Skip to content

K
Kernel

openeuler / Kernel
1 年多 前同步成功

通知 8
Star 0
Fork 0
K
Kernel
切换分支/标签
  1. 28 7月, 2018 1 次提交
    • E
      ARC: dma [non-IOC] setup SMP_CACHE_BYTES and cache_line_size · eb277739
      由 Eugeniy Paltsev 提交于 
      As for today we don't setup SMP_CACHE_BYTES and cache_line_size for
ARC, so they are set to L1_CACHE_BYTES by default. L1 line length
(L1_CACHE_BYTES) might be easily smaller than L2 line (which is
usually the case BTW). This breaks code.

For example this breaks ethernet infrastructure on HSDK/AXS103 boards
with IOC disabled, involving manual cache flushes
Functions which alloc and manage sk_buff packet data area rely on
SMP_CACHE_BYTES define. In the result we can share last L2 cache
line in sk_buff linear packet data area between DMA buffer and
some useful data in other structure. So we can lose this data when
we invalidate DMA buffer.

   sk_buff linear packet data area
                |
                |
                |         skb->end        skb->tail
                V            |                |
                             V                V
----------------------------------------------.
      packet data            | <tail padding> |  <useful data in other struct>
----------------------------------------------.

---------------------.--------------------------------------------------.
     SLC line        |             SLC (L2 cache) line (128B)           |
---------------------.--------------------------------------------------.
        ^                                     ^
        |                                     |
     These cache lines will be invalidated when we invalidate skb
     linear packet data area before DMA transaction starting.

This leads to issues painful to debug as it reproduces only if
(sk_buff->end - sk_buff->tail) < SLC_LINE_SIZE and
if we have some useful data right after sk_buff->end.

Fix that by hardcode SMP_CACHE_BYTES to max line length we may have.
Signed-off-by: NEugeniy Paltsev <Eugeniy.Paltsev@synopsys.com>
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
      eb277739
  3. 20 7月, 2018 1 次提交
  5. 08 6月, 2018 1 次提交
    • L
      mm: introduce ARCH_HAS_PTE_SPECIAL · 3010a5ea
      由 Laurent Dufour 提交于 
      Currently the PTE special supports is turned on in per architecture
header files.  Most of the time, it is defined in
arch/*/include/asm/pgtable.h depending or not on some other per
architecture static definition.

This patch introduce a new configuration variable to manage this
directly in the Kconfig files.  It would later replace
__HAVE_ARCH_PTE_SPECIAL.

Here notes for some architecture where the definition of
__HAVE_ARCH_PTE_SPECIAL is not obvious:

arm
 __HAVE_ARCH_PTE_SPECIAL which is currently defined in
arch/arm/include/asm/pgtable-3level.h which is included by
arch/arm/include/asm/pgtable.h when CONFIG_ARM_LPAE is set.
So select ARCH_HAS_PTE_SPECIAL if ARM_LPAE.

powerpc
__HAVE_ARCH_PTE_SPECIAL is defined in 2 files:
 - arch/powerpc/include/asm/book3s/64/pgtable.h
 - arch/powerpc/include/asm/pte-common.h
The first one is included if (PPC_BOOK3S & PPC64) while the second is
included in all the other cases.
So select ARCH_HAS_PTE_SPECIAL all the time.

sparc:
__HAVE_ARCH_PTE_SPECIAL is defined if defined(__sparc__) &&
defined(__arch64__) which are defined through the compiler in
sparc/Makefile if !SPARC32 which I assume to be if SPARC64.
So select ARCH_HAS_PTE_SPECIAL if SPARC64

There is no functional change introduced by this patch.

Link: http://lkml.kernel.org/r/1523433816-14460-2-git-send-email-ldufour@linux.vnet.ibm.comSigned-off-by: NLaurent Dufour <ldufour@linux.vnet.ibm.com>
Suggested-by: NJerome Glisse <jglisse@redhat.com>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Acked-by: NDavid Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: "Aneesh Kumar K . V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: Rich Felker <dalias@libc.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Vineet Gupta <vgupta@synopsys.com>
Cc: Palmer Dabbelt <palmer@sifive.com>
Cc: Albert Ou <albert@sifive.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Robin Murphy <robin.murphy@arm.com>
Cc: Christophe LEROY <christophe.leroy@c-s.fr>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      3010a5ea
  7. 19 5月, 2018 1 次提交
  9. 09 5月, 2018 2 次提交
  11. 06 2月, 2018 1 次提交
  13. 09 1月, 2018 1 次提交
  15. 22 11月, 2017 1 次提交
    • V
      ARC: perf: avoid vmalloc backed mmap · 82385732
      由 Vineet Gupta 提交于 
      For non-alising Dcache, vmalloc is not needed.

vmalloc triggers additonal D-TLB Misses in the perf interrupt code path
making it slightly inefficient as evident from hackbench runs below.

| [ARCLinux]# perf stat -e dTLB-load-misses --repeat 5 hackbench
| Running with 10*40 (== 400) tasks.
| Time: 35.060
| ...
| Performance counter stats for 'hackbench' (5 runs):

Before:      399235      dTLB-load-misses ( +-  2.08% )
After :      397676      dTLB-load-misses ( +-  2.27% )
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
      82385732
  17. 12 10月, 2017 1 次提交
  19. 04 10月, 2017 1 次提交
  21. 02 9月, 2017 2 次提交
    • A
      ARC: [plat-hsdk] initial port for HSDK board · a518d637
      由 Alexey Brodkin 提交于 
      This initial port adds support of ARC HS Development Kit board with some
basic features such serial port, USB, SD/MMC and Ethernet.

Essentially we run Linux kernel on all 4 cores (i.e. utilize SMP) and
heavily use IO Coherency for speeding-up DMA-aware peripherals.

Note as opposed to other ARC boards we link Linux kernel to
0x9000_0000 intentionally because cores 1 and 3 configured with DCCM
situated at our more usual link base 0x8000_0000. We still can use
memory region starting at 0x8000_0000 as we reallocate DCCM in our
platform code.

Note that PAE remapping for DMA clients does not work due to an RTL bug,
so CREG_PAE register must be programmed to all zeroes, otherwise it will
cause problems with DMA to/from peripherals even if PAE40 is not used.
Acked-by: NRob Herring <robh@kernel.org>
Signed-off-by: NAlexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: NEugeniy Paltsev <Eugeniy.Paltsev@synopsys.com>
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
      a518d637
    • E
      ARC: mm: Decouple RAM base address from kernel link address · 9ed68785
      由 Eugeniy Paltsev 提交于 
      	[Needed for HSDK]

Currently the first page of system (hence RAM base) is assumed to be
@ CONFIG_LINUX_LINK_BASE, where kernel itself is linked.

However is case of HSDK platform, for reasons explained in that patch,
this is not true. kernel needs to be linked @ 0x9000_0000 while DDR
is still wired at 0x8000_0000. To properly account for this 256M of RAM,
we need to introduce a new option and base page frame accountiing off of
it.
Signed-off-by: NEugeniy Paltsev <Eugeniy.Paltsev@synopsys.com>
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
[vgupta: renamed  CONFIG_KERNEL_RAM_BASE_ADDRESS => CONFIG_LINUX_RAM_BASE
       : simplified changelog]
      9ed68785
  23. 04 8月, 2017 1 次提交
  25. 06 5月, 2017 1 次提交
  27. 27 4月, 2017 1 次提交
  29. 21 4月, 2017 1 次提交
    • V
      ARCv2: entry: save Accumulator register pair (r58:59) if present · 3d5e8012
      由 Vineet Gupta 提交于 
      Accumulator is present in configs with FPU and/or DSP MPY (mpy > 6)

Instead of doing this in pt_regs (and thus every kernel entry/exit),
this could have been done in context switch (and for user task only) as
currently kernel doesn't clobber these registers for its own accord.
However we will soon start using 64-bit multiply instructions for kernel
which can clobber these. Also gcc folks also plan to start using these
as GPRs, hence better to always save/restore them
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
      3d5e8012
  31. 29 3月, 2017 1 次提交
  33. 07 2月, 2017 2 次提交
    • V
      ARC: [plat-*] ARC_HAS_COH_CACHES no longer relevant · 8ba605b6
      由 Vineet Gupta 提交于 
      A typical SMP system expects cache coherency. Initial NPS platform
support was slated to be SMP w/o cache coherency.

However it seems the platform now selects that option, so there is no
point in keeping it around.
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
      8ba605b6
    • Y
      ARCv2: intc: Rework the build time irq count information · f33b8cdd
      由 Yuriy Kolerov 提交于 
      Currently Kconfig knob ARC_NUMBER_OF_INTERRUPTS is used as indicator of
hard irq count. But it is flawed that it doesn't affect
 - NR_IRQS     : for number of virtual interrupts
 - NR_CPU_IRQS : for number of hardware interrupts

Moreover the actual hardware irq count might still not be same as
ARC_NUMBER_OF_INTERRUPTS. So use the information availble in the
Build Configuration Registers and get rid of the Kconfig option.

We still need "some" build time info about irq count to set up
sufficient number of vector table entries. This is done with a
sufficiently large NR_CPU_IRQS which will eventually be used soley for
that purpose (subsequent patches will remove its usage elsewhere)

So to summarize what this patch does:

  * NR_CPU_IRQS defines a maximum number of hardware interrupts.
  * Remove ARC_NUMBER_OF_INTERRUPTS option and create interrupts
    table for all possible hardware interrupts.
  * Increase a maximum number of virtual IRQs to 512. ARCv2 can
    support 240 interrupts in the core interrupts controllers
    and 128 interrupts in IDU. Thus 512 virtual IRQs must be
    enough for most configurations of boards.

This patch leads to NR_CPU_IRQS in 2 places, to reduce the overall
churn. The next patch will remove the 2nd definition anyways.
Signed-off-by: NYuriy Kolerov <yuriy.kolerov@synopsys.com>
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
[vgupta: reworked the changelog a bit]
      f33b8cdd
  35. 19 1月, 2017 1 次提交
  37. 19 12月, 2016 1 次提交
  39. 01 12月, 2016 2 次提交
  41. 29 10月, 2016 1 次提交
  43. 17 10月, 2016 2 次提交
    • D
      ARC: [build] Support gz, lzma compressed uImage · 27f3d2a3
      由 Daniel Mentz 提交于 
      Add support for lzma compressed uImage.

Support for gzip was already available but could not be enabled because
we were missing CONFIG_HAVE_KERNEL_GZIP in arch/arc/Kconfig.
Signed-off-by: NDaniel Mentz <danielmentz@google.com>
Cc: linux-snps-arc@lists.infradead.org
Cc: Vineet Gupta <Vineet.Gupta1@synopsys.com>
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
      27f3d2a3
    • V
      ARCv2: intc: untangle SMP, MCIP and IDU · 3ce0fefc
      由 Vineet Gupta 提交于 
      The IDU intc is technically part of MCIP (Multi-core IP) hence
historically was only available in a SMP hardware build (and thus only
in a SMP kernel build). Now that hardware restriction has been lifted,
so a UP kernel needs to support it.

This requires breaking mcip.c into parts which are strictly SMP
(inter-core interrupts) and IDU which in reality is just another
intc and thus has no bearing on SMP.

This change allows IDU in UP builds and with a suitable device tree, we
can have the cascaded intc system

    ARCv2 core intc <---> ARCv2 IDU intc <---> periperals
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
      3ce0fefc
  45. 01 10月, 2016 2 次提交
  47. 02 6月, 2016 1 次提交
  49. 31 5月, 2016 2 次提交
  51. 21 5月, 2016 1 次提交
    • Z
      lib/GCD.c: use binary GCD algorithm instead of Euclidean · fff7fb0b
      由 Zhaoxiu Zeng 提交于 
      The binary GCD algorithm is based on the following facts:
	1. If a and b are all evens, then gcd(a,b) = 2 * gcd(a/2, b/2)
	2. If a is even and b is odd, then gcd(a,b) = gcd(a/2, b)
	3. If a and b are all odds, then gcd(a,b) = gcd((a-b)/2, b) = gcd((a+b)/2, b)

Even on x86 machines with reasonable division hardware, the binary
algorithm runs about 25% faster (80% the execution time) than the
division-based Euclidian algorithm.

On platforms like Alpha and ARMv6 where division is a function call to
emulation code, it's even more significant.

There are two variants of the code here, depending on whether a fast
__ffs (find least significant set bit) instruction is available.  This
allows the unpredictable branches in the bit-at-a-time shifting loop to
be eliminated.

If fast __ffs is not available, the "even/odd" GCD variant is used.

I use the following code to benchmark:

	#include <stdio.h>
	#include <stdlib.h>
	#include <stdint.h>
	#include <string.h>
	#include <time.h>
	#include <unistd.h>

	#define swap(a, b) \
		do { \
			a ^= b; \
			b ^= a; \
			a ^= b; \
		} while (0)

	unsigned long gcd0(unsigned long a, unsigned long b)
	{
		unsigned long r;

		if (a < b) {
			swap(a, b);
		}

		if (b == 0)
			return a;

		while ((r = a % b) != 0) {
			a = b;
			b = r;
		}

		return b;
	}

	unsigned long gcd1(unsigned long a, unsigned long b)
	{
		unsigned long r = a | b;

		if (!a || !b)
			return r;

		b >>= __builtin_ctzl(b);

		for (;;) {
			a >>= __builtin_ctzl(a);
			if (a == b)
				return a << __builtin_ctzl(r);

			if (a < b)
				swap(a, b);
			a -= b;
		}
	}

	unsigned long gcd2(unsigned long a, unsigned long b)
	{
		unsigned long r = a | b;

		if (!a || !b)
			return r;

		r &= -r;

		while (!(b & r))
			b >>= 1;

		for (;;) {
			while (!(a & r))
				a >>= 1;
			if (a == b)
				return a;

			if (a < b)
				swap(a, b);
			a -= b;
			a >>= 1;
			if (a & r)
				a += b;
			a >>= 1;
		}
	}

	unsigned long gcd3(unsigned long a, unsigned long b)
	{
		unsigned long r = a | b;

		if (!a || !b)
			return r;

		b >>= __builtin_ctzl(b);
		if (b == 1)
			return r & -r;

		for (;;) {
			a >>= __builtin_ctzl(a);
			if (a == 1)
				return r & -r;
			if (a == b)
				return a << __builtin_ctzl(r);

			if (a < b)
				swap(a, b);
			a -= b;
		}
	}

	unsigned long gcd4(unsigned long a, unsigned long b)
	{
		unsigned long r = a | b;

		if (!a || !b)
			return r;

		r &= -r;

		while (!(b & r))
			b >>= 1;
		if (b == r)
			return r;

		for (;;) {
			while (!(a & r))
				a >>= 1;
			if (a == r)
				return r;
			if (a == b)
				return a;

			if (a < b)
				swap(a, b);
			a -= b;
			a >>= 1;
			if (a & r)
				a += b;
			a >>= 1;
		}
	}

	static unsigned long (*gcd_func[])(unsigned long a, unsigned long b) = {
		gcd0, gcd1, gcd2, gcd3, gcd4,
	};

	#define TEST_ENTRIES (sizeof(gcd_func) / sizeof(gcd_func[0]))

	#if defined(__x86_64__)

	#define rdtscll(val) do { \
		unsigned long __a,__d; \
		__asm__ __volatile__("rdtsc" : "=a" (__a), "=d" (__d)); \
		(val) = ((unsigned long long)__a) | (((unsigned long long)__d)<<32); \
	} while(0)

	static unsigned long long benchmark_gcd_func(unsigned long (*gcd)(unsigned long, unsigned long),
								unsigned long a, unsigned long b, unsigned long *res)
	{
		unsigned long long start, end;
		unsigned long long ret;
		unsigned long gcd_res;

		rdtscll(start);
		gcd_res = gcd(a, b);
		rdtscll(end);

		if (end >= start)
			ret = end - start;
		else
			ret = ~0ULL - start + 1 + end;

		*res = gcd_res;
		return ret;
	}

	#else

	static inline struct timespec read_time(void)
	{
		struct timespec time;
		clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &time);
		return time;
	}

	static inline unsigned long long diff_time(struct timespec start, struct timespec end)
	{
		struct timespec temp;

		if ((end.tv_nsec - start.tv_nsec) < 0) {
			temp.tv_sec = end.tv_sec - start.tv_sec - 1;
			temp.tv_nsec = 1000000000ULL + end.tv_nsec - start.tv_nsec;
		} else {
			temp.tv_sec = end.tv_sec - start.tv_sec;
			temp.tv_nsec = end.tv_nsec - start.tv_nsec;
		}

		return temp.tv_sec * 1000000000ULL + temp.tv_nsec;
	}

	static unsigned long long benchmark_gcd_func(unsigned long (*gcd)(unsigned long, unsigned long),
								unsigned long a, unsigned long b, unsigned long *res)
	{
		struct timespec start, end;
		unsigned long gcd_res;

		start = read_time();
		gcd_res = gcd(a, b);
		end = read_time();

		*res = gcd_res;
		return diff_time(start, end);
	}

	#endif

	static inline unsigned long get_rand()
	{
		if (sizeof(long) == 8)
			return (unsigned long)rand() << 32 | rand();
		else
			return rand();
	}

	int main(int argc, char **argv)
	{
		unsigned int seed = time(0);
		int loops = 100;
		int repeats = 1000;
		unsigned long (*res)[TEST_ENTRIES];
		unsigned long long elapsed[TEST_ENTRIES];
		int i, j, k;

		for (;;) {
			int opt = getopt(argc, argv, "n:r:s:");
			/* End condition always first */
			if (opt == -1)
				break;

			switch (opt) {
			case 'n':
				loops = atoi(optarg);
				break;
			case 'r':
				repeats = atoi(optarg);
				break;
			case 's':
				seed = strtoul(optarg, NULL, 10);
				break;
			default:
				/* You won't actually get here. */
				break;
			}
		}

		res = malloc(sizeof(unsigned long) * TEST_ENTRIES * loops);
		memset(elapsed, 0, sizeof(elapsed));

		srand(seed);
		for (j = 0; j < loops; j++) {
			unsigned long a = get_rand();
			/* Do we have args? */
			unsigned long b = argc > optind ? strtoul(argv[optind], NULL, 10) : get_rand();
			unsigned long long min_elapsed[TEST_ENTRIES];
			for (k = 0; k < repeats; k++) {
				for (i = 0; i < TEST_ENTRIES; i++) {
					unsigned long long tmp = benchmark_gcd_func(gcd_func[i], a, b, &res[j][i]);
					if (k == 0 || min_elapsed[i] > tmp)
						min_elapsed[i] = tmp;
				}
			}
			for (i = 0; i < TEST_ENTRIES; i++)
				elapsed[i] += min_elapsed[i];
		}

		for (i = 0; i < TEST_ENTRIES; i++)
			printf("gcd%d: elapsed %llu\n", i, elapsed[i]);

		k = 0;
		srand(seed);
		for (j = 0; j < loops; j++) {
			unsigned long a = get_rand();
			unsigned long b = argc > optind ? strtoul(argv[optind], NULL, 10) : get_rand();
			for (i = 1; i < TEST_ENTRIES; i++) {
				if (res[j][i] != res[j][0])
					break;
			}
			if (i < TEST_ENTRIES) {
				if (k == 0) {
					k = 1;
					fprintf(stderr, "Error:\n");
				}
				fprintf(stderr, "gcd(%lu, %lu): ", a, b);
				for (i = 0; i < TEST_ENTRIES; i++)
					fprintf(stderr, "%ld%s", res[j][i], i < TEST_ENTRIES - 1 ? ", " : "\n");
			}
		}

		if (k == 0)
			fprintf(stderr, "PASS\n");

		free(res);

		return 0;
	}

Compiled with "-O2", on "VirtualBox 4.4.0-22-generic #38-Ubuntu x86_64" got:

  zhaoxiuzeng@zhaoxiuzeng-VirtualBox:~/develop$ ./gcd -r 500000 -n 10
  gcd0: elapsed 10174
  gcd1: elapsed 2120
  gcd2: elapsed 2902
  gcd3: elapsed 2039
  gcd4: elapsed 2812
  PASS
  zhaoxiuzeng@zhaoxiuzeng-VirtualBox:~/develop$ ./gcd -r 500000 -n 10
  gcd0: elapsed 9309
  gcd1: elapsed 2280
  gcd2: elapsed 2822
  gcd3: elapsed 2217
  gcd4: elapsed 2710
  PASS
  zhaoxiuzeng@zhaoxiuzeng-VirtualBox:~/develop$ ./gcd -r 500000 -n 10
  gcd0: elapsed 9589
  gcd1: elapsed 2098
  gcd2: elapsed 2815
  gcd3: elapsed 2030
  gcd4: elapsed 2718
  PASS
  zhaoxiuzeng@zhaoxiuzeng-VirtualBox:~/develop$ ./gcd -r 500000 -n 10
  gcd0: elapsed 9914
  gcd1: elapsed 2309
  gcd2: elapsed 2779
  gcd3: elapsed 2228
  gcd4: elapsed 2709
  PASS

[akpm@linux-foundation.org: avoid #defining a CONFIG_ variable]
Signed-off-by: NZhaoxiu Zeng <zhaoxiu.zeng@gmail.com>
Signed-off-by: NGeorge Spelvin <linux@horizon.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      fff7fb0b
  53. 09 5月, 2016 4 次提交
  55. 05 5月, 2016 1 次提交
    • V
      ARC: support HIGHMEM even without PAE40 · 26f9d5fd
      由 Vineet Gupta 提交于 
      Initial HIGHMEM support on ARC was introduced for PAE40 where the low
memory (0x8000_0000 based) and high memory (0x1_0000_0000) were
physically contiguous. So CONFIG_FLATMEM sufficed (despite a peipheral
hole in the middle, which wasted a bit of struct page memory, but things
worked).

However w/o PAE, highmem was not possible and we could only reach
~1.75GB of DDR. Now there is a use case to access ~4GB of DDR w/o PAE40
The idea is to have low memory at canonical 0x8000_0000 and highmem
at 0 so enire 4GB address space is available for physical addressing
This needs additional platform/interconnect mapping to convert
the non contiguous physical addresses into linear bus adresses.

From Linux point of view, non contiguous divide means FLATMEM no
longer works and DISCONTIGMEM is needed to track the pfns in the 2
regions.

This scheme would also work for PAE40, only better in that we don't
waste struct page memory for the peripheral hole.

The DT description will be something like

    memory {
        ...
        reg = <0x80000000 0x200000000   /* 512MB: lowmem */
               0x00000000 0x10000000>;  /* 256MB: highmem */
   }
Signed-off-by: NNoam Camus <noamc@ezchip.com>
Signed-off-by: NVineet Gupta <vgupta@synopsys.com>
      26f9d5fd
  57. 27 4月, 2016 2 次提交