From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from linux.microsoft.com (linux.microsoft.com [13.77.154.182]) by mx.groups.io with SMTP id smtpd.web11.4639.1682543214627010442 for ; Wed, 26 Apr 2023 14:06:54 -0700 Authentication-Results: mx.groups.io; dkim=pass header.i=@linux.microsoft.com header.s=default header.b=kJEXkjb9; spf=pass (domain: linux.microsoft.com, ip: 13.77.154.182, mailfrom: osde@linux.microsoft.com) Received: from [10.137.194.171] (unknown [131.107.1.171]) by linux.microsoft.com (Postfix) with ESMTPSA id 2166821C33C7; Wed, 26 Apr 2023 14:06:54 -0700 (PDT) DKIM-Filter: OpenDKIM Filter v2.11.0 linux.microsoft.com 2166821C33C7 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linux.microsoft.com; s=default; t=1682543214; bh=VLPlIki1JK2P+8Zs2+qU38WCItiD8oVEbdfug1XrTgc=; h=Date:Subject:To:Cc:References:From:In-Reply-To:From; b=kJEXkjb9wPSQcj6V398LgS88zXzDGtp/XXSjLmnUvyIfd+yBLwyvnGOtKF0OO2AVl CZZ/bFZ39p/CyeRiLIxvv/TgqxaHybo0tv2AaV0Gix7FUv9T7zp4ml1sHBPQzqJNk4 96YhU8HtBrF23M6qob8/MYv5BwVCaC/9l/PHrpRc= Message-ID: Date: Wed, 26 Apr 2023 14:06:53 -0700 MIME-Version: 1.0 User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:102.0) Gecko/20100101 Thunderbird/102.10.0 Subject: Re: [edk2-devel] [PATCH v1 1/2] Add the volatile keyword to NvmExpressDxe's Passthru CQs and SQs. To: devel@edk2.groups.io, hao.a.wu@intel.com, "Ni, Ray" Cc: "Wang, Jian J" , "Gao, Liming" , Michael Kubacki , Sean Brogan References: <20230419234108.10243-1-osde@linux.microsoft.com> <20230419234108.10243-2-osde@linux.microsoft.com> From: "Oliver Smith-Denny" In-Reply-To: Content-Language: en-US Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 7bit Hi Hao, Thanks for the review! For the Sq, I agree, currently some metadata is read from the queue, but it is not fields that are going to change (such as SGL usage). The thought process there was in case we interact with the HW queue differently in the future. I will drop the Sq change in v2 of this patch. For the Cq, I think the safer option is to mark the whole structure as volatile, because there are other bits that we read out of there that the HW updates, for example the status code here: // // Check the NVMe cmd execution result // if (Status != EFI_TIMEOUT) { if ((Cq->Sct == 0) && (Cq->Sc == 0)) { Status = EFI_SUCCESS; Without the structure marked as volatile, I believe the compiler could optimize the code such that it only reads these metadata fields at the beginning of this function, potentially before they are set by the HW. I do not believe there is much of a performance downside to marking the structure vs individual fields. I am curious to get your feedback here, as well. My goal would be to have a robust solution here so we don't play whack a mole as different compilers make different choices, but obviously without too much overhead :) Thanks, Oliver On 4/25/2023 11:32 PM, Wu, Hao A wrote: > Thanks Oliver, > > For the Submission Queue pointer "Sq", I think it is being used to format the command that will be sent to the NVME controller. > NvmExpressPassThru() does not read back its content for checking after the command gets submitted. > My opinion is that it might be not necessary to add volatile attribute for it. > > For the Completion Queue pointer "Cq", I am not sure which of the following is better: > a) Introduce a volatile pointer to "Cq->Pt", or > b) Mark "Cq" as volatile > Would like to get your feedback on this. Thanks. > > Best Regards, > Hao Wu > >> -----Original Message----- >> From: devel@edk2.groups.io On Behalf Of Oliver >> Smith-Denny >> Sent: Thursday, April 20, 2023 11:48 PM >> To: devel@edk2.groups.io; Ni, Ray >> Cc: Wu, Hao A ; Wang, Jian J ; >> Gao, Liming ; Michael Kubacki >> ; Sean Brogan >> Subject: Re: [edk2-devel] [PATCH v1 1/2] Add the volatile keyword to >> NvmExpressDxe's Passthru CQs and SQs. >> >> Hi Ray, >> >> This is not a pure copy from HW to SW memory, we are also polling the CQ to >> see if a transaction has completed: >> >> // >> // Wait for completion queue to get filled in. >> // >> Status = EFI_TIMEOUT; >> while (EFI_ERROR (gBS->CheckEvent (TimerEvent))) { >> if (Cq->Pt != Private->Pt[QueueId]) { >> Status = EFI_SUCCESS; >> break; >> } >> } >> >> >> What we have seen happen is that without the volatile keyword, the compiler >> can move the Cq->Pt read outside of the loop and only do register compares >> inside the loop, i.e. we end up going the full timeout even if the CQ reports it is >> finished. >> >> Here is the issue that was filed on the project Mu side: >> https://github.com/microsoft/mu_basecore/issues/324. >> >> Thanks, >> Oliver >> >> On 4/19/2023 5:48 PM, Ni, Ray wrote: >>> If it's to copy from hw to sw memory, why do we need volatile? >>> >>> Thanks, >>> Ray >>> >>>> -----Original Message----- >>>> From: devel@edk2.groups.io On Behalf Of Oliver >>>> Smith-Denny >>>> Sent: Thursday, April 20, 2023 7:41 AM >>>> To: devel@edk2.groups.io >>>> Cc: Wu, Hao A ; Ni, Ray ; Wang, >>>> Jian J ; Gao, Liming >>>> ; Michael Kubacki >>>> ; Sean Brogan >>>> >>>> Subject: [edk2-devel][PATCH v1 1/2] Add the volatile keyword to >>>> NvmExpressDxe's Passthru CQs and SQs. >>>> >>>> This updates the relevant functions that expect a non-volatile >>>> >>>> structure to be passed to them to take casts of the CQ and SQ, >>>> >>>> now that they are volatile. >>>> >>>> >>>> >>>> Cc: Hao A Wu >>>> >>>> Cc: Ray Ni >>>> >>>> Cc: Jian J Wang >>>> >>>> Cc: Liming Gao >>>> >>>> Cc: Michael Kubacki >>>> >>>> Cc: Sean Brogan >>>> >>>> Signed-off-by: Oliver Smith-Denny >>>> >>>> --- >>>> >>>> MdeModulePkg/Bus/Pci/NvmExpressDxe/NvmExpressPassthru.c | 10 >>>> +++++----- >>>> >>>> 1 file changed, 5 insertions(+), 5 deletions(-) >>>> >>>> >>>> >>>> diff --git a/MdeModulePkg/Bus/Pci/NvmExpressDxe/NvmExpressPassthru.c >>>> b/MdeModulePkg/Bus/Pci/NvmExpressDxe/NvmExpressPassthru.c >>>> >>>> index f37baa626a16..1a7e39500ac0 100644 >>>> >>>> --- a/MdeModulePkg/Bus/Pci/NvmExpressDxe/NvmExpressPassthru.c >>>> >>>> +++ b/MdeModulePkg/Bus/Pci/NvmExpressDxe/NvmExpressPassthru.c >>>> >>>> @@ -459,8 +459,8 @@ NvmExpressPassThru ( >>>> >>>> EFI_STATUS Status; >>>> >>>> EFI_STATUS PreviousStatus; >>>> >>>> EFI_PCI_IO_PROTOCOL *PciIo; >>>> >>>> - NVME_SQ *Sq; >>>> >>>> - NVME_CQ *Cq; >>>> >>>> + volatile NVME_SQ *Sq; >>>> >>>> + volatile NVME_CQ *Cq; >>>> >>>> UINT16 QueueId; >>>> >>>> UINT16 QueueSize; >>>> >>>> UINT32 Bytes; >>>> >>>> @@ -581,7 +581,7 @@ NvmExpressPassThru ( >>>> >>>> return EFI_INVALID_PARAMETER; >>>> >>>> } >>>> >>>> >>>> >>>> - ZeroMem (Sq, sizeof (NVME_SQ)); >>>> >>>> + ZeroMem ((VOID *)Sq, sizeof (NVME_SQ)); >>>> >>>> Sq->Opc = (UINT8)Packet->NvmeCmd->Cdw0.Opcode; >>>> >>>> Sq->Fuse = (UINT8)Packet->NvmeCmd->Cdw0.FusedOperation; >>>> >>>> Sq->Cid = Private->Cid[QueueId]++; >>>> >>>> @@ -815,14 +815,14 @@ NvmExpressPassThru ( >>>> >>>> // Dump every completion entry status for debugging. >>>> >>>> // >>>> >>>> DEBUG_CODE_BEGIN (); >>>> >>>> - NvmeDumpStatus (Cq); >>>> >>>> + NvmeDumpStatus ((NVME_CQ *)Cq); >>>> >>>> DEBUG_CODE_END (); >>>> >>>> } >>>> >>>> >>>> >>>> // >>>> >>>> // Copy the Respose Queue entry for this command to the callers >>>> response buffer >>>> >>>> // >>>> >>>> - CopyMem (Packet->NvmeCompletion, Cq, sizeof >>>> (EFI_NVM_EXPRESS_COMPLETION)); >>>> >>>> + CopyMem (Packet->NvmeCompletion, (VOID *)Cq, sizeof >>>> (EFI_NVM_EXPRESS_COMPLETION)); >>>> >>>> } else { >>>> >>>> // >>>> >>>> // Timeout occurs for an NVMe command. Reset the controller to >>>> abort the >>>> >>>> -- >>>> >>>> 2.39.2 >>>> >>>> >>>> >>>> >>>> >>>> -=-=-=-=-=-= >>>> Groups.io Links: You receive all messages sent to this group. >>>> View/Reply Online (#103263): >>>> https://edk2.groups.io/g/devel/message/103263 >>>> Mute This Topic: https://groups.io/mt/98378948/1712937 >>>> Group Owner: devel+owner@edk2.groups.io >>>> Unsubscribe: https://edk2.groups.io/g/devel/unsub [ray.ni@intel.com] >>>> -=-=-=-=-=-= >>>> >>> >>> >>> >>> >>> >> >> >> >> > > > > > >