RE: Patch avoid call strlen repeatedly in loop.
От | Ranier VF |
---|---|
Тема | RE: Patch avoid call strlen repeatedly in loop. |
Дата | |
Msg-id | MN2PR18MB2927BDD3EF56FCA70ABA4A13E37A0@MN2PR18MB2927.namprd18.prod.outlook.com обсуждение исходный текст |
Ответ на | Re: Patch avoid call strlen repeatedly in loop. (Mark Dilger <hornschnorter@gmail.com>) |
Список | pgsql-hackers |
Hi Mark, "In general, writing a string with snprintf and then calling strlen on that same string is not guaranteed to give the same lengths. You can easily construct a case where they differ: char foo[3] = {0}; int foolen; foolen = snprintf(foo, sizeof(foo), "%s", "xxxxxxxx"); printf("strlen(foo) = %u, foolen = %u, foo = '%s'\n", strlen(foo), foolen, foo); Using standard snprintf (and not pg_snprintf), I get: strlen(foo) = 2, foolen = 8, foo = 'xx'" Well, I've been using snprintf, no problem for several years now. But what you reported, I would easily solve with an assert. assert(foolen == strlen(foo)); To make sure things would stay under control. "I think the way it is coded is easier to read, and probably more robust against future changes, even if your proposed change happens to be safe today." I find it amazing that software I admire so much, such as PostgreSQL, makes extensive and heavy use of functions like strlen. Speed makes a lot of difference, for some people it is above safety. Maybe that's why PostgreSQL loses some battles against MySQL. Not using strlen is for educational purposes as well. Allowing is to encourage use! So stupid things such as: #define CheckComplicatedStuff (a, b) (strlen (a) > strlen (b)) for (;;) { if CheckComplicatedStuff (x, y) { break; } } They start to contaminate all the code. Using features like strlen, the programmer begins to create easy shortcuts, but in the end, they are very slow. Maybe that's why I have things in my code like: char sql [4096]; PQexec (cn, sql); While MySQL for example, would look like this: char sql [4096]; int sql_len; sql_len = snprintf (sql, sizeof (sql), "INSERT ..."); mysql_real_query (cn, sql, sql_len); "A bit more information about the performance problem you are encountering might make it easier to understand the motivation for this patch." My motivation? Speed. Win from MySQL, always. Anyway I'm redoing the patch with your suggestion. What about other functions that make extensive use of strlen? Thank you. Ranier Vilela ________________________________________ De: Mark Dilger <hornschnorter@gmail.com> Enviado: sábado, 9 de novembro de 2019 00:12 Para: Ranier VF; pgsql-hackers@lists.postgresql.org Assunto: Re: Patch avoid call strlen repeatedly in loop. On 11/8/19 9:41 AM, Ranier VF wrote: > --- \dll\postgresql-12.0\a\backend\libpq\auth.c Mon Sep 30 17:06:55 2019 > +++ auth.c Fri Nov 08 14:27:17 2019 > @@ -1815,6 +1815,7 @@ > char ident_user[IDENT_USERNAME_MAX + 1]; > pgsocket sock_fd = PGINVALID_SOCKET; /* for talking to Ident server */ > int rc; /* Return code from a locally called function */ > + int ident_query_len; > bool ident_return; > char remote_addr_s[NI_MAXHOST]; > char remote_port[NI_MAXSERV]; > @@ -1913,7 +1914,7 @@ > } > > /* The query we send to the Ident server */ > - snprintf(ident_query, sizeof(ident_query), "%s,%s\r\n", > + ident_query_len = snprintf(ident_query, sizeof(ident_query), "%s,%s\r\n", > remote_port, local_port); > > /* loop in case send is interrupted */ > @@ -1921,7 +1922,7 @@ > { > CHECK_FOR_INTERRUPTS(); > > - rc = send(sock_fd, ident_query, strlen(ident_query), 0); > + rc = send(sock_fd, ident_query, ident_query_len, 0); Hello Ranier, In general, writing a string with snprintf and then calling strlen on that same string is not guaranteed to give the same lengths. You can easily construct a case where they differ: char foo[3] = {0}; int foolen; foolen = snprintf(foo, sizeof(foo), "%s", "xxxxxxxx"); printf("strlen(foo) = %u, foolen = %u, foo = '%s'\n", strlen(foo), foolen, foo); Using standard snprintf (and not pg_snprintf), I get: strlen(foo) = 2, foolen = 8, foo = 'xx' Perhaps an analysis of the surrounding code would prove that in all cases this particular snprintf will return the same result that strlen(ident_query) would return, but I don't care to do the analysis. I think the way it is coded is easier to read, and probably more robust against future changes, even if your proposed change happens to be safe today. As for calling strlen(ident_query) just once, caching that result, and then looping, I don't immediately see a problem, but I also don't expect that loop to run more than one iteration except under unusual instances. Do you find that send() gets interrupted a lot? Is strlen(ident_query) taking long enough to be significant compared to how long send() takes? A bit more information about the performance problem you are encountering might make it easier to understand the motivation for this patch. -- Mark Dilger
В списке pgsql-hackers по дате отправления: