| Message ID | 1433791592.2291.83.camel@HansenPartnership.com |
|---|---|
| State | Rejected |
| Headers | show |
On 8 June 2015 at 21:26, James Bottomley <James.Bottomley@hansenpartnership.com> wrote: > On Mon, 2015-06-08 at 13:27 +0200, Ard Biesheuvel wrote: >> The rtc-efi driver declares that the EFI 'epoch' is 1/1/1998, but >> the UEFI spec does not define it at all. It does define a minimum >> of 1900 for the 'Year' member of the EFI_TIME struct, so let's use >> 1900 as the minimum year and not 1998. >> >> This prevents rtc_read_time() failures when the RTC that backs the >> EFI time services is set to a date before 1998. >> >> Cc: Alessandro Zummo <a.zummo@towertech.it> >> Cc: Alexandre Belloni <alexandre.belloni@free-electrons.com> >> Cc: rtc-linux@googlegroups.com >> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> >> --- >> drivers/rtc/rtc-efi.c | 13 +++++++------ >> 1 file changed, 7 insertions(+), 6 deletions(-) >> >> diff --git a/drivers/rtc/rtc-efi.c b/drivers/rtc/rtc-efi.c >> index cb989cd00b14..4a93b0bbf22c 100644 >> --- a/drivers/rtc/rtc-efi.c >> +++ b/drivers/rtc/rtc-efi.c >> @@ -25,9 +25,12 @@ >> >> #define EFI_ISDST (EFI_TIME_ADJUST_DAYLIGHT|EFI_TIME_IN_DAYLIGHT) >> /* >> - * EFI Epoch is 1/1/1998 >> + * The UEFI spec does not literally define an epoch, but it does >> + * define EFI_TIME::Year as having a value between 1900 and 9999. >> + * (UEFI spec v2.5 paragraph 7.3) >> + * So let's use 1900 as the EFI epoch year. >> */ >> -#define EFI_RTC_EPOCH 1998 >> +#define EFI_RTC_EPOCH 1900 >> >> /* >> * returns day of the year [0-365] >> @@ -40,8 +43,6 @@ compute_yday(efi_time_t *eft) >> } >> /* >> * returns day of the week [0-6] 0=Sunday >> - * >> - * Don't try to provide a year that's before 1998, please ! >> */ >> static int >> compute_wday(efi_time_t *eft) >> @@ -60,9 +61,9 @@ compute_wday(efi_time_t *eft) >> ndays += compute_yday(eft); >> >> /* >> - * 4=1/1/1998 was a Thursday >> + * 1=1/1/1900 was a Monday >> */ >> - return (ndays + 4) % 7; >> + return (ndays + 1) % 7; >> } > > This is a bit nuts: increasing the for loop by a huge amount every time > we call the date just so we can cope with the odd early date. Why not > just count backwards for the exception case? That way the epoch becomes > an optimisation point, every year actually works, but years far before > the epoch get penalised in the loop, while our current "efficiency" > remains ... perhaps plus points if we want to move the epoch to 2015? > The 'epoch' is defined by the UEFI spec as starting at 1900, and changing it to 2015 to optimize this for loop is equally silly imo. Note that the core purpose of the patch is to change the cutoff point below which years are rejected, but that context does not appear in the diff. rtc_time64_to_tm() already has a better way of counting the leap days between two years. I'd be happy to propose a followup patch that gets rid of the for loop entirely. But those are two separate issues imo.
On Mon, 2015-06-08 at 22:15 +0200, Ard Biesheuvel wrote: > On 8 June 2015 at 21:26, James Bottomley > <James.Bottomley@hansenpartnership.com> wrote: > > On Mon, 2015-06-08 at 13:27 +0200, Ard Biesheuvel wrote: > >> The rtc-efi driver declares that the EFI 'epoch' is 1/1/1998, but > >> the UEFI spec does not define it at all. It does define a minimum > >> of 1900 for the 'Year' member of the EFI_TIME struct, so let's use > >> 1900 as the minimum year and not 1998. > >> > >> This prevents rtc_read_time() failures when the RTC that backs the > >> EFI time services is set to a date before 1998. > >> > >> Cc: Alessandro Zummo <a.zummo@towertech.it> > >> Cc: Alexandre Belloni <alexandre.belloni@free-electrons.com> > >> Cc: rtc-linux@googlegroups.com > >> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> > >> --- > >> drivers/rtc/rtc-efi.c | 13 +++++++------ > >> 1 file changed, 7 insertions(+), 6 deletions(-) > >> > >> diff --git a/drivers/rtc/rtc-efi.c b/drivers/rtc/rtc-efi.c > >> index cb989cd00b14..4a93b0bbf22c 100644 > >> --- a/drivers/rtc/rtc-efi.c > >> +++ b/drivers/rtc/rtc-efi.c > >> @@ -25,9 +25,12 @@ > >> > >> #define EFI_ISDST (EFI_TIME_ADJUST_DAYLIGHT|EFI_TIME_IN_DAYLIGHT) > >> /* > >> - * EFI Epoch is 1/1/1998 > >> + * The UEFI spec does not literally define an epoch, but it does > >> + * define EFI_TIME::Year as having a value between 1900 and 9999. > >> + * (UEFI spec v2.5 paragraph 7.3) > >> + * So let's use 1900 as the EFI epoch year. > >> */ > >> -#define EFI_RTC_EPOCH 1998 > >> +#define EFI_RTC_EPOCH 1900 > >> > >> /* > >> * returns day of the year [0-365] > >> @@ -40,8 +43,6 @@ compute_yday(efi_time_t *eft) > >> } > >> /* > >> * returns day of the week [0-6] 0=Sunday > >> - * > >> - * Don't try to provide a year that's before 1998, please ! > >> */ > >> static int > >> compute_wday(efi_time_t *eft) > >> @@ -60,9 +61,9 @@ compute_wday(efi_time_t *eft) > >> ndays += compute_yday(eft); > >> > >> /* > >> - * 4=1/1/1998 was a Thursday > >> + * 1=1/1/1900 was a Monday > >> */ > >> - return (ndays + 4) % 7; > >> + return (ndays + 1) % 7; > >> } > > > > This is a bit nuts: increasing the for loop by a huge amount every time > > we call the date just so we can cope with the odd early date. Why not > > just count backwards for the exception case? That way the epoch becomes > > an optimisation point, every year actually works, but years far before > > the epoch get penalised in the loop, while our current "efficiency" > > remains ... perhaps plus points if we want to move the epoch to 2015? > > > > The 'epoch' is defined by the UEFI spec as starting at 1900, and > changing it to 2015 to optimize this for loop is equally silly imo. The spec doesn't define an epoch in the sense we usually mean it, it just says the year in EFI_TIME should be between 1900 and 9999. So I don't think EFI_RTC_EPOCH in the driver is really related. > Note that the core purpose of the patch is to change the cutoff point > below which years are rejected, but that context does not appear in > the diff. Both patches achieve that ... I just think doing a for loop from 1900 is very inefficient. > rtc_time64_to_tm() already has a better way of counting the leap days > between two years. I'd be happy to propose a followup patch that gets > rid of the for loop entirely. > But those are two separate issues imo. Getting rid of the loop is better ... but creating a problem and then fixing it is still not really good operating procedure. Just propose the final fix and we can apply that. James
On 8 June 2015 at 22:27, James Bottomley <James.Bottomley@hansenpartnership.com> wrote: > On Mon, 2015-06-08 at 22:15 +0200, Ard Biesheuvel wrote: >> On 8 June 2015 at 21:26, James Bottomley >> <James.Bottomley@hansenpartnership.com> wrote: >> > On Mon, 2015-06-08 at 13:27 +0200, Ard Biesheuvel wrote: >> >> The rtc-efi driver declares that the EFI 'epoch' is 1/1/1998, but >> >> the UEFI spec does not define it at all. It does define a minimum >> >> of 1900 for the 'Year' member of the EFI_TIME struct, so let's use >> >> 1900 as the minimum year and not 1998. >> >> >> >> This prevents rtc_read_time() failures when the RTC that backs the >> >> EFI time services is set to a date before 1998. >> >> >> >> Cc: Alessandro Zummo <a.zummo@towertech.it> >> >> Cc: Alexandre Belloni <alexandre.belloni@free-electrons.com> >> >> Cc: rtc-linux@googlegroups.com >> >> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org> >> >> --- >> >> drivers/rtc/rtc-efi.c | 13 +++++++------ >> >> 1 file changed, 7 insertions(+), 6 deletions(-) >> >> >> >> diff --git a/drivers/rtc/rtc-efi.c b/drivers/rtc/rtc-efi.c >> >> index cb989cd00b14..4a93b0bbf22c 100644 >> >> --- a/drivers/rtc/rtc-efi.c >> >> +++ b/drivers/rtc/rtc-efi.c >> >> @@ -25,9 +25,12 @@ >> >> >> >> #define EFI_ISDST (EFI_TIME_ADJUST_DAYLIGHT|EFI_TIME_IN_DAYLIGHT) >> >> /* >> >> - * EFI Epoch is 1/1/1998 >> >> + * The UEFI spec does not literally define an epoch, but it does >> >> + * define EFI_TIME::Year as having a value between 1900 and 9999. >> >> + * (UEFI spec v2.5 paragraph 7.3) >> >> + * So let's use 1900 as the EFI epoch year. >> >> */ >> >> -#define EFI_RTC_EPOCH 1998 >> >> +#define EFI_RTC_EPOCH 1900 >> >> >> >> /* >> >> * returns day of the year [0-365] >> >> @@ -40,8 +43,6 @@ compute_yday(efi_time_t *eft) >> >> } >> >> /* >> >> * returns day of the week [0-6] 0=Sunday >> >> - * >> >> - * Don't try to provide a year that's before 1998, please ! >> >> */ >> >> static int >> >> compute_wday(efi_time_t *eft) >> >> @@ -60,9 +61,9 @@ compute_wday(efi_time_t *eft) >> >> ndays += compute_yday(eft); >> >> >> >> /* >> >> - * 4=1/1/1998 was a Thursday >> >> + * 1=1/1/1900 was a Monday >> >> */ >> >> - return (ndays + 4) % 7; >> >> + return (ndays + 1) % 7; >> >> } >> > >> > This is a bit nuts: increasing the for loop by a huge amount every time >> > we call the date just so we can cope with the odd early date. Why not >> > just count backwards for the exception case? That way the epoch becomes >> > an optimisation point, every year actually works, but years far before >> > the epoch get penalised in the loop, while our current "efficiency" >> > remains ... perhaps plus points if we want to move the epoch to 2015? >> > >> >> The 'epoch' is defined by the UEFI spec as starting at 1900, and >> changing it to 2015 to optimize this for loop is equally silly imo. > > The spec doesn't define an epoch in the sense we usually mean it, it > just says the year in EFI_TIME should be between 1900 and 9999. So I > don't think EFI_RTC_EPOCH in the driver is really related. > OK. >> Note that the core purpose of the patch is to change the cutoff point >> below which years are rejected, but that context does not appear in >> the diff. > > Both patches achieve that ... I just think doing a for loop from 1900 is > very inefficient. > Yours allows years < 1900. I know this is all of little importance, just pointing it out since retaining the test but changing the limit was the point of the whole patch. Allowing any year instead is also fine by me, but I do think it makes sense to test against the UEFI specified range. Or, since this is Linux, we could settle on 1970 as well. Mind you, I am just trying to get rid of the errors when booting on a machine whose RTC is uninitialized, e.g., when it has lost power, so anything that achieves that is acceptable to me. >> rtc_time64_to_tm() already has a better way of counting the leap days >> between two years. I'd be happy to propose a followup patch that gets >> rid of the for loop entirely. >> But those are two separate issues imo. > > Getting rid of the loop is better ... but creating a problem and then > fixing it is still not really good operating procedure. Just propose > the final fix and we can apply that. > Yep, makes sense. I will spin a v2 that gets rid of the 'epoch' mention and the for loop.
diff --git a/drivers/rtc/rtc-efi.c b/drivers/rtc/rtc-efi.c index cb989cd..73f1ba6 100644 --- a/drivers/rtc/rtc-efi.c +++ b/drivers/rtc/rtc-efi.c @@ -40,8 +40,6 @@ compute_yday(efi_time_t *eft) } /* * returns day of the week [0-6] 0=Sunday - * - * Don't try to provide a year that's before 1998, please ! */ static int compute_wday(efi_time_t *eft) @@ -49,16 +47,25 @@ compute_wday(efi_time_t *eft) int y; int ndays = 0; - if (eft->year < EFI_RTC_EPOCH) { - pr_err("EFI year < " __stringify(EFI_RTC_EPOCH) ", invalid date\n"); - return -1; + if (unlikely(eft->year < EFI_RTC_EPOCH)) { + /* unlikely event, just count backwards */ + for (y = eft->year; y < EFI_RTC_EPOCH; y++) + ndays -= 365 + (is_leap_year(y) ? 1 : 0); + } else { + for (y = EFI_RTC_EPOCH; y < eft->year; y++) + ndays += 365 + (is_leap_year(y) ? 1 : 0); } - for (y = EFI_RTC_EPOCH; y < eft->year; y++) - ndays += 365 + (is_leap_year(y) ? 1 : 0); - ndays += compute_yday(eft); + if (ndays < 0) + /* + * make ndays positive by adding a multiple of seven because + * clock arithmetic using % doesn't work with negative + * numbers + */ + ndays += ((-ndays)/7 + 1)*7; + /* * 4=1/1/1998 was a Thursday */