On Fri, 15 Nov 2024 at 16:28, Simon Glass <[email protected]> wrote: > > Hi Ilias, > > On Fri, 15 Nov 2024 at 06:58, Ilias Apalodimas > <[email protected]> wrote: > > > > > > > > On Fri, 15 Nov 2024 at 15:54, Ilias Apalodimas > > <[email protected]> wrote: > >> > >> > >> > >> On Fri, 15 Nov 2024 at 15:53, Simon Glass <[email protected]> wrote: > >>> > >>> Hi Ilias, > >>> > >>> On Thu, 14 Nov 2024 at 00:04, Ilias Apalodimas > >>> <[email protected]> wrote: > >>> > > >>> > Hi Simon, > >>> > > >>> > On Tue, 12 Nov 2024 at 16:11, Simon Glass <[email protected]> wrote: > >>> > > > >>> > > The cache-flush function is incorrect which causes a crash in the > >>> > > remoteproc tests with arm64. > >>> > > > >>> > > Fix both problems by using map_sysmem() to convert an address to a > >>> > > pointer and map_to_sysmem() to convert a pointer to an address. > >>> > > > >>> > > Also update the image-loader's cache-flushing logic. > >>> > > > >>> > > Signed-off-by: Simon Glass <[email protected]> > >>> > > Fixes: 3286d223fd7 ("sandbox: implement invalidate_icache_all()") > >>> > > --- > >>> > > > >>> > > (no changes since v2) > >>> > > > >>> > > Changes in v2: > >>> > > - Drop message about EFI_LOADER > >>> > > > >>> > > arch/sandbox/cpu/cache.c | 8 +++++++- > >>> > > drivers/remoteproc/rproc-elf-loader.c | 18 +++++++++++------- > >>> > > lib/efi_loader/efi_image_loader.c | 3 ++- > >>> > > 3 files changed, 20 insertions(+), 9 deletions(-) > >>> > > > >>> > > diff --git a/arch/sandbox/cpu/cache.c b/arch/sandbox/cpu/cache.c > >>> > > index c8a5e64214b..96b3da47e8e 100644 > >>> > > --- a/arch/sandbox/cpu/cache.c > >>> > > +++ b/arch/sandbox/cpu/cache.c > >>> > > @@ -4,12 +4,18 @@ > >>> > > */ > >>> > > > >>> > > #include <cpu_func.h> > >>> > > +#include <mapmem.h> > >>> > > #include <asm/state.h> > >>> > > > >>> > > void flush_cache(unsigned long addr, unsigned long size) > >>> > > { > >>> > > + void *ptr; > >>> > > + > >>> > > + ptr = map_sysmem(addr, size); > >>> > > + > >>> > > /* Clang uses (char *) parameters, GCC (void *) */ > >>> > > - __builtin___clear_cache((void *)addr, (void *)(addr + size)); > >>> > > + __builtin___clear_cache(map_sysmem(addr, size), ptr + size); > >>> > > + unmap_sysmem(ptr); > >>> > > } > >>> > > > >>> > > void invalidate_icache_all(void) > >>> > > diff --git a/drivers/remoteproc/rproc-elf-loader.c > >>> > > b/drivers/remoteproc/rproc-elf-loader.c > >>> > > index ab1836b3f07..0b3941b7798 100644 > >>> > > --- a/drivers/remoteproc/rproc-elf-loader.c > >>> > > +++ b/drivers/remoteproc/rproc-elf-loader.c > >>> > > @@ -6,6 +6,7 @@ > >>> > > #include <dm.h> > >>> > > #include <elf.h> > >>> > > #include <log.h> > >>> > > +#include <mapmem.h> > >>> > > #include <remoteproc.h> > >>> > > #include <asm/cache.h> > >>> > > #include <dm/device_compat.h> > >>> > > @@ -180,6 +181,7 @@ int rproc_elf32_load_image(struct udevice *dev, > >>> > > unsigned long addr, ulong size) > >>> > > for (i = 0; i < ehdr->e_phnum; i++, phdr++) { > >>> > > void *dst = (void *)(uintptr_t)phdr->p_paddr; > >>> > > void *src = (void *)addr + phdr->p_offset; > >>> > > + ulong dst_addr; > >>> > > > >>> > > if (phdr->p_type != PT_LOAD) > >>> > > continue; > >>> > > @@ -195,10 +197,11 @@ int rproc_elf32_load_image(struct udevice *dev, > >>> > > unsigned long addr, ulong size) > >>> > > if (phdr->p_filesz != phdr->p_memsz) > >>> > > memset(dst + phdr->p_filesz, 0x00, > >>> > > phdr->p_memsz - phdr->p_filesz); > >>> > > - flush_cache(rounddown((unsigned long)dst, > >>> > > ARCH_DMA_MINALIGN), > >>> > > - roundup((unsigned long)dst + > >>> > > phdr->p_filesz, > >>> > > + dst_addr = map_to_sysmem(dst); > >>> > > + flush_cache(rounddown(dst_addr, ARCH_DMA_MINALIGN), > >>> > > + roundup(dst_addr + phdr->p_filesz, > >>> > > ARCH_DMA_MINALIGN) - > >>> > > - rounddown((unsigned long)dst, > >>> > > ARCH_DMA_MINALIGN)); > >>> > > + rounddown(dst_addr, ARCH_DMA_MINALIGN)); > >>> > > } > >>> > > > >>> > > return 0; > >>> > > @@ -377,6 +380,7 @@ int rproc_elf32_load_rsc_table(struct udevice > >>> > > *dev, ulong fw_addr, > >>> > > const struct dm_rproc_ops *ops; > >>> > > Elf32_Shdr *shdr; > >>> > > void *src, *dst; > >>> > > + ulong dst_addr; > >>> > > > >>> > > shdr = rproc_elf32_find_rsc_table(dev, fw_addr, fw_size); > >>> > > if (!shdr) > >>> > > @@ -398,10 +402,10 @@ int rproc_elf32_load_rsc_table(struct udevice > >>> > > *dev, ulong fw_addr, > >>> > > (ulong)dst, *rsc_size); > >>> > > > >>> > > memcpy(dst, src, *rsc_size); > >>> > > - flush_cache(rounddown((unsigned long)dst, ARCH_DMA_MINALIGN), > >>> > > - roundup((unsigned long)dst + *rsc_size, > >>> > > - ARCH_DMA_MINALIGN) - > >>> > > - rounddown((unsigned long)dst, ARCH_DMA_MINALIGN)); > >>> > > + dst_addr = map_to_sysmem(dst); > >>> > > + flush_cache(rounddown(dst_addr, ARCH_DMA_MINALIGN), > >>> > > + roundup(dst_addr + *rsc_size, ARCH_DMA_MINALIGN) - > >>> > > + rounddown(dst_addr, ARCH_DMA_MINALIGN)); > >>> > > > >>> > > return 0; > >>> > > } > >>> > > diff --git a/lib/efi_loader/efi_image_loader.c > >>> > > b/lib/efi_loader/efi_image_loader.c > >>> > > index 0ddf69a0918..bb58cf1badb 100644 > >>> > > --- a/lib/efi_loader/efi_image_loader.c > >>> > > +++ b/lib/efi_loader/efi_image_loader.c > >>> > > @@ -13,6 +13,7 @@ > >>> > > #include <efi_loader.h> > >>> > > #include <log.h> > >>> > > #include <malloc.h> > >>> > > +#include <mapmem.h> > >>> > > #include <pe.h> > >>> > > #include <sort.h> > >>> > > #include <crypto/mscode.h> > >>> > > @@ -977,7 +978,7 @@ efi_status_t efi_load_pe(struct > >>> > > efi_loaded_image_obj *handle, > >>> > > } > >>> > > > >>> > > /* Flush cache */ > >>> > > - flush_cache((ulong)efi_reloc, > >>> > > + flush_cache(map_to_sysmem(efi_reloc), > >>> > > >>> > Shouldn't we now umap that address on the err: tag? > >>> > >>> This address should be in sandbox's DRAM, so there is no mapping > >>> created. The mapping is needed for things that are not in the emulated > >>> DRAM, such as stack variables. > >> > >> > >> Isn't there a ref counter that gets increased for these though? > >> Why would we not want to decrease it properly? > > > > > > Ah scratch that, the refcnt is not for the sandbox DRAM. But in any case, I > > don't like seeing that magic of people knowing what they should unmap or > > not. > > I would prefer calling the unmap call unconditionally, since the function > > deals with sandbox memory correctly already > > Yes, that's OK with me and is clearer, as you say.
Thanks Simon, with the unmap call added Reviewed-by: Ilias Apalodimas <[email protected]> > > Regards, > Simon
