crash after NX error
On my two socket POWER9 system (powernv) with 842 zwap set up, I recently got a crash with the Ubuntu kernel (I haven't tried with upstream, and this is the first time the system has died like this, so I'm not sure how repeatable it is).
[ 2.891463] zswap: loaded using pool 842-nx/zbud ... [15626.124646] nx_compress_powernv: ERROR: CSB still not valid after 5000000 us, giving up : 00 00 00 00 00000000 [16868.932913] Unable to handle kernel paging request for data at address 0x6655f67da816cdb8 [16868.933726] Faulting instruction address: 0xc000000000391600
cpu 0x68: Vector: 380 (Data Access Out of Range) at [c000001c9d98b9a0]
pc: c000000000391600: kmem_cache_alloc+0x2e0/0x340
lr: c0000000003915ec: kmem_cache_alloc+0x2cc/0x340
sp: c000001c9d98bc20
msr: 900000000280b033
dar: 6655f67da816cdb8
current = 0xc000001ad43cb400
paca = 0xc00000000fac7800 softe: 0 irq_happened: 0x01
pid = 8319, comm = makeLinux version 4.15.0-50-generic (buildd@bos02-ppc64el-006) (gcc version 7.3.0 (Ubuntu 7.3.0-16ubuntu3)) #54-Ubuntu SMP Mon May 6 18:55:18 UTC 2019 (Ubuntu 4.15.0-50.54-generic 4.15.18)
68:mon> t
[c000001c9d98bc20] c0000000003914d4 kmem_cache_alloc+0x1b4/0x340 (unreliable) [c000001c9d98bc80] c0000000003b1e14 __khugepaged_enter+0x54/0x220 [c000001c9d98bcc0] c00000000010f0ec copy_process.isra.5.part.6+0xebc/0x1a10 [c000001c9d98bda0] c00000000010fe4c _do_fork+0xec/0x510 [c000001c9d98be30] c00000000000b584 ppc_clone+0x8/0xc
--- Exception: c00 (System Call) at 00007afe9daf87f4
SP (7fffca606880) is in userspace
So, it looks like there could be a problem in the error path, plausibly fixed by this patch:
commit 656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5
Author: Haren Myneni <[email protected]> Date: Wed Jun 13 00:32:40 2018 -0700
crypto/nx: Initialize 842 high and normal RxFIFO control registers
NX increments readOffset by FIFO size in receive FIFO control register when CRB is read. But the index in RxFIFO has to match with the corresponding entry in FIFO maintained by VAS in kernel. Otherwise NX may be processing incorrect CRBs and can cause CRB timeout.
VAS FIFO offset is 0 when the receive window is opened during initialization. When the module is reloaded or in kexec boot, readOffset in FIFO control register may not match with VAS entry. This patch adds nx_coproc_init OPAL call to reset readOffset and queued entries in FIFO control register for both high and normal FIFOs.
$ git describe --contains 656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5 v4.19-rc1~24^2~50
Which was never backported to any stable release, so probably needs to be for v4.14 through v4.18. Notably, Ubuntu is on v4.15 and it doesn't seem to have picked up the patch. I'm opening an Ubuntu bug for this.
Haren, is this something you can drive through the stable process (assuming my above crash looks like this failure)?
Signature · 2 lines
Stewart Smith OPAL Architect, IBM.
Quoted 60 lines · Stewart Smith
On my two socket POWER9 system (powernv) with 842 zwap set up, I recently got a crash with the Ubuntu kernel (I haven't tried with upstream, and this is the first time the system has died like this, so I'm not sure how repeatable it is).
[ 2.891463] zswap: loaded using pool 842-nx/zbud ... [15626.124646] nx_compress_powernv: ERROR: CSB still not valid after 5000000 us, giving up : 00 00 00 00 00000000 [16868.932913] Unable to handle kernel paging request for data at address 0x6655f67da816cdb8 [16868.933726] Faulting instruction address: 0xc000000000391600cpu
0x68: Vector: 380 (Data Access Out of Range) at [c000001c9d98b9a0]pc: c000000000391600: kmem_cache_alloc+0x2e0/0x340 lr: c0000000003915ec: kmem_cache_alloc+0x2cc/0x340 sp: c000001c9d98bc20 msr: 900000000280b033 dar: 6655f67da816cdb8 current = 0xc000001ad43cb400 paca = 0xc00000000fac7800 softe: 0 irq_happened: 0x01 pid = 8319, comm = makeLinux version 4.15.0-50-generic (buildd@bos02-ppc64el-006) (gcc version 7.3.0 (Ubuntu 7.3.0-16ubuntu3)) #54-Ubuntu SMP Mon May 6 18:55:18 UTC 2019 (Ubuntu 4.15.0-50.54-generic 4.15.18)
68:mon> t
[c000001c9d98bc20]c0000000003914d4kmem_cache_alloc+0x1b4/0x340(unreliable) [c000001c9d98bc80]c0000000003b1e14__khugepaged_enter+0x54/0x220[c000001c9d98bcc0]c00000000010f0eccopy_process.isra.5.part.6+0xebc/0x1a10 [c000001c9d98bda0]c00000000010fe4c_do_fork+0xec/0x510[c000001c9d98be30]c00000000000b584ppc_clone+0x8/0xc
--- Exception: c00 (System Call) at00007afe9daf87f4
SP (7fffca606880) is in userspaceSo, it looks like there could be a problem in the error path, plausibly fixed by this patch:
commit
656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5Author: Haren Myneni <[email protected]> Date: Wed Jun 13 00:32:40 2018 -0700
crypto/nx: Initialize 842 high and normal RxFIFO control registersNX increments readOffset by FIFO size in receive FIFO control register when CRB is read. But the index in RxFIFO has to match with the corresponding entry in FIFO maintained by VAS in kernel. Otherwise NX may be processing incorrect CRBs and can cause CRB timeout.
VAS FIFO offset is 0 when the receive window is opened during initialization. When the module is reloaded or in kexec boot, readOffset in FIFO control register may not match with VAS entry. This patch adds
nx_coproc_initOPAL call to reset readOffset and queued entries in FIFO control register for both high and normal FIFOs.Signed-off-by Haren Myneni <[email protected]>[mpe: Fixup uninitialized variable warning]Signed-off-by Michael Ellerman <[email protected]>$ git describe --contains 656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5 v4.19-rc1~24^2~50
Quoted 7 lines
Which was never backported to any stable release, so probably needs to be for v4.14 through v4.18. Notably, Ubuntu is on v4.15 and it doesn't seem to have picked up the patch. I'm opening an Ubuntu bug for this.
Haren, is this something you can drive through the stable process (assuming my above crash looks like this failure)?
Thanks Stewart. Missed this in stable releases and I will work on it. Merged in Ubuntu 18.04.x kernel recently and will be in the next update.
Also need
commit 6e708000ec2c93c2bde6a46aa2d6c3e80d4eaeb9
Author: Haren Myneni <[email protected]> Date: Wed Jun 13 00:28:57 2018 -0700
powerpc/powernv: Export opal_check_token symbol
Export opal_check_token symbol for modules to check the availability of OPAL calls before using them.
Quoted 56 lines · Stewart Smith
On my two socket POWER9 system (powernv) with 842 zwap set up, I recently got a crash with the Ubuntu kernel (I haven't tried with upstream, and this is the first time the system has died like this, so I'm not sure how repeatable it is).
[ 2.891463] zswap: loaded using pool 842-nx/zbud ... [15626.124646] nx_compress_powernv: ERROR: CSB still not valid after 5000000 us, giving up : 00 00 00 00 00000000 [16868.932913] Unable to handle kernel paging request for data at address 0x6655f67da816cdb8 [16868.933726] Faulting instruction address: 0xc000000000391600cpu
0x68: Vector: 380 (Data Access Out of Range) at [c000001c9d98b9a0]pc: c000000000391600: kmem_cache_alloc+0x2e0/0x340 lr: c0000000003915ec: kmem_cache_alloc+0x2cc/0x340 sp: c000001c9d98bc20 msr: 900000000280b033 dar: 6655f67da816cdb8 current = 0xc000001ad43cb400 paca = 0xc00000000fac7800 softe: 0 irq_happened: 0x01 pid = 8319, comm = makeLinux version 4.15.0-50-generic (buildd@bos02-ppc64el-006) (gcc version 7.3.0 (Ubuntu 7.3.0-16ubuntu3)) #54-Ubuntu SMP Mon May 6 18:55:18 UTC 2019 (Ubuntu 4.15.0-50.54-generic 4.15.18)
68:mon> t
[c000001c9d98bc20]c0000000003914d4kmem_cache_alloc+0x1b4/0x340(unreliable) [c000001c9d98bc80]c0000000003b1e14__khugepaged_enter+0x54/0x220[c000001c9d98bcc0]c00000000010f0eccopy_process.isra.5.part.6+0xebc/0x1a10 [c000001c9d98bda0]c00000000010fe4c_do_fork+0xec/0x510[c000001c9d98be30]c00000000000b584ppc_clone+0x8/0xc
--- Exception: c00 (System Call) at00007afe9daf87f4
SP (7fffca606880) is in userspaceSo, it looks like there could be a problem in the error path, plausibly fixed by this patch:
commit
656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5Author: Haren Myneni <[email protected]> Date: Wed Jun 13 00:32:40 2018 -0700
crypto/nx: Initialize 842 high and normal RxFIFO control registersNX increments readOffset by FIFO size in receive FIFO control register when CRB is read. But the index in RxFIFO has to match with the corresponding entry in FIFO maintained by VAS in kernel. Otherwise NX may be processing incorrect CRBs and can cause CRB timeout.
VAS FIFO offset is 0 when the receive window is opened during initialization. When the module is reloaded or in kexec boot, readOffset in FIFO control register may not match with VAS entry. This patch adds
nx_coproc_initOPAL call to reset readOffset and queued entries in FIFO control register for both high and normal FIFOs.Signed-off-by Haren Myneni <[email protected]>[mpe: Fixup uninitialized variable warning]Signed-off-by Michael Ellerman <[email protected]>
Quoted 6 lines
$ git describe --contains 656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5 v4.19-rc1~24^2~50Which was never backported to any stable release, so probably needs to be for v4.14 through v4.18.
Yeah the P9 NX support went in in:
b0d6c9bab5e4 ("crypto/nx: Add P9 NX support for 842 compression engine")
Which was: v4.14-rc1~119^2~21, so first released in v4.14.
I'm actually less interested in that and more interested in the subsequent crash. The time stamps are miles apart though, did we just leave some corrupted memory after the NX failed and then hit it later? Or did we not correctly signal to the upper level APIs that the request failed.
I think we need to do some testing with errors injected into the wait_for_csb() path, to ensure that failures there are not causing corrupting in zswap. Haren have you done any testing of error injection?
cheers
Quoted 76 lines · Michael Ellerman
Quoted 56 lines · Stewart Smith
On my two socket POWER9 system (powernv) with 842 zwap set up, I recently got a crash with the Ubuntu kernel (I haven't tried with upstream, and this is the first time the system has died like this, so I'm not sure how repeatable it is).
[ 2.891463] zswap: loaded using pool 842-nx/zbud ... [15626.124646] nx_compress_powernv: ERROR: CSB still not valid after 5000000 us, giving up : 00 00 00 00 00000000 [16868.932913] Unable to handle kernel paging request for data at address 0x6655f67da816cdb8 [16868.933726] Faulting instruction address: 0xc000000000391600cpu
0x68: Vector: 380 (Data Access Out of Range) at [c000001c9d98b9a0]pc: c000000000391600: kmem_cache_alloc+0x2e0/0x340 lr: c0000000003915ec: kmem_cache_alloc+0x2cc/0x340 sp: c000001c9d98bc20 msr: 900000000280b033 dar: 6655f67da816cdb8 current = 0xc000001ad43cb400 paca = 0xc00000000fac7800 softe: 0 irq_happened: 0x01 pid = 8319, comm = makeLinux version 4.15.0-50-generic (buildd@bos02-ppc64el-006) (gcc version 7.3.0 (Ubuntu 7.3.0-16ubuntu3)) #54-Ubuntu SMP Mon May 6 18:55:18 UTC 2019 (Ubuntu 4.15.0-50.54-generic 4.15.18)
68:mon> t
[c000001c9d98bc20]c0000000003914d4kmem_cache_alloc+0x1b4/0x340(unreliable) [c000001c9d98bc80]c0000000003b1e14__khugepaged_enter+0x54/0x220[c000001c9d98bcc0]c00000000010f0eccopy_process.isra.5.part.6+0xebc/0x1a10 [c000001c9d98bda0]c00000000010fe4c_do_fork+0xec/0x510[c000001c9d98be30]c00000000000b584ppc_clone+0x8/0xc
--- Exception: c00 (System Call) at00007afe9daf87f4
SP (7fffca606880) is in userspaceSo, it looks like there could be a problem in the error path, plausibly fixed by this patch:
commit
656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5Author: Haren Myneni <[email protected]> Date: Wed Jun 13 00:32:40 2018 -0700
crypto/nx: Initialize 842 high and normal RxFIFO control registersNX increments readOffset by FIFO size in receive FIFO control register when CRB is read. But the index in RxFIFO has to match with the corresponding entry in FIFO maintained by VAS in kernel. Otherwise NX may be processing incorrect CRBs and can cause CRB timeout.
VAS FIFO offset is 0 when the receive window is opened during initialization. When the module is reloaded or in kexec boot, readOffset in FIFO control register may not match with VAS entry. This patch adds
nx_coproc_initOPAL call to reset readOffset and queued entries in FIFO control register for both high and normal FIFOs.Signed-off-by Haren Myneni <[email protected]>[mpe: Fixup uninitialized variable warning]Signed-off-by Michael Ellerman <[email protected]>Quoted 6 lines
$ git describe --contains 656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5 v4.19-rc1~24^2~50Which was never backported to any stable release, so probably needs to be for v4.14 through v4.18.
Yeah the P9 NX support went in in:
b0d6c9bab5e4("crypto/nx: Add P9 NX support for 842 compression engine")Which was: v4.14-rc1~119^2~21, so first released in v4.14.
I'm actually less interested in that and more interested in the subsequent crash. The time stamps are miles apart though, did we just leave some corrupted memory after the NX failed and then hit it later? Or did we not correctly signal to the upper level APIs that the request failed.
Quoted 4 lines
I think we need to do some testing with errors injected into the
wait_for_csb()path, to ensure that failures there are not causing corrupting in zswap. Haren have you done any testing of error injection?
So, things died pretty heavily overnight (requiring e2fsck) with a *lot* of those wait_for_csb() errors in the log.
It certainly *looks* like there's corruption around, as one of the CI jobs that failed around that time got "internal compiler error" which is usually a good sign that things have gone poorly somewhere.
Signature · 2 lines
Stewart Smith OPAL Architect, IBM.
Quoted 76 lines · Michael Ellerman
Quoted 56 lines · Stewart Smith
On my two socket POWER9 system (powernv) with 842 zwap set up, I recently got a crash with the Ubuntu kernel (I haven't tried with upstream, and this is the first time the system has died like this, so I'm not sure how repeatable it is).
[ 2.891463] zswap: loaded using pool 842-nx/zbud ... [15626.124646] nx_compress_powernv: ERROR: CSB still not valid after 5000000 us, giving up : 00 00 00 00 00000000 [16868.932913] Unable to handle kernel paging request for data at address 0x6655f67da816cdb8 [16868.933726] Faulting instruction address: 0xc000000000391600cpu
0x68: Vector: 380 (Data Access Out of Range) at [c000001c9d98b9a0]pc: c000000000391600: kmem_cache_alloc+0x2e0/0x340 lr: c0000000003915ec: kmem_cache_alloc+0x2cc/0x340 sp: c000001c9d98bc20 msr: 900000000280b033 dar: 6655f67da816cdb8 current = 0xc000001ad43cb400 paca = 0xc00000000fac7800 softe: 0 irq_happened: 0x01 pid = 8319, comm = makeLinux version 4.15.0-50-generic (buildd@bos02-ppc64el-006) (gcc version 7.3.0 (Ubuntu 7.3.0-16ubuntu3)) #54-Ubuntu SMP Mon May 6 18:55:18 UTC 2019 (Ubuntu 4.15.0-50.54-generic 4.15.18)
68:mon> t
[c000001c9d98bc20]c0000000003914d4kmem_cache_alloc+0x1b4/0x340(unreliable) [c000001c9d98bc80]c0000000003b1e14__khugepaged_enter+0x54/0x220[c000001c9d98bcc0]c00000000010f0eccopy_process.isra.5.part.6+0xebc/0x1a10 [c000001c9d98bda0]c00000000010fe4c_do_fork+0xec/0x510[c000001c9d98be30]c00000000000b584ppc_clone+0x8/0xc
--- Exception: c00 (System Call) at00007afe9daf87f4
SP (7fffca606880) is in userspaceSo, it looks like there could be a problem in the error path, plausibly fixed by this patch:
commit
656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5Author: Haren Myneni <[email protected]> Date: Wed Jun 13 00:32:40 2018 -0700
crypto/nx: Initialize 842 high and normal RxFIFO control registersNX increments readOffset by FIFO size in receive FIFO control register when CRB is read. But the index in RxFIFO has to match with the corresponding entry in FIFO maintained by VAS in kernel. Otherwise NX may be processing incorrect CRBs and can cause CRB timeout.
VAS FIFO offset is 0 when the receive window is opened during initialization. When the module is reloaded or in kexec boot, readOffset in FIFO control register may not match with VAS entry. This patch adds
nx_coproc_initOPAL call to reset readOffset and queued entries in FIFO control register for both high and normal FIFOs.Signed-off-by Haren Myneni <[email protected]>[mpe: Fixup uninitialized variable warning]Signed-off-by Michael Ellerman <[email protected]>Quoted 6 lines
$ git describe --contains 656ecc16e8fc2ab44b3d70e3fcc197a7020d0ca5 v4.19-rc1~24^2~50Which was never backported to any stable release, so probably needs to be for v4.14 through v4.18.
Yeah the P9 NX support went in in:
b0d6c9bab5e4("crypto/nx: Add P9 NX support for 842 compression engine")Which was: v4.14-rc1~119^2~21, so first released in v4.14.
I'm actually less interested in that and more interested in the subsequent crash. The time stamps are miles apart though, did we just leave some corrupted memory after the NX failed and then hit it later? Or did we not correctly signal to the upper level APIs that the request failed.
Quoted 3 lines
I think we need to do some testing with errors injected into the
wait_for_csb()path, to ensure that failures there are not causing corrupting in zswap. Haren have you done any testing of error injection?
The code path returns error code from wait_for_csb() properly to upper level APIs. In the case of decompression case, upon failure the request will fall back to SW 842.
If NX is involved in this crash, the compression request may be successful with invalid CRB (mismatch FIFO entries in NX and VAS). Then SW 842 may be decompressed invalid data which might cause corruption later when accessing it.
I will try to reproduce the issue with 4.14 kernel,
Thanks
Haren
Quoted 3 lines
cheers