Re: [PATCH v2 1/3] scsi: Fix a scsi_show_rq() NULL pointer dereference
On Mon, Dec 11, 2017 at 10:11:29PM -0500, Martin K. Petersen wrote: > > Hi Ming, > > > This patch allocates one array for T10_PI_TYPE2_PROTECTION command, > > size of each element is SD_EXT_CDB_SIZE, and the length is > > host->can_queue, then we can retrieve one command buffer runtime > > via rq->tag. > > > > So we can avoid to allocate the command buffer runtime, also the > > recent use-after-free report[1] in scsi_show_rq() can be fixed too. > > I'm still mulling over the pros and cons of this one for 4.16+... Hi Martin, This patch can't work in case of real multiple hw queues, but can be fixed without much work. Even we can convert the big allocation into page by page allocation if there is case of huge tag space. Anyway if you think this approach is good, please let me know, and I am happy to cook V2 for review. Thanks, Ming
Re: [PATCH v2 1/3] scsi: Fix a scsi_show_rq() NULL pointer dereference
Hi Ming, > This patch allocates one array for T10_PI_TYPE2_PROTECTION command, > size of each element is SD_EXT_CDB_SIZE, and the length is > host->can_queue, then we can retrieve one command buffer runtime > via rq->tag. > > So we can avoid to allocate the command buffer runtime, also the > recent use-after-free report[1] in scsi_show_rq() can be fixed too. I'm still mulling over the pros and cons of this one for 4.16+... -- Martin K. Petersen Oracle Linux Engineering
Re: [PATCH v2 1/3] scsi: Fix a scsi_show_rq() NULL pointer dereference
Bart, > Avoid that scsi_show_rq() triggers a NULL pointer dereference if > called after sd_uninit_command(). Swap the NULL pointer assignment > and the mempool_free() call in sd_uninit_command() to make it less > likely that scsi_show_rq() triggers a use-after-free. Note: even > with these changes scsi_show_rq() can trigger a use-after-free but > that's a lesser evil than e.g. suppressing debug information for > T10-PI commands completely. This patch fixes the following oops: Applied to 4.15/scsi-fixes. Thanks, Bart. -- Martin K. Petersen Oracle Linux Engineering
Re: [PATCH v2 1/3] scsi: Fix a scsi_show_rq() NULL pointer dereference
Hi Martin, On Fri, Dec 08, 2017 at 04:44:55PM +0800, Ming Lei wrote: > Hi Martin, > > On Thu, Dec 07, 2017 at 09:46:21PM -0500, Martin K. Petersen wrote: > > > > Ming, > > > > > As I explained in [1], the use-after-free is inevitable no matter if > > > clearing 'SCpnt->cmnd' before mempool_free() in sd_uninit_command() or > > > not, so we need to comment the fact that cdb may point to garbage > > > data, and this function(especially __scsi_format_command() has to > > > survive that, so that people won't be surprised when kasan complains > > > use-after-free, and guys will be careful when they try to change the > > > code in future. > > > > Longer term we really need to get rid of the separate CDB allocation. It > > was a necessary evil when I did it. And not much of a concern since I > > did not expect anybody sane to use Type 2 (it's designed for use inside > > disk arrays). > > > > However, I keep hearing about people using Type 2 drives. Some vendors > > source drives formatted that way and use the same SKU for arrays and > > standalone servers. > > > > So we should really look into making it possible for a queue to have a > > bigger than 16-byte built-in CDB. For Type 2 devices, 32-byte reads and > > writes are a prerequisite. So it would be nice to be able to switch a > > queue to a larger allocation post creation (we won't know the type until > > after READ CAPACITY(16) has been sent). > > I am wondering why you don't make __cmd[] in scsi_request defined as 32bytes? > Even for some hosts with thousands of tag, the memory waste is still not > too much. > > Or if you prefer to do post creation, we have sbitmap_queue now, which can > help to build a pre-allocated memory pool easily, and its allocation/free is > pretty efficient. Or something like the following patch? I run several IO tests over scsi_debug(dif=2, dix=1), and looks it works without any problem. >From 7731af623af164c6be451d9c543ce6b70e7e66b8 Mon Sep 17 00:00:00 2001 From: Ming LeiDate: Fri, 8 Dec 2017 17:35:18 +0800 Subject: [PATCH] SCSI: pre-allocate command buffer for T10_PI_TYPE2_PROTECTION This patch allocates one array for T10_PI_TYPE2_PROTECTION command, size of each element is SD_EXT_CDB_SIZE, and the length is host->can_queue, then we can retrieve one command buffer runtime via rq->tag. So we can avoid to allocate the command buffer runtime, also the recent use-after-free report[1] in scsi_show_rq() can be fixed too. [1] https://marc.info/?l=linux-block=151030452715642=2 Signed-off-by: Ming Lei --- drivers/scsi/hosts.c | 1 + drivers/scsi/sd.c| 55 drivers/scsi/sd.h| 4 ++-- drivers/scsi/sd_dif.c| 32 ++-- include/scsi/scsi_host.h | 2 ++ 5 files changed, 49 insertions(+), 45 deletions(-) diff --git a/drivers/scsi/hosts.c b/drivers/scsi/hosts.c index fe3a0da3ec97..74f55b8f16fe 100644 --- a/drivers/scsi/hosts.c +++ b/drivers/scsi/hosts.c @@ -350,6 +350,7 @@ static void scsi_host_dev_release(struct device *dev) if (parent) put_device(parent); + kfree(shost->cmd_ext_buf); kfree(shost); } diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c index 24fe68522716..853eb57ad4ad 100644 --- a/drivers/scsi/sd.c +++ b/drivers/scsi/sd.c @@ -131,9 +131,6 @@ static DEFINE_IDA(sd_index_ida); * object after last put) */ static DEFINE_MUTEX(sd_ref_mutex); -static struct kmem_cache *sd_cdb_cache; -static mempool_t *sd_cdb_pool; - static const char *sd_cache_types[] = { "write through", "none", "write back", "write back, no read (daft)" @@ -1026,6 +1023,13 @@ static int sd_setup_flush_cmnd(struct scsi_cmnd *cmd) return BLKPREP_OK; } +static char *sd_get_ext_buf(struct scsi_device *sdp, struct scsi_cmnd *SCpnt) +{ + struct request *rq = SCpnt->request; + + return >host->cmd_ext_buf[rq->tag * SD_EXT_CDB_SIZE]; +} + static int sd_setup_read_write_cmnd(struct scsi_cmnd *SCpnt) { struct request *rq = SCpnt->request; @@ -1168,12 +1172,7 @@ static int sd_setup_read_write_cmnd(struct scsi_cmnd *SCpnt) protect = 0; if (protect && sdkp->protection_type == T10_PI_TYPE2_PROTECTION) { - SCpnt->cmnd = mempool_alloc(sd_cdb_pool, GFP_ATOMIC); - - if (unlikely(SCpnt->cmnd == NULL)) { - ret = BLKPREP_DEFER; - goto out; - } + SCpnt->cmnd = sd_get_ext_buf(sdp, SCpnt); SCpnt->cmd_len = SD_EXT_CDB_SIZE; memset(SCpnt->cmnd, 0, SCpnt->cmd_len); @@ -1318,12 +1317,6 @@ static void sd_uninit_command(struct scsi_cmnd *SCpnt) if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) __free_page(rq->special_vec.bv_page); - - if (SCpnt->cmnd != scsi_req(rq)->cmd) { - mempool_free(SCpnt->cmnd, sd_cdb_pool); - SCpnt->cmnd = NULL;
Re: [PATCH v2 1/3] scsi: Fix a scsi_show_rq() NULL pointer dereference
Hi Martin, On Thu, Dec 07, 2017 at 09:46:21PM -0500, Martin K. Petersen wrote: > > Ming, > > > As I explained in [1], the use-after-free is inevitable no matter if > > clearing 'SCpnt->cmnd' before mempool_free() in sd_uninit_command() or > > not, so we need to comment the fact that cdb may point to garbage > > data, and this function(especially __scsi_format_command() has to > > survive that, so that people won't be surprised when kasan complains > > use-after-free, and guys will be careful when they try to change the > > code in future. > > Longer term we really need to get rid of the separate CDB allocation. It > was a necessary evil when I did it. And not much of a concern since I > did not expect anybody sane to use Type 2 (it's designed for use inside > disk arrays). > > However, I keep hearing about people using Type 2 drives. Some vendors > source drives formatted that way and use the same SKU for arrays and > standalone servers. > > So we should really look into making it possible for a queue to have a > bigger than 16-byte built-in CDB. For Type 2 devices, 32-byte reads and > writes are a prerequisite. So it would be nice to be able to switch a > queue to a larger allocation post creation (we won't know the type until > after READ CAPACITY(16) has been sent). I am wondering why you don't make __cmd[] in scsi_request defined as 32bytes? Even for some hosts with thousands of tag, the memory waste is still not too much. Or if you prefer to do post creation, we have sbitmap_queue now, which can help to build a pre-allocated memory pool easily, and its allocation/free is pretty efficient. Thanks, Ming
Re: [PATCH v2 1/3] scsi: Fix a scsi_show_rq() NULL pointer dereference
Ming, > As I explained in [1], the use-after-free is inevitable no matter if > clearing 'SCpnt->cmnd' before mempool_free() in sd_uninit_command() or > not, so we need to comment the fact that cdb may point to garbage > data, and this function(especially __scsi_format_command() has to > survive that, so that people won't be surprised when kasan complains > use-after-free, and guys will be careful when they try to change the > code in future. Longer term we really need to get rid of the separate CDB allocation. It was a necessary evil when I did it. And not much of a concern since I did not expect anybody sane to use Type 2 (it's designed for use inside disk arrays). However, I keep hearing about people using Type 2 drives. Some vendors source drives formatted that way and use the same SKU for arrays and standalone servers. So we should really look into making it possible for a queue to have a bigger than 16-byte built-in CDB. For Type 2 devices, 32-byte reads and writes are a prerequisite. So it would be nice to be able to switch a queue to a larger allocation post creation (we won't know the type until after READ CAPACITY(16) has been sent). Last I looked at this it was not entirely trivial given how we tag things on to the end. But that really is my preferred fix. -- Martin K. Petersen Oracle Linux Engineering
Re: [PATCH v2 1/3] scsi: Fix a scsi_show_rq() NULL pointer dereference
On Tue, Dec 05, 2017 at 04:57:51PM -0800, Bart Van Assche wrote: > Avoid that scsi_show_rq() triggers a NULL pointer dereference if > called after sd_uninit_command(). Swap the NULL pointer assignment > and the mempool_free() call in sd_uninit_command() to make it less > likely that scsi_show_rq() triggers a use-after-free. Note: even > with these changes scsi_show_rq() can trigger a use-after-free but > that's a lesser evil than e.g. suppressing debug information for > T10-PI commands completely. This patch fixes the following oops: > > BUG: unable to handle kernel NULL pointer dereference at (null) > IP: scsi_format_opcode_name+0x1a/0x1c0 > CPU: 1 PID: 1881 Comm: cat Not tainted 4.14.0-rc2.blk_mq_io_hang+ #516 > Call Trace: > __scsi_format_command+0x27/0xc0 > scsi_show_rq+0x5c/0xc0 > __blk_mq_debugfs_rq_show+0x116/0x130 > blk_mq_debugfs_rq_show+0xe/0x10 > seq_read+0xfe/0x3b0 > full_proxy_read+0x54/0x90 > __vfs_read+0x37/0x160 > vfs_read+0x96/0x130 > SyS_read+0x55/0xc0 > entry_SYSCALL_64_fastpath+0x1a/0xa5 > > Fixes: 0eebd005dd07 ("scsi: Implement blk_mq_ops.show_rq()") > Reported-by: Ming Lei> Signed-off-by: Bart Van Assche > Cc: James E.J. Bottomley > Cc: Martin K. Petersen > Cc: Ming Lei > Cc: Christoph Hellwig > Cc: Hannes Reinecke > Cc: Johannes Thumshirn > Cc: sta...@vger.kernel.org > --- > drivers/scsi/scsi_debugfs.c | 6 -- > drivers/scsi/sd.c | 4 +++- > 2 files changed, 7 insertions(+), 3 deletions(-) > > diff --git a/drivers/scsi/scsi_debugfs.c b/drivers/scsi/scsi_debugfs.c > index 01f08c03f2c1..c3765d29fd3f 100644 > --- a/drivers/scsi/scsi_debugfs.c > +++ b/drivers/scsi/scsi_debugfs.c > @@ -8,9 +8,11 @@ void scsi_show_rq(struct seq_file *m, struct request *rq) > { > struct scsi_cmnd *cmd = container_of(scsi_req(rq), typeof(*cmd), req); > int msecs = jiffies_to_msecs(jiffies - cmd->jiffies_at_alloc); > - char buf[80]; > + const u8 *const cdb = READ_ONCE(cmd->cmnd); > + char buf[80] = "(?)"; > > - __scsi_format_command(buf, sizeof(buf), cmd->cmnd, cmd->cmd_len); > + if (cdb) > + __scsi_format_command(buf, sizeof(buf), cdb, cmd->cmd_len); > seq_printf(m, ", .cmd=%s, .retries=%d, allocated %d.%03d s ago", buf, > cmd->retries, msecs / 1000, msecs % 1000); > } As I explained in [1], the use-after-free is inevitable no matter if clearing 'SCpnt->cmnd' before mempool_free() in sd_uninit_command() or not, so we need to comment the fact that cdb may point to garbage data, and this function(especially __scsi_format_command() has to survive that, so that people won't be surprised when kasan complains use-after-free, and guys will be careful when they try to change the code in future. Once this comment is added, with or without clearing 'SCpnt->cmnd' before mempool_free(), I am fine with this patch. [1] https://marc.info/?l=linux-block=151252302112512=2 > diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c > index d175c5c5ccf8..d841743b2107 100644 > --- a/drivers/scsi/sd.c > +++ b/drivers/scsi/sd.c > @@ -1284,6 +1284,7 @@ static int sd_init_command(struct scsi_cmnd *cmd) > static void sd_uninit_command(struct scsi_cmnd *SCpnt) > { > struct request *rq = SCpnt->request; > + u8 *cmnd; > > if (SCpnt->flags & SCMD_ZONE_WRITE_LOCK) > sd_zbc_write_unlock_zone(SCpnt); > @@ -1292,9 +1293,10 @@ static void sd_uninit_command(struct scsi_cmnd *SCpnt) > __free_page(rq->special_vec.bv_page); > > if (SCpnt->cmnd != scsi_req(rq)->cmd) { > - mempool_free(SCpnt->cmnd, sd_cdb_pool); > + cmnd = SCpnt->cmnd; > SCpnt->cmnd = NULL; > SCpnt->cmd_len = 0; > + mempool_free(cmnd, sd_cdb_pool); > } > } > > -- > 2.15.0 > -- Ming
[PATCH v2 1/3] scsi: Fix a scsi_show_rq() NULL pointer dereference
Avoid that scsi_show_rq() triggers a NULL pointer dereference if called after sd_uninit_command(). Swap the NULL pointer assignment and the mempool_free() call in sd_uninit_command() to make it less likely that scsi_show_rq() triggers a use-after-free. Note: even with these changes scsi_show_rq() can trigger a use-after-free but that's a lesser evil than e.g. suppressing debug information for T10-PI commands completely. This patch fixes the following oops: BUG: unable to handle kernel NULL pointer dereference at (null) IP: scsi_format_opcode_name+0x1a/0x1c0 CPU: 1 PID: 1881 Comm: cat Not tainted 4.14.0-rc2.blk_mq_io_hang+ #516 Call Trace: __scsi_format_command+0x27/0xc0 scsi_show_rq+0x5c/0xc0 __blk_mq_debugfs_rq_show+0x116/0x130 blk_mq_debugfs_rq_show+0xe/0x10 seq_read+0xfe/0x3b0 full_proxy_read+0x54/0x90 __vfs_read+0x37/0x160 vfs_read+0x96/0x130 SyS_read+0x55/0xc0 entry_SYSCALL_64_fastpath+0x1a/0xa5 Fixes: 0eebd005dd07 ("scsi: Implement blk_mq_ops.show_rq()") Reported-by: Ming LeiSigned-off-by: Bart Van Assche Cc: James E.J. Bottomley Cc: Martin K. Petersen Cc: Ming Lei Cc: Christoph Hellwig Cc: Hannes Reinecke Cc: Johannes Thumshirn Cc: sta...@vger.kernel.org --- drivers/scsi/scsi_debugfs.c | 6 -- drivers/scsi/sd.c | 4 +++- 2 files changed, 7 insertions(+), 3 deletions(-) diff --git a/drivers/scsi/scsi_debugfs.c b/drivers/scsi/scsi_debugfs.c index 01f08c03f2c1..c3765d29fd3f 100644 --- a/drivers/scsi/scsi_debugfs.c +++ b/drivers/scsi/scsi_debugfs.c @@ -8,9 +8,11 @@ void scsi_show_rq(struct seq_file *m, struct request *rq) { struct scsi_cmnd *cmd = container_of(scsi_req(rq), typeof(*cmd), req); int msecs = jiffies_to_msecs(jiffies - cmd->jiffies_at_alloc); - char buf[80]; + const u8 *const cdb = READ_ONCE(cmd->cmnd); + char buf[80] = "(?)"; - __scsi_format_command(buf, sizeof(buf), cmd->cmnd, cmd->cmd_len); + if (cdb) + __scsi_format_command(buf, sizeof(buf), cdb, cmd->cmd_len); seq_printf(m, ", .cmd=%s, .retries=%d, allocated %d.%03d s ago", buf, cmd->retries, msecs / 1000, msecs % 1000); } diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c index d175c5c5ccf8..d841743b2107 100644 --- a/drivers/scsi/sd.c +++ b/drivers/scsi/sd.c @@ -1284,6 +1284,7 @@ static int sd_init_command(struct scsi_cmnd *cmd) static void sd_uninit_command(struct scsi_cmnd *SCpnt) { struct request *rq = SCpnt->request; + u8 *cmnd; if (SCpnt->flags & SCMD_ZONE_WRITE_LOCK) sd_zbc_write_unlock_zone(SCpnt); @@ -1292,9 +1293,10 @@ static void sd_uninit_command(struct scsi_cmnd *SCpnt) __free_page(rq->special_vec.bv_page); if (SCpnt->cmnd != scsi_req(rq)->cmd) { - mempool_free(SCpnt->cmnd, sd_cdb_pool); + cmnd = SCpnt->cmnd; SCpnt->cmnd = NULL; SCpnt->cmd_len = 0; + mempool_free(cmnd, sd_cdb_pool); } } -- 2.15.0