From mboxrd@z Thu Jan 1 00:00:00 1970 From: Yuanhan Liu Subject: Re: [ [PATCH v2] 05/13] virtio: change io_base datatype from uint32_t to uint64_type Date: Thu, 17 Dec 2015 15:19:27 +0800 Message-ID: <20151217071927.GA29571@yliu-dev.sh.intel.com> References: <1450098032-21198-1-git-send-email-sshukla@mvista.com> <1450098032-21198-6-git-send-email-sshukla@mvista.com> <20151216134850.GU29571@yliu-dev.sh.intel.com> <20151216142326.GV29571@yliu-dev.sh.intel.com> <20151216145846.GX29571@yliu-dev.sh.intel.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: dev@dpdk.org To: Santosh Shukla Return-path: Received: from mga09.intel.com (mga09.intel.com [134.134.136.24]) by dpdk.org (Postfix) with ESMTP id 1C2DA68A5 for ; Thu, 17 Dec 2015 08:19:07 +0100 (CET) Content-Disposition: inline In-Reply-To: List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" On Wed, Dec 16, 2015 at 08:35:58PM +0530, Santosh Shukla wrote: > On Wed, Dec 16, 2015 at 8:28 PM, Yuanhan Liu > wrote: > > On Wed, Dec 16, 2015 at 08:09:40PM +0530, Santosh Shukla wrote: > >> On Wed, Dec 16, 2015 at 7:53 PM, Yuanhan Liu > >> wrote: > >> > On Wed, Dec 16, 2015 at 07:31:57PM +0530, Santosh Shukla wrote: > >> >> On Wed, Dec 16, 2015 at 7:18 PM, Yuanhan Liu > >> >> wrote: > >> >> > On Mon, Dec 14, 2015 at 06:30:24PM +0530, Santosh Shukla wrote: > >> >> >> In x86 case io_base to store ioport address not more than 65535 ioports. i.e..0 > >> >> >> to ffff but in non-x86 case in particular arm64 it need to store more than 32 > >> >> >> bit address so changing io_base datatype from 32 to 64. > >> >> >> > >> >> >> Signed-off-by: Santosh Shukla > >> >> >> --- > >> >> >> drivers/net/virtio/virtio_ethdev.c | 2 +- > >> >> >> drivers/net/virtio/virtio_pci.h | 4 ++-- > >> >> >> 2 files changed, 3 insertions(+), 3 deletions(-) > >> >> >> > >> >> >> diff --git a/drivers/net/virtio/virtio_ethdev.c b/drivers/net/virtio/virtio_ethdev.c > >> >> >> index d928339..620e0d4 100644 > >> >> >> --- a/drivers/net/virtio/virtio_ethdev.c > >> >> >> +++ b/drivers/net/virtio/virtio_ethdev.c > >> >> >> @@ -1291,7 +1291,7 @@ eth_virtio_dev_init(struct rte_eth_dev *eth_dev) > >> >> >> return -1; > >> >> >> > >> >> >> hw->use_msix = virtio_has_msix(&pci_dev->addr); > >> >> >> - hw->io_base = (uint32_t)(uintptr_t)pci_dev->mem_resource[0].addr; > >> >> >> + hw->io_base = (uint64_t)(uintptr_t)pci_dev->mem_resource[0].addr; > >> >> > > >> >> > I'd suggest to move the io_base assignment (and cast) into virtio_ioport_init() > >> >> > so that we could do the correct cast there, say cast it to uint32_t for > >> >> > X86, and uint64_t for others. > >> >> > > >> >> > >> >> Ok. > >> >> > >> >> This was deliberately done considering your 1.0 virtio spec patch do > >> >> care for uint64_t types and in arm64 case, If I plan to use those > >> >> future patches, IMO it make more sense to me keep it in uint64_t way; > >> > > >> > I did different cast, 32 bit for legacy virtio pci device, and 64 bit > >> > for modern virtio pci device. > >> > > >> >> Also in x86 case max address could of type 0x1000-101f and so forth; > >> >> changing data-type to uint64_t default wont effect such address, > >> >> right? > >> > > >> > Right, but what's the harm of doing the right cast? :) > >> > > >> > >> Agree. > >> > >> >> And hw->io_base by looking at virtio_pci.h function like > >> >> inb/outb etc.. takes io_base address as unsigned long types which is > >> >> arch dependent; i.e.. 4 byte for 32 bit and 8 for 64 bit so the lower > >> >> level rd/wr apis are taking care of data-types accordingly. > >> > > >> > Didn't get it. inb/outb takes "unsigned short" arguments, but not > >> > "unsigned long". > >> > > >> > >> sys/io.h in x86 case using unsigned short int types.. > >> > >> include/asm-generic/io.h for arm64 using it unsigned long (from linux > >> header files) > >> > >> In such case keeping > >> #define VIRTIO_PCI_REG_ADDR(hw, reg) \ > >> (unsigned short)((hw)->io_base + (reg)) > >> > >> would be x86 specific and what I thought and used in this patch is > >> > >> #define VIRTIO_PCI_REG_ADDR(hw, reg) \ > >> (unsigned long)((hw)->io_base + (reg)) > >> > >> to avoid ifdef ARM or non-x86..clutter, I know data-type is not right > >> fit for x86 sys/io.h but considering possible address inside > >> hw->io_base, wont effect functionality and performance my any mean. > >> That is why at virtio_ethdev_init() i choose to keep it in hw->io_base > >> = (uint64_t) types. > >> > >> Otherwise I'll have to duplicate VIRTIO_PCI_REG_XXX definition for > >> non-x86 case, Pl. suggest better alternative. Thanks > > > > > > My understanding is that if you have done the right cast in the first > > time (at the io_base assignment), casting from a short type to a longer > > type will not matter: the upper bits will be filled with zero. > > > > So, I guess we are fine here. I'm thinking that the extra cast in > > VIRTIO_PCI_REG_ADDR() is not necessary, as C will do the right > > cast for different inb(), say cast it to "unsigned short" for x86, > > and "unsigned long" for your arm implementation. The same to > > other io helpers. > > > > so to summarize and correct me if i misunderstood, > keep hw->io_base = (uint64_t) I still want a different explicit cast for x86 and non-x86. And actually, we should cast it to (unsigned short) but not (uint32_t) for x86, don't we? On the other hand, we may cast it to uint64_t unconditionally, and then have an explicit sanity check for io_base for x86, say if ((unsigned short)hw->io_base != hw->io_base) { PMD_INIT_LOG(ERR, "invalid io port: %"PRIx64, ...); return -1; } It's better than the (unsigned short) cast, as the later simply hides issue when something went wrong, though it's not rare. What do you think of that? > and remove extra cast {i.e.. (unsigned short) for x86 or (unsigned > long) for non-x86/arm64 case} in VIRTIO_PCI_REG_ADDR(). Technically speaking, yes, we don't need this kind of cast. --yliu