From mboxrd@z Thu Jan 1 00:00:00 1970 From: Tom Lendacky Subject: Re: [RFC PATCH v2 06/32] x86/pci: Use memremap when walking setup data Date: Fri, 3 Mar 2017 15:15:34 -0600 Message-ID: References: <148846752022.2349.13667498174822419498.stgit@brijesh-build-machine> <148846759008.2349.8274808454274279039.stgit@brijesh-build-machine> <20170303204209.GA31767@bhelgaas-glaptop.roam.corp.google.com> Mime-Version: 1.0 Content-Type: text/plain; charset="windows-1252"; format=flowed Content-Transfer-Encoding: 7bit Cc: , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , To: Bjorn Helgaas , Brijesh Singh Return-path: In-Reply-To: <20170303204209.GA31767@bhelgaas-glaptop.roam.corp.google.com> Sender: owner-linux-mm@kvack.org List-Id: linux-crypto.vger.kernel.org On 3/3/2017 2:42 PM, Bjorn Helgaas wrote: > On Thu, Mar 02, 2017 at 10:13:10AM -0500, Brijesh Singh wrote: >> From: Tom Lendacky >> >> The use of ioremap will force the setup data to be mapped decrypted even >> though setup data is encrypted. Switch to using memremap which will be >> able to perform the proper mapping. > > How should callers decide whether to use ioremap() or memremap()? > > memremap() existed before SME and SEV, and this code is used even if > SME and SEV aren't supported, so the rationale for this change should > not need the decryption argument. When SME or SEV is active an ioremap() will remove the encryption bit from the pagetable entry when it is mapped. This allows MMIO, which doesn't support SME/SEV, to be performed successfully. So my take is that ioremap() should be used for MMIO and memremap() for pages in RAM. > >> Signed-off-by: Tom Lendacky >> --- >> arch/x86/pci/common.c | 4 ++-- >> 1 file changed, 2 insertions(+), 2 deletions(-) >> >> diff --git a/arch/x86/pci/common.c b/arch/x86/pci/common.c >> index a4fdfa7..0b06670 100644 >> --- a/arch/x86/pci/common.c >> +++ b/arch/x86/pci/common.c >> @@ -691,7 +691,7 @@ int pcibios_add_device(struct pci_dev *dev) >> >> pa_data = boot_params.hdr.setup_data; >> while (pa_data) { >> - data = ioremap(pa_data, sizeof(*rom)); >> + data = memremap(pa_data, sizeof(*rom), MEMREMAP_WB); > > I can't quite connect the dots here. ioremap() on x86 would do > ioremap_nocache(). memremap(MEMREMAP_WB) would do arch_memremap_wb(), > which is ioremap_cache(). Is making a cacheable mapping the important > difference? The memremap(MEMREMAP_WB) will actually check to see if it can perform a __va(pa_data) in try_ram_remap() and then fallback to the arch_memremap_wb(). So it's actually the __va() vs the ioremap_cache() that is the difference. Thanks, Tom > >> if (!data) >> return -ENOMEM; >> >> @@ -710,7 +710,7 @@ int pcibios_add_device(struct pci_dev *dev) >> } >> } >> pa_data = data->next; >> - iounmap(data); >> + memunmap(data); >> } >> set_dma_domain_ops(dev); >> set_dev_domain_options(dev); >> -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: email@kvack.org