Y2038 January 2016

y2038@lists.linaro.org

12 participants
26 discussions

[PATCH RESEND] AFS: Correctly use 64-bit time for UUID

by Arnd Bergmann

UUID calculation uses 'struct timespec' whose seconds will overflow in year 2038 and beyond for 32-bit systems. This patch removes the dependency on 'struct timespec' by using ktime_get_real(). While the patch does not fix a 'bug' as such, it is part of a larger effort to remove instances of 'struct timespec' and other data-structures suffering from y2038 problem from the kernel. Suggested-by: Arnd Bergmann <arnd(a)arndb.de> Signed-off-by: Tina Ruchandani <ruchandani.tina(a)gmail.com> Signed-off-by: Arnd Bergmann <arnd(a)arndb.de> diff --git a/fs/afs/main.c b/fs/afs/main.c index 35de0c04729f..129ff432391c 100644 --- a/fs/afs/main.c +++ b/fs/afs/main.c @@ -14,6 +14,7 @@ #include <linux/init.h> #include <linux/completion.h> #include <linux/sched.h> +#include <linux/ktime.h> #include "internal.h" MODULE_DESCRIPTION("AFS Client File System"); @@ -37,7 +38,6 @@ struct workqueue_struct *afs_wq; */ static int __init afs_get_client_UUID(void) { - struct timespec ts; u64 uuidtime; u16 clockseq; int ret; @@ -48,9 +48,7 @@ static int __init afs_get_client_UUID(void) if (ret < 0) return ret; - getnstimeofday(&ts); - uuidtime = (u64) ts.tv_sec * 1000 * 1000 * 10; - uuidtime += ts.tv_nsec / 100; + uuidtime = ktime_divns(ktime_get_real(), 100); uuidtime += AFS_UUID_TO_UNIX_TIME; afs_uuid.time_low = uuidtime; afs_uuid.time_mid = uuidtime >> 32;

8 years, 10 months

You have new fax, document 000294396

by Interfax

You have received a new fax. Please check your fax document in the attachment to this e-mail. Quality: 400 DPI Processed in: 33 seconds Author: Richard Franklin Date: Wed, 20 Jan 2016 22:49:43 +0300 Filesize: 150 Kb Document name: task.000294396.doc Pages number: 10 Thanks for choosing Interfax!

8 years, 10 months

You have 1 new fax, document 0000708965

by noreply＠unverified.beget.ru

You have received a new fax. Please check your fax document in the attachment to this e-mail. File size: 163 Kb Scan quality: 100 DPI Fax name: scanned_0000708965.doc Pages scanned: 6 Scanned by: Ken Gonzalez Scanned at: Tue, 19 Jan 2016 17:05:09 +0300 Scanned in: 22 seconds Thank you for using Interfax!

8 years, 10 months

Re: [Y2038] [RFC 02/15] vfs: Change all structures to support 64 bit time

by Arnd Bergmann

On Saturday 16 January 2016 12:14:22 Andreas Dilger wrote: > >> > >> Sure, and nfs is a pain because of all it's internal use of > >> timespecs, too. > > > > lustre is probably the worst. > > Lustre currently only has one-second granularity in a 64-bit field, > so it doesn't really care about the difference between timespec or > timespec64 at all. > > The only other uses are for measuring relative times, so the 64-bitness > shouldn't really matter. > > Could you please point out what issues exist so they can be fixed. It's not really a bug that needs to be fixed, but more the general issue of referencing inode->i_?time and attr->ia_?time and passing them around. When we change the types in the inode and iattr from timespec to timespec64, all assigments need to be modified, and lustre has more of those assignments than any other file system I'm aware of. Arnd

8 years, 10 months

[RFC 00/15] Add 64 bit timestamp support

by Deepa Dinamani

This is an update to Arnd Bergmann's RFC patch series: https://lkml.org/lkml/2014/5/30/669 . The syscalls and runtime libraries will be handled in separate patch series. The filling of max and min timestamps for individual filesystems can be leveraged from the patch series: https://lkml.org/lkml/2015/11/20/413 1. Objective: To transition all file system code to use 64 bit time. This translates to using timespec64 across the fs code for any timestamp representation. 2. Problem Description: struct timespec cannot represent times after year 2038 on 32 bit systems. The alternative to timespec in the kernel world is timespec64 to get around the above problem. This will be the UAPI exposed interface. 3. Design objectives: The goal of this approach was to come up with small manageable patches to address the above problem. Preferably, a single patch per filesystem that can be merged independently. Also, a more generic approach that all the individual filesystems could follow was preferred. 4. Solution: The solution incorporated in the patch series involves stages defined below: 4.1. CONFIG_FS_USES_64BIT_TIME This new config is defined to #ifdef code that is required to support 64 bit times. 4.2. struct inode times: The struct inode saves {a,c,m}timestamps as struct timespec. This leads to 2 problems: a. The size of the structure depends on whether the machine is 32/ 64 bit. b. y2038 problem described above. struct timespec64 also has the same problem as in (a) above. Choosing scalar types to store these timestamps(time64 and s32) solves both the above problems. Adding accessors to access these timestamps would hide the internal data type so that this can be changed according to config in (4.1). 4.3. struct inode_timespec: Use inode_timespec for all other timestamp representation throughout VFS and individual filesystems. inode_timespec is aliased to timespec or timespec64 based on config defined in (4.1). Using timespec64 in this stage would require ifdef-ing the code, which is spread throughout the code. 4.4. Enable config in (4.1). 4.5. Convert inode_timespec to timespec64: Drop all references to inode_timespec in (4.3). Replace it with timespec64. 5. Alternate Solution: Steps involved are: 5.1. Change VFS code to handle both timespec64 and timespec: There are a few API's in the VFS that take timestamps as arguments: generic_update_time(), inode->i_op->update_time(), lease_get_mtime(), fstack_copy_attr_all(), setattr_copy(), generic_fillattr. The attr and kstat properties could have accessors like inode. But, the other functions will have to maintain copies or be updated simultaenously along with other affecting filesystems. 5.2.struct timespec64: Change individual fs to use timespec64. 5.3. struct inode times: The struct inode saves {a,c,m}timestamps as struct timespec. This leads to 2 problems: a. The size of the structure depends on whether the machine is 32/ 64 bit. b. y2038 problem described above. struct timespec64 also has the same problem as in (a) above. Choosing scalar types to store these timestamps(time64 and s32) solves both the above problems. Change individual filesystems to use macros to access inode times. Inode macros can assume conversion from timespec64 always. 5.4. VFS: Change vfs code also as above in (1) and (2). 5.5. Drop timespec This involves dropping support for any api/ struct changes to make vfs use only timespec64 for timestamps. 6. Rationale: The advantage of the method described in (5) is that we do not have inode_timespec aliases only to be dropped later. But, the method suffers from disadvantages: a. As mentioned in (5.1), our process is affected as all the filesystems using each api must be changed simultaenously or VFS should have a copy. b. While individual filesystems are being changed, VFS will have to have 2 copies of a few apis. This will mean that at this time, any new code being added might use either. This might lead to confusion. Misc: 7. Range check: Patches include range check and clamping of timestamps. This topic did not have a conclusion on the previous RFC. 7.1. Rationale The method incorporated in the patch series is based on following principles: a. Linux does not impose any fixed format for on-disk inodes. LKML discussion is still ongoing concerning the best handling of file systems used or updated after their expiration date. b. Linux cannot surmise the side effects to a file system because of the wrong timestamps as each fs saves timestamps differently. c. Individual filesystems must be able to say what to do when timestamps are clamped. 7.2. Solution Based on the above principles, the solution is described below: 7.2.1. There are 2 instances that the solution handles differently: a. While mounting a filesystem: A filesystem that has already exceeded the range of its timestamp fields. b. While doing operations on a mounted filesystem: Timestamps start getting clamped after the filesystem is mounted. 7.2.2. In both the above cases, a function is invoked as per the callbacks registered by filesystems. 8. Testing This is a proof of concept implementation. I want to get some feedback before I convert rest of the file systems. I've done some initial testing based on the patches below on x86 64 bit arch. Testing was mainly done on the root filesystem. mount, stat, touch, read, write system calls were used for testing timestamp clamps and other functionality. Patches 8-15 are only included to provide a complete picture. Deepa Dinamani (15): fs: add Kconfig entry CONFIG_FS_USES_64BIT_TIME vfs: Change all structures to support 64 bit time kernel: time: Add macros and functions to support 64 bit time vfs: Add support for vfs code to use 64 bit time fs: cifs: Add support for cifs to use 64 bit time fs: fat: convert fat to 64 bit time fs: ext4: convert to use 64 bit time fs: Enable 64 bit time fs: cifs: replace inode_timespec with timespec64 fs: fat: replace inode_timespec with timespec64 fs: ext4: replace inode_timespec with timespec64 vfs: remove inode_timespec and timespec references kernel: time: change inode_timespec to timespec64 vfs: Remove inode_timespec aliases fs: Drop CONFIG_FS_USES_64BIT_TIME fs/attr.c | 15 ++--- fs/bad_inode.c | 10 ++- fs/binfmt_misc.c | 7 +- fs/cifs/cache.c | 16 +++-- fs/cifs/cifsencrypt.c | 2 +- fs/cifs/cifsglob.h | 6 +- fs/cifs/cifsproto.h | 9 +-- fs/cifs/cifssmb.c | 17 +++-- fs/cifs/file.c | 9 ++- fs/cifs/inode.c | 65 ++++++++++++------- fs/cifs/netmisc.c | 26 ++++---- fs/ext4/acl.c | 3 +- fs/ext4/ext4.h | 44 +++++++------ fs/ext4/extents.c | 25 ++++++-- fs/ext4/ialloc.c | 9 ++- fs/ext4/inline.c | 10 ++- fs/ext4/inode.c | 16 +++-- fs/ext4/ioctl.c | 16 +++-- fs/ext4/namei.c | 40 ++++++++---- fs/ext4/super.c | 6 +- fs/ext4/xattr.c | 2 +- fs/fat/dir.c | 7 +- fs/fat/fat.h | 8 ++- fs/fat/file.c | 10 ++- fs/fat/inode.c | 46 ++++++++++---- fs/fat/misc.c | 7 +- fs/fat/namei_msdos.c | 40 +++++++----- fs/fat/namei_vfat.c | 41 ++++++++---- fs/inode.c | 53 +++++++++++----- fs/libfs.c | 50 ++++++++++++--- fs/locks.c | 5 +- fs/nsfs.c | 6 +- fs/pipe.c | 6 +- fs/posix_acl.c | 2 +- fs/stack.c | 6 +- fs/stat.c | 6 +- fs/super.c | 10 +++ fs/utimes.c | 6 +- include/linux/fs.h | 101 +++++++++++++++++++++++++---- include/linux/fs_stack.h | 9 +-- include/linux/stat.h | 6 +- include/linux/time64.h | 4 ++ kernel/time/time.c | 162 ++++++++++++++++++++++++++++++++++++++++++----- 43 files changed, 691 insertions(+), 253 deletions(-) -- 1.9.1

8 years, 10 months

You have received a new fax, document 0000515383

by Interfax

You have a new fax! Please download attached fax document. Pages number: 6 File name: fax_0000515383.doc Scanned: Thu, 14 Jan 2016 15:31:25 +0300 Quality: 300 DPI Sender: Greg Coffman Filesize: 223 Kb Processed in: 12 seconds Thank you for using Interfax!

8 years, 10 months

[PATCH 6/6] scsi: bfa: bfa_cs: Remove struct timeval

by Amitoj Kaur Chawla

32 bit systems using 'struct timeval' will break in the year 2038, so we replace the code appropriately. This patch replaces struct timeval and do_gettimeofday() with monotonic ktime_get_ns(). Since we are interested in microseconds portion of the time, we can use NSEC_PER_USEC macro but this would lead to expensive division for a frequently used function. Alternatively a bit shift has been done which performs a division of 1024 instead of 1000 which slightly alters the value returned by this function. Signed-off-by: Amitoj Kaur Chawla <amitoj1606(a)gmail.com> --- drivers/scsi/bfa/bfa_cs.h | 8 +------- 1 file changed, 1 insertion(+), 7 deletions(-) diff --git a/drivers/scsi/bfa/bfa_cs.h b/drivers/scsi/bfa/bfa_cs.h index 91a8aa3..daee860 100644 --- a/drivers/scsi/bfa/bfa_cs.h +++ b/drivers/scsi/bfa/bfa_cs.h @@ -32,13 +32,7 @@ #define BFA_TRC_MAX (4 * 1024) #endif -#define BFA_TRC_TS(_trcm) \ - ({ \ - struct timeval tv; \ - \ - do_gettimeofday(&tv); \ - (tv.tv_sec*1000000+tv.tv_usec); \ - }) +#define BFA_TRC_TS(_trcm) (ktime_get_ns() >> 10) #ifndef BFA_TRC_TS #define BFA_TRC_TS(_trcm) ((_trcm)->ticks++) -- 1.9.1

8 years, 10 months

[PATCH 5/6] scsi: bfa: bfa_port: Remove 'struct timeval'

by Amitoj Kaur Chawla

32 bit systems using 'struct timeval' will break in the year 2038, so we modify the code appropriately. We only need to find elapsed seconds rather than absolute time and we only care about full seconds so it's better to use monotonic time, so using ktime_get_seconds() also makes the code more efficient and more robust against a concurrent settimeofday(). stats_reset_time variable has been changed to 'time64_t' type to store the monotonic time returned by ktime_get_seconds(). Signed-off-by: Amitoj Kaur Chawla <amitoj1606(a)gmail.com> --- drivers/scsi/bfa/bfa_port.c | 16 ++++------------ drivers/scsi/bfa/bfa_port.h | 2 +- 2 files changed, 5 insertions(+), 13 deletions(-) diff --git a/drivers/scsi/bfa/bfa_port.c b/drivers/scsi/bfa/bfa_port.c index 8ea7697..0c86ea7 100644 --- a/drivers/scsi/bfa/bfa_port.c +++ b/drivers/scsi/bfa/bfa_port.c @@ -95,14 +95,12 @@ bfa_port_get_stats_isr(struct bfa_port_s *port, bfa_status_t status) port->stats_busy = BFA_FALSE; if (status == BFA_STATUS_OK) { - struct timeval tv; - memcpy(port->stats, port->stats_dma.kva, sizeof(union bfa_port_stats_u)); bfa_port_stats_swap(port, port->stats); - do_gettimeofday(&tv); - port->stats->fc.secs_reset = tv.tv_sec - port->stats_reset_time; + port->stats->fc.secs_reset = ktime_get_seconds() - + port->stats_reset_time; } if (port->stats_cbfn) { @@ -123,16 +121,13 @@ bfa_port_get_stats_isr(struct bfa_port_s *port, bfa_status_t status) static void bfa_port_clear_stats_isr(struct bfa_port_s *port, bfa_status_t status) { - struct timeval tv; - port->stats_status = status; port->stats_busy = BFA_FALSE; /* * re-initialize time stamp for stats reset */ - do_gettimeofday(&tv); - port->stats_reset_time = tv.tv_sec; + port->stats_reset_time = ktime_get_seconds(); if (port->stats_cbfn) { port->stats_cbfn(port->stats_cbarg, status); @@ -470,8 +465,6 @@ void bfa_port_attach(struct bfa_port_s *port, struct bfa_ioc_s *ioc, void *dev, struct bfa_trc_mod_s *trcmod) { - struct timeval tv; - WARN_ON(!port); port->dev = dev; @@ -493,8 +486,7 @@ bfa_port_attach(struct bfa_port_s *port, struct bfa_ioc_s *ioc, /* * initialize time stamp for stats reset */ - do_gettimeofday(&tv); - port->stats_reset_time = tv.tv_sec; + port->stats_reset_time = ktime_get_seconds(); bfa_trc(port, 0); } diff --git a/drivers/scsi/bfa/bfa_port.h b/drivers/scsi/bfa/bfa_port.h index 2fcab6b..5f6a9d0 100644 --- a/drivers/scsi/bfa/bfa_port.h +++ b/drivers/scsi/bfa/bfa_port.h @@ -35,7 +35,7 @@ struct bfa_port_s { bfa_port_stats_cbfn_t stats_cbfn; void *stats_cbarg; bfa_status_t stats_status; - u32 stats_reset_time; + time64_t stats_reset_time; union bfa_port_stats_u *stats; struct bfa_dma_s stats_dma; bfa_boolean_t endis_pending; -- 1.9.1

8 years, 10 months

[PATCH 4/6] scsi: bfa: Remove use of time_t

by Amitoj Kaur Chawla

32 bit systems using 'time_t' will break in the year 2038, so we modify the code appropriately. This patch removes the cast to 'time_t' in the assignment statement since we are eventually removing the time_t definition from the kernel as an effort to solve the y2038 problem. This change will avoid the build error but the code is still broken and requires a change in the ioctl interface. Further, since the variable io_profile_start_time will break in 2038 or 2106 depending on user space interpreting it as signed or unsigned, comments have been added to highlight the same. Signed-off-by: Amitoj Kaur Chawla <amitoj1606(a)gmail.com> --- Only apply this patch if it's seen as acceptable that the io_profile_start_time remains truncated to 32 bits in IOCMD_ITNIM_GET_IOPROFILE. If this is something that needs to be fixed by adding a replacement vendor command, leave the cast in place as a reminder. drivers/scsi/bfa/bfa_defs_svc.h | 4 ++++ drivers/scsi/bfa/bfa_fcpim.c | 3 ++- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/drivers/scsi/bfa/bfa_defs_svc.h b/drivers/scsi/bfa/bfa_defs_svc.h index 638f441f..8ab7964 100644 --- a/drivers/scsi/bfa/bfa_defs_svc.h +++ b/drivers/scsi/bfa/bfa_defs_svc.h @@ -1211,6 +1211,10 @@ struct bfa_itnim_ioprofile_s { u32 clock_res_mul; u32 clock_res_div; u32 index; + /* + * Overflow in 2038 or 2106 depending on user space interpreting it as + * signed or unsigned. + */ u32 io_profile_start_time; /* IO profile start time */ u32 iocomps[BFA_IOBUCKET_MAX]; /* IO completed */ struct bfa_itnim_latency_s io_latency; diff --git a/drivers/scsi/bfa/bfa_fcpim.c b/drivers/scsi/bfa/bfa_fcpim.c index 6730340..56df8d0 100644 --- a/drivers/scsi/bfa/bfa_fcpim.c +++ b/drivers/scsi/bfa/bfa_fcpim.c @@ -1478,7 +1478,8 @@ bfa_itnim_get_ioprofile(struct bfa_itnim_s *itnim, return BFA_STATUS_IOPROFILE_OFF; itnim->ioprofile.index = BFA_IOBUCKET_MAX; - itnim->ioprofile.io_profile_start_time = (u32)(time_t) + /* io_profile_start_time will overflow in 2038 or 2106 */ + itnim->ioprofile.io_profile_start_time = (u32) bfa_io_profile_start_time(itnim->bfa); itnim->ioprofile.clock_res_mul = bfa_io_lat_clock_res_mul; itnim->ioprofile.clock_res_div = bfa_io_lat_clock_res_div; -- 1.9.1

8 years, 10 months

[PATCH 3/6] scsi: bfa: Use 'ktime_get_real_seconds()'

by Amitoj Kaur Chawla

This patch replaces u32 time storing variable with u64 variable to store the 64 bit seconds value returned by ktime_get_real_seconds(). A cast to time_t has also been added to the assignment statement which will cause a build error when we eventually remove the time_t definition from the kernel. We can't change the layout of the structure retrieving profile data as it is being used in a vendor specific command that can get sent from user space. Signed-off-by: Amitoj Kaur Chawla <amitoj1606(a)gmail.com> --- drivers/scsi/bfa/bfa_fcpim.c | 6 +++--- drivers/scsi/bfa/bfa_fcpim.h | 4 ++-- drivers/scsi/bfa/bfad_bsg.c | 6 +++--- 3 files changed, 8 insertions(+), 8 deletions(-) diff --git a/drivers/scsi/bfa/bfa_fcpim.c b/drivers/scsi/bfa/bfa_fcpim.c index d7385d1..6730340 100644 --- a/drivers/scsi/bfa/bfa_fcpim.c +++ b/drivers/scsi/bfa/bfa_fcpim.c @@ -468,7 +468,7 @@ bfa_ioim_profile_start(struct bfa_ioim_s *ioim) } bfa_status_t -bfa_fcpim_profile_on(struct bfa_s *bfa, u32 time) +bfa_fcpim_profile_on(struct bfa_s *bfa, u64 time) { struct bfa_itnim_s *itnim; struct bfa_fcpim_s *fcpim = BFA_FCPIM(bfa); @@ -1478,8 +1478,8 @@ bfa_itnim_get_ioprofile(struct bfa_itnim_s *itnim, return BFA_STATUS_IOPROFILE_OFF; itnim->ioprofile.index = BFA_IOBUCKET_MAX; - itnim->ioprofile.io_profile_start_time = - bfa_io_profile_start_time(itnim->bfa); + itnim->ioprofile.io_profile_start_time = (u32)(time_t) + bfa_io_profile_start_time(itnim->bfa); itnim->ioprofile.clock_res_mul = bfa_io_lat_clock_res_mul; itnim->ioprofile.clock_res_div = bfa_io_lat_clock_res_div; *ioprofile = itnim->ioprofile; diff --git a/drivers/scsi/bfa/bfa_fcpim.h b/drivers/scsi/bfa/bfa_fcpim.h index e693af6..274e3af 100644 --- a/drivers/scsi/bfa/bfa_fcpim.h +++ b/drivers/scsi/bfa/bfa_fcpim.h @@ -135,7 +135,7 @@ struct bfa_fcpim_s { struct bfa_fcpim_del_itn_stats_s del_itn_stats; bfa_boolean_t ioredirect; bfa_boolean_t io_profile; - u32 io_profile_start_time; + u64 io_profile_start_time; bfa_fcpim_profile_t profile_comp; bfa_fcpim_profile_t profile_start; }; @@ -309,7 +309,7 @@ bfa_status_t bfa_fcpim_port_iostats(struct bfa_s *bfa, struct bfa_itnim_iostats_s *stats, u8 lp_tag); void bfa_fcpim_add_stats(struct bfa_itnim_iostats_s *fcpim_stats, struct bfa_itnim_iostats_s *itnim_stats); -bfa_status_t bfa_fcpim_profile_on(struct bfa_s *bfa, u32 time); +bfa_status_t bfa_fcpim_profile_on(struct bfa_s *bfa, u64 time); bfa_status_t bfa_fcpim_profile_off(struct bfa_s *bfa); #define bfa_fcpim_ioredirect_enabled(__bfa) \ diff --git a/drivers/scsi/bfa/bfad_bsg.c b/drivers/scsi/bfa/bfad_bsg.c index 023b9d4..92f5b0f 100644 --- a/drivers/scsi/bfa/bfad_bsg.c +++ b/drivers/scsi/bfa/bfad_bsg.c @@ -15,6 +15,7 @@ * General Public License for more details. */ +#include <linux/ktime.h> #include <linux/uaccess.h> #include "bfad_drv.h" #include "bfad_im.h" @@ -2093,13 +2094,12 @@ bfad_iocmd_fcpim_cfg_profile(struct bfad_s *bfad, void *cmd, unsigned int v_cmd) { struct bfa_bsg_fcpim_profile_s *iocmd = (struct bfa_bsg_fcpim_profile_s *)cmd; - struct timeval tv; unsigned long flags; - do_gettimeofday(&tv); spin_lock_irqsave(&bfad->bfad_lock, flags); if (v_cmd == IOCMD_FCPIM_PROFILE_ON) - iocmd->status = bfa_fcpim_profile_on(&bfad->bfa, tv.tv_sec); + iocmd->status = bfa_fcpim_profile_on(&bfad->bfa, + ktime_get_real_seconds()); else if (v_cmd == IOCMD_FCPIM_PROFILE_OFF) iocmd->status = bfa_fcpim_profile_off(&bfad->bfa); spin_unlock_irqrestore(&bfad->bfad_lock, flags); -- 1.9.1

8 years, 10 months

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

Y2038 January 2016