Fix degraded parity writes for EC N+K when K>1

Fixes possible `calc_rmw_parity_ec(): Assertion `bufs[i][curbuf[i]].buf' failed` error
rm-left-on-dead
Vitaliy Filippov 2022-11-20 00:39:59 +03:00
parent 6e6f6ecbb0
commit 09a8864686
2 changed files with 119 additions and 32 deletions

View File

@ -478,15 +478,18 @@ void* calc_rmw(void *request_buf, osd_rmw_stripe_t *stripes, uint64_t *read_osd_
{
if (write_osd_set[role] != 0)
{
write_parity = 1;
write_parity++;
if (write_osd_set[role] != read_osd_set[role])
{
start = 0;
end = chunk_size;
for (int r2 = pg_minsize; r2 < role; r2++)
{
stripes[r2].write_start = start;
stripes[r2].write_end = end;
if (write_osd_set[r2] != 0)
{
stripes[r2].write_start = start;
stripes[r2].write_end = end;
}
}
}
stripes[role].write_start = start;
@ -555,7 +558,7 @@ void* calc_rmw(void *request_buf, osd_rmw_stripe_t *stripes, uint64_t *read_osd_
}
}
// Allocate read buffers
void *rmw_buf = alloc_read_buffer(stripes, pg_size, (write_parity ? pg_size-pg_minsize : 0) * (end - start));
void *rmw_buf = alloc_read_buffer(stripes, pg_size, write_parity * (end - start));
// Position write buffers
uint64_t buf_pos = 0, in_pos = 0;
for (int role = 0; role < pg_size; role++)
@ -804,13 +807,11 @@ void calc_rmw_parity_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize,
calc_rmw_parity_copy_mod(stripes, pg_size, pg_minsize, read_osd_set, write_osd_set, chunk_size, bitmap_granularity, start, end);
if (end != 0)
{
int i;
for (i = pg_minsize; i < pg_size; i++)
{
int write_parity = 0;
for (int i = pg_minsize; i < pg_size; i++)
if (write_osd_set[i] != 0)
break;
}
if (i < pg_size)
write_parity++;
if (write_parity > 0)
{
// Calculate new coding chunks
buf_len_t bufs[pg_size][3];
@ -830,8 +831,11 @@ void calc_rmw_parity_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize,
}
for (int i = pg_minsize; i < pg_size; i++)
{
bufs[i][nbuf[i]++] = { .buf = stripes[i].write_buf, .len = end-start };
positions[i] = start;
if (write_osd_set[i] != 0)
{
bufs[i][nbuf[i]++] = { .buf = stripes[i].write_buf, .len = end-start };
positions[i] = start;
}
}
uint32_t pos = start;
while (pos < end)
@ -839,31 +843,37 @@ void calc_rmw_parity_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize,
uint32_t next_end = end;
for (int i = 0; i < pg_size; i++)
{
assert(curbuf[i] < nbuf[i]);
assert(bufs[i][curbuf[i]].buf);
data_ptrs[i] = (uint8_t*)bufs[i][curbuf[i]].buf + pos-positions[i];
uint32_t this_end = bufs[i][curbuf[i]].len + positions[i];
if (next_end > this_end)
next_end = this_end;
if (i < pg_minsize || write_osd_set[i] != 0)
{
assert(curbuf[i] < nbuf[i]);
assert(bufs[i][curbuf[i]].buf);
data_ptrs[i] = (uint8_t*)bufs[i][curbuf[i]].buf + pos-positions[i];
uint32_t this_end = bufs[i][curbuf[i]].len + positions[i];
if (next_end > this_end)
next_end = this_end;
}
}
assert(next_end > pos);
for (int i = 0; i < pg_size; i++)
{
uint32_t this_end = bufs[i][curbuf[i]].len + positions[i];
if (next_end >= this_end)
if (i < pg_minsize || write_osd_set[i] != 0)
{
positions[i] += bufs[i][curbuf[i]].len;
curbuf[i]++;
uint32_t this_end = bufs[i][curbuf[i]].len + positions[i];
if (next_end >= this_end)
{
positions[i] += bufs[i][curbuf[i]].len;
curbuf[i]++;
}
}
}
#ifdef WITH_ISAL
ec_encode_data(
next_end-pos, pg_minsize, pg_size-pg_minsize, matrix->isal_data,
next_end-pos, pg_minsize, write_parity, matrix->isal_data,
(uint8_t**)data_ptrs, (uint8_t**)data_ptrs+pg_minsize
);
#else
jerasure_matrix_encode(
pg_minsize, pg_size-pg_minsize, OSD_JERASURE_W, matrix->je_data,
pg_minsize, write_parity, OSD_JERASURE_W, matrix->je_data,
(char**)data_ptrs, (char**)data_ptrs+pg_minsize, next_end-pos
);
#endif
@ -871,16 +881,19 @@ void calc_rmw_parity_ec(osd_rmw_stripe_t *stripes, int pg_size, int pg_minsize,
}
for (int i = 0; i < pg_size; i++)
{
data_ptrs[i] = stripes[i].bmp_buf;
if (i < pg_minsize || write_osd_set[i] != 0)
{
data_ptrs[i] = stripes[i].bmp_buf;
}
}
#ifdef WITH_ISAL
ec_encode_data(
bitmap_size, pg_minsize, pg_size-pg_minsize, matrix->isal_data,
bitmap_size, pg_minsize, write_parity, matrix->isal_data,
(uint8_t**)data_ptrs, (uint8_t**)data_ptrs+pg_minsize
);
#else
jerasure_matrix_encode(
pg_minsize, pg_size-pg_minsize, OSD_JERASURE_W, matrix->je_data,
pg_minsize, write_parity, OSD_JERASURE_W, matrix->je_data,
(char**)data_ptrs, (char**)data_ptrs+pg_minsize, bitmap_size
);
#endif

View File

@ -20,6 +20,7 @@ void test11();
void test12();
void test13();
void test14();
void test15();
int main(int narg, char *args[])
{
@ -47,6 +48,8 @@ int main(int narg, char *args[])
test13();
// Test 14
test14();
// Test 15
test15();
// End
printf("all ok\n");
return 0;
@ -706,7 +709,7 @@ void test13()
/***
13. basic jerasure 2+1 test
14. basic jerasure 2+1 test
calc_rmw(offset=128K-4K, len=8K, osd_set=[1,2,0], write_set=[1,2,3])
= {
read: [ [ 0, 128K ], [ 0, 128K ], [ 0, 0 ] ],
@ -727,13 +730,13 @@ void test14()
osd_num_t write_osd_set[3] = { 1, 2, 3 };
osd_rmw_stripe_t stripes[3] = {};
unsigned bitmaps[3] = { 0 };
// Test 13.0
// Test 14.0
void *write_buf = malloc_or_die(8192);
split_stripes(2, 128*1024, 128*1024-4096, 8192, stripes);
assert(stripes[0].req_start == 128*1024-4096 && stripes[0].req_end == 128*1024);
assert(stripes[1].req_start == 0 && stripes[1].req_end == 4096);
assert(stripes[2].req_start == 0 && stripes[2].req_end == 0);
// Test 13.1
// Test 14.1
void *rmw_buf = calc_rmw(write_buf, stripes, osd_set, 3, 2, 3, write_osd_set, 128*1024, bmp);
for (int i = 0; i < 3; i++)
stripes[i].bmp_buf = bitmaps+i;
@ -750,7 +753,7 @@ void test14()
assert(stripes[0].write_buf == write_buf);
assert(stripes[1].write_buf == (uint8_t*)write_buf+4096);
assert(stripes[2].write_buf == rmw_buf);
// Test 13.2 - encode
// Test 14.2 - encode
set_pattern(write_buf, 8192, PATTERN3);
set_pattern(stripes[0].read_buf, 128*1024-4096, PATTERN1);
set_pattern(stripes[1].read_buf, 128*1024-4096, PATTERN2);
@ -767,7 +770,7 @@ void test14()
assert(stripes[0].write_buf == write_buf);
assert(stripes[1].write_buf == (uint8_t*)write_buf+4096);
assert(stripes[2].write_buf == rmw_buf);
// Test 13.3 - decode and verify
// Test 14.3 - decode and verify
osd_num_t read_osd_set[4] = { 0, 2, 3 };
memset(stripes, 0, sizeof(stripes));
split_stripes(2, 128*1024, 0, 128*1024, stripes);
@ -802,3 +805,74 @@ void test14()
free(write_buf);
use_ec(3, 2, false);
}
/***
15. EC 2+2 partial overwrite with 1 missing stripe
calc_rmw(offset=64K+28K, len=4K, osd_set=[1,2,3,0], write_set=[1,2,3,0])
= {
read: [ [ 28K, 32K ], [ 0, 0 ], [ 0, 0 ], [ 0, 0 ] ],
write: [ [ 0, 0 ], [ 28K, 32K ], [ 28K, 32K ], [ 0, 0 ] ],
input buffer: [ write1 ],
rmw buffer: [ write2, read0 ],
}
***/
void test15()
{
const int bmp = 64*1024 / 4096 / 8;
use_ec(4, 2, true);
osd_num_t osd_set[4] = { 1, 2, 3, 0 };
osd_num_t write_osd_set[4] = { 1, 2, 3, 0 };
osd_rmw_stripe_t stripes[4] = {};
unsigned bitmaps[4] = { 0 };
// Test 15.0
void *write_buf = malloc_or_die(4096);
split_stripes(2, 64*1024, (64+28)*1024, 4096, stripes);
assert(stripes[0].req_start == 0 && stripes[0].req_end == 0);
assert(stripes[1].req_start == 28*1024 && stripes[1].req_end == 32*1024);
assert(stripes[2].req_start == 0 && stripes[2].req_end == 0);
assert(stripes[3].req_start == 0 && stripes[3].req_end == 0);
// Test 15.1
void *rmw_buf = calc_rmw(write_buf, stripes, osd_set, 4, 2, 3, write_osd_set, 64*1024, bmp);
for (int i = 0; i < 4; i++)
stripes[i].bmp_buf = bitmaps+i;
assert(rmw_buf);
assert(stripes[0].read_start == 28*1024 && stripes[0].read_end == 32*1024);
assert(stripes[1].read_start == 0 && stripes[1].read_end == 0);
assert(stripes[2].read_start == 0 && stripes[2].read_end == 0);
assert(stripes[3].read_start == 0 && stripes[3].read_end == 0);
assert(stripes[0].write_start == 0 && stripes[0].write_end == 0);
assert(stripes[1].write_start == 28*1024 && stripes[1].write_end == 32*1024);
assert(stripes[2].write_start == 28*1024 && stripes[2].write_end == 32*1024);
assert(stripes[3].write_start == 0 && stripes[3].write_end == 0);
assert(stripes[0].read_buf == (uint8_t*)rmw_buf+4*1024);
assert(stripes[1].read_buf == NULL);
assert(stripes[2].read_buf == NULL);
assert(stripes[3].read_buf == NULL);
assert(stripes[0].write_buf == NULL);
assert(stripes[1].write_buf == (uint8_t*)write_buf);
assert(stripes[2].write_buf == rmw_buf);
assert(stripes[3].write_buf == NULL);
// Test 15.2 - encode
set_pattern(write_buf, 4*1024, PATTERN1);
set_pattern(stripes[0].read_buf, 4*1024, PATTERN2);
memset(stripes[0].bmp_buf, 0, bmp);
memset(stripes[1].bmp_buf, 0, bmp);
calc_rmw_parity_ec(stripes, 4, 2, osd_set, write_osd_set, 64*1024, bmp);
assert(*(uint32_t*)stripes[2].bmp_buf == 0x80);
assert(stripes[0].write_start == 0 && stripes[0].write_end == 0);
assert(stripes[1].write_start == 28*1024 && stripes[1].write_end == 32*1024);
assert(stripes[2].write_start == 28*1024 && stripes[2].write_end == 32*1024);
assert(stripes[3].write_start == 0 && stripes[3].write_end == 0);
assert(stripes[0].write_buf == NULL);
assert(stripes[1].write_buf == (uint8_t*)write_buf);
assert(stripes[2].write_buf == rmw_buf);
assert(stripes[3].write_buf == NULL);
check_pattern(stripes[2].write_buf, 4*1024, PATTERN1^PATTERN2); // first parity is always xor :)
// Done
free(rmw_buf);
free(write_buf);
use_ec(3, 2, false);
}