| Message ID | 1601303017-8176-2-git-send-email-pmorel@linux.ibm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | s390x: css: pv: css test adaptation for PV | expand |
On Mon, 28 Sep 2020 16:23:34 +0200 Pierre Morel <pmorel@linux.ibm.com> wrote: > Some architectures need allocations to be done under a > specific address limit to allow DMA from I/O. > > We propose here a very simple page allocator to get > pages allocated under this specific limit. > > The DMA page allocator will only use part of the available memory > under the DMA address limit to let room for the standard allocator. > > Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> > --- > lib/alloc_dma_page.c | 57 ++++++++++++++++++++++++++++++++++++++++++++ > lib/alloc_dma_page.h | 24 +++++++++++++++++++ > lib/s390x/sclp.c | 2 ++ > s390x/Makefile | 1 + > 4 files changed, 84 insertions(+) > create mode 100644 lib/alloc_dma_page.c > create mode 100644 lib/alloc_dma_page.h (...) > diff --git a/lib/alloc_dma_page.h b/lib/alloc_dma_page.h > new file mode 100644 > index 0000000..85e1d2f > --- /dev/null > +++ b/lib/alloc_dma_page.h > @@ -0,0 +1,24 @@ > +/* > + * Page allocator for DMA definitions > + * > + * Copyright (c) IBM, Corp. 2020 > + * > + * Authors: > + * Pierre Morel <pmorel@linux.ibm.com> > + * > + * This code is free software; you can redistribute it and/or modify it > + * under the terms of the GNU Library General Public License version 2. > + */ > +#ifndef _ALLOC_DMA_PAGE_H_ > +#define _ALLOC_DMA_PAGE_H_ > + > +#include <asm/page.h> > + > +void put_dma_page(void *dma_page); > +void *get_dma_page(void); > +phys_addr_t dma_page_alloc_init(phys_addr_t start_pfn, phys_addr_t nb_pages); > + > +#define DMA_MAX_PFN (0x80000000 >> PAGE_SHIFT) > +#define DMA_ALLOC_RATIO 8 Hm, shouldn't the architecture be able to decide where a dma page can be located? Or am I misunderstanding? > + > +#endif /* _ALLOC_DMA_PAGE_H_ */ (...)
On 9/28/20 5:31 PM, Cornelia Huck wrote: > On Mon, 28 Sep 2020 16:23:34 +0200 > Pierre Morel <pmorel@linux.ibm.com> wrote: > >> Some architectures need allocations to be done under a >> specific address limit to allow DMA from I/O. >> >> We propose here a very simple page allocator to get >> pages allocated under this specific limit. >> >> The DMA page allocator will only use part of the available memory >> under the DMA address limit to let room for the standard allocator. >> >> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> >> --- >> lib/alloc_dma_page.c | 57 ++++++++++++++++++++++++++++++++++++++++++++ >> lib/alloc_dma_page.h | 24 +++++++++++++++++++ >> lib/s390x/sclp.c | 2 ++ >> s390x/Makefile | 1 + >> 4 files changed, 84 insertions(+) >> create mode 100644 lib/alloc_dma_page.c >> create mode 100644 lib/alloc_dma_page.h > > (...) > >> diff --git a/lib/alloc_dma_page.h b/lib/alloc_dma_page.h >> new file mode 100644 >> index 0000000..85e1d2f >> --- /dev/null >> +++ b/lib/alloc_dma_page.h >> @@ -0,0 +1,24 @@ >> +/* >> + * Page allocator for DMA definitions >> + * >> + * Copyright (c) IBM, Corp. 2020 >> + * >> + * Authors: >> + * Pierre Morel <pmorel@linux.ibm.com> >> + * >> + * This code is free software; you can redistribute it and/or modify it >> + * under the terms of the GNU Library General Public License version 2. >> + */ >> +#ifndef _ALLOC_DMA_PAGE_H_ >> +#define _ALLOC_DMA_PAGE_H_ >> + >> +#include <asm/page.h> >> + >> +void put_dma_page(void *dma_page); >> +void *get_dma_page(void); >> +phys_addr_t dma_page_alloc_init(phys_addr_t start_pfn, phys_addr_t nb_pages); >> + >> +#define DMA_MAX_PFN (0x80000000 >> PAGE_SHIFT) >> +#define DMA_ALLOC_RATIO 8 > > Hm, shouldn't the architecture be able to decide where a dma page can > be located? Or am I misunderstanding? Before we start any other discussion on this patch we should clear up if this is still necessary after Claudio's alloc revamp. I think he added options to request special types of memory. > >> + >> +#endif /* _ALLOC_DMA_PAGE_H_ */ > (...) >
On 9/29/20 9:19 AM, Janosch Frank wrote: > On 9/28/20 5:31 PM, Cornelia Huck wrote: >> On Mon, 28 Sep 2020 16:23:34 +0200 >> Pierre Morel <pmorel@linux.ibm.com> wrote: >> >>> Some architectures need allocations to be done under a >>> specific address limit to allow DMA from I/O. >>> >>> We propose here a very simple page allocator to get >>> pages allocated under this specific limit. >>> >>> The DMA page allocator will only use part of the available memory >>> under the DMA address limit to let room for the standard allocator. >>> ...snip... > > Before we start any other discussion on this patch we should clear up if > this is still necessary after Claudio's alloc revamp. > > I think he added options to request special types of memory. Isn't it possible to go on with this patch series. It can be adapted later to the changes that will be introduced by Claudio when it is final.
On 11/5/20 1:16 PM, Pierre Morel wrote: > > > On 9/29/20 9:19 AM, Janosch Frank wrote: >> On 9/28/20 5:31 PM, Cornelia Huck wrote: >>> On Mon, 28 Sep 2020 16:23:34 +0200 >>> Pierre Morel <pmorel@linux.ibm.com> wrote: >>> >>>> Some architectures need allocations to be done under a >>>> specific address limit to allow DMA from I/O. >>>> >>>> We propose here a very simple page allocator to get >>>> pages allocated under this specific limit. >>>> >>>> The DMA page allocator will only use part of the available memory >>>> under the DMA address limit to let room for the standard allocator. >>>> > > ...snip... > >> >> Before we start any other discussion on this patch we should clear up if >> this is still necessary after Claudio's alloc revamp. >> >> I think he added options to request special types of memory. > > Isn't it possible to go on with this patch series. > It can be adapted later to the changes that will be introduced by > Claudio when it is final. > > Pierre, that's outside of my jurisdiction, you're adding code to the common code library. I've set Paolo CC, let's see if he finds this thread :)
On Thu, Nov 05, 2020 at 03:15:47PM +0100, Janosch Frank wrote: > On 11/5/20 1:16 PM, Pierre Morel wrote: > > > > > > On 9/29/20 9:19 AM, Janosch Frank wrote: > >> On 9/28/20 5:31 PM, Cornelia Huck wrote: > >>> On Mon, 28 Sep 2020 16:23:34 +0200 > >>> Pierre Morel <pmorel@linux.ibm.com> wrote: > >>> > >>>> Some architectures need allocations to be done under a > >>>> specific address limit to allow DMA from I/O. > >>>> > >>>> We propose here a very simple page allocator to get > >>>> pages allocated under this specific limit. > >>>> > >>>> The DMA page allocator will only use part of the available memory > >>>> under the DMA address limit to let room for the standard allocator. > >>>> > > > > ...snip... > > > >> > >> Before we start any other discussion on this patch we should clear up if > >> this is still necessary after Claudio's alloc revamp. > >> > >> I think he added options to request special types of memory. > > > > Isn't it possible to go on with this patch series. > > It can be adapted later to the changes that will be introduced by > > Claudio when it is final. > > > > > > Pierre, that's outside of my jurisdiction, you're adding code to the > common code library. > > I've set Paolo CC, let's see if he finds this thread :) > I'll also try to find some time to revisit this. Thanks, drew
On 05/11/20 15:15, Janosch Frank wrote: >> Isn't it possible to go on with this patch series. >> It can be adapted later to the changes that will be introduced by >> Claudio when it is final. >> >> > Pierre, that's outside of my jurisdiction, you're adding code to the > common code library. > > I've set Paolo CC, let's see if he finds this thread:) > I have queued Claudio's series already, so let's start from there. Paolo
On 11/6/20 2:25 PM, Paolo Bonzini wrote: > On 05/11/20 15:15, Janosch Frank wrote: >>> Isn't it possible to go on with this patch series. >>> It can be adapted later to the changes that will be introduced by >>> Claudio when it is final. >>> >>> >> Pierre, that's outside of my jurisdiction, you're adding code to the >> common code library. >> >> I've set Paolo CC, let's see if he finds this thread:) >> > > I have queued Claudio's series already, so let's start from there. > > Paolo > OK, thanks Pierre
diff --git a/lib/alloc_dma_page.c b/lib/alloc_dma_page.c new file mode 100644 index 0000000..6a16e38 --- /dev/null +++ b/lib/alloc_dma_page.c @@ -0,0 +1,57 @@ +/* + * Page allocator for DMA + * + * Copyright (c) IBM, Corp. 2020 + * + * Authors: + * Pierre Morel <pmorel@linux.ibm.com> + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU Library General Public License version 2. + */ +#include <libcflat.h> +#include <asm/spinlock.h> +#include <alloc_dma_page.h> + +static struct spinlock lock; +static void *dma_freelist = 0; + +void put_dma_page(void *dma_page) +{ + spin_lock(&lock); + *(void **)dma_page = dma_freelist; + dma_freelist = dma_page; + spin_unlock(&lock); +} + +void *get_dma_page(void) +{ + void *p = NULL; + + spin_lock(&lock); + if (!dma_freelist) + goto end_unlock; + + p = dma_freelist; + dma_freelist = *(void **)dma_freelist; + +end_unlock: + spin_unlock(&lock); + return p; +} + +phys_addr_t dma_page_alloc_init(phys_addr_t start, phys_addr_t end) +{ + int start_pfn = start >> PAGE_SHIFT; + int nb_pfn = ((end - start) >> PAGE_SHIFT) - 1; + int max, pfn; + + max = start_pfn + nb_pfn / DMA_ALLOC_RATIO; + if (max > DMA_MAX_PFN) + max = DMA_MAX_PFN; + + for (pfn = start_pfn; pfn < max; pfn++) + put_dma_page((void *)((unsigned long) pfn << PAGE_SHIFT)); + + return (phys_addr_t)pfn << PAGE_SHIFT; +} diff --git a/lib/alloc_dma_page.h b/lib/alloc_dma_page.h new file mode 100644 index 0000000..85e1d2f --- /dev/null +++ b/lib/alloc_dma_page.h @@ -0,0 +1,24 @@ +/* + * Page allocator for DMA definitions + * + * Copyright (c) IBM, Corp. 2020 + * + * Authors: + * Pierre Morel <pmorel@linux.ibm.com> + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU Library General Public License version 2. + */ +#ifndef _ALLOC_DMA_PAGE_H_ +#define _ALLOC_DMA_PAGE_H_ + +#include <asm/page.h> + +void put_dma_page(void *dma_page); +void *get_dma_page(void); +phys_addr_t dma_page_alloc_init(phys_addr_t start_pfn, phys_addr_t nb_pages); + +#define DMA_MAX_PFN (0x80000000 >> PAGE_SHIFT) +#define DMA_ALLOC_RATIO 8 + +#endif /* _ALLOC_DMA_PAGE_H_ */ diff --git a/lib/s390x/sclp.c b/lib/s390x/sclp.c index 4054d0e..9c95ca5 100644 --- a/lib/s390x/sclp.c +++ b/lib/s390x/sclp.c @@ -19,6 +19,7 @@ #include "sclp.h" #include <alloc_phys.h> #include <alloc_page.h> +#include <alloc_dma_page.h> extern unsigned long stacktop; @@ -35,6 +36,7 @@ static void mem_init(phys_addr_t mem_end) phys_addr_t freemem_start = (phys_addr_t)&stacktop; phys_addr_t base, top; + freemem_start = dma_page_alloc_init(freemem_start, mem_end); phys_alloc_init(freemem_start, mem_end - freemem_start); phys_alloc_get_unused(&base, &top); base = (base + PAGE_SIZE - 1) & -PAGE_SIZE; diff --git a/s390x/Makefile b/s390x/Makefile index 9144d57..109ef9f 100644 --- a/s390x/Makefile +++ b/s390x/Makefile @@ -52,6 +52,7 @@ cflatobjs += lib/alloc_phys.o cflatobjs += lib/alloc_page.o cflatobjs += lib/vmalloc.o cflatobjs += lib/alloc_phys.o +cflatobjs += lib/alloc_dma_page.o cflatobjs += lib/s390x/io.o cflatobjs += lib/s390x/stack.o cflatobjs += lib/s390x/sclp.o
Some architectures need allocations to be done under a specific address limit to allow DMA from I/O. We propose here a very simple page allocator to get pages allocated under this specific limit. The DMA page allocator will only use part of the available memory under the DMA address limit to let room for the standard allocator. Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> --- lib/alloc_dma_page.c | 57 ++++++++++++++++++++++++++++++++++++++++++++ lib/alloc_dma_page.h | 24 +++++++++++++++++++ lib/s390x/sclp.c | 2 ++ s390x/Makefile | 1 + 4 files changed, 84 insertions(+) create mode 100644 lib/alloc_dma_page.c create mode 100644 lib/alloc_dma_page.h