This will be fixed on next release, 2.2.26. We won't backport it to latest stable, sorry.
Aki
On 18.10.2016 15:36, Pierre Jaury wrote:
Hello,
I can confirm the issue is fixed. Do you have a policy to backport the patch at least to the latest stable?
Regards,
On 10/17/2016 11:45 PM, Aki Tuomi wrote:
>>>> On 10/16/2016 11:16 PM, Pierre Jaury wrote: >>>>> Hello, >>>>> >>>>> I am using a dict proxy for my sieve extdata plugin to access some >>>>> fields from an SQLite database (autoreply text and other >>>>> database-configured items). >>>>> >>>>> All tests are performed against version 2.2.25. >>>>> >>>>> $ dovecot --version >>>>> 2.2.25 (7be1766) >>>>> >>>>> My configuration looks like: >>>>> >>>>> dict { >>>>> sieve = sqlite:/etc/dovecot/pigeonhole-sieve.dict >>>>> } >>>>> >>>>> [...] >>>>> sieve_extdata_dict_uri = proxy::sieve >>>>> >>>>> I am able to read pretty much any attribute without any issue, except >>>>> when the value contains a special character like "\r" or "\n". By using >>>>> the doveadm dict client, I narrowed it down to the dictionary management >>>>> part (either server, protocol or client). >>>>> >>>>> I was suspecting escaping functions from "lib/strescape.c" (mostly >>>>> str_tabescape and its counterpart, used by "lib-dict/client.c"), so I >>>>> monitored socket communications. It seems that escaping is done properly >>>>> on the server and the socket is not an issue either. >>>>> >>>>> The following strace dump results from running "doveadm dict get" >>>>> against the dict socket: >>>>> >>>>> connect(8, {sa_family=AF_UNIX, sun_path="..."}, 110) = 0 >>>>> fstat(8, {st_mode=S_IFSOCK|0777, st_size=0, ...}) = 0 >>>>> [...] >>>>> write(8, "H2\t0\t0\tadmin@domain.tld\tsieve\n", 30) = 30 >>>>> [...] >>>>> read(8, "Otest\1r\1ntest\n", 8192) = 14 >>>>> >>>>> Indeed "\1r" and "\1n" are the escape sequences used by >>>>> "lib/strescape.c". I went deeped and debugged the call to "dict_lookup" >>>>> performed by doveadm. Indeed the client gets the proper string from the >>>>> socket and to my surprise, it is properly unescaped. >>>>> >>>>> Then, in "client_dict_lookup" ("lib-dict/dict-client.c"), the call to >>>>> "p_strdup" returns an empty string (null byte set at the target address). >>>>> >>>>> Before the call to the dict "->lookup" attribute (client_dict_lookup): >>>>> >>>>> RAX: 0x7ffff73a37c0 (push r14) >>>>> RBX: 0x6831b8 ("priv/reply_body") >>>>> RCX: 0x7fffffffe240 --> 0x682a60 --> 0x6831b8 ("priv/reply_body") >>>>> RDX: 0x6831b8 ("priv/reply_body") >>>>> RSI: 0x683288 --> 0x7ffff7653120 --> 0x7ffff73ea620 ([...]) >>>>> RDI: 0x690ad0 --> 0x7ffff7400713 --> 0x75250079786f7270 ('proxy') >>>>> >>>>> 0x7ffff73a1f10
: mov rcx,r11 (value_r) >>>>> 0x7ffff73a1f13 : mov rdx,r8 (key) >>>>> 0x7ffff73a1f16 : mov rsi,r10 (pool) >>>>> 0x7ffff73a1f19 : mov rdi,r9 (dict) >>>>> 0x7ffff73a1f1c : add rsp,0x8 >>>>> => 0x7ffff73a1f20 : jmp rax >>>>> >>>>> Before the call to p_strdup in "client_dict_lookup": >>>>> >>>>> RSI: 0x6832d8 ("test\r\ntest") (lookup.result.value) >>>>> RDI: 0x683288 --> 0x7ffff7653120 --> [...] (pool) >>>>> RAX: 0x0 (result) >>>>> >>>>> 0x7ffff73a384f: nop >>>>> 0x7ffff73a3850: mov rsi,QWORD PTR [rsp+0x8] >>>>> 0x7ffff73a3855: mov rdi,r14 >>>>> => 0x7ffff73a3858: call 0x7ffff736d3c0 p_strdup@plt >>>>> 0x7ffff73a385d: mov QWORD PTR [r13+0x0],rax >>>>> 0x7ffff73a3861: mov rsi,QWORD PTR [rsp+0x18] >>>>> 0x7ffff73a3866: xor rsi,QWORD PTR fs:0x28 >>>>> 0x7ffff73a386f: mov eax,ebx >>>>> >>>>> After the call: >>>>> >>>>> 0x7ffff73a3850: mov rsi,QWORD PTR [rsp+0x8] >>>>> 0x7ffff73a3855: mov rdi,r14 >>>>> 0x7ffff73a3858: call 0x7ffff736d3c0 p_strdup@plt >>>>> => 0x7ffff73a385d: mov QWORD PTR [r13+0x0],rax >>>>> 0x7ffff73a3861: mov rsi,QWORD PTR [rsp+0x18] >>>>> 0x7ffff73a3866: xor rsi,QWORD PTR fs:0x28 >>>>> 0x7ffff73a386f: mov eax,ebx >>>>> 0x7ffff73a3871: jne 0x7ffff73a38da >>>>> >>>>> RSI: 0x0 >>>>> RDI: 0x6832d8 --> 0x0 >>>>> RAX: 0x6832d8 --> 0x0 (result) >>>>> >>>>> It is worth noting that I can reproduce the exact same execution flow >>>>> with a non-multiline result string (lookup.result.value) that is >>>>> properly copied by "p_strdup" and returned in RAX, then displayed by >>>>> doveadm. >>>>> >>>>> I am not familiar with the pooling mechanism hidden behind the call to >>>>> p_strdump and not quite sure why this behaviour is emerging. Maybe I am >>>>> even miles away from an understanding of the issue here, but it sounds >>>>> to me like something is wrong in the way "p_strdup" performs the copy. >>>>> >>>>> Hope this helps, >>>>> kaiyou. >>>>> >>>>> >>>>> Fixed with https://github.com/dovecot/core/commit/4f051c3082080b9d69ef12c3720c683cff34b... Aki Tuomi