch_wants_vmalloc_huge_always(void) +{ + return system_supports_bbml2_noabort(); +} + #define arch_vmap_pud_supported arch_vmap_pud_supported static inline bool arch_vmap_pud_supported(pgprot_t prot) { diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c index 5135f2d66958..b800e3a3fe85 100644 --- a/arch/arm64/mm/pageattr.c +++ b/arch/arm64/mm/pageattr.c @@ -163,8 +163,6 @@ static int change_memory_common(unsigned long addr, int numpages, * we are operating on does not result in such splitting. * * Let's restrict ourselves to mappings created by vmalloc (or vmap). - * Disallow VM_ALLOW_HUGE_VMAP mappings to guarantee that only page - * mappings are updated and splitting is never needed. * * So check whether the [addr, addr + size) interval is entirely * covered by precisely one VM area that has the VM_ALLOC flag set. @@ -172,7 +170,7 @@ static int change_memory_common(unsigned long addr, int numpages, area = find_vm_area((void *)addr); if (!area || end > (unsigned long)kasan_reset_tag(area->addr) + area->size || - ((area->flags & (VM_ALLOC | VM_ALLOW_HUGE_VMAP)) != VM_ALLOC)) + ((area->flags & VM_ALLOC) != VM_ALLOC)) return -EINVAL; if (!numpages) diff --git a/include/linux/vmalloc.h b/include/linux/vmalloc.h index eb54b7b3202f..b0f04f7e8cfa 100644 --- a/include/linux/vmalloc.h +++ b/include/linux/vmalloc.h @@ -84,6 +84,13 @@ struct vmap_area { unsigned long flags; /* mark type of vm_map_ram area */ }; +#ifndef arch_wants_vmalloc_huge_always +static inline bool arch_wants_vmalloc_huge_always(void) +{ + return false; +} +#endif + /* archs that select HAVE_ARCH_HUGE_VMAP should override one or more of these */ #ifndef arch_vmap_p4d_supported static inline bool arch_vmap_p4d_supported(pgprot_t prot) diff --git a/mm/vmalloc.c b/mm/vmalloc.c index ddd9294a4634..99da3d256360 100644 --- a/mm/vmalloc.c +++ b/mm/vmalloc.c @@ -3857,7 +3857,8 @@ void *__vmalloc_node_range_noprof(unsigned long size, unsigned long align, return NULL; } - if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) { + if (vmap_allow_huge && ((arch_wants_vmalloc_huge_always()) || + (vm_flags & VM_ALLOW_HUGE_VMAP))) { /* * Try huge pages. Only try for PAGE_KERNEL allocations, * others like modules don't yet expect huge pages in @@ -3871,6 +3872,9 @@ void *__vmalloc_node_range_noprof(unsigned long size, unsigned long align, shift = arch_vmap_pte_supported_shift(size); align = max(original_align, 1UL << shift); + + /* If arch wants huge by default, set flag unconditionally */ + vm_flags |= VM_ALLOW_HUGE_VMAP; } again: -- 2.30.2