Re: Use km_alloc(9) in drm
On Sat, Feb 05, 2022 at 11:44:03AM +0100, Mark Kettenis wrote:
> We want to get rid of the uvm_km_valloc() interfaces in favour of
> km_alloc(). This changes the calls in drm(4) over. The kv_physwait
> struct is made static to prevent collission with a symbol in
> vm_machdep.c on some architectures. The goal is to move this into
> uvm/uvm_km.c eventually.
>
> Just to make sure I didn't screw the conversion up somehow a few tests
> on a mix of inteldrm(4) and amdgpu(4) systems would be good.
I am running it since Feb 5 on amd64 (amdgpu0: RAVEN2 3 CU rev 0x09)
and see no particular new problem.
> Index: dev/pci/drm/drm_linux.c
> ===
> RCS file: /cvs/src/sys/dev/pci/drm/drm_linux.c,v
> retrieving revision 1.89
> diff -u -p -r1.89 drm_linux.c
> --- dev/pci/drm/drm_linux.c 21 Jan 2022 23:49:36 - 1.89
> +++ dev/pci/drm/drm_linux.c 5 Feb 2022 10:40:22 -
> @@ -564,6 +564,11 @@ __pagevec_release(struct pagevec *pvec)
> pagevec_reinit(pvec);
> }
>
> +static struct kmem_va_mode kv_physwait = {
> + .kv_map = _map,
> + .kv_wait = 1,
> +};
> +
> void *
> kmap(struct vm_page *pg)
> {
> @@ -572,7 +577,7 @@ kmap(struct vm_page *pg)
> #if defined (__HAVE_PMAP_DIRECT)
> va = pmap_map_direct(pg);
> #else
> - va = uvm_km_valloc_wait(phys_map, PAGE_SIZE);
> + va = (vaddr_t)km_alloc(PAGE_SIZE, _physwait, _none, _waitok);
> pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), PROT_READ | PROT_WRITE);
> pmap_update(pmap_kernel());
> #endif
> @@ -589,7 +594,7 @@ kunmap_va(void *addr)
> #else
> pmap_kremove(va, PAGE_SIZE);
> pmap_update(pmap_kernel());
> - uvm_km_free_wakeup(phys_map, va, PAGE_SIZE);
> + km_free((void *)va, PAGE_SIZE, _physwait, _none);
> #endif
> }
>
> @@ -624,7 +629,8 @@ vmap(struct vm_page **pages, unsigned in
> paddr_t pa;
> int i;
>
> - va = uvm_km_valloc(kernel_map, PAGE_SIZE * npages);
> + va = (vaddr_t)km_alloc(PAGE_SIZE * npages, _any, _none,
> + _nowait);
> if (va == 0)
> return NULL;
> for (i = 0; i < npages; i++) {
> @@ -645,7 +651,7 @@ vunmap(void *addr, size_t size)
>
> pmap_remove(pmap_kernel(), va, va + size);
> pmap_update(pmap_kernel());
> - uvm_km_free(kernel_map, va, size);
> + km_free((void *)va, size, _any, _none);
> }
>
> bool
>
--
Sebastien Marie
Re: Use km_alloc(9) in drm
On Sat, Feb 05, 2022 at 02:26:48PM +0100, Mark Kettenis wrote:
> > Date: Sat, 5 Feb 2022 22:34:10 +1100
> > From: Jonathan Gray
> >
> > On Sat, Feb 05, 2022 at 11:44:03AM +0100, Mark Kettenis wrote:
> > > We want to get rid of the uvm_km_valloc() interfaces in favour of
> > > km_alloc(). This changes the calls in drm(4) over. The kv_physwait
> > > struct is made static to prevent collission with a symbol in
> > > vm_machdep.c on some architectures. The goal is to move this into
> > > uvm/uvm_km.c eventually.
> > >
> > > Just to make sure I didn't screw the conversion up somehow a few tests
> > > on a mix of inteldrm(4) and amdgpu(4) systems would be good.
> > >
> > > ok?
> > >
> > >
> > > Index: dev/pci/drm/drm_linux.c
> > > ===
> > > RCS file: /cvs/src/sys/dev/pci/drm/drm_linux.c,v
> > > retrieving revision 1.89
> > > diff -u -p -r1.89 drm_linux.c
> > > --- dev/pci/drm/drm_linux.c 21 Jan 2022 23:49:36 - 1.89
> > > +++ dev/pci/drm/drm_linux.c 5 Feb 2022 10:40:22 -
> > > @@ -564,6 +564,11 @@ __pagevec_release(struct pagevec *pvec)
> > > pagevec_reinit(pvec);
> > > }
> > >
> > > +static struct kmem_va_mode kv_physwait = {
> > > + .kv_map = _map,
> > > + .kv_wait = 1,
> > > +};
> > > +
> > > void *
> > > kmap(struct vm_page *pg)
> > > {
> > > @@ -572,7 +577,7 @@ kmap(struct vm_page *pg)
> > > #if defined (__HAVE_PMAP_DIRECT)
> > > va = pmap_map_direct(pg);
> > > #else
> > > - va = uvm_km_valloc_wait(phys_map, PAGE_SIZE);
> > > + va = (vaddr_t)km_alloc(PAGE_SIZE, _physwait, _none, _waitok);
> > > pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), PROT_READ | PROT_WRITE);
> > > pmap_update(pmap_kernel());
> > > #endif
> > > @@ -589,7 +594,7 @@ kunmap_va(void *addr)
> > > #else
> > > pmap_kremove(va, PAGE_SIZE);
> > > pmap_update(pmap_kernel());
> > > - uvm_km_free_wakeup(phys_map, va, PAGE_SIZE);
> > > + km_free((void *)va, PAGE_SIZE, _physwait, _none);
> > > #endif
> > > }
> > >
> > > @@ -624,7 +629,8 @@ vmap(struct vm_page **pages, unsigned in
> > > paddr_t pa;
> > > int i;
> > >
> > > - va = uvm_km_valloc(kernel_map, PAGE_SIZE * npages);
> > > + va = (vaddr_t)km_alloc(PAGE_SIZE * npages, _any, _none,
> > > + _nowait);
> >
> > this could be kd_waitok as the linux vmap() uses might_sleep()
> > but I suppose that would be a change in behaviour/different diff
>
> Actually that doesn't make a difference. Specifying kd_waitok will
> only make it wait for physical pages. But we're passing kp_none so no
> physical pages are allocated.
>
> Yes, this is confusing and continues to trip people up.
ah right
tested by starting X and chromium on
amd64
amdgpu0: VEGA10 56 CU rev 0x01
inteldrm0: msi, BROADWELL, gen 8
i386
X and crack-attack (chromium won't start due to illegal instruction)
radeondrm0: RV200
inteldrm0: apic 1 int 16, I85X, gen 2
ok jsg@
>
> >
> > > if (va == 0)
> > > return NULL;
> > > for (i = 0; i < npages; i++) {
> > > @@ -645,7 +651,7 @@ vunmap(void *addr, size_t size)
> > >
> > > pmap_remove(pmap_kernel(), va, va + size);
> > > pmap_update(pmap_kernel());
> > > - uvm_km_free(kernel_map, va, size);
> > > + km_free((void *)va, size, _any, _none);
> > > }
> > >
> > > bool
> > >
> > >
> >
>
>
Re: Use km_alloc(9) in drm
> Date: Sat, 5 Feb 2022 22:34:10 +1100
> From: Jonathan Gray
>
> On Sat, Feb 05, 2022 at 11:44:03AM +0100, Mark Kettenis wrote:
> > We want to get rid of the uvm_km_valloc() interfaces in favour of
> > km_alloc(). This changes the calls in drm(4) over. The kv_physwait
> > struct is made static to prevent collission with a symbol in
> > vm_machdep.c on some architectures. The goal is to move this into
> > uvm/uvm_km.c eventually.
> >
> > Just to make sure I didn't screw the conversion up somehow a few tests
> > on a mix of inteldrm(4) and amdgpu(4) systems would be good.
> >
> > ok?
> >
> >
> > Index: dev/pci/drm/drm_linux.c
> > ===
> > RCS file: /cvs/src/sys/dev/pci/drm/drm_linux.c,v
> > retrieving revision 1.89
> > diff -u -p -r1.89 drm_linux.c
> > --- dev/pci/drm/drm_linux.c 21 Jan 2022 23:49:36 - 1.89
> > +++ dev/pci/drm/drm_linux.c 5 Feb 2022 10:40:22 -
> > @@ -564,6 +564,11 @@ __pagevec_release(struct pagevec *pvec)
> > pagevec_reinit(pvec);
> > }
> >
> > +static struct kmem_va_mode kv_physwait = {
> > + .kv_map = _map,
> > + .kv_wait = 1,
> > +};
> > +
> > void *
> > kmap(struct vm_page *pg)
> > {
> > @@ -572,7 +577,7 @@ kmap(struct vm_page *pg)
> > #if defined (__HAVE_PMAP_DIRECT)
> > va = pmap_map_direct(pg);
> > #else
> > - va = uvm_km_valloc_wait(phys_map, PAGE_SIZE);
> > + va = (vaddr_t)km_alloc(PAGE_SIZE, _physwait, _none, _waitok);
> > pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), PROT_READ | PROT_WRITE);
> > pmap_update(pmap_kernel());
> > #endif
> > @@ -589,7 +594,7 @@ kunmap_va(void *addr)
> > #else
> > pmap_kremove(va, PAGE_SIZE);
> > pmap_update(pmap_kernel());
> > - uvm_km_free_wakeup(phys_map, va, PAGE_SIZE);
> > + km_free((void *)va, PAGE_SIZE, _physwait, _none);
> > #endif
> > }
> >
> > @@ -624,7 +629,8 @@ vmap(struct vm_page **pages, unsigned in
> > paddr_t pa;
> > int i;
> >
> > - va = uvm_km_valloc(kernel_map, PAGE_SIZE * npages);
> > + va = (vaddr_t)km_alloc(PAGE_SIZE * npages, _any, _none,
> > + _nowait);
>
> this could be kd_waitok as the linux vmap() uses might_sleep()
> but I suppose that would be a change in behaviour/different diff
Actually that doesn't make a difference. Specifying kd_waitok will
only make it wait for physical pages. But we're passing kp_none so no
physical pages are allocated.
Yes, this is confusing and continues to trip people up.
>
> > if (va == 0)
> > return NULL;
> > for (i = 0; i < npages; i++) {
> > @@ -645,7 +651,7 @@ vunmap(void *addr, size_t size)
> >
> > pmap_remove(pmap_kernel(), va, va + size);
> > pmap_update(pmap_kernel());
> > - uvm_km_free(kernel_map, va, size);
> > + km_free((void *)va, size, _any, _none);
> > }
> >
> > bool
> >
> >
>
Re: Use km_alloc(9) in drm
On Sat, Feb 05, 2022 at 11:44:03AM +0100, Mark Kettenis wrote:
> We want to get rid of the uvm_km_valloc() interfaces in favour of
> km_alloc(). This changes the calls in drm(4) over. The kv_physwait
> struct is made static to prevent collission with a symbol in
> vm_machdep.c on some architectures. The goal is to move this into
> uvm/uvm_km.c eventually.
>
> Just to make sure I didn't screw the conversion up somehow a few tests
> on a mix of inteldrm(4) and amdgpu(4) systems would be good.
>
> ok?
>
>
> Index: dev/pci/drm/drm_linux.c
> ===
> RCS file: /cvs/src/sys/dev/pci/drm/drm_linux.c,v
> retrieving revision 1.89
> diff -u -p -r1.89 drm_linux.c
> --- dev/pci/drm/drm_linux.c 21 Jan 2022 23:49:36 - 1.89
> +++ dev/pci/drm/drm_linux.c 5 Feb 2022 10:40:22 -
> @@ -564,6 +564,11 @@ __pagevec_release(struct pagevec *pvec)
> pagevec_reinit(pvec);
> }
>
> +static struct kmem_va_mode kv_physwait = {
> + .kv_map = _map,
> + .kv_wait = 1,
> +};
> +
> void *
> kmap(struct vm_page *pg)
> {
> @@ -572,7 +577,7 @@ kmap(struct vm_page *pg)
> #if defined (__HAVE_PMAP_DIRECT)
> va = pmap_map_direct(pg);
> #else
> - va = uvm_km_valloc_wait(phys_map, PAGE_SIZE);
> + va = (vaddr_t)km_alloc(PAGE_SIZE, _physwait, _none, _waitok);
> pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), PROT_READ | PROT_WRITE);
> pmap_update(pmap_kernel());
> #endif
> @@ -589,7 +594,7 @@ kunmap_va(void *addr)
> #else
> pmap_kremove(va, PAGE_SIZE);
> pmap_update(pmap_kernel());
> - uvm_km_free_wakeup(phys_map, va, PAGE_SIZE);
> + km_free((void *)va, PAGE_SIZE, _physwait, _none);
> #endif
> }
>
> @@ -624,7 +629,8 @@ vmap(struct vm_page **pages, unsigned in
> paddr_t pa;
> int i;
>
> - va = uvm_km_valloc(kernel_map, PAGE_SIZE * npages);
> + va = (vaddr_t)km_alloc(PAGE_SIZE * npages, _any, _none,
> + _nowait);
this could be kd_waitok as the linux vmap() uses might_sleep()
but I suppose that would be a change in behaviour/different diff
> if (va == 0)
> return NULL;
> for (i = 0; i < npages; i++) {
> @@ -645,7 +651,7 @@ vunmap(void *addr, size_t size)
>
> pmap_remove(pmap_kernel(), va, va + size);
> pmap_update(pmap_kernel());
> - uvm_km_free(kernel_map, va, size);
> + km_free((void *)va, size, _any, _none);
> }
>
> bool
>
>
Use km_alloc(9) in drm
We want to get rid of the uvm_km_valloc() interfaces in favour of
km_alloc(). This changes the calls in drm(4) over. The kv_physwait
struct is made static to prevent collission with a symbol in
vm_machdep.c on some architectures. The goal is to move this into
uvm/uvm_km.c eventually.
Just to make sure I didn't screw the conversion up somehow a few tests
on a mix of inteldrm(4) and amdgpu(4) systems would be good.
ok?
Index: dev/pci/drm/drm_linux.c
===
RCS file: /cvs/src/sys/dev/pci/drm/drm_linux.c,v
retrieving revision 1.89
diff -u -p -r1.89 drm_linux.c
--- dev/pci/drm/drm_linux.c 21 Jan 2022 23:49:36 - 1.89
+++ dev/pci/drm/drm_linux.c 5 Feb 2022 10:40:22 -
@@ -564,6 +564,11 @@ __pagevec_release(struct pagevec *pvec)
pagevec_reinit(pvec);
}
+static struct kmem_va_mode kv_physwait = {
+ .kv_map = _map,
+ .kv_wait = 1,
+};
+
void *
kmap(struct vm_page *pg)
{
@@ -572,7 +577,7 @@ kmap(struct vm_page *pg)
#if defined (__HAVE_PMAP_DIRECT)
va = pmap_map_direct(pg);
#else
- va = uvm_km_valloc_wait(phys_map, PAGE_SIZE);
+ va = (vaddr_t)km_alloc(PAGE_SIZE, _physwait, _none, _waitok);
pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), PROT_READ | PROT_WRITE);
pmap_update(pmap_kernel());
#endif
@@ -589,7 +594,7 @@ kunmap_va(void *addr)
#else
pmap_kremove(va, PAGE_SIZE);
pmap_update(pmap_kernel());
- uvm_km_free_wakeup(phys_map, va, PAGE_SIZE);
+ km_free((void *)va, PAGE_SIZE, _physwait, _none);
#endif
}
@@ -624,7 +629,8 @@ vmap(struct vm_page **pages, unsigned in
paddr_t pa;
int i;
- va = uvm_km_valloc(kernel_map, PAGE_SIZE * npages);
+ va = (vaddr_t)km_alloc(PAGE_SIZE * npages, _any, _none,
+ _nowait);
if (va == 0)
return NULL;
for (i = 0; i < npages; i++) {
@@ -645,7 +651,7 @@ vunmap(void *addr, size_t size)
pmap_remove(pmap_kernel(), va, va + size);
pmap_update(pmap_kernel());
- uvm_km_free(kernel_map, va, size);
+ km_free((void *)va, size, _any, _none);
}
bool
