fragment transmission while retrieving memory - Hafnium

17 Apr 2023

Hi, Jens, Olivier,
update for optee:
https://github.com/OP-TEE/optee_os/pull/5966 <https://github.com/OP-TEE/optee_os/pull/5966 >
update for hafnium:

diff --git a/src/api.c b/src/api.c
index 5f699478..7e43e305 100644
--- a/src/api.c
+++ b/src/api.c
@@ -3191,7 +3191,8 @@ struct ffa_value api_ffa_mem_frag_rx(ffa_memory_handle_t handle,
 return ffa_error(FFA_INVALID_PARAMETERS);
 }
 }
-
+ if (to->id==0x8001 && to->mailbox.state != MAILBOX_STATE_EMPTY)
+ to->mailbox.state = MAILBOX_STATE_EMPTY;
 to_locked = vm_lock(to);
 if (vm_is_mailbox_busy(to_locked)) {
diff --git a/src/ffa_memory.c b/src/ffa_memory.c
index 3e978d23..59e71ff3 100644
--- a/src/ffa_memory.c
+++ b/src/ffa_memory.c
@@ -2385,7 +2385,7 @@ struct ffa_value ffa_memory_retrieve_continue(struct vm_locked to_locked,
 &fragment_length);
 CHECK(remaining_constituent_count == 0);
 to_locked.vm->mailbox.recv_size = fragment_length;
- to_locked.vm->mailbox.recv_sender = HF_HYPERVISOR_VM_ID;
+ to_locked.vm->mailbox.recv_sender = to_locked.vm->id;
 to_locked.vm->mailbox.recv_func = FFA_MEM_FRAG_TX_32;
 to_locked.vm->mailbox.state = MAILBOX_STATE_FULL;
It seems that to->mailbox.state needs to be set to MAILBOX_STATE_EMPTY firstly. 
I have not figured out why, but large TA could be loaded successfully with above fix now.
Regards,
Yuye.
------------------------------------------------------------------
发件人：梅建强(禹夜) meijianqiang.mjq@alibaba-inc.com
发送时间：2023年4月17日(星期一) 02:16
收件人：Jens Wiklander jens.wiklander@linaro.org
抄　送：Olivier Deprez Olivier.Deprez@arm.com; hafnium hafnium@lists.trustedfirmware.org; op-tee op-tee@lists.trustedfirmware.org
主　题：fragment transmission while retrieving memory
Hi, Jens,
I have implement the missing support for fragmented mem retrieve response as you suggested.
I will push fix into the community code soon and please help review it later.
Thanks.
Regards,
Yuye.
------------------------------------------------------------------
发件人：Jens Wiklander jens.wiklander@linaro.org
发送时间：2023年4月14日(星期五) 14:23
收件人：梅建强(禹夜) meijianqiang.mjq@alibaba-inc.com
抄　送：Olivier Deprez Olivier.Deprez@arm.com; hafnium hafnium@lists.trustedfirmware.org; op-tee op-tee@lists.trustedfirmware.org
主　题：Re: fragment transmission while retrieving memory
Hi Yuye,
On Fri, Apr 14, 2023 at 4:29 AM 梅建强(禹夜)
meijianqiang.mjq@alibaba-inc.com wrote:
...
Hi, Jens, Olivier,
If I understand properly, for this specific issue both optee and hafnium need to do further fix.
The following discussion and questions are based on figure 18.4 in the FFA spec.
As I saw in the optee code, the retrieve mechanism would start with this code:
retrieve_desc = spmc_retrieve_req(cookie);
And in the function spmc_retrieve_req,
optee sets the total length to the same size as fragment0 length before sending FFA_MEM_FRAG_REQ to hafnium,
which means hafnium will only retrieve the fragment0 memory region.
In the example process,
optee will determines the number of TX sized fragments of descriptor before sending FFA_MEM_FRAG_REQ to hafnium.
I have a question here.
How should optee calculate this quantity?
Hafnium then receives the FFA_MEM_FRAG_REQ.
As I saw in the hafnium code, hafnium blocks the case that total length>fragment length.
if (fragment_length != length) {
dlog_verbose("Fragmentation not yet supported.\n");
return ffa_error(FFA_INVALID_PARAMETERS);
}
Obviously it is not what we expected.
Then following the example process, hafnium should allocate handle and use it to associate fragments.
I didn't find the corresponding implementation in hafnium code for this step.
And I want to know how to implement the associate action here.
After that, optee lacks the implementation of FFA_MEM_FRAG_RX and FFA_MEM_FRAG_TX ABI in CFG_CORE_SEL2_SPMC enable case.
Is there any support plan for the implementation of fragmented memory retrieve in the optee community?
Does the hafnium community have a plan to implement it cooperatively?
If I've understood this correctly we have two different cases of
eventual fragmentation of descriptors. The first is the descriptor
passed with FFA_MEM_RETRIEVE_REQ_32, this one is not likely to need
fragmentation support soon. The other is the descriptor received with
FFA_MEM_RETRIEVE_RESP, I believe this is where you've run into
trouble.
From OP-TEE point of view, it would make sense if you took lead on
implementing this. I can help with review etc.
Thanks,
Jens
...
Thanks a lot for your support.
Regards,
Yuye.

发件人：Olivier Deprez Olivier.Deprez@arm.com
发送时间：2023年4月13日(星期四) 15:52
收件人：梅建强(禹夜) meijianqiang.mjq@alibaba-inc.com
抄 送：hafnium hafnium@lists.trustedfirmware.org; Jens Wiklander jens.wiklander@linaro.org
主 题：Re: fragment transmission while retrieving memory
Hi Yuye, Jens,
For the record, and If I understand properly the last comment in the github issue:
OP-TEE is missing the implementation for receipt of a fragmented retrieve response
There is no further fix to be done in Hafnium for this specific issue at this moment, please let me know otherwise.
Thanks, Olivier.

From: 梅建强(禹夜) meijianqiang.mjq@alibaba-inc.com
Sent: 12 April 2023 09:27
To: Olivier Deprez Olivier.Deprez@arm.com
Cc: hafnium hafnium@lists.trustedfirmware.org; Jens Wiklander jens.wiklander@linaro.org
Subject: Re: fragment transmission while retrieving memory
Hi, Olivier,
In our setup,
Hafnium commit: 997476a74571aec4f1a23590d45edf516f3934f4
optee version: 3.20.0
Thanks.
Regards,
Yuye.

发件人：Olivier Deprez Olivier.Deprez@arm.com
发送时间：2023年4月12日(星期三) 15:22
收件人：梅建强(禹夜) meijianqiang.mjq@alibaba-inc.com
抄 送：hafnium hafnium@lists.trustedfirmware.org; Jens Wiklander jens.wiklander@linaro.org
主 题：Re: fragment transmission while retrieving memory
Hi Yuye,
AFAIK concerning Hafnium, fragmented mem sharing is supported for FFA_MEM_SHARE/LEND/DONATE and FFA_MEM_RETRIEVE_RESP.
(For the sake of clarity, this isn't supported for FFA_MEM_RETRIEVE_REQ but this shouldn't be a concern as this limitation only exists in the case of mem sharing to multiple borrowers. In your case of a single borrower the mem retrieve req. shouldn't have to be fragmented).
Can you tell which hafnium commit hash is used in your setup?
At the moment, I cannot tell if the issue described concern a miss in Hafnium or OP-TEE.
I need to dig a bit further into both implementations and I'll let you know.
Regards,
Olivier.

From: 梅建强(禹夜) meijianqiang.mjq@alibaba-inc.com
Sent: 12 April 2023 08:21
To: Olivier Deprez Olivier.Deprez@arm.com
Cc: hafnium hafnium@lists.trustedfirmware.org; Jens Wiklander jens.wiklander@linaro.org
Subject: fragment transmission while retrieving memory
Hi, Olivier,
Recently, I've been working on this issue.
https://github.com/OP-TEE/optee_os/issues/5943 <https://github.com/OP-TEE/optee_os/issues/5943 >
Do you know any differences between hafnium's current implementation of fragment transmission while retrieving memory and the example process described in FF-A 1.1 Figure 18.4?
Regards,
Yuye.

    