twoskip does a full munmap/mmap for every transaction in foreach

[brong]

Discovered by Variomedia AG when they upgraded to 2.5.8 from 2.4 series and used twoskip as the default mailboxes.db.

You can trivially see this on Linux as well with an strace:

brong@wot:/src/cyrus-imapd$ /usr/cyrus/bin/cyr_dbtool -n /tmp/skiplist.db skiplist set a a
brong@wot:/src/cyrus-imapd$ /usr/cyrus/bin/cyr_dbtool -n /tmp/skiplist.db skiplist set b b
brong@wot:/src/cyrus-imapd$ /usr/cyrus/bin/cyr_dbtool -n /tmp/skiplist.db skiplist set c c
brong@wot:/src/cyrus-imapd$ /usr/cyrus/bin/cyr_dbtool -n /tmp/twoskip.db twoskip set a a
brong@wot:/src/cyrus-imapd$ /usr/cyrus/bin/cyr_dbtool -n /tmp/twoskip.db twoskip set b b
brong@wot:/src/cyrus-imapd$ /usr/cyrus/bin/cyr_dbtool -n /tmp/twoskip.db twoskip set c c
brong@wot:/src/cyrus-imapd$ /usr/cyrus/bin/cyr_dbtool -n /tmp/skiplist.db skiplist show
a a
b b
c c
brong@wot:/src/cyrus-imapd$ /usr/cyrus/bin/cyr_dbtool -n /tmp/twoskip.db twoskip show
a a
b b
c c

Here's the interesting bit of the skiplist show under strace:

mmap(NULL, 16384, PROT_READ, MAP_SHARED, 5, 0) = 0x7f37c1856000
sendto(3, "<182>Jul 15 22:54:20 025212C1/cy"..., 123, MSG_NOSIGNAL, NULL, 0) = 123
fcntl(5, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fcntl(5, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(5, {st_mode=S_IFREG|0600, st_size=232, ...}) = 0
stat("/tmp/skiplist.db", {st_mode=S_IFREG|0600, st_size=232, ...}) = 0
fcntl(5, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
writev(1, [{"a", 1}, {"\t", 1}, {"a", 1}, {"\n", 1}], 4a a
) = 4
fcntl(5, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(5, {st_mode=S_IFREG|0600, st_size=232, ...}) = 0
stat("/tmp/skiplist.db", {st_mode=S_IFREG|0600, st_size=232, ...}) = 0
fcntl(5, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
writev(1, [{"b", 1}, {"\t", 1}, {"b", 1}, {"\n", 1}], 4b b
) = 4
fcntl(5, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(5, {st_mode=S_IFREG|0600, st_size=232, ...}) = 0
stat("/tmp/skiplist.db", {st_mode=S_IFREG|0600, st_size=232, ...}) = 0
fcntl(5, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
writev(1, [{"c", 1}, {"\t", 1}, {"c", 1}, {"\n", 1}], 4c c
) = 4
fcntl(5, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(5, {st_mode=S_IFREG|0600, st_size=232, ...}) = 0
stat("/tmp/skiplist.db", {st_mode=S_IFREG|0600, st_size=232, ...}) = 0
fcntl(5, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
munmap(0x7f37c1856000, 16384) = 0

And the interesting bit of the twoskip show under strace:

mmap(NULL, 16384, PROT_READ, MAP_SHARED, 4, 0) = 0x7f1389e5f000
fcntl(4, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(4, {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
stat("/tmp/twoskip.db", {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
munmap(0x7f1389e5f000, 16384) = 0
mmap(NULL, 16384, PROT_READ, MAP_SHARED, 4, 0) = 0x7f1389e5f000
fcntl(4, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fcntl(4, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(4, {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
stat("/tmp/twoskip.db", {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
munmap(0x7f1389e5f000, 16384) = 0
mmap(NULL, 16384, PROT_READ, MAP_SHARED, 4, 0) = 0x7f1389e5f000
fcntl(4, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
writev(1, [{"a", 1}, {"\t", 1}, {"a", 1}, {"\n", 1}], 4a a
) = 4
fcntl(4, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(4, {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
stat("/tmp/twoskip.db", {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
munmap(0x7f1389e5f000, 16384) = 0
mmap(NULL, 16384, PROT_READ, MAP_SHARED, 4, 0) = 0x7f1389e5f000
fcntl(4, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
writev(1, [{"b", 1}, {"\t", 1}, {"b", 1}, {"\n", 1}], 4b b
) = 4
fcntl(4, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(4, {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
stat("/tmp/twoskip.db", {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
munmap(0x7f1389e5f000, 16384) = 0
mmap(NULL, 16384, PROT_READ, MAP_SHARED, 4, 0) = 0x7f1389e5f000
fcntl(4, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
writev(1, [{"c", 1}, {"\t", 1}, {"c", 1}, {"\n", 1}], 4c c
) = 4
fcntl(4, F_SETLKW, {l_type=F_RDLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
fstat(4, {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
stat("/tmp/twoskip.db", {st_mode=S_IFREG|0600, st_size=536, ...}) = 0
munmap(0x7f1389e5f000, 16384) = 0
mmap(NULL, 16384, PROT_READ, MAP_SHARED, 4, 0) = 0x7f1389e5f000
fcntl(4, F_SETLKW, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=0, l_len=0}) = 0
close(4) = 0
munmap(0x7f1389e5f000, 16384) = 0

....

The fix will be to make lib/mappedfile.c less agressive about re-mmaping.

[brong]

Fixed in 2.5 and master