While the first kernel boots, it registers memory regions for fadump
such as:
* CPU state data (has to be populated by the platform)
* HPTE state data (has to be populated by the platform)
* Real Mode Regions (platform should copy it to requested
destination addresses)
* OS defined regions (such as parameter save area)
Platform is also expected to modify the 'bytes_dumped' to the length of
data preserved/copied by platform (ideally same as the source length
passed by kernel).
The kernel passes source address and length for the memory regions, and
a destination address to where the memory is to be copied.
Implement the preserving/copying of the Real Mode Regions and the
Parameter Save Area in QEMU Pseries
Signed-off-by: Aditya Gupta <adityag@linux.ibm.com>
---
hw/ppc/spapr_rtas.c | 117 ++++++++++++++++++++++++++++++++++++++++-
include/hw/ppc/spapr.h | 27 +++++++++-
2 files changed, 142 insertions(+), 2 deletions(-)
diff --git a/hw/ppc/spapr_rtas.c b/hw/ppc/spapr_rtas.c
index 01c82375f03d..9b29cadab2c9 100644
--- a/hw/ppc/spapr_rtas.c
+++ b/hw/ppc/spapr_rtas.c
@@ -344,6 +344,120 @@ static void rtas_ibm_set_system_parameter(PowerPCCPU *cpu,
struct fadump_metadata fadump_metadata;
bool is_next_boot_fadump;
+/* Preserve the memory locations registered for fadump */
+static bool fadump_preserve_mem(void)
+{
+ struct rtas_fadump_mem_struct *fdm = &fadump_metadata.registered_fdm;
+ uint64_t next_section_addr;
+ int dump_num_sections, data_type;
+ uint64_t src_addr, src_len, dest_addr;
+ void *copy_buffer;
+
+ assert(fadump_metadata.fadump_registered);
+ assert(fadump_metadata.fdm_addr != -1);
+
+ /* Read the fadump header passed during fadump registration */
+ cpu_physical_memory_read(fadump_metadata.fdm_addr,
+ &fdm->header, sizeof(fdm->header));
+
+ /* Verify that we understand the fadump header version */
+ if (fdm->header.dump_format_version != cpu_to_be32(FADUMP_VERSION)) {
+ /*
+ * Dump format version is unknown and likely changed from the time
+ * of fadump registration. Back out now.
+ */
+ return false;
+ }
+
+ dump_num_sections = be16_to_cpu(fdm->header.dump_num_sections);
+
+ if (dump_num_sections > FADUMP_MAX_SECTIONS) {
+ qemu_log_mask(LOG_GUEST_ERROR,
+ "FADUMP: Too many sections: %d\n", fdm->header.dump_num_sections);
+ return false;
+ }
+
+ next_section_addr =
+ fadump_metadata.fdm_addr +
+ be32_to_cpu(fdm->header.offset_first_dump_section);
+
+ /*
+ * Handle all sections
+ *
+ * CPU State Data and HPTE regions are handled in their own cases
+ *
+ * RMR regions and any custom OS reserved regions such as parameter
+ * save area, are handled by simply copying the source region to
+ * destination address
+ */
+ for (int i = 0; i < dump_num_sections; ++i) {
+ /* Read the fadump section from memory */
+ cpu_physical_memory_read(next_section_addr,
+ &fdm->rgn[i], sizeof(fdm->rgn[i]));
+
+ next_section_addr += sizeof(fdm->rgn[i]);
+
+ data_type = be16_to_cpu(fdm->rgn[i].source_data_type);
+ src_addr = be64_to_cpu(fdm->rgn[i].source_address);
+ src_len = be64_to_cpu(fdm->rgn[i].source_len);
+ dest_addr = be64_to_cpu(fdm->rgn[i].destination_address);
+
+ /* Reset error_flags & bytes_dumped for now */
+ fdm->rgn[i].error_flags = 0;
+ fdm->rgn[i].bytes_dumped = 0;
+
+ if (be32_to_cpu(fdm->rgn[i].request_flag) != FADUMP_REQUEST_FLAG) {
+ qemu_log_mask(LOG_UNIMP,
+ "FADUMP: Skipping copying region as not requested\n");
+ continue;
+ }
+
+ switch (data_type) {
+ case FADUMP_CPU_STATE_DATA:
+ /* TODO: Add CPU state data */
+ break;
+ case FADUMP_HPTE_REGION:
+ /* TODO: Add hpte state data */
+ break;
+ case FADUMP_REAL_MODE_REGION:
+ case FADUMP_PARAM_AREA:
+ /* Skip copy if source and destination are same (eg. param area) */
+ if (src_addr != dest_addr) {
+ copy_buffer = g_malloc(src_len + 1);
+ if (copy_buffer == NULL) {
+ qemu_log_mask(LOG_GUEST_ERROR,
+ "FADUMP: Failed allocating memory for copying reserved memory regions\n");
+ fdm->rgn[i].error_flags =
+ cpu_to_be16(FADUMP_ERROR_LENGTH_EXCEEDS_SOURCE);
+
+ continue;
+ }
+
+ /* Copy the source region to destination */
+ cpu_physical_memory_read(src_addr, copy_buffer, src_len);
+ cpu_physical_memory_write(dest_addr, copy_buffer, src_len);
+ g_free(copy_buffer);
+ }
+
+ /*
+ * Considering cpu_physical_memory_write would have copied the
+ * complete region
+ */
+ fdm->rgn[i].bytes_dumped = cpu_to_be64(src_len);
+
+ break;
+ default:
+ qemu_log_mask(LOG_GUEST_ERROR,
+ "FADUMP: Skipping unknown source data type: %d\n", data_type);
+
+ fdm->rgn[i].error_flags =
+ cpu_to_be16(FADUMP_ERROR_INVALID_DATA_TYPE);
+ }
+ }
+
+ return true;
+}
+
static void trigger_fadump_boot(target_ulong spapr_retcode)
{
/*
@@ -353,7 +467,8 @@ static void trigger_fadump_boot(target_ulong spapr_retcode)
*/
pause_all_vcpus();
- if (true /* TODO: Preserve memory registered for fadump */) {
+ /* Preserve the memory locations registered for fadump */
+ if (!fadump_preserve_mem()) {
/* Failed to preserve the registered memory regions */
rtas_st(spapr_retcode, 0, RTAS_OUT_HW_ERROR);
diff --git a/include/hw/ppc/spapr.h b/include/hw/ppc/spapr.h
index efa2f891a8a7..a80704187583 100644
--- a/include/hw/ppc/spapr.h
+++ b/include/hw/ppc/spapr.h
@@ -776,7 +776,32 @@ void push_sregs_to_kvm_pr(SpaprMachineState *spapr);
#define FADUMP_CMD_UNREGISTER 2
#define FADUMP_CMD_INVALIDATE 3
-#define FADUMP_VERSION 1
+#define FADUMP_VERSION 1
+
+/*
+ * The Firmware Assisted Dump Memory structure supports a maximum of 10 sections
+ * in the dump memory structure. Presently, three sections are used for
+ * CPU state data, HPTE & Parameters area, while the remaining seven sections
+ * can be used for boot memory regions.
+ */
+#define FADUMP_MAX_SECTIONS 10
+#define RTAS_FADUMP_MAX_BOOT_MEM_REGS 7
+
+/* Firmware provided dump sections */
+#define FADUMP_CPU_STATE_DATA 0x0001
+#define FADUMP_HPTE_REGION 0x0002
+#define FADUMP_REAL_MODE_REGION 0x0011
+
+/* OS defined sections */
+#define FADUMP_PARAM_AREA 0x0100
+
+/* Dump request flag */
+#define FADUMP_REQUEST_FLAG 0x00000001
+
+/* Dump status flag */
+#define FADUMP_ERROR_INVALID_DATA_TYPE 0x8000
+#define FADUMP_ERROR_INVALID_SOURCE_ADDR 0x4000
+#define FADUMP_ERROR_LENGTH_EXCEEDS_SOURCE 0x2000
/*
* The Firmware Assisted Dump Memory structure supports a maximum of 10 sections
--
2.48.1
On 2/17/25 12:47, Aditya Gupta wrote:
> While the first kernel boots, it registers memory regions for fadump
> such as:
> * CPU state data (has to be populated by the platform)
> * HPTE state data (has to be populated by the platform)
> * Real Mode Regions (platform should copy it to requested
> destination addresses)
> * OS defined regions (such as parameter save area)
>
> Platform is also expected to modify the 'bytes_dumped' to the length of
> data preserved/copied by platform (ideally same as the source length
> passed by kernel).
>
> The kernel passes source address and length for the memory regions, and
> a destination address to where the memory is to be copied.
>
> Implement the preserving/copying of the Real Mode Regions and the
> Parameter Save Area in QEMU Pseries
>
> Signed-off-by: Aditya Gupta <adityag@linux.ibm.com>
> ---
> hw/ppc/spapr_rtas.c | 117 ++++++++++++++++++++++++++++++++++++++++-
> include/hw/ppc/spapr.h | 27 +++++++++-
> 2 files changed, 142 insertions(+), 2 deletions(-)
>
> diff --git a/hw/ppc/spapr_rtas.c b/hw/ppc/spapr_rtas.c
> index 01c82375f03d..9b29cadab2c9 100644
> --- a/hw/ppc/spapr_rtas.c
> +++ b/hw/ppc/spapr_rtas.c
> @@ -344,6 +344,120 @@ static void rtas_ibm_set_system_parameter(PowerPCCPU *cpu,
> struct fadump_metadata fadump_metadata;
> bool is_next_boot_fadump;
>
> +/* Preserve the memory locations registered for fadump */
> +static bool fadump_preserve_mem(void)
> +{
> + struct rtas_fadump_mem_struct *fdm = &fadump_metadata.registered_fdm;
> + uint64_t next_section_addr;
> + int dump_num_sections, data_type;
> + uint64_t src_addr, src_len, dest_addr;
> + void *copy_buffer;
> +
> + assert(fadump_metadata.fadump_registered);
> + assert(fadump_metadata.fdm_addr != -1);
> +
> + /* Read the fadump header passed during fadump registration */
> + cpu_physical_memory_read(fadump_metadata.fdm_addr,
> + &fdm->header, sizeof(fdm->header));
> +
> + /* Verify that we understand the fadump header version */
> + if (fdm->header.dump_format_version != cpu_to_be32(FADUMP_VERSION)) {
> + /*
> + * Dump format version is unknown and likely changed from the time
> + * of fadump registration. Back out now.
> + */
> + return false;
> + }
> +
> + dump_num_sections = be16_to_cpu(fdm->header.dump_num_sections);
> +
> + if (dump_num_sections > FADUMP_MAX_SECTIONS) {
> + qemu_log_mask(LOG_GUEST_ERROR,
> + "FADUMP: Too many sections: %d\n", fdm->header.dump_num_sections);
> + return false;
> + }
> +
> + next_section_addr =
> + fadump_metadata.fdm_addr +
> + be32_to_cpu(fdm->header.offset_first_dump_section);
> +
> + /*
> + * Handle all sections
> + *
> + * CPU State Data and HPTE regions are handled in their own cases
> + *
> + * RMR regions and any custom OS reserved regions such as parameter
> + * save area, are handled by simply copying the source region to
> + * destination address
> + */
> + for (int i = 0; i < dump_num_sections; ++i) {
> + /* Read the fadump section from memory */
> + cpu_physical_memory_read(next_section_addr,
> + &fdm->rgn[i], sizeof(fdm->rgn[i]));
> +
> + next_section_addr += sizeof(fdm->rgn[i]);
> +
> + data_type = be16_to_cpu(fdm->rgn[i].source_data_type);
> + src_addr = be64_to_cpu(fdm->rgn[i].source_address);
> + src_len = be64_to_cpu(fdm->rgn[i].source_len);
> + dest_addr = be64_to_cpu(fdm->rgn[i].destination_address);
> +
> + /* Reset error_flags & bytes_dumped for now */
> + fdm->rgn[i].error_flags = 0;
> + fdm->rgn[i].bytes_dumped = 0;
> +
> + if (be32_to_cpu(fdm->rgn[i].request_flag) != FADUMP_REQUEST_FLAG) {
> + qemu_log_mask(LOG_UNIMP,
> + "FADUMP: Skipping copying region as not requested\n");
> + continue;
> + }
> +
> + switch (data_type) {
> + case FADUMP_CPU_STATE_DATA:
> + /* TODO: Add CPU state data */
> + break;
> + case FADUMP_HPTE_REGION:
> + /* TODO: Add hpte state data */
> + break;
> + case FADUMP_REAL_MODE_REGION:
> + case FADUMP_PARAM_AREA:
> + /* Skip copy if source and destination are same (eg. param area) */
> + if (src_addr != dest_addr) {
> + copy_buffer = g_malloc(src_len + 1);
> + if (copy_buffer == NULL) {
> + qemu_log_mask(LOG_GUEST_ERROR,
> + "FADUMP: Failed allocating memory for copying reserved memory regions\n");
> + fdm->rgn[i].error_flags =
> + cpu_to_be16(FADUMP_ERROR_LENGTH_EXCEEDS_SOURCE);
> +
> + continue;
> + }
> +
> + /* Copy the source region to destination */
> + cpu_physical_memory_read(src_addr, copy_buffer, src_len);
> + cpu_physical_memory_write(dest_addr, copy_buffer, src_len);
> + g_free(copy_buffer);
> + }
> +
> + /*
> + * Considering cpu_physical_memory_write would have copied the
> + * complete region
> + */
> + fdm->rgn[i].bytes_dumped = cpu_to_be64(src_len);
Is this really valid for FADUMP_PARAM_AREA where we intend to skip copy?
> +
> + break;
> + default:
> + qemu_log_mask(LOG_GUEST_ERROR,
> + "FADUMP: Skipping unknown source data type: %d\n", data_type);
> +
> + fdm->rgn[i].error_flags =
> + cpu_to_be16(FADUMP_ERROR_INVALID_DATA_TYPE);
> + }
> + }
> +
> + return true;
> +}
> +
> static void trigger_fadump_boot(target_ulong spapr_retcode)
> {
> /*
> @@ -353,7 +467,8 @@ static void trigger_fadump_boot(target_ulong spapr_retcode)
> */
> pause_all_vcpus();
>
> - if (true /* TODO: Preserve memory registered for fadump */) {
> + /* Preserve the memory locations registered for fadump */
> + if (!fadump_preserve_mem()) {
This change can be avoided as suggested in previous patch.
> /* Failed to preserve the registered memory regions */
> rtas_st(spapr_retcode, 0, RTAS_OUT_HW_ERROR);
>
> diff --git a/include/hw/ppc/spapr.h b/include/hw/ppc/spapr.h
> index efa2f891a8a7..a80704187583 100644
> --- a/include/hw/ppc/spapr.h
> +++ b/include/hw/ppc/spapr.h
> @@ -776,7 +776,32 @@ void push_sregs_to_kvm_pr(SpaprMachineState *spapr);
> #define FADUMP_CMD_UNREGISTER 2
> #define FADUMP_CMD_INVALIDATE 3
>
> -#define FADUMP_VERSION 1
> +#define FADUMP_VERSION 1
This change can be avoided if taken care initially.
Thanks
Harsh
> +
> +/*
> + * The Firmware Assisted Dump Memory structure supports a maximum of 10 sections
> + * in the dump memory structure. Presently, three sections are used for
> + * CPU state data, HPTE & Parameters area, while the remaining seven sections
> + * can be used for boot memory regions.
> + */
> +#define FADUMP_MAX_SECTIONS 10
> +#define RTAS_FADUMP_MAX_BOOT_MEM_REGS 7
> +
> +/* Firmware provided dump sections */
> +#define FADUMP_CPU_STATE_DATA 0x0001
> +#define FADUMP_HPTE_REGION 0x0002
> +#define FADUMP_REAL_MODE_REGION 0x0011
> +
> +/* OS defined sections */
> +#define FADUMP_PARAM_AREA 0x0100
> +
> +/* Dump request flag */
> +#define FADUMP_REQUEST_FLAG 0x00000001
> +
> +/* Dump status flag */
> +#define FADUMP_ERROR_INVALID_DATA_TYPE 0x8000
> +#define FADUMP_ERROR_INVALID_SOURCE_ADDR 0x4000
> +#define FADUMP_ERROR_LENGTH_EXCEEDS_SOURCE 0x2000
>
> /*
> * The Firmware Assisted Dump Memory structure supports a maximum of 10 sections
On 05/03/25 12:10, Harsh Prateek Bora wrote:
>
>
> On 2/17/25 12:47, Aditya Gupta wrote:
>> <...snip...>
>>
>> + /* Reset error_flags & bytes_dumped for now */
>> + fdm->rgn[i].error_flags = 0;
>> + fdm->rgn[i].bytes_dumped = 0;
>> +
>> + if (be32_to_cpu(fdm->rgn[i].request_flag) !=
>> FADUMP_REQUEST_FLAG) {
>> + qemu_log_mask(LOG_UNIMP,
>> + "FADUMP: Skipping copying region as not requested\n");
>> + continue;
>> + }
>> +
>> + switch (data_type) {
>> + case FADUMP_CPU_STATE_DATA:
>> + /* TODO: Add CPU state data */
>> + break;
>> + case FADUMP_HPTE_REGION:
>> + /* TODO: Add hpte state data */
>> + break;
>> + case FADUMP_REAL_MODE_REGION:
>> + case FADUMP_PARAM_AREA:
>> + /* Skip copy if source and destination are same (eg.
>> param area) */
>> + if (src_addr != dest_addr) {
>> + copy_buffer = g_malloc(src_len + 1);
>> + if (copy_buffer == NULL) {
>> + qemu_log_mask(LOG_GUEST_ERROR,
>> + "FADUMP: Failed allocating memory for
>> copying reserved memory regions\n");
>> + fdm->rgn[i].error_flags =
>> + cpu_to_be16(FADUMP_ERROR_LENGTH_EXCEEDS_SOURCE);
>> +
>> + continue;
>> + }
>> +
>> + /* Copy the source region to destination */
>> + cpu_physical_memory_read(src_addr, copy_buffer,
>> src_len);
>> + cpu_physical_memory_write(dest_addr, copy_buffer,
>> src_len);
>> + g_free(copy_buffer);
>> + }
>> +
>> + /*
>> + * Considering cpu_physical_memory_write would have
>> copied the
>> + * complete region
>> + */
>> + fdm->rgn[i].bytes_dumped = cpu_to_be64(src_len);
>
> Is this really valid for FADUMP_PARAM_AREA where we intend to skip copy?
>
Yes I think it's good to keep it. Because that's an optimisation i did
to skip the copy if src and dest are same.
But the actual copy depends on the OS passing us the
"FADUMP_REQUEST_FLAG" in the region's request flag.
If the flag is set, i am expecting the kernel asked us to copy, and
hence the .bytes_dumped should be same as the number of bytes asked by
kernel to copy ideally.
If the flag is not set, we return early, so we let the .bytes_dumped be 0.
>> +
>> + break;
>> + default:
>> + qemu_log_mask(LOG_GUEST_ERROR,
>> + "FADUMP: Skipping unknown source data type: %d\n",
>> data_type);
>> +
>> + fdm->rgn[i].error_flags =
>> + cpu_to_be16(FADUMP_ERROR_INVALID_DATA_TYPE);
>> + }
>> + }
>> +
>> + return true;
>> +}
>> +
>> static void trigger_fadump_boot(target_ulong spapr_retcode)
>> {
>> /*
>> @@ -353,7 +467,8 @@ static void trigger_fadump_boot(target_ulong
>> spapr_retcode)
>> */
>> pause_all_vcpus();
>> - if (true /* TODO: Preserve memory registered for fadump */) {
>> + /* Preserve the memory locations registered for fadump */
>> + if (!fadump_preserve_mem()) {
>
> This change can be avoided as suggested in previous patch.
Agreed, will do it in previous patch.
>
>> /* Failed to preserve the registered memory regions */
>> rtas_st(spapr_retcode, 0, RTAS_OUT_HW_ERROR);
>> diff --git a/include/hw/ppc/spapr.h b/include/hw/ppc/spapr.h
>> index efa2f891a8a7..a80704187583 100644
>> --- a/include/hw/ppc/spapr.h
>> +++ b/include/hw/ppc/spapr.h
>> @@ -776,7 +776,32 @@ void push_sregs_to_kvm_pr(SpaprMachineState
>> *spapr);
>> #define FADUMP_CMD_UNREGISTER 2
>> #define FADUMP_CMD_INVALIDATE 3
>> -#define FADUMP_VERSION 1
>> +#define FADUMP_VERSION 1
>
> This change can be avoided if taken care initially.
Nice, I missed this diff. Will fix it in the patch which introduced this.
Thanks,
- Aditya Gupta
>
> Thanks
> Harsh
>> +
>> +/*
>> + * The Firmware Assisted Dump Memory structure supports a maximum of
>> 10 sections
>> + * in the dump memory structure. Presently, three sections are used for
>> + * CPU state data, HPTE & Parameters area, while the remaining seven
>> sections
>> + * can be used for boot memory regions.
>> + */
>> +#define FADUMP_MAX_SECTIONS 10
>> +#define RTAS_FADUMP_MAX_BOOT_MEM_REGS 7
>> +
>> +/* Firmware provided dump sections */
>> +#define FADUMP_CPU_STATE_DATA 0x0001
>> +#define FADUMP_HPTE_REGION 0x0002
>> +#define FADUMP_REAL_MODE_REGION 0x0011
>> +
>> +/* OS defined sections */
>> +#define FADUMP_PARAM_AREA 0x0100
>> +
>> +/* Dump request flag */
>> +#define FADUMP_REQUEST_FLAG 0x00000001
>> +
>> +/* Dump status flag */
>> +#define FADUMP_ERROR_INVALID_DATA_TYPE 0x8000
>> +#define FADUMP_ERROR_INVALID_SOURCE_ADDR 0x4000
>> +#define FADUMP_ERROR_LENGTH_EXCEEDS_SOURCE 0x2000
>> /*
>> * The Firmware Assisted Dump Memory structure supports a maximum
>> of 10 sections
© 2016 - 2026 Red Hat, Inc.