On Tue, Sep 02, 2025 at 07:37:22PM +0000, Viacheslav Dubeyko wrote: > On Sat, 2025-08-30 at 21:28 +0800, Guan-Chun Wu wrote: > > Previously, ceph_base64_encode() used a bitstream approach, handling one > > input byte at a time and performing extra bit operations. While correct, > > this method was suboptimal. > > > > Sounds interesting! > > Is ceph_base64_decode() efficient then? > Do we have something in crypto library of Linux kernel? Maybe we can use > something efficient enough from there? > Hi Viacheslav, FYI, we already have base64 encode/decode implementations in lib/base64.c. As discussed in another thread [1], I think we can put the optimized version there and have users switch to call the library functions. [1]: https://lore.kernel.org/lkml/38753d95-8503-4b72-9590-cb129aa49a41@xxxxxxxx/ Hi Guan-Chun, I was also trying optimizing base64 performance, but I saw your patch first. Happy to help if you need any assistance! Regards, Kuan-Wei > > This patch processes input in 3-byte blocks, mapping directly to 4 output > > characters. Remaining 1 or 2 bytes are handled according to standard Base64 > > rules. This reduces computation and improves performance. > > > > So, why namely 3-byte blocks? Could you please explain in more details your > motivation and improved technique in commit message? How exactly your technique > reduces computation and improves performance? > > > Performance test (5 runs) for ceph_base64_encode(): > > > > 64B input: > > ------------------------------------------------------- > > > Old method | 123 | 115 | 137 | 119 | 109 | avg ~121 ns | > > ------------------------------------------------------- > > > New method | 84 | 83 | 86 | 85 | 84 | avg ~84 ns | > > ------------------------------------------------------- > > > > 1KB input: > > -------------------------------------------------------- > > > Old method | 1217 | 1150 | 1146 | 1149 | 1149 | avg ~1162 ns | > > -------------------------------------------------------- > > > New method | 776 | 772 | 772 | 774 | 770 | avg ~773 ns | > > -------------------------------------------------------- > > > > Signed-off-by: Guan-Chun Wu <409411716@xxxxxxxxxxxxxx> > > --- > > Tested on Linux 6.8.0-64-generic x86_64 > > with Intel Core i7-10700 @ 2.90GHz > > > > I assume that it is still the commit message. So, I think this portion should be > before Signed-off-by. > > > Test is executed in the form of kernel module. > > > Test script: > > > > Is it finally script or kernel module? As far as I can see, it is not complete > source code. So, I am not sure that everybody will be capable to build and test > this module. > > What's about to introduce this as Kunit test or self-test that can be used by > everybody in CephFS kernel client for testing and checking performance? I am > working on initial set of Kunit tests for CephFS kernel client right now. > > > static int encode_v1(const u8 *src, int srclen, char *dst) > > { > > u32 ac = 0; > > int bits = 0; > > int i; > > char *cp = dst; > > > > for (i = 0; i < srclen; i++) { > > ac = (ac << 8) | src[i]; > > bits += 8; > > do { > > bits -= 6; > > *cp++ = base64_table[(ac >> bits) & 0x3f]; > > } while (bits >= 6); > > } > > if (bits) > > *cp++ = base64_table[(ac << (6 - bits)) & 0x3f]; > > return cp - dst; > > } > > > > static int encode_v2(const u8 *src, int srclen, char *dst) > > { > > u32 ac = 0; > > int i = 0; > > char *cp = dst; > > > > while (i + 2 < srclen) { > > ac = ((u32)src[i] << 16) | ((u32)src[i + 1] << 8) | (u32)src[i + 2]; > > *cp++ = base64_table[(ac >> 18) & 0x3f]; > > *cp++ = base64_table[(ac >> 12) & 0x3f]; > > *cp++ = base64_table[(ac >> 6) & 0x3f]; > > *cp++ = base64_table[ac & 0x3f]; > > i += 3; > > } > > > > switch (srclen - i) { > > case 2: > > ac = ((u32)src[i] << 16) | ((u32)src[i + 1] << 8); > > *cp++ = base64_table[(ac >> 18) & 0x3f]; > > *cp++ = base64_table[(ac >> 12) & 0x3f]; > > *cp++ = base64_table[(ac >> 6) & 0x3f]; > > break; > > case 1: > > ac = ((u32)src[i] << 16); > > *cp++ = base64_table[(ac >> 18) & 0x3f]; > > *cp++ = base64_table[(ac >> 12) & 0x3f]; > > break; > > } > > return cp - dst; > > } > > > > static void run_test(const char *label, const u8 *data, int len) > > { > > char *dst1, *dst2; > > int n1, n2; > > u64 start, end; > > > > dst1 = kmalloc(len * 2, GFP_KERNEL); > > dst2 = kmalloc(len * 2, GFP_KERNEL); > > > > if (!dst1 || !dst2) { > > pr_err("%s: Failed to allocate dst buffers\n", label); > > goto out; > > } > > > > pr_info("[%s] input size = %d bytes\n", label, len); > > > > start = ktime_get_ns(); > > n1 = encode_v1(data, len, dst1); > > end = ktime_get_ns(); > > pr_info("[%s] encode_v1 time: %lld ns\n", label, end - start); > > > > start = ktime_get_ns(); > > n2 = encode_v2(data, len, dst2); > > end = ktime_get_ns(); > > pr_info("[%s] encode_v2 time: %lld ns\n", label, end - start); > > > > if (n1 != n2 || memcmp(dst1, dst2, n1) != 0) > > pr_err("[%s] Mismatch detected between encode_v1 and encode_v2!\n", label); > > else > > pr_info("[%s] Outputs are identical.\n", label); > > > > out: > > kfree(dst1); > > kfree(dst2); > > } > > --- > > fs/ceph/crypto.c | 33 ++++++++++++++++++++++----------- > > 1 file changed, 22 insertions(+), 11 deletions(-) > > > > diff --git a/fs/ceph/crypto.c b/fs/ceph/crypto.c > > index 3b3c4d8d401e..a35570fd8ff5 100644 > > --- a/fs/ceph/crypto.c > > +++ b/fs/ceph/crypto.c > > @@ -27,20 +27,31 @@ static const char base64_table[65] = > > int ceph_base64_encode(const u8 *src, int srclen, char *dst) > > { > > u32 ac = 0; > > - int bits = 0; > > - int i; > > + int i = 0; > > char *cp = dst; > > > > - for (i = 0; i < srclen; i++) { > > - ac = (ac << 8) | src[i]; > > - bits += 8; > > - do { > > - bits -= 6; > > - *cp++ = base64_table[(ac >> bits) & 0x3f]; > > - } while (bits >= 6); > > + while (i + 2 < srclen) { > > Frankly speaking, I am not completely happy about hardcoded constants. As a > result, it makes code hard to understand, modify and support. Could you please > introduce named constants instead of hardcoded numbers? > > > > + ac = ((u32)src[i] << 16) | ((u32)src[i + 1] << 8) | (u32)src[i + 2]; > > + *cp++ = base64_table[(ac >> 18) & 0x3f]; > > + *cp++ = base64_table[(ac >> 12) & 0x3f]; > > + *cp++ = base64_table[(ac >> 6) & 0x3f]; > > + *cp++ = base64_table[ac & 0x3f]; > > + i += 3; > > + } > > + > > + switch (srclen - i) { > > + case 2: > > + ac = ((u32)src[i] << 16) | ((u32)src[i + 1] << 8); > > + *cp++ = base64_table[(ac >> 18) & 0x3f]; > > + *cp++ = base64_table[(ac >> 12) & 0x3f]; > > + *cp++ = base64_table[(ac >> 6) & 0x3f]; > > + break; > > + case 1: > > + ac = ((u32)src[i] << 16); > > + *cp++ = base64_table[(ac >> 18) & 0x3f]; > > + *cp++ = base64_table[(ac >> 12) & 0x3f]; > > + break; > > } > > - if (bits) > > - *cp++ = base64_table[(ac << (6 - bits)) & 0x3f]; > > return cp - dst; > > } > > > > Let me test your patch and check that it doesn't introduce regression(s). > > Thanks, > Slava.
