Bug 120111
| Summary: | [crypto] [marvell-cesa] Driver hangs on Armada 385 device | ||
|---|---|---|---|
| Product: | Drivers | Reporter: | Jussi Keränen (jussike) |
| Component: | Platform | Assignee: | drivers_other |
| Status: | NEW --- | ||
| Severity: | normal | CC: | jussike, thomas.petazzoni |
| Priority: | P1 | ||
| Hardware: | ARM | ||
| OS: | Linux | ||
| Kernel Version: | 4.8-rc2 | Subsystem: | |
| Regression: | No | Bisected commit-id: | |
| Attachments: | Kernel configuration | ||
Some updates: Linux 4.8-rc2 fixes random locks when tdma is disabled. Important commit is "ARM: mvebu: fix HW I/O coherency related deadlocks". Still the device always completely freezes without the patch provided above. The patch is not really workaround because it kills too much performance. This bug seems to be a memory mapping related problem. I have used mem= kernel-parameter and that triggers the problem. Some mem-values works and others not. Working values are: (with device with 1 GiB of RAM) 256M (and up to 60k below) 512M (and up to 60k below) 768M (and up to 60k below) 1024M (and up to 60k below) All other tried values causes the problem described above. |
Created attachment 219721 [details] Kernel configuration Steps to reproduce: * Armada 385 board (88F6820) * Linux 4.4.10 (or later up to Linux 4.7-rc3) * Marvell-cesa enabled on kernel configuration (full conf as attachment) * Device tree: armada-385-db-ap.dts # cryptsetup benchmark --cipher aes-cbc Actual results: Kernel messages during boot: marvell-cesa f1090000.crypto: CESA device successfully registered mv_xor f1060800.xor: Marvell shared XOR driver mv_xor f1060800.xor: Marvell XOR (Registers Mode): ( xor cpy intr ) mv_xor f1060900.xor: Marvell shared XOR driver mv_xor f1060900.xor: Marvell XOR (Registers Mode): ( xor cpy intr ) Cryptsetup benchmark never completes and the device locks up. Expected results: Cryptsetup benchmark completed. Additional information: After some testing and debugging, I can workaround this issue by disabling TDMA: --- a/drivers/crypto/marvell/cesa.c +++ b/drivers/crypto/marvell/cesa.c @@ -239,7 +239,7 @@ static const struct mv_cesa_caps armada_xp_caps = { .ncipher_algs = ARRAY_SIZE(armada_370_cipher_algs), .ahash_algs = armada_370_ahash_algs, .nahash_algs = ARRAY_SIZE(armada_370_ahash_algs), - .has_tdma = true, + .has_tdma = false, }; static const struct of_device_id mv_cesa_of_match_table[] = { After patching I can run the cryptsetup benchmark with following results: # Tests are approximate using memory only (no storage IO). # Algorithm | Key | Encryption | Decryption aes-cbc 256b 51.0 MiB/s 50.5 MiB/s Performance is slightly better than without acceleration which is approx. 25 MiB/s. The device still hangs after a couple of crypto tasks, e.g. when writing a few hundred MiB of data to btrfs which is on top of luks. I'm using Linux-yocto 4.4.10 and tried to update drivers/crypto/marvell and drivers/dma/mv_xor.* to Linux 4.7-rc3-version. I haven't noticed any difference with 4.4.10 -version. I'm primarily using a custom dts, but I tested with an unmodified armada-385-db-ap.dts with the same results.