hifn_795x in Linux-2.6.27-rc7
Hi all,
I'm trying to use a Soekris Hifn 7955 based pci card to accelerate dm-crypt device on vanilla 2.6.27-rc7 LK.
# cryptsetup luksFormat -c aes-cbc-plain -s 256 /dev/generic/test completes successfully but the volume is unreadable. Dmesg is
hifn795x 0000:03:04.0: PCI INT A -> GSI 16 (level, low) -> IRQ 16
hifn795x: assuming 66MHz clock speed, override with hifn_pll_ref=ext<frequency>
hifn0: AES 128 ECB test has been successfully passed. Driver for HIFN 795x crypto accelerator chip has been successfully registered.nommu_map_single: overflow 1fe047e00+512 of device mask ffffffffnommu_map_single: overflow 1fe047e00+512 of device mask ffffffffnommu_map_single: overflow 1fe047e00+512 of device mask ffffffff
The same works when I use software AES module. The hardware is a dual opteron 252 with 8 GB RAM on a tyan S2882D mainboard. OS is Debian Lenny i386. Kernel built from vanilla sources from kernel.org. The .config is attached.
Regards,
Dimitri Puzin
Hi Dimitri.
Quoted 1 line · Dimitri Puzin
nommu_map_single: overflow 1fe047e00+512 of device mask ffffffff
Does attached patch fix the problem?
drivers/crypto/hifn_795x.c +1 -1
@@ -2593,7 +2593,7 @@ static int hifn_probe(struct pci_dev *pdev, const struct pci_device_id *id)return err;pci_set_master (pdev); - err =pci_set_dma_mask (pdev,DMA_32BIT_MASK ); + err =pci_set_dma_mask (pdev,DMA_64BIT_MASK );if (err) goto err_out_disable_pci_device;
Signature · 1 line
Evgeniy Polyakov
drivers/crypto/hifn_795x.c static int hifn_probe(struct pci_dev *pdev, const struct pci_device_id *id) · line 2598
Quoted 12 lines · Evgeniy Polyakov
drivers/crypto/hifn_795x.c +1
-1@@ -2593,7 +2593,7 @@ static int hifn_probe(struct pci_dev *pdev, const struct pci_device_id *id)return err;pci_set_master (pdev); - err =pci_set_dma_mask (pdev,DMA_32BIT_MASK ); + err =pci_set_dma_mask (pdev,DMA_64BIT_MASK );if (err) goto err_out_disable_pci_device;
Actually not, I checked documentation, and although HIFN claims to have some kind of 64 bit support, all addresses are 32-bit wide only, so I really wonder if HIFN supports 64 bit.
There is no remapping or address mask parameter in the crypto stack, so that it would not try to provide page to the devices, which do not support it.
Signature · 1 line
Evgeniy Polyakov
Quoted 6 lines · Evgeniy Polyakov
Hi Dimitri.
Quoted 1 line · Dimitri Puzin
nommu_map_single: overflow 1fe047e00+512 of device mask ffffffffDoes attached patch fix the problem?
With this patch applied it still doesn't work as expected. The overflow messages are gone however syslog shows
[ 120.924266] hifn0: abort: c: 0, s: 1, d: 0, r: 0.
when doing cryptsetup luksFormat as in original e-mail. At this point cryptsetup hangs and can't be killed with -SIGKILL. I've attached SysRq-t dump of this condition.
Regards,
Dimitri Puzin
Quoted 14 lines
drivers/crypto/hifn_795x.c +1
-1@@ -2593,7 +2593,7 @@ static int hifn_probe(struct pci_dev *pdev, const struct pci_device_id *id)return err;pci_set_master (pdev); - err =pci_set_dma_mask (pdev,DMA_32BIT_MASK ); + err =pci_set_dma_mask (pdev,DMA_64BIT_MASK );if (err) goto err_out_disable_pci_device;
Quoted 6 lines · Dimitri Puzin
With this patch applied it still doesn't work as expected. The overflow messages are gone however syslog shows
[ 120.924266] hifn0: abort: c: 0, s: 1, d: 0, r: 0.when doing cryptsetup luksFormat as in original e-mail. At this point cryptsetup hangs and can't be killed with -SIGKILL. I've attached SysRq-t dump of this condition.
Yes, I was wrong with the patch: HIFN does not support 64-bit addresses afaics.
Attached patch should not allow HIFN to be registered on 64-bit arch, so crypto layer will fallback to the software algorithms.
drivers/crypto/hifn_795x.c +5 -0
@@ -2787,6 +2787,11 @@ static int __devinit hifn_init(void)unsigned int freq;int err; +if (sizeof (dma_addr_t) >4 ) { +printk (KERN_INFO "HIFN supports only 32-bit addresses.\n" ); +return -EINVAL ; + } +if (strncmp (hifn_pll_ref,"ext" ,3 ) &&strncmp (hifn_pll_ref,"pci" ,3 )) { printk(KERN_ERR "hifn795x: invalid hifn_pll_ref clock, "
Signature · 1 line
Evgeniy Polyakov
David Miller
maintainerFrom: Evgeniy Polyakov <[email protected]>
Date: Tue, 23 Sep 2008 22:18:42 +0400
Quoted 8 lines · Dimitri Puzin
Quoted 6 lines
With this patch applied it still doesn't work as expected. The overflow messages are gone however syslog shows
[ 120.924266] hifn0: abort: c: 0, s: 1, d: 0, r: 0.when doing cryptsetup luksFormat as in original e-mail. At this point cryptsetup hangs and can't be killed with -SIGKILL. I've attached SysRq-t dump of this condition.
Quoted 7 lines
Yes, I was wrong with the patch: HIFN does not support 64-bit addresses afaics.
Attached patch should not allow HIFN to be registered on 64-bit arch, so crypto layer will fallback to the software algorithms.
Signed-off-by Evgeniy Polyakov <[email protected]>
Well, it will disallow HIFN on 32-bit systems using highmem too. :-)
In the end, you will need to use bouncebuffers or similar, it seems.
Quoted 9 lines · David Miller
Quoted 7 lines
Yes, I was wrong with the patch: HIFN does not support 64-bit addresses afaics.
Attached patch should not allow HIFN to be registered on 64-bit arch, so crypto layer will fallback to the software algorithms.
Signed-off-by Evgeniy Polyakov <[email protected]>Well, it will disallow HIFN on 32-bit systems using highmem too. :-)
Yes, but only 64g, where dma_addr_t is 8 bytes.
Quoted 1 line
In the end, you will need to use bouncebuffers or similar, it seems.
Maybe, we need to check, since allocation of multiple smaller pages from 4 gb area, copy data between them, hardware crypto, backward copy and page freeing can be slower than software crypto. Although we can preallocate buffers for several common sizes, although this will not help dm-crypto which can submit up to 31 pages in single request on top of ext3 in the common case.
--
Evgeniy Polyakov
Herbert Xu
maintainerQuoted 7 lines · Evgeniy Polyakov
Maybe, we need to check, since allocation of multiple smaller pages from 4 gb area, copy data between them, hardware crypto, backward copy and page freeing can be slower than software crypto. Although we can preallocate buffers for several common sizes, although this will not help dm-crypto which can submit up to 31 pages in single request on top of ext3 in the common case.
IPsec is going to be lowmem most of the time. So you could allocate a software backup similar to what padlock-sha does and use that within hifn if you get a highmem request.
Of course not registering it works for me too. I don't think we need to lose too much sleep over people who can afford this much memory but still insists on 32-bit. They can afford to hire a kernel engineer to fix this :)
Cheers,
Signature · 4 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txtDavid Miller
maintainerFrom: Evgeniy Polyakov <[email protected]>
Date: Wed, 24 Sep 2008 06:53:08 +0400
Quoted 6 lines
Maybe, we need to check, since allocation of multiple smaller pages from 4 gb area, copy data between them, hardware crypto, backward copy and page freeing can be slower than software crypto. Although we can preallocate buffers for several common sizes, although this will not help dm-crypto which can submit up to 31 pages in single request on top of ext3 in the common case.
I think you will need to preallocate a pool of pages.
Herbert Xu
maintainerQuoted 5 lines · Evgeniy Polyakov
Attached patch should not allow HIFN to be registered on 64-bit arch, so crypto layer will fallback to the software algorithms.
Signed-off-by Evgeniy Polyakov <[email protected]>
Patch applied.
Signature · 4 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txtQuoted 6 lines · Dimitri Puzin
With this patch applied it still doesn't work as expected. The overflow messages are gone however syslog shows
[ 120.924266] hifn0: abort: c: 0, s: 1, d: 0, r: 0.when doing cryptsetup luksFormat as in original e-mail. At this point cryptsetup hangs and can't be killed with -SIGKILL. I've attached SysRq-t dump of this condition.
There are a few known problems left in the HIFN driver that I have patches for but unfortunately didn't manage to clean them up and submit them yet.
Looking at my queue, what's still broken is roughly:
[HIFN]: Have HW invalidate src and dest descriptors after processing
The descriptors need to be invalidated after processing for ring cleanup to work properly and to avoid using an old destination descriptor when the src and cmd descriptors are already set up and the dst descriptor isn't.
That one is actually probably OK but I didn't send it upstream yet because I didn't finish all follow-up patches that depend on how this works.
Further:
[HIFN]: Fix DMA setup
without a changelog :) IIRC the problem is that HIFN sets up each DMA descriptor for the full request length, even though it should only cover on SG entry. The result is memory corruption.
[HIFN]: Don't copy src sg list
also without a changelog. I think this one was just an optimization, the source descriptors don't have alignment or size restrictions and thus we don't need to linearize it first.
[HIFN]: Fix request context corruption
HIFN uses the transform context to store per-request data, which breaks when more than one request is outstanding. Move per request members from struct hifn_context to a new struct hifn_request_context and convert the code to use this.
and:
[HIFN]: Fix queue processing
again without a changelog. The problem this patch fixed was missing crypto backlog handling, causing crashes with dm_crypt under load.
If someone is interested in picking this up I could send my old patches, I probably won't manage to do it myself in the next time.
Hi Patrick.
Quoted 18 lines · Patrick McHardy
There are a few known problems left in the HIFN driver that I have patches for but unfortunately didn't manage to clean them up and submit them yet.
Looking at my queue, what's still broken is roughly:
[HIFN]: Have HW invalidate src and dest descriptors after processing
The descriptors need to be invalidated after processing for ring cleanup to work properly and to avoid using an old destination descriptor when the src and cmd descriptors are already set up and the dst descriptor isn't.
Signed-off-by Patrick McHardy <[email protected]>That one is actually probably OK but I didn't send it upstream yet because I didn't finish all follow-up patches that depend on how this works.
Quoted 7 lines
Further:
[HIFN]: Fix DMA setup
without a changelog :) IIRC the problem is that HIFN sets up each DMA descriptor for the full request length, even though it should only cover on SG entry. The result is memory corruption.
Can not tell without patch itself, but code operates on the number of bytes provided from the higher levels, not full dma size (0x3ff bytes).
Quoted 5 lines
[HIFN]: Don't copy src sg list
also without a changelog. I think this one was just an optimization, the source descriptors don't have alignment or size restrictions and thus we don't need to linearize it first.
Also can not say for sure, but there should not be src linearization, instead we copy data from src to aligned dst and then perform crypto processin in place, and then copy data back.
Quoted 6 lines
[HIFN]: Fix request context corruption
HIFN uses the transform context to store per-request data, which breaks when more than one request is outstanding. Move per request members from struct
hifn_contextto a new structhifn_request_contextand convert the code to use this.
Its per-tfm, so things like key and iv are ok, and others should only be accessed under the lock, but there may be problems.
Quoted 6 lines
and:
[HIFN]: Fix queue processing
again without a changelog. The problem this patch fixed was missing crypto backlog handling, causing crashes with
dm_cryptunder load.
Yup, that's messy :)
Quoted 2 lines
If someone is interested in picking this up I could send my old patches, I probably won't manage to do it myself in the next time.
Please do, if they work for you, we can just apply them, but I would like first to check them out.
Thank you!
Signature · 1 line
Evgeniy Polyakov
Quoted 6 lines · Evgeniy Polyakov
Hi Patrick.
On Wed, Sep 24, 2008 at 06:40:29PM +0200, Patrick McHardy ([email protected]) wrote:
[HIFN]: Fix DMA setup
Quoted 8 lines
without a changelog :) IIRC the problem is that HIFN sets up each DMA descriptor for the full request length, even though it should only cover on SG entry. The result is memory corruption.
Can not tell without patch itself, but code operates on the number of bytes provided from the higher levels, not full dma size (
0x3ffbytes).
Yes, but it uses that size for every descriptor.
I'm going to send you the patches in private (they're quite large) so you can have a look.
Quoted 12 lines
Quoted 7 lines
[HIFN]: Don't copy src sg list
also without a changelog. I think this one was just an optimization, the source descriptors don't have alignment or size restrictions and thus we don't need to linearize it first.
Also can not say for sure, but there should not be src linearization, instead we copy data from src to aligned dst and then perform crypto processin in place, and then copy data back.
Right, I remember. The copying to the (temporary) destination area is not necessary, the chip can just read it from the source are directly.
Quoted 12 lines
Quoted 8 lines
[HIFN]: Fix request context corruption
HIFN uses the transform context to store per-request data, which breaks when more than one request is outstanding. Move per request members from struct
hifn_contextto a new structhifn_request_contextand convert the code to use this.Its per-tfm, so things like key and iv are ok, and others should only be accessed under the lock, but there may be problems.
Things like op, mode, ... are per request and are set without any locking.
Quoted 6 lines
Quoted 3 lines
If someone is interested in picking this up I could send my old patches, I probably won't manage to do it myself in the next time.
Please do, if they work for you, we can just apply them, but I would like first to check them out.
I'm not sure whether they're ready for applying. IIRC I still got stalls occasionally, but I don't remember the details. They made things better for me though :)
Anyways, patches coming up in a minute.
Quoted 8 lines · Patrick McHardy
Quoted 2 lines
Please do, if they work for you, we can just apply them, but I would like first to check them out.
I'm not sure whether they're ready for applying. IIRC I still got stalls occasionally,
but I don't remember the details. They made things better for me though :)Anyways, patches coming up in a minute.
I will review them and push to Herbert if things are ok. Thanks a lot, Patrick!
Signature · 1 line
Evgeniy Polyakov