From patchwork Mon Jun 26 01:24:56 2017 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Timofey Titovets X-Patchwork-Id: 9808503 Return-Path: Received: from mail.wl.linuxfoundation.org (pdx-wl-mail.web.codeaurora.org [172.30.200.125]) by pdx-korg-patchwork.web.codeaurora.org (Postfix) with ESMTP id A0D0360329 for ; Mon, 26 Jun 2017 01:25:19 +0000 (UTC) Received: from mail.wl.linuxfoundation.org (localhost [127.0.0.1]) by mail.wl.linuxfoundation.org (Postfix) with ESMTP id A3F3626E39 for ; Mon, 26 Jun 2017 01:25:19 +0000 (UTC) Received: by mail.wl.linuxfoundation.org (Postfix, from userid 486) id 9860C2807B; Mon, 26 Jun 2017 01:25:19 +0000 (UTC) X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on pdx-wl-mail.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-6.3 required=2.0 tests=BAYES_00, DKIM_ADSP_CUSTOM_MED, DKIM_SIGNED, FREEMAIL_FROM, RCVD_IN_DNSWL_HI, RCVD_IN_SORBS_SPAM, T_DKIM_INVALID autolearn=ham version=3.3.1 Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.wl.linuxfoundation.org (Postfix) with ESMTP id CBAC726E39 for ; Mon, 26 Jun 2017 01:25:18 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751942AbdFZBZR (ORCPT ); Sun, 25 Jun 2017 21:25:17 -0400 Received: from mail-wr0-f193.google.com ([209.85.128.193]:35952 "EHLO mail-wr0-f193.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751919AbdFZBZM (ORCPT ); Sun, 25 Jun 2017 21:25:12 -0400 Received: by mail-wr0-f193.google.com with SMTP id 77so27205915wrb.3 for ; Sun, 25 Jun 2017 18:25:12 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20161025; h=from:to:cc:subject:date:message-id:in-reply-to:references; bh=7kMG/i5pRQjQ4BKFdbSs4Ud08OkoP0SeEfp3uF5tJ34=; b=T9NzJfgmJRhBWPQKxBext3aX0oiGzN5yjW1tDlZpos8zEoePBzVA9LLwM4WzuOTvIZ lpXuxoWEUhA1WPostqS3Efy6bka1lpZP8F2j2gdu6pXwVd/BVj9W1cnWgCmGsfbjs39V CouqHLxkxLh4I5umEuEd1hR4GXhfKJt/qI/V2XBarVLVrcMeZ4yADsUyQ/IAl+rUfVwx 3KdG0Ld5LJB79uVuKc3073+YwnH4jbhJncF5H3YGAoMBQrOtl9qlSMQbvuV9eZC5sttP 6sMgudX9JhKdROODGw+QuCORNhtNwPCVt5kAE6l4L1jseV2DEJ3Z3z4qcimdPs4Kkrft bHhg== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:from:to:cc:subject:date:message-id:in-reply-to :references; bh=7kMG/i5pRQjQ4BKFdbSs4Ud08OkoP0SeEfp3uF5tJ34=; b=elVAAtBeGoCfHiaGZUNzlp0IWwYbYurd/oTbai2jtAW1F92e7egpkx+sagbFJIyr1t 1juogJAHW+2SduKbEAahttvfCVi3m6RZWjgaseBKRHNVEZYxKdYSeW/bC3h1jKcKSDeR /6DEfpGRjDhk//o5pajvfUpizr1ALBSKPixZFgYDWYleexCjbctlGShShrTYyrQ0Tq2O GjltbbMd+LjNRBvICjaSM+L/K6x7Bhvc3xtkN503G1RriW1sKcMbYgEeW1L0tE5xq0Db VCIKVSJRgJfm7uGu3yrpdvve1fpOdoY7dtV8U232Sb8HMFyP08QvPiib1f+fSkeTTV47 QhOA== X-Gm-Message-State: AKS2vOyHRu9b1iC7lftyEj9rWx6jZ12qH6lz4FLsYGB2TH0u3WFFORC7 FWiKF88e29uJBHaD X-Received: by 10.223.172.67 with SMTP id v61mr13178563wrc.112.1498440310947; Sun, 25 Jun 2017 18:25:10 -0700 (PDT) Received: from titovetst-beplan.itransition.corp (nat3-minsk-pool-46-53-180-190.telecom.by. [46.53.180.190]) by smtp.gmail.com with ESMTPSA id 19sm16386909wrx.26.2017.06.25.18.25.10 (version=TLS1_2 cipher=ECDHE-RSA-AES128-GCM-SHA256 bits=128/128); Sun, 25 Jun 2017 18:25:10 -0700 (PDT) From: Timofey Titovets To: linux-btrfs@vger.kernel.org Cc: Timofey Titovets Subject: [RFC PATCH v2 2/2] Btrfs: add heuristic method for make decision compress or not compress Date: Mon, 26 Jun 2017 04:24:56 +0300 Message-Id: <20170626012456.19872-3-nefelim4ag@gmail.com> X-Mailer: git-send-email 2.13.1 In-Reply-To: <20170626012456.19872-1-nefelim4ag@gmail.com> References: <20170626012456.19872-1-nefelim4ag@gmail.com> Sender: linux-btrfs-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-btrfs@vger.kernel.org X-Virus-Scanned: ClamAV using ClamSMTP Add a heuristic computation before compression, for avoiding load resource heavy compression workspace, if data are probably can't be compressed. Signed-off-by: Timofey Titovets --- fs/btrfs/Makefile | 2 +- fs/btrfs/heuristic.c | 264 +++++++++++++++++++++++++++++++++++++++++++++++++++ fs/btrfs/heuristic.h | 13 +++ fs/btrfs/inode.c | 37 +++++--- 4 files changed, 301 insertions(+), 15 deletions(-) create mode 100644 fs/btrfs/heuristic.c create mode 100644 fs/btrfs/heuristic.h -- 2.13.1 -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html diff --git a/fs/btrfs/Makefile b/fs/btrfs/Makefile index 128ce17a80b0..8386095c9032 100644 --- a/fs/btrfs/Makefile +++ b/fs/btrfs/Makefile @@ -9,7 +9,7 @@ btrfs-y += super.o ctree.o extent-tree.o print-tree.o root-tree.o dir-item.o \ export.o tree-log.o free-space-cache.o zlib.o lzo.o \ compression.o delayed-ref.o relocation.o delayed-inode.o scrub.o \ reada.o backref.o ulist.o qgroup.o send.o dev-replace.o raid56.o \ - uuid-tree.o props.o hash.o free-space-tree.o + uuid-tree.o props.o hash.o free-space-tree.o heuristic.o log2_lshift16.o btrfs-$(CONFIG_BTRFS_FS_POSIX_ACL) += acl.o btrfs-$(CONFIG_BTRFS_FS_CHECK_INTEGRITY) += check-integrity.o diff --git a/fs/btrfs/heuristic.c b/fs/btrfs/heuristic.c new file mode 100644 index 000000000000..0fd90b5627aa --- /dev/null +++ b/fs/btrfs/heuristic.c @@ -0,0 +1,264 @@ +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "heuristic.h" +/* Precalculated log2 realization */ +#include "log2_lshift16.h" + +/* For shannon full integer entropy calculation */ +#define BUCKET_SIZE (1 << 8) + +struct _backet_item { + u8 padding; + u8 symbol; + u16 count; +}; + + +/* For sorting */ +static int compare(const void *lhs, const void *rhs) +{ + struct _backet_item *l = (struct _backet_item *)(lhs); + struct _backet_item *r = (struct _backet_item *)(rhs); + + return r->count - l->count; +} + +/* + * For good compressible data + * u8set size over sample + * will be small <= 64 + */ +static u32 _symbset_calc(const struct _backet_item *bucket) +{ + u32 a = 0; + u32 symbset_size = 0; + + for (; a < BUCKET_SIZE && symbset_size <= 64; a++) { + if (bucket[a].count) + symbset_size++; + } + return symbset_size; +} + + +/* + * Try calculate coreset size + * i.e. how many symbols use 90% of input data + * < 50 - good compressible data + * > 200 - bad compressible data + * For right & fast calculation bucket must be reverse sorted + */ +static u32 _coreset_calc(const struct _backet_item *bucket, + const u32 sum_threshold) +{ + u32 a = 0; + u32 coreset_sum = 0; + + for (a = 0; a < 201 && bucket[a].count; a++) { + coreset_sum += bucket[a].count; + if (coreset_sum > sum_threshold) + break; + } + return a; +} + +static u64 _entropy_perc(const struct _backet_item *bucket, + const u32 sample_size) +{ + u64 a, p; + u64 entropy_sum = 0; + u64 entropy_max = LOG2_RET_SHIFT*8; + + for (a = 0; a < BUCKET_SIZE && bucket[a].count > 0; a++) { + p = bucket[a].count; + p = div64_u64(p*LOG2_ARG_SHIFT, sample_size); + entropy_sum += -p*log2_lshift16(p); + } + + entropy_sum = div64_u64(entropy_sum, LOG2_ARG_SHIFT); + return div64_u64(entropy_sum*100, entropy_max); +} + +/* Pair distance from random distribution */ +static u32 _rnd_dist(const struct _backet_item *bucket, + const u32 coreset_size, const u8 *sample, u32 sample_size) +{ + u32 a, b; + u8 pair_a[2], pair_b[2]; + u32 pairs_count; + u64 sum = 0; + u64 buf1, buf2; + + for (a = 0; a < coreset_size-1; a++) { + pairs_count = 0; + pair_a[0] = bucket[a].symbol; + pair_a[1] = bucket[a+1].symbol; + pair_b[1] = bucket[a].symbol; + pair_b[0] = bucket[a+1].symbol; + for (b = 0; b < sample_size-1; b++) { + u16 *pair_c = (u16 *) &sample[b]; + + if (pair_c == (u16 *) pair_a) + pairs_count++; + else if (pair_c == (u16 *) pair_b) + pairs_count++; + } + buf1 = bucket[a].count*bucket[a+1].count; + buf1 = div64_u64(buf1*100000, (sample_size*sample_size)); + buf2 = pairs_count*2*100000; + buf2 = div64_u64(pairs_count, sample_size); + sum += (buf1 - buf2)*(buf1 - buf2); + } + + return div64_u64(sum, 2048); +} + +/* + * Algorithm description + * 1. Get subset of data for fast computation + * 2. Scan bucket for symbol set + * - symbol set < 64 - data will be easy compressible, return + * 3. Try compute coreset size (symbols count that use 90% of input data) + * - reverse sort bucket + * - sum cells until we reach 90% threshold, + * incriment coreset size each time + * - coreset_size < 50 - data will be easy compressible, return + * > 200 - data will be bad compressible, return + * in general this looks like data compression ratio 0.2 - 0.8 + * 4. Compute shannon entropy + * - shannon entropy count of bytes and can't count pairs & entropy_calc + * so assume: + * - usage of entropy can lead to false negative + * so for prevent that (in bad) case it's useful to "count" pairs + * - entropy are not to high < 70% easy compressible, return + * - entropy are high < 90%, try count pairs, + * if there is any noticeable amount, compression are possible, return + * - entropy are high > 90%, try count pairs, + * if there is noticeable amount, compression are possible, return + */ + +#define READ_SIZE 16 + +enum compression_advice btrfs_compress_heuristic(struct inode *inode, + u64 start, u64 end) +{ + enum compression_advice ret = COMPRESS_NONE; + u64 bytes_len = end - start; + u64 index = start >> PAGE_SHIFT; + u64 end_index = end >> PAGE_SHIFT; + struct page *page; + u8 *input_data; + u64 offset_count, shift, sample_size; + u64 a, b, c; + struct _backet_item bucket[BUCKET_SIZE]; + u8 *sample; + + /* + * In data: 128K 64K 32K 4K + * Sample: 4096b 3072b 2048b 1024b + * Avoid allocating array bigger then 4kb + */ + + offset_count = 64 + bytes_len/512; + + if (bytes_len >= 96*1024) + offset_count = 256; + + shift = bytes_len/offset_count; + sample_size = offset_count*READ_SIZE; + + /* + * speedup by copy data to sample array +30% + * I think it's because of memcpy speed and + * cpu cache hits + */ + sample = kmalloc(sample_size, GFP_NOFS); + if (!sample) + goto out; + + memset(&bucket, 0, sizeof(bucket)); + + /* Preset symbols */ + for (a = 0; a < BUCKET_SIZE; a++) + bucket[a].symbol = a; + + /* Read small subset of data 1024b-4096b */ + a = 0; b = 0; c = 0; + while (index <= end_index) { + page = find_get_page(inode->i_mapping, index); + BUG_ON(!page); /* Pages should be in the extent_io_tree */ + input_data = kmap(page); + c = 0; + while (c < c+PAGE_SIZE && a < bytes_len && b < sample_size) { + memcpy(&sample[b], &input_data[c], READ_SIZE); + c += shift; + a += shift; + b += READ_SIZE; + } + kunmap(page); + put_page(page); + index++; + } + + for (a = 0; a < sample_size; a++) + bucket[sample[a]].count++; + + + a = _symbset_calc(bucket); + if (a < 64) { + ret = COMPRESS_COST_EASY; + goto out; + } + + /* Sort in reverse order */ + sort(bucket, BUCKET_SIZE, sizeof(u32), &compare, NULL); + + a = _coreset_calc(bucket, sample_size*90/100); + + if (a < 50) { + ret = COMPRESS_COST_EASY; + goto out; + } + + if (a > 200) { + ret = COMPRESS_NONE; + goto out; + } + + /* + * Okay, code fail to fast detect data type + * Let's calculate entropy + */ + b = _entropy_perc(bucket, sample_size); + if (b < 70) { + ret = COMPRESS_COST_MEDIUM; + goto out; + } + + a = _rnd_dist(bucket, a, sample, sample_size); + if (b < 90) { + if (a > 0) + ret = COMPRESS_COST_MEDIUM; + else + ret = COMPRESS_NONE; + goto out; + } else { + if (a > 10) + ret = COMPRESS_COST_HARD; + else + ret = COMPRESS_NONE; + goto out; + } + +out: + kfree(sample); + return ret; +} diff --git a/fs/btrfs/heuristic.h b/fs/btrfs/heuristic.h new file mode 100644 index 000000000000..f978bb3c21b1 --- /dev/null +++ b/fs/btrfs/heuristic.h @@ -0,0 +1,13 @@ +#include +#include +#include + +enum compression_advice { + COMPRESS_NONE, + COMPRESS_COST_EASY, + COMPRESS_COST_MEDIUM, + COMPRESS_COST_HARD +}; + +enum compression_advice btrfs_compress_heuristic(struct inode *inode, + u64 start, u64 end); diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c index ef3c98c527c1..d5a13e3e0c6c 100644 --- a/fs/btrfs/inode.c +++ b/fs/btrfs/inode.c @@ -60,6 +60,7 @@ #include "props.h" #include "qgroup.h" #include "dedupe.h" +#include "heuristic.h" struct btrfs_iget_args { struct btrfs_key *location; @@ -458,16 +459,16 @@ static noinline void compress_file_range(struct inode *inode, unsigned long total_compressed = 0; unsigned long total_in = 0; int i; - int will_compress; + bool will_compress; int compress_type = fs_info->compress_type; - int redirty = 0; + bool redirty = false; inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1, SZ_16K); actual_end = min_t(u64, isize, end + 1); again: - will_compress = 0; + will_compress = false; nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1; BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0); nr_pages = min_t(unsigned long, nr_pages, @@ -510,15 +511,6 @@ static noinline void compress_file_range(struct inode *inode, */ if (inode_need_compress(inode)) { WARN_ON(pages); - pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS); - if (!pages) { - /* just bail out to the uncompressed code */ - goto cont; - } - - if (BTRFS_I(inode)->force_compress) - compress_type = BTRFS_I(inode)->force_compress; - /* * we need to call clear_page_dirty_for_io on each * page in the range. Otherwise applications with the file @@ -529,7 +521,24 @@ static noinline void compress_file_range(struct inode *inode, * dirty again later on. */ extent_range_clear_dirty_for_io(inode, start, end); - redirty = 1; + redirty = true; + + ret = btrfs_compress_heuristic(inode, start, end); + + /* Heuristic say: dont try compress that */ + if (!ret) + goto cont; + + pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS); + if (!pages) { + /* just bail out to the uncompressed code */ + goto cont; + } + + if (BTRFS_I(inode)->force_compress) + compress_type = BTRFS_I(inode)->force_compress; + + ret = btrfs_compress_pages(compress_type, inode->i_mapping, start, pages, @@ -552,7 +561,7 @@ static noinline void compress_file_range(struct inode *inode, PAGE_SIZE - offset); kunmap_atomic(kaddr); } - will_compress = 1; + will_compress = true; } } cont: