On 2022-03-07 10:48, Corentin Labbe wrote:
Le Mon, Mar 07, 2022 at 09:59:29AM +0200, Gilad Ben-Yossef a �crit :On Sun, Mar 6, 2022 at 11:49 PM Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote:On Fri, Mar 04, 2022 at 02:30:06PM +0100, Corentin Labbe wrote:Hello I got: [ 17.563793] ------------[ cut here ]------------ [ 17.568492] DMA-API: ccree e6601000.crypto: device driver frees DMA memory with different direction [device address=0x0000000078fe5800] [size=8 bytes] [mapped with DMA_TO_DEVICE] [unmapped with DMA_BIDIRECTIONAL]The direction argument during unmap must match whatever direction you used during the original map call.Yes, of course. I changed one but forgot the other. Corentin, could you be kind and check that this solves the original problem and does not produce new warnings? diff --git a/drivers/crypto/ccree/cc_buffer_mgr.c b/drivers/crypto/ccree/cc_buffer_mgr.c index 11e0278c8631..31cfe014922e 100644 --- a/drivers/crypto/ccree/cc_buffer_mgr.c +++ b/drivers/crypto/ccree/cc_buffer_mgr.c @@ -356,12 +356,14 @@ void cc_unmap_cipher_request(struct device *dev, void *ctx, req_ctx->mlli_params.mlli_dma_addr); } - dma_unmap_sg(dev, src, req_ctx->in_nents, DMA_BIDIRECTIONAL); - dev_dbg(dev, "Unmapped req->src=%pK\n", sg_virt(src)); - if (src != dst) { - dma_unmap_sg(dev, dst, req_ctx->out_nents, DMA_BIDIRECTIONAL); + dma_unmap_sg(dev, src, req_ctx->in_nents, DMA_TO_DEVICE); + dma_unmap_sg(dev, dst, req_ctx->out_nents, DMA_FROM_DEVICE); dev_dbg(dev, "Unmapped req->dst=%pK\n", sg_virt(dst)); + dev_dbg(dev, "Unmapped req->src=%pK\n", sg_virt(src)); + } else { + dma_unmap_sg(dev, src, req_ctx->in_nents, DMA_BIDIRECTIONAL); + dev_dbg(dev, "Unmapped req->src=%pK\n", sg_virt(src)); } } @@ -377,6 +379,7 @@ int cc_map_cipher_request(struct cc_drvdata *drvdata, void *ctx, u32 dummy = 0; int rc = 0; u32 mapped_nents = 0; + int src_direction = (src != dst ? DMA_TO_DEVICE : DMA_BIDIRECTIONAL); req_ctx->dma_buf_type = CC_DMA_BUF_DLLI; mlli_params->curr_pool = NULL; @@ -399,7 +402,7 @@ int cc_map_cipher_request(struct cc_drvdata *drvdata, void *ctx, } /* Map the src SGL */ - rc = cc_map_sg(dev, src, nbytes, DMA_BIDIRECTIONAL, &req_ctx->in_nents, + rc = cc_map_sg(dev, src, nbytes, src_direction, &req_ctx->in_nents, LLI_MAX_NUM_OF_DATA_ENTRIES, &dummy, &mapped_nents); if (rc) goto cipher_exit; @@ -416,7 +419,7 @@ int cc_map_cipher_request(struct cc_drvdata *drvdata, void *ctx, } } else { /* Map the dst sg */ - rc = cc_map_sg(dev, dst, nbytes, DMA_BIDIRECTIONAL, + rc = cc_map_sg(dev, dst, nbytes, DMA_FROM_DEVICE, &req_ctx->out_nents, LLI_MAX_NUM_OF_DATA_ENTRIES, &dummy, &mapped_nents); if (rc)Hello I still get the warning: [ 433.406230] ------------[ cut here ]------------ [ 433.406326] DMA-API: ccree e6601000.crypto: cacheline tracking EEXIST, overlapping mappings aren't supported [ 433.406386] WARNING: CPU: 7 PID: 31074 at /home/clabbe/linux-next/kernel/dma/debug.c:571 add_dma_entry+0x1d0/0x288 [ 433.406434] Modules linked in: [ 433.406458] CPU: 7 PID: 31074 Comm: kcapi Not tainted 5.17.0-rc6-next-20220303-00130-g30042e47ee47-dirty #54 [ 433.406473] Hardware name: Renesas Salvator-X board based on r8a77950 (DT) [ 433.406484] pstate: 40000005 (nZcv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--) [ 433.406498] pc : add_dma_entry+0x1d0/0x288 [ 433.406510] lr : add_dma_entry+0x1d0/0x288 [ 433.406522] sp : ffff800015da3690 [ 433.406531] x29: ffff800015da3690 x28: 0000000000000000 x27: 0000000000000000 [ 433.406562] x26: 0000000000000000 x25: ffff80000b4c7bc0 x24: ffff80000b4c7000 [ 433.406593] x23: 0000000000000000 x22: 00000000ffffffef x21: ffff80000a9b6000 [ 433.406623] x20: ffff0004c0af5c00 x19: ffff80000b420000 x18: ffffffffffffffff [ 433.406653] x17: 6c7265766f202c54 x16: 534958454520676e x15: 000000000000022e [ 433.406683] x14: ffff800015da3380 x13: 00000000ffffffea x12: ffff80000b4be010 [ 433.406713] x11: 0000000000000001 x10: 0000000000000001 x9 : ffff80000b4a6028 [ 433.406743] x8 : c0000000ffffefff x7 : 0000000000017fe8 x6 : ffff80000b4a5fd0 [ 433.406773] x5 : ffff0006ff795c48 x4 : 0000000000000000 x3 : 0000000000000027 [ 433.406802] x2 : 0000000000000023 x1 : 8ca4e4fbf4b87900 x0 : 0000000000000000 [ 433.406833] Call trace: [ 433.406841] add_dma_entry+0x1d0/0x288 [ 433.406854] debug_dma_map_sg+0x150/0x398 [ 433.406869] __dma_map_sg_attrs+0x9c/0x108 [ 433.406889] dma_map_sg_attrs+0x10/0x28 [ 433.406904] cc_map_sg+0x80/0x100 [ 433.406924] cc_map_cipher_request+0x178/0x3c8 [ 433.406939] cc_cipher_process+0x210/0xb58 [ 433.406953] cc_cipher_encrypt+0x2c/0x38 [ 433.406967] crypto_skcipher_encrypt+0x44/0x78 [ 433.406986] skcipher_recvmsg+0x36c/0x420 [ 433.407003] ____sys_recvmsg+0x90/0x280 [ 433.407024] ___sys_recvmsg+0x88/0xd0 [ 433.407038] __sys_recvmsg+0x6c/0xd0 [ 433.407049] __arm64_sys_recvmsg+0x24/0x30 [ 433.407061] invoke_syscall+0x44/0x100 [ 433.407082] el0_svc_common.constprop.3+0x90/0x120 [ 433.407096] do_el0_svc+0x24/0x88 [ 433.407110] el0_svc+0x4c/0x100 [ 433.407131] el0t_64_sync_handler+0x90/0xb8 [ 433.407145] el0t_64_sync+0x170/0x174 [ 433.407160] irq event stamp: 5624 [ 433.407168] hardirqs last enabled at (5623): [<ffff80000812f6a8>] __up_console_sem+0x60/0x98 [ 433.407191] hardirqs last disabled at (5624): [<ffff800009c9a060>] el1_dbg+0x28/0x90 [ 433.407208] softirqs last enabled at (5570): [<ffff8000097e62f8>] lock_sock_nested+0x80/0xa0 [ 433.407226] softirqs last disabled at (5568): [<ffff8000097e62d8>] lock_sock_nested+0x60/0xa0 [ 433.407241] ---[ end trace 0000000000000000 ]--- [ 433.407381] DMA-API: Mapped at: [ 433.407396] debug_dma_map_sg+0x16c/0x398 [ 433.407416] __dma_map_sg_attrs+0x9c/0x108 [ 433.407436] dma_map_sg_attrs+0x10/0x28 [ 433.407455] cc_map_sg+0x80/0x100 [ 433.407475] cc_map_cipher_request+0x178/0x3c8 BUT I start to thing this is a bug in DMA-API debug. My sun8i-ss driver hit the same warning: [ 142.458351] WARNING: CPU: 1 PID: 90 at kernel/dma/debug.c:597 add_dma_entry+0x2ec/0x4cc [ 142.458429] DMA-API: sun8i-ss 1c15000.crypto: cacheline tracking EEXIST, overlapping mappings aren't supported [ 142.458455] Modules linked in: ccm algif_aead xts cmac [ 142.458563] CPU: 1 PID: 90 Comm: 1c15000.crypto- Not tainted 5.17.0-rc6-next-20220307-00132-g39dad568d20a-dirty #223 [ 142.458581] Hardware name: Allwinner A83t board [ 142.458596] unwind_backtrace from show_stack+0x10/0x14 [ 142.458627] show_stack from 0xf0abdd1c [ 142.458646] irq event stamp: 31747 [ 142.458660] hardirqs last enabled at (31753): [<c019316c>] __up_console_sem+0x50/0x60 [ 142.458688] hardirqs last disabled at (31758): [<c0193158>] __up_console_sem+0x3c/0x60 [ 142.458710] softirqs last enabled at (31600): [<c06990c8>] sun8i_ss_handle_cipher_request+0x300/0x8b8 [ 142.458738] softirqs last disabled at (31580): [<c06990c8>] sun8i_ss_handle_cipher_request+0x300/0x8b8 [ 142.458758] ---[ end trace 0000000000000000 ]--- [ 142.458771] DMA-API: Mapped at: Yes the mapped at is empty just after. And the sequence of DMA operations in my driver is simple, so I cannot see how any overlap could occur.
The "overlap" is in the sense of having more than one mapping within the same cacheline:
[ 142.458120] DMA-API: add_dma_entry start P=ba79f200 N=ba79f D=ba79f200 L=10 DMA_FROM_DEVICE attrs=0 [ 142.458156] DMA-API: add_dma_entry start P=445dc010 N=445dc D=445dc010 L=10 DMA_TO_DEVICE attrs=0
[ 142.458178] sun8i-ss 1c15000.crypto: SRC 0/1/1 445dc000 len=16 bi=0 [ 142.458215] sun8i-ss 1c15000.crypto: DST 0/1/1 ba79f200 len=16 bi=0[ 142.458234] DMA-API: add_dma_entry start P=ba79f210 N=ba79f D=ba79f210 L=10 DMA_FROM_DEVICE attrs=0
This actually illustrates exactly the reason why this is unsupportable. ba79f200 is mapped for DMA_FROM_DEVICE, therefore subsequently mapping ba79f210 for DMA_TO_DEVICE may cause the cacheline covering the range ba79f200-ba79f23f to be written back over the top of data that the device has already started to write to memory. Hello data corruption.
Separate DMA mappings should be from separate memory allocations, respecting ARCH_DMA_MINALIGN.
Thanks, Robin.
