|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: [EXTERNAL] [PATCH v4 08/14] remove remaining uses of iommu_legacy_map/unmap
> -----Original Message-----
> From: Jan Beulich <jbeulich@xxxxxxxx>
> Sent: 06 August 2020 11:29
> To: Paul Durrant <paul@xxxxxxx>
> Cc: xen-devel@xxxxxxxxxxxxxxxxxxxx; Durrant, Paul <pdurrant@xxxxxxxxxxxx>;
> Andrew Cooper
> <andrew.cooper3@xxxxxxxxxx>; Wei Liu <wl@xxxxxxx>; Roger Pau Monné
> <roger.pau@xxxxxxxxxx>; George
> Dunlap <george.dunlap@xxxxxxxxxx>; Ian Jackson <ian.jackson@xxxxxxxxxxxxx>;
> Julien Grall
> <julien@xxxxxxx>; Stefano Stabellini <sstabellini@xxxxxxxxxx>; Jun Nakajima
> <jun.nakajima@xxxxxxxxx>;
> Kevin Tian <kevin.tian@xxxxxxxxx>
> Subject: RE: [EXTERNAL] [PATCH v4 08/14] remove remaining uses of
> iommu_legacy_map/unmap
>
> CAUTION: This email originated from outside of the organization. Do not click
> links or open
> attachments unless you can confirm the sender and know the content is safe.
>
>
>
> On 04.08.2020 15:42, Paul Durrant wrote:
> > The 'legacy' functions do implicit flushing so amend the callers to do the
> > appropriate flushing.
> >
> > Unfortunately, because of the structure of the P2M code, we cannot remove
> > the per-CPU 'iommu_dont_flush_iotlb' global and the optimization it
> > facilitates. It is now checked directly iommu_iotlb_flush(). Also, it is
> > now declared as bool (rather than bool_t) and setting/clearing it are no
> > longer pointlessly gated on is_iommu_enabled() returning true. (Arguably
> > it is also pointless to gate the call to iommu_iotlb_flush() on that
> > condition - since it is a no-op in that case - but the if clause allows
> > the scope of a stack variable to be restricted).
> >
> > NOTE: The code in memory_add() now fails if the number of pages passed to
> > a single call overflows an unsigned int. I don't believe this will
> > ever happen in practice.
>
> I.e. you don't think adding 16Tb of memory in one go is possible?
> I wouldn't bet on that ...
>
I've re-worked previous patches to use unsigned long so I don't need this
restriction any more.
> > --- a/xen/arch/x86/mm.c
> > +++ b/xen/arch/x86/mm.c
> > @@ -2446,10 +2446,16 @@ static int cleanup_page_mappings(struct page_info
> > *page)
> >
> > if ( d && unlikely(need_iommu_pt_sync(d)) && is_pv_domain(d) )
> > {
> > - int rc2 = iommu_legacy_unmap(d, _dfn(mfn), PAGE_ORDER_4K);
> > + unsigned int flush_flags = 0;
> > + int err;
> >
> > + err = iommu_unmap(d, _dfn(mfn), PAGE_ORDER_4K, 1,
> > &flush_flags);
> > if ( !rc )
> > - rc = rc2;
> > + rc = err;
> > +
> > + err = iommu_iotlb_flush(d, _dfn(mfn), PAGE_ORDER_4K, 1,
> > flush_flags);
> > + if ( !rc )
> > + rc = err;
> > }
>
> Wasn't the earlier change to add flushing in the error case to
> allow to simplify code like this to
>
> if ( d && unlikely(need_iommu_pt_sync(d)) && is_pv_domain(d) )
> {
> unsigned int flush_flags = 0;
> int err;
>
> err = iommu_unmap(d, _dfn(mfn), PAGE_ORDER_4K, 1, &flush_flags);
> if ( !err )
> err = iommu_iotlb_flush(d, _dfn(mfn), PAGE_ORDER_4K, 1,
> flush_flags);
> if ( !rc )
> rc = err;
> }
>
> ?
Yes.
>
> > @@ -1441,9 +1446,16 @@ int clear_identity_p2m_entry(struct domain *d,
> > unsigned long gfn_l)
> >
> > if ( !paging_mode_translate(d) )
> > {
> > - if ( !is_iommu_enabled(d) )
> > - return 0;
> > - return iommu_legacy_unmap(d, _dfn(gfn_l), PAGE_ORDER_4K);
> > + unsigned int flush_flags = 0;
> > + int err;
> > +
> > + ret = iommu_unmap(d, _dfn(gfn_l), PAGE_ORDER_4K, 1, &flush_flags);
> > +
> > + err = iommu_iotlb_flush(d, _dfn(gfn_l), PAGE_ORDER_4K, 1,
> > flush_flags);
> > + if ( !ret )
> > + ret = err;
> > +
> > + return ret;
> > }
>
> Similarly here then.
>
Yes.
> > --- a/xen/arch/x86/x86_64/mm.c
> > +++ b/xen/arch/x86/x86_64/mm.c
> > @@ -1413,21 +1413,22 @@ int memory_add(unsigned long spfn, unsigned long
> > epfn, unsigned int pxm)
> > !iommu_use_hap_pt(hardware_domain) &&
> > !need_iommu_pt_sync(hardware_domain) )
> > {
> > - for ( i = spfn; i < epfn; i++ )
> > - if ( iommu_legacy_map(hardware_domain, _dfn(i), _mfn(i),
> > - PAGE_ORDER_4K,
> > - IOMMUF_readable | IOMMUF_writable) )
> > - break;
> > - if ( i != epfn )
> > - {
> > - while (i-- > old_max)
> > - /* If statement to satisfy __must_check. */
> > - if ( iommu_legacy_unmap(hardware_domain, _dfn(i),
> > - PAGE_ORDER_4K) )
> > - continue;
> > + unsigned int flush_flags = 0;
> > + unsigned int n = epfn - spfn;
> > + int rc;
> >
> > + ret = -EOVERFLOW;
> > + if ( spfn + n != epfn )
> > + goto destroy_m2p;
> > +
> > + rc = iommu_map(hardware_domain, _dfn(i), _mfn(i),
> > + PAGE_ORDER_4K, n, IOMMUF_readable | IOMMUF_writable,
> > + &flush_flags);
> > + if ( !rc )
> > + rc = iommu_iotlb_flush(hardware_domain, _dfn(i),
> > PAGE_ORDER_4K, n,
> > + flush_flags);
> > + if ( rc )
> > goto destroy_m2p;
> > - }
> > }
>
> Did you mean to use "ret" here instead of introducing "rc"?
>
The previous code did not set ret in the case of an iommu op failure but that
does appear to be a mistake. I will use ret, as you suggest, but I will call it
out in the commit description too.
> > --- a/xen/common/grant_table.c
> > +++ b/xen/common/grant_table.c
> > @@ -1225,11 +1225,23 @@ map_grant_ref(
> > kind = IOMMUF_readable;
> > else
> > kind = 0;
> > - if ( kind && iommu_legacy_map(ld, _dfn(mfn_x(mfn)), mfn, 0, kind) )
> > + if ( kind )
> > {
> > - double_gt_unlock(lgt, rgt);
> > - rc = GNTST_general_error;
> > - goto undo_out;
> > + dfn_t dfn = _dfn(mfn_x(mfn));
> > + unsigned int flush_flags = 0;
> > + int err;
> > +
> > + err = iommu_map(ld, dfn, mfn, 0, 1, kind, &flush_flags);
> > + if ( !err )
> > + err = iommu_iotlb_flush(ld, dfn, 0, 1, flush_flags);
>
> Question of 0 vs PAGE_ORDER_4K again.
>
> > @@ -1473,21 +1485,25 @@ unmap_common(
> > if ( rc == GNTST_okay && gnttab_need_iommu_mapping(ld) )
> > {
> > unsigned int kind;
> > + dfn_t dfn = _dfn(mfn_x(op->mfn));
> > + unsigned int flush_flags = 0;
> > int err = 0;
> >
> > double_gt_lock(lgt, rgt);
> >
> > kind = mapkind(lgt, rd, op->mfn);
> > if ( !kind )
> > - err = iommu_legacy_unmap(ld, _dfn(mfn_x(op->mfn)), 0);
> > + err = iommu_unmap(ld, dfn, 0, 1, &flush_flags);
> > else if ( !(kind & MAPKIND_WRITE) )
> > - err = iommu_legacy_map(ld, _dfn(mfn_x(op->mfn)), op->mfn, 0,
> > - IOMMUF_readable);
> > -
> > - double_gt_unlock(lgt, rgt);
> > + err = iommu_map(ld, dfn, op->mfn, 0, 1, IOMMUF_readable,
> > + &flush_flags);
> >
> > + if ( !err )
> > + err = iommu_iotlb_flush(ld, dfn, 0, 1, flush_flags);
> > if ( err )
> > rc = GNTST_general_error;
> > +
> > + double_gt_unlock(lgt, rgt);
> > }
>
> While moving the unlock ahead of the flush would be somewhat troublesome
> in the map case, it seems straightforward here. Even if this gets further
> adjusted by a later patch, it should imo be done here - the later patch
> may also go in much later.
>
Ok.
> > --- a/xen/common/memory.c
> > +++ b/xen/common/memory.c
> > @@ -824,8 +824,7 @@ int xenmem_add_to_physmap(struct domain *d, struct
> > xen_add_to_physmap *xatp,
> > xatp->gpfn += start;
> > xatp->size -= start;
> >
> > - if ( is_iommu_enabled(d) )
> > - this_cpu(iommu_dont_flush_iotlb) = 1;
> > + this_cpu(iommu_dont_flush_iotlb) = true;
>
> Just like you replace the original instance here, ...
>
> > @@ -845,6 +844,8 @@ int xenmem_add_to_physmap(struct domain *d, struct
> > xen_add_to_physmap *xatp,
> > }
> > }
> >
> > + this_cpu(iommu_dont_flush_iotlb) = false;
> > +
> > if ( is_iommu_enabled(d) )
> > {
> > int ret;
>
> ... I'm sure you meant to also remove the original instance from
> down below here.
I did indeed. Thanks for spotting.
>
> > @@ -364,7 +341,7 @@ int iommu_iotlb_flush(struct domain *d, dfn_t dfn,
> > unsigned int page_order,
> > int rc;
> >
> > if ( !is_iommu_enabled(d) || !hd->platform_ops->iotlb_flush ||
> > - !page_count || !flush_flags )
> > + !page_count || !flush_flags || this_cpu(iommu_dont_flush_iotlb) )
> > return 0;
>
> The patch description ought to assure the safety of this change: So
> far, despite the flag set callers of iommu_iotlb_flush() (which
> may be unaware of the flag's state) did get what they did ask for.
> The change relies on there not being any such uses.
>
Ok, I'll call it out.
> > @@ -370,15 +362,12 @@ void iommu_dev_iotlb_flush_timeout(struct domain *d,
> > struct pci_dev *pdev);
> >
> > /*
> > * The purpose of the iommu_dont_flush_iotlb optional cpu flag is to
> > - * avoid unecessary iotlb_flush in the low level IOMMU code.
> > - *
> > - * iommu_map_page/iommu_unmap_page must flush the iotlb but somethimes
> > - * this operation can be really expensive. This flag will be set by the
> > - * caller to notify the low level IOMMU code to avoid the iotlb flushes.
> > - * iommu_iotlb_flush/iommu_iotlb_flush_all will be explicitly called by
> > - * the caller.
> > + * avoid unecessary IOMMU flushing while updating the P2M.
>
> Correct the spelling of "unnecessary" at the same time?
>
Oh yes. Will do.
Paul
> Jan
|
 |
Lists.xenproject.org is hosted with RackSpace, monitoring our |