Re: [Xen-devel] [PATCH 4/7] xen/9pfs: connect to the backend
On Wed, 8 Mar 2017, Julien Grall wrote: > Hi Stefano, > > On 08/03/17 00:49, Stefano Stabellini wrote: > > On Tue, 7 Mar 2017, Julien Grall wrote: > > > On 03/06/2017 08:01 PM, Stefano Stabellini wrote: > > > > + if (ring->bytes == NULL) > > > > + goto error; > > > > + for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) > > > > + ring->intf->ref[i] = > > > > gnttab_grant_foreign_access(dev->otherend_id, > > > > pfn_to_gfn(virt_to_pfn((void*)ring->bytes) + i), 0);. > > > > > > Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). > > > > OK > > > > > > > Also, this is not going to work on 64K kernel because you will grant > > > access to > > > noncontiguous memory (e.g 0-4K, 64K-68K,...). > > > > By using virt_to_gfn like you suggested, the loop will correctly iterate > > on a 4K by 4K basis, even on a 64K kernel: > > > > ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, > > XEN_9PFS_RING_ORDER - (PAGE_SHIFT - XEN_PAGE_SHIFT)); > > for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) > > ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, > > virt_to_gfn((void*)ring->bytes) + i, 0); > > BTW, the cast (void *) is not necessary. > > > > > where XEN_9PFS_RING_ORDER specifies the order at 4K granularity. Am I > > missing something? > > I think it is fine. You could move virt_to_gfn(...) outside and avoid to do > the translation everytime. All right, thanks.
Re: [Xen-devel] [PATCH 4/7] xen/9pfs: connect to the backend
On Wed, 8 Mar 2017, Julien Grall wrote: > Hi Stefano, > > On 08/03/17 00:49, Stefano Stabellini wrote: > > On Tue, 7 Mar 2017, Julien Grall wrote: > > > On 03/06/2017 08:01 PM, Stefano Stabellini wrote: > > > > + if (ring->bytes == NULL) > > > > + goto error; > > > > + for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) > > > > + ring->intf->ref[i] = > > > > gnttab_grant_foreign_access(dev->otherend_id, > > > > pfn_to_gfn(virt_to_pfn((void*)ring->bytes) + i), 0);. > > > > > > Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). > > > > OK > > > > > > > Also, this is not going to work on 64K kernel because you will grant > > > access to > > > noncontiguous memory (e.g 0-4K, 64K-68K,...). > > > > By using virt_to_gfn like you suggested, the loop will correctly iterate > > on a 4K by 4K basis, even on a 64K kernel: > > > > ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, > > XEN_9PFS_RING_ORDER - (PAGE_SHIFT - XEN_PAGE_SHIFT)); > > for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) > > ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, > > virt_to_gfn((void*)ring->bytes) + i, 0); > > BTW, the cast (void *) is not necessary. > > > > > where XEN_9PFS_RING_ORDER specifies the order at 4K granularity. Am I > > missing something? > > I think it is fine. You could move virt_to_gfn(...) outside and avoid to do > the translation everytime. All right, thanks.
Re: [Xen-devel] [PATCH 4/7] xen/9pfs: connect to the backend
Hi Stefano, On 08/03/17 00:49, Stefano Stabellini wrote: On Tue, 7 Mar 2017, Julien Grall wrote: On 03/06/2017 08:01 PM, Stefano Stabellini wrote: + if (ring->bytes == NULL) + goto error; + for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) + ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, pfn_to_gfn(virt_to_pfn((void*)ring->bytes) + i), 0);. Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). OK Also, this is not going to work on 64K kernel because you will grant access to noncontiguous memory (e.g 0-4K, 64K-68K,...). By using virt_to_gfn like you suggested, the loop will correctly iterate on a 4K by 4K basis, even on a 64K kernel: ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, XEN_9PFS_RING_ORDER - (PAGE_SHIFT - XEN_PAGE_SHIFT)); for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, virt_to_gfn((void*)ring->bytes) + i, 0); BTW, the cast (void *) is not necessary. where XEN_9PFS_RING_ORDER specifies the order at 4K granularity. Am I missing something? I think it is fine. You could move virt_to_gfn(...) outside and avoid to do the translation everytime. Cheers, -- Julien Grall
Re: [Xen-devel] [PATCH 4/7] xen/9pfs: connect to the backend
Hi Stefano, On 08/03/17 00:49, Stefano Stabellini wrote: On Tue, 7 Mar 2017, Julien Grall wrote: On 03/06/2017 08:01 PM, Stefano Stabellini wrote: + if (ring->bytes == NULL) + goto error; + for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) + ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, pfn_to_gfn(virt_to_pfn((void*)ring->bytes) + i), 0);. Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). OK Also, this is not going to work on 64K kernel because you will grant access to noncontiguous memory (e.g 0-4K, 64K-68K,...). By using virt_to_gfn like you suggested, the loop will correctly iterate on a 4K by 4K basis, even on a 64K kernel: ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, XEN_9PFS_RING_ORDER - (PAGE_SHIFT - XEN_PAGE_SHIFT)); for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, virt_to_gfn((void*)ring->bytes) + i, 0); BTW, the cast (void *) is not necessary. where XEN_9PFS_RING_ORDER specifies the order at 4K granularity. Am I missing something? I think it is fine. You could move virt_to_gfn(...) outside and avoid to do the translation everytime. Cheers, -- Julien Grall
Re: [Xen-devel] [PATCH 4/7] xen/9pfs: connect to the backend
On Tue, 7 Mar 2017, Julien Grall wrote: > Hi Stefano, > > On 03/06/2017 08:01 PM, Stefano Stabellini wrote: > > +static int xen_9pfs_front_alloc_dataring(struct xenbus_device *dev, > > + struct xen_9pfs_dataring *ring) > > +{ > > + int i; > > + int ret = -ENOMEM; > > + > > + init_waitqueue_head(>wq); > > + spin_lock_init(>lock); > > + INIT_WORK(>work, p9_xen_response); > > + > > + ring->intf = (struct xen_9pfs_data_intf *) __get_free_page(GFP_KERNEL > > | __GFP_ZERO); > > + if (!ring->intf) > > + goto error; > > + memset(ring->intf, 0, XEN_PAGE_SIZE); > > + ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, > > XEN_9PFS_RING_ORDER); > > The ring order will be in term of Xen page size and not Linux. So you are > going to allocate much more memory than expected on 64KB kernel. I'll fix. > > + if (ring->bytes == NULL) > > + goto error; > > + for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) > > + ring->intf->ref[i] = > > gnttab_grant_foreign_access(dev->otherend_id, > > pfn_to_gfn(virt_to_pfn((void*)ring->bytes) + i), 0);. > > Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). OK > Also, this is not going to work on 64K kernel because you will grant access to > noncontiguous memory (e.g 0-4K, 64K-68K,...). By using virt_to_gfn like you suggested, the loop will correctly iterate on a 4K by 4K basis, even on a 64K kernel: ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, XEN_9PFS_RING_ORDER - (PAGE_SHIFT - XEN_PAGE_SHIFT)); for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, virt_to_gfn((void*)ring->bytes) + i, 0); where XEN_9PFS_RING_ORDER specifies the order at 4K granularity. Am I missing something? > We have various helper to break-down the page for you, see > gnttab_for_one_grant, gnttab_foreach_grant, gnttab_count_grant, > xen_for_each_gfn (though this one it is internal to xlate_mmu.c so far) > > Please use them to avoid any further. > > > + ring->ref = gnttab_grant_foreign_access(dev->otherend_id, > > pfn_to_gfn(virt_to_pfn((void*)ring->intf)), 0); > > Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). Sure > > + ring->ring.in = ring->bytes; > > + ring->ring.out = ring->bytes + XEN_9PFS_RING_SIZE; > > + > > + ret = xenbus_alloc_evtchn(dev, >evtchn); > > + if (ret) > > + goto error; > > + ring->irq = bind_evtchn_to_irqhandler(ring->evtchn, > > xen_9pfs_front_event_handler, > > + 0, "xen_9pfs-frontend", ring); > > + if (ring->irq < 0) { > > + xenbus_free_evtchn(dev, ring->evtchn); > > + ret = ring->irq; > > + goto error; > > + } > > return 0; > > + > > +error: > > + if (ring->intf != NULL) > > + kfree(ring->intf); > > + if (ring->bytes != NULL) > > + kfree(ring->bytes); > > + return ret; > > }
Re: [Xen-devel] [PATCH 4/7] xen/9pfs: connect to the backend
On Tue, 7 Mar 2017, Julien Grall wrote: > Hi Stefano, > > On 03/06/2017 08:01 PM, Stefano Stabellini wrote: > > +static int xen_9pfs_front_alloc_dataring(struct xenbus_device *dev, > > + struct xen_9pfs_dataring *ring) > > +{ > > + int i; > > + int ret = -ENOMEM; > > + > > + init_waitqueue_head(>wq); > > + spin_lock_init(>lock); > > + INIT_WORK(>work, p9_xen_response); > > + > > + ring->intf = (struct xen_9pfs_data_intf *) __get_free_page(GFP_KERNEL > > | __GFP_ZERO); > > + if (!ring->intf) > > + goto error; > > + memset(ring->intf, 0, XEN_PAGE_SIZE); > > + ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, > > XEN_9PFS_RING_ORDER); > > The ring order will be in term of Xen page size and not Linux. So you are > going to allocate much more memory than expected on 64KB kernel. I'll fix. > > + if (ring->bytes == NULL) > > + goto error; > > + for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) > > + ring->intf->ref[i] = > > gnttab_grant_foreign_access(dev->otherend_id, > > pfn_to_gfn(virt_to_pfn((void*)ring->bytes) + i), 0);. > > Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). OK > Also, this is not going to work on 64K kernel because you will grant access to > noncontiguous memory (e.g 0-4K, 64K-68K,...). By using virt_to_gfn like you suggested, the loop will correctly iterate on a 4K by 4K basis, even on a 64K kernel: ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, XEN_9PFS_RING_ORDER - (PAGE_SHIFT - XEN_PAGE_SHIFT)); for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, virt_to_gfn((void*)ring->bytes) + i, 0); where XEN_9PFS_RING_ORDER specifies the order at 4K granularity. Am I missing something? > We have various helper to break-down the page for you, see > gnttab_for_one_grant, gnttab_foreach_grant, gnttab_count_grant, > xen_for_each_gfn (though this one it is internal to xlate_mmu.c so far) > > Please use them to avoid any further. > > > + ring->ref = gnttab_grant_foreign_access(dev->otherend_id, > > pfn_to_gfn(virt_to_pfn((void*)ring->intf)), 0); > > Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). Sure > > + ring->ring.in = ring->bytes; > > + ring->ring.out = ring->bytes + XEN_9PFS_RING_SIZE; > > + > > + ret = xenbus_alloc_evtchn(dev, >evtchn); > > + if (ret) > > + goto error; > > + ring->irq = bind_evtchn_to_irqhandler(ring->evtchn, > > xen_9pfs_front_event_handler, > > + 0, "xen_9pfs-frontend", ring); > > + if (ring->irq < 0) { > > + xenbus_free_evtchn(dev, ring->evtchn); > > + ret = ring->irq; > > + goto error; > > + } > > return 0; > > + > > +error: > > + if (ring->intf != NULL) > > + kfree(ring->intf); > > + if (ring->bytes != NULL) > > + kfree(ring->bytes); > > + return ret; > > }
Re: [Xen-devel] [PATCH 4/7] xen/9pfs: connect to the backend
Hi Stefano, On 03/06/2017 08:01 PM, Stefano Stabellini wrote: +static int xen_9pfs_front_alloc_dataring(struct xenbus_device *dev, + struct xen_9pfs_dataring *ring) +{ + int i; + int ret = -ENOMEM; + + init_waitqueue_head(>wq); + spin_lock_init(>lock); + INIT_WORK(>work, p9_xen_response); + + ring->intf = (struct xen_9pfs_data_intf *) __get_free_page(GFP_KERNEL | __GFP_ZERO); + if (!ring->intf) + goto error; + memset(ring->intf, 0, XEN_PAGE_SIZE); + ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, XEN_9PFS_RING_ORDER); The ring order will be in term of Xen page size and not Linux. So you are going to allocate much more memory than expected on 64KB kernel. + if (ring->bytes == NULL) + goto error; + for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) + ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, pfn_to_gfn(virt_to_pfn((void*)ring->bytes) + i), 0);. Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). Also, this is not going to work on 64K kernel because you will grant access to noncontiguous memory (e.g 0-4K, 64K-68K,...). We have various helper to break-down the page for you, see gnttab_for_one_grant, gnttab_foreach_grant, gnttab_count_grant, xen_for_each_gfn (though this one it is internal to xlate_mmu.c so far) Please use them to avoid any further. + ring->ref = gnttab_grant_foreign_access(dev->otherend_id, pfn_to_gfn(virt_to_pfn((void*)ring->intf)), 0); Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). + ring->ring.in = ring->bytes; + ring->ring.out = ring->bytes + XEN_9PFS_RING_SIZE; + + ret = xenbus_alloc_evtchn(dev, >evtchn); + if (ret) + goto error; + ring->irq = bind_evtchn_to_irqhandler(ring->evtchn, xen_9pfs_front_event_handler, + 0, "xen_9pfs-frontend", ring); + if (ring->irq < 0) { + xenbus_free_evtchn(dev, ring->evtchn); + ret = ring->irq; + goto error; + } return 0; + +error: + if (ring->intf != NULL) + kfree(ring->intf); + if (ring->bytes != NULL) + kfree(ring->bytes); + return ret; } Cheers, -- Julien Grall
Re: [Xen-devel] [PATCH 4/7] xen/9pfs: connect to the backend
Hi Stefano, On 03/06/2017 08:01 PM, Stefano Stabellini wrote: +static int xen_9pfs_front_alloc_dataring(struct xenbus_device *dev, + struct xen_9pfs_dataring *ring) +{ + int i; + int ret = -ENOMEM; + + init_waitqueue_head(>wq); + spin_lock_init(>lock); + INIT_WORK(>work, p9_xen_response); + + ring->intf = (struct xen_9pfs_data_intf *) __get_free_page(GFP_KERNEL | __GFP_ZERO); + if (!ring->intf) + goto error; + memset(ring->intf, 0, XEN_PAGE_SIZE); + ring->bytes = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, XEN_9PFS_RING_ORDER); The ring order will be in term of Xen page size and not Linux. So you are going to allocate much more memory than expected on 64KB kernel. + if (ring->bytes == NULL) + goto error; + for (i = 0; i < (1 << XEN_9PFS_RING_ORDER); i++) + ring->intf->ref[i] = gnttab_grant_foreign_access(dev->otherend_id, pfn_to_gfn(virt_to_pfn((void*)ring->bytes) + i), 0);. Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). Also, this is not going to work on 64K kernel because you will grant access to noncontiguous memory (e.g 0-4K, 64K-68K,...). We have various helper to break-down the page for you, see gnttab_for_one_grant, gnttab_foreach_grant, gnttab_count_grant, xen_for_each_gfn (though this one it is internal to xlate_mmu.c so far) Please use them to avoid any further. + ring->ref = gnttab_grant_foreign_access(dev->otherend_id, pfn_to_gfn(virt_to_pfn((void*)ring->intf)), 0); Please use virt_to_gfn rather than pfn_to_gfn(virt_to_pfn). + ring->ring.in = ring->bytes; + ring->ring.out = ring->bytes + XEN_9PFS_RING_SIZE; + + ret = xenbus_alloc_evtchn(dev, >evtchn); + if (ret) + goto error; + ring->irq = bind_evtchn_to_irqhandler(ring->evtchn, xen_9pfs_front_event_handler, + 0, "xen_9pfs-frontend", ring); + if (ring->irq < 0) { + xenbus_free_evtchn(dev, ring->evtchn); + ret = ring->irq; + goto error; + } return 0; + +error: + if (ring->intf != NULL) + kfree(ring->intf); + if (ring->bytes != NULL) + kfree(ring->bytes); + return ret; } Cheers, -- Julien Grall