| Message ID | 20220926065407.2389-1-wangdeming@inspur.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | vfio: Use filp instead of fd | expand |
On Mon, Sep 26, 2022 at 02:54:07AM -0400, Deming Wang wrote: > The function of kvm_vfio_group_set_spapr_tce and kvm_vfio_group_del > use fd indirectly.But,it only be used for fd.file. So,we can directly > use the struct of file instead. > > Signed-off-by: Deming Wang <wangdeming@inspur.com> NAK. fget() is for the cases when we must keep the reference across the syscall boundary/pass to another thread/etc. If fdget() is applicable, it's a better alternative. And I would suggest you to look at the generated code - it pretty much turns into struct file *file; int need_fput; r = __fdget(fd); need_fput = r & 3; r &= ~3; file = (struct file *)r; .... if (unlikely(need_fput)) fput(file); Note that we are *not* actually passing a structure out of a function, etc. - fdget() is inlined and out-of-line part returns unsigned long. Lower two bits carry flags, the rest - file pointer. Rearrangement into struct fd is done in the caller and compiler manages to dissolve that struct into a couple of local variables. Incidentally, pretty much the only thing in struct fd besides struct file pointer is that "have we failed to skip bumping the reference count?" flag. Between fdget() and fdput() only the file pointer is used - in all users. fdput() uses the flag to decide whether it needs to bother with refcount at all.
On Mon, Sep 26, 2022 at 02:54:07AM -0400, Deming Wang wrote: > The function of kvm_vfio_group_set_spapr_tce and kvm_vfio_group_del > use fd indirectly.But,it only be used for fd.file. So,we can directly > use the struct of file instead. > > Signed-off-by: Deming Wang <wangdeming@inspur.com> > --- > virt/kvm/vfio.c | 24 ++++++++++++------------ > 1 file changed, 12 insertions(+), 12 deletions(-) I thought about changing this too when I was looking at this. fdget/fdput includes a tiny micro-optimization that is legal here, however I doubt anyone cares about performance on this path. Reviewed-by: Jason Gunthorpe <jgg@nvidia.com> Jason
On Mon, Sep 26, 2022 at 02:07:43PM -0300, Jason Gunthorpe wrote: > On Mon, Sep 26, 2022 at 02:54:07AM -0400, Deming Wang wrote: > > The function of kvm_vfio_group_set_spapr_tce and kvm_vfio_group_del > > use fd indirectly.But,it only be used for fd.file. So,we can directly > > use the struct of file instead. > > > > Signed-off-by: Deming Wang <wangdeming@inspur.com> > > --- > > virt/kvm/vfio.c | 24 ++++++++++++------------ > > 1 file changed, 12 insertions(+), 12 deletions(-) > > I thought about changing this too when I was looking at > this. fdget/fdput includes a tiny micro-optimization that is legal > here, however I doubt anyone cares about performance on this path. Microoptimization or not, I'd rather keep fget() limited to cases where we really need it. There are non-trivial cases and having them easy to find is a good thing. Again, the preferred way to do descriptor lookups is fdget() family, not fget() one.
diff --git a/virt/kvm/vfio.c b/virt/kvm/vfio.c index ce1b01d02c51..3be84d82f905 100644 --- a/virt/kvm/vfio.c +++ b/virt/kvm/vfio.c @@ -178,11 +178,11 @@ static int kvm_vfio_group_del(struct kvm_device *dev, unsigned int fd) { struct kvm_vfio *kv = dev->private; struct kvm_vfio_group *kvg; - struct fd f; + struct file *filp; int ret; - f = fdget(fd); - if (!f.file) + filp = fget(fd); + if (!filp) return -EBADF; ret = -ENOENT; @@ -190,7 +190,7 @@ static int kvm_vfio_group_del(struct kvm_device *dev, unsigned int fd) mutex_lock(&kv->lock); list_for_each_entry(kvg, &kv->group_list, node) { - if (kvg->file != f.file) + if (kvg->file != filp) continue; list_del(&kvg->node); @@ -207,7 +207,7 @@ static int kvm_vfio_group_del(struct kvm_device *dev, unsigned int fd) mutex_unlock(&kv->lock); - fdput(f); + fput(filp); kvm_vfio_update_coherency(dev); @@ -221,14 +221,14 @@ static int kvm_vfio_group_set_spapr_tce(struct kvm_device *dev, struct kvm_vfio_spapr_tce param; struct kvm_vfio *kv = dev->private; struct kvm_vfio_group *kvg; - struct fd f; + struct file *filp; int ret; if (copy_from_user(¶m, arg, sizeof(struct kvm_vfio_spapr_tce))) return -EFAULT; - f = fdget(param.groupfd); - if (!f.file) + filp = fget(param.groupfd); + if (!filp) return -EBADF; ret = -ENOENT; @@ -238,13 +238,13 @@ static int kvm_vfio_group_set_spapr_tce(struct kvm_device *dev, list_for_each_entry(kvg, &kv->group_list, node) { struct iommu_group *grp; - if (kvg->file != f.file) + if (kvg->file != filp) continue; grp = kvm_vfio_file_iommu_group(kvg->file); if (WARN_ON_ONCE(!grp)) { ret = -EIO; - goto err_fdput; + goto err_fput; } ret = kvm_spapr_tce_attach_iommu_group(dev->kvm, param.tablefd, @@ -252,9 +252,9 @@ static int kvm_vfio_group_set_spapr_tce(struct kvm_device *dev, break; } -err_fdput: +err_fput: mutex_unlock(&kv->lock); - fdput(f); + fput(filp); return ret; } #endif
The function of kvm_vfio_group_set_spapr_tce and kvm_vfio_group_del use fd indirectly.But,it only be used for fd.file. So,we can directly use the struct of file instead. Signed-off-by: Deming Wang <wangdeming@inspur.com> --- virt/kvm/vfio.c | 24 ++++++++++++------------ 1 file changed, 12 insertions(+), 12 deletions(-)