On Thu, Feb 04, 2021 at 01:07:06PM -0600, Eric Blake wrote: > We have multiple clients of qemu_strtosz (qemu-io, the opts visitor, > the keyval visitor), and it gets annoying that edge-case testing is > impacted by implicit rounding to 53 bits of precision due to parsing > with strtod(). As an example posted by Rich Jones: > $ nbdkit memory $(( 2**63 - 2**30 )) --run \ > 'build/qemu-io -f raw "$uri" -c "w -P 3 $(( 2**63 - 2**30 - 512 )) 512" ' > write failed: Input/output error > > because 9223372035781033472 got rounded to 0x7fffffffc0000000 which is > out of bounds. > > It is also worth noting that our existing parser, by virtue of using > strtod(), accepts decimal AND hex numbers, even though test-cutils > previously lacked any coverage of the latter. We do have existing > clients that expect a hex parse to work (for example, iotest 33 using > qemu-io -c "write -P 0xa 0x200 0x400"), but strtod() parses "08" as 8 > rather than as an invalid octal number, so we know there are no > clients that depend on octal. Our use of strtod() also means that > "0x1.8k" would actually parse as 1536 (the fraction is 8/16), rather > than 1843 (if the fraction were 8/10); but as this was not covered in > the testsuite, I have no qualms forbidding hex fractions as invalid, > so this patch declares that the use of fractions is only supported > with decimal input, and enhances the testsuite to document that. > > Our previous use of strtod() meant that -1 parsed as a negative; now > that we parse with strtoull(), negative values can wrap around module > 2^64, so we have to explicitly check whether the user passed in a '-'. > > We also had no testsuite coverage of "1.1e0k", which happened to parse > under strtod() but is unlikely to occur in practice; as long as we are > making things more robust, it is easy enough to reject the use of > exponents in a strtod parse. > > The fix is done by breaking the parse into an integer prefix (no loss > in precision), rejecting negative values (since we can no longer rely > on strtod() to do that), determining if a decimal or hexadecimal parse > was intended (with the new restriction that a fractional hex parse is > not allowed), and where appropriate, using a floating point fractional > parse (where we also scan to reject use of exponents in the fraction). > The bulk of the patch is then updates to the testsuite to match our > new precision, as well as adding new cases we reject (whether they > were rejected or inadvertenly accepted before). > > Signed-off-by: Eric Blake <ebl...@redhat.com> >
> diff --git a/util/cutils.c b/util/cutils.c > index 0b5073b33012..0234763bd70b 100644 > --- a/util/cutils.c > +++ b/util/cutils.c > @@ -241,10 +241,21 @@ static int64_t suffix_mul(char suffix, int64_t unit) > } > > /* > - * Convert string to bytes, allowing either B/b for bytes, K/k for KB, > - * M/m for MB, G/g for GB or T/t for TB. End pointer will be returned > - * in *end, if not NULL. Return -ERANGE on overflow, and -EINVAL on > - * other error. > + * Convert size string to bytes. > + * > + * Allow either B/b for bytes, K/k for KB, M/m for MB, G/g for GB or > + * T/t for TB, with scaling based on @unit, and with @default_suffix > + * implied if no explicit suffix was given. > + * > + * The end pointer will be returned in *end, if not NULL. If there is > + * no fraction, the input can be decimal or hexadecimal; if there is a > + * fraction, then the input must be decimal and there must be a suffix > + * (possibly by @default_suffix) larger than Byte, and the fractional > + * portion may suffer from precision loss or rounding. The input must > + * be positive. Even though the test suite gives some illustrations, I think we should document here the patterns we're intending to support. IIUC, we aim for [quote] The size parsing supports the following syntaxes - 12345 - decimal, bytes - 12345{bBkKmMgGtT} - decimal, scaled bytes - 12345.678 - fractional decimal, bytes - 12345.678{bBkKmMgGtT} - fractional decimal, scaled bytes - 0x7FEE - hex, bytes The following are intentionally not supported - octal - fractional hex - floating point exponents [/quote] > + * > + * Return -ERANGE on overflow (with *@end advanced), and -EINVAL on > + * other error (with *@end left unchanged). > */ Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|