On Mon, 13 May 2019 11:41:36 +0200 Cornelia Huck <cohuck@xxxxxxxxxx> wrote: > On Fri, 26 Apr 2019 20:32:41 +0200 > Halil Pasic <pasic@xxxxxxxxxxxxx> wrote: > > > As virtio-ccw devices are channel devices, we need to use the dma area > > for any communication with the hypervisor. > > > > This patch addresses the most basic stuff (mostly what is required for > > virtio-ccw), and does take care of QDIO or any devices. > > "does not take care of QDIO", surely? I did not bother making the QDIO library code use dma memory for anything that is conceptually dma memory. AFAIK QDIO is out of scope for prot virt for now. If one were to do some emulated qdio with prot virt guests, one wound need to make a bunch of things shared. > (Also, what does "any devices" > mean? Do you mean "every arbitrary device", perhaps?) What I mean is: this patch takes care of the core stuff, but any particular device is likely to have to do more -- that is it ain't all the cio devices support prot virt with this patch. For example virtio-ccw needs to make sure that the ccws constituting the channel programs, as well as the data pointed by the ccws is shared. If one would want to make vfio-ccw DASD pass-through work under prot virt, one would need to make sure, that everything that needs to be shared is shared (data buffers, channel programs). Does is clarify things? > > > > > An interesting side effect is that virtio structures are now going to > > get allocated in 31 bit addressable storage. > > Hm... > > > > > Signed-off-by: Halil Pasic <pasic@xxxxxxxxxxxxx> > > --- > > arch/s390/include/asm/ccwdev.h | 4 +++ > > drivers/s390/cio/ccwreq.c | 8 ++--- > > drivers/s390/cio/device.c | 65 +++++++++++++++++++++++++++++++++------- > > drivers/s390/cio/device_fsm.c | 40 ++++++++++++------------- > > drivers/s390/cio/device_id.c | 18 +++++------ > > drivers/s390/cio/device_ops.c | 21 +++++++++++-- > > drivers/s390/cio/device_pgid.c | 20 ++++++------- > > drivers/s390/cio/device_status.c | 24 +++++++-------- > > drivers/s390/cio/io_sch.h | 21 +++++++++---- > > drivers/s390/virtio/virtio_ccw.c | 10 ------- > > 10 files changed, 148 insertions(+), 83 deletions(-) > > (...) > > > diff --git a/drivers/s390/virtio/virtio_ccw.c b/drivers/s390/virtio/virtio_ccw.c > > index 6d989c360f38..bb7a92316fc8 100644 > > --- a/drivers/s390/virtio/virtio_ccw.c > > +++ b/drivers/s390/virtio/virtio_ccw.c > > @@ -66,7 +66,6 @@ struct virtio_ccw_device { > > bool device_lost; > > unsigned int config_ready; > > void *airq_info; > > - u64 dma_mask; > > }; > > > > struct vq_info_block_legacy { > > @@ -1255,16 +1254,7 @@ static int virtio_ccw_online(struct ccw_device *cdev) > > ret = -ENOMEM; > > goto out_free; > > } > > - > > vcdev->vdev.dev.parent = &cdev->dev; > > - cdev->dev.dma_mask = &vcdev->dma_mask; > > - /* we are fine with common virtio infrastructure using 64 bit DMA */ > > - ret = dma_set_mask_and_coherent(&cdev->dev, DMA_BIT_MASK(64)); > > - if (ret) { > > - dev_warn(&cdev->dev, "Failed to enable 64-bit DMA.\n"); > > - goto out_free; > > - } > > This means that vring structures now need to fit into 31 bits as well, > I think? Nod. > Is there any way to reserve the 31 bit restriction for channel > subsystem structures and keep vring in the full 64 bit range? (Or am I > fundamentally misunderstanding something?) > At the root of this problem is that the DMA API basically says devices may have addressing limitations expressed by the dma_mask, while our addressing limitations are not coming from the device but from the IO arch: e.g. orb.cpa and ccw.cda are 31 bit addresses. In our case it depends on how and for what is the device going to use the memory (e.g. buffers addressed by MIDA vs IDA vs direct). Virtio uses the DMA properties of the parent, that is in our case the struct device embedded in struct ccw_device. The previous version (RFC) used to allocate all the cio DMA stuff from this global cio_dma_pool using the css0.dev for the DMA API interactions. And we set *css0.dev.dma_mask == DMA_BIT_MASK(31) so e.g. the allocated ccws are 31 bit addressable. But I was asked to change this so that when I allocate DMA memory for a channel program of particular ccw device, a struct device of that ccw device is used as the first argument of dma_alloc_coherent(). Considering void *dma_alloc_attrs(struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag, unsigned long attrs) { const struct dma_map_ops *ops = get_dma_ops(dev); void *cpu_addr; WARN_ON_ONCE(dev && !dev->coherent_dma_mask); if (dma_alloc_from_dev_coherent(dev, size, dma_handle, &cpu_addr)) return cpu_addr; /* let the implementation decide on the zone to allocate from: */ flag &= ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM); that is the GFP flags dropped that implies that we really want cdev->dev restricted to 31 bit addressable memory because we can't tell (with the current common DMA code) hey but this piece of DMA mem you are abot to allocate for me must be 31 bit addressable (using GFP_DMA as we usually do). So, as described in the commit message, the vring stuff being forced into ZONE_DMA is an unfortunate consequence of this all. A side note: making the subchannel device 'own' the DMA stuff of a ccw device (something that was discussed in the RFC thread) is tricky because the ccw device may outlive the subchannel (all that orphan stuff). So the answer is: it is technically possible (e.g. see RFC) but it comes at a price, and I see no obviously brilliant solution. Regards, Halil > > - > > vcdev->config_block = kzalloc(sizeof(*vcdev->config_block), > > GFP_DMA | GFP_KERNEL); > > if (!vcdev->config_block) { >