On Sat, 3 Jul 2021, Christos Zoulas wrote:
Interesting. snprintf should not matter. Something else is going on.
I think we should investigate why it fails instead of using the big
signal hammer.
Well, this only happens with xz(1) *and* _only_ if the final pipe-write
gets interrupted. Neither gzip(1) nor bzip2(1) exhibit this issue
because I think, of the trio, only xz does non-blocking I/O.
This almost feels that NetBSD-10 blocker PR #56113, except that only
an interruption of the final pipe-write causes xz(1) to misbehave.
In any case, setting _IONBF on the popen()ed stream fixes the issue
(or, at least, makes it go into remission). Patch below:
---START---
diff -urN ftp.orig/fetch.c ftp/fetch.c
--- ftp.orig/fetch.c 2020-07-11 00:29:38.000000000 +0000
+++ ftp/fetch.c 2021-07-04 22:22:19.047393346 +0000
@@ -1253,6 +1253,28 @@
}
#endif
+/**
+ * Write buffer safe from interrupts.
+ */
+static size_t
+fullwrite(FILE *fp, const void *buf, size_t size)
+{
+ size_t n, len = size;
+ const char* p = buf;
+
+ for (; len > 0; len -= n, p += n) {
+ errno = 0;
+ n = fwrite(p, 1, len, fp);
+ if (n == 0) {
+ if (errno == EINTR)
+ clearerr(fp);
+ else
+ return 0;
+ }
+ }
+ return size;
+}
+
/*
* Retrieve URL, via a proxy if necessary, using HTTP.
* If proxyenv is set, use that for the proxy, otherwise try ftp_proxy or
@@ -1526,6 +1548,7 @@
warn("Can't execute `%s'", savefile + 1);
goto cleanup_fetch_url;
}
+ setvbuf(fout, NULL, _IONBF, 0); /* for xz(1) */
closefunc = pclose;
}
}
@@ -1650,8 +1673,7 @@
}
bytes += flen;
bufrem -= flen;
- if (fwrite(xferbuf, sizeof(char), flen, fout)
- != flen) {
+ if (fullwrite(fout, xferbuf, flen) != flen) {
warn("Writing `%s'", savefile);
goto cleanup_fetch_url;
}
---END---
-RVP
