openwrtv4/target/linux/apm821xx/patches-4.14/020-0017-crypto-crypto4xx-add-backlog-queue-support.patch
Christian Lamparter 780477d17c apm821xx: backport crypto4xx patches from 4.15
This patch backports changes to crypto4xx in order
to get the crypto4xx operational.

Signed-off-by: Christian Lamparter <chunkeey@gmail.com>
2018-01-12 08:00:04 +01:00

161 lines
5.8 KiB
Diff

From 8ef8d195430ca3542d0434cf25e5115484b9fa32 Mon Sep 17 00:00:00 2001
From: Christian Lamparter <chunkeey@gmail.com>
Date: Wed, 4 Oct 2017 01:00:09 +0200
Subject: [PATCH 17/25] crypto: crypto4xx - add backlog queue support
Previously, If the crypto4xx driver used all available
security contexts, it would simply refuse new requests
with -EAGAIN. CRYPTO_TFM_REQ_MAY_BACKLOG was ignored.
in case of dm-crypt.c's crypt_convert() function this was
causing the following errors to manifest, if the system was
pushed hard enough:
| EXT4-fs warning (dm-1): ext4_end_bio:314: I/O error -5 writing to ino ..
| EXT4-fs warning (dm-1): ext4_end_bio:314: I/O error -5 writing to ino ..
| EXT4-fs warning (dm-1): ext4_end_bio:314: I/O error -5 writing to ino ..
| JBD2: Detected IO errors while flushing file data on dm-1-8
| Aborting journal on device dm-1-8.
| EXT4-fs error : ext4_journal_check_start:56: Detected aborted journal
| EXT4-fs (dm-1): Remounting filesystem read-only
| EXT4-fs : ext4_writepages: jbd2_start: 2048 pages, inode 498...; err -30
(This did cause corruptions due to failed writes)
To fix this mess, the crypto4xx driver needs to notifiy the
user to slow down. This can be achieved by returning -EBUSY
on requests, once the crypto hardware was falling behind.
Note: -EBUSY has two different meanings. Setting the flag
CRYPTO_TFM_REQ_MAY_BACKLOG implies that the request was
successfully queued, by the crypto driver. To achieve this
requirement, the implementation introduces a threshold check and
adds logic to the completion routines in much the same way as
AMD's Cryptographic Coprocessor (CCP) driver do.
Note2: Tests showed that dm-crypt starved ipsec traffic.
Under load, ipsec links dropped to 0 Kbits/s. This is because
dm-crypt's callback would instantly queue the next request.
In order to not starve ipsec, the driver reserves a small
portion of the available crypto contexts for this purpose.
Signed-off-by: Christian Lamparter <chunkeey@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
---
drivers/crypto/amcc/crypto4xx_core.c | 47 ++++++++++++++++++++++++++++++------
drivers/crypto/amcc/crypto4xx_core.h | 3 ++-
2 files changed, 41 insertions(+), 9 deletions(-)
--- a/drivers/crypto/amcc/crypto4xx_core.c
+++ b/drivers/crypto/amcc/crypto4xx_core.c
@@ -39,6 +39,7 @@
#include <crypto/ctr.h>
#include <crypto/sha.h>
#include <crypto/scatterwalk.h>
+#include <crypto/internal/skcipher.h>
#include "crypto4xx_reg_def.h"
#include "crypto4xx_core.h"
#include "crypto4xx_sa.h"
@@ -573,8 +574,10 @@ static u32 crypto4xx_ablkcipher_done(str
dst->offset, dst->length, DMA_FROM_DEVICE);
}
crypto4xx_ret_sg_desc(dev, pd_uinfo);
- if (ablk_req->base.complete != NULL)
- ablk_req->base.complete(&ablk_req->base, 0);
+
+ if (pd_uinfo->state & PD_ENTRY_BUSY)
+ ablkcipher_request_complete(ablk_req, -EINPROGRESS);
+ ablkcipher_request_complete(ablk_req, 0);
return 0;
}
@@ -591,9 +594,10 @@ static u32 crypto4xx_ahash_done(struct c
crypto4xx_copy_digest_to_dst(pd_uinfo,
crypto_tfm_ctx(ahash_req->base.tfm));
crypto4xx_ret_sg_desc(dev, pd_uinfo);
- /* call user provided callback function x */
- if (ahash_req->base.complete != NULL)
- ahash_req->base.complete(&ahash_req->base, 0);
+
+ if (pd_uinfo->state & PD_ENTRY_BUSY)
+ ahash_request_complete(ahash_req, -EINPROGRESS);
+ ahash_request_complete(ahash_req, 0);
return 0;
}
@@ -704,6 +708,7 @@ u32 crypto4xx_build_pd(struct crypto_asy
struct pd_uinfo *pd_uinfo = NULL;
unsigned int nbytes = datalen, idx;
u32 gd_idx = 0;
+ bool is_busy;
/* figure how many gd is needed */
num_gd = sg_nents_for_len(src, datalen);
@@ -734,6 +739,31 @@ u32 crypto4xx_build_pd(struct crypto_asy
* already got must be return the original place.
*/
spin_lock_irqsave(&dev->core_dev->lock, flags);
+ /*
+ * Let the caller know to slow down, once more than 13/16ths = 81%
+ * of the available data contexts are being used simultaneously.
+ *
+ * With PPC4XX_NUM_PD = 256, this will leave a "backlog queue" for
+ * 31 more contexts. Before new requests have to be rejected.
+ */
+ if (req->flags & CRYPTO_TFM_REQ_MAY_BACKLOG) {
+ is_busy = ((dev->pdr_head - dev->pdr_tail) % PPC4XX_NUM_PD) >=
+ ((PPC4XX_NUM_PD * 13) / 16);
+ } else {
+ /*
+ * To fix contention issues between ipsec (no blacklog) and
+ * dm-crypto (backlog) reserve 32 entries for "no backlog"
+ * data contexts.
+ */
+ is_busy = ((dev->pdr_head - dev->pdr_tail) % PPC4XX_NUM_PD) >=
+ ((PPC4XX_NUM_PD * 15) / 16);
+
+ if (is_busy) {
+ spin_unlock_irqrestore(&dev->core_dev->lock, flags);
+ return -EBUSY;
+ }
+ }
+
if (num_gd) {
fst_gd = crypto4xx_get_n_gd(dev, num_gd);
if (fst_gd == ERING_WAS_FULL) {
@@ -888,11 +918,12 @@ u32 crypto4xx_build_pd(struct crypto_asy
sa->sa_command_1.bf.hash_crypto_offset = 0;
pd->pd_ctl.w = ctx->pd_ctl;
pd->pd_ctl_len.w = 0x00400000 | datalen;
- pd_uinfo->state = PD_ENTRY_INUSE;
+ pd_uinfo->state = PD_ENTRY_INUSE | (is_busy ? PD_ENTRY_BUSY : 0);
+
wmb();
/* write any value to push engine to read a pd */
writel(1, dev->ce_base + CRYPTO4XX_INT_DESCR_RD);
- return -EINPROGRESS;
+ return is_busy ? -EBUSY : -EINPROGRESS;
}
/**
@@ -997,7 +1028,7 @@ static void crypto4xx_bh_tasklet_cb(unsi
tail = core_dev->dev->pdr_tail;
pd_uinfo = &core_dev->dev->pdr_uinfo[tail];
pd = &core_dev->dev->pdr[tail];
- if ((pd_uinfo->state == PD_ENTRY_INUSE) &&
+ if ((pd_uinfo->state & PD_ENTRY_INUSE) &&
pd->pd_ctl.bf.pe_done &&
!pd->pd_ctl.bf.host_ready) {
pd->pd_ctl.bf.pe_done = 0;
--- a/drivers/crypto/amcc/crypto4xx_core.h
+++ b/drivers/crypto/amcc/crypto4xx_core.h
@@ -44,7 +44,8 @@
#define PPC4XX_LAST_SD (PPC4XX_NUM_SD - 1)
#define PPC4XX_SD_BUFFER_SIZE 2048
-#define PD_ENTRY_INUSE 1
+#define PD_ENTRY_BUSY BIT(1)
+#define PD_ENTRY_INUSE BIT(0)
#define PD_ENTRY_FREE 0
#define ERING_WAS_FULL 0xffffffff