[PATCH] kexec-tools: ppc64: avoid adding coherent memory regions to crash memory ranges
Alistair Popple
alistair at popple.id.au
Thu Aug 17 21:56:30 PDT 2017
Hi,
I haven't reviewed the code in depth but skipping memory regions marked
ibm,coherent-device-memory is the right thing to do as the links to this memory
will usually go down prior during a kexec.
Also device-drivers typically put this memory into ZONE_MOVABLE which means
there will be no kernel allocations present in this memory.
Regards,
Alistair
On Thu, 17 Aug 2017 06:01:51 PM Hari Bathini wrote:
> Accelerator devices like GPU and FPGA cards contain onboard memory. This
> onboard memory is represented as a memory only NUMA node, integrating it
> with core memory subsystem. Since, the link through which these devices
> are integrated to core memory goes down after a system crash and they are
> meant for user workloads, avoid adding coherent device memory regions to
> crash memory ranges. Without this change, makedumpfile tool tries to save
> unaccessible coherent device memory regions, crashing the system.
>
> Signed-off-by: Hari Bathini <hbathini at linux.vnet.ibm.com>
> ---
> kexec/arch/ppc64/crashdump-ppc64.c | 64 +++++++++++++++++++++++++++++++++++-
> kexec/arch/ppc64/kexec-ppc64.h | 1 +
> 2 files changed, 63 insertions(+), 2 deletions(-)
>
> diff --git a/kexec/arch/ppc64/crashdump-ppc64.c b/kexec/arch/ppc64/crashdump-ppc64.c
> index 13995bf..7ea3983 100644
> --- a/kexec/arch/ppc64/crashdump-ppc64.c
> +++ b/kexec/arch/ppc64/crashdump-ppc64.c
> @@ -181,6 +181,53 @@ static int get_dyn_reconf_crash_memory_ranges(void)
> return 0;
> }
>
> +/*
> + * For a given memory node, check if it is mapped to system RAM or
> + * to onboard memory on accelerator device like GPU card or such.
> + */
> +static int is_coherent_device_mem(const char *fname)
> +{
> + char fpath[PATH_LEN];
> + char buf[32];
> + DIR *dmem;
> + FILE *file;
> + struct dirent *mentry;
> + int cnt, ret = 0;
> +
> + strcpy(fpath, fname);
> + if ((dmem = opendir(fpath)) == NULL) {
> + perror(fpath);
> + return -1;
> + }
> +
> + while ((mentry = readdir(dmem)) != NULL) {
> + if (strcmp(mentry->d_name, "compatible"))
> + continue;
> +
> + strcat(fpath, "/compatible");
> + if ((file = fopen(fpath, "r")) == NULL) {
> + perror(fpath);
> + ret = -1;
> + break;
> + }
> + if ((cnt = fread(buf, 1, 32, file)) < 0) {
> + perror(fpath);
> + fclose(file);
> + ret = -1;
> + break;
> + }
> + if (!strncmp(buf, "ibm,coherent-device-memory", 26)) {
> + ret = 1;
> + break;
> + }
> + fclose(file);
> + }
> +
> + closedir(dmem);
> + return ret;
> +}
> +
> +
> /* Reads the appropriate file and retrieves the SYSTEM RAM regions for whom to
> * create Elf headers. Keeping it separate from get_memory_ranges() as
> * requirements are different in the case of normal kexec and crashdumps.
> @@ -196,12 +243,12 @@ static int get_crash_memory_ranges(struct memory_range **range, int *ranges)
> {
>
> char device_tree[256] = "/proc/device-tree/";
> - char fname[256];
> + char fname[PATH_LEN];
> char buf[MAXBYTES];
> DIR *dir, *dmem;
> FILE *file;
> struct dirent *dentry, *mentry;
> - int n, crash_rng_len = 0;
> + int n, ret, crash_rng_len = 0;
> unsigned long long start, end;
> int page_size;
>
> @@ -240,6 +287,19 @@ static int get_crash_memory_ranges(struct memory_range **range, int *ranges)
> continue;
> strcpy(fname, device_tree);
> strcat(fname, dentry->d_name);
> +
> + ret = is_coherent_device_mem(fname);
> + if (ret == -1) {
> + closedir(dir);
> + goto err;
> + } else if (ret == 1) {
> + /*
> + * Avoid adding this memory region as it is not
> + * mapped to system RAM.
> + */
> + continue;
> + }
> +
> if ((dmem = opendir(fname)) == NULL) {
> perror(fname);
> closedir(dir);
> diff --git a/kexec/arch/ppc64/kexec-ppc64.h b/kexec/arch/ppc64/kexec-ppc64.h
> index 633ae77..434b4bf 100644
> --- a/kexec/arch/ppc64/kexec-ppc64.h
> +++ b/kexec/arch/ppc64/kexec-ppc64.h
> @@ -1,6 +1,7 @@
> #ifndef KEXEC_PPC64_H
> #define KEXEC_PPC64_H
>
> +#define PATH_LEN 256
> #define MAXBYTES 128
> #define MAX_LINE 160
> #define CORE_TYPE_ELF32 1
>
More information about the kexec
mailing list