x86_64: Bug in new out of line put_user()
The new out of line put_user() assembly on x86_64 changes %rcx without telling GCC about it causing things like:
http://bugme.osdl.org/show_bug.cgi?id=4515
See to it that %rcx is not changed (made it consistent with get_user()).
arch/x86_64/lib/getuser.S +2 -2
@@ -78,9 +78,9 @@ __get_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_get_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_get_user + jae 40f subq $7,%rcx 4: movq (%rcx),%rdx xorl %eax,%eax
arch/x86_64/lib/putuser.S +18 -9
@@ -46,36 +46,45 @@ __put_user_2: GET_THREAD_INFO(%r8) addq $1,%rcx - jc bad_put_user + jc 20f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -2: movw %dx,-1(%rcx) + jae 20f +2: decq %rcx + movw %dx,(%rcx) xorl %eax,%eax ret +20: decq %rcx + jmp bad_put_user .p2align 4 .globl __put_user_4 __put_user_4: GET_THREAD_INFO(%r8) addq $3,%rcx - jc bad_put_user + jc 30f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -3: movl %edx,-3(%rcx) + jae 30f +3: subq $3,%rcx + movl %edx,(%rcx) xorl %eax,%eax ret +30: subq $3,%rcx + jmp bad_put_user .p2align 4 .globl __put_user_8 __put_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_put_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -4: movq %rdx,-7(%rcx) + jae 40f +4: subq $7,%rcx + movq %rdx,(%rcx) xorl %eax,%eax ret +40: subq $7,%rcx + jmp bad_put_user bad_put_user: movq $(-EFAULT),%rax
Quoted 6 lines · Alexander Nyberg
The new out of line
put_user()assembly onx86_64changes %rcx without telling GCC about it causing things like:http://bugme.osdl.org/show_bug.cgi?id=4515
See to it that %rcx is not changed (made it consistent with
get_user()).
Damn. I actually fixed this before submission, but it looks like the old patch staid in the queue :-( Sorry for you having to debug it again.
Linus, can can you please apply the patch?
-Andi
Quoted 81 lines
Signed-off-by Alexander Nyberg <[email protected]>Index:
test/arch/x86_64/lib/getuser.S
===================================================================arch/x86_64/lib/getuser.S +2
-2@@ -78,9 +78,9 @@ __get_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_get_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_get_user + jae 40f subq $7,%rcx 4: movq (%rcx),%rdx xorl %eax,%eaxIndex:
test/arch/x86_64/lib/putuser.S
===================================================================arch/x86_64/lib/putuser.S +18
-9@@ -46,36 +46,45 @@ __put_user_2: GET_THREAD_INFO(%r8) addq $1,%rcx - jc bad_put_user + jc 20f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -2: movw %dx,-1(%rcx) + jae 20f +2: decq %rcx + movw %dx,(%rcx) xorl %eax,%eax ret +20: decq %rcx + jmp bad_put_user .p2align 4 .globl __put_user_4 __put_user_4: GET_THREAD_INFO(%r8) addq $3,%rcx - jc bad_put_user + jc 30f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -3: movl %edx,-3(%rcx) + jae 30f +3: subq $3,%rcx + movl %edx,(%rcx) xorl %eax,%eax ret +30: subq $3,%rcx + jmp bad_put_user .p2align 4 .globl __put_user_8 __put_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_put_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -4: movq %rdx,-7(%rcx) + jae 40f +4: subq $7,%rcx + movq %rdx,(%rcx) xorl %eax,%eax ret +40: subq $7,%rcx + jmp bad_put_user bad_put_user: movq $(-EFAULT),%rax
Hello,
Quoted 4 lines · Alexander Nyberg
The new out of line
put_user()assembly onx86_64changes %rcx without telling GCC about it causing things like:
Thank you, this patch fixes the message queue problem.
Best regards,
Nicolas
Quoted 83 lines
See to it that %rcx is not changed (made it consistent with
get_user()).Signed-off-by Alexander Nyberg <[email protected]>Index:
test/arch/x86_64/lib/getuser.S
===================================================================arch/x86_64/lib/getuser.S +2
-2@@ -78,9 +78,9 @@ __get_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_get_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_get_user + jae 40f subq $7,%rcx 4: movq (%rcx),%rdx xorl %eax,%eaxIndex:
test/arch/x86_64/lib/putuser.S
===================================================================arch/x86_64/lib/putuser.S +18
-9@@ -46,36 +46,45 @@ __put_user_2: GET_THREAD_INFO(%r8) addq $1,%rcx - jc bad_put_user + jc 20f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -2: movw %dx,-1(%rcx) + jae 20f +2: decq %rcx + movw %dx,(%rcx) xorl %eax,%eax ret +20: decq %rcx + jmp bad_put_user .p2align 4 .globl __put_user_4 __put_user_4: GET_THREAD_INFO(%r8) addq $3,%rcx - jc bad_put_user + jc 30f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -3: movl %edx,-3(%rcx) + jae 30f +3: subq $3,%rcx + movl %edx,(%rcx) xorl %eax,%eax ret +30: subq $3,%rcx + jmp bad_put_user .p2align 4 .globl __put_user_8 __put_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_put_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -4: movq %rdx,-7(%rcx) + jae 40f +4: subq $7,%rcx + movq %rdx,(%rcx) xorl %eax,%eax ret +40: subq $7,%rcx + jmp bad_put_user bad_put_user: movq $(-EFAULT),%rax
Hello Alexander,
I have other kind of problems with this patch...
With 2.6.12-rc2 + your patch, when I run OpenOffice (a 32-bit application), I get this in dmesg :
Unable to handle kernel paging request at 00002aaaad9280b0 RIP:
<ffffffff801f5d70>{__put_user_4+32}
PGD 0
Oops: 0002 [1] SMP
CPU 0
Modules linked in: cpufreq_ondemand sch_sfq sch_htb ipt_CONNMARKipt_ipp2p ipt_connmark ipt_mark ipt_tos ipt_length ipt_MARK iptable_filter iptable_mangle ip_tables powernow_k8 tun nvidia snd_pcm_oss snd_mixer_oss snd_via82xx snd_ac97_codec snd_pcm snd_timer snd_page_alloc snd_mpu401_uart snd_rawmidi snd_seq_device snd cx8800 v4l1_compat v4l2_common cx88xx i2c_algo_bit video_buf ir_common btcx_risc tveeprom videodev tuner usbhid w83627hf i2c_sensor i2c_isa i2c_core usb_storage uhci_hcd ehci_hcd usbcore sd_mod scsi_mod
Pid: 11236, comm: soffice.bin Tainted: P 2.6.12-rc2
RIP: 0010:[<ffffffff801f5d70>] <ffffffff801f5d70>{__put_user_4+32}
RSP: 0000:ffff81001d32fea0 EFLAGS: 00010202
RAX: 0000000000000003 RBX: ffff8100255b0f70 RCX: 00002aaaad9280b0
RDX: 0000000000000000 RSI: ffff810033534940 RDI: 00002aaaad9280b0
RBP: 0000000000000000 R08: ffff81001d32e000 R09: 0000000000000002
R10: ffff81001d32e000 R11: 00000000558740dc R12: 0000000000000000
R13: ffff810033534940 R14: 0000000000000001 R15: 0000000000000000
FS: 00002aaaae0ff820(0000) GS:ffffffff80489840(0000)knlGS:0000000000000000
CS: 0010 DS: 002b ES: 002b CR0: 000000008005003b
CR2: 00002aaaad9280b0 CR3: 0000000025150000 CR4: 00000000000006e0 Process soffice.bin (pid: 11236, threadinfo ffff81001d32e000, task ffff8100255b0f70)
Stack: ffffffff80133414 ffff810019f11138 ffff810033534940
ffff8100255b0f70
ffff8100255b0f70 0000000000000000 ffffffff80137c0a
0000000000000000
0000000000000000 ffff8100255b15a4
Call Trace:<ffffffff80133414>{mm_release+116} <ffffffff80137c0a>{exit_mm
+42}
<ffffffff8013802f>{do_exit+351} <ffffffff80138b8c>{do_group_exit
+252}
<ffffffff80122ab1>{ia32_sysret+0} It is easily reproducible on my system, I just have to start and exit OpenOffice 1.1.4.
I got similar messages with 2.6.12-rc3 + your patch (please tell me if you also want these messages).
I also had one system freeze and one reboot (when I clicked on a link in Firefox (32-bit app too), I don't know if it is related). But I was unable to reproduce these crashes, so I'm not sure whether it is caused by rc3 or by your patch.
Please note that these problems did not occur with 2.6.12-rc2 + these patches reverted:
- x86_64-give-out-of-line-get_user-better-calling.patch
- x86_64-move-put_user-out-of-line.patch
- x86_64-remove-stale-unused-file.patch
Best regards,
Nicolas
Quoted 88 lines · Alexander Nyberg
The new out of line
put_user()assembly onx86_64changes %rcx without telling GCC about it causing things like:http://bugme.osdl.org/show_bug.cgi?id=4515
See to it that %rcx is not changed (made it consistent with
get_user()).Signed-off-by Alexander Nyberg <[email protected]>Index:
test/arch/x86_64/lib/getuser.S
===================================================================arch/x86_64/lib/getuser.S +2
-2@@ -78,9 +78,9 @@ __get_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_get_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_get_user + jae 40f subq $7,%rcx 4: movq (%rcx),%rdx xorl %eax,%eaxIndex:
test/arch/x86_64/lib/putuser.S
===================================================================arch/x86_64/lib/putuser.S +18
-9@@ -46,36 +46,45 @@ __put_user_2: GET_THREAD_INFO(%r8) addq $1,%rcx - jc bad_put_user + jc 20f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -2: movw %dx,-1(%rcx) + jae 20f +2: decq %rcx + movw %dx,(%rcx) xorl %eax,%eax ret +20: decq %rcx + jmp bad_put_user .p2align 4 .globl __put_user_4 __put_user_4: GET_THREAD_INFO(%r8) addq $3,%rcx - jc bad_put_user + jc 30f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -3: movl %edx,-3(%rcx) + jae 30f +3: subq $3,%rcx + movl %edx,(%rcx) xorl %eax,%eax ret +30: subq $3,%rcx + jmp bad_put_user .p2align 4 .globl __put_user_8 __put_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_put_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -4: movq %rdx,-7(%rcx) + jae 40f +4: subq $7,%rcx + movq %rdx,(%rcx) xorl %eax,%eax ret +40: subq $7,%rcx + jmp bad_put_user bad_put_user: movq $(-EFAULT),%rax
Quoted 78 lines · Alexander Nyberg
The new out of line
put_user()assembly onx86_64changes %rcx without telling GCC about it causing things like:http://bugme.osdl.org/show_bug.cgi?id=4515
See to it that %rcx is not changed (made it consistent with
get_user()).Signed-off-by Alexander Nyberg <[email protected]>Index:
test/arch/x86_64/lib/getuser.S
===================================================================arch/x86_64/lib/getuser.S +2
-2@@ -78,9 +78,9 @@ __get_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_get_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_get_user + jae 40f subq $7,%rcx 4: movq (%rcx),%rdx xorl %eax,%eaxIndex:
test/arch/x86_64/lib/putuser.S
===================================================================arch/x86_64/lib/putuser.S +16
-9@@ -46,36 +46,45 @@ __put_user_2: GET_THREAD_INFO(%r8) addq $1,%rcx - jc bad_put_user + jc 20f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -2: movw %dx,-1(%rcx) + jae 20f +2: decq %rcx + movw %dx,(%rcx) xorl %eax,%eax ret +20: decq %rcx + jmp bad_put_user .p2align 4 .globl __put_user_4 __put_user_4: GET_THREAD_INFO(%r8) addq $3,%rcx - jc bad_put_user + jc 30f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -3: movl %edx,-3(%rcx) + jae 30f +3: subq $3,%rcx + movl %edx,(%rcx) xorl %eax,%eax ret +30: subq $3,%rcx + jmp bad_put_user .p2align 4 .globl __put_user_8 __put_user_8: GET_THREAD_INFO(%r8) addq $7,%rcx - jc bad_put_user + jc 40f cmpq threadinfo_addr_limit(%r8),%rcx - jae bad_put_user -4: movq %rdx,-7(%rcx) + jae 40f +4: subq $7,%rcx + movq %rdx,(%rcx)
arch/x86_64/lib/putuser.S
Quoted 10 lines
arch/x86_64/lib/putuser.S +2
-1xorl %eax,%eax ret +40: subq $7,%rcx + jmp bad_put_user bad_put_user: movq $(-EFAULT),%rax -
This patch has a serious bug. The 2, 3 and 4 labels must be on the mov instructions in order to catch faults.
--
Brian Gerst
Brian, thanks for seeing this. (me goes hiding...)
The labels after the last put_user patch were misplaced so exceptions on the real mov instructions would not be handled.
arch/x86_64/lib/putuser.S +6 -6
@@ -49,8 +49,8 @@ jc 20f cmpq threadinfo_addr_limit(%r8),%rcx jae 20f -2: decq %rcx - movw %dx,(%rcx) + decq %rcx +2: movw %dx,(%rcx) xorl %eax,%eax ret 20: decq %rcx
@@ -64,8 +64,8 @@ jc 30f cmpq threadinfo_addr_limit(%r8),%rcx jae 30f -3: subq $3,%rcx - movl %edx,(%rcx) + subq $3,%rcx +3: movl %edx,(%rcx) xorl %eax,%eax ret 30: subq $3,%rcx
@@ -79,8 +79,8 @@ jc 40f cmpq threadinfo_addr_limit(%r8),%rcx jae 40f -4: subq $7,%rcx - movq %rdx,(%rcx) + subq $7,%rcx +4: movq %rdx,(%rcx) xorl %eax,%eax ret 40: subq $7,%rcx