prior to version 1.1.0, the difference between pthread_self (the
public function) and __pthread_self (the internal macro or inline
function) was that the former would lazily initialize the thread
pointer if it was not already initialized, whereas the latter would
crash in this case. since lazy initialization is no longer supported,
use of pthread_self no longer makes sense; it simply generates larger,
slower code.

such kernels cannot support threads, but the thread pointer is also
important for other purposes, most notably stack protector. without a
valid thread pointer, all code compiled with stack protector will
crash. the same applies to any use of thread-local storage by
applications or libraries.

the concept of this patch is to fall back to using the modify_ldt
syscall, which has been around since linux 1.0, to setup the gs
segment register. since the kernel does not have a way to
automatically assign ldt entries, use of slot zero is hard-coded. if
this fallback path is used, __set_thread_area returns a positive value
(rather than the usual zero for success, or negative for error)
indicating to the caller that the thread pointer was successfully set,
but only for the main thread, and that thread creation will not work
properly. the code in __init_tp has been changed accordingly to record
this result for later use by pthread_create.

the results of a dns query, whether it's performed as part of one of
the standard name-resolving functions or directly by res_send, should
be a function of the query, not of the particular nameserver that
responds to it. thus, all responses which indicate a failure or
refusal by the nameserver, as opposed to a positive or negative result
for the query, should be ignored.

the strategy used is to re-issue the query immediately (but with a
limit on the number of retries, in case the server is really broken)
when a response code of 2 (server failure, typically transient) is
seen, and otherwise take no action on bad responses (which generally
indicate a misconfigured nameserver or one which the client does not
have permission to use), allowing the normal retry interval to apply
and of course accepting responses from other nameservers queried in
parallel.

empirically this matches the traditional resolver behavior for
nameservers that respond with a code of 2 in the case where there is
just a single nameserver configured. the behavior diverges when
multiple nameservers are available, since musl is querying them in
parallel. in this case we are mildly more aggressive at retrying.

the way this is implemented, it also allows explicit setting of
TZ=/etc/localtime even for suid programs. this is not a problem
because /etc/localtime is a trusted path, much like the trusted
zoneinfo search path.

reading the variadic mode argument is only valid when the O_CREAT flag
is present. this probably does not matter, but is needed for formal
correctness, and could affect LTO or other full-program analysis.

since there is no easy way to detect whether open honored or ignored
the O_CLOEXEC flag, the optimal solution to providing a fallback is
simply to make the fcntl syscall to set the close-on-exec flag
immediately after open returns.

the fcntl function is heavy, so make the syscall directly instead.
also, avoid the code size and runtime overhead of querying the old
flags, since it's reasonable to assume nothing will be set on a
newly-created socket. this code is only used on old kernels which lack
proper atomic close-on-exec support, so future changes that might
invalidate such an assumption do not need to be considered.

as usual, this is non-atomic, but better than producing an error or
failing to set the close-on-exec flag at all.

the input name is validated, the other parameters are assumed to be
valid (the list of already compressed names are not checked for
infinite reference loops or out-of-bound offsets).

names are handled case-sensitively for now.

trailing . should be accepted in domain name strings by convention
(RFC 1034), host name lookup accepts "." but rejects empty "", res_*
interfaces also accept empty name following existing practice.

this condition could only happen due to malloc failure.

the fdopen operation is also moved to take place after the unlink to
minimize the window during which a link to the file exists in the
directory table.

A domain name is at most 255 bytes long (RFC 1035), but the string
representation is two bytes smaller so the strlen maximum is 253.

Due to an error introduced in commit fcc522c9,
checking of the remaining output buffer space was not performed correctly,
allowing malformed input to write past the end of the buffer.

In addition, the loop detection logic failed to account for the possibility
of infinite loops with no output, which would hang the function.

The output size is now limited more strictly so only names with valid length
are accepted.

this also affects the legacy getservbyport family, which uses
getnameinfo as its backend.

this also affects the legacy gethostbyaddr family, which uses
getnameinfo as its backend.

some other minor changes associated with the refactoring of source
files are also made; in particular, the resolv.conf parser now uses
the same code that's used elsewhere to handle ip literals, so as a
side effect it can now accept a scope id for nameserver addressed with
link-local scope.

the service and protocol functions are defined also in other files,
and the protocol ones are actually non-nops elsewhere, so the weak
definitions in ent.c could have prevented the strong definitions from
getting pulled in and used in some static programs.

the old implementation preallocated a buffer in order to try to avoid
calling vsnprintf more than once. not only did this potentially lead
to memory fragmentation from trimming with realloc; it also pulled in
realloc/free, which otherwise might not be needed in a static linked
program.

for all address types, a scope_id specified as a decimal value is
accepted. for addresses with link-local scope, a string containing the
interface name is also accepted.

some changes are made to error handling to avoid unwanted fallbacks in
the case where the scope_id is invalid: if an earlier name lookup
backend fails with an error rather than simply "0 results", this
failure now suppresses any later attempts with other backends.

in getnameinfo, a light "itoa" type function is added for generating
decimal scope_id results, and decimal port strings for services are
also generated using this function now so as not to pull in the
dependency on snprintf.

in netdb.h, a definition for the NI_NUMERICSCOPE flag is added. this
is required by POSIX (it was previously missing) and needed to allow
callers to suppress interface-name lookups.

the return value is unsigned, so negative results for "errors" do not
make sense; 0 is the value reserved for when the interface name does
not exist.

previously, all failures to obtain at least one address were treated
as nonexistant names (EAI_NONAME). this failed to account for the
possibility of transient failures (no response at all, or a response
with rcode of 2, server failure) or permanent failures that do not
indicate the nonexistence of the requested name. only an rcode of 3
should be treated as an indication of nonexistence.

since the buffer passed always has an actual size of 512 bytes, the
maximum possible response packet size, no out-of-bounds access was
possible; however, reading past the end of the valid portion of the
packet could cause the parser to attempt to process junk as answer
content.

when wcsrtombs stopped due to hitting zero remaining space in the
output buffer, it was wrongly clearing the position pointer as if it
had completed the conversion successfully.

this commit rearranges the code somewhat to make a clear separation
between the cases of ending due to running out of output buffer space,
and ending due to reaching the end of input or an illegal sequence in
the input. the new branches have been arranged with the hope of
optimizing more common cases, too.

the old resolver code used a function __ipparse which contained the
logic for inet_addr and inet_aton, which is needed in getaddrinfo.
this was phased out in the resolver overhaul in favor of directly
using inet_aton and inet_pton as appropriate.

this commit cleans up some stuff that was left behind.

this is the third phase of the "resolver overhaul" project.

this commit removes all of the old dns code, and switches the
__lookup_name backend (used by getaddrinfo, etc.) and the getnameinfo
function to use the newly implemented __res_mkquery and __res_msend
interfaces. for parsing the results, a new callback-based __dns_parse
function, based on __dns_get_rr from the old dns code, is used.

this is the second phase of the "resolver overhaul" project.

the key additions in this commit are the __res_msend and __res_mkquery
functions, which have been factored so as to provide a backend for
both the legacy res_* functions and the standard getaddrinfo and
getnameinfo functions. the latter however are still using the old
backend code; there is code duplication which still needs to be
removed, and this will be the next phase of the resolver overhaul.

__res_msend is derived from the old __dns_doqueries function, but
generalized to send arbitrary caller-provided packets in parallel
rather than producing the parallel queries itself. this allows it to
be used (completely trivially) as a backend for res_send. the
factored-out query generation code, with slightly more generality, is
now part of __res_mkquery.

iptables and ipsec-tools among others require these to function
properly.

this bug was introduced in the recent resolver overhaul commits. it
likely had visible symptoms. these were probably limited to wrongly
accepting truncated versions of over-long names (vs rejecting them),
as opposed to stack-based overflows or anything more severe, but no
extensive checks were made. there have been no releases where this bug
was present.

now that host and service lookup have been separated in the backend,
there's no need for service lookup functions to pull in the host
lookup code. moreover, dynamic allocation is no longer needed, so this
function should now be async-signal-safe. it's also significantly
smaller.

one change in getservbyname is also made: knowing that getservbyname_r
needs only two character pointers in the caller-provided buffer, some
wasted bss can be avoided.

these changes reduce the size of the function somewhat and remove many
of its dependencies, including free. in principle it should now be
async-signal-safe, but this has not been verified in detail.

minor changes to error handling are also made.

this is the first phase of the "resolver overhaul" project.

conceptually, the results of getaddrinfo are a direct product of a
list of address results and a list of service results. the new code
makes this explicit by computing these lists separately and combining
the results. this adds support for services that have both tcp and udp
versions, where the caller has not specified which it wants, and
eliminates a number of duplicate code paths which were all producing
the final output addrinfo structures, but in subtly different ways,
making it difficult to implement any of the features which were
missing.

in addition to the above benefits, the refactoring allows for legacy
functions like gethostbyname to be implemented without using the
getaddrinfo function itself. such changes to the legacy functions have
not yet been made, however.

further improvements include matching of service alias names from
/etc/services (previously only the primary name was supported),
returning multiple results from /etc/hosts (previously only the first
matching line was honored), and support for the AI_V4MAPPED and AI_ALL
flags.

features which remain unimplemented are IDN translations (encoding
non-ASCII hostnames for DNS lookup) and the AI_ADDRCONFIG flag.

at this point, the DNS-based name resolving code is still based on the
old interfaces in __dns.c, albeit somewhat simpler in its use of them.
there may be some dead code which could already be removed, but
changes to this layer will be a later phase of the resolver overhaul.

int8_t, u_int8_t, etc types are moved under _BSD_SOURCE

CONCAT(0x1p,LDBL_MANT_DIG) is not safe outside of libc,
use 2/LDBL_EPSILON instead.

fix was proposed by Morten Welinder.

from linux/in.h and linux/in6.h uapi headers the following
missing socket options were added:

IP_NODEFRAG - used with customized ipv4 headers
IPV6_RECVPATHMTU - for ipv6 path mtu
IPV6_PATHMTU - for ipv6 path mtu
IPV6_DONTFRAG - for ipv6 path mtu
IPV6_ADDR_PREFERENCES - RFC5014 Source Address Selection
IPV6_MINHOPCOUNT - RFC5082 Generalized TTL Security Mechanism
IPV6_ORIGDSTADDR - used by tproxy
IPV6_RECVORIGDSTADDR - used by tproxy
IPV6_TRANSPARENT - used by tproxy
IPV6_UNICAST_IF - ipv6 version of IP_UNICAST_IF

and socket option values:

IP_PMTUDISC_OMIT - value for IP_MTU_DISCOVER option, new in linux 3.14
IPV6_PMTUDISC_OMIT - same for IPV6_MTU_DISCOVER
IPV6_PMTUDISC_INTERFACE - ipv6 version of IP_PMTUDISC_INTERFACE
IPV6_PREFER_* - flags for IPV6_ADDR_PREFERENCES

not added: ipv6 flow info and flow label related definitions.
(it's unclear if libc should define these and namespace polluting
type name is involved so they are not provided for now)

this was introduced to query BPF extension support with getsockopt
in linux 3.14, commit ea02f9411d9faa3553ed09ce0ec9f00ceae9885e

linux 3.14 introduced sched_getattr and sched_setattr syscalls in
commit d50dde5a10f305253cbc3855307f608f8a3c5f73

and the related SCHED_DEADLINE scheduling policy in
commit aab03e05e8f7e26f51dee792beddcb5cca9215a5

but struct sched_attr "extended scheduling parameters data structure"
is not yet exported to userspace (necessary for using the syscalls)
so related uapi definitions are not added yet.